Citation
Advancing the Understanding of Healthcare Utilization and Costs among Patients with Musculoskeletal Pain

Material Information

Title:
Advancing the Understanding of Healthcare Utilization and Costs among Patients with Musculoskeletal Pain
Creator:
Lentz, Trevor A
Publisher:
University of Florida
Publication Date:
Language:
English

Thesis/Dissertation Information

Degree:
Doctorate ( Ph.D.)
Degree Grantor:
University of Florida
Degree Disciplines:
Rehabilitation Science
Committee Chair:
GEORGE,STEVEN
Committee Co-Chair:
BENECIUK,JASON M
Committee Members:
MARLOW,NICOLE MARGUERITE
FILLINGIM,ROGER BENTON

Subjects

Subjects / Keywords:
musculoskeletal
pain
rehabilitation
value

Notes

General Note:
Musculoskeletal pain is a prevalent condition with far-reaching public health and economic effects. Treatment pathways that more effectively manage high impact musculoskeletal pain are needed, however must be sensitive to value-based healthcare reform. Individuals at risk for low value care include those with high levels of healthcare utilization or complex medical needs. These individuals represent high priority targets for comparative effectiveness research, but limited knowledge of their characteristics and healthcare utilization patterns has constrained research progress. Therefore, the overall goals of these dissertation projects are to use three different datasets to 1) identify characteristics of populations with musculoskeletal pain at risk for low value care due to high healthcare costs and/or complex medical needs, and 2) define healthcare utilization patterns for these populations. The first project used data from the Medical Expenditure Panel Survey (MEPS) to identify individuals at risk of persistently high pain-related healthcare expenditures over the 2 years of the survey. Robust predictors of persistently high expenditures were older age, greater missed work days, greater pain interference, having private insurance, having a musculoskeletal injury diagnosis, number of musculoskeletal diagnoses, and use of prescription medication for pain. The second project identified predictors of additional healthcare utilization following an episode of physical therapy for musculoskeletal pain. This study used data from the OPT-IN OSPRO study, a longitudinal cohort study of patients seeking outpatient physical therapy. Intensity of pain and its course over the early phases of rehabilitation were important predictors of subsequent healthcare utilization in the year following physical therapy. Psychological factors, disability, and comorbidities predicted use of specific services. The third project used data from the Medicare Current Beneficiary Survey (MCBS) with matching claims to identify co-morbidity subgroups and patterns of healthcare utilization among individuals with osteoarthritis or low back pain. Latent class analysis identified 5 subgroups that were similar between pain conditions: low comorbidity, pulmonary/hypothyroidism/anemia, hypertension with diabetes, complicated hypertension/renal/anemia, and complex cardiac/high comorbidity. Greater disability and higher costs were associated with higher disease burden (defined by both number and combination of specific comorbidities) profiles. Implications of each projects' findings for clinical practice, research and policy are discussed.

Record Information

Source Institution:
UFRGP
Rights Management:
All applicable rights reserved by the source institution and holding location.
Embargo Date:
8/31/2018

Downloads

This item has the following downloads:


Full Text

PAGE 1

ADVANCING THE UNDERSTANDING OF HEALTHCARE UTILIZATION AND COSTS AMONG PATIENTS WITH MUSCULOSKELETAL PAIN By TREVOR ANTHONY LENTZ A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2017

PAGE 2

2017 Trevor Anthony Lentz

PAGE 3

To Kristen, Madeline and Emma, f or being by my side every step of the way.

PAGE 4

ACKNOWLEDGMENTS I wish to thank my parents for teaching me the value of hard work and perseverance. Thank you to Dr. Steven George for nourishing an inquisitive spirit many years ago. He helped me develop the tools to succeed and gave me the freedom to explore. His vision, humility and productivity have set an example that I will continue to follow thr oughout my career I am fortunate to call him a colleague, mentor and friend. I would like to thank Drs. Mark Bishop and Terese Chmielewski for starting me down the path of clinical research. They opened many doors that have led me to where I am today. Tha nk you to my dissertation committee, Drs. Jason Beneciuk, Nic ole Marlow and Roger Fillingim, for their guidance and consultation as I wandered into new and exciting territory with my research. I would also like to thank the UF Pain S quad, especially Joel Bialosky, Carolina Valencia, Roy Coronado, Corey Simon, Katie Butera, Meryl Alapp attu and Chaz Penza for their invaluable advice, friendship and support. I would like to acknowledge the University of Florida, Foundation for Physical Therapy and Brooks Rehabilitation for their generous support of my training over the past 4 years. I would like to thank my daughters, Madeline and Emma. With a song or a dance, they could provide much needed perspective and the motivation to just keep swimming Finally, I wou ld like to thank my wife, Kristen. She selflessly supported my dream in every way imaginable. She is the reason this was possible. 4

PAGE 5

TABLE OF CONTENTS page ACKNOWLEDGMENTS .................................................................................................. 4 LIST OF TABLES ............................................................................................................ 8 LIST OF FIGURES ........................................................................................................ 10 ABSTRACT ................................................................................................................... 11 CHAPTER 1 IMPACT OF MUSCULOSKELETAL PAIN AND THE NEED TO IMPROVE VALUE OF PAIN CARE .......................................................................................... 13 Personal and Economic Burden of Musculoskeletal Pain ....................................... 13 Public Health Crisis of Pain ..................................................................................... 15 Impact #1: Increasing Strain on Healthcare Resources ................................... 15 Impact #2: High Concentration of Utilization and Spending in Vulnerable Po pulations with Complex Conditions ........................................................... 16 Impact #3: Avoidable Morbidity and Mortality Associated with Opioid Use ...... 17 Changing Economic Conditions in Healthcare and Implications for the Treatment of Musculoskeletal Pain ...................................................................... 18 Value Model for Physical Therapy .......................................................................... 20 Using the Value Model to Inform Comparative Effectiveness Research for Musculoskeletal Pain ........................................................................................... 22 2 PREDICTORS OF PERSISTENTLY HIGH COST HEALTHCARE UTILIZATION FOR MUSCULOSKELETAL PAIN .......................................................................... 26 Introduction ............................................................................................................. 26 Methods .................................................................................................................. 30 Dataset ............................................................................................................. 30 Study Sample ................................................................................................... 30 Expenditure Summaries ................................................................................... 32 Predictor Variables ........................................................................................... 33 Statistical Analysis ............................................................................................ 38 Results .................................................................................................................... 41 Descriptive Analysis ......................................................................................... 41 Univariate Analysis ........................................................................................... 42 Multivariate Analysis ......................................................................................... 43 Cluster Analysis ................................................................................................ 46 Discriminant Function Analysis ......................................................................... 47 Cluster Comparisons ........................................................................................ 48 Discussion .............................................................................................................. 48 Future Research ............................................................................................... 55 5

PAGE 6

3 RISK MODEL DEVELOPMENT FOR ADDITIONAL HEALTHCARE USE FOLLOWING AN EPISODE OF PHYSICAL THERAPY FOR MUSCULOSKELETAL PAIN .................................................................................. 82 Introduction ............................................................................................................. 82 Methods .................................................................................................................. 85 Dataset ............................................................................................................. 85 Patient Population ............................................................................................ 86 Healthcare Utilization Predictors ...................................................................... 87 Intervention ....................................................................................................... 90 Healthcare Utilization Outcomes ...................................................................... 90 Statistical Analysis ............................................................................................ 91 Sample Size ..................................................................................................... 94 Results .................................................................................................................... 95 Utilization of Addit ional Healthcare Services over 12 Months ........................... 96 Intensity of Utilization Reported at 6Month and 12Month Follow Up .............. 96 Utilization of Opioids ......................................................................................... 97 Utilization of Injection. ....................................................................................... 98 Ut ilization of Surgery ........................................................................................ 99 Utilization of Diagnostic Tests or Imaging. ........................................................ 99 Utilization of Emergency Room ...................................................................... 100 Results of Imputed Models and IPAW Analyses ............................................ 101 Discussion ............................................................................................................ 102 4 IDENTIFICATION OF COMORBIDITY SUBGROUPS AMONG OLDER ADULTS SEEKING HEALTHCARE FOR MUSCULOSKELETAL ........................ 120 Methods ................................................................................................................ 123 Dataset ........................................................................................................... 123 Analytic File Development .............................................................................. 124 Subjects .......................................................................................................... 124 Identification of Diagnoses ............................................................................. 125 Assignment of Comorbidities .......................................................................... 1 25 Sociodemographic Information ....................................................................... 126 Health Related Information ............................................................................. 126 Healthcare Expenditures ................................................................................ 127 Statistical Analysis .......................................................................................... 127 Results .................................................................................................................. 130 Subgroup Analysis ......................................................................................... 130 Subgroup Comparisons on Patient Reported Function and Expenditures ..... 135 Discussion ............................................................................................................ 137 Future Research ............................................................................................. 145 5 CONCLUSION ...................................................................................................... 163 Summary of Findings and Future Directions ......................................................... 163 6

PAGE 7

Persistent High Cost Healthcare Utilization is Concentrated among a Small Percentage of Individuals with Musculoskeletal Pain .................................. 163 Comorbidity Complexity is an Important Contributor to Utilization and Costs for Musculoskeletal Pain ............................................................................. 164 Self Reported Health Measures and Change in Symptoms over Time Influence Costs and Utilization, but Are Under Represented in Survey and Claims Data ................................................................................................. 165 Psychological Factors Do Not Contribute Substantially to Healthcare Costs and Utilization After Considering Other Predisposing and Enabling Factors ........................................................................................................ 166 A Considerable Proportion of Patients Report Additional Healthcare Utilization after an Episode of Physical Therapy for Musculoskeletal Pain. 167 APPENDIX A LIST OF MUSCULOSKELETAL CONDITION ICD9 CODES .............................. 169 B DEFINITION AND ANALY SIS CODING OF CATEGORICAL VARIABLES .......... 172 LIST OF REFERENCES ............................................................................................. 176 BIOGRAPHICAL SKETCH .......................................................................................... 195 7

PAGE 8

LIST OF TABLES Table page 2 1 Predictor variables .............................................................................................. 59 2 2 Unweighted and weighted percentile group sizes ............................................... 60 2 3 Weighted percentile group means for demographic and healthrelated information. ......................................................................................................... 61 2 4 Demographic info rmation for the sample (total) and each group (none, low, medium, high) with weighted population percentage estimates. ......................... 62 2 5 Unweighted frequency table of diseases of the musculoskeletal system and connective tissue ICD 9 codes. .......................................................................... 64 2 6 Unweighted frequency table of musculoskeletal injury ICD 9 codes. ................. 65 2 7 Weighted mean annual expenditures for musculoskeletal pain diagnoses and a ll healthcare by percentile groups. .................................................................... 66 2 8 Univariate logistic regression results (15% expenditure percentile criteria) ........ 69 2 9 Significant results from multivariate logistic regression (15% expenditure percentile criteria)* .............................................................................................. 71 2 10 Significant results from multivariate logistic regression sensitivity analysis (10% expenditure percentile criteria)* ................................................................. 72 2 11 Significant results from multivariate logistic regression sensitivity analysis (20% expenditure percentile criteria)* ................................................................. 73 2 12 Logistic regression results for sensitivity analysis using non specific musculoskeletal pain ICD 9 codes (15% expenditure percentile criteria)* .......... 74 2 13 Cluster profiles .................................................................................................... 75 2 14 Eigenvalues from cluster analysis ...................................................................... 76 2 15 Standardized Canonical Discriminant Function Coefficients .............................. 76 2 16 Pooled withingroups correlations between discrim inating variables and standardized canonical discriminant functions .................................................... 76 2 17 Distribution of categorical variables among clusters* ......................................... 77 2 18 Descriptive analysis for continuous variables among clusters* ........................... 79 8

PAGE 9

2 19 Distribution of expenditure percentile groups by cluster membership ................. 80 3 1 Demographic information for the full, complete case, and incomplete follow up cohorts. ........................................................................................................ 109 3 2 Health related information for the full, complete case, and incomplete follow up cohorts. ........................................................................................................ 110 3 3 OSPRO questionnaire scores for the full, compl ete case, and incomplete follow up cohorts. ............................................................................................. 111 3 4 Frequency of healthcare utilization reported at 6month and 12month follow up ..................................................................................................................... 112 3 5 Summary of variables that contributed to prediction of any additional healthcare utilization ......................................................................................... 113 3 6 Summary of variables that contributed to prediction of utilization intensity at 6 and 12month follow up* ................................................................................... 114 3 7 Summary of variables that contributed to prediction of opioid utilization .......... 115 3 8 Summary of variables that contributed to prediction of injection utilization* ..... 116 3 9 Summary of variables that contributed to prediction of surgery utilization* ....... 117 3 10 Summary of variables that contribut ed to prediction of diagnostic tests and imaging utilization* ............................................................................................ 118 3 11 Summary of variables that contributed to prediction of emergency room utilization ........................................................................................................ 119 4 1 ICD9 CM codes used to identify osteoarthritis ................................................ 146 4 2 Demographic information of the sample by diagnostic category ....................... 147 4 3 Sample and weighted prevalence of individual comorbidities by diagnostic category ............................................................................................................ 148 4 4 Characteristics of those assigned to each osteoarthritis subgroup ................... 150 4 5 Characteristics of those assigned to each low back pain subgroup .................. 152 4 6 Sample and weighted prevalence of comorbidities by diagnostic category (without claims restrictions) .............................................................................. 153 9

PAGE 10

LIST OF FIGURES Figure page 1 1 Model of value development in physical therapy ................................................ 24 1 2 Integration of dissertation projects into framework for identifying best practice for patients with musculoskeletal pain. ............................................................... 25 2 1 Sample Selection Flowchart ............................................................................... 58 2 2 Weighted mean annual expenditures for musculoskeletal pain diagnoses and all healthcare by percentile groups. .................................................................... 67 2 3 Weighted mean percentage of total musculoskeletal pain expenditures attributable to each event type by percentile group. ........................................... 68 2 4 Weighted Mean Annual Expenditures for musculoskeletal conditions among 6 demographic clusters with 95% confidence interval ........................................ 81 4 1 Comorbidity classification structure for osteoarthritis diagnosis category. ........ 149 4 2 Comorbidity classification structure for low back pain diagnosis category. ....... 151 4 3 Comorbidity classification structure for osteoarthritis only diagnosis category (without claims restrictions). ............................................................................. 155 4 4 Comorbidity classification structure for low back pain only diagnosis category (without claims restrictions). ............................................................................. 156 4 5 Comorbidity classification structure for combined osteoarthritis/low back pain diagnosis category (without claims restrictions) ............................................... 157 4 6 Osteoarthritis subgroup longitudinal profiles. .................................................... 158 4 7 Low back pain subgroup longitudinal profiles. .................................................. 159 4 8 Osteoarthritis only subgroup longitudina l profiles (without claims restrictions). 160 4 9 Low back pain only subgroup longitudinal profiles (without claims restrictions). ...................................................................................................... 161 4 10 Combined osteoarthritis and low back pain subgroup longitudinal profiles (without claims restrictions). ............................................................................. 162 10

PAGE 11

Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy ADVANCING THE UNDERSTANDING OF HEALTHCARE UTILIZATION AND COSTS AMONG PATIENTS WITH MUSCULOSKELETAL PAIN By Trevor Anthony Lentz August 2017 Chair: Steven Z. George Major: Rehabilitation Science Musculoskeletal pain is a prevalent condition with far reaching public health and economic effects Treatment pathways that more effectively manage high impact musculoskeletal pain are needed, however must be sensitive to valuebased healthcare reform. Ind ividuals at risk for low value care include those with high levels of healthcare utilization or complex medical needs. T hese individuals represent high priority targets for comparative effectiveness research, but limited knowledge of their characteristics and healthcare utilization patterns has constrained research progress Therefore, t he overall goal s of these dissertation projects are to use three different datasets to 1) identify characteristics of populations with musculoskeletal pain at risk for low value care due to high healthcare costs and/or complex medical needs and 2) define healthcare utilization patterns for these populations. The first project used data from the Medical Expenditure Panel Survey (MEPS) to identify individuals at risk of persistently high pain related healthcare expenditures over the 2 years of the survey. R obust predictors of persistently high expenditures were older age, greater missed work days, greater pain interference, having private insurance, 11

PAGE 12

having a musculoskeletal injury diagnosis, number of musculoskeletal diagnoses, and use of prescription medication for pain. The second project identified predictors of additional healthcare utilization following an episode of physical t herapy for musculoskeletal pain. This study use d data from the OPT IN OSPRO study, a longitudinal cohort study of patients seeking outpatient physical t herapy. Intensity of pain and its course over the early phases of rehabilitation were important predictors of subsequent healthcare utilization in the year following physical therapy. Psychological factors, disability, and comorbidities predicted use of specific services. The third project used data from the Medicare Current Beneficiary Survey (MCBS) wi th matching claims to i dentify co morbidity subgroups and patterns of healthcare utilization among individuals with osteoarthritis or low back pain. Latent class analysis identified 5 subgroups that were similar between pain conditions : low comorbidity pulmonary/hypothyroidism/anemia, hy pertension with diabetes complicated hypertension/renal/anemia, and complex cardiac/high comorbidity Greater disability and higher costs were associated with higher disease burden (defined by both number and combination of specific comorbidities) profiles. Implications of each projects findings for clinical practice, research and policy are discussed. 12

PAGE 13

CHAPTER 1 IMPACT OF MUSCULOSKELETAL PAIN AND THE NEED TO IMPROVE VALUE OF PAIN CARE Personal and Economic Burden of Musculoskeletal Pain Chronic pain is among the most prevalent, disabling, and costly conditions experienced by individuals throughout the world. (Institute of Medicine (US) Committee on Advancing Pain Research, Care, and Education, 2011) In the United States it is estimated that 126.1 million adults reported some pain in the previous 3 months, with 25.3 million adults (11.2%) suffering from chronic pain. (Nahin, 2015) Prevalence rates of pain associated with musculoskeletal conditions are of particular importance for rehabilitation science. Age adjusted prevalence of adults repor ting musculoskeletal pain in the last 3 months is highest for low back pain (28.1%) followed by knee pain (19.5%) and neck pain (15.1%). (Institute of Medicine (US) Committee on Advancing Pain Research, Care, and Education, 2011) Evidence also suggests that pain prevalence may vary based on sex and race, with women and African Americans experiencing higher prevalence rates for some pain conditions (Bartley & Fillingim, 2013; Lawrence et al., 2008) In 2011, the Institute of Medicine (now National Academy of Medicine) report Relievi ng Pain in America: A Blueprint for Transforming Prevention, Care, Education, and Research classified chronic pain as a high priority public health problem. (Institute of Medicine (US) Committee on Advancing Pain Research, Care, and Education, 2011) Consistent with this classification is that chronic pain substantially exceeds the prevalence and costs of many other public health problems, such as heart disease, cancer, and diabetes. (Institute of Medicine (US) Committee on Advancing Pain Research, Ca re, and Education, 2011) Of particular concern is that estimates suggest 13

PAGE 14

the prevalence of some pain conditions such as chronic and nonchronic LBP is increasing (by 64% and 29% respectively from 20002007 ). (Smith, Davis, Stano, & Whedon, 2013) This trend is partially attributable to the growing population of older adults, since prevalence of pain peaks in older adulthood. (Institute of Medicine (US) Committee on Advancing Pain Research, Care, and Education, 2011; Smith et al., 2013) Direct costs associated with pain treatment contribute substantially to overall healthcare expenditures in the US, with estimates from $261 to $300 billion annually. (Gaskin & Richard, 2012) W hen considering the value of lost work and lower wages, the total societal cost of pain ranges from $560 to $635 billion, with a significant proportion of those costs attributed to musculoskeletal pain. (Gaskin & Richard, 2012) The i ndividual burden associated with pain is often complicated by social and contextual factors that affect the availability and consistency of healthcare. For instance, provision of care is often fragmented and patients may have difficulty accessing the appropriate spectrum of services. (Von Korff et al., 2016) W ide variability in clinical practices related to prevention, assessment, and treatment also contribute to inconsistencies in outcomes and costs. (Institute of Medicine (US) Committee on Advancing Pain Research, Care, and Education, 2011) Moreover, pain management i n the US healthcare system is episodic and treatment most often coincides with the onset of pain or exacerbation of chronic symptoms. While a front loaded management approach might be appropriate for some with acute pain, it is discordant with the nature o f care needed for chronic or persistent pain. This discordance is important because chronic or persistent symptoms most often contribute to the public health burden. (Blyth, Van Der Windt, & Croft, 2015; A. D. Woolf & Pfleger, 2003) 14

PAGE 15

Public Health C risis of Pain The increasing prevalence of pain coupled with its high economic costs and so cietal burden has created a public health problem with far reaching impacts on healthcare services. (Goldberg & McGee, 2011) A few of these impacts are outlined below. Impa ct #1: Increasing S trai n on Healthcare R esources Expansion of healthcare coverage granted by the Affordable Care Act (ACA) and the rising number of Medicare beneficiaries have contributed to an increase in the healthcareseeking population at risk for disability due to chronic pain. (Blumenthal, Stremikis, & Cutler, 2013; Smith et al., 2013) In addition, there is a growing emphasis on collaborative, multi disciplinary care to better address the biopsychosocial needs of patients with chronic pain. However, multidisciplinary rehabilitation can be costly and consumes many healthcare resources. (Kamper et al., 2015; Von Korff et al., 2016) Efforts to expand multidisciplinary care, alongside growth in the healthcareseeking population and a high incidence rate of avoidable care (M oynihan, Doust, & Henry, 2012; Stewart et al., 2015) has contributed to considerable strain on healthcare providers and resources associated with the treatment of musculoskeletal pain. Strain is especially concentrated among providers that commonly treat patients with pain, such as physical therapists. Indeed, in the United States, the largest proportion of direct medical costs for low back pain are spent on physical therapy (17.0%) and inpatient services (17.0%), followed by pharmacy (13.0%) and primar y care (13.0%). (Dagenais, Caro, & Haldeman, 2008) For healthcare practitioners that manage musculoskeletal pain health services research informed by pain prevalence, best practice, and 15

PAGE 16

healthcare payment reform issues will drive improved health policy, clinical practice, and health workforce planning. (Smith et al., 2013) Impact #2: High Concentration of Utilization and Spending in Vulnerable Populations with Complex C onditions H ealthc are spending is concentrated among a small proportion of the population with high levels of healthcare utilization (Stanton & Rutherford, 2005) For example, among patients with chronic low back pain in a Medical Expenditure Panel Survey (MEPS) study, the top decile of healthcare utilizers accounted for 57% of costs, and the top quintile accounted for 72% of costs. (Smith et al., 2013) Utilization and related healthcare costs are disproportionately higher for those with indicators f or chronic and/or complex pain conditions compared to those with acute pain and few risk factors for developing chronic pain. (Mller Schwefe et al., 2011) For instance, Smith et al. reported 94% of patients with chronic back pain use healthcare services, whereas only 75% of those with nonchronic back pain use healthcare services (odds ratio, 4.7; 95% confidence interval, 4.0 5.6). (Smith et al., 2013) In another study using MEPS data, Stockbridge et al. found that moderate and severe chronic pain related interference was associated with a $3,707 and $5,804 increase in expendi tures over no pain interference and a $2,218 and $4,315 increase over nonchronic interference, respectively. (Stockbridge, Suzuki, & Pagn, 2015) The presence of comorbid chronic conditions is also attributable to higher healthcare spending and utilization. (Chi, Le e, & Wu, 2011b, 2011a; Meraya, Raval, & Sambamoorthi, 2015) The exact relationship between comorbidities and spending for musculoskeletal pain conditions is unknown, but one study suggests high rates of spending and out of pocket expenses among those wi th arthritis and 2 or more highl y 16

PAGE 17

prevalent chronic conditions ( diabetes mellitus, heart disease, or hypertension) (Meraya et al., 2015) The potential influence of comorbidities on spending and outcomes a ssociated with chronic pain is i mportant considering an estimated 118 million (1 in 2) adults live with at least 1 chronic condition and 60 million live with 2 or more. (B. W. Ward, Schiller, & Goodman, 2014) High utilization high cost individuals with musculoskeletal pain are important targets for comparative and cost effecti veness research, particularly if they have comorbidities that make them vulnerable to poor outcomes. Impact #3: Avoidable Morbidity and M ortality Associated with Opioid Use Opioids are commonly prescribed for chronic pain conditions and may be effective for short term pain relief, however longterm effectiveness is questionable (Trescot et al., 2008) The rising incidence of opio i d related overdose and death has become a significant public health concern. (Larochelle et al. 2016) From 1999 to 2014, the rates of drug related d eath s nearly tripled, with over 60% attributed to opioid use. (Rudd, Seth, David, & Scholl, 2016) This is a particular concern for older adults, since nearly 40% of adults aged 65 and older experience polypharmacy (tak ing > or equal to 5 prescription drugs). (Charlesworth, Smit, Lee, Alramadhan, & Odden, 2015; Guthrie, Makubate, Hernandez Santiago, & Dreischulte, 2015) The increase in overdose mortality may be attributable to increased opio i d prescribing(Modarai et al., 2013) and has led to calls for more research into safer, non pharmacological treatment alternatives. (Larochelle, Liebschutz, Zhang, Ross Degnan, & Wharam, 2016) Comparative effectiveness studies are needed to 1) identify which patients are likely to benefit most from long term opioid use, 2) optimal opioid prescription parameters, and 3) which patients are at the h ighest risk for adverse events due to opioid use. 17

PAGE 18

Recognizing the growing extent of pain as a public health crisis, the National Academy of Medicines Interagency Pain Research Coordinating Committee (IPRCC) recently outlined priorities for pain research, treatment, education and policy in th e National Pain Strategy ( NPS). (Von Korff et al., 2016) In particular, it has called upon clinicians and researchers to 1) identify methods for reducing the incidence of acute to chronic pain transition, and 2) develop effective management strategies for high impact chronic pain. Pain research has improved our understanding of acute and chronic pain as separate yet related health conditions (C. J. Woolf, 2010) Acute pain is considered part of a normal response to a noxious stimulus that serves to protect the body from damage or potential harm. (C. J. Woolf, 2010) On the other hand, chronic pain is now viewed as a nervous system disease that develops following an acute pain episode, but is d istinct in that it no longer serves a protective function. Chronic pain conditions are often resistant to treatment and there is growing interest in identifying more cost effective pain treatment methods. Therefore, methods for prevention and treatment of chronic pain will require fundamentally different treatment approaches. Changing Economic Conditions in Healthcare and Implications for the Treatment of Musculoskeletal Pain R ising healthcare costs coupled with the increasing burden of musculoskeletal pai n has spurred growing concern over the need to improve value of care. In the context of healthcare, value is often defined as the relationship between quality or outcomes of services and their associated costs (Jewell, Moore, & Goldstein, 2013) M ore specifically, value may be defined as the health outcomes achieved per dollar spent. (Porter, 2010) Concern for improving value of care is not spec ific to the treatment of pai n. Rather, it reflects a much broader ongoing shift in how healthcare is sustainably 18

PAGE 19

delivered and financed. Historically, reimbursement structures have shaped the way healthcare is delivered in the United States. (Institu te of Medicine (US) Roundtable on Evidence Based Medicine, 2010) Fee for service models, thirdparty payment systems and poor cost transparency have contributed to healthcare delivery that aspires to provide the best outcomes, but is highly cost insens itive. The result has been unrestrained and often wasteful spending without incentives for clinicians and researchers to identify and utilize cost effective treatment options. In response, the Department of Health and Human Services has partnered with federal and private payers to advance valuebased payment methods that support the Triple Aim, of better care, better health, and lower costs. (Berwick, Nolan, & Whittington, 2008; Burwell, 2015) Many of these methods aim to improve value by distributing financial risk among providers to incentivize collaboration across healthcare disciplines. As a result, value of service has become an increasingly important metric by which healthcare services are compared and selected by patients, providers, payers, and poli cy makers. The shift to value based purchasing will have significant implications for the development of sustainable treatment pathways to prevent and manage chronic pain. Beyond the need to improve pain related outcomes is the pragmatic concern for doing so while optimizing costs. Thus, healthcare providers will be well positioned to meet payment reform initiatives by establishing and improving upon the value of care they deliver. Establishing value is typi cally accomplished through comparative and cost effectiveness studies. However, these approaches stop short of describing the components and process es that drive quality and costs, which is a necessary step for improving the value of existing services or developing new ones. The Value Model for 19

PAGE 20

Phy sical Therapy (Figure 11) describes how value emerges from interactions among the health care system, health care services organizations, the physical therapist, and the patient. (Lentz, Harman, Marlow, & George, 2017) It can be used to guide the development and evaluation of pain management pathways. Value Model for Physical Therapy The Value Model for Physical Therapy integrates utilization, economic and biopsychosocial models of healthcare(Andersen, 1995; Babitsch, Gohl, & von Lengerke, 2012; Fordyce, 1994; Leeuw et al., 2007) as well as existing research on healthcare quality and costs. It outlines interactions among modifiable and nonmodifiable components related to the patient, physical therapist health services organiz ation (HSO) and healthcare system that give rise to value. The Model shares features that are universal to value models. Among these universal features is the emergence of value through interactions between different components. T he healthcare system, HSO, physical therapist and patient exemplify structures which interact to create processes representing the provision of care. These processes then give rise to outcomes that define value, such as satisfaction, function and perceived health status. Consistent with complex systems, we propose that quality and costs arise not as a direct function of each structure (i.e. patient, physical therapist HSO, or healthcare system), but rather from the interaction of these structures. These interactions are dynamic and distinct yet interdependent, with feedforward (solid arrows) and feedback (dashed arrows) mechanisms. I n this Model the most direct pathways are those with closest proximity to interactions involving the patient. We propose that proximal interactions create more direct and immediate effects on quality and costs because of the strong influence they 20

PAGE 21

have on outcomes and utilization. The patient therapist interaction represents the clinical encounter where a critical driver of healt hcare utilization, perceived need, is assessed. Through this interaction, the therapist evaluates the health status of the patient and determines which services are necessary (i.e. evaluated need). Because this is the closest interaction to value in our Mo del and one highly influenced by provider behavior, we suggest that leverage points here represent the best opportunity to make direct and immediate improvements to value. Direct pathways are also easier to monitor and measure making them attractive targets for research and quality improvement applications. Conversely, interactions at the healthcare system and organization level will produce few immediate or direct contributions to quality and costs, but can have substantial indirect and downstream effects by influencing the behavior of organizations and providers. Improving health policy is one pathway to enhancing value at the organizationsystem interaction level (Rundell et al., 2015) but more difficult for individual provid ers to influence and monitor. The Model can be applied to pain care in two ways. First, value in the context of muscu loskeletal pain is defined by outcomes that closely align with the National Academy of Medicines priorities. For those with acute pain, value is defined as the prevention of chronic pain. For those with chronic pain, value is defined by effective pain man agement. To meet the contrasting needs of each group, providers, health services organizations (HSO) and healthcare systems must interact in different ways to achieve high value care. Second, model components that are evidence supported for influencing costs and outcomes for those with muscul oskeletal pain are emphasized. Readers are 21

PAGE 22

directed to the Value Model development paper for further description of pain related components and characteristics. (Lentz et al., 2017) Using the Val ue Model to Inform Comparative Effectiveness Research for Musculoskeletal P ain Comparative effectiveness research is designed to inform healthcare decisions by providing evidence on the effectiveness, benefits, and harms of different treatment options. (AHRQ Effective Health Care P rogram, n.d.) It is conducted most often in pragmatic, real world settings to compare interventions in a population of interest. However, it may also be used to assess the relative effectiveness of an intervention among different patient populations. As the importance of healthcare value has grown, patient populations at risk for low value care have become high priority targets for comparative effectiveness research. Those with high painrelated healthcare costs and/or complex medical needs may have an e levated risk of experiencing low value care. Research to better defin e these patients is an important precursor for comparative effectiveness studies aimed at identifying optimal pain management pathways. T he overall goal s of these dissertation projects a re to use 3 different datasets to 1) identify characteristics of populations with musculoskeletal pain classified as being at risk for low value care, and 2) define healthcare utilization patterns for these populations. For each project, the Value Model is applied as a framework to guide the selection of candidate variables for analysis We a prioiri defined 3 at risk groups to study, including: 1. Individuals that have persistently high direct healthcare expenditures related to musculoskeletal pain 2. Individuals that utilize additional healthcare following an episode of physical therapy for musculoskeletal pain 3. Individuals with chronic comorbid conditions alongside a musculoskeletal pain condition(Benjamin, 2010) 22

PAGE 23

The projects are important because they will 1) improve prospective identification of those at risk for low value care, 2) inform cohort selection for comparative and cost effectiveness research and 3) inform the development of treatment pathways that more effectively and efficiently address the unique healthcare needs of patients with musculoskeletal pain. Figure 1 2 depicts how these projects fit into a broader framework for identifying best practice and optimizing value of care. Identifying the most cost effective pain treatment strategies for those at risk for low value care will directly address priorities for pain research, treatment, education and policy, as well as healthcare reform initiatives promoting improved value. 23

PAGE 24

Figure 11. Model of value development in physical therapy. *Characteristics of each component relative to value based physical therapy are provided in Lentz et al. PTJ 2017 24

PAGE 25

Figure 12. Integration of dissertation projects into framework for identifying best practice for patients with musculoskeletal pain. 25

PAGE 26

CHAPTER 2 PREDICTORS OF PERSISTENTLY HIGH COST HEAL THCARE UTILIZATION FOR MUSCULOSKELETAL PAIN Introduction Healthcare costs in the US are concentrated among a small proportion of the population with high lev els of healthcare utilization. (Mitchell, 2001; Riley, 2007) Often called high cost utilizers, these individuals are an important target of research and healthcare policy to improve patient management and reduce preventable costs. For instance, th e US federal government has funded the State Innovation Model s Initiative to promote the planning, design, testing, and evaluation of new service delivery models that improve value for those with high utilization and healthcare costs (Emeche, 2015; Hughes, Peltz, & Conway, 2015) These new delivery models focus on reducing preventable healthcare utilization, employing datadriven strategies, improving stakeholder engagement, and redesign ing clinical pathways (Emeche, 2015) An important consideration when studying high cost utilizers is the persistence of high costs. Isolated high cost episodes may be appropriate in some situations when associated with cost effective pain treatment options, such as total joint replacement. (Daigle, Weinstein, Katz, & Losina, 2012) However, it is expected that these high cost episodes will reduce downstream utilization and precede a period of minimal healthcare spending for the condition. Indeed, Johnson et al. (Johnson et al., 2015) reported in a study of high healthcare utilizers, only a small percentage of those who had high utilization initially continued to have high levels of utilization a year later. Therefore, those which have persistently high expenditures and utilization are potentially more important targets of valuebased delivery model redesigns, (Chang et 26

PAGE 27

al., 2016) since the inc idence of unhelpful or wasteful healthcare utilization may be particularly high in this population segment. Prior research has examined predictors of high cost utilization; however, most studies assess point high cost utilization rather than persistent high cost utilization. Those with persistently high cost utilization are a research priority because of their unique ability to identify potential system inefficiencies linked to low value care (Emeche, 2015) To our knowledge, persistent high cost utilization studies have not been performed in populations with musculoskeletal pain. (Fleishman & Cohen, 2010) Point high cost studies in other patient populations have identified factors such as older age, Caucasian race lower household income, lower education level, obesity, smoking status, phy sical inactivity, poorer physical and mental health status, and existence of comorbid chronic disease as predictors of high cost use. (Chechulin et al., 2014; Fitzpatrick et al., 2015) It is unclear whether these same factors contribute to persistence of high healthcare costs and use among those with musculoskeletal pain. Additionally, many prior studies on high cost utilizers were performed in countries with government funded, singlepayer healthcare systems that provide universal healthcare coverage. (Chechulin et al., 2014; Fitzpatrick et al., 2015; Heslop, Athan, Gardner, Diers, & Poh, 2005; Rosella et al., 2014) As a result, many of the predisposing or enabling factors that could affect healthcare use and costs in the US, such as income and insurance coverage, are largely unexplored. These factors are likely to have a significant influence on individual healthcare costs and use, particularly related to musculoskeletal pain. 27

PAGE 28

The National Pain Strategy has called for new delivery models that improve value of care because management of chronic musculoskeletal pa in is particularly suscep tible to high cost utilization. (Von Korff et al., 2016) On reason for this susceptibility is that psychosocial factors often contribute to persistence of painrelated disability (Hill & Fritz, 2011; Nicholas, Linton, Watson, Main, & Decade of the Flags Working Group, 2011; D. Turk & Monarch, 2002) and have a known influence on healthcare utilization in the general population. For instance, i n a study comparing point high cost users to persistent high cost users, Chang et al. reported recurrent medication usage and higher prevalence of chronic an d psychosocial conditions in those with persistent high costs. (Chang et al., 2016) Other studies have reported a higher prevalence of mental illness, including mood disorders, anxiety and depression among high cost users. (Chechulin, Nazerian, Rais, & Malikov, 2014; Hensel, Taylor, Fung, & Vigod, 2016) As val ue based care models become more prevalent early identificat ion and selective targeting of those at risk for persistently high costs due to musculoskeletal pain is a top priority so that limited healthcare resources can be more efficiently and effectively distributed. Factors that contribute to high healthcare utilization and costs are multidimensional, and represent characteristics of the patient, provider, healthcare organization, and healthcare system. (Lentz et al., 2017) Risk factors for high cost utilization can be used to inform the development of pathway redesign to optimize value, with modifiable risk factors representing important leverage points for reducing expenditures. Defining these fact ors is a critical first step toward developing best practice pathways for musculoskeletal pain in response to the National Pain Strategy. 28

PAGE 29

This analysis will directly inform future studies and pragmatic trials that test clinical pathways tailored to the unique needs and characteristics of those at risk for persistently high cost utilization associated with musculoskeletal pain. Two complem entary aims were developed to identify the sociodemographic and healthrelated characteristics associated with persistently high cost utilization for musculoskeletal pain. The first aim was to identify sociodemographic and health related f actors that predicted classification into the highest 15th percentile of annual expenditures for treatment associated with musculoskeletal pain over 2 consecutive years (i.e. persistent high cost utilization) The second aim was to assess healthcare expenditures across patient subgroups characterized by empirical combinations of sociodemographic and healthrelated fact ors. Both aims explore alternative approaches to identifying clinical populations at risk for high healthcare spending due to musculoskeletal pain. This analysis will determine whether identifiable high cost patient subgroups exists or if persistent high cost utilizers are better identified by individual predictors. Geographic region and type of insurance predict healthcare utilization in the general population and we hypothesize these to predict higher costs in populations with musculoskeletal pain as well. (Committee on Geographic Variation in Health Care Spending and Promotion of High Value Care, Board on Health Care Services, & Institute of Medicine, 2013; Nyman, 2004) Furthermore, we hypothesize that patient level factors such as high psychological distress, depre ssion, pain interference, and increased comorbidities will predict high cost group membership, as these are often associated with elevated perceived need and poorer treatment outcomes (Andersen, 29

PAGE 30

1995; Bair, Robinson, Katon, & Kroenke, 2003; Hart et al., 2011; Ritzwoller, Crounse, Shetterly, & Rublee, 2006) Accordingly, we hypothesize that subgroups defined by greater number of musculoskeletal conditions and greater access to care (i.e. insurance) will exhibit higher odds of persistent high cost utilization. Methods Dataset This study used Public Use File Household Component data from the five most recent panels (Panels 1317, years 20082013) of the Medical Expenditure Panel Survey (MEPS). Administered by the Agency for Healthc a re Research and Quality (AHRQ), MEPS is a set of largescale surveys of families and indi viduals, their medical providers and employers across the United States that includes data on demographic characteristics, health conditions, health status, use of medical care services, charges and payments, access to care, satisfaction with care, health insurance coverage, income, and employment (J. Cohen, 1997) Subjects are enrolled in panels, each of which includes data collection in 5 rounds over 2 calendar years. MEPS uses a stratified, multistage sampling design. Survey weights are provided to account for differential sample selection probabilities and adjust for unit level attrition and personlevel nonresponse to provide nationally representative estimates of the US noninstitutionalized civilian population. (S. Cohen, 2009) Study Sample Survey respondents with musculoskeletal conditions were identified using medical conditions files. The medical conditions file includes diagnosis level detail on medical conditions reported by MEPS respondents in each calendar year. Medical conditions are collected from multiple sections of the survey: Condition Enumeration 30

PAGE 31

(CE), Medical Events (ME) and/or Disability Days (DD). The CE section asks respondents to identify specific physical or mental health problems during the interview reference period, regardless of whether medical care was sought for these problems. In most ME sections, respondents are asked to identify the conditions that are associated with healthcare events they reported. Finally, conditions that caused respondents to miss school or work, or spend more than half a day in bed, are captured in the DD section. Each record in a MEPS annual medical condition file represents a current condition reported as existing for a MEPS respondent at any time during the year. ( J. Cohen, 1997) Medical conditions reported in each section are recorded as verbatim text and then coded by professional coders to fully specified International Classification of Diseases, 9th Revision, Clinical Modification (ICD9 CM) codes Respondents who reported at least one musculoskeletal condition of interest in the index year of the panel (Year 1) were considered for inclusion in the study. Selection of ICD 9 codes to be included in the study followed definitions from The Burden of Musculoskele tal Diseases in the United States: Prevalence, Societal and Economic Costs (BMUS), 3rd edition, which is produced by the United States Bone and Joint Initiative (USBJI).(United States Bone and Joint Initiative, 2014) Our intent was to identify high expenditures for musculoskeletal conditions where expenditures are not exp ected to be persistently high. Therefore, medically complex musculoskeletal conditions that are expected to produce high expenditures over prolonged periods were excluded. Exam ples of such conditions include spinal cord injury, amputation, congenital deformities and cancer. MEPS Public Use Files exclude fully specified ICD 9 codes to protect respondent confidentiality but include ICD 9 codes truncated to the first 31

PAGE 32

three digits. Threedigit ICD 9 codes included and excluded from the analytic sample are listed in Appendix A Respondents were also excluded from the final analytic dataset if they were not in scope for the entire panel, < 18 years of age upon entry into the survey, did not provide data for all 5 rounds of the panel, were ineligible to complete the Adult Self Administered Quest ionnaire (SAQ) in Year 1 or had a proxy complete the Adult SAQ. The Adult SAQ is a supplemental paper questionnaire administered once per year to all household respondents 18 years old and older that includes questions from the Consumer Assessment of Healt h Plans (CAHPS), the SF 12, the EuroQol 5D, and attitude items. Respondents with an incomplete SAQ were excluded because m ultiple items from the SAQ were used as independent variables in this study. Finally, respondents that reported zero expenditures in Year 1 were also excluded. Our intent was to identify predictors of persistent high cost utilization among those seeking healthcare. Inclusion was conditional upon nonzero expenditures in the first year so that two years of expenditure data were available following initiation of healthcare for all respondents. The selection process for the final analytic sample is outlined in Figure 2 1. Expenditure Summaries We used direct method for estimating expenditures related to MSK pain. (Coughlan, Yeh, ONeill, & Frick, 2014) MEPS includes separate, full year event files for prescribed medicine, office based medical provider visits, outpatient department visits, emergency room visits, inpatient hospital stays and home health visits. For each event, 12 sources of payment are reported: self/family, Medicare, Medicaid, private insurance, Veterans /CHAMPVA, TRICARE, other federal sources, state and local (non32

PAGE 33

federal) government sources, W orkers Compensation, other private insurance, other public insurance, and other insurance. The sum of 12 payment sources is provided in a summary field for each event. Payments for over the counter drugs are not included in event files. Healthcare events were linked to each musculoskeletal condition through the conditionevent crosswalk. Summary costs for officebased expenditures, outpatient visits, prescribed medicine events, inpatient services, emergency room visits, and home health visits related to tr eatments for musculoskeletal pain conditions were used in constructing annualized cost summaries. Personlevel overall expenditure summaries for musculoskeletal pain were developed for each of the two years in the panel. For Year 2 expenditure estimates, only those events linked to conditions reported in Panel Year 1 were considered. This approach excluded expenditures for musculoskeletal conditions that were reported in Panel Year 2 but not Year 1. For Year 2, summary costs were set at $0 if no events were reported for a condition. For each year, an expenditure percentile rank was assigned to each respondent based on expenditure summaries for the entire sample for that year. Percentiles were only assigned for r espondents with nonzero expenditures. Predictor V ariables Sociodemographic and healthrelated variables were selected from the longitudinal panel data file. Unless otherwise noted, all selected predictor variables were collected from Year 1 data. Selecti on of variables was based on the Value Model of Musculoskeletal Pain and Andersen Behavioral Model to include sociodemographic and healthrelated variables with the potential to influence healthcare utilization. (Andersen, 33

PAGE 34

1995; Lentz et al., 2017) Predictor variables are described below and summarized in Table 2 1. Sociodemographic variables. Age, sex, race, ethnicity, years of education, body mass index (BMI) smoking s tatus, poverty category, employment status, metropolitan statistical area (MSA), and census region were included as sociodemographic variables. Category descriptions and analysis coding procedures for all categorical variables are listed in Appendix B Co morbidities We used the Deyo adaptation of the Charlson Comorbidity Index to determine comorbidity burden. (Deyo, Cherkin, & Ciol, 1992) The index adaptation accounts for disease severity and c omorbid conditions in studies of outcome and resource use employing administrative databases. The presence of comorbid conditions was determined from ICD 9 codes listed in respondents medic al condition files for Year 1. Days of work missed due to i llness Sick leave can substantially contribute to the economic burden of musculoskeletal pain. (Gaskin & Richard, 2012) Moreover, high initial amounts of sick leave have been shown to predict future sick leave (Karjalainen et al., 2003) and may help to predict persistence of high healthcare costs. To measure sick leave in this study, the number of times the respondent lost a half day or more from work because of illness, injury, or mental or emotional problems were recorded for each of the 5 rounds in Year 1 and summed for total days of work missed. General health status General health status was assessed with the SF 12, a component of the adult SAQ. The SF 12 is a generic, self reported health survey that assesses 8 domains of mental and physical health, including physical functioning, role 34

PAGE 35

limitations due to physical problems, bodily pain, general health perceptions, energy and vitality, social functioning, role limitations due to emotional problems, and mental health. (J. Ware, Kosinski, & Keller, 1996) Mental and Physical Component subscores were collec ted for analysis. (Soley Bori et al., 2015) Pain i nterference. P ain interference with work and daily activities was assessed using the following SF 12 question from the Adult SAQ: During past 4 weeks, how much has pain interfered with normal work outside the home and housework? Responses include None, a little bit, moderately, quite a bit or extremely. This item is commonly used as a measure of pain interference in populationbased studies on musculoskeletal pain conditions (Karjalainen et al., 2003; Thomas, Peat, Harris, Wilkie, & Croft, 2004) General psychological d istress. The Adult SAQ includes six mental healthrelated questions, using the K 6 scale developed by Kessler and colleagues (Kessler et al., 2002) Developed for use in the annual US National Health Interview Survey and National Household Survey on Drug Abuse, the K 6 scale assesses non specific psychological di stress during the past 30 days and can be used to screen for individuals with mental illness at the population level. (Prochaska, Sung, Max, Shi, & Ong, 2012) The K6 summary score provides an index to measure general psychological distress with higher values indicating greater psychological distress. The scale has demonstrated excellent internal c onsistency and reliability (Cronbach's alpha=0.89) (Kessler et al., 2002) as well as construct validity across gender s(Drapeau et al., 2010) and major sociodemographic subsamples (Kessler et al., 2002) 35

PAGE 36

Depression Depression can have substantial effects on functional status, symptomrelated difficulty, sick days, and healthcare utilization. (Kroenke, Spitzer, & Williams, 2003) The Adult SAQ includes two additional mental health questions from the Patient Health Questionnaire (PHQ 2). These questions assess the frequency of the respondents depressed mood and decreased interest in usual activities: 1) During the past two weeks, how often have you been bothered by having little interest or pleasure in doing things, and 2) During the past two weeks, how often have you been bothered by feeling down, depressed, or hopeless. Scores are summed to provide an overall assessment of depression. Perceived health status and attitudes about h ealth. Self reported physical and mental health status were assessed with questions asking r espondents to report their perceived health status and perceived mental health status compared to others T he Adult SAQ also includes questions that ascertain certain health related attitudes As an estimate of self efficacy for managing ones own condition, respondents were asked if they felt they can overcome illness without help from a medically tra ined person. Self efficacy is important, potentially modifiable characteristic that can influence healthcare utilization and outcomes and may have significant effects on costs (Cross, March, Lapsley, Byrne, & Brooks, 2006; Greene, Hibbard, Sacks, Overton, & Parrotta, 2015; Hibbard, Greene, & Overton, 2013) Total musculoskeletal pain diagnoses. The existence of multiple musculoskeletal pain diagnoses is likely to influence utilization and spending in this analysis. Therefore, to better understand the effect of multiple conditions, and to control for this effect on other variables in the model, we developed a summary count of 36

PAGE 37

musculoskeletal pain conditions reported in Year 1 of the panel. These conditions were identified by ICD 9 code using the same methodology when identifying eligible participants for the study. D iagnosis type. Our study eligibility criteria were designed to reduce the likelihood that medical complexity would substantially influence costs. However, certain diagnoses, such as fractures, sprains, and other acute injuries, are likely to be associated with higher initial costs, particularly if surgery is performed. As a result, those with musculoskeletal injuries might be more likely to be classified as high cost utilizers, despite the appropriate and necessary use of high cost interventions. To explore the effects of diagnosis type in the model, diagnoses were separated into two types based on ICD 9 code specifications: 1) Diseases of The Musculoskeletal System and Connective Tissue diagnosis only ( ICD9 codes 715739 ) and 2) Musculoskeletal Injury (ICD9 codes 805 959). For the purposes of analysis, respondents were classified into a musculoskeletal disease only group or a musculoskeletal injury with or without a musculoskeletal disease group. Significance of this variable in multivariate analysis might suggest the need to examine these two groups separately to reduce the effects of high costs services that might be considered generally appropriate. Usual care p rovider. For usual source of care, the MEPS Household Component access to care section as ks respondents whether there is a particular doctors office, clinic, health center, or other place they usually go to if sick or in need of health advice. This measure of access to care was included since existence of a usual care provider has been associ ated with reduced emergency room visits in studies of frequent healthcare utilizers. (L. J. Harris et al., 2016; Shi, 2012) 37

PAGE 38

Health in surance v ariables. MEPS provides information on monthly payer status for each of the following: TRICARE, Medicare, Medicaid/SCHIP, Other Public A Insurance, Other Public B Insurance, or private insurance. MEPS also includes summary measures that indicate w hether or not a person has any insurance in a month. E ach respondent was categorized as being privately insured all year, publicly insured all year, uninsured part of the year and either privately or publicly insured the remainder or uninsured all year. (Horner Johnson, Dobbertin, Lee, Andresen, & Exper t Panel on Disability and Health Disparities, 2014) Number of prescription medications for p ain. An important potentially modifiable factor is use of prescription medication for pain management. Escalating abuse of prescription pain medication has prompted efforts to understand how pain medication use influences outcomes and costs across pain conditions. Total number of prescription medication (opioid and non opioid medications) for pain events were summed for Year 1 and included as a predictor of high c ost utilization. Statistical Analysis Prediction m odel for persistently high e xpenditures R espondents were classified into one of three groups depending on their expenditure levels in Year 1 and Year 2 of the panel. T hose in the top 15% of expenditures specific to events with a musculoskeletal diagnosis in both years were defined as the high healthcare expenditure (HIGH) group (de Oliveira, Cheng, Vigod, Rehm, & Kurdyak, 2016; Garfinkel, Riley, & Iannacchione, 1988; Hartmann, Jacobs, Eberhard, von Lengerke, & Amelung, 2016; Leininger, Saloner, & Wherry, 2015) T hose in the bottom 15% of expenditures across both years were defined as the low healthcare expenditure (LOW) group. T hose between the highest and lowest 38

PAGE 39

expenditure groups (70%) were defined as the medium healthcare expenditure (MEDIUM) group. Respondents who had expenditures in Year 1 of the panel and zero expenditures in Year 2 of the panel were allocated to the LOW or MEDIUM group based on their percenti le expenditure from year 1 Prior to multivariate analysis, univar iate logistic regression (PROC SURVEYLOGISTIC) was used to assess the relationship between each predictor and higher order expenditure group membership. M ultivariate ordinal logistic regression (PROC SURVEYLOGISTIC) analysis was then used to determine pred ictors of group membership when accounting for all other predictors in the model. Proportional odds assumption was tested to determine appropriateness of ordinal logistic regression. In the event of a proportional odds assumption violation, we planned to use a generalized logit (multinomial) model and compare results to ordinal regression for determination of predictor robustness. First order correlations between all independent variables were assessed to determine potential for multicollinearity in the model. All correlations were < 0.5, therefore each independent variable considered for the analysis was included in the multivariate regression model. As an additional assessment of multicollinearity, variance i nflation f actor (VIF) and t olerance were calcula ted for each variable after applying sampling weights but without accounting for the complex sample design (PROC REG) (Rudolf J. Freud & Ramon C. Littell, 2000) Missing v alues accounted for less than 5 % of the observations for all study variables. Therefore, rather than imputing missing variables, listwise deletion was used to analyze available data. O verall model fit was examined with chi square analysis, Nagelkerke r2 and c statistic values when appropriate. Comparison of adjusted odds ratios (OR) and 95 % 39

PAGE 40

confidence interval (CI) were used to determine the relative strength of each predictor. Robustness of model predictors was tested with sensitivity analyses by alternatively defining the percentile r anks for group membership. Analyses were conducted with 10%/80%/10% and 20%/60%/20% definitions for group membership ( HIGH / MEDIUM / LOW respectively), as these thresholds have been used in other studies to define high healthcare expenditures (de Oliveira et al., 2016; Guilcher, Bronskill, Guan, & Wodchis, 2016; Maeng et al., 2015; Rosella et al., 2014; Wodchis, Austin, & Henry, 2016) Expenditures among p atient s ubgroups Two step cluster analysis with log likelihood distance measuring was used to identify population subgroups based on key sociodemographic and healthrelated factors. Two step clustering was used because it is the ideal clustering approach with large datasets and/or when solutions based on mixtures of continuous and categorical variables are needed. Factors considered for inclusion in cluster analysis w ere: age, sex, race, ethnicity, insurance and number of musculoskeletal conditions. The number of clusters were not specified a priori, and cluster solutions based on largest ratio of distance measures for both Schwarzs Bayesian Criterion (BIC) and Akaike s Information Criterion (AIC) were examined for convergence. Cluster solution quality was assessed with a silhouette measure of cohesion and separation (a goodness of fit measure) ratio of largest to smallest cluster, smallest cluster > 25% criterion and number of excluded cases. Total expenditures for musculoskeletal pain among the clustered subgroups were compared with one way ANOVA and Bonferroni post hoc adjustment Discriminant function analysis with cross validated classification 40

PAGE 41

was performed to determine accuracy of cluster classification, as well as identify which variable( s ) best differentiated subgroup classification. All prediction model and cost estimation analyses were conducted using PROC SURVEY procedures (SAS v.9.3, SAS In stitute Inc., Cary, NC, USA) to account for the complex sampling design and employed personlevel SAQ sampling weights to adjust for questionnaire non response. (Machlin, Yu, & Zodet, 2005) Taylor series linearization was used for variance estimation. For cluster analysis, SPSS (Version) was used due to its ability to efficiently perform twostep clustering. Alpha was set at p=0.05 for all analy ses. The University of Florida Institutional Review Board approved this study. Results Descriptive Analysis The study included 12,985 respondents who met inclusion criteria over 5 panels, representing 150 792 697 individuals in the US population. Weighted and unweighted percentile group sizes are listed in Table 2 2. Approximately, 4% of the sample, representing 5,896,762 individuals, were classified in the HIGH expenditure group, whereas 11%, representing 15, 963 948 individuals, were classified in the LOW expenditure group. Demographic information and health related information for the sample are listed in Tables 2 3 and 2 4. Significant percentile group differences were observed for all variables except sex, BMI, smoking status, geographic region, education, and metropolitan statistical area. Diagnosis ( ICD9 ) code frequency for the sample is listed separately for diseases of the musculoskeletal system (Table 25) and musculoskeletal injuries ( Table 2 6 ). Mean annual m usculoskeletal expenditures wer e $21.22 (median = $20.41) for the LOW group, $1,431.08 (median = $420.06) for the MEDIUM group, and $12,547 (median = $8036.64) for the HIGH group (Table 2 7 and 41

PAGE 42

Figure 2 2). The proportion of total healthcare expenditures attributable to musculoskeletal pain was significantly different among groups with musculoskeletal pain expenditures comprising < 1% of overall healthcare expenditures for the LOW group, 12.5% for the MEDIUM group, and 52% for the HIGH group. Distribution of musculoskeletal pain expendi tures attributable to each event type per year for the 3 groups is presented in Figure 2 3. The majority of expenditures in the LOW group were prescription medication and officebased medical provider events, with prescription medication comprising over 75% of the costs in Year 2. Office based medical provider events contributed to between 4045% of costs in the MEDIUM group and represented the largest single event type category for both years in this group. Inpatient procedures contributed to a larger perc entage of musculoskeletal pain expenditures in the HIGH group compared to other groups. Univariate Analysis Univariate regression analy sis results are listed in Table 2 8 All individual univariate analyses violated the proportional odds assumption of ord inal regression. Therefore, ordinal regression was followed up with a generalized logit (multinomial) model that relaxes the proportional odds requirement and has a full set of parameters for each generalized logit. Results of both the ordinal and generali zed logit analyses are provided with the assumption that the true odds ratios fall between results of both methods. Factors that were consistently associated with higher order expenditure group membership across each method were older age, greater missed work days, greater pain interference, higher psychological distress, more depression, poorer perceived mental health, presence of a usual care provider, being insured, having a 42

PAGE 43

musculoskeletal injury diagnosis, greater number of musculoskeletal diagnoses, and greater use of prescription medication for pain. Being black and reporting better physical health status were consistently associated with lower odds of higher order group membership. Census region was also associated with group membership, with those living in the Midwest having lower odds of higher order group membership compared to those in the Northeast. Additional factors such as being unemployed, reporting an inability to overcome ills without medical help, and having greater comorbidities were also associated with higher order expenditure group membership, however these relations hips were less consistent and more sensitive to the type of analysis method used. Likewise, Hispanic ethnicity and better mental health status were associated with lower odds of higher expenditure group membership, but the strength and significance of the associations varied based on regression method. Multivariate Analysis Results for the multivariate ordinal regression showed a violation of the proportional odds assumption meaning that one or more explanatory variables exerted an effect on each cumulativ e logit that was dependent on cutoff value. Therefore, results of both ordinal and multinomial regression are presented for each predictor ( Table 29 ). Examination of regression diagnostics demonstrated no concern for multicollinearity (all VIFs < 10 and t olerances >0.1) among model predictors. Assessment of fit statistics showed good model fit for ordinal (Wald chi square = 29.66, p<.001, Nagelkerke r2 = 0.18) and multinomial (Wald chi square = 18.40, p<.001, Nagelkerke r2 = 0.20) regression models, with a lower AIC for the multinomial model, suggesting slight improvement in fit over the ordinal model. 43

PAGE 44

Variables associated with increased odds of being in the HIGH group compared to LOW group in multivariate models were: older age, greater missed work days, greater pain interference, having private insurance, having a musculoskeletal injury diagnosis, greater number of musculoskeletal diagnoses, and greater use of prescription medication for pain. Variables associated with decreased odds of being in the HIGH group compared to LOW grou p were being black and having low income compared to being poor/near poor. Better physical health (SF 8 physical component scale) was also associated with decreased odds of HIGH group membership, however was not associated with decreased odds of MEDIUM compared to LOW group membership. Interestingly, poorer perceived physical health status was associated with lower odds of being in the MEDIUM compared to LOW group, however poorer perceived mental health status was associated with h igher odds of being in the MEDIUM compared to LOW group. Those in the Midwest census region were less likely to be in the HIGH group compared to LOW group, but no less likely to be in the MEDIUM compared to LOW group. Missed work days, total number of musc uloskeletal conditions and insurance type were associated with the highest odds of higher order expenditure group membership. Sensitivity analyses using alternative categorizations for percentile groups (10% and 20%) demonstrated findings that were general ly consistent with the primary analysis Results of the sensitivity analyses are reported in Tables 210 and 2 11. Across all three analyses, age, missed work days, pain interference, insurance, diagnosis type, total number of musculoskeletal diagnoses, and total prescription medications were consistent predictors of group allocation. Odds associated with race, 44

PAGE 45

poverty level, census region, physical health (SF 8 physical component scale), and perceived health status were also similar across analyses, but continued to show differences in results between multinomial and ordinal regression. When using a 10% threshold to define high and low cost utilization, being female and having higher BMI were associated with lower odds of HIGH group membership, whereas high er education level and living in a metropolitan area were associated with higher odds of HIGH group membership. When using a 20% threshold, BMI was also associated with lower odds of high group membership, as was living in the South compared to the Northeast. At this threshold, smokers had lower odds of being in the MEDIUM group, as did those living in the West compared to Northeast. Diagnosis type was significant throughout all multivariate models, with odds of higher expenditure group membership associat ed with a musculoskeletal injury. Due to the potential influence of costlier, yet generally necessary surgical events often associated with injury, we repeated the primary analysis with respondents who only reported musculoskeletal disease diagnoses (ICD 9 codes 715739). The sample comprised of 8,112 respondents (1 ,092 in LOW, 6,753 in MEDIUM, 267 in HIGH) representing approximately 93.6 million individuals and 62.5 % of the original sample. Results of the analysis are listed in Table 2 12 Effects of age, race, poverty level, census region, missed work days, pain interference, physical health (SF 8 physical component scale), insurance, number of musculoskeletal diagnoses, and use of prescription medication for pain were similar compared to the primary analysis. Notable differences were that pain interference, perceived physical health status and perceived mental health status were not significant predictors when considering only those without 45

PAGE 46

musculoskeletal injury diagnosis codes. However, like the ot her sensitivity analyses, higher BMI was associated with lower odds of HIGH group membership. Missed work days, total number of musculoskeletal conditions and insurance type continued to be associated with the highest odds of higher order expenditure group membership. Cluster Analysis The initial cluster solution with 6 inputs yielded poor cluster quality, with the silhouette measure of cohesion and separation less than 0.2. To improve cluster quality, we sequentially reduced the inputs by one variable at a time based on importance measures provided by SPSS to identify a better cluster solution. Removing age and sex from the analysis improved the silhouette measure of cohesion and separation above the threshold for good cluster quality ( >.05 ). (Mooi & Sarstedt, 2011) The largest change in Schwarz Bayesian (BIC) values and ratio of BIC distance measures suggested a 6 cluster solution with 4 inputs. Using Akaike information criterion (AIC) instead of BIC produced the same 6cluster solution. Cluster sizes ranged from 11.8% 22.2% of the sample, with the ratio of largest to smallest cluster size equaling 1.88, suggest ing adequate cluster size. Cluster profiles based on the 4 cluster analysis inputs are list ed in Table 2 13. We assessed whether removing age and sex reduced the capability of these common demographic clusters to identify high expenditure group membership. To do this, we fit a linear regression model predicting mean musculoskeletal panel expenditures from cluster membership, age and sex. Additional variance explained by age or sex in the model would suggest these factors should be included in defining cluster membership despite a poorer quality cluster solution. Age, but not sex, improved variance explained in mean m usculoskeletal expenditures (p>.05). However, adding 46

PAGE 47

age back into the analysis did not change cluster membership for any subject. Therefore, age was excluded from the final cluster analysis to produce a more parsimonious solution. Discriminant Function Analysis Discriminant function analysis resulted in 5 discriminant functions for the 6 clusters. Equality of group means tests suggested that each predictor contributed discriminate among the 5 cluster solutions. Function 1 accounted for 70% and function 2 accounted for 17% of the total varianc e in the relationship between predictors and clusters ( Table 2 14 ) The pooled within groups correlations between discriminating variables and standardized canonical discriminant functions, as well as the standardized canonical discriminant function coeffi cients are listed in Tables 2 15 and 216 In discriminant function analysis, the first function is considered to be most important as it accounts for the greatest proportion of discriminating power among functions (Tabachnick & Fidell, 2006) B ased on the standardized coefficients for the first discriminant function, ethnicity demonstrated the strongest relationship with the discriminant function suggesting this variable was the strongest predictor of cluster allocation. Together, the functi ons were able to correctly classify 98.8% of grouped cases, (95.4% of the cluster 3, 99.0% of cluster 4, 99.9% of cluster 5, and 100% of clusters 1, 2, and 6) 47

PAGE 48

Cluster C omparisons D escriptive information on each of the 6 cluster s is included in Tables 2 17 and 2 1 8 The proportion of HIGH percentile group membership was significantly higher in cluster 1 (7% of cluster) and cluster 4 (5.4% of the cluster) compared to other clusters ( Table 219 ). The proportion of LOW percentile group membership was significantly higher in cluster 2 (14.0%), cluster 5 (15.7%), and cluster 6 (15.3%) compared to other clusters. Mean annual expenditure comparisons among clusters are shown in Figure 4. Weighted mean panel expenditures were significantly higher for cluster 1 ($2841.93, p<.01) than all other clusters. Weighted mean panel expenditures were lowest for cluster 6 ($891.20) compared to all other clusters. For each cluster, mean annual expenditures were lower in Year 2 compared to Year 1. Discussion Those with persistentl y high expenditures (top 15% of expenditures specific to events with a musculoskeletal diagnosis in both years) represented approximately 4% of respondents reporting a musculoskeletal pain diagnosis. This finding is consistent with Johnson et al. (Johnson et al., 2015) who found the majority of those that experience initial episodes of high cost utilization do not continue to have high expenditures over time. Therefore, persi stence of high expenditures is concentrated among a relatively small subset of individuals who experience an incidence of high expenditures due to musculoskeletal pain. Consistent predictors of HIGH group membership across all analyses were older age, greater missed work days, greater pain interference, having private insurance, having a musculoskeletal injury diagnosis, number of musculoskeletal diagnoses, and use of prescription medication for pain. Summarizing these findings, high expenditures are concentrated among individuals with 48

PAGE 49

more complex and disabling pain who have access to private insurance. These results also provide support for application of established healthcare ut ilization models (e.g. Andersen model) in musculoskeletal pain populations since predictors identified in this study represent a variety of predisposing characteristics (e.g. age), enabling resources (e.g. insurance), and needs (e.g. pain interference, diagnosis, number of conditions). Of particular interest are pain interference and use of prescription medication for pain, as these represent potentially modifiable factors that could be targeted through care pathway redesign. We used cluster analysis as an alternative method for identifying groups of individuals with persistently high expenditures as opposed to identifying individual variables predictive of high expenditures. This pragmatic approach was taken to clarify whether pathways aimed at reducing high cost utilization may be better targeted toward individual risk factors, or toward groups defined by combinations of specific sociodemographic and/or healthrelated factors This analysis found high costs concent rated among white, non Hispanic individu als with multiple musculoskeletal conditions and private insurance. This group had the highest weighted mean annualized costs in both years and is consistent with findings from the first aim that showed race, insurance coverage and number of musculoskeletal conditions to be important variablelevel predictors of expenditures. Interestingly, the cluster with the lowest expenditures across both years was also c omprised of white, nonHispanic individuals with private insurance. However, this low cost cluster o nly reported a single musculoskeletal diagnosis. This suggests, not surprisingly, that number of musculoskeletal diagnoses is a critically important factor in driving costs particularly among individuals who are white 49

PAGE 50

and non Hispanic The cluster approac h may be more clinically feasible than assessing individual risk factors, as it demonstrates how the predictive variables are distributed in subgroups of patients. For example, cluster analysis could be a useful method of developing screening methods that target groups that are likely to have higher expenditures. However, this approach may yield lower accuracy for determining an individuals risk for high expenditures, and be less useful for informing personalized treatment pathways. Sensitivity analyses su ggest the effects of characteristics such as sex, education, BMI, smoking status and area of residence are dependent upon the definition of high cost utilization. Interestingly, pain interference was not a significant predictor of expenditure group members hip when considering only those without a musculoskeletal injury diagnosis. Pain interference may help drive costs when associated with a musculoskeletal injury, but is less likely to contribute to high cost utilization when associated with more nonspeci f ic musculoskeletal conditions. This finding may reflect the nature of healthcare use for musculoskeletal pain and how diagnoses are classified. The musculoskeletal injury group included acute pain diagnoses for which individuals are more likely to seek high cost services (e.g. use of emergency services, inpatient care, imaging and/or surgery for fractures). The musculoskeletal condition category included a greater proportion of diagnoses that tend to be chronic. Those with chronic conditions may be more lik ely to utilize conservative treatments that are typically lower cost (e.g. physical therapy) compared to higher cost services that were common in the HIGH and MEDIUM groups (e.g. inpatient care). 50

PAGE 51

Comparison of our results to prior high cost utilization studies of overall healthcare is challenging since most have been conducted in single paye r, universal health care systems. Within government sponsored healthcare systems, enabling factors such as insurance or income can significantly affect likelihood of high cost utilization These studies have consistently shown that being older, having multiple chronic conditions, reporting poorer self perceived health, and being of lower socio economic status are important predictors of high cost utilization. (Calver et al., 2006; Guilcher et al., 2016; Rosella et al., 2014) Some have also shown a relationship between persistence of high cost utilization and multiple drug therapies or recurrent medication us e (Chang et al., 2016; Guilche r et al., 2016) Many of our findings align with those of prior high cost utilization studies of overall healthcare. For instance, we found odds of higher expenditures associated wi th being older, having multiple musculoskeletal conditions, greater number of prescription pain medications, and missed work days (M. Charlson, Wells, Ullman, King, & Shmukler, 2014; Chechulin et al., 2014; Merkesdal & Mau, 2005; Otani & Baden, 2009; Rundell et al., 2017) O ur study also found that some proxy measures for socioeconomic status used, such as education, poverty level, and insurance status were either not associated with expenditure group, or associated such that higher socioeconomic status was rel ated to higher spending. These results compare favorably to prior literature that has assessed the influence of insurance and education on healthcare expenditures for musculoskeletal pain in the US. For instance, Luo et al. (Luo, Pietrobon, Sun, Liu, & Hey, 2004) reported higher per capita expenditures related to low back pain among those who were medically insured. In a study on rheumatoid arthritis, Clarke et al. 51

PAGE 52

reported higher direct costs associated with r eceiving more formal education. (A. E. Clarke et al., 1999) These relationships highlight the enabling nature of insurance and income that favor higher spending in the US healthcare system. Our analyses produced some notable findings related to the influence of psychological factors on expenditures. Psychol ogical distress, depression and poor mental health were associated with greater odds of higher order expenditure group membership in univariate but not multivariate analyses This suggests that other factors like pain interference, number of musculoskeletal pain conditions, or insurance fully mediate the relationship between psychological characteristics and healthcare utilization These findings contrast with traditional pain outcomes studies that generally show a more robust relationship between psycholog ical factors and clinical outcomes. However, readers are cautioned that psychological measures used in this study were not specific to pain. Additionally, our models controlled for variables commonly used as clinical outcomes in studies on pain (e.g. pain interference, self reported function, or pain intensity). Therefore, while psychological characteristics might be useful for predicting utilization when considered in isolation, they become less informative when considered alongside other painrelated variables. Of particular interest was the influence of prescription medication on the persistence of high costs. It might be expected that prescription medication costs and overall costs would be highly correlated, but this analysis demonstrated that those with higher prescription medication use in the first year are more likely to continue to have high costs in the following year. Prescription medication use f or pain was a significant predictor even after accounting for pain interference. Therefore, use of prescription 52

PAGE 53

medication for pain has a significant effect on persistently high spending that is independent of pain interference alone. Our study design did not directly assess appropriate use of prescription medication for pain (opioid and nonopioid) but does suggest that high utilization has an important downstream economic impac t s, in addition to known adverse public health effects. (J. L. Clarke, Skoufalos, & Scranton, 2016; Morris & Mir, 2015) While identifying predictors of LOW expenditure group membership was not the primary focus of this analysis, our results a llow us to comment on characteristics associated with its membership. Characteristics of the LOW group may provide insight on factors that protect against unnecessary or wasteful spending, a desirable outcome when the intent is to control costs and optimiz e value. Characteristics of the LOW group may also illustrate where access to care is limited a potentially detrimental situation leading to health disparities. Race and insurance status were both associated with LOW group membership, with black and unins ured respondents having higher odds of LOW group membership. Interestingly, low income individuals were less likely than individuals classified as poor or near poor to be in the MEDIUM or HIGH expenditure group. It is plausible that access to insurance cov erage or other economic demands impact healthcare usage disproportionately in this group compared to those at worse levels of poverty who have greater access to public services and subsidies. While our study design is not able to differentiate among met and unmet healthcare needs, it suggests further health disparities research in musculoskeletal pain is warranted. Strengths of this study include the use of longitudinal data over 2 years to assess predictors of persistence of high cost utilization rather th an point high cost utilization. 53

PAGE 54

MEPS is a dataset rich in variables representing a variety of predisposing and enabling factors, as well as other perceived and evaluated need characteristics Survey weights and statistical procedures were used to account f or differential sample selection probabilities and adjust for nonresponse to provide nationally representative estimates that are suitable for informing broader health policy. We explored two different approaches for identifying those with persistently high costs, one that considered individual predictors, and another that determined whether those with persistently high costs can be identified as members of common demographic cohorts Convergence of results from these two approaches lends strength to the fi ndings. We also took multiple steps to assess the sensitivity of our findings, by alternately defining high cost utilization, examining the stability of results based on diagnosis category and exploring multiple regression methodologies. As a result, predi ctors that are consistent across sensitivity analyses can be viewed as robust. This analysis had some limitations that should be conside red when interpreting results. First, the survey design did not allow for assessment of self reported measures to coinc ide with onset of the injury or condition. All demographic and health related variables were collected from the earliest data collection point in the panel. However, in some cases, self reported variables may have been measured prior to the episode of care. This would most significantly impact variables related to pain diagnoses, such as pain interference, general health, perceived health status, psychological distress or depression. Readers are also cautioned that this analysis excludes medically complex pain conditions that are more likely to contribute to persistently high costs. Therefore, 54

PAGE 55

the results of this study cannot be extended to medically complex pain conditions and predictors of high cost utilization may be different in those populations. Second, we performed cluster analysis in the second aim using common demographic factors. The goal was to derive easily identifiable patient subgroups and determi ne whether high cost utilization could be identified within these subgroups. It is possible that clustering on other variables may have provided greater discrimination between high and low expenditure groups, however we wanted to have limited a priori crit eria for group selection, as this approach best suited our aim. Third, some expenditures are not included in the MEPS survey, such as over the counter medications and durable medical equipment. Indirect costs, such as lost wages due to missed work days, a re also not included. The direct method of expenditure calculation is likely to provide a low end estimate of costs for musculoskeletal pain. (Martin et al., 2008) Moreover, this appr oach is likely to substantially underestimate the total economic burden of musculoskeletal pain, as indirect costs are estimated to comprise almost 75% of total costs attributable to musculoskeletal pain. (Martin et al., 2008; M. M. Ward, 2002) While it was our goal to specifically examine direct healthcare costs, we acknowledge that different factors may predict total costs. However, results are important for any stakeholder who may be interested in reducing avoidable direc t costs, such as patients, providers, health services organizations and payers. Future Research Our findings have important implications for future research and healthcare policy. We identified predictors of high healthcare expenditures, but these should not be confused with predictors of wasteful or necessarily preventable spending. While we propose that the presence of persistently high costs may indicate suboptimal healthcare 55

PAGE 56

management of those with uncomplicated musculoskeletal pain conditions, the st udy design is not such that we can determine the appropriateness of treatment or i dentify unhelpful utilization. As Newton and LeFabvre (Newton & Lefebvre, 2015) stress, the quality of care and patient experience must not be ignored in an effort blindly directed at reducing expenditures. Perhaps an appropriate conclusion given our findings is that cost reduction policies may be most beneficial if targeted toward de incentivizing use of low value services, such as imaging for acute, uncomplicated low back pain (Chou, Qaseem, Owens, Shekelle, & Clinical Guidelines Committee of the American College of Physicians, 2011) These policies may be especially impactful among those with private insurance at risk for high expenditures. An important direction for future research is to identify instances of unhelpful utilization that contribute to high costs. Distinguishing wastefu l from helpful utilization, particularly in survey or claims data with limited patient reported outcomes, is extremely difficult. Therefore, these findings represent a starting point for studying the circumstances under which wasteful utilization may be most likely to occur. The current study focused on predictors of persistently high cost utilization, however other metrics for health services use are worth examining in this population. For instance, individuals with excessively high use of particular services (i.e. super utilizers) have become an important research target, particularly in an effort to reduce unnecessary emergency department use. We provided descriptive information for different services among expenditure groups, however our methodology was not similar enough to common super utilizer analyses to make any meaningful direct comparisons of results. Future research should examine high expenditure groups more closely to 56

PAGE 57

distinguish reasons for high cost utilization for musculoskeletal pain, possibly through a person centered attribution approach. (Guilcher et al., 2016) Understanding whether high cost utilization is more strongly driven by volume of utilization, cost per utilization episode, utilization of specific types of services or all of the above, will better inform health policy and pathway redesign for treatment of musculoskeletal pain. 57

PAGE 58

Figure 2 1. Sample Selection Flowchart 58

PAGE 59

Table 2 1 Predictor variables Category Variable name Sociodemographic variables Age Sex Race Ethnicity Years of education BMI Smoking status Poverty category Employment status Metropolitan Statistical Area Census region Health related variables Charlson comorbidity index Days of work missed due to illness General Health Status Pain interference General psychological distress Depression Perceived physical and mental health status Attitudes about health Total musculoskeletal diagnoses Diagnosis type Insurance and healthcare expenditure variables Usual Care provider Health insurance Number of prescription medications 59

PAGE 60

Table 2 2. Unweighted and weighted percentile group sizes Percentile Group Frequency Weighted Frequency Standard Error of Weighted Frequency Percent Standard Error of Percent Low 1 504 15 963 948 58 2318 10.6 0.4 Medium 10 983 128 931 988 3 002 650 85.5 0.4 High 498 5 896 762 337 695 3.9 0.2 Total 12 985 150 792 697 3 295 002 100 .0 60

PAGE 61

Table 2 3. Weighted percentile group means for demographic and health related information. Data are mean standard deviation (range) PCS = physical component subscale of the SF 12, MCS = mental component subscale of the SF 12. Variable* Low N=1504 Medium N=10983 High N=498 Total N=12 985 p value Age (years) 50.8

PAGE 62

Table 2 4. Demographic information for the sample (total) and each group (none, low, medium, high) with weighted population percentage estimates. Variable* Low N=1504 Medium N=10983 High N=498 Total N=12 985 p value Sex Male 558 (39.0) 4 026 (39.6) 156 (35.4) 4 740 (39.4) 0.35 Female 946 (61.0) 6 957 (60.4) 342 (64.6) 8 245 (60.6) Race White 1 060 (82.2) 8 164 (85.8) 380 (87.7) 9 604 (85.5) <.001 Black 340 (13.0) 2 035 (9.3) 87(8.4) 2 462 (9.7) Other 104 (4.8) 784 (4.9) 31 (3.9) 919 (4.8) Ethnicity Hispanic 310 (10.6) 1 803 (8.2) 79 (8.9) 2 192 (8.5) 0.0 04 Non Hispanic 1 194 (89.4) 9 180 (91.8) 419 (91.1) 10 793 (91.5) Poverty category Poor or Near Poor 397 (16.8) 2 508 (15.5) 151 (20.9) 3 056 (15.8) <.001 Low Income 266 (16.9) 1 638 (12.9) 73 (13.2) 1 977 (13.3) Middle Income 433 (30.2) 3 172 (28.3) 126 (25.9) 3 731 (28.4) High Income 408 (36.1) 3 665 (43.3) 148 (40.1) 4 221 (42.4) Geographic r egion Northeast 210 (15.0) 1 754 (18.4) 83 (20.3) 2 047 (18.1) 0.07 Midwest 353 (25.6) 2 492 (23.7) 101 (20.4) 2 946 (23.8) South 563 (35.4) 4 018 (34.8) 170 (31.6) 4 751 (34.7) West 378 (24.0) 2 719 (23.2) 144 (27.7) 3 241 (23.5) Smoking status Yes 297 (19.5) 2 001 (17.7) 112 (22.3) 2 410 (18.1) 0.08 No 1 170 (80.5) 8 731 (82.2) 376 (77.7) 10 277 (81.9) Education High school diploma or less 1 017 (60.8) 6 882 (57.2) 317 (58.8) 8 216 (57.7) 0.11 Some college/college degree 487 (39.2) 4 101 (42.8) 181 (41.2) 4 769 (42.3) Employment Empl oyed or have a job to return to 840 (58.0) 5 805 (56.2) 187 (41.9) 6 832 (55.8) <.001 Unemployed 660 (42.0) 5 147 (43.8) 310 (58.1) 6 117 (44.1) *Data are sample size (weighted population percentage estimates) 62

PAGE 63

Table 2 4. (continued). Variable* Low N=1504 Medium N=10983 High N=498 Total N=12 985 p value Pain i nterference Not at all, a little bit, or moderately 1 255 (86.0) 7 965 (76.8) 169 (43.1) 9 389 (76.5) <.001 Quite a bit or extremely 230 (14.0) 2 872 (23.2) 320 (56.9) 3 422 (23.5) Overcome Disagree strongly, disagree somewhat or uncertain 1 209 (80.6) 8 930 (82.3) 434 (89.0) 10 573 (82.4) 0.01 Agree somewhat or agree strongly 258 (19.4) 1 743 (17.7) 44 (11.0) 2 045 (17.6) Insurance Uninsured 382 (20.4) 1 879 (14.1) 47 (8.8) 2 308 (14.6) <.001 Public Insurance 424 (24.7) 3 258 (25.4) 214 (35.7) 3 896 (25.7) Private Insurance 698 (54.8) 5 846 (60.5) 237 (55.5) 6 781 (59.7) Health Excellent, very good, or good 1 134 (79.0) 8 032 (78.0) 242 (54.7) 9 408 (77.2) <.001 Fair or poor 370 (21.0) 2 942 (22.0) 256 (45.3) 3 568 (22.8) Mental h ealth Excellent, very good, or good 1 368 (92.8) 9 726 (90.4) 383 (83.0) 11 477 (90.4) <.001 Fair or poor 134 (7.2) 1 253 (9.6) 113 (17.0) 1 500 (9.6) Usual care p rovider Yes 1 264 (85.7) 9 694 (89.5) 462 (94.6) 11 420 (89.3) <.001 No 233 (14.3) 1 232 (10.5) 35 (5.4) 1 500 (10.7) Missed

PAGE 64

Table 2 5. Unweighted frequency table of diseases of the musculoskeletal system and connective tissue ICD9 codes. ICD 9 diagnostic c ategory ICD9 code Low (n=1504) M edium (n=10 983) High (n=498) Total Diseases of the Musculoskeletal system and connective t issue 719 433 3 470 206 4 109 724 260 2 403 207 2 870 716 240 2 176 179 2 595 715 140 1 907 169 2 216 729 165 1 440 154 1 759 733 65 804 31 900 722 28 726 117 871 728 58 562 57 677 726 44 527 35 606 723 37 425 40 502 727 16 331 22 369 734 6 183 29 218 737 4 80 8 92 721 6 68 6 80 717 2 49 8 59 735 1 47 2 50 738 2 40 7 49 718 1 32 3 36 736 0 33 3 36 730 2 15 2 19 725 1 6 3 10 64

PAGE 65

Table 2 6 Unweighted frequency table of musculoskeletal injury ICD 9 codes. ICD 9 diagnostic c ategory ICD 9 code Low (n=1504) Medium (n=10 983) High (n=498) Total Musculoskeletal i njuries 959 86 1 076 96 1 258 847 40 354 16 410 845 35 318 8 361 840 10 286 27 323 848 28 210 11 249 924 17 206 8 231 844 20 190 9 219 836 2 133 9 144 842 17 109 2 128 825 8 107 6 121 839 8 87 4 99 826 9 79 3 91 805 3 71 14 88 814 2 74 6 82 824 2 75 4 81 923 6 69 5 80 922 6 63 3 72 816 5 62 0 67 843 3 49 4 56 815 2 47 2 51 822 0 44 5 49 831 2 46 0 48 812 0 46 0 46 818 0 43 3 46 827 4 34 8 46 820 0 27 3 30 810 1 21 0 22 835 0 15 7 22 846 3 18 1 22 829 1 18 2 21 837 1 9 5 15 834 2 11 1 14 808 0 10 1 11 813 0 9 0 9 821 0 5 3 8 823 0 7 1 8 832 0 6 0 6 841 2 3 0 5 811 0 4 0 4 838 0 4 0 4 837 0 1 0 1 65

PAGE 66

Table 2 7 Weighted mean annual expenditures for musculoskeletal pain diagnoses and all healthcare by percentile groups. Group Variable Mean (Dollars) Median Standard error of m ean 95% CI for m ean Upper Lower Low (n=1 504) Year 1 musculoskeletal expenditures 36.43 35.00 0.73 35.00 37.86 Year 2 musculoskeletal expenditures 6.01 0 0.50 5.03 6.99 Mean t otal musculoskeletal expenditures 21.22 20.41 0.43 20.38 22.06 Year 1 overall healthcare e xpenditures 5 135.50 1 677.60 407.12 4 332.79 5 938.21 Year 2 overall healthcare e xpenditures 6 405.08 1 874.79 603.84 5 214.51 7 595.66 Medium (n=10, 983) Year 1 musculoskeletal expenditures 2 078.33 505.64 89.27 1 902.32 2 254.35 Year 2 musculoskeletal expenditures 783.83 60.97 37.95 709.00 858.66 Mean t otal musculoskeletal expenditures 1 431.08 420.06 47.39 1 337.65 1 524.52 Year 1 overall healthcare e xpenditures 7 548.99 3 719.39 162.66 7 228.28 7 869.69 Year 2 o verall healthcare e xpenditures 7 852.32 3 343.37 182.28 7 492.92 8 211.72 High (n=498) Year 1 m usculoskeletal expenditures 13 511.00 7 161.42 949.24 11 639.17 15 382.32 Year 2 m usculoskeletal expenditures 11 584.00 6 617.36 807.55 9 991.38 13 175.82 Mean t otal m usculoskeletal expenditures 12 547.00 8 036.64 646.47 11 272.56 13 821.79 Year 1 overall healthcare e xpenditures 23 848.00 15 936.00 1 308.09 21 268.71 26 426.93 Year 2 overall healthcare e xpenditures 24 647.00 16 917.00 1 154.65 22 370.39 26 923.54 66

PAGE 67

Figure 2 2. Weighted mean annual expenditures for musculoskeletal pain diagnoses and all healthcare by percentile groups. 0 5000 10000 15000 20000 25000 30000 LOW MEDIUM HIGHAnnual Expenditures for MSK Conditons and All Healthcare Year 1 MSK Expenditures Year 1 Overall Expenditures Year 2 MSK Expenditures Year 2 Overall Expenditures67

PAGE 68

Figure 2 3. Weighted mean percentage of total musculoskeletal pain expenditures attributable to each event type by percentile group. 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Low (Year 1) Low (Year 2) Medium (Year 1) Medium (Year 2) High (Year 1) High (Year 2)Percentage of Total Annual Musculoskeletal Pain Expenditures Percentile Group Emergency Room Home Health Office-Based Medical Provider Outpatient Inpatient Precription Medication68

PAGE 69

Table 2 8 Univariate logistic regression results (15% expenditure percentile criteria) Variable Group Generalized logit Ordinal regression Odds ratio 95% CI Odds ratio 95% CI Lower Upper Lower Upper Age Medium 1.01 1.01 1.02 1.01 1.01 1.02 High 1.02 1.02 1.03 Sex (female vs male) Medium 0.98 0.85 1.13 1.03 0.91 1.17 High 1.15 0.87 1.5 0 Race (Black vs White) Medium 0.68 0.57 0.82 0.7 0 0.6 0 0.83 High 0.59 0.44 0.81 Race (Other vs White) Medium 0.98 0.75 1.28 0.91 0.73 1.14 High 0.74 0.42 1.31 Ethnicity (Hispanic vs. non Hispanic) Medium 0.74 0.62 0.87 0.81 0.68 0.95 High 0.82 0.57 1.2 0 Education (college degree/some college vs. high school) Medium 1.15 1 .00 1.34 1.11 0.97 1.26 High 1.12 0.84 1.49 Body mass index (BMI) Medium 1 .00 0.99 1.01 1.01 0.99 1.02 High 1.02 0.99 1.04 Smoking status (smoker vs. non smoker) Medium 0.88 0.76 1.02 0.99 0.86 1.15 High 1.18 0.86 1.6 0 Poverty level (low income vs poor/near poor) Medium 0.81 0.64 1.02 0.76 0.61 0.95 High 0.61 0.4 0 0.94 Poverty level (middle income vs. poor/near poor) Medium 1.02 0.84 1.23 0.91 0.76 1.09 High 0.73 0.51 1.04 Poverty level (high income vs. poor/near poor) Medium 1.28 1.05 1.57 1.08 0.89 1.3 0 High 0.92 0.64 1.32 Employment (unemployed vs employed) Medium 1.08 0.93 1.25 1.25 1.1 0 1.42 High 1.87 1.41 2.47 Metropolitan statistical area (MSA) (Non MSA vs MSA) Medium 0.92 0.75 1.13 0.89 0.74 1.07 High 0.76 0.51 1.13 Census region (Midwest vs Northeast) Medium 0.74 0.57 0.97 0.74 0.6 0 0.93 High 0.57 0.36 0.91 Census region (South vs Northeast) Medium 0.82 0.64 1.07 0.83 0.67 1.02 High 0.7 0 0.45 1.08 Census region (West vs Northeast) Medium 0.81 0.62 1.04 0.88 0.7 0 1.09 High 0.87 0.57 1.33 Charlson comorbidity index Medium 1.06 0.99 1.14 1.08 1.02 1.14 High 1.16 1.04 1.29 Missed work days

PAGE 70

Table 2 8 Continued. Variable Group Generalized logit Ordinal regression Odds ratio 95%CI Odds ratio 95% CI Lower Upper Upper Lower SF 8 Physical Component Scale Medium 0.98 0.97 0.99 0.96 0.96 0.97 High 0.92 0.91 0.93 SF 8 Mental Component Scale Medium 1 .00 0.99 1.01 0.99 0.98 0.99 High 0.97 0.96 0.98 Pain interference (quite a bit/extremely vs moderately/a little bit/not at all) Medium 1.85 1.53 2.23 3.01 2.56 3.55 High 8.3 0 6.32 10.89 Distress (K 6 Scale) Medium 1.01 1 .00 1.02 1.05 1.03 1.06 High 1.1 0 1.08 1.13 Depression (PHQ 2 Score) Medium 1.05 1 .00 1.09 1.16 1.1 0 1.21 High 1.35 1.26 1.45 Perceived health status (fair/poor vs excellent/very good/good) Medium 1.05 0.9 0 1.23 1.58 1.36 1.84 High 3.05 2.37 3.93 Perceived mental health status (fair/poor vs excellent/very good/good) Medium 1.32 1.05 1.66 1.62 1.33 1.97 High 2.6 0 1.86 3.62 Can overcome ills (disagree strongly/somewhat vs uncertain/agree somewhat/agree strongly) Medium 1.13 0.93 1.39 1.27 1.07 1.52 High 2.09 1.28 3.41 Have usual health provider (yes vs. no) Medium 1.39 1.14 1.7 0 1.51 1.27 1.8 0 High 2.85 1.7 0 4.77 Insurance (public vs uninsured) Medium 1.5 0 1.24 1.82 1.76 1.46 2.11 High 3.41 2.2 0 5.29 Insurance (private vs uninsured) Medium 1.61 1.37 1.9 0 1.6 0 1.38 1.87 High 2.43 1.62 3.66 Diagnosis (musculoskeletal injury vs musculoskeletal disease only) Medium 1.81 1.53 2.14 1.72 1.51 1.95 High 2.57 1.99 3.31 Total musculoskeletal conditions in Year 1 Medium 2.91 2.53 3.34 2.11 1.99 2.25 High 5.38 4.62 6.27 Total prescription medications in Year 1 Medium 1.29 1.24 1.34 1.3 0 1.26 1.34 High 1.66 1.57 1.75 70

PAGE 71

Table 2 9 Significant results from m ultivariate logistic regression (15% expenditure percentile criteria) Variable Group Generalized logit Ordinal regression Odds ratio 95% CI Odds ratio 95% CI Lower Upper Lower Upper Age Medium 1.01 1.01 1.02 1.01 1.01 1.01 High 1.02 1.01 1.03 Race (Black vs White) Medium 0.7 0 0.58 0.85 0.72 0.61 0.85 High 0.58 0.4 0 0.84 Poverty level (low income vs poor/near poor ) Medium 0.7 0 0.55 0.9 0 0.69 0.55 0.86 High 0.48 0.28 0.81 Census region (Midwest vs Northeast) Medium 0.76 0.57 1.01 0.74 0.58 0.94 High 0.49 0.3 0 0.8 0 Missed work days

PAGE 72

Table 2 10. Significant results from multivariate logistic regression sensitivity analysis (10% expenditure percentile criteria) Variable Group Generalized Logit Ordinal Regression Odds ratio 95% CI Odds ratio 95% CI Lower Upper Lower Upper AGE Medium 1.01 1 .00 1.02 1.01 1.01 1.02 High 1.04 1.02 1.05 Sex (female vs male) Medium 0.88 0.73 1.05 0.83 0.71 0.97 High 0.62 0.41 0.92 Race (Black vs White) Medium 0.78 0.64 0.96 0.81 0.68 0.98 High 0.87 0.56 1.34 Education (college degree/some college vs high school) Medium 1.2 0 0.98 1.48 1.21 1.01 1.45 High 1.43 0.9 0 2.29 Body mass index (BMI) Medium 0.99 0.98 1 .00 0.99 0.98 0.99 High 0.97 0.95 0.99 Poverty level (high income vs poor/near poor ) Medium 1.42 1.05 1.93 1.28 0.94 1.73 High 1.19 0.56 2.53 Metropolitan statistical area (MSA) (Non MSA vs MSA) Medium 0.97 0.77 1.22 0.89 0.73 1.09 High 0.58 0.35 0.96 Census region (South vs Northeast) Medium 0.81 0.6 0 1.11 0.75 0.57 0.98 High 0.46 0.24 0.88 Missed work days

PAGE 73

Table 2 11. Significant results from m ultivariate l ogistic regression sensitivity analysis (20% expenditure percentile criteria) Variable Group Generalized Logit Ordinal Regression Odds ratio 95% CI Odds ratio 95% CI Lower Upper Lower Upper Age Medium 1.01 1.01 1.01 1.01 1.01 1.01 High 1.01 1 .00 1.02 Race (Black vs White) Medium 0.75 0.64 0.87 0.74 0.64 0.85 High 0.59 0.41 0.83 Body mass index (BMI) Medium 0.99 0.98 1 .00 0.99 0.98 0.99 High 0.98 0.96 0.99 Smoking status (smoker vs non smoker) Medium 0.84 0.71 0.98 0.87 0.74 1.01 High 0.81 0.57 1.14 Poverty level Medium 0.79 0.62 1.02 0.73 0.59 0.91 High 0.5 0 0.32 0.76 Census region (South vs Northeast) Medium 0.84 0.69 1.01 0.83 0.69 0.99 High 0.66 0.43 0.99 Census region (West vs Northeast) Medium 0.79 0.66 0.95 0.87 0.71 1.06 High 0.82 0.53 1.27 Missed work days

PAGE 74

Table 2 12. Logistic regression results for sensitivity analysis using nonspecific musculoskeletal pain ICD 9 codes (15% expenditure percentile criteria) Variable Percentile Group Generalized Logit Ordinal Regression Odds ratio 95% CI Odds ratio 95% CI Lower Upper Lower Upper Age Medium 1.01 1.00 1.02 1.01 1.01 1.02 High 1.01 1.00 1.03 Race (Black vs White) Medium 0.72 0.57 0.90 0.70 0.57 0.87 High 0.48 0.29 0.81 Body mass i ndex (BMI) Medium 1.00 0.98 1.01 0.99 0.98 0.99 High 0.97 0.95 0.99 Poverty level (low Income vs poor/n ear p oor ) Medium 0.77 0.58 1.03 0.74 0.57 0.95 High 0.49 0.26 0.93 Census region (Midwest vs Northeast) Medium 0.87 0.66 1.16 0.78 0.60 1.01 High 0.43 0.23 0.77 Missed work days

PAGE 75

Table 2 13. Cluster profiles Variable White, Non Hispanic, Private Insurance, multiple condition ( Cluster 1 ) White, Hispanic, Mixed Insurance, Mixed condition ( Cluster 2 ) Non White, Non Hispanic, Mixed insurance, Mixed condition ( Cluster 3 ) Older, White, Non Hispanic, Public insurance, mixed condition ( Cluster 4 ) Younger, Primarily White, Non Hispanic, Uninsured, multiple condition (C luster 5 ) White, Non Hispanic, Private Insurance, one condition ( Cluster 6 ) n = 2 145 (16.5%) n = 2 060 (15.9%) n = 2 879 (22.2%) n = 1 985 (15.3%) n = 1 529 (11.8%) n = 2 387 (18.4%) Age (SEM) 54.67 (0.44) 49.71 (0.52) 51.45 (0.44) 66.06 (0.44) 44.27 (0.46) 51.16 (0.43) Ethnicity Non Hispanic (100%) Hispanic (0%) Non Hispanic (0%) Hispanic (100%) Non Hispanic (95.4%) Hispanic (4.6%) Non Hispanic (100%) Hispanic (0%) Non Hispanic (100%) Hispanic (0%) Non Hispanic (100%) Hispanic (0%) Insurance Private (100%) Public (0%) Uninsured (0%) Private (36.2%) Public (34.3%) Uninsured (29.5%) Private (52.2%) Public (41.9%) Uninsured (5.9%) Private (0%) Public (100%) Uninsured (0%) Private (0%) Public (0%) Uninsured (100%) Private (100%) Public (0%) Uninsured (0%) Race White (100%) Black (0%) Other (0%) White (100%) Black (0%) Other (0%) White (0%) Black (68.1%) Other (31.9%) White (100%) Black (0%) Other (0%) White (67.2%) Black (32.8%) Other (0%) White (100%) Black (0%) Other (0%) Total MSK conditions One (0%) Two or more (100%) One (60.5%) Two or more (39.4%) One (57.1%) Two or more (42.9%) One (44.1%) Two or more (55.9%) One (55.9%) Two or more (44.1%) One (100%) Two or more (0%) 75

PAGE 76

Table 2 14. Eigenvalues from cluster analysis Function Eigenvalue % of Variance Cumulative % Canonical Correlation 1 14.350 70.0 70.0 .967 2 3.533 17.2 87.2 .883 3 2.069 10.1 97.3 .821 4 0 .552 2.7 100.0 .596 Table 2 15. Standardized Canonical Discriminant Function Coefficients Function 1 2 3 4 Insurance 0 .044 0 .082 1.003 0 .005 Race 0 .281 0 .957 0 .120 0 .035 T otal musculoskeletal conditions 0 .027 0 .017 0 .000 1.000 Ethnicity 0 .991 0 .180 0 .087 0 .026 Table 2 16. Pooled within groups correlations between discriminating variables and standardized canonical discriminant functions Function 1 2 3 4 Ethnicity 0 .958 0 .283 0 .018 0 .031 Race 0 .186 0 .978 0 .088 0 .012 Insurance 0 .061 0 .142 0 .988 0 .004 T otal musculoskeletal conditions 0 .018 0 .041 0 .007 0 .999 76

PAGE 77

Table 2 17. Distribution of categorical variables among clusters* Variable Cluster 1 (n=2145) Cluster 2 (n=2060) Cluster 3 (n=2879) Cluster 4 (n=1985) Cluster 5 (n=1529) Cluster 6 (n=2387) n % n % n % n % n % n % Sex Male 814 40.3 716 37.6 955 35.8 619 32.9 610 43.9 1026 43.4 Female 1331 59.7 1344 62.4 1924 64.2 1366 67.1 919 56.1 1361 56.6 Poverty category Poor or Near Poor 134 5.4 689 27.0 870 24.6 666 26.1 578 29.8 119 3.9 Low Income 203 8.4 413 19.1 428 13.6 390 19.7 316 19.0 227 8.6 Middle Income 648 27.9 584 29.2 843 30.6 486 25.7 408 29.5 762 29.1 High Income 1160 58.3 374 24.7 738 31.2 443 28.5 227 21.7 1279 58.4 Geographic region Northeast 369 20.5 330 14.6 490 17.7 303 18.4 164 13.1 391 19.2 Midwest 624 26.6 205 9.2 481 17.5 543 24.7 386 25.2 707 27.3 South 641 30.6 634 31.4 1291 43.1 730 36.2 697 38.8 758 32.4 West 511 22.3 891 44.8 617 21.8 409 20.7 282 22.9 531 21.1 Smoking status Yes 301 14.2 262 13.1 531 17.9 434 19.9 553 36.2 329 13.9 No 1813 85.8 1759 86.9 2235 82.1 1506 80.1 936 63.8 2028 86.1 Education High school diploma or less 1030 47.1 1608 72.9 1843 60.3 1433 68.8 1102 69.3 1200 47.9 Some college/College degree 1115 52.9 452 27.1 1036 39.7 552 31.2 427 30.7 1187 52.1 Employment Employed or have a job to return to 1367 65.0 1080 55.1 1441 54.4 332 17.5 894 62.2 1718 73.5 Unemployed 777 35.0 970 44.9 1435 45.6 1649 82.5 620 37.8 666 26.5 Pain i nterference Not at all, a little bit, or moderately 1603 76.6 1435 72.6 1981 73.0 1198 64.4 1053 70.7 2119 90.6 Quite a bit or extremely 525 23.4 587 27.4 854 27.0 756 35.6 460 29.3 240 9.4 Overcome Disagree strongly, disagree somewhat or uncertain 1758 84.5 1769 86.9 2403 85.4 1664 86.3 1158 76.2 1821 77.5 Agree somewhat or agree strongly 337 15.5 239 13.1 371 14.6 250 13.7 333 23.8 515 22.5 Perceived health status Excellent, very good, or good 1723 81.4 1283 66.1 1950 70.9 1267 67.1 1043 70.6 2142 90.1 Fair or poor 421 18.6 776 33.9 926 29.1 718 32.9 485 29.4 242 9.9 Perceived mental health status Excellent, very good, or good 1970 92.5 1764 86.4 2492 88.1 1676 86.4 1292 85.0 2283 96.0 Fair or poor 173 7.5 294 13.6 387 11.9 308 13.6 235 15.0 103 4.0 *Chi -square tests significant at p<.05 for group differences among all variables. 77

PAGE 78

Table 2 17 (continued). Variable Cluster 1 (n=2145) Cluster 2 (n=2060) Cluster 3 (n=2879) Cluster 4 (n=1985) Cluster 5 (n=1529) Cluster 6 (n=2387) n % n % n % n % n % n % Usual care p rovider Yes 1997 93.0 1726 84.9 2562 88.2 1868 94.9 1120 74.4 2147 90.3 No 139 7.0 318 15.1 299 11.8 113 5.1 401 25.6 230 9.7 Missed 1 work day in Year 1 Yes 1656 76.8 1641 79.3 2315 78.6 1897 95.7 1186 77.6 2043 85.2 No 489 23.2 419 20.7 564 21.4 88 4.3 343 22.4 344 14.8 Diagnosis Musculoskeletal d isease only 1301 59.6 1541 73.2 2062 70.8 1431 72.1 930 60.7 1788 73.9 Musculoskeletal i njury only 86 4.1 288 14.6 403 15.2 166 8.5 326 21.3 599 26.1 Musculoskeletal disease and i njury 758 36.3 231 12.3 414 14.0 388 19.5 273 18.0 0 0.0 Metropolitan Statistical Area (MSA) Non MSA 347 16.7 166 7.1 355 10.2 533 24.6 307 21.4 427 18.3 MSA 1798 83.3 1894 92.9 2524 89.8 1452 75.4 1222 78.6 1960 81.7 *Chi -square tests significant at p<.05 for group differences among all variables. 78

PAGE 79

Table 2 18. Descriptive analysis for continuous variables among clusters* Variable Cluster Membership Cluster 1 (n=2145) Cluster 2 (n=2060) Cluster 3 (n=2879) Cluster 4 (n=1985) Cluster 5 (n=1529) Cluster 6 (n=2387) Mean SEM Mean SEM Mean SEM Mean SEM Mean SEM Mean SEM Body mass index (BMI) 28.89 0.19 29.42 0.17 29.77 0.2 0 29.11 0.23 29.37 0.21 27.9 0.15 Charlson comorbidity index 0.55 0.02 0.51 0.03 0.56 0.02 0.88 0.03 0.46 0.03 0.51 0.02 SF 12 physical function subscale 43.87 0.25 43.72 0.37 42.67 0.32 37.88 0.33 43.43 0.44 49.11 0.21 SF 12 mental function subscale 51.05 0.24 47.73 0.33 48.99 0.26 49.1 0.3 46.72 0.39 51.77 0.2 0 Distress 3.93 0.1 5.15 0.16 4.75 0.12 5.15 0.14 5.76 0.18 3.12 0.09 Depression 0.8 0 0.03 1.21 0.05 1.11 0.04 1.2 0 0.04 1.39 0.06 0.61 0.02 Total Musculoskeletal c onditions 2.66 0.03 1.61 0.03 1.67 0.03 1.98 0.03 1.78 0.04 1 .00 0 .01 Mean t otal Prescriptions (Year 1) 2.15 0.07 1.62 0.07 1.84 0.06 2.11 0.09 1.95 0.1 0 0.75 0.03 Mean t otal Musculoskeletal E xpenditures (Year 1) 3941.27 284.66 2116.31 285.57 2304.26 210.86 2153.56 138.71 1820.49 173.95 1235.34 104.55 Mea n t otal Musculoskeletal E xpenditures (Year 2) 1742.6 131.96 942.31 96.87 1088.08 165.51 1496.81 157.06 727.91 84.49 547.05 74.89 Mean t otal Musculoskeletal E xpenditures (Panel) 2841.93 166.55 1529.31 155.36 1696.17 168.95 1825.18 111.69 1274.2 110.86 891.2 67.18 *ANOVA omnibus tests significant at p<.05 for group differences among all variables.79

PAGE 80

Table 2 19. Distribution of expenditure percentile groups by cluster membership Percentile Group Cluster Number Total 1 2 3 4 5 6 Low Count 73 289 345 191 24 0 366 1504 % within Cluster Number 3.4 14.0 12.0 9.6 15.7 15.3 11.6 % of Total 0.6 2.2 2.7 1.5 1.8 2.8 11.6 Medium Count 1921 1697 b 2425 1686 1255 1999 10983 % within Cluster Number 89.6 82.4 84.2 84. 9 82.1 83.7 84.6 % of Total 14.8 13.1 18.7 13.0 9.7 15.4 84.6 High Count 151 74 109 108 34 22 498 % within Cluster Number 7.0 3.6 3.8 5.4 2.2 0.9 3.8 % of Total 1.2 0.6 0.8 0.8 0.3 0.2 3.8 Total Count 2145 2060 2879 1985 1529 2387 12985 % within Cluster Number 100.0 100.0 100.0 100.0 100.0 100.0 100.0 % of Total 16.5 15.9 22.2 15.3 11.8 18.4 100.0 Each subscript letter denotes a subset of Cluster Number categories whose column proportions do not differ significantly from each other at the .05 level.80

PAGE 81

Figure 2 4. Weighted Mean Annual Expenditures for musculoskeletal conditions among 6 demographic clusters with 95% confidence interval 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 1 2 3 4 5 6Mean Expenditures (dollars) Cluster Membership Total Year 1 MSK Costs Total Year 2 MSK Costs 81

PAGE 82

CHAPTER 3 RIS K MODEL DEVEL OPMENT FOR ADDITIONAL HEALTHCARE USE FOLLOWING AN EPISODE OF PHYSICAL THERAPY FOR MUSCULOSKELETAL PAIN Introduction Physical therapy is a non pharmacological treatment approach for musculoskeletal pain and its effectiveness for improving pain and disability has been established for many musculoskeletal pain conditions. (Boyles, Toy, Mellon, Hayes, & Hammer, 2011; Brge, Monnin, Berchtold, & Allet, 2016; G. D. Deyle et al., 2000; Gail D. Deyle et al., 2005; Freburger, Carey, & Holmes, 2006; Kuhn et al., 2013) Non pharmacological treatment is now recommended as fro ntline management for musculoskeletal pain, which will lead to increased utilization of physical therapy services. (J. L. Clarke et al., 2016; Dowell, Haegerich, & Chou, 2016; Von Korff et al., 2016) As frontline providers, physical therapists must be able to identify patients who are at risk for escalation of care beyond the physical therapy episode (e.g. subsequent use of surgery, injection, opioids, etc.) Prospective identification of those who are likely to need additional healthcare services will allow providers to distribute limited healthcare r esources more effectively and provide tailored interventions so that value of care is improved for this high risk group Little is known about which factors drive addit ional pain related healthcare utilization following physical therapy. Given the dynamic nature of the provider patient interaction, it is likely that time varying factors may have a significant influence on the perceived need for additional services. For instance, change in pain or psychological distress in early treatment has been shown to predict later disability (Wideman, Hill, et al., 2012; Wideman, Scott, Martel, & Sullivan, 2012) bu t the effects of those changes on subsequent healthcare utilization are unknown. Understanding the relative predictive 82

PAGE 83

strength of baseline versus change scores for timevarying factors could have significant implications for clinical practice. For example, if changes in psychological distress or pain interference in early rehabilitation are stronger predictors of subsequent utilization than at baseline, it would suggest the need to emphasize treatment monitoring throughout the clinical encounter. Theoreti cal models suggest enabling factors (e.g. insurance), predisposing factors (e.g. age, sex) and perceived need (e.g. self reported functional status, pain) may be important drivers of healthcare utilization. (Andersen, 1995) However the relative importance of these factors in driving the need for pain related healthcare services beyond physical therapy is also unclear. A st ronger influence of modifiable factors might imply the need to build interventions that address those factors directly. For less modifiable factors, patients should be matched to stratified care pathways that maximize the probability of patient defined suc cess. An important potential contributor to painrelated healthcare utilization following physical therapy is comorbidities. The influence of comorbidities on symptom progression and disability for patients with musculoskeletal pain is well understood. (Alrwaily et al., 2016; de Rooij et al., 2014; Peter et al., 2015; Ritzwoller et al., 2006; Rundell et al., 2016, 2017; van Dijk et al., 2008; Vogeli et al., 2007) However, we do not yet fully understand how information on como rbidities can be used in isolation or alongside other assessments to identify risk for healthcare utilization beyond physical therapy Comorbidities are prevalent among patients seeking physical therapy (Boissonnault, 1999) therefore understanding their utility for prospectively identifying those at risk for pain related healthcare subsequent to physical therapy is w arranted. 83

PAGE 84

This study will identify predictors of additional healthcare use following an episode of physical therapy for musculoskeletal pain. Enabling factors and comorbidities as well as the influence of baseline versus change scores for timevarying per ceived needs are not commonly examined in outcome prediction studies of musculoskeletal pain. Therefore, this project will lead to the development of potentially novel risk models for health services utilization due to musculoskeletal pain. The results of this project will be particularly important for clinical decisionmaking in capitated or value based payment settings where service efficiency and effectiveness are prioritized. Prospective identification of those in need of additional health services wil l also facilitate future studies aimed at 1) developing more valuable clinical pathways for these patients, and 2) distinguishing wasteful versus appropriate additional healthcare utilization. The primary aim of this study was to determine predictors of additional healthcare utili zation intensity following an episode of physical t herapy for musculoskeletal pain. This aim will improve the ability to predict additional healthcare service needs following physical therapy, which can then be used to more efficiently direct care. This aim will also aid in identifying a highutilization patient subgroup for comparative effectiveness research aimed at avoiding unnecessary/unwarranted escalation of care for those seeking musculoskeletal pain and receiving nonpharmac ological treatment first In a secondary aim, we determined predictors of additional utilization of specific healthcare services (e.g. imaging, surgery, prescription pain medication) following an episode of physical t herapy for musculoskeletal pain. This aim will clarify whether 84

PAGE 85

certain risk factors are servicespecific or more globally related to utilization. This information will also be important for developing strategies to reduce unnecessary use. We hypothesized that higher baseline pain interference, disability, and comorbidities as well as an increase in pain and psychological distress over 4 weeks would be associated with a higher healthcare utilization. Our hypothesis is based on prior studies that suggest these factors are associated with poor functional outcomes after rehabilitation, or related to higher levels of healthcare utilization in the general population. (Andersen, 1995; Bair et al., 2003; Beneciuk, Fritz, & George, 2014; George et al., 201 1; Wideman, Scott, et al., 2012; Wideman & Sullivan, 2011) We also hypothesize that change in psychological distress, pain, and function will be stronger predictors of specific service utilization than baseline measures. Surgery and emergency room visit s will be predicted by a combination of healthrelated and demographic factors, as both represent resourceintensive services that are more sensitive factors such as insurance coverage, access, and geographic location. Methods Dataset This study used data from the Orthopedic Physical Therapy Investigative Networks (OPT IN) Optimal Screening for Prediction of Referral and Outcome (OSPRO) validation cohort study, a longitudinal study of patients with knee, shoulder, back or neck pain seeking Physical Thera py. The aim of the OSPRO study was to develop and validate review of systems (i.e. evidence of systematic involvement) and yellow flag (i.e. pain related distress) screening tools for use in outpatient orthopedic physical therapy. This study is part of the planned validation analysis of the tools for predicting healthcare utilization. Data collection occurred at initial physical therapy 85

PAGE 86

evaluation (hereafter, baseline), and again at 4 weeks, 6 months, and 12 months All data were collected using an online p ortal (Research Electronic Data Capture, REDCap). P articipants were notified of a pending assessment by an email that directed them back to the study website to complete their follow up assessment. Baseline information included patient demographics, medica l history, injury characteristics, symptoms, disability, and expectations. Follow up data collection included symptoms, disability and healthcare utilization. The University of Florida Institutional Review Board approved this study. Patient Population A co nvenience sample of participants was recruited from participating OPT IN clinical sites during their initial outpatient physical therapy evaluation. The OPT IN includes from 9 clinical sites around the United States (Gainesville, FL; Jacksonville, FL; Greenville, SC; Philadelphia, PA; Terra Haute, IN; Chicago, IL; Boulder, CO; Los Angeles, CA; and Portland, OR). Clinics were selected based on different sociodemographic strata, geographic location and representation of urban and rural communities. The OP T IN clinics that participated in data collection represented 5 of 8 geographic regions for the United States including the Mideast, Southeast, Great Lakes, Rocky Mountain States, and Far West. Eligibility c riteria were intentionally broad since our intent was to develop assessment tool s with wide clinical application. Using narrow eligibility criteria would have exclude d a significant number of patients commonly seen by orthopedic physical therapists, result ing in limited application of these tools. 86

PAGE 87

Inclusion criteria Patients between the ages of 18 and 75 years of age were eligible to participate in this study if they were: 1) seeking outpatient physical therapy treatment for musculoskeletal pain, 2) had primary complaints involving the cervical spine, lumbar spine, shoulder or knee, and 3) able to read and comprehend English language. Exclusion criteria Patients were excluded from study participation for any diagnosis indicative of 1) widespread chronic pain syndrome (e.g. fibromyalgia or irritable bowel syndrome), 2) neuropathic pain syndrome (e.g. complex regional pain syndrome or diabetic neuropathy), 3) psychiatric history (currently under the care of a mental health care provider or taking multiple ps ychiatric medications), 4) cancer (currently receiving treatment for active cancer), 5) neurological disorder (e.g. stroke, spinal cord injury, or traumatic brain injury). Healthcare Utilization P redictors Sociodemographic and health related information Participants completed a standard intake form previously used in our clinical studies includ ing: age, sex, race, and insurance provider type. Health related variables included anatomical region of pain (low back, neck, shoulder, knee) and whether the patient had undergone surgery for their primary pain complaints (yes or no) Due to small cell sizes, race was dichotomized as white or nonwhite. Additionally, insurance type was collapsed to 3 categories: private, public, and other (no insurance, workers compensation, or other commercial insurance). 87

PAGE 88

Persistent or chronic pain status At intake, pain status was determined using established definitions that account for the duration of pain and activity limitations (Carey et al., 2010; Freburger et al., 2009) using the fol lowing two questions: 1) How long have you been experiencing your current painful symptoms? and 2) Have you experienced ANY pain and activity limitations every day for the past 3 months? Responses to question 1 of greater than 90 days or responses to question 2 of Yes were used to classify patients as chronic at intake. Pain intensity Pain intensity was assessed by the numerical pain rating scale (NPRS) ranging from 0 (no pain) to 10 (worst pain imaginable). (Bolton, 1999; Childs, Piva, & Fritz, 2005; Jensen, Turner, Romano, & Fisher, 1999) Participants rated their current pain intensity, as well as their best (lowest) and worst (highest) pain intensity over the past 24 hours. Quality of life and r egion specific disability Self report of functional status was assessed at intake and 1year follow up with two measures: 1) the Med ical Outcomes Study 8 item Short Form Health Survey (SF 8), which is a general quality of life measure that has physical and mental health domains (J. E. Ware, Kosinski, Dewey, & Gandek, 2001) and 2) the Neck Disability Index (NDI) (H. Vernon & Mior, 1991; Howard Vernon, 2008) Oswestry Disability Questionnaire (ODQ) (J. M. Fritz & Irrgang, 2001; Hudson Cook, Tomes Nicholson, & Breen, 1989) Quick Disability of Arm Shoulder and Hand (DASH) (Beaton, Wright, Katz, & Upper Extremity Collaborative Group, 2005) o r International Knee Documentation Committee (IKDC) Subjective Knee Form (Irrgang et al., 2001) as condition specific 88

PAGE 89

measures for cervical, low back, shoulder and knee pain, respectively. Region specific disa bility measures were z transformed for purposes of analysis. Charlson comorbidity index Health history was determined with the Charlson Comorbidity Index (M. E. Charlson, Pompei, A les, & MacKenzie, 1987) The Charlson Comorbidity Index lists 19 medical conditions that participants are asked to indicate whether they have ever been diagnosed with by a physician. OSPRO Review of Systems tool (OSPROROS) The OSPRO ROS is a review of symptoms screening tool for use in outpatient orthopedic physical therapy settings. (George et al., 2015) This measure includes standard symptom descriptors to aid with identification of systemic origins of musculoskeletal pain. The full length 23item version of the OSPRO ROS is capable of identifying 100% of red flag responders (i.e. indicating yes to at least one systemic symptom on a questionnaire) on outpatient orthopedic physical therapy. A shorter, 10 item version is also available that has been shown to identify approximately 95% of redflag responders. OSPRO Yellow Flag Tool (OSPROYF) The OSPRO YF is a yellow flag assessment tool that includes items from pain vulnerability domains (negative affect and fear avoidance) and pain resilience domains (positive affect and self efficacy) to aid with identification of pain associated psychological distress in outpatient orthopedic physical therapy settings. (Lentz et al., 2016) T he full length OSPRO YF has 17 items, however a shortened 10item version is also available with an acceptable trade off in accuracy 89

PAGE 90

Intervention All physical therapy treatment was provided at the discretion of the treating clinician. The duration of the episode, the number of physical therapy visits, and individual treatment parameters (type, intensity, duration, frequency) were not collected. Healthcare U ti lization Outcomes Health care utilization was assessed wi th questions derived from previous population based studies involving musculoskeletal pain that have used survey methods for follow up assessment (Carey et al., 2010; Freburger et al., 2009) Patients were asked whether they used any of the following healthcare services in the time following their physical therapy treatment at 6and 12 month follow up: 1. Opioid painkillers (eg. Vicodin, Lortab, Hydrocodone, Fentanyl, Percocet, Oxycontin, Oxycodone, tramadol, Ultram, Diludid, etc) 2. Injections 3. Surgery 4. Diagnostic tests or Imaging (eg. xray, MRI, CT scan, nerve conduction test etc.) 5. Emergency room visits Yes responses were followed by questions regarding the quantity of services utilized (i.e. number of opioid painkillers number of diagnostic tests or number of emergency room visits) All utilization questions were answered on a categorical scale ( 0, 1, 2 5, 5 10, or >10 ) indicating the quantity of a particular service received during the applicable follow up timeframe. At 6 month f ollow up, patients reported their use of services for the previous 2 months, allowing a timeframe of 4 months from from initial evaluation for them to complete physical therapy. At 12month follow up, patients reported their use of services over the previ ous 6 months since their last survey. This method provided an 8 month overall follow up period after physical therapy and was performed to minimize recall bias 90

PAGE 91

Statistical Analysis We developed models to separately predict: 1) the dichotomous outcome of no healthcare utilization versus any healthcare utilization over the course of the entire follow up period, 2) the intensity of healthcare utilization at 6 month follow up and between 6month and 12month follow up, and 3) the utilization of specific servi ces over the course of the entire follow up period. We decided to develop separate models since each of outcome predicted by these models might have unique policy implications. For instance, those who utilize no additional services might represent a low r isk group for which physical therapy alone might be particularly appropriate or there is low risk of escalating care. Conversely, identifying predictors of high intensity versus low intensity utilization would have implications for mobilization of healthcare resources to better meet patient needs. Finally, predicting use of specific services would inform policy where reduction of specific services is a high priority, such as utilization of opioid painkillers or unnecessary use of emergency room serv ices. All prediction models used a hierarchical design with age, sex, race, anatomical region of pain, insurance, chronicity of pain, pain intensity, comorbidities (from Charlson index), disability, 10item OSPRO YF and 10item OSPRO ROS at baseline enter ed in the first step. In the second step, the remaining items from the OSPRO YF and OSPRO ROS were entered to determine whether full length versions of the surveys provided better prediction over shortened versions. In the final step, b aseline to 4 week ch ange in pain intensity, regionspecific disability, and OSPRO YF score were added, since early changes in these variables may be associated with improved prediction of outcomes over baseline variables alone. (Beneciuk et al., 2014) This approach allowed 91

PAGE 92

for as sessment of the relative value of treatment monitoring for the prediction of healthcare utilization outcomes. For the first analysis, binary logistic regression was used to determine predictors of any healthcare utilization following physical therapy, with the dependent variable defined as reporting one or more utilization events for any of the potential healthcare services. For the second analysis, we were interested in both the volume and type of healthcare utilization as an indicator of utilization intensity in both the early phase (utilization reported at 6 months) and late phase (utilization reported between 6 months and 12 months) following physical therapy. Because utilization of individual services was measured on a categorical scale and could not be directly added, utilization intensity categories were developed using both volume and service type criteria. Patients that reported no healthcare utilization of any type during the follow up phase were included in the NONE category. Those who reported utilizing any of the healthcare services (except surgery) no more than once were categorized as LOW. For instance, patients who sought onetime follow up care by their physician and received imaging and/or treatment (e.g. injection and/or prescription medic ation) were included in the LOW category. All other patients, including those who reported any surgery or greater than 1 episode of any healthcare service during the follow up period, were classified as HIGH. This approach served to provide a measure of ut ilization intensity, as well as collapse higher utilization categories that were less prevalent in the sample. Separate utilization intensity measures were calculated for the early and late phases following physical therapy, with separate ordinal regression models developed to predict category membership during each phase. The proportional odds assumption was tested to 92

PAGE 93

determine appropriateness of ordinal logistic regression for each model. We modeled early and late phases separately to determine whether persistent or longterm utilization is driven by different factors than what drives utilization in more acute phas es of rehabilitation. For the final analyses, utilization of specific services was dichotomized for each service, with report of any utilization categorized as YES and all others categorized as NO. Specific service utilization over early and late phases following physical therapy were collapsed to create a single dichotomous utilization indicator over the course of the entire study follow up peri od. Separate multivariate binary logistic regression models were then fitted for the dichotomous indicator of each healthcare service (e.g. opioid use injection, imaging, surgery, and ER visits ). O verall fit for each model was examined with Hosmer & Lemeshow test, chi square and pseudor2 values (e.g. Nagelkerke) when appropriate. Comparison of adjusted odds ratios (OR) and 95 % confidence interval (CI) were used to determine the relative strength of each predictor. Multicollinearity was assessed using variance i nflation f actor (VIF) and t olerance, where VIFs < 10 and tolerances >0.1 suggested no significant collinearity among independent variables. The OPT IN data collection forms required complete data from respondents before they were allowed to proceed to subsequent survey pages. Therefore, the occurrence of missing data for independent predictor variables was minimal (<1% of sample). However, for subjects who were lost to follow up, we planned to take three approaches to account for missing data. F irst, demographic and baseline health variables would be compared between those with complete follow up at 1 year and 93

PAGE 94

those without follow up at 1 year to identify any potential group differences related to completion of follow up. Second, sensitivity anal yses would be conducted wi th missing outcomes imputed by multiple imputation, a n iterative form of stochastic imputation. When necessary, analyses will use 20 imputations and pooled B weights and odds ratios will be reported when available. Regression resu lts using imputation will be compared with those obtained from complete case only analysis to assess the potential influence of missing data on the findings and identify robust predictors Finally, sensitivity analyses would be conducted by repeating each analysis using inverse probability of attrition weighting (IPAW) This approach is used to account for attrition related selection bias in longitudinal studies by more heavily weighting observations associated with a lower probability of study completion. (Weuve et al., 2012) Thus, the resulting analysis is compensated for under representation of subjects who are more likely to be lost to follow up. IPAW produces smaller effect estimate biases than more conventional methods that adjust for baseline predict ors of attrition (Hernn, Hernndez Daz, & Robin s, 2004) Briefly, logistic regression will be performed to identify predictors of attrition using an opportunistic approach that optimizes model fit, with an area under the curve (AUC) target value of > 0.7. Then, inverse of predicted probabilities for r emaining in the study will be used to weight observations, and all analyses will be repeated. Regression results using IPAW will be compared with those obtained from complete case only analysis to assess the potential influence of missing data on the findi ngs and identify robust predictors. Sample Size For logistic regression analyses, event per variable values of 10 or greater are suggested, since overfitting will weaken the probability that original findings will be 94

PAGE 95

reproduced in an independent sample. (Kent, Keating, & Leboeuf Yde, 2010; Peduzzi, Concato, Kemper, Holford, & Feinstein, 1996) With 18 potent ial predictors, a sample of n=180 reporting healthcare utilization at follow up would be s ufficient for the proposed analyses. However, this estimate is conservative. Other methods for determining sample size for prediction analyses suggest 10 subjects per independent variable or N>104 + number of independent predictors. (Green, 1991; R. J. Harris, 2001) For these less conservative estimates, the projected study sample size is sufficient for the proposed analyses. Results Four hundred and forty subjects were recruited at initial evaluation. Follow up at 4 weeks was 75.0%, at 6 months was 69.0% and at 12 months was 65.2%. For primary analyses, only those with complete follow up data at each time point were considered (n=246). Baseline demographics and health related characteristics for the entire sample, as well as those who did and did not complete all follow up are presented in Tabl es 3 1 through 33. Those who did not complete follow up were younger, more likely to be nonwhite, had less than college degree, more likely to have sudden symptom onset, had higher pain intensity, and higher psychological distress measured by the OSPRO Y F. Overall, 43.1% of the sample utilized at least one healthcare service following the physical therapy episode. Distribution of utilization for specific services is provided in Table 3 4. For utilization intensity, at 6 months, n=140 subjects (56.9%) rep orted no additional healthcare utilization while n=60 (24.4%) were classified as LOW and n=46 (18.7%) were classified as HIGH. At 12 months, n=140 subjects (56.9%) reported no additional healthcare utilization, n=56 (22.8%) were classified as LOW and n=50 95

PAGE 96

(20.3%) were classified as HIGH. For multivariate analyses, all VIFs were less than 10 and tolerance values greater than 0.1 suggesting no significant multicollinearity among independent variables. Utilization of Additional Healthcare S ervices o ver 12 M ont hs Prediction model results for use of any additional healthcar e services are listed in Table 3 5 In the first step, baseline demographics and health related information contributed significantly to the model (chi square=49.46, df=15, p<.001, Nagelkerke r2 = 0.27). In this step, the only significant individual predictor of utilization was presence of chronic symptoms (OR=2.42, p=.022). Addition of remaining OSPRO ROS and OSPRO YF did not improve model fit (chi square =1.62, df=2, p=.45, change in r2=.01). Addition of change scores in the final step significantly improved model fit (chi square = 13.31, df=4, p=.01, change in r2 =.06). In the final full multivariate model, baseline pain (OR=1.52, p<.01) and change in pain from baseline to 4week follow up (O R=1.55, p<.01) were significant predictors of ANY additional healthcare utilization (Nagelkerke r2=.34, Hosmer and Lemeshow test chi square=4.24, df=8, p=.834, classification=74.0). The final parsimonious model included chronic symptoms (OR=2.19, p=.04), baseline pain (OR=1.45, p<.01), CCI (OR=1.23, p=.07), baseline disability (OR=1.48, p=.05), and change in pain from baseline to 4week follow up (OR=1.45, p<.01) (r2=.30, Hosmer and Lemeshow test chi square =4.49, df=8, p=.81, classification=69.5). Where change in pain, pain intensity, or disability were significant predictors, higher utilization was associated with worsening of pain, higher pain intensity, or higher disability. Intensity of U tilizatio n Reported at 6 M onth and 12M onth Follow U p O rdinal re gression results for intensity of utilization are listed in Table 3 6 For intensity of utilization reported at 6 months, the final multivariate model included 96

PAGE 97

baseline pain intensity (OR=1.55, p<.01) and baseline to 4 week change in pain intensity (OR=1.5 7, p<.01). The parallel lines test was not significant, suggesting the presence of proportional odds and confirming the appropriateness of ordinal regression. C hi square test (chi square=82.09, df=21, p<.001) and variance explained ( Nagelkerke r2= .36 ) sug gested good model fit For utilization intensity reported at 12 month follow up, baseline pain intensity (OR=1.52, p<.01) and baseline to 4 week change in pain intensity (OR=1.52, p<.01) were significant predictors. Chi square test (chi square=75.81, df=21, p<.001) and variance explained ( Nagelkerke r2= .34) suggested good model fit. The parallel lines test was not significant, indicating the presence of proportional odds and confirming the appropriateness of ordinal regression. Utilization of O pioids Prediction model results for use of opioids are listed in Table 3 7 In the first step, baseline pain intensity (OR=1.36, p=.04) and CCI (OR= 1.39, p=.02) were significant predictors (chi square = 44.94, df=15, p<.01, Nagelkerke r2= 0.29). Addition of rema ining OSPRO ROS and OSPRO YF did not improve model fit in the second step (chi square = 0.24, df=2, p=.89, change in r2<.01). Addition of change scores in the final step significantly improved model fit (chi square = 22.75, df=4, p<.01, change in r2=.13). In the final full multivariate model, baseline pain (OR=1.74, p<.01), CCI (OR = 1.48, p=.01), change in additional 7item OSPRO YF score (OR=.86, p=.03) and change in pain from baseline to 4 week follow up (OR=1.52, p=.02) were significant predictors of opioid use (Nagelkerke = .42, Hosmer and Lemeshow chi square=1.78, df=8, p=.99, classification 86.5). In the final parsimonious model, chronicity (OR=2.80, p=.08), baseline pain (OR=1.70, p<.01), CCI (OR = 1.54, p<.01), change in additional 7 item 97

PAGE 98

OSPRO YF s core (OR=.92, p=.08) and change in pain from baseline to 4week follow up (OR=1.71, p<.01) were significant predictors of opioid use (Nagelkerke r2= .33, Hosmer and Lemeshow chi square= 9.00, df=8, p=.34, classification=84.8). In these models, higher pain, CCI and increase in pain were associated with higher utilization while worsening of psychological distress as measured by the additional 7item OSPRO YF was associated with low er odds of opioid utilization. Utilization of I njection. Prediction model res ults for use of injection are listed in Table 3 8 In the first step, baseline demographics and health related information contributed significantly to the model (chi square=41.56, df=15, p<.001, Nagelkerke r2 = 0.29). Baseline disability (OR=2.31, p=.01) was the only significant predictor in this step. Like previous models, addition of remaining OSPRO ROS and OSPRO YF did not improve model fit (chi square = 3.47, df=2, p=.18, change in r2=.03). Addition of change scores in the final step also did not signi ficantly improve model fit (chi square = 6.07, df=4, p=.19, change in r2=.03). The final full multivariate model did not include any individual predictors significant at alpha=.05 (Nagelkerke r2=.35, Hosmer and Lemeshow chi square=5.04, df=8, p=.75, classi fication=85.2) however a few approach significance including baseline pain (OR=1.44, p=.07) and change in pain from baseline to 4week follow up (OR=1.47, p=.06). In the final parsimonious model, race (OR=0.17, p=.09), chronicity (OR= 3.57, p=.05), and baseline disability (OR=2.14, p<.01) were significant predictors of injection use (Nagelkerke r2=.22, Hosmer and Lemeshow chi square =16.14, df=8, p=.04, classification=85.2). In this model, more chronic symptoms and higher baseline disability were associate d with higher utilization. 98

PAGE 99

Utilization of S urgery Prediction model results for use of surgery are listed in Table 39 In the first step, baseline demographics and health related information were not significant predictors in the model (chi square=24.33, df=15, p=.06, Nagelkerke r2 = 0.23), however the contribution of baseline disability approached significance (OR = 1.97, p=.08). Addition of the remaining OSPRO ROS and OSPRO YF items also did not improve model fit in the second step (chi square = 0.28, df =2, p=.25, change in r2=.03). However, addition of change scores in the final step significantly improved model fit (chi square = 19.37, df=4, p<.01, change in r2=.17). In the final full multivariate model, only change in 10item OSPRO YF score (OR=1.16, p=.05) was a significant predictor, such that worsening of psychological distress was associated with higher odds of subsequent surgery ( Nagelkerke r2=.43, Hosmer and Lemeshow chi square= 11.72, df=8, p=.164, classification=93.7) In the final parsimonious model, baseline disability (OR = 3.13, p<.01), change in 10item OSPRO YF score (OR=1.12, p=.03), and baseline to 4 week change in disability (OR=3.04, p<.01) were significant predictors of subsequent surgery (Nagelkerke r2=. 30, Hosmer and Lemeshow chi squ are =16.63, df=8, p=.03, classification=92.4). In this model, higher baseline disability, as well as increases in disability and psychological distress over the first 4 weeks were associated with higher odds of surgery. Utilization of Diagnostic T ests or I maging. Prediction model results for use of diagnostic tests or imaging are listed in Table 3 10. In the first step, baseline demographics and health related information contributed significantly to the model (chi square=43.61, df=15, p<.01, Nagelkerke r2= 0.25). In this step, CCI (OR = 1.34, p=.02) and baseline disability (OR = 2.06, p<.01) were significant 99

PAGE 100

contributors to the model, with higher CCI and disability associated with higher odds of utilization. Addition of remaining OSPRO ROS and OSPRO YF di d not improve model fit (chi square = 3.41, df=2, p=.18, change in r2=.02). Addition of change scores in the final step significantly improved model fit (chi square = 13.64, df=4, p<.01, change in r2=.07). In the final full multivariate model, CCI (OR=1.42, p=.01), baseline disability (OR = 2.20, p=.03) and baseline to 4week change in pain intensity (OR = 1.53, p<.01) were significant predictors (Nagelkerke r2=.34, Hosmer and Lemeshow chi square=6.31, df=8, p=.613, classification=75.3). The final parsimoni ous model included chronicity of symptoms (OR=1.82, p=.16), CCI (OR=1.35, p=.01), baseline disability (OR = 2.25, p<.01), and baseline to 4 week change in pain intensity (OR=1.41, p<.01) (Nagelkerke r2=.29, Hosmer and Lemeshow chi square =26.38, df=8, p=.01, classification=76.7 ) In this model, chronic symptoms, higher CCI, higher disability and worsening pain were associated with higher odds of utilization of diagnostic tests or imaging. U tilization of Emergency R oom Prediction model results for emergenc y room utilization are listed in Table 3 11. In the first step, baseline demographics and health related information contributed significantly to the model (chi square=38.12, df=15, p<.01, Nagelkerke r2= 0.40). In this step, age (OR = 0.92, p<.01), anatomi cal location of pain (OR=0.5, p=.04), insurance (OR=6.34, p=.02) and baseline disability (OR = 2.95, p=.02) were significant contributors to the model. In this step, higher utilization was associated with younger age, lack of insurance or Workers Compensa tion compared to private insurance, and higher baseline disability. Compared to those with low back pain, subjects with knee pain were less likely to utilize the emergency room. Addition of the remaining OSPRO ROS and OSPRO YF items did not improve model f it (chi square = 5.56, df=2, p=.06, 100

PAGE 101

change in r2=.05) in the second step. In the final full multivariate model, age (OR = 0.92, p<.01), anatomical location of pain (OR=0.2, p=.03), insurance (OR=13.60, p=.02), surgery for primary complaint (OR=18.06, p=.02) and baseline to 4week change in pain intensity (OR = 2.25, p=.02) were significant contributors to the model. As in the prior step, higher utilization was associated with younger age and lack of insurance or Workers Compensation compared to private ins urance. Higher utilization was also associated with history of surgery for the primary complaint and increase in pain intensity from baseline to 4 weeks (Nagelkerke r2= .56, Hosmer and Lemeshow chi square=17.94, df=8, p=.02, classification= .96). The final parsimonious model included age (OR = 0.91, p<.01), anatomical location of pain (OR=0.06, p=.06), insurance (OR=8.99, p=.01), chronicity (OR=5.81, p=.16), surgery for primary complaint (OR=5.43, p=.07), baseline disability (OR=4.88, p<.01), baseline addit ional OSPRO YF items (OR=.82, p=0.03) and baseline to 4week change in pain intensity (OR = 1.77, p<.01) as significant contributors to the model (Nagelkerke r2=.48, Hosmer and Lemeshow chi square =15.13, df=8, p=.06, classification=95.5 ) All relationships were as previously described, however higher psychological distress as measured by the additional 7 items from the OSPRO YF was associated with lower odds of emergency room utilization. Results of Imputed M odels and IPAW A nalyses Due to the large number of missing utilization outcomes, analyses using multiple imputation and IPAW were undertaken for comparison with complete case analyses. This approach would assess the robustness of model predictors against potential bias introduced by loss to follow up. For multiple imputation, t wenty imputed datasets were created and reanalyzed using the same methods as previously described. Pooled 101

PAGE 102

coefficients of the full and parsimonious models were assessed for binary logistic analyses and full model results were ass essed for ordinal analyses. Since backward selection did not consistently return the same parsimonious model for each imputed set only predictors that were identified in at least 50% of the imputed models were considered when deriving parsimonious model pooled estimates. (Wood, White, & Royston, 2008) Imputation analysis results are listed with their respective outcome in Tables 3 5 through 312 P ooled model estimates for 6and 12month utilization intensity failed to identify any significant predictors in multivariate analysis Re sults of IPAW analyses are also listed with their respective outcome in Tables 35 through 312. These results are largely consistent with results of the complete case analysis. Discussion I n general, i ntensity of pain and its course over the early phases of rehabilitation are important predictors of subsequent healthcare utilization incidence and intensity following an episode of physical therapy for musculoskeletal pain. There are specific cases where course of pain associated psychological distress added to the prediction of healthcare utilization (e.g. surgery). These findings highlight the importance of monitoring pain outcomes throughout early phases of rehabilitation and incorporating time varying factors into risk prediction models to better identif y and manage those at risk for subsequent healthcare. For patients who show signs of worsening pain in this early timeframe treatment pathways may need to be altered to reduce the risk of subsequent healthcare utilization. Aside from pain intensity predi ctors of utilization are highly outcomespecific We focus on interpreting predictors that were identified in the majority of analytic models (e.g. full multivariate, parsimonious, imputed) for each type of healthcare service as they are more robust and most likely to be reproduced in future 102

PAGE 103

studies. Healthrelated variables were generally stronger predictors than sociodemographic factors across models, although some sociodemographic factors were related to use of a few specif ic services. For instance, younger age was an important predictor of emergency room utilization, as was insurance, which supports existing literature that shows a strong relationship between emergency room use and insurance status (Gandhi, Grant, & Sabik 2014; Tang, Stein, Hsia, Maselli, & Gonzales, 2010) The identification of sociodemographic and health related predictors supports the application of conceptual, multidimensional healthcare utilization models to musculoskeletal pain populations. An im portant health related predictor was comorbidity level which was associated with higher odds of subsequent opioid use and di agnostic tests and imaging. Opioids and diagnostic tests and imaging were the two most common subsequent healthcare services utiliz ed following physical therapy. Approximately 42% reported opioid use and 70% reported use of diagnostic tests and imaging among those who had any su bsequent healthcare utilization. Use of these services is particularly concerning as they may heighten the r isk for downstream escalation of care. Given the growing public health concern over opioid use and the desire to avoid wasteful, invasive interventions these results suggest the importance of considering comorbidities for pathway selection and outcomes pr ediction i n outpatient physical therapy. An additional consideration regarding this finding is that longterm opioid management might be appropriate for a subset of patients who fail to respond to other treatment approaches alone. While our study design is unable to determine appropriateness of subsequent 103

PAGE 104

opioid utilization in this subset of patients, recent opioid prescription guidelines have been published to address this question. (Dowell et al., 2016) Physical therapy is purported to be an effective, nonpharmacological approach for managing musculoskeletal pain, however approximately 43% of subjects in our study reported subsequent healthcare utilization for their painrelated condition in the year following initial physical therapy evaluation. Some additional healthcare utilization is expected, especially among individuals that are on longterm pain management pathways due to chronic or persistent symptoms Readers are also cautioned that treatment was not c ontrolled or recorded in this study and the degree of compliance with established clinical practice guidelines is unknown. We acknowledge there is no established benchmark to use in determining whether this percentage is high, low, or comparable to outcomes using other health services. Yet with almost half of subjects utilizing additional healthcare for their symptoms across a national sample of practice types and locations it is apparent that opportunities exist to improve the effectiveness of current phy sical therapy approaches for musculoskeletal pain. This finding is particularly notable given recent efforts to define physical therapy as an effective first line, non pharmacological treatment option against more invasive or higher risk services, such as surgery or opioid use, respectively. Attempts to minimize the incidence of subsequent healthcare utilization through physical therapy pathway redesign is a viable target for healthcare systems. One potential approach for physical therapy pathway re design might be to spread out fo rmal care over longer periods with booster sessions rather than using front loaded approaches more com mon to management of acute pain (Abb ott et al., 2015) Alternatively, effectiveness might be improved through greater 104

PAGE 105

provider patient partnerships, where principles of self management are instilled while providing for consultation w ith the PT when symptoms worsen. Psychological distress was not a consistent predictor of utilization in all models except surgery This finding contrasts with our hypotheses and results from many prior studies that have use d development of chronic pain or disability as an outcome. A potential reason for this is that unique effects of psychological distress on healthcare utilization become less important when considered along with other sociodemographic and health related factors that more significantly enable or predispose someone to healthcare use. More direc tly stated, the influence of enabling factors such as access to care, insurance, and financial status could attenuate the effects of psychological distress on careseeking behavior. Another potential reason is that change in pain was also included as a predictor in the models. Change in pain is often associated with psychological distress and studied as a dependent variable in most pain outcomes studies. As a result, we cannot directly compare our results to those of more traditional pain outcomes studies. An alternative explanation for the finding is that the OSPRO YF as a composite measure of psychological distress may not be sensitive enough to predict utilization if utilization is a function of specific psychological characteristics, such as catastrophiz ing or depression. Future studies might consider using more specific psychological constructs to assess their differential effects on healthcare utilization. Despite unexpected findings regarding psychological distress, the incidence of subsequent surgery was predicted by an early increase in general psychological distress measured by the 10item OSPRO YF tool. This might suggest that patients and/or healthcare providers inappropriately infer that an increase in psychological 105

PAGE 106

distress requires a surgical s olution. At the same time, it highlights an important opportunity to potentially reduce the likelihood of future surgery by implementing strategies that address psychological distress in rehabilitation. Importantly, risk was identified using a shortened version of the OSPROYF tool, highlighting its utility over longer versions of the tool. On the other hand, neither versions of the OSPRO ROS consistently predicted utilization outcomes in multivariate analyses As a result, when considering use of the OSPRO tools to screen for risk of subsequent healthcare utilization after accounting for other key variables we can only currently recommend use of the 10item OSPRO YF to predict risk of subsequent surgery. The primary strength of the study is longitudinal f ollow up at multiple time points for a patient population seeking physical therapy for a variety of musculoskeletal pain conditions. For all models except emergency room use, anatomical location of pain was not a significant predictor, suggesting results are widely applicable across anatomical regions of pain. Another strength of this cohort study is the assessment of multiple healthcare utilization outcomes. This approach allowed us to examine predictors for a variety of different utilization outcomes that are of interest when establishing health policy. In addition, it allows for assessment of utilization outcomes compared to painor disability related outcomes that are more commonly assessed in cohort studies. The use of multiple screening tools (i.e. ye llow flags and review of systems) along with other healthrelated assessments (e.g. comorbidities) is also a strength of this study. Y ellow flags and comorbidities are not often assessed in physical therapy but have the potential to impact rehabilitation outcomes, including the need for additional healthcare. A final strength is inclusion of multiple sociodemographic, healthrelated and 106

PAGE 107

psychosocial factors as potential predictors. Healthcare outcomes and utilization exhibit emergent properties that requir e the consideration of multiple, competing factors to fully explain. (Lentz et al., 2017) However, explained variance estimates in our models suggest that further research is necessary to identify additional factors contributing to healthcare utilization following physical therapy. The primary limitation of the study is the high number of subjects lost to follow up. Traditional survey studies often report follow up rates around onethird, while the goal for follow up in most c ohort studies is at least 80%. In our study, follow up was 75% at 4 weeks, 69% at 6 months and 65% at 12 months. We attempted to account for loss to follow up in our models with multiple imputation and IPAW which are robust strategies for conducting analyses with missing data. (Hernn et al., 2004; Sterne et al., 2009) We observed only modest convergence of results among complete case and imputed analyses likely a function of imputing nearly 1/3 of the utilization outcomes. Results were more concordant among complete case and IPAW analyses. However, we focus our discussion on interpretation of predictors that were identified most consistently through multiple analytic approaches. A second limitation is the lack of detailed information on individual treatments received by patients in the cohort. The deci sion to not collect individual level treatment information was driven by the primary goals of the cohort study to develop tools that broadly predicted outcomes. There were also logistic hurdles in attempting to consistently track individual level treatment information from clinics participating in different health systems. The result is that we are prohibited from making conclusions regarding which characteristics of the clinical encounter might influence subsequent utilization or determine levels of pain or disability immediately 107

PAGE 108

following physical therapy. Moreover, we were unable to account for number of visits or whether patients completed their course of physical therapy. A final limitation to consider is the use of patient recall to measure utilization. To mitigate recall bias, we used two follow up points, at 6 and 12 months. We took this approach to ensure that recall greater than 6 months was not required. However, under or over reporting of utilization is often a concern with studies requiring subj ect recall. Medical record and claims data were not available for these subjects, but could be utilized in future studies to confirm results of our analyses Physical therapy is a common treatment for musculoskeletal pain and there is value in prospectively identifying patients who may seek out subsequent healthcare services for their condition, especially if nonpharmacological care (i.e. physical therapy) is going to be used more for frontline management of acute and chronic pain conditions The benefit of early identification is that physical therapy treatment pathways could be modified or supplemented to better address patients individual needs with the goal of reducing the need for future healthcare. Pain intensity at baseline and its change over the course of early rehabilitation are important predictors of subs equent healthcare utilization. Comorbidities, psychological distress and certain sociodemographic factors predict specific healthcare utilization and may be important additions to healthcare prediction models following physical therapy Future research is needed to better understand how to incorporate these risk models into clinical decision making so that value of physical therapy as a front line treatment for musculoskeletal pain i s optimized, and ways of appropriately deescalating risk for management of musculoskeletal pain are refined. 108

PAGE 109

Table 3 1. Demographic information for the full, complete case, and incomplete follow up cohorts. Variable Label Full Cohort at baseline Completed follow up (n=246) Did not complete follow up (n=194) p value Demographic i nformation Age MeanSD 45.06 15.82 46.59 16.00 43.15 15.43 0.02 Median (min, max) 45 (18 75) 47 (18 75) 42 (18 74) Sex (1 missing) Male 164 (37.3%) 85 (34.6%) 79 (40.7%) 0.20 Female 275 (62.5%) 160 (65.0%) 115 (59.3%) Race (7 missing) White 343 (78.0%) 200 (81.3%) 143 (73.7%) 0.05 Non white 90 (20.5%) 42 (17.1%) 48 (24.7%) Ethnic ity (33 missing) Hispanic or Latino 31 (7.0%) 20 (8.1%) 11 (5.7%) 0.36 Not Hispanic or Latino 376 (85.5%) 211 (85.8%) 165 (85.1%) Education (6 missing) Less than college 161 (36.6%) 71 (28.9%) 90 (46.4%) <.0 0 1 College graduate 273 (62.0%) 172 (69.9%) 101 (52.1%) Income (66 missing) $35,000 or less 112 (25.5%) 62 (25.2%) 50 (25.8%) 0.30 $35,000 to $70,000 106 (24.1%) 59 (24.0%) 47 (24.2%) Greater than 70,000 156 (35.5%) 99 (40.2%) 57 (29.4%) Insurance (26 missing) Private 273 (62.0%) 156 (63.4%) 117 (60.3%) 0.70 Public 75 (17.0%) 46 (18.7%) 29 (14.9%) Other 66 (15.0%) 36 (14.6%) 30 (15.5%) Geographic region Southeast 275 (62.5%) 146 (59.3%) 129 (66.5%) 0.10 Midwest 47 (10.7%) 23 (9.3%) 24 (12.4%) West 98 (22.3%) 65 (26.4%) 33 (17.0%) Northeast 20 (4.5%) 12 (4.9%) 8 (4.1%) 109

PAGE 110

Table 3 2. Health related information for the full, complete case, and incomplete follow up cohorts. Variable Label Full Cohort at baseline (n=440) Completed follow up (n=246) Did not complete follow up (n=194) p value Health related i nformation Anatomical region Neck 98 (22.3%) 48 (19.5%) 50 (25.8%) 0.27 Low Back 118 (26.8%) 66 (26.8%) 52 (26.8%) Shoulder 107 (24.3%) 59 (24.0%) 48 (24.7%) Knee 117 (26.6%) 73 (29.7%) 44 (22.7%) Symptom onset Gradual 239 (54.3%) 146 (59.3%) 93 (47.9%) 0.03 Sudden 138 (31.4%) 65 (26.4%) 73 (37.6%) Traumatic 63 (14.3%) 35 (14.2%) 28 (14.4%) Duration of symptoms MeanSD 398.58 1715.80 379.79 1999.77 423.01 1259.33 0.80 Median 90 (0 29565) 90 (1 29565) 90 (0 10000) Surgery for condition Yes 83 (18.9%) 44 (17.9%) 39 (20.1%) 0.56 No 357 (81.1%) 202 (82.1%) 155 (79.9%) Work related injury (32 missing ) Yes 63 (14.3%) 30 (12.2%) 33 (17.0%) 0.36 No 345 (78.4%) 198 (80.5%) 147 (75.8%) Chronicity Acute 101 (23.0%) 65 (26.4%) 36 (18.6%) 0.05 Chronic 324 (73.6%) 173 (70.3%) 151 (77.8%) Pain Intensity MeanSD 4.22 1.98 3.94 1.72 4.58 2.21 0.01 Median (min, max) 4 (09.67) 4 (08) 4.5 (0 9.7) Disability MeanSD 0 1.0 .06 .97 .08 1.03 0.16 Median (min, max) .16 ( 2.41 3.14) .23 ( 2.21 2.94) .01 ( 2.41 3.14) CCI MeanSD .66 1.47 .68 1.62 .63 1.25 0.76 Median (min, max) 0 (013) 0 (013) 0 (08) 110

PAGE 111

Table 3 3. OSPRO questionnaire scores for the full, complete case, and incomplete follow up cohorts. Variable Label Full Cohort at baseline (n=440) Completed follow up (n=246) Did not complete follow up (n=194) p value OSPRO ROS 10 MeanSD 2.68 2.38 2.52 2.24 2.89 2.55 0.11 Median (min, max) 2 (0 10) 2 (0 10) 2.5 (0 10) OSPRO ROS 13 MeanSD 1.25 1.80 1.14 1.52 1.38 2.09 0.17 Median (min, max) 1 (0 12) 1 (0 7) 1 (0 12) OSPRO YF 10 MeanSD 17.43 6.69 16.87 6.46 18.15 6.91 0.04 Median (min, max) 17 (4 47) 16 (4 40) 17 (4 47) OSPRO YF 7 MeanSD 14.92 5.51 14.41 4.93 15.57 6.12 0.03 Median (min, max) 15 (3 34) 14 (3 28) 16 (3 34) 111

PAGE 112

Table 3 4. Frequency of healthcare utilization reported at 6month and 12 month follow up Label Utilization reported at 6 month follow up Utilization reported at 12 month follow up Dichotomous indicator for any healthcare utilization over entire follow up Utilization volume 0 1 2 5 5 10 >10 0 1 2 5 5 10 >10 No Yes Opioid painkillers 209 18 16 1 2 204 19 16 7 0 201 45 Injection 212 17 14 2 1 217 17 12 0 0 206 40 Surgery 240 4 2 0 0 231 13 2 0 0 227 19 Diagnostic Tests or Imaging 183 40 22 1 0 188 26 28 4 0 172 74 Emergency room 237 7 2 0 0 232 11 2 1 0 228 18 Any utilization 140 106 Intensity of utilization None Low High None Low High 140.0 60.0 46.0 140.0 56.0 50.0 112

PAGE 113

Table 3 5. Summary of variables that contributed to prediction of any additional healthcare utilization Model Analysis Variable B S.E. Sig. Odds Ratio Full multivariate Complete c ase Baseline p ain 0.42 0.14 <0. 0 01 1.52 Change in pain (b aseline to 4 week) 0.44 0.15 <0. 0 01 1.55 Imputed Change in pain (b aseline to 4 week) 0.31 0.15 0.04 1.36 Inverse probability weight CCI 0.31 0.15 0.04 1.36 Baseline p ain 0.39 0.16 0.01 1.48 Change in pain (b aseline to 4 week) 0.47 0.17 <0. 0 01 1.59 Parsimonious Complete Case Chronicity 0.78 0.38 0.04 2.19 Baseline p ain 0.37 0.12 <0. 0 01 1.45 Change in pain (b aseline to 4 week) 0.37 0.11 <0. 0 01 1.45 Imputed Baseline d isability 0.36 0.14 0.01 1.43 Change in pain (b aseline to 4 week) 0.21 0.1 0 0.03 1.23 Inverse probability weight CCI 0.31 0.14 0.03 1.36 Baseline d isability 0.9 0 0.19 <0. 0 01 2.47 Change in pain (b aseline to 4 week) 0.25 0.1 0 0.02 1.28 113

PAGE 114

Table 3 6 Summary of variables that contributed to prediction of utilization intensity at 6 and 12month follow up* Time Model Variable Estimate S.E. Sig. Odds Ratio 6 mon th follow up Full multivariate complete c ase Baseline p ain 0.42 0.13 <0. 0 01 1.52 Change in pain (b aseline to 4 week) 0.42 0.13 <0. 0 01 1.53 Inverse probability weight Baseline p ain 0 .42 0 .14 <0. 0 01 1.52 CCI 0 .21 0 .10 0 .0 4 1.23 Change in p ain (b aseline to 4 week) 0 .5 1 0 .1 5 <0. 0 01 1.66 12 month follow up Full multivariate complete c ase Baseline p ain 0.44 0.13 <0. 0 01 1.55 Change in p ain (b aseline to 4 week) 0.45 0.14 <0. 0 01 1.57 Inverse probability weight Baseline p ain 0 .41 0 .1 4 < 0 0 01 1.51 CCI 0 .2 0 0 10 0 .04 1.23 Change in p ain (b aseline to 4 week) 0 .4 2 0 .14 < 0 0 01 1.52 *No significant predictors were identified in the pooled imputed variable analysis 114

PAGE 115

Table 3 7 Summary of variables that contributed to prediction of opioid utilization Model Analysis Variable B S.E. Sig. Odds Ratio Full multivariate Complete case CCI 0.39 0.15 0.01 1.48 Baseline p ain 0.55 0.2 0 0.01 1.74 Change in pain (b aseline to 4 week) 0.42 0.19 0.02 1.52 Ch ange in additional 7 items from OSPRO YF (b aseline to 4 week) 0.16 0.07 0.03 0.85 Imputed CCI 0.29 0.13 0.03 1.33 Inverse probability weight CCI 0.42 0.16 <0. 0 01 1.52 Baseline p ain 0.52 0.2 0 0.01 1.68 Change in Pain (b aseline to 4 week) 0.4 0.2 0 0.04 1.49 Ch ange in additional 7 items from OSPRO YF (b aseline to 4 week) 0.18 0.08 0.02 0.83 Parsimonious Complete case CCI 0.43 0.13 <0. 0 01 1.54 Baseline p ain 0.53 0.13 <0. 0 01 1.7 0 Change in pain (b aseline to 4 week) 0.54 0.14 <0. 0 01 1.71 Imputed CCI 0.28 0.12 0.02 1.32 Inverse probabi lity weight CCI 0.47 0.14 <0. 0 01 1.6 0 Baseline p ain 0.57 0.15 <0. 0 01 1.76 Change in pain (b aseline to 4 week) 0.53 0.15 <0. 0 01 1.7 0 115

PAGE 116

Table 3 8 Summary of variables that contributed to prediction of injection utili zation* Model Analysis Variable B S.E. Sig. Odds Ratio Full multivariate Inverse probability weight Race (n on white compared to white) 2.56 1.32 0.05 0.08 Baseline p ain 0.42 0.21 0.05 1.52 Parsimonious Complete case Chronicity 1.27 0.65 0.05 3.57 Baseline d isability 0.76 0.2 0 <0.01 2.14 Imputed Baseline d isability 0.49 0.19 0.01 1.63 Inverse probability weight Baseline d isability 0.76 0.21 <0.01 2.14 *No significant predictors were identified in the full multivariate complete case or imputed variable analyses 116

PAGE 117

Table 3 9 Summary of variables that contributed to prediction of surgery utilization* Model Analysis Variable B S.E. Sig. Odds Ratio Full multivariate Complete case Change in 10 item OSPRO YF score (b aseline to 4week) 0.15 0.08 0.05 1.17 Inverse probability weight Change in 10 item OSPRO YF score (b aseline to 4week) 0.17 0.09 0.05 1.18 Parsimonious Complete case Baseline d isability 1.14 0.28 < .001 3.13 Change in 10 item OSPRO YF score (b aseline to 4week) 0.12 0.05 0.03 1.12 Chan ge in disability (b aseline to 4 week) 1.11 0.42 0.01 3.04 Imputed Baseline p ain 0.24 0.11 0.03 1.27 Inverse probability weight Baseline d isability 1.18 0.31 <0.0 0 1 3.25 Change in 10 item OSPRO YF score (b aseline to 4week) 0.13 0.06 0.02 1.14 Chan ge in disability (b aseline to 4 week) 1.12 0.44 0.01 3.05 *No significant predictors were identified in the full multivariate imputed variable analysis 117

PAGE 118

Table 3 10. Summary of variables that contributed to prediction of diagnostic tests and imaging utilization* Model Analysis Variable B S.E. Sig. Odds Ratio Full multivariate Complete case CCI 0.35 0.13 0.01 1.42 Baseline d isability 0.79 0.35 0.03 2.2 0 Change in pain (b aseline to 4 week) 0.43 0.16 0.01 1.53 Inverse probability weight CCI 0.39 0.15 <0.0 0 1 1.47 Baseline d isability 0.95 0.38 0.01 2.58 Change in pain (b aseline to 4 week) 0.48 0.17 <0.0 0 1 1.61 Parsimonious Complete case CCI 0.3 0.12 0.02 1.35 Baseline d isability 0.81 0.18 <0.0 0 1 2.25 Change in pain (b aseline to 4 week) 0.34 0.11 <0.0 0 1 1.41 Imputed Baseline d isability 0.43 0.16 0.01 1.53 Inverse probability weight CCI 0.37 0.13 <0.0 0 1 1.45 Baseline d isability 0.98 0.19 <0.0 0 1 2.66 Change in pain (b aseline to 4 week) 0.37 0.11 <0.0 0 1 1.45 *No significant predictors were identified in the full multivariate imputed variable analysis 118

PAGE 119

Table 3 11. Summary of variables that contributed to prediction of emergency room utilization Model Analysis Variable B S.E. Sig. Odds Ratio Full multivariate Complete case Age 0.1 0 0.04 0.01 0.9 0 Location of symptoms** 4.02 1.82 0.03 0.02 Insurance 2.61 1.08 0.02 13.6 Surgery for primary complaint 2.89 1.24 0.02 18.06 Change in pain (b aseline to 4 week) 0.81 0.35 0.02 2.25 Inverse probability weight Age 0.12 0.04 <0.0 0 1 0.89 Location of symptoms ** 4.61 2.12 0.03 0.01 Insurance 3.93 1.39 <0.0 0 1 50.78 Surgery for primary complaint 3.51 1.5 0 0.02 33.49 Change in pain (b aseline to 4 week) 1.03 0.43 0.02 2.81 Parsimonious Complete case Age 0.09 0.03 <0.0 0 1 0.91 Insurance 2.2 0 0.88 0.01 8.99 Baseline d isability 1.59 0.47 <0.0 0 1 4.88 Addi tional 7 items from OSPRO YF at b aseline 0.2 0 0.09 0.03 0.82 Change in pain (b aseline to 4 week) 0.57 0.22 0.01 1.77 Imputed Baseline d isability 0.41 0.2 0 0.05 1.51 Inv erse probability weight Age 0.06 0.03 0.03 0.94 Location of symptoms ** 3.52 1.53 0.02 0.03 Insurance 2.58 0.86 <0.0 0 1 13.15 Baseline d isability 1.2 0 0.34 <0.0 0 1 3.33 Cha nge in 10 item OSPRO YF score (b aseline to 4 week) 0.14 0.07 0.05 1.15 Change in pain (b aseline to 4 week) 0.46 0.21 0.03 1.59 *No significant predictors were identified in the full multivariate imputed variable analysis **Knee compared to low back Other compared to private insurance Surgery compared to no surgery 119

PAGE 120

CHAPTER 4 IDENTIFICATION OF COMORBIDITY SUBGROUPS AMONG OLDER ADULTS SEEKING HEALTHCARE FOR MUSCULOSKELETAL Musculoskeletal pain is a common condition that contributes to significant personal disability and societal economic burden. (Institute of Medicine (US) Commit tee on Advancing Pain Research, Care, and Education, 2011) Disability and economic burden due to musculoskeletal pain may be exacerbated by the presence of chronic comorbid conditions, which can independently influence the trajectories of perceived heal th status, functional impairment and disability. (Ritzwoller et al., 2006; van Dijk et al., 2008; Vogeli et al., 2007) For instance, Rundell et al. have shown that comorbid low back pain and knee osteoarthritis is associated with worse longterm disability and healthrelated quality of life than is associated with either condition alone. (Rundell et al., 2017) Moreover, there is evidence to suggest t hat comorbidities are more prevalent among those with musculoskeletal pain conditions compared to painfree counterparts. (Gore, Tai, Sadosky, Leslie, & St acey, 2011) Among individuals with musculoskeletal pain, older adults may be particularly susceptible to the moderating effects of comorbidities Currently, it is estimated that 62% of the population over 65 years of age have multiple chronic conditions with that number projected to be 81 million people by 2020. (Vogeli et al., 2007) While the effects of comorbidity on hea lth and function are clear the existence of distinct comorbidity subgroups among older adults with musculoskeletal pain, and the differential effects of comorbidity subgroupings on disability and cost trajectories are unknown. Comorbidity distinctions are often made based on the nature of the health condition, the relative importance of the cooccurring conditions, and the chronology of condition presentation. (Valderas, Starfield, Sibbald, Salisbury, & Roland, 2009) 120

PAGE 121

C omorbidity subgroups have been previously identified among the general population of older adults (Whitson et al., 2016) In a study of Medicare beneficiaries, Whitson et al. identified 6 distinct groups : Minimal Dise ase Class, Non Vascular Class (excess prevalence in cancer, osteoporosis, arthritis, arrhythmia, COPD, and psychiatric disorders), Vascular Class (excess prevalence in HTN, DM, and stroke), Cardio Stroke Cancer Class (excess prevalence in CHF, CHD, arrhythmia, stroke and to a lesser extent HTN, DM, cancer), Major Neurologic Disease Class (excess prevalence in AD, Parkinsons disease, psychiatric disorders), and Very Sick Class (above average prevalence of all 13 conditions) (Whitson et al., 2016) However, it is unknown if these co morbidity subgroups are similar among a more defined population of older adults with musculoskeletal pain. Moreover, the consistency of comorbidity subgroups among different musculoskeletal pain condit ions (e.g. low back pain and osteoarthritis, or OA ) is unclear Studying the differential effects of comorbidity subgroupings on musculoskeletal pain is important for multiple reasons. First, value assessments of musculoskeletal pain treatments are made complicated since healthcare quality utilization a nd costs may differ substantially based on the type and number of co existing health conditions. (Ritzwoller et al., 2006) As more payment models move toward value based reimbursement, understanding the impact of different combinations of co existing health conditions on treatment outcomes and cost trajectories is critica l for evaluating cost effectiveness A ccordingly, Meghani et al. have stated that better defining comorbidity subgroups is an important precursor for effectiveness research (Meghani et al., 2013) 121

PAGE 122

Second, comorbidities are an important, yet understudied contributor to patient heterogeneity that may substantially influence the effectiveness of tailored treatment programs (Saragiotto, Maher, Hancock, & Koes, 2017) Existing research often controls for the effects of comorbidities on treatment outcomes, but these methodological approaches rarely provide direction on how treatment should be modified to account for those effects. In response, the US Department of Health and Human Services has called for identification of comorbidity subgroups as an initial step to ward improving population health status (Parekh, Goodman, Gordon, Koh, & HHS Interagency Workgroup on Multiple Chronic Conditions, 2011) Within a healthcare system that often lacks multi disciplinary collaboration and resources for complex patient centered care, patients with musculoskeletal pain and comorbidities are at risk for low value care. (Druaz Luyet et al., 2015) Identification of comorbidity subgroups will have implications for coordination of care and development of comorbidity specific treatme nt pathways that might improve the effectiveness of current precisionbased pain management approaches The primary aim of this study was to derive empirically based comorbidity classification subgroups separately in older adults with 1) OA and 2) low back pain (LBP) These conditions were selected because of their high prevalence in older adults and they are common reasons to seek health care. (Docking et al., 2011; Y. Zhang & Jordan, 2010) Latent class cluster analysis, a model based method closely analogous to clus ter analysis but suitable for categorical data, was used to identify latent (unobservable) clusters and to classify individuals into their most likely clusters. (Formann & Kohlmann, 1996) We planned to investigate convergence of co122

PAGE 123

morbidity subgroup types between the two pain conditions as an initial assessment of subgroup generalizability. In a secondary aim, disability and cost trajectories for each subgroup were examined over 3 years We hy pothesized that comorbidity subgroups would fit the following profiles: low comorbidities, mental health comorbidities, and physical plus mental health (complex) comorbidities. This hypothesis was informed by results of Whitson et al. who reported a hierar chical structure of increasingly complex comorbidity profiles. (Whitson et al., 2016) Given the psychosocial nature of musculoskeletal pain, we predicted that mental health factors would define their own unique subgroup while separately contributing to a more complex subgroup that had high concurrent rates of physical comorbidities We also hypothesized that clusters with mental health comorbidities would have higher costs and poorer patient reported outcomes than clusters where mental health conditions were absent, since diagnoses like depression and anxiety are often associated with poorer outcomes related to musculoskeletal pain. Finally, we hypothesized comorbidity profiles that included a combination of physical and mental health comorbidities would have the highest disability and costs over 3 years. The rationale for this hypothesis is that physical and mental health comorbidities would have an additive effect that would more strongly influence functional outcomes and utilization than additive effects of physical comorbidities alone. Methods Dataset This analysis utilize d 5 years (20062010) of the 100% Cost and Use dataset from Medicare Current Beneficiary Survey ( MCBS ) a longitudinal, nationally representative in person survey of randomly sampled Medicare beneficiaries as well as 123

PAGE 124

matching Medicare claims data. Survey reported data include information on the use and cost of all types of medical services, as well as information on supplementary health insurance, living arrangements, income, health status, and physical functioning. Medicare claims data includes use and cost inform ation on inpatient hospitalizations, outpatient hospital care, physician services, home health care, durable medical equipment, skilled nursing home services, hospice care, and other medical services. Cost and Use files contain a combination of survey reported data, M edicare Fee for Service claims data, and other healthcare utilization data from CMS administrative files. New participants are enrolled annually in rotating panels (rounds) and remain in the study for 12 rounds of data collection over 3 years. Analytic File D evelopment Complete person level files were constructed using Medicare physician, inpatient, outpatient skilled nursing facility and home health claims files as well as survey reported data, across the 3 full years that participants were i ncluded in the MCBS. Diagnoses were derived using claims data only. We excluded survey reported episodes without matching claims data to ensure that all diagnoses were assigned by a medical professional. Subjects Individuals 65 years of age or older with a diagnosis of OA or LBP and that provided data through all 3 years of the survey were included in the analysis. International Classification of Diseases, Ninth Revision, Clinical Modification ( ICD9 CM ) codes in claims files from each participants index year in the survey were used to ident ify OA or LBP diagnoses (Table 4 1 ) and determine the presence of 124

PAGE 125

comorbidities (Julie M. Fritz, Childs, Wainner, & Flynn, 2012; United States Bone and Joint Initiative, 2014) Iden tification of D iagnoses A notable criticism of using claims data to identify diagnoses is that it may overestimate presence of disease due to the inclusion of ruleout diagnoses. Ruleout diagnoses are non validated diagnostic codes most often found in physician and outpatient claims files. (C. N. Klabunde, Potosky, Legler, & Warren, 2000) Physici ans may record a diagnosis to ruleout when ordering tests or making referrals as part of the diagnostic process. Hence, ruleout diagnoses may or may not represent the actual presence of a condition. To remove ruleout diagnoses from the analysis, we fo llowed up all analyses using a dataset that employed a commonly used rule out algorithm (C. N. Klabunde et al., 2000; Carrie N. Klabunde, Harlan, & Warren, 2006) This algorithm requires that for physician and outpatient claims, a patient's diagnoses must appear on at least two different claims that are more than 30 days apart. Conditions that do not appear on two different claims are considered to be "rule out" diagnoses, and are not counted among the individuals actual medical diagnoses. In a follow up analysis we considered a diagnosis to be present if listed in any diagnosis field on any claim. This process identified diagnoses without claims restrictions W e used this approach as a form of sensitivity analysis to test the stability of comorbidity clusters when using a more liberal identification criterion for medical diagnoses. Assignment of C omorbidities The Elixhauser ICD 9 CM coding algorithm defined by Quan et al. (Quan et al., 2005) was use d to identify the following comorbidities from claims in each respondents index year of entry into the MCBS: Congestive heart failure, valvular disease, 125

PAGE 126

pulmonary circulation disorders, peripheral vascular disorders, hypertension (uncomplicated), hypertension (complicated), paralysis, other neurological disorders, chronic pulmonary disease, diabetes (uncomplicated), diabetes (complicated), hypothyroidism, renal failure, liver disease, peptic ulcer disease, AIDS/HIV, lymphoma, metastatic cancer, solid tumor without metastasis, rheumatoid arthritis, coagulopathy, obesity, weight loss, fluid and electrolyte disorders, blood loss anemia, deficiency anemia, alcohol abuse, drug abuse, psychoses, and depression. We chose this comorbidity coding method because it i ncludes more ICD 9 CM codes and identifies a variety of mental health conditions ( e.g. alcohol and drug abuse, psychoses, depression) which were of particular interest given the influence of psychological var iables on painrelated outcomes (Hill & Fritz, 2011; Nicholas et al., 2011; D. C. Turk & Gatchel, 2002) Sociodemograph ic I nformation Age, sex race, income, census division education were collected from the index year of entry into the MCBS for each individual. Health Related I nformation Self reported measures of general health and healthrelated function are collected annually in the MCBS. To assess general health, respondents are asked What is your general health compared to others the same age Potential responses are Excellent Very good Good Fair and Poor Functional limitations were assessed using the Nagi disability scale. (Nagi, 1976) Questions measure difficulty with : s tooping/crouching, walking 23 blocks lifting objects weighing up to 10 lbs. reaching above shoulder and w r iting or handling small objects. Potential responses are No difficulty, A little difficulty, Some difficulty, A lot of 126

PAGE 127

difficulty, Cannot do it. A total score for the Nagi disability scale is determined by summing item responses, with total scores ranging from 525 and higher scores indicating higher levels of disability (Nagi, 1976) Healthcare E xpenditures Two different annual healthcare expenditure estimates were collected for each respondent. First, annual all cause medica l expenditure totals for each respondent were retrieved from the person level summary file for each year in the MCBS. Second, annual expenditures attributable to OA and LBP (diagnosis specific costs) were calculated from inpatient, outpatient, home health, physician, skilled nursing and durable medical equipment claims data. Claims for prescription drug benefits, which are paid through Medicare Part D, were not included in our sample data and do not appear in the analyses. Medicare payments listed on claims with an OA or LBP ICD 9 code associated in any diagnosis code field were summed for a diagnosis specific annual expenditure summary. Total diagnosis specific costs were a sum of the following: payments made by Medicare, payments made on behalf of a benef i ciary by a primary payer other than Medicare deductibles and coinsurance payments All cost estimates were adjusted to 2010 costs using consumer price indices. (Agency for Healthcare Research and Quality (US), n.d.) Statistical Analysis An a priori decision was made to include a separate OA plus LBP group (hereafter, COMBINED group) if a substantial percentage of the sample had both diagnoses (> 20% sample). In this case, three mutually exclusive groups would be inclu ded in the analysis: OA only, LBP only and COMBINED 127

PAGE 128

Identification of comorbidity subgroups Comorbidity subgroups were identified using latent class analysis (LCA). L atent class analysis is a probability clustering technique that identifies unobserved latent classes defined by the distribution of binary diagnosis indicators, in this case the presence or absence of a particular comorbidity diagnosis. (Hagenaars & McCutcheon, 2009) Therefore, each latent class represents a cluster of comorbidities that occurs most commonly in a subgroup of the populat ion. Latent class analysis is viewed as the optimal method of classification when using binary indicators and provides 1) prevalence estimates for latent classes, and 2) conditional response probabilities for each respondent. Latent Gold software (Statisti cal Innovations, Belmont, MA) was used to conduct LCA using survey weights, strata and primary sampling units to account for the complex survey design of the MCBS All t hirty comorbidity indicators included in the Elixhauser index were considered for LCA modeling. To allow for proper identification of a model, sparsely distributed indicators were excluded if present in <5% of the analytic sample. This approach assured that a sufficient number of respondents with each diagnosis were included in the analysis L atent class analysis assumes conditional independence, which states that within latent classes, each variable is statistically independent from every other variable. (Vermunt & Magidson, 2004) This assumption holds that identified latent classes should explain associations between all indicator variables in a well fit model. For this analysis, statistic al independence was not a tenable assumption among some conditions (e.g. CHF and fluid retention), leading to the possibility that model fit indices would be too high. In such cases, appropriate model fit can be achieved by relaxing the 128

PAGE 129

local independence assumption to separately model significant bivariate relationships. This approach allows the model to account for excessive associations between indicators that are not explainable by the latent classes (Vermunt & Magidson, 2004) Conditional dependence is identified in Latent Gold software by assessing bootstrapped bivariate residuals (BVR), a formal measure of the extent to which the observed association between 2 variables is reproduced by a model (O berski, Kollenburg, & Vermunt, 2013; Vermunt & Magidson, 2003, 2005) Stated more directly, BVRs measure the strength of association between indicators, after those associations are accounted for in the model. Large bivariate residuals (>2) are considered to be indicative of conditional dependence and suggest the model is not adequately capturing those relationships (Vermunt & Magidson, 2005) In these cases, model indicator pairs were allowed to correlate by including direct effects of those correlations in the model. Mul tiple statistical and nonstatistical criteria were used when selecting optimal class size. Model parsimony was assessed for models of differing numbers of classes using Bayesian Information Criterion (BIC), and Akaike Information Criterion (AIC), based on the 2LL of the model and adjusted for number of parameters and sample size (Vermunt & Magidson, 2003) with lower values indicating more parsimonious models. BIC values were weighted more heavily when identifying number of classes since BIC tends to produce more parsimonious models while AIC tends to produce larger models at the risk of defining spurious classes. Due to the large number of indicators with some known to have low prevalence rates, the potential for sparsely populated tables raised concerns for validity of the goodness of fit chi square statistic and associated pvalues. (Ong & Dulmen, 2006) To overcome this limitation, 129

PAGE 130

bootstrapped (500 iterations) Cressie Read statistic estimates were generated, with p values >.05 suggesting better goodness of fit.(Langeheine, Pannekoek, & Van, 1996) Additional criteria used to determine optimal class solution were misclassification error rate, class size, and interpretability of classes, particularly where other criterion did not produce a clearly superior model. (Nylund, Asparouhov, & Muthn, 2007) After selecting a best fit latent class model, subjects were assigned group membership based on highest posterior probability. Subgroup comparisons on patient reported outcomes and expenditures Comorbdity subgroups were compared across time points on composite functional limitation scores (Nagi disability scale), as well as adjusted all cause healthcare and diagnosis specific (LBP or OA) costs using PROC SURVEYREG. Because the Medicare Current Beneficiary Survey uses a stratified, multistage sampling scheme and oversamples certain population groups, each person had an unequal probability of being included in the survey. To obtain nationally representative population estimates, we conduct ed the analyses using MCBS threeyear backward longitudinal weight s, strata and primary sampling units t o account for the complex multistage sample design and the number of nonresponses. (Briesacher, Tjia, Doubeni, Chen, & Rao, 2012) These analyses were performed using s urvey methods in SAS version 9.3 (SAS Institute Inc., Cary, NC). The University of Florida Institutional Review Board approved this study. Results Subgroup A nalysis The analytic dataset included 12,608 respondents first enrolled in MCBS during the time from 20062008. Of those, 10,470 were 65 years or older and 7,640 were still 130

PAGE 131

enrolled by the end of the 3year survey. Thus, this study included 7,640 respondents who m et inclusion criteria, representing 30,787,466 i ndividuals in the US population. With claims restrictions, 723 respondents (9.5 % of the sample) had an OA diagnosis representing 2,663,029 individuals and 617 (8.1%) had a LBP diagnosis representing 2,418,650 individuals. Of those, 158 (2.1%) reported both an OA and LBP diagnosis (COMBINED), representing 596,799 individuals. Of respondents with OA or LBP (n=1,182) those with a COMBINED diagnosis represented 13.4% of the sample. Because the number of COMBINED diagnoses was lower than our a priori 20% threshold when using the claims restriction criterion, we analyzed 2 separate, nonmutually exclusive groups (OA, n=723; LBP, n=617). Demographic and health related information for the sample are listed in Tables 42 and 43 Model fit estimates for those with OA identified a 4class solution based on lowest sample sized adjusted BIC ( 761.52) after adding direct effects of significant BVRs to the model. However, the BIC value for a 5 class solution was only slightl y higher ( 752.21, bootstrapped CressieRead statistic p=0.24), than the 4 class model and additional fit statistics such as classification error (0.15), class interpretability and BVRs < 2 suggested a better fit compared to 4class solution. Thus, a 5 cl ass model was chosen as the final solution for LBP. Prior to model identification, pulmonary circulation disorders, paralysis, liver disease, peptic ulcer disease, AIDS/HIV, lymphoma, metastatic cancer, coagulopathy, obesity, weight loss, blood loss anemia, alcohol abuse, drug abuse and psychoses were removed from the analysis due to low prevalence (<5%). U ncomplicated hypertension was the most common comorbidity for OA and LBP and probability for the condition was high among all subgroups. Therefore, 131

PAGE 132

in de fining subgroup names, hypertension was not heavily weighted. The final subgroups were identified as (% of sample with subgroup classification): low comorbidity (47.6%), pulmonary/hypothyroidism/anemia (28.6%), hypertension with diabetes (12.2%), complicat ed hypertension/renal/anemia (5.8%), and complex cardiac/high comorbidity (5.8%) (Figure 41) Selected demographic information for the subgroups is provided in Table 4 4 Model fit estimates for those with LBP identified a 5class solution based on lowest sample sized adjusted BIC ( 858.82) after adding direct effects of significant BVRs to the model. Additional fit statistics such as bootstrapped Cressie Read statistic (p=0.02), classification error (0.11), class interpretability and BVRs < 2 suggested a better fitting model than 4or 6 class solutions. Prior to model identification, pulmonary circulation disorders, paralysis, liver disease, peptic ulcer disease, AIDS/HIV, lymphoma, metastatic cancer, coagulopathy, obesity, weight loss, blood loss anemi a, alcohol abuse, drug abuse and psychoses were removed from the analysis due to low prevalence (<5%). The final subgroups were identified as (% of sample with subgroup classification): low comorbidity (54.4%), pulmonary/hypothyroidism/anemia (21.8%), hypertension with diabetes (15.0%), complicated hypertension/renal/anemia (5.5%), and complex cardiac/high comorbidity (3.3%) (Figure 42). Selected demographic information for the subgroups is provided in Table 45 Subgroup a nalysis without claims restrictions Without claims restrictions, 1,648 respondents (21.6% of the sample) had an OA diagnosis representing 6,272,230 individuals and 1,374 (18.0%) had a LBP diagnosis representing 5,305,900 individuals. Of those, 601 (7.9%) reported both an OA and LBP diagnosis (COMBINED), representing 2,298,148 individuals. Respondents with a 132

PAGE 133

COMBINED diagnosis represented 24.8% of the sample that reported at least one of the two conditions of interest in the analysis (OA and/or LBP, n=2,421). Because the n umber of COMBINED diagnoses exceeded our a priori 20% threshold when using the no claims restriction criterion, we analyzed 3 separate, mutually exclusive groups (OA only, n=1,047; LBP only, n=773; COMBINED, n=601). Model fit estimates for those with OA i dentified a 5class solution based on lowest sample sized adjusted BIC ( 196.12) after adding direct effects of significant BVRs to the model. Additional fit statistics such as bootstrapped Cressie Read statistic (p=0.92), classification error (0.24) and BVRs < 2 suggested a better fitting model than 4 or 6class solutions. Prior to model identification, pulmonary circulation disorders, paralysis, liver disease, peptic ulcer disease, AIDS/HIV, lymphoma, metastatic cancer, blood loss anemia, alcohol abuse and drug abuse were removed from the analysis due to low prevalence (<5%). Uncomplicated hypertension was the most common comorbidity and probability for the condition was high among all subgroups. Therefore, in defining subgroup names, uncontrolled hyper tension was given less consideration. The final subgroups were identified as (% of sample with subgroup classification): low comorbidity (34.4%), hypertension with diabetes (26.2%), pulmonary/hypothyroidism/anemia (24.7%), neuro/psychological (9.1%), and c omplex cardiac/high comorbidity (5.6%) (Figure 43) Model fit estimates for those with LBP identified a 4class solution based on lowest sample sized adjusted BIC ( 801.31) after adding direct effects of significant BVRs to the model. However, the BIC v alue for a 5 class solution was only slightly higher ( 759.57) than the 4class model and additional fit statistics such as bootstrapped 133

PAGE 134

Cressie Read statistic (p=0.55), classification error (0.14) and BVRs < 2 suggested a better fit compared to 4class s olution. Thus, a 5class model was chosen as the final solution for LBP. Prior to model identification, pulmonary circulation disorders, paralysis, liver disease, peptic ulcer disease, AIDS/HIV, lymphoma, metastatic cancer, coagulopathy, obesity, weight loss, blood loss anemia, alcohol abuse, drug abuse and psychoses were removed from the analysis due to low prevalence (<5%). Again, uncomplicated hypertension was the most common comorbidity and probability for the condition was high among all subgroups. The refore, in defining subgroup names, uncontrolled hypertension was not heavily weighted. The final subgroups were identified as (% of sample with subgroup classification): low comorbidity (44.8%), cardiac disease without diabetes (20.6%), hypertension with diabetes (16.0%), complex cardiac/high comorbidity (11.5%), and depression/rheumatism (7.2%) (Figure 4 4). Model fit estimates for those with COMBINED identified a 5class solution based on lowest sample sized adjusted BIC (1180.83) after adding direct eff ects of significant BVRs to the model. Additional fit statistics such as bootstrapped Cressie Read statistic (p=0.06), classification error (0.15) and BVRs < 2 suggested a better fitting model than 4 or 6class solutions. Prior to model identification, pulmonary circulation disorders, paralysis, liver disease, peptic ulcer disease, AIDS/HIV, lymphoma, metastatic cancer, blood loss anemia, alcohol abuse and drug abuse were removed from the analysis due to low prevalence (<5%). Uncomplicated hypertension w as again the most common comorbidity and was weighted accordingly when naming subgroups. The final subgroups were identified as (% of sample with subgroup classification): low comorbidity (43.2%), pulmonary/hypothyroidism/anemia (20.4%), complicated 134

PAGE 135

hypert ension without diabetes (14.9%), hypertension with diabetes (12.6%), and complex cardiac/high comorbidity (8.9%) (Figure 4 5). Exploratory analysis including number of comorbidities Following methodology of Whitson et al we explored whether adding number of comorbidities to each model would improve model fit and reduce classification error. (Whitson et al., 2016) For each analysis, we included an additional term representi ng the comorbidity count. Results demonstrated that for each model, adding a comorbidity count did not improve model fit by any criteria and led to greater classification error in all cases. Subgroup Comparisons on Patient Reported Function and E xpenditures Subgroup contrasts at each time point were assessed. In the OA diagnostic group, Nagi disability scores were significantly higher as disease burden increased. Likewise, all cause healthcare costs wer e significantly higher for subgroups with hi gher disease burden (Figure 4 8b). For diagnosis specific costs, the complex cardiac/high comorbidity subgroup costs were similar to those of the low comorbidity group in Year 1 and lower than all other sub groups in Year 2. In Year 3 there were no signific ant differences in average costs between subgroups (Figure 4 8c). In the LBP diagnostic group, similar to the OA diagnostic group, higher disability and all cause costs were observed as a function of higher disease burden ( Figures 4 7a and 47 b). L ike with the OA diagnostic group, spending among the complex cardiac/high comorbidity sub group was not consistently higher than all other sub groups, despite a general trend toward subgroups with higher disease burden having higher spending. This was not the case in Year 3 where the complex cardiac/high comorbidity had the lowest spending ( Figure 4 7 c). 135

PAGE 136

Subgroup comparisons on patient reported function and expenditures (without claims restrictions) In the OA only diagnostic group, Nagi disability scores were significantly higher as disease burden increased, with no significant differences between the neuro/psychological and complex cardiac/high comorbidity subgroups at Years 1 and 3 (Figure 48 a). Likewise, all cause healthcare costs were significantly higher for subgroups with higher disease burden, with no significant differences between the neuro/psychological and complex cardiac/high comorbidity subgroups at Years 2 and 3 ( Figure 48 b). For diagnosis specific costs, the neuro/psychological subgroup had higher spending in Year 1 compared to all other subgroups except cardiac/high comorbidity. The complex cardiac/high comorbidity subgroup costs decreased dramatically in Year 2, and in Year 3 there were no significant di fferences in average costs between subgroups ( Figure 48 c). In the LBP only diagnostic group, subgroups with higher disease burden were associated with higher Nagi disability scores. Interestingly, the subgroup defined as depression/rheumatism demonstrated similar Nagi disability scores as the low comorbidity sub group ( Figure 49 a). All cause healthcare costs for this diagnostic group showed a pattern similar to the Nagi disability scores, with the depression/rheumatism subgroup having similar average costs to subgroups with lower comorbidity complexity ( Figure 49 b). Higher disease burden was also associated with higher diagnosis specific costs. At Year 1, the complex cardiac/high comorbidity sub group had a significantly higher mean costs, which was largely due to a small number of outliers with excessively high costs (>$20,000). In Years 2 & 3, subgroups had more similar costs, albeit differences were statistically significant ( Figure 49 c). 136

PAGE 137

For the COMBINED diagnostic group, complex cardiac/high comorbidit y and the subgroup defined by higher probabilities of pulmonary/hypothyroid/anemia had the highest level of disabil ity across all years (Figure 4 10a). For all cause costs, the complex cardiac/high comorbidity subgroup demonstrated high levels of average spending in Year 1, followed by a decrease in Years 2 and 3 to levels that continued to be significantly higher than all other subgroups. All subgroup differences in spending through Years 2 & 3 were significant ( Figure 4 10b). For diagnosis specific costs, all subgroups experienced high costs in Year 1, with the expected hierarchy of costs based on disease burden, followed by reductions in average costs in Years 2 & 3. As expected, average diagnosis specific costs for the COMBINED subgroups tended to be higher compared to the OA only and LBP only subgroups ( Figure 410c). Discussion Comorbidity subgroups for individuals with OA and LBP are defined by the presence of similar specific comorbid conditions and classification is not significantly improved by use of a simple comorbidity count. However, a clear hierarchy of disease burden was observed, with the majority of respondents falling into subgroups that were defined by lower probabilities of most comorbidities. E ven among older adults individuals with higher disease burden comprise a small proportion of the population. However, these individuals tended to report poorer health and function, along with experiencing higher costs. Our findings support prior literature that report costs are concentrated among a small, more medically complex subset of the population. (Mitchell, 2001; Riley, 2007) With the exception of the complex cardiac/high comorbidity sub group, we observed a trend toward higher diagnosis specific spending among those with higher disease burden. This might suggest that although higher 137

PAGE 138

spending for musculoskeletal pain is associated with higher disease burden, utilization of pain related services wanes as treatment for the most serious medical conditions (e.g. congestive heart failure) tak es priority. Alternatively, this finding may indicate that individuals with the highest disease burden are poor candidates for costlier invasive procedu res (e.g. surgery) or simply do not pursue them We observed convergence in subgroup profiles between individ uals with OA and those with LBP, but failed to identify distinct mental health subgroups. Interestingly, probabilities for psychoses and depression were highest in sub groups defined by higher disease burden, suggesting a high comorbidity phenotype is likely to also have characteristics of poorer mental health. The co existence of high medical comorbidity and poor mental health is an important finding. It calls into question the extent to which physiologic load due to medical comorbidities confounds the relationship between mental health and pain related outcomes Moreover, it suggests effective pain treatment pathways may need t o address both mental health and medical comorbidities. Current pain treatment guidelines stress the importance of addressing mental health, but do not often consider the presence of medical comorbidity. In analyses without claims restrictions, we were abl e to identify distinct subgroups defined by higher probabilities of mental health conditions. The discrepancy in results between analytic methods might be a function of using claims restrictions that bias against inclusion of condensed mental health treatm ent episodes commonly provided in outpatient settings. For the unrestricted claims analysis, mental health subgroups had high levels of condition specific spending for OA, but comparatively low conditionspecific spending for LBP. 138

PAGE 139

These results would need to be corroborated in future studies given the methoddependent differences in comorbidity profiles Diagnosis specific costs exhibited an interesting trend, with high average costs in the initial year followed by sharp decreases in the subsequent two years. There are two plausible explanations for this trend. First, respondents were required to have spending for OA or LBP in their first panel year to be included in the analysis Given the episodic nature of musculoskeletal pain, we would expect that some individuals would go on to have less need for healthcare utilization in Years 2 and/or 3. In fact, our data showed that approximately 40% of those who had diagnosis specific costs in Year 1 had no diagnosis specific costs in Year 2, contributing $0 in spending toward the subgroup average in that year. This finding highlights the episodic nature of utilization and spending for musculoskeletal pain and at least partially explains the apparent reduction in average subgroup diagnosis related costs. A second possible explanation is that individuals with acute or chronic pain were captured in cost estimates for Year 1, whereas those who sought care for an acute pain condition in Year 2 or Year 3, but not in Year 1 were excluded from the study. Assuming high costs are more likely to be associated with initial treatment of acute pain episodes (e.g. due to diagnostic testing and front loaded treatments), Year 1 estimates are expected to be higher. Understanding distinctions between costs in Year 1 and costs in subsequent years is important because larger, more consistent subgroup differences in diagnosis specific costs are present in Year 1 compared to other years. H igher diagnosis specific costs were generally associated with higher disease burden in Year 1 but not in subsequent years. This suggest s that when utilization for musculoskeletal 139

PAGE 140

pain occurs, costs tend to be higher for subgroups with higher disease burden. But the episodic nature of utilization and spending for musculoskeletal pain may be universal and less dependent on disease burden. Importantly we observed that high costs in Year 1 for even the complex cardiac/high comorbidity subgroup did not remain high over time. However higher variability in costs within and between years for subgroups with higher disease burden suggest s poorer stability and predictability of costs in these subgroups. An important implication of episodic utilization and higher cost variability is that estimating treatment value and effectiveness particularly am ong individuals with higher disease burden, may be more challenging. Future investigations to better explain factors underlying the initiation of treatment episodes for musculoskeletal pain are needed. To our knowledge, this is the first study to define and compare comorbidity subgroups in individu als with musculoskeletal pain. The study by Whitson et al. identified comorbidity subgroups in a general cohort study of Medicare fee for service beneficiaries using methodology that is most similar to ours among other studies that have examined comorbidity classifications. (Gore, Sadosky, Stacey, Tai, & Leslie, 2012; Gore et al., 2011; Kirkness, Yu, & Asche, 2008; Leite et al., 2011; Whitson et al., 2016) Similar to Whitson et al., we identified s eparate minimal and complex comorbid it y subgroups as well as general vascular and nonvascular subgroups However, we did not observe subgroups defined by higher probabilities of cancer or neurological disorders. This might be due to studying a different cohort, which was limited to those with two common types of musculoske letal pain diagnoses. Contrasting results may also be due to differences in methodology (Carrie N. Klabunde et al., 2006; Carrie N. 140

PAGE 141

Klabunde, Warren, & Legler, 2002) We used claims data over 1 year to identify comorbidity subgroups, while Whitson et al. used self repor ted measures of comorbidity included in the MCBS survey A notable strength of our approach is that the Elixhauser index includes a wide range of comorbidities, including mental health comorbidities, which were of particular interest in this study. Our app roach also ensured that diagnoses were made by a physician and were less subject to recall or bias. However, this more stringent identification may also be considered a limitation, as respondents would have needed to have at least one claim with a diagnosi s documented during the one year look back period t o be considered a comorbidity. Despite some differences in comorbidity classifications between the analyses, both studies report a highly prevalent low comorbidity group and a far less prevalent group defi ned by high disease burden These high and low groups might be common to many medical conditions, with other, more nuanced comorbidity subgroups dependent on specific conditions or patient populations. Our findings have important implications for clini cal practice and healthcare policy. First, a small, yet distinct subgroup of medically complex individuals with high costs and poor function was identified in both diagnostic groups using both claims criterion. This high disease burden sub group may represent a distinct pain phenotype as identified by Kittelson et al (Kittelson, Stevens Lapsley, & Schmiege, 2016) Despite the small proportion of those categorized in the high comorbidity subgroups (approximately 6%), these individuals may cont ribute disproportionately to the public health burden of musculoskeletal pain. While this study is not designed to directly assess the value of care currently provided for these individuals, it suggests individuals 141

PAGE 142

with moderate to high disease burden are ideal targets for com parative effectiveness research. Studies should be designed to develop and test treatment pathways aimed at optimizing cost and disability related outcomes for this subgroup. Second, our findings show that type and severity of comorbi dities not just the number, are associated with differential levels of spending related to musculoskeletal pain. Therefore, overall disease burden should be considered in cost effectiveness estimates and in the development of treatment pathways to prevent and manage chronic musculoskeletal pain I mproved clinical outcomes might be recognized through precision based guidelines that more thoroughly consider the moderating effects of comorbidities on costs and disability in older adults. Finally, t he highly episodic nature of healthcare utilization for musculoskeletal pain suggest s that both healthcare providers and policymakers develop delivery models that more effectively address long term management for those with chronic symptoms. (Atun, 2015) An important strength of this study is the use of dataset that provides nationally representative estimates for older adults with OA and/or LBP. This improves external validity of the findings, which have im portant implications for the development of treatment pathways and national healthcare policy addressing an aging population. However, these results do not reflect comorbidity classifications among younger individuals or in other pain conditions, which may be significantly different. Other strengths include empirical subgrouping using latent class analysis, which allows for more robust modeling with dichotomous indicators that violate t he assumption of independence. 142

PAGE 143

Some limitations should be considered when interpreting the results of this study. First, individuals were not defined at a common point in disease. As a result, the data represent only a cross sectional assessment of disease progression for a heterogeneous group of individuals. Moreover, changes in disability and costs over time represent group averages and cannot be used to make inferences about the typical progression of disease for individuals. However, our intent was not to model individual change, but rather use averages to compare the relative costs and disability among comorbidity subgroups. Mixed model growth curve analysis of individuals at a common point in their disease progression would provide greater insight into the usual course of disability and costs for those with varying comorb idity profiles. A second limitation is the use of a ruleout algorithm that was potentially too restrictive We chose this algorithm as it is common methodology when using claims to identify comorbidities (C. N. Klabunde et al., 2000) However, this approach may have led to under identification of some diagnoses for which providers are less likely to rely on clinical tests to confirm (e.g. low back pain). With respect to OA, using a far less stringent ruleout algorithm with a 2 year look back period, the chronic condition warehouse (CCW) prevalence rates for OA have been reported between 2830% for the timeframe of 20062010. (Chronic Condition Data Warehouse (CCW), 2016) These rates also include rheumatic conditions and spondylolysis diagnoses that were not included in our definition of OA. Prior to exclusion of subjects with incomplete data, prevalence of OA in our sample was approximately 33% without claims res trictions, and 18% with claims restrictions. Considering the differences in methodologies, our prevalence rates are slightly lower than previous estimates which is expected. (J. X. Zhang, Iwashyna, & 143

PAGE 144

Christakis, 1999) While it is unlikely these differences would invalidate comorbidity classifications, it is possible that some individuals, like those who rarely seek healthcare for pain, may be under represented in the current analysis. Although this algorithm may have been restrictive, we felt a consistent criterion was necessary for accurately identifying diagnoses. Therefore, an appropriate interpretation might be that results with out c laims restrictions act as a sensitivity analysis for the primary analysis with claims restrictions. More robust comorbidity classifications are identified through both methods. A third limitation is that without Part D claims data, we were unable to assess diagnosis specific group differences in spending and utilization for prescription medication. In musculoskeletal pain populations, prescription medication use can be substantial (Toblin, Mack, Perveen, & Paulozzi, 2011) and its use among different comorbidity groups, particularly those with more complex comorbidity profiles, is an important consideration for future study. The increase in opioid use and its associated public health crisis further underscores the need for future research that includes prescription medication costs Finally, we should note that although we were unable to consistently ident ify specific mental health subgroups, psychological assessment was limited to diagnosed psychological conditions. This dataset did not include information on more nuanced measures of painrelated psychological distress (e.g. fear avoidance beliefs, pain s elf efficacy, pain catastrophizing) that do not meet the strict diagnostic criteri a for a mental health condition. While psychological disorders and depression have been linked to the development and maintenance of musculoskeletal pain, more nuanced measur es of 144

PAGE 145

pain related psychological distress are of ten discussed in this context. Therefore, the inability to identify specific mental health subgroups should not be conflated with the necessary absence of subgroups that exhibit high levels of pain related ps ychological distress defined by more nuanced criteria. Future Research In addition to exploring prescription medication use in future research, studies should also assess whether comorbidity subgroups differ based on sex and age. We identified differences in sex distribution among some comorbidity subgroups but did not perform separate analyses based on sex due to the large number of sparse comorbidity indicators and relative sample size. Future work should determine whether pain management pathways may provide better value if sensitive to sex, in addition to comorbidity profile. Another important direction for f uture research is assess ing the type, timing, and pattern of healthcare utilization among comorbidity subgroups particularly those with higher dis ease burden. Defining specific care pathways for those with higher disease burden will allow for the comparison of these pathways to determine which provides the most value and for whom. As a result, healthcare providers may be better equipped to predict t reatment outcomes and therefore make more informed personalized treatment decisions. 145

PAGE 146

Table 4 1. ICD9 CM codes used to identify osteoarthritis Condition Code Description Osteoarthiritis 715.x0 Osteoarthritis, site unspecified 715.x1 Osteoarthritis of the shoulder region 715.x2 Osteoarthritis of the upper arm 715.x3 Osteoarthritis of the forearm 715.x4 Osteoarthritis of the hand 715.x5 Osteoarthritis of the hip/pelvis 715.x6 Osteoarthritis of the knee/lower leg 715.x7 Osteoarthritis of the ankle/foot 715.x8 Osteoarthritis, other specified sites 715.89 Osteoarthritis involving, or with mention of more than one site, but not specified as generalized, multiple sites 721.3 Lumbosacral spondylosis without myelopathy 722.1 Lumbar disc displacement 722.52 Lumbar/ lumbosacral disc displacement 722.73 Lumbar disc disease with myelopathy 722.93 Other disc disorder lumbar region Low back pain 724.02 Spinal stenosis lumbar 724.2 Lumbago 724.3 Sciatica 724.4 Thoracic or lumbosacral neuritis or radiculitis, unspecified 724.5 Backache, unspecified 756.11 Spondylolysis, lumbosacral region 756.12 Spondylolesthesis 846.0 Sprain lumbosacral 846.1 Sprain sacroiliac 846.8 Sprain other specified sites of sacroiliac region 846.9 Sprain unspecified site of sacroiliac region 847.2 Sprain lumbar region 847.3 Sprain sacrum 146

PAGE 147

Table 4 2 Demographic information of th e sample by diagnostic category Factor Osteoarthritis (n=723) Low back pain (n=617) Parameter* Weighted fq Percent SE Parameter* Weighted fq Percent SE Age 76.96 0.27 74.97 0.3 Sex Males 218 820042 30.79 1.8 220 856859 35.43 1.94 Females 505 1842987 69.21 1.8 397 1561791 64.57 1.94 Race Don't know 1 3461 0.01 0.13 1 3461 0.14 0.14 American Indian 8 26590 0.05 1 7 27029 1.12 0.49 Asian/Pacific Islander 19 76736 0.11 2.88 11 60912 2.52 0.83 Black/African American 55 182837 0.15 6.87 29 116849 4.83 0.92 White 622 2309944 0.4 86.74 555 2154448 89.08 1.49 More than one 15 53602 0.06 2.01 10 39961 1.65 0.53 Other 3 9859 0.03 0.37 4 15990 0.66 0.4 Income $25,000 or less 397 1366297 51.31 5.13 278 1045388 43.22 5.93 $25,001 $50,000 219 847241 31.82 5.16 221 879622 36.36 5.89 $50,001 or more 107 449491 16.88 1.69 118 493639 20.41 2.24 Census New England 24 94873 3.56 1.56 16 69148 2.86 1.35 Middle Atlantic 109 432552 16.24 2.04 89 339964 14.06 1.72 East North Central 118 424042 15.92 2.19 109 417653 17.27 2.73 West North Central 58 209032 7.85 1.52 43 166465 6.88 1.76 South Atlantic 175 604655 22.71 3.13 141 563480 23.3 3.61 East South Central 60 214547 8.06 2.1 43 165407 6.84 1.44 West South Central 61 202637 7.61 2.28 75 282986 11.7 3.7 Mountain 42 150066 5.64 1.62 37 138778 5.74 1.68 Pacific 69 310532 11.66 2.54 62 268694 11.11 2.75 Puerto Rico 7 20093 0.75 0.22 2 6075 0.25 0.25 Education Dont know 8 18858 0.71 0.25 3 8869 0.37 0.22 Refused 1 3559 0.13 0.13 0 No schooling 13 43304 1.63 0.53 3 9377 0.39 0.22 Nursery to 8th grade 91 293028 11 1.1 53 191834 7.93 1.11 9 th to 12 th grade, no diploma 101 358901 13.48 1.36 72 261611 10.82 1.33 High school graduate 212 781274 29.34 2.24 195 780313 32.26 2.37 Vocational, Tech, Business, etc. 51 204831 7.69 1.02 50 208054 8.6 1.17 Some college, no degree 103 395841 14.86 1.59 99 381796 15.79 1.73 Associates degree 18 74541 2.8 0.68 25 101741 4.21 0.86 Bachelor's degree or higher 125 488892 18.36 2.52 117 475056 19.65 2.77 Nagi Disability Score (5 25) 11.74 0.22 11.22 0.21 General Health (1 5) 2.72 0.06 2.76 0.06 *Parameter is frequency for all variables except for age where value is mean age in years 147

PAGE 148

Table 4 3 Sample and weighted prevalence of individual comorbidities by diagnostic category Comorbidity Osteoarthritis (n=723) Low back pain (n=617) Sample f requency W eighted f requency SE (freq) % SE (%) Sample f requency Weighted Frequency SE (freq) % SE (%) Congestive heart failure 104 347058 36956 13.03 1.39 62 212735 26678 8.8 1.03 Valvular disease 58 197715 26998 7.42 1.06 38 142874 20561 5.91 0.82 Pulmonary circulation disorders 13 45975 13370 1.73 0.49 9 33429 11302 1.38 0.46 Peripheral vascular disorders 96 320923 39928 12.05 1.45 55 193030 30075 7.98 1.22 Hypertension (uncomplicated) 531 1940750 90275 72.88 1.96 364 1376514 82297 56.91 2.06 Hypertension (complicated) 91 311152 33772 11.68 1.32 53 208053 32719 8.6 1.3 Paralysis 23 73299 15794 2.75 0.6 10 36566 11944 1.51 0.49 Other neurological disorders 81 251730 34818 9.45 1.25 36 128603 22376 5.32 0.83 Chronic pulmonary disease 141 510501 45119 19.17 1.57 96 376450 37638 15.56 1.58 Diabetes (uncomplicated) 182 660420 49825 24.8 1.66 131 507300 43177 20.97 1.67 Diabetes (complicated) 44 155073 26939 5.82 0.98 39 156965 29650 6.49 1.2 Hypothyroidism 155 543566 42997 20.41 1.44 87 332046 41922 13.73 1.54 Renal failure 58 199586 25397 7.49 0.97 34 130416 24461 5.39 0.96 Liver disease 6 20470 8779 0.77 0.32 7 26666 10508 1.1 0.43 Peptic ulcer disease 1 3931 3931 0.15 0.15 0 0 0 0 0 AIDS/HIV 0 0 0 0 0 0 0 0 0 0 Lymphoma 2 9197 6540 0.35 0.25 6 25415 11252 1.05 0.47 Metastatic cancer 1 2816 2816 0.11 0.11 2 6500 4637 0.27 0.19 Solid tumor without metastasis 50 188186 26356 7.07 0.96 42 154371 23277 6.38 0.9 Rheumatoid arthritis 56 213008 34613 8 1.3 48 190887 29760 7.89 1.05 Coagulopathy 18 61682 14977 2.32 0.57 17 57280 14802 2.37 0.61 Obesity 32 131674 24583 4.94 0.87 15 58034 16559 2.4 0.68 Weight loss 16 49037 10930 1.84 0.42 5 15075 8024 0.62 0.32 Fluid and electrolyte disorders 92 322343 34609 12.1 1.25 53 192245 27563 7.95 1.03 Blood loss anemia 21 74556 17781 2.8 0.66 11 34981 10679 1.45 0.44 Deficiency anemia 170 588830 45704 22.11 1.55 90 328753 39072 13.59 1.55 Alcohol abuse 2 6292 4511 0.24 0.17 1 3671 3671 0.15 0.15 Drug abuse 5 15138 5699 0.57 0.21 4 12255 6213 0.51 0.25 Psychoses 35 117063 23125 4.4 0.84 22 89467 21601 3.7 0.85 Depression 77 265558 31743 9.97 1.16 44 146102 22656 6.04 0.88 148

PAGE 149

Figure 4 1 Comorbidity classification structure for osteoarthritis diagnosis categor y. Abbreviations: CHF = Congestive heart failure, VALVDIS = Valvular disease, PVD = Peripheral vascular disorders, HTN_UN = Hypertension (uncomplicated), HTN_COMP = Hypertension (complicated), NE URO = Other neurological disorders, PULM_CHR = Chronic pulmonary disease, DIAB_UN = Diabetes (uncomplicated), DIAB_COM = Diabetes (complicated), HYPOTHY = Hypothyroidism, RENAL = Renal failure, TUMOR = Solid tumor without metastasis, RHEUM = Rheumatoid art hritis, FLUID = Fluid and electrolyte disorders, ANEMIA_D = Deficiency anemia, DEPRES = Depression 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Probability of Class Inclusion Comorbidity Low comorbidity Pulmonary/Hypothyroidism/Anemia Hypertension with diabetes Complicated hypertension/Renal/Anemia Complex cardiac/High comorbidity149

PAGE 150

Table 4 4 Characteristics of those assigned to each osteoarthritis subgroup Variable Low comorbidity (n=348) Pulmonary/ hypothyroidism/ anemia (n=184) Hypertension with diabetes (n=108) Complicated hypertension/renal/ anemia (n=42) Complex cardiac/high comorbidity (n=41) Freq* % SE (%) Freq* % SE (%) Freq* % SE (%) Freq* % SE (%) Freq* % SE (%) Age (years) 75.74 (0.35) 79.74 (0.53) 76.07 (0.73) 78.41 (1.26) 77.61 (1.18) Sex Male 117 34.3 3.2 37 20.0 2.6 34 31.1 4.9 18 44.2 8.3 12 30.4 8.2 Female 231 65.8 3.2 147 79.9 2.6 74 68.9 4.9 24 55.7 8.3 29 69.6 8.2 Race Don't know 0 0. 0 0.0 0.0 0.0 0.0 0 0.0 0.0 1 2.4 2.4 0 0.0 0.0 American Indian 5 1.2 0.6 1 0.7 0.7 0 0.0 0.0 1 1.8 1.9 1 2.1 2.1 Asian/Pacific Islander 9 2.7 1.6 3 2.5 1.5 3 3 1.6 0 0.0 0.0 4 9.1 4.3 Black/African American 20 5.4 1.1 15 6.8 2.0 13 10.6 2.4 3 7.5 4.4 4 9.9 4.9 White 303 87.8 1.8 162 88.4 2.6 90 84.7 3.1 36 86.5 5.5 31 75.4 6.9 More than one 9 2.4 0.9 2 1.1 0.7 2 1.7 1.2 1 1.7 1.7 1 3.4 3.4 Other 2 0. 5 0. 4 1 0.6 0. 6 0 0.0 0.0 0 0.0 0.0 0 0.0 0.0 *Freq is frequency for all variables except for age where value is mean age in years **SE (freq) is standard error of the frequency for all variables except for age where value is standard error of age in years 150

PAGE 151

Figure 4 2 Comorbidity classific ation structure for low back pain diagnosis category A bbreviations: CHF = Congestive heart failure, VALVDIS = Valvular disease, PVD = Peripheral vascular disorders, HTN_UN = Hypertension (uncomplicated), HTN_COMP = Hypertension (complicated), NEURO = Othe r neurological disorders, PULM_CHR = Chronic pulmonary disease, DIAB_UN = Diabetes (uncomplicated), DIAB_COM = Diabetes (complicated), HYPOTHY = Hypothyroidism, RENAL = Renal failure, TUMOR = Solid tumor without metastasis, RHEUM = Rheumatoid arthritis, FLUID = Fluid and electrolyte disorders, ANEMIA_D = Deficiency anemia, DEPRES = Depression 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Probability of Class Inclusion Comorbidity Low comorbidity Pulmonary/Hypothyroidism/Anemia Hypertension with diabetes Complicated hypertension/Renal/Anemia Complex cardiac/High comorbidity151

PAGE 152

Table 4 5 Characteristics of those assigned to each low back pain subgroup Variable Low comorbidity (n=343) Pulmonary/ h ypothyroidism/ a nemia (n=128) Hypertension with diabetes (n=96) Complicated hypertension/renal/ a nemia (n=30) Complex cardiac/h igh comorbidity (n=20) Freq* % SE (%) Freq* % SE (%) Freq* % SE (%) Freq* % SE (%) Freq* % SE (%) Age (Years) 74.26 (0.39) 77.61 (0.64) 74.4 3 (0.48) 75.35 (1.41) 73.65 (0.97) Sex Male 125 36.0 2.7 39 29.3 3.7 42 43.0 5.3 9 31.9 9.8 5 30.8 7.8 Female 218 64.0 2.7 89 70.7 3.7 54 57.0 5.3 21 68.1 9.8 15 69.2 7.8 Race Don't know 0 0.0 0.0 0 0.0 0.0 0 0.0 0.0 1 2.9 2.9 0 0.0 0.0 American Indian 5 1.4 0.7 2 1.8 0.9 0 0.0 0.0 0 0.0 0.0 0 0.0 0.0 Asian/Pacific Islander 6 2.8 1.1 4 3.9 2.1 1 1.1 0.1 0 0.0 0.0 0 0.0 0.0 Black/African American 9 2.7 1.0 7 5.7 2.4 7 8.3 3.0 4 12.6 5.8 2 10.0 6.6 White 318 91.7 1.7 113 87.2 3.2 84 86.6 3.3 23 77.1 7.9 17 83.6 8.4 More than one 4 1.2 0.7 1 0.7 0.7 3 2.9 1.2 1 3.7 3.7 1 6.4 6.1 Other 1 0.3 0.3 1 0.8 0.8 1 1.2 0.1 1 3.7 3.6 0 0.0 0.0 *Freq is frequency for all variables except for age where value is mean age in years **SE (freq) is standard error of the frequency for all variables except for age where v alue is standard error of age in years. 152

PAGE 153

Table 4 6 Sample and weighted prevalence of comorbidities by diagnostic category ( without claims restrictions) Comorbidity Osteoarthritis only (n=1,047) Low back pain only (n=773) Combined osteoarthritis/low back pain (n=601) Sample frequency Weighted frequency % SE (%) Sample frequency Weighted frequency % SE (%) Sample frequency Weighted frequency % SE (%) Congestive heart failure 152 488907 12.3 1.0 112 402136 13.4 1.6 109 378204 16.5 1.6 Valvular disease 181 659312 16.6 1.5 102 368684 12.3 1.3 123 450642 19.6 1.8 Pulmonary circulation disorders 38 134048 3.4 0.6 34 134257 4.5 0.9 23 97234 4.2 0.9 Peripheral vascular disorders 226 802720 20.2 1.2 151 540268 18.0 1.4 151 561132 24.4 1.9 Hypertension (uncomplicated) 816 3056981 76.9 1.5 544 2074236 69.0 1.8 489 1849954 80.5 2.0 Hypertension (complicated) 169 609517 15.3 1.3 123 438261 14.6 1.3 111 417105 18.2 1.7 Paralysis 24 82244 2.1 0.4 9 29385 1.0 0.3 12 39843 1.7 0.5 Other neurological disorders 134 418431 10.5 1.1 77 256420 8.5 1.1 81 276717 12.0 1.3 Chronic pulmonary disease 258 954698 24.0 1.4 179 678824 22.6 1.4 172 650987 28.3 2.0 Diabetes (uncomplicated) 284 1057384 26.6 1.3 230 854918 28.4 1.5 198 740373 32.2 2.3 Diabetes (complicated) 89 315002 7.9 0.9 72 250490 8.3 1.0 64 268246 11.7 1.7 Hypothyroidism 275 1000102 25.2 1.6 176 661185 22.0 1.3 176 626237 27.3 2.0 Renal failure 93 318589 8.0 0.8 51 181353 6.0 0.8 62 238568 10.4 1.2 Liver disease 17 71067 1.8 0.4 10 37349 1.2 0.4 10 32547 1.4 0.5 Peptic ulcer disease 3 10953 0.3 0.2 2 5408 0.2 0.1 1 2890 0.1 0.1 AIDS/HIV 0 0 0.0 0.0 1 2540 0.1 0.1 1 3510 0.2 0.2 Lymphoma 9 31524 0.8 0.3 9 37701 1.3 0.4 4 13674 0.6 0.3 Metastatic cancer 7 20975 0.5 0.2 10 34837 1.2 0.3 9 33751 1.5 0.5 Solid tumor without metastasis 100 384374 9.7 1.0 96 337051 11.2 1.1 84 311759 13.6 1.4 Rheumatoid arthritis 80 300852 7.6 0.8 57 227998 7.6 0.9 83 315232 13.7 1.6 Coagulopathy 60 211447 5.3 0.7 26 92173 3.1 0.6 31 112975 4.9 0.9 153

PAGE 154

Table 4 6 Continued. Comorbidity Osteoarthritis only (n=1,047) Low back pain only (n=773) Combined osteoarthritis/low back pain (n=601) Sample frequency Weighted frequency % SE (%) Sample frequency Weighted frequency % SE (%) Sample frequency Weighted frequency % SE (%) Obesity 69 285224 7.2 0.8 27 108030 3.6 0.7 42 177303 7.7 1.1 Weight loss 57 204124 5.1 0.7 31 108211 3.6 0.7 37 125622 5.5 0.9 Fluid and electrolyte disorders 166 596134 15.0 1.1 93 321168 10.7 1.2 110 402318 17.5 1.4 Blood loss anemia 26 89530 2.3 0.4 16 60151 2.0 0.5 18 66388 2.9 0.7 Deficiency anemia 290 1029323 25.9 1.4 157 560396 18.6 1.5 183 678996 29.6 2.0 Alcohol abuse 4 16586 0.4 0.2 4 14893 0.5 0.3 5 15728 0.7 0.3 Drug abuse 1 2640 0.1 0.1 2 6627 0.2 0.2 8 24865 1.1 0.4 Psychoses 60 189924 4.8 0.6 35 133575 4.4 0.9 43 164858 7.2 1.2 Depression 101 362278 9.1 1.0 74 283009 9.4 1.3 87 309349 13.5 1.4 154

PAGE 155

Figure 43 Comorbidity classification structure for osteoarthritis only diagnosis category (without claims restrictions). Abbreviations: CHF = Congestive heart failure, VALVDIS = Valvular disease, PVD = Peripheral vascular disorders, HTN_UN = Hypertension (uncomplicated), HTN_COMP = Hypertension (complicated), NEURO = Other neurological disorders, PULM_CHR = Chronic pulmonary disease, DIAB_UN = Diabetes (uncomplicated), DIAB_COM = Diabetes (complicated), HYPOTHY = Hypothyroidism, RENAL = Renal failure, TUMOR = Solid t umor without metastasis, RHEUM = Rheumatoid arthritis, COAG = Coagulopathy, OBESITY = Obesity, WEIGHTL = Weight loss, FLUID = Fluid and electrolyte disorders, ANEMIA_D = Deficiency anemia, PSYCH = Psychoses, DEPRES = Depression 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Probability of Class Inclusion Comorbidity Low comorbidity Hypertension with diabetes Pulmonary/Hypothyroidism/Anemia Neuro/Psychological Complex cardiac/High comorbidity155

PAGE 156

Figure 4 4 Comorbidity classification structure for low back pain only diagnosis category (without claims restrictions). Abbreviations: CHF = Congestive heart failure, VALVDIS = Valvular disease, PVD = Peripheral vascular disorders, HTN_UN = Hypertension (uncomplicat ed), HTN_COMP = Hypertension (complicated), NEURO = Other neurological disorders, PULM_CHR = Chronic pulmonary disease, DIAB_UN = Diabetes (uncomplicated), DIAB_COM = Diabetes (complicated), HYPOTHY = Hypothyroidism, RENAL = Renal failure, TUMOR = Solid tu mor without metastasis, RHEUM = Rheumatoid arthritis, FLUID = Fluid and electrolyte disorders, ANEMIA_D = Deficiency anemia, DEPRES = Depression 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Probability of Class Inclusion Comorbidity Low comorbidity Cardiac disease without diabetes Hypertension with diabetes Complex cardiac/High comorbidity Depression/Rheumatism156

PAGE 157

Figure 4 5 Comorbidity classification structure for combined ost eoarthritis/low back pain diagnosis category (without claims restrictions) Abbreviations: CHF = Congestive heart failure, VALVDIS = Valvular disease, PVD = Peripheral vascular disorders, HTN_UN = Hypertension (uncomplicated), HTN_COMP = Hypertension (comp licated), NEURO = Other neurological disorders, PULM_CHR = Chronic pulmonary disease, DIAB_UN = Diabetes (uncomplicated), DIAB_COM = Diabetes (complicated), HYPOTHY = Hypothyroidism, RENAL = Renal failure, TUMOR = Solid tumor without metastasis, RHEUM = Rheumatoid arthritis, COAG = Coagulopathy, OBESITY = Obesity, WEIGHTL = Weight loss, FLUID = Fluid and electrolyte disorders, ANEMIA_D = Deficiency anemia, PSYCH = Psychoses, DEPRES = Depression 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Probability of Class Inclusion Comorbidity Low comorbidity Pulmonary/Hypothyroidism/Anemia Complicated hypertension without diabetes Hypertension with diabetes Complex cardiac/High comorbidity157

PAGE 158

A B C Figure 46 Osteoarthritis subgroup longitudinal p rofiles. A) NAGI disability score, B) Annual all healthcare costs, C) Annual costs attributed to treatment of osteoarthritis 158

PAGE 159

A B C Figure 47 Low back pain subgroup longitudinal profiles. A) NAGI disability score, B) Annual all healthcare costs, C) Annual costs attributed to treatment of low back pain 159

PAGE 160

A B C Figure 48 Osteoarthritis only subgroup longitudinal profiles (without claims restrictions). A) NAGI disability score, B) Annual all healthcare costs, C) Annual costs attributed to treatment of osteoarthritis 160

PAGE 161

A B C Figure 49 Low back painonly subgroup longitudinal profiles (without claims restrictions). A) NAGI disability score, B) Annual all healthcare costs, C) Annual costs attributed to treatment of low back pain 161

PAGE 162

A B C Figure 410. Combined osteoarthritis and low back pain subgroup longitudinal profiles (without claims restrictions). A) NAGI disability score, B) Annual all healthcare costs, C) Annual costs attributed to treatment of osteoarthritis and low back pain 162

PAGE 163

C HAPTER 5 CONCL USION Summary of Findings and Future Directions These dissertation projects studied healthcare utilization and costs among populations with musculoskeletal pain that were a priori defined as at risk for low value care Syntheses of the major findings and their implications for healthcare delivery, research, and policy are listed below. Persistent High Cost Healthcare Utilization is C oncentrated among a Small P ercentage of Individuals with Musculoskeletal P ain I ndividuals with more disabling pain and in mul tiple anatomical regions who have private insurance are at risk of persistently high pain related expenditures In the US healthcare system, those with private insurance tend to have higher expenditures However, evidence suggests that in the contest of healthcare for musculoskeletal pain, more intense utilization or selection of costlier services do not necessarily yield greater value. (Julie M. Fritz, Brennan, & Hunter, 2015; Julie M. Fritz et al., 2012) F ut ure research should better define how the type and temporal characteristics of utilization vary as a function of insurance coverage and whether these characteristics impact value This information will assist healthcare providers and policymakers to develop delivery models that provide high value care regardless of insurance status. A n important potential policy consideration is how to direct patients toward low er cost high er value pathways even if they have the access to and means for pursuing costl ier services. The concentration of persistently high costs among a small percentage of the population might also indicate the presence of a healthcareseeking trait among some individuals with musculoskeletal pain. In a preliminary analysis of healthcare costs 163

PAGE 164

within 2 years following hip surgery, Cook and Rhon found pre operative healthcare seeking behavior to predict post operative costs for a variety of hiprelated and nonhiprelated healthcare costs. (Cook & Rhon, 2017) T his and other similar studies (Ot ani & Baden, 2009) suggest an important, understudied behavioral trait related to healthcare entitlement and expectation that drives utilization independent of symptom severity or comorbid conditions We did not measure healthcare seeking behavior direc tly in these projects, but future studies should more thoroughly assess its influence and make suggestions for clinical management. Comorbidity C omplexity is an Important Contributor to U tilization and C osts for M usculoskeletal P ain Predictors of utilization and costs in these projects were multidimensional supporting the application of established models of healthcare utilization (e.g. Andersen Model) to musculoskeletal pain. However, some predictors, such as comorbidity complexity, stood out due to their consistency among findings in all 3 projects Both chronic pain and multiple comorbidity have been classified as significant emerging public health concerns (Benjamin, 2010; Goldberg & McGee, 2011) making their coexistence especially concerning. The causal relationship between medical comorbidities and pain is not well understood. However, higher physiologic load may reduce resilience to pain, or at least make individuals more likely to seek help for it This finding is important since comorbidities are likely to moderate treatment effects, influencing both pati ent reported outcomes and cos ts. Yet most guidelines for the management of pain do not consider modifications to interventions or prognosis based on the presence of comorbid conditions. 164

PAGE 165

Higher disease burden may also be associated with escalation of care that is more likely to have negative downstream health and financial consequences. I n the OSPRO project comorbidity level and change in pain predicted subsequent opioid use following physical therapy. Using MEPS data, we observed that opioid use and higher comorbidity levels predict ed persistently high expenditures for musculoskeletal pain. Together these results point to individuals with complex comorbidities as a high priority initial target for comparative effectiveness research. Of particular interest are those with complex cardi ac morbidity or complicated renal disease/hypertension/anemia. Understanding which treatment pathways are most effective for these individuals might lead to reduced risk of opioid use and high costs among a particularly vulnerable population. One potential approach for reducing risk of opioid use and persistently high spending is to develop pathways that incorporate both early pain reduction techniques and interventions that address concomitant needs of comorbid chronic conditions. Self Reported Health Measures and Change in S ymptoms ov er T ime I nfluence Costs and U tilization, but A re Under Represented in Survey and Claims D ata Pain is a common reason to seek care initially, and in these analyses we found that the continuation of pain was a common reason to escalate care for musculoskeletal pain. S elf reported pain and disability, as well as changes in those variables over time, were important predictors of healthcare utilization. The importance of self reported data is supported by prior research showing improved prediction of costs compared to using demographic information alone. (Fleishman & Cohen, 2010) However, c laims based and medical record datasets often lack patient reported information, and rarely include l ongitudinal assessment of that informati on In the OSPRO study, change in pain was an important, nearly universal predictor of subsequent healthcare utilization. High 165

PAGE 166

priority comparative effectiveness studies for musculoskeletal pain would first identify patients who were resistant to early changes in pain, then determine which healthcare pathways mitigate their risk for care escalation. Yet without this type of information in most claims based or medical record datasets, the ability to perform comparative effectiveness research on existing treatment approaches provided in clinical settings is nearly impossible. Therefore, it is necessary to prospectively develop standardized datasets that include patient reported information and can be pragmatically administered for multiple pain conditions acro ss the episode of care. It is only after this type of data collection is implemented on a large scale that useful comparative effectiveness analyses can be conducted. Organizations that adopt this data collection approach will be ideally suited to conduct comparative effectiveness studies aimed at optimizing value and integrate their findings into clinical care. Psychological Factors Do Not Contribute Substantially to Healthcare Costs and U tiliz ation After Considering Other P redisposing and Enabling Factor s Poorer mental health can contribute to persistently high healthcare utilization for musculoskeletal pain and may suggest higher risk for use of certain healthcare services, such as surgery following physical therapy. However, those with poorer mental hea lth may not comprise a unique subgroup, complicating the delivery of risk stratified care based on psychological distress. Interestingly, characteristics like depression and psychological distress were strongly associated with utilization and costs in univ ariate analyses, but these relationships were attenuated when other factors such as pain, disability, comorbid conditions and insurance were included in the models. In the context of healthcare utilization for pain, psychological distress does not appear t o be a strong, unique predictor after considering other predisposing and enabling 166

PAGE 167

characteristics. These findings can be contrasted with studies that assess the influence of psychological distress on clinical outcomes such as pain intensity or disability, wh ere distress is a more reliable predictor. Reasons for this disparate finding have been previously discussed in the Dissertation but include 1) the use of psychological measures in survey and claims analyses that were not specific to pain, 2) controlling for variables commonly used as outcomes in clinical pain studies (e.g. pain interference, self reported function or pain intensity), and 3) the inclusion of enabling factors such as insurance that could more significantly influence variation in utilization compared to psychological distress. A Considerable Proportion of Patients R eport A dditional Healthcare U tilization after an Episode of Physical Therapy for Musculoskeletal P ain The relatively high proportion of patients seeking subsequent healthcare following physical therapy should be a concern for physical therapists. It is inevitable that some patients will not respond well to physical therapy or require more persistent care f or chronic symptoms. It is also possible that physical therapy does deliver added value even if additional care is required. The acceptable rate for subsequent opioid use following physical therapy is unknown, and we acknowledge that the proportion would l ikely be much higher for patients that do not seek physical therapy at all. However, the high proportion of individuals seeking additional care within 1 year of physical therapy initial evaluation suggests there is room to improve the effectiveness of care that is currently delivered. We should further study those who seek additional care to determine how physical therapists can better meet their needs. These patients may be more likely to have chronic symptoms; therefore, physical therapists should develop and test sustainable long term management pathways Moreover, physical therapists need 167

PAGE 168

methods of better identifying those who are likely to experience escalation of care following physical therapy and those who are not. Before advocating a physical ther apy first approach for musculoskeletal conditions, physical therapists need to make certain that they can identify and deliver the right treatment for the right patient at the appropr iate time. These treatments must also be provided reliably across providers so that consistency of outcomes is improved. Physical therapists are uniquely positioned as nonpharmacological providers of musculoskeletal pain care and non pharmacological care is strongly advocated for in recent practice guidelines. (J L. Clarke et al., 2016; Dowell et al., 2016; Von Korff et al., 2016) However, fully embracing that frontline position requires diligence to identify shortcomings of care and build more effective delivery models that reduce risk for care escalation and maximize likelihood of good clinical outcomes 168

PAGE 169

APPENDIX A LIST OF MUSCULOSKELETAL CONDITION ICD9 CODES ICD 9 codes included in the analysis 715 : Osteoarthrosis and allied disorders 716 : Other and unspecified arthroplasties 717 : Internal derangement of knee 718 : Other derangement of joint 719 : Other and unspecified disorders of joint 720 : Ankylosing spondylitis and other inflammatory spondylopathies 721 : Spondylosis and allied disorders 722 : I ntervertebral disc disorders 723 : Other disorder of cervical region 724 : Other and unspecified disorders of back 725 : Polymyalgia rheumatica 726 : Peripheral enthesopathies and allied syndromes 727 : Synovitis and tenosynovitis 728 : Disorders of muscle, ligament, and fascia 729 : Other disorders of soft tissue 730 : Acute osteomyelitis 731 : Osteitis deformans and osteopathies associated with other disorders classified elsewhere 732 : Osteochondropathies 733 : Other disorders of bone and cartilage (Osteoporosis; pathologic fracture, cyst, necrosis of bone, malunion and nonunion of fracture) 734 : Flat foot 735 : Acquired deformities of toe 736 : Acquired deformities of forearm 737 : Curvature of spine 738 : Other acquired deformity (of musculoskeletal system), spondylolisthesis 739 : Nonallopathic lesions, not elsewhere classified 805 : Fracture of vertebral column without mention of spinal cord injury 808 : Fracture of pelvis (Acetabulum, closed) 809 : Ill defined fractures of bones and trunk 810 : Fracture of clavicle (closed) 811 : Fracture of scapula (closed) 812 : Fracture of humerus (Upper end, closed) 813 : Fracture of radius and ulna (Upper end, closed) 814 : Fracture of carpal bone(s) (Closed) 815 : Fracture of metacarpal bone(s) (Closed) 816 : Fracture of one or more phalanges of hand (Closed) 817 : Multiple fractures of hand bones 818 : Ill defined fractures of upper limb 820 : Fracture of neck of femur (transcervical fracture, closed) 821 : Fracture of other and unspecified parts of femur (Shaft or unspecified part, closed) 822 : Fracture of patella 169

PAGE 170

823 : Fracture of tibia and fibula, upper end (closed) 824 : Fracture of ankle 825 : Fracture of one or more tarsal and metatarsal bones 826 : Fracture of one or more phalanges of foot 827 : Other, multiple, and ill defined fractures of lower limb 829 : Fractures of unspecified bones 831 : Dislocation of shoulder 832 : Dislocation of elbow 833 : Dislocation of wrist 834 : Dislocation of finger 835 : Dislocation of hip 836 : Dislocation of knee 837 : Dislocation of ankle 838 : Dislocation of foot 839 : Other, multiple, and ill defined dislocations 840 : Sprains and strains of shoulder and upper arm 841 : Sprains and strains of elbow and forearm 842 : Sprains and strains of wrist and hand 843 : Sprains and strains of hip and thigh 844 : Sprains and strains of knee and leg 845 : Sprains and strains of ankle and foot 846 : Sprains and strains of sacroiliac region 847 : Sprains and strains of other and unspecified parts of back 848 : Other and ill defined s prains and strains 922 : Contusion of trunk 923 : Contusion of upper limb 924 : Contusion of lower limb and of other and unspecified sites 954 : Injury to other nerve(s) of trunk, excluding shoulder and pelvic girdles 955 : Injury to peripheral nerve(s) of shoulder girdle and upper limb 956 : Injury to peripheral nerve(s) of pelvic girdle and lower limb 959 : Injury, other and unspecified (to musculoskeletal system) ICD 9 codes excluded from the analysis 135 : Sarcoidosis 170 : Malignant neoplasm of bone and articular cartilage 171 : Malignant neoplasm of connective and other soft tissue 198 : Secondary malignant neoplasm of bone and bone marrow 203 : Multiple myeloma and immunoproliferative neoplasms 213 : Benign neoplasm of bone and articular cartilage 215 : Other benign neoplasm of connective and other soft tissue 238 : Neoplasm of uncertain behavior of other and unspecified sites and tissues; Connective and other soft tissue; Bone soft tissue and skin 239.2 : Neoplasms of unspecified nature; Bone soft tissue and skin 274 : Gout; Gouty arthroplathy 354 : Mononeuritis of upper limb and mononeuritis multiplex 710 : Diffuse diseases of connective tissue 170

PAGE 171

711 : Arthropathy associated with infections 712 : Crystal arthropathies 713 : Arthropathy associated wit h other disorders classified elsewhere 714 : Rheumatoid arthritis and other inflammatory polyarthropathies 741 : Spina bifida 754 : Certain congenital musculoskeletal deformities 755 : Other congenital anomalies of limbs (Polydactyly) 756 : Other congenita l musculoskeletal anomalies 806 : Fracture of vertebral column with mention of spinal cord injury 807 : Fracture of vertebral column with mention of spinal cord injury 819 : Multiple fractures involving both upper limbs, and upper limb with rib(s) and ster num 875 : Open wound of chest (wall) 876 : Open wound of back 877 : Open wound of buttock 879 : Open wound of other and unspecified sites (except limbs) 880 : Open wound of shoulder and upper arm 881 : Open wound of elbow, forearm, and wrist 882 : Open wound of hand except finger(s) alone 883 : Open wound of finger(s) 884 : Multiple and unspecified open wound of upper limb 885 : Traumatic amputation of thumb 886 : Traumatic amputation of other finger(s) 887 : Traumatic amputation of arm and hand (complete) (partial) 890 : Open wound of hip and thigh 891 : Open wound of knee, leg [except thigh], and ankle 892 : Open wound of foot except toe(s) alone 893 : Open wound of toe(s) 894 : Multiple and unspecified open wound of lower limb 895 : Traumatic amputation of toe(s) 896 : Traumatic amputation of foot (complete) (partial) 897 : Traumatic amputation of leg(s) (complete) (partial) 926 : Crushing injury of trunk 927 : Crushing injury of upper limb 928 : Crushing injury of lower limb 929 : Crushing injury of multi ple and unspecified sites 996 : Complications peculiar to certain specified procedures V43.6 : Organ or tissue replaced by other means (joint) V54 : Other orthopaedic aftercare V67 : Follow up examination, following surgery 171

PAGE 172

APPENDIX B DEFINITION AND ANALYSIS CODING OF C ATEGORICAL VARIABLES Sex [SEX] 1. Male (reference) 2. Female Race [RACEX] 1. White 2. Black 3. American Indian/Alaska Native 4. Asian 5. Native Hawaiian/Pacific Islander 6. Multiple race reported Race groups were collapsed for analysis: 1. White (reference) 2. Black 3. Other (American Indian/Alaska Native, Asian, Native Hawaiian/Pacific Islander, multiple races). Ethnicity [HISPANX] 1. Hispanic 2. Non Hispanic (reference) Years of Education [HIDEG and EDRECODE] 1. No degree 2. GED 3. High school diploma 4. Bachelors degree 5. Masters degree 6. Doctorate degree 7. Other degree Years of education was collapsed for analysis: 1. High school diploma or less (reference) 2. Bachelors degree or more. Smoking Status [ADSMOK2] 1. Yes 2. No (reference) Poverty Category [POVCAT] 1. Negative or Poor: Per sons in families with income less than or equal to the poverty line and includes those who reported negative income. 2. Near poor: Persons in families with income over the poverty line through 125 percent of the poverty line. 172

PAGE 173

3. Low income: Persons in families w ith income over 125 percent through 200 percent of the poverty line. 4. Middle income: Persons in families with income over 200 percent through 400 percent of the poverty line. 5. High income: Persons in families with income over 400 percent of the poverty line. Poverty category was collapsed for analysis: 1. Poor/Negative/Near Poor 2. Low Income 3. Middle Income 4. High Income (reference) Employment Status (Round 1) [EMPST1] 1. Employed 2. Job to return to 3. Job during the reference period 4. Not employed Employment status was collapsed for analysis: 1. Employed (Employed, Job to return to, Job during reference period) (reference) 2. Unemployed (not employed) Metropolitan Statistical Area (MSA ) [MSAY1] 1. Non MSA 2. MSA (reference) Census region [REGIONY1] 1. Northeast (reference) 2. Midwest 3. South 4. West Days of Work Missed due to Illness [DDNWRK1 DDNWRK5] The number of times the respondent lost a half day or more from work because of illness, injury, or mental or emotional problems were recorded for each of the 5 rounds in Year 1 and summed for total days of work missed. Days of work missed was dichotomiz ed for analysis: 1. no missed days (reference) 2. one of more days missed Pain Interference [ADPAIN2] Pain interference with work and daily activities was assessed using the following SF 12 question from the Adult SAQ: During past 4 weeks, how much has pain interfered with normal work outside the home and housework? [ 173

PAGE 174

Possible responses include: 1. None 2. A little bit 3. Moderately 4. Quite a bit 5. Extremely Pain interference was dichotomized for analysis: 1. None/A little bit/Moderately (reference) 2. Quite a bit/Extremely Perceived Health Status [RTHLTH1] Perceived health status was reported as fair/poor (reference) or excellent/very good/good. [Lee, 2014] 1. Excellent 2. Very good 3. Good 4. Fair 5. Poor Perceived health status was dichotomized for analysis: 1. Excellent/Very good/Good (reference) 2. Fair/Poor Perceived Mental Health Status [MNHLTH1] Perceived mental health status was reported as fair/poor (reference) or excellent/very good/good. [Lee, 2014] 1. Excellent 2. Very good 3. Good 4. Fair 5. Poor Perceived health status was dichotomized for analysis: 1. Excellent/Very good/Good (reference) 2. Fair/Poor Can Overcome Ills [ADOVER2] The Adult SAQ includes a question that ascertains whether the respondent believes they can overcome illness without help from a medically trained p erson. 1. Disagree strongly 2. Disagree somewhat 3. Uncertain 4. Agree somewhat 5. Agree strongly 174

PAGE 175

Overcome ills was dichotomized for analysis: 1. Disagree strongly/Disagree somewhat 2. Uncertain/Agree somewhat/Agree strongly (reference) Usual Healthcare Provider [HAVEUS2] 1. Yes 2. No (reference) Health Insurance [PRIVATEY1, MEDICARE Y1, MEDICAIDY1, OTHP UBAY1, OTHPUBBY1, TRICAREY1 INSURANYY1] MEPS provides information on monthly payer status for each of the following: TRICARE, Medicare, Medicaid/SCHIP, Other Public A Insurance, Other Public B Insurance, or private insurance. MEPS also includes summary measures that indicate whether or not a person has any insurance in a month. Each respondent was categorized as being privately insured all year, publicly insured all year, uninsur ed part of the year and either privately or publicly insured the remainder or uninsured all year. [PMC4254135] Insurance status was collapsed for analysis: 1. Privately insured all year 2. Publicly insured all year 3. Uninsured part of the year and either privat ely or publicly insured the remainder, or uninsured all year (reference) Diagnosis [ICD9CODX] 1. Diseases of The Musculoskeletal System and Connective Tissue diagnosis/diagnoses only (reference) 2. Musculoskeletal Injury diagnosis with or without Diseases of the Musculoskeletal System and Connective Tissue diagnosis/diagnoses 175

PAGE 176

LIST OF REFERENCES Abbott, J. H., Chapple, C. M., Fitzgerald, G. K., Fritz, J. M., Childs, J. D., Harcombe, H., & Stout, K. (2015). Th e Incremental Effects of Manual Therapy or Booster Sessions in Addition to Exercise Therapy for Knee Osteoarthritis: A Randomized Clinical Trial. The Journal of Orthopaedic and Sports Physical Therapy 45(12), 975 983. https://doi.org/10.2519/jospt.2015.6015 Agency for Healthcare Research and Quality (US). (n.d.). Using Appropriate Price Indices for Analyses of Health Care Expendittures or Income Across Multiple Years. Retrieved from https://meps.ahrq.gov/about_meps/Price_Index.shtml AHRQ Effective Health Care Program. (n.d.). What Is Comparative Effectiveness Research. Retrieved June 1, 2017, from http://effectivehealthcare.ahrq.gov/index.cfm/what is comparativeeffectiveness research1/ Alrwaily, M., Timko, M., Schneider, M., Stevans, J., Bise, C., Harihar an, K., & Delitto, A. (2016). Treatment Based Classification System for Low Back Pain: Revision and Update. Physical Therapy 96 (7), 1057 1066. https://doi.org/10.2522/ptj.20150345 Andersen, R. M. (1995). Revisiting the behavioral model and access to medic al care: does it matter? Journal of Health and Social Behavior 36(1), 1 10. Atun, R. (2015). Transitioning health systems for multimorbidity. Lancet (London, England) 386 (9995), 721 722. https://doi.org/10.1016/S01406736(14)622546 Babitsch, B., Gohl, D ., & von Lengerke, T. (2012). Rerevisiting Andersens Behavioral Model of Health Services Use: a systematic review of studies from 19982011. Psycho Social Medicine, 9 Doc11. https://doi.org/10.3205/psm000089 Bair, M. J., Robinson, R. L., Katon, W., & Kr oenke, K. (2003). Depression and pain comorbidity: a literature review. Archives of Internal Medicine, 163(20), 2433 2445. https://doi.org/10.1001/archinte.163.20.2433 Bartley, E. J., & Fillingim, R. B. (2013). Sex differences in pain: a brief review of cl inical and experimental findings. British Journal of Anaesthesia, 111(1), 52 58. https://doi.org/10.1093/bja/aet127 Beaton, D. E., Wright, J. G., Katz, J. N., & Upper Extremity Collaborative Group. (2005). Development of the QuickDASH: comparison of three item reduction approaches. The Journal of Bone and Joint Surgery. American Volume, 87 (5), 1038 1046. https://doi.org/10.2106/JBJS.D.02060 176

PAGE 177

Beneciuk, J. M., Fritz, J. M., & George, S. Z. (2014). The STarT Back Screening Tool for prediction of 6month clinic al outcomes: relevance of change patterns in outpatient physical therapy settings. The Journal of Orthopaedic and Sports Physical Therapy 44 (9), 656 664. https://doi.org/10.2519/jospt.2014.5178 Benjamin, R. M. (2010). Multiple chronic conditions: a public health challenge. Public Health Reports (Washington, D.C.: 1974) 125 (5), 626 627. https://doi.org/10.1177/003335491012500502 Berwick, D. M., Nolan, T. W., & Whittington, J. (2008). The triple a im: care, health, and cost. Health Affairs (Project Hope) 27(3), 759 769. https://doi.org/10.1377/hlthaff.27.3.759 Blumenthal, D., Stremikis, K., & Cutler, D. (2013). Health care spending--a giant slain or sleeping? The New England Journal of Medicine, 369 (26), 2551 2557. https://doi.org/10.1056/NEJMhpr1310415 Blyth, F. M., Van Der Windt, D. A., & Croft, P. R. (2015). Chronic Disabling Pain: A Significant Public Health Problem. American Journal of Preventive Medicine, 49(1), 98 101. https://doi.org/10.1016 /j.amepre.2015.01.008 Boissonnault, W. G. (1999). Prevalence of comorbid conditions, surgeries, and medication use in a physical therapy outpatient population: a multicentered study. The Journal of Orthopaedic and Sports Physical Therapy 29 (9), 506519 525. https://doi.org/10.2519/jospt.1999.29.9.506 Bolton, J. E. (1999). Accuracy of recall of usual pain intensity in back pain patients. Pain 83(3), 533 539. Boyles, R., Toy, P., Mellon, J., Hayes, M., & Hammer, B. (2011). Effectiveness of manual physical t herapy in the treatment of cervical radiculopathy: a systematic review. The Journal of Manual & Manipulative Therapy 19 (3), 135 142. https://doi.org/10.1179/2042618611Y.0000000011 Briesacher, B. A., Tjia, J., Doubeni, C. A., Chen, Y., & Rao, S. R. (2012). Methodological issues in using multiple years of the Medicare current beneficiary survey. Medicare & Medicaid Research Review 2 (1). https://doi.org/10.5600/mmrr.002.01.a04 Brge, E., Monnin, D., Berchtold, A., & Allet, L. (2016). CostEffectiveness of Ph ysical Therapy Only and of Usual Care for Various Health Conditions: Systematic Review. Physical Therapy 96 (6), 774 786. https://doi.org/10.2522/ptj.20140333 Burwell, S. M. (2015). Setting value based payment goals --HHS efforts to improve U.S. health care. The New England Journal of Medicine, 372(10), 897 899. https://doi.org/10.1056/NEJMp1500445 177

PAGE 178

Calver, J., Brameld, K. J., Preen, D. B., Alexia, S. J., Boldy, D. P., & McCaul, K. A. (2006). Highcost users of hospital beds in Western Australia: a population based record linkage study. The Medical Journal of Australia, 184(8), 393 397. Carey, T. S., Freburger, J. K., Holmes, G. M., Jackman, A., Knauer, S., Wallace, A., & Darter, J. (2010). Race, care seeking, and utilization for chronic back and neck pain: population perspectives. The Journal of Pain: Official Journal of the American Pain Society 11(4), 343 350. https://doi.org/10.1016/j.jpain.2009.08.003 Chang, H. Y., Boyd, C. M., Leff, B., Lemke, K. W., Bodycombe, D. P., & Weiner, J. P. (2016). Identifying Consistent Highcost Users in a Health Plan: Comparison of Alternative Prediction Models. Medical Care, 54(9), 852 859. https://doi.org/10.1097/MLR.0000000000000566 Charlesworth, C. J., Smit, E., Lee, D. S. H., Alramadhan, F., & Odden, M. C. (2015). Polypharmacy Among Adults Aged 65 Years and Older in the United States: 19882010. The Journals of Gerontology. Series A, Biological Sciences and Medical Sciences 70(8), 989 995. https://doi.org/10.1093/gerona/glv013 Charlson, M. E., Pompei, P., Ales, K. L., & MacKenzie, C. R. (1987). A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. Journal of Chronic Diseases 40(5), 373 383. Charlson, M., Wells, M. T., Ullman, R., King, F., & Shmukler, C. (2014). The Charl son comorbidity index can be used prospectively to identify patients who will incur high future costs. PloS One 9 (12), e112479. https://doi.org/10.1371/journal.pone.0112479 Chechulin, Y., Nazerian, A., Rais, S., & Malikov, K. (2014). Predicting patients w ith high risk of becoming highcost healthcare users in Ontario (Canada). Healthcare Policy = Politiques De Sante, 9 (3), 68 79. Chi, M., Lee, C., & Wu, S. (2011a). Multiple morbidity combinations impact on medical expenditures among older adults. Archives of Gerontology and Geriatrics 52(3), e210214. https://doi.org/10.1016/j.archger.2010.11.013 Chi, M., Lee, C., & Wu, S. (2011b). The prevalence of chronic conditions and medical expenditures of the elderly by chronic condition indicator (CCI). Archives of Gerontology and Geriatrics 52(3), 284 289. https://doi.org/10.1016/j.archger.2010.04.017 Childs, J. D., Piva, S. R., & Fritz, J. M. (2005). Responsiveness of the numeric pain rating scale in patients with low back pain. Spine, 30(11), 1331 1334. 178

PAGE 179

Chou, R ., Qaseem, A., Owens, D. K., Shekelle, P., & Clinical Guidelines Committee of the American College of Physicians. (2011). Diagnostic imaging for low back pain: advice for highvalue health care from the American College of Physicians. Annals of Internal Medicine 154(3), 181 189. https://doi.org/10.7326/000348191543 201102010 00008 Chronic Condition Data Warehouse (CCW). (2016). Musculoskeletal and Joint Conditions Trends, 2005 2014. Retrieved from https://www.ccwdata.org/web/guest/medicarecharts/medicare chronic conditioncharts Clarke, A. E., Levinton, C., Joseph, L., Penrod, J., Zowall, H., Sibley, J. T., Esdaile, J. M. (1999). Predicting the short term direct medical costs incurred by patients with rheumatoid arthritis. The Journal of Rheumatology 26(5), 1068 1075. Clarke, J. L., Skoufalos, A., & Scranton, R. (2016). The American Opioid Epidemic: Population Health Implications and Potential Solutions. Report from the National Stakeholder Panel Population Health Management 19 Suppl 1, S1 10. https://doi.org/10.1089/pop.2015.0144 Cohen, J. (1997). Sample Design of the Medical Expenditure Panel Survey Household Component (MEPS Methodology Report No. 2. No. AHCPR publication 97 0027). Rockville ( MD): Agency for Health Care Policy and Research. Cohen, S. (2009). Testing for Survey Attrition Effects on Estimates of Chronic Disease Costs in a National Longitudinal Medical Expenditure Survey Agency for Healthcare Research and Quality. Committee on Geographic Variation in Health Care Spending and Promotion of High Value Care, Board on Health Care Services, & Institute of Medicine. (2013). Variation in Health Care Spending: Target Decision Making, Not Geography (J. P. Newhouse, A. M. Garber, R. P. Grah am, M. A. McCoy, M. Mancher, & A. Kibria, Eds.). Washington (DC): National Academies Press (US). Retrieved from http://www.ncbi.nlm.nih.gov/books/NBK201647/ Cook, C. E., & Rhon, D. (2017). Does Health CareSeeking Behavior 12 Months Prior to Hip Surgery Dr ive Health Care Costs 2 Years After Surgery? Journal of Orthopaedic & Sports Physical Therapy 47 (1), A7. Coughlan, D., Yeh, S. T., ONeill, C., & Frick, K. D. (2014). Evaluating direct medical expenditures estimation methods of adults using the medical ex penditure panel survey: an example focusing on head and neck cancer. Value in Health: The Journal of the International Society for Pharmacoeconomics and Outcomes Research, 17 (1), 90 97. https://doi.org/10.1016/j.jval.2013.10.004 179

PAGE 180

Cross, M. J., March, L. M., Lapsley, H. M., Byrne, E., & Brooks, P. M. (2006). Patient self efficacy and health locus of control: relationships with health status and arthritis related expenditure. Rheumatology (Oxford, England) 45 (1), 92 96. https://doi.org/10.1093/rheumatology/k ei114 Dagenais, S., Caro, J., & Haldeman, S. (2008). A systematic review of low back pain cost of illness studies in the United States and internationally. The Spine Journal: Official Journal of the North American Spine Society 8 (1), 8 20. https://doi.org /10.1016/j.spinee.2007.10.005 Daigle, M. E., Weinstein, A. M., Katz, J. N., & Losina, E. (2012). The costeffectiveness of total joint arthroplasty: a systematic review of published literature. Best Practice & Research. Clinical Rheumatology 26(5), 649 65 8. https://doi.org/10.1016/j.berh.2012.07.013 de Oliveira, C., Cheng, J., Vigod, S., Rehm, J., & Kurdyak, P. (2016). Patients With High Mental Health Costs Incur Over 30 Percent More Costs Than Other HighCost Patients. Health Affairs (Project Hope) 35(1) 36 43. https://doi.org/10.1377/hlthaff.2015.0278 de Rooij, M., van der Leeden, M., Avezaat, E., Hkkinen, A., Klaver, R., Maas, T., Dekker, J. (2014). Development of comorbidity adapted exercise protocols for patients with knee osteoarthritis. Clinical Interventions in Aging, 9 829 842. https://doi.org/10.2147/CIA.S55705 Druaz Luyet, A., NGoran, A. A., Tandjung, R., Frey, P., Zeller, A., Haller, D. M., Herzig, L. (2015). Multimorbidity in primary care: protocol of a national cross sectional study i n Switzerland. BMJ Open, 5 (10), e009165. https://doi.org/10.1136/bmjopen2015009165 Deyle, G. D., Allison, S. C., Matekel, R. L., Ryder, M. G., Stang, J. M., Gohdes, D. D., Garber, M. B. (2005). Physical therapy treatment effectiveness for osteoarthriti s of the knee: a randomized comparison of supervised clinical exercise and manual therapy procedures versus a home exercise program. Physical Therapy 85 (12), 1301 1317. Deyle, G. D., Henderson, N. E., Matekel, R. L., Ryder, M. G., Garber, M. B., & Allison S. C. (2000). Effectiveness of manual physical therapy and exercise in osteoarthritis of the knee. A randomized, controlled trial. Annals of Internal Medicine, 132(3), 173 181. Deyo, R. A., Cherkin, D. C., & Ciol, M. A. (1992). Adapting a clinical comorbidity index for use with ICD 9 CM administrative databases. Journal of Clinical Epidemiology 45(6), 613 619. 180

PAGE 181

Docking, R. E., Fleming, J., Brayne, C., Zhao, J., Macfarlane, G. J., Jones, G. T., & Cambridge City over 75s Cohort Study collaboration. (2011). Epidemiology of back pain in older adults: prevalence and risk factors for back pain onset. Rheumatology (Oxford, England) 50(9), 1645 1653. https://doi.org/10.1093/rheumatology/ker175 Dowell, D., Haegerich, T. M., & Chou, R. (2016). CDC Guideline for Pr escribing Opioids for Chronic Pain--United States, 2016. JAMA, 315(15), 1624 1645. https://doi.org/10.1001/jama.2016.1464 Drapeau, A., BeaulieuPrvost, D., Marchand, A., Boyer, R., Prville, M., & Kairouz, S. (2010). A lifecourse and time perspective on the construct validity of psychological distress in women and men. Measurement invariance of the K6 across gender. BMC Medical Research Methodology 10, 68. https://doi.org/10.1186/1471 228810 68 Emeche, U. (2015). Is a strategy focused on super utilizers equal to the task of health care system transformation? Yes. Annals of Family Medicine, 13(1), 6 7. https://doi.org/10.1370/afm.1746 Fitzpatrick, T., Rosella, L. C., Calzavara, A., Petch, J., Pinto, A. D., Manson, H., Wodchis, W. P. (2015). Looking Beyond Income and Education: Socioeconomic Status Gradients Among Future High Cost Users of Health Care. American Journal of Preventive Medicine 49(2), 161 171. https://doi.org/10.1016/j.amepre.2015.02.018 Fleishman, J. A., & Cohen, J. W. (2010). Using information on clinical conditions to predict high cost patients. Health Services Research, 45 (2), 532 552. https://doi.org/10.1111/j.14756773.2009.01080.x Fordyce, W. E. (1994). Pain and suffering: what is the unit? Quality of Life Research: An International Journal of Quality of Life Aspects of Treatment, Care and Rehabilitation, 3 Suppl 1, S51 56. Formann, A K., & Kohlmann, T. (1996). Latent class analysis in medical research. Statistical Methods in Medical Research, 5 (2), 179 211. https://doi.org/10.1177/096228029600500205 Freburger, J. K., Carey, T. S., & Holmes, G. M. (2006). Effectiveness of physical the rapy for the management of chronic spine disorders: a propensity score approach. Physical Therapy 86(3), 381 394. Freburger, J. K., Holmes, G. M., Agans, R. P., Jackman, A. M., Darter, J. D., Wallace, A. S., Carey, T. S. (2009). The rising prevalence of chronic low back pain. Archives of Internal Medicine, 169(3), 251 258. https://doi.org/10.1001/archinternmed.2008.543 181

PAGE 182

Fritz, J. M., Brennan, G. P., & Hunter, S. J. (2015). Physical Therapy or Advanced Imaging as First Management Strategy Following a New C onsultation for Low Back Pain in Primary Care: Associations with Future Health Care Utilization and Charges. Health Services Research, 50(6), 1927 1940. https://doi.org/10.1111/1475 6773.12301 Fritz, J. M., Childs, J. D., Wainner, R. S., & Flynn, T. W. (20 12). Primary care referral of patients with low back pain to physical therapy: impact on future health care utilization and costs. Spine, 37(25), 2114 2121. https://doi.org/10.1097/BRS.0b013e31825d32f5 Fritz, J. M., & Irrgang, J. J. (2001). A comparison of a modified Oswestry Low Back Pain Disability Questionnaire and the Quebec Back Pain Disability Scale. Physical Therapy 81 (2), 776 788. Gandhi, S. O., Grant, L. P., & Sabik, L. M. (2014). Trends in nonemergent use of emergency departments by health insurance status. Medical Care Research and Review: MCRR, 71(5), 496 521. https://doi.org/10.1177/1077558714541481 Garfinkel, S. A., Riley, G. F., & Iannacchione, V. G. (1988). High cost users of medical care. Health Care Financing Review 9 (4), 41 52. Gaskin, D J., & Richard, P. (2012). The economic costs of pain in the United States. The Journal of Pain: Official Journal of the American Pain Society 13(8), 715 724. https://doi.org/10.1016/j.jpain.2012.03.009 George, S. Z., Beneciuk, J. M., Bialosky, J. E., Le ntz, T. A., Zeppieri, G., Pei, Q., & Wu, S. S. (2015). Development of a Review of Systems Screening Tool for Orthopaedic Physical Therapists: Results From the Optimal Screening for Prediction of Referral and Outcome (OSPRO) Cohort. The Journal of Orthopaedic and Sports Physical Therapy 45 (7), 512 526. https://doi.org/10.2519/jospt.2015.5900 George, S. Z., Coronado, R. A., Beneciuk, J. M., Valencia, C., Werneke, M. W., & Hart, D. L. (2011). Depressive symptoms, anatomical region, and clinical outcomes for p atients seeking outpatient physical therapy for musculoskeletal pain. Physical Therapy 91 (3), 358 372. https://doi.org/10.2522/ptj.20100192 Goldberg, D. S., & McGee, S. J. (2011). Pain as a global public health priority. BMC Public Health, 11, 770. https: //doi.org/10.1186/1471245811770 Gore, M., Sadosky, A., Stacey, B. R., Tai, K.S., & Leslie, D. (2012). The burden of chronic low back pain: clinical comorbidities, treatment patterns, and health care costs in usual care settings. Spine, 37(11), E668 677. https://doi.org/10.1097/BRS.0b013e318241e5de 182

PAGE 183

Gore, M., Tai, K.S., Sadosky, A., Leslie, D., & Stacey, B. R. (2011). Clinical comorbidities, treatment patterns, and direct medical costs of patients with osteoarthritis in usual care: a retrospective claim s database analysis. Journal of Medical Economics 14(4), 497 507. https://doi.org/10.3111/13696998.2011.594347 Green, S. B. (1991). How Many Subjects Does It Take To Do A Regression Analysis. Multivariate Behavioral Research, 26(3), 499 510. https://doi.o rg/10.1207/s15327906mbr2603_7 Greene, J., Hibbard, J. H., Sacks, R., Overton, V., & Parrotta, C. D. (2015). When patient activation levels change, health outcomes and costs change, too. Health Affairs (Project Hope) 34(3), 431 437. https://doi.org/10.1377/hlthaff.2014.0452 Guilcher, S. J. T., Bronskill, S. E., Guan, J., & Wodchis, W. P. (2016). Who Are the High Cost Users? A Method for PersonCentred Attribution of Health Care Spending. PloS One 11 (3), e0149179. https://doi.org/10.1371/journal.pone.0149179 Guthrie, B., Makubate, B., Hernandez Santiago, V., & Dreischulte, T. (2015). The rising tide of polypharmacy and drug drug interactions: population database analysis 19952010. BMC Medicine, 13, 74. https://doi.org/10.1186/s12916 015 0322 7 Hagenaars, J. A., & McCutcheon, A. L. (Eds.). (2009). Applied Latent Class Analysis (1 edition). Cambridge: Cambridge University Press. Harris, L. J., Graetz, I., Podila, P. S. B., Wan, J., Waters, T. M., & Bailey, J. E. (2016). Characteristics of Hospital and Emergenc y Care Super utilizers with Multiple Chronic Conditions. The Journal of Emergency Medicine, 50(4), e203214. https://doi.org/10.1016/j.jemermed.2015.09.002 Harris, R. J. (2001). A Primer of Multivariate Statistics (3 edition). Mahwah, N.J: Psychology Press. Hart, D. L., Werneke, M. W., Deutscher, D., George, S. Z., Stratford, P. W., & Mioduski, J. E. (2011). Using intake and change in multiple psychosocial measures to predict functional status outcomes in people wi th lumbar spine syndromes: a preliminary analysis. Physical Therapy 91 (12), 1812 1825. https://doi.org/10.2522/ptj.20100377 Hartmann, J., Jacobs, S., Eberhard, S., von Lengerke, T., & Amelung, V. (2016). Analysing predictors for future high cost patients using German SHI data to identify starting points for prevention. European Journal of Public Health, 26(4), 549 555. https://doi.org/10.1093/eurpub/ckv248 Hensel, J. M., Taylor, V. H., Fung, K., & Vigod, S. N. (2016). Rates of Mental Illness and Addiction among HighCost Users of Medical Services in Ontario. Canadian 183

PAGE 184

Journal of Psychiatry. Revue Canadienne De Psychiatrie, 61(6), 358 366. https://doi.org/10.1177/0706743716644764 Hernn, M. A., Hernndez Daz, S., & Robins, J. M. (2004). A structural approach to selection bias. Epidemiology (Cambridge, Mass.) 15(5), 615 625. Heslop, L., Athan, D., Gardner, B., Diers, D., & Poh, B. C. (2005). An analysis of highcost users at an Australian public health service organ ization. Health Services Management Research, 18(4), 232 243. https://doi.org/10.1258/095148405774518633 Hibbard, J. H., Greene, J., & Overton, V. (2013). Patients with lower activation associated with higher costs; delivery systems should know their patients scores. Health Affairs (Project Hope) 32(2), 216 222. https://doi.org/10.1377/hlthaff.2012.1064 Hill, J. C., & Fritz, J. M. (2011). Psychosocial influences on low back pain, disability, and response to treatment. Physical Therapy 91(5), 712 721. h ttps://doi.org/10.2522/ptj.20100280 Horner Johnson, W., Dobbertin, K., Lee, J. C., Andresen, E. M., & Expert Panel on Disability and Health Disparities. (2014). Disparities in health care access and receipt of preventive services by disability type: analys is of the medical expenditure panel survey. Health Services Research, 49 (6), 1980 1999. https://doi.org/10.1111/1475 6773.12195 HudsonCook, N., Tomes Nicholson, K., & Breen, A. (1989). A revised Oswestry disability questionnaire. In M. Roland & J. Jenner (Eds.), Back Pain: New Approaches to Rehabilitation and Education (pp. 187 204). New York, NY: Manchester University Press. Hughes, L. S., Peltz, A., & Conway, P. H. (2015). State innovation model initiative: a state led approach to accelerating health car e system transformation. JAMA, 313(13), 1317 1318. https://doi.org/10.1001/jama.2015.2017 Institute of Medicine (US) Committee on Advancing Pain Research, Care, and Education. (2011). Relieving Pain in America: A Blueprint for Transforming Prevention, Care, Education, and Research. Washington (DC): National Academies Press (US). Retrieved from http://www.ncbi.nlm.nih.gov/books/NBK91497/ Institute of Medicine (US) Roundtable on EvidenceBased Medicine. (2010). The Healthcare Imperative: Lowering Costs and Im proving Outcomes: Workshop Series Summary (P. L. Yong, R. S. Saunders, & L. Olsen, Eds.). Washington (DC): National Academies Press (US). Retrieved from http://www.ncbi.nlm.nih.gov/books/NBK53920/ 184

PAGE 185

Irrgang, J. J., Anderson, A. F., Boland, A. L., Harner, C. D., Kurosaka, M., Neyret, P., Shelborne, K. D. (2001). Development and validation of the international knee documentation committee subjective knee form. The American Journal of Sports Medicine, 29(5), 600 613. https://doi.org/10.1177/03635465010290051301 Jensen, M. P., Turner, J. A., Romano, J. M., & Fisher, L. D. (1999). Comparative reliability and validity of chronic pain intensity measures. Pain 83(2), 157 162. Jewell, D. V., Moore, J. D., & Goldstein, M. S. (2013). Delivering the physical therapy value proposition: a call to action. Physical Therapy 93 (1), 104 114. https://doi.org/10.2522/ptj.20120175 Johnson, T. L., Rinehart, D. J., Durfee, J., Brewer, D., Batal, H., Blum, J., Gabow, P. (2015). For many patients who use large amounts of health c are services, the need is intense yet temporary. Health Affairs (Project Hope) 34(8), 1312 1319. https://doi.org/10.1377/hlthaff.2014.1186 Kamper, S. J., Apeldoorn, A. T., Chiarotto, A., Smeets, R. J. E. M., Ostelo, R. W. J. G., Guzman, J., & van Tulder, M. W. (2015). Multidisciplinary biopsychosocial rehabilitation for chronic low back pain: Cochrane systematic review and metaanalysis. BMJ (Clinical Research Ed.) 350, h444. Karjalainen, K., Malmivaara, A., Mutanen, P., Pohjolainen, T., Roine, R., & Hurr i, H. (2003). Outcome determinants of subacute low back pain. Spine, 28(23), 2634 2640. https://doi.org/10.1097/01.BRS.0000099097.61495.2E Kent, P., Keating, J. L., & Leboeuf Yde, C. (2010). Research methods for subgrouping low back pain. BMC Medical Research Methodology 10, 62. https://doi.org/10.1186/1471 228810 62 Kessler, R. C., Andrews, G., Colpe, L. J., Hiripi, E., Mroczek, D. K., Normand, S. L. T., Zaslavsky, A. M. (2002). Short screening scales to monitor population prevalences and trends in non specific psychological distress. Psychological Medicine, 32(6), 959 976. Kirkness, C. S., Yu, J., & Asche, C. V. (2008). The effect on comorbidity and pain in patients with osteoarthritis. Journal of Pain & Palliative Care Pharmacotherapy 22(4), 336 348. https://doi.org/10.1080/15360280802536649 Kittelson, A. J., Stevens Lapsley, J. E., & Schmiege, S. J. (2016). Determination of Pain Phenotypes in Knee Osteoarthritis: A Latent Class Analysis Using Data From the Osteoarthritis Initiative. Arthritis Care & Research, 68 (5), 612 620. https://doi.org/10.1002/acr.22734 Klabunde, C. N., Harlan, L. C., & Warren, J. L. (2006). Data sources for measuring comorbidity: a comparison of hospital records and medicare claims for cancer patients. Medical Care, 44 (10), 921 928. https://doi.org/10.1097/01.mlr.0000223480.52713.b9 185

PAGE 186

Klabunde, C. N., Potosky, A. L., Legler, J. M., & Warren, J. L. (2000). Development of a comorbidity index using physician claims data. Journal of Clinical Epidemiology 53(12), 1258 1267. Klabunde, C. N., Warren, J. L., & Legler, J. M. (2002). Assessing comorbidity using claims data: an overview. Medical Care, 40(8 Suppl), IV 26 35. https://doi.org/10.1097/01.MLR.0000020936.03651.2D Kroenke, K., Spitzer, R. L., & Williams, J. B. W. (2003). The Patient Health Questionnaire2: validity of a two item depression screener. Medical Care, 41(11), 1284 1292. https://doi.org/10.1097/01.MLR.0000093487.78664.3C Kuhn, J. E., Dunn, W. R., Sanders, R., An, Q., Baumgarten, K. M., Bishop, J. Y., MOON Shou lder Group. (2013). Effectiveness of physical therapy in treating atraumatic full thickness rotator cuff tears: a multicenter prospective cohort study. Journal of Shoulder and Elbow Surgery 22(10), 1371 1379. https://doi.org/10.1016/j.jse.2013.01.026 Langeheine, R., Pannekoek, J., & Van, D. P. F. (1996). Bootstrapping Goodness of Fit Measures in Categorical Data Analysis. Sociological Methods and Research, 24(4), 492 516. Larochelle, M. R., Liebschutz, J. M., Zhang, F., Ross Degnan, D., & Wharam, J. F. (2016). Opioid Prescribing After Nonfatal Overdose and Association With Repeated Overdose. Annals of Internal Medicine, 165 (5), 376 377. https://doi.org/10.7326/L16 0168 Lawrence, R. C., Felson, D. T., Helmick, C. G., Arnold, L. M., Choi, H., Deyo, R. A., N ational Arthritis Data Workgroup. (2008). Estimates of the prevalence of arthritis and other rheumatic conditions in the United States. Part II. Arthritis and Rheumatism 58(1), 26 35. https://doi.org/10.1002/art.23176 Leeuw, M., Goossens, M. E. J. B., Lin ton, S. J., Crombez, G., Boersma, K., & Vlaeyen, J. W. S. (2007). The fear avoidance model of musculoskeletal pain: current state of scientific evidence. Journal of Behavioral Medicine, 30(1), 77 94. https://doi.org/10.1007/s10865006 90850 Leininger, L. J., Saloner, B., & Wherry, L. R. (2015). Predicting highcost pediatric patients: derivation and validation of a population based model. Medical Care, 53(8), 729 735. https://doi.org/10.1097/MLR.0000000000000391 Leite, A. A., Costa, A. J. G., Lima, B. de A. M. de, Padilha, A. V. L., Albuquerque, E. C. de, & Marques, C. D. L. (2011). Comorbidities in patients with osteoarthritis: frequency and impact on pain and physical function. Revista Brasileira De Reumatologia, 51(2), 118 123. 186

PAGE 187

Lentz, T. A ., Beneciuk, J. M., Bialosky, J. E., Zeppieri, G., Dai, Y., Wu, S. S., & George, S. Z. (2016). Development of a Yellow Flag Assessment Tool for Orthopaedic Physical Therapists: Results From the Optimal Screening for Prediction of Referral and Outcome (OSPR O) Cohort. The Journal of Orthopaedic and Sports Physical Therapy 46 (5), 327 343. https://doi.org/10.2519/jospt.2016.6487 Lentz, T. A., Harman, J. S., Marlow, N. M., & George, S. Z. (2017). Application of a Value Model for the Prevention and Management of Chronic Musculoskeletal Pain by Physical Therapists. Physical Therapy 97(3), 354 364. https://doi.org/10.2522/ptj.20160167 Luo, X., Pietrobon, R., Sun, S. X., Liu, G. G., & Hey, L. (2004). Estimates and patterns of direct health care expenditures among i ndividuals with back pain in the United States. Spine, 29(1), 79 86. https://doi.org/10.1097/01.BRS.0000105527.13866.0F Machlin, S., Yu, W., & Zodet, M. (2005). Computing Standard Errors for MEPS Estimates Rockville, MD: Agency for Healthcare Research and Quality. Retrieved from https://meps.ahrq.gov/survey_comp/standard_errors.jsp Maeng, D. D., Stewart, W. F., Yan, X., Boscarino, J. A., Mardekian, J., Harnett, J., & Von Korff, M. R. (2015). Use of electronic health records for early detection of high cost low back pain patients. Pain Research & Management 20 (5), 234 240. Martin, B. I., Deyo, R. A., Mirza, S. K., Turner, J. A., Comstock, B. A., Hollingworth, W., & Sullivan, S. D. (2008). Expenditures and health status among adults with back and neck probl ems. JAMA, 299(6), 656 664. https://doi.org/10.1001/jama.299.6.656 Meghani, S. H., Buck, H. G., Dickson, V. V., Hammer, M. J., Rabelo Silva, E. R., Clark, R., & Naylor, M. D. (2013). The conceptualization and measurement of comorbidity: a review of the int erprofessional discourse. Nursing Research and Practice 2013, 192782. https://doi.org/10.1155/2013/192782 Meraya, A. M., Raval, A. D., & Sambamoorthi, U. (2015). Chronic condition combinations and health care expenditures and out of pocket spending burden among adults, Medical Expenditure Panel Survey, 2009 and 2011. Preventing Chronic Disease, 12, E12. https://doi.org/10.5888/pcd12.140388 Merkesdal, S., & Mau, W. (2005). Prediction of costs of illness in patients with low back pain undergoing orthopedic o utpatient rehabilitation. International Journal of Rehabilitation Research. Internationale Zeitschrift Fur Rehabilitationsforschung. Revue Internationale De Recherches De Readaptation, 28(2), 119 126. 187

PAGE 188

Mitchell, E. M. (2001). Concentration of Health Expenditures in the U.S. Civilian Noninstitutionalized Population, 2014. In Medical Expenditure Panel Survey Statistical Briefs Rockville (MD): Agency for Healthcare Research and Quality (US). Retrieved from http://www.ncbi.nlm.nih.gov/books/NBK425792/ Modarai, F., Mack, K., Hicks, P., Benoit, S., Park, S., Jones, C., Paulozzi, L. (2013). Relationship of opioid prescription sales and overdoses, North Carolina. Drug and Alcohol Dependence, 132(1 2), 81 86. https://doi.org/10.1016/j.drugalcdep.2013.01.006 Mooi, E., & Sarstedt, M. (2011). A Concise Guide to Market Research: The Process, Data, and Methods Using IBM SPSS Statistics Springer Science & Business Media. Morris, B. J., & Mir, H. R. (2015). The opioid epidemic: impact on orthopaedic surgery. The Journal of the American Academy of Orthopaedic Surgeons 23 (5), 267 271. https://doi.org/10.5435/JAAOS D 14 00163 Moynihan, R., Doust, J., & Henry, D. (2012). Preventing overdiagnosis: how to stop harming the healthy. BMJ (Clinical Research Ed.) 344, e3502. Mller Schwefe, G., Freytag, A., Her, A., Schiffhorst, G., Becker, A., Casser, H.R., Treede, R. D. (2011). Healthcare utilization of back pain patients: results of a claims data analysis. Journal of Medical Economics 14(6), 816 823. https://doi.org/10 .3111/13696998.2011.625067 Nagi, S. Z. (1976). An epidemiology of disability among adults in the United States. The Milbank Memorial Fund Quarterly. Health and Society 54(4), 439 467. Nahin, R. L. (2015). Estimates of pain prevalence and severity in adult s: United States, 2012. The Journal of Pain: Official Journal of the American Pain Society 16(8), 769 780. https://doi.org/10.1016/j.jpain.2015.05.002 Newton, W. P., & Lefebvre, A. (2015). Is a strategy focused on super utilizers equal to the task of heal th care system transformation? No. Annals of Family Medicine, 13(1), 8 9. https://doi.org/10.1370/afm.1747 Nicholas, M. K., Linton, S. J., Watson, P. J., Main, C. J., & Decade of the Flags Working Group. (2011). Early identification and management of psy chological risk factors (yellow flags) in patients with low back pain: a reappraisal. Physical Therapy 91 (5), 737 753. https://doi.org/10.2522/ptj.20100224 Nylund, K. L., Asparouhov, T., & Muthn, B. O. (2007). Deciding on the Number of Classes in Laten t Class Analysis and Growth Mixture Modeling: A Monte Carlo Simulation Study. Structural Equation Modeling: A Multidisciplinary Journal 14(4), 535 569. https://doi.org/10.1080/10705510701575396 188

PAGE 189

Nyman, J. A. (2004). Is moral hazard inefficient? The polic y implications of a new theory. Health Affairs (Project Hope) 23 (5), 194 199. https://doi.org/10.1377/hlthaff.23.5.194 Oberski, D. L., Kollenburg, G. H., & Vermunt, J. K. (2013). A Monte Carlo Evaluation of Three Methods to Detect Local Dependence in Bina ry Data Latent Class Models. Adv. Data Anal. Classif. 7 (3), 267 279. https://doi.org/10.1007/s11634013 01462 Ong, A. D., & Dulmen, M. H. M. V. (2006). Oxford Handbook of Methods in Positive Psychology Oxford University Press. Otani, K., & Baden, W. W. (2009). Healthcare cost and predictive factors: highand low utilization model development. Health Marketing Quarterly 26 (3), 198 208. https://doi.org/10.1080/07359680903263599 Parekh, A. K., Goodman, R. A., Gordon, C., Koh, H. K., & HHS In teragency Workgroup on Multiple Chronic Conditions. (2011). Managing multiple chronic conditions: a strategic framework for improving health outcomes and quality of life. Public Health Reports (Washington, D.C.: 1974) 126 (4), 460 471. https://doi.org/10.1 177/003335491112600403 Peduzzi, P., Concato, J., Kemper, E., Holford, T. R., & Feinstein, A. R. (1996). A simulation study of the number of events per variable in logistic regression analysis. Journal of Clinical Epidemiology 49(12), 1373 1379. Peter, W. F., Dekker, J., Tilbury, C., Tordoir, R. L., Verdegaal, S. H. M., Onstenk, R., Vliet Vlieland, T. P. M. (2015). The association between comorbidities and pain, physical function and quality of life following hip and knee arthroplasty. Rheumatology International 35 (7), 1233 1241. https://doi.org/10.1007/s0029601532117 Porter, M. E. (2010). What is value in health care? The New England Journal of Medicine, 363(26), 2477 2481. https://doi.org/10.1056/NEJMp1011024 Prochaska, J. J., Sung, H. Y., Max, W., S hi, Y., & Ong, M. (2012). Validity study of the K6 scale as a measure of moderate mental distress based on mental health treatment need and utilization. International Journal of Methods in Psychiatric Research, 21 (2), 88 97. https://doi.org/10.1002/mpr.1349 Quan, H., Sundararajan, V., Halfon, P., Fong, A., Burnand, B., Luthi, J. C., Ghali, W. A. (2005). Coding algorithms for defining comorbidities in ICD 9 CM and ICD 10 administrative data. Medical Care, 43(11), 1130 1139. Riley, G. F. (2007). Longterm t rends in the concentration of Medicare spending. Health Affairs (Project Hope) 26(3), 808 816. https://doi.org/10.1377/hlthaff.26.3.808 189

PAGE 190

Ritzwoller, D. P., Crounse, L., Shetterly, S., & Rublee, D. (2006). The association of comorbidities, utilization and c osts for patients identified with low back pain. BMC Musculoskeletal Disorders 7 72. https://doi.org/10.1186/147124747 72 Rosella, L. C., Fitzpatrick, T., Wodchis, W. P., Calzavara, A., Manson, H., & Goel, V. (2014). Highcost health care users in Ontario, Canada: demographic, socioeconomic, and health status characteristics. BMC Health Services Research 14, 532. https://doi.org/10.1186/s12913 014 0532 2 Rudd, R. A., Seth, P., David, F., & Scholl, L. (2016). Increases in Drug and OpioidInvolved Overdose Deaths United States, 2010 2015. MMWR. Morbidity and Mortality Weekly Report 65. https://doi.org/10.15585/mmwr.mm655051e1 Rudolf J. Freud, & Ramon C. Littell. (2000). SAS System for Regression: Third Edition. SAS Institute. Rundell, S. D., Goode, A. P., Friedly, J. L., Jarvik, J. G., Sullivan, S. D., & Bresnahan, B. W. (2015). Role of Health Services Research in Producing High Value Rehabilitation Care. Physical Therapy 95 (12), 1703 1711. https://doi.org/10.2522/ptj.20150033 Rundell, S. D., Goode, A. P., Suri, P., Heagerty, P. J., Comstock, B. A., Friedly, J. L., Jarvik, J. G. (2017). Effect of Comorbid Knee and Hip Osteoarthritis on Longitudinal Clinical and Health Care Use Outcomes in Older Adults With New Visits for Back Pain. Archives of Physical Medicine and Rehabilitation, 98(1), 43 50. https://doi.org/10.1016/j.apmr.2016.06.022 Rundell, S. D., Sherman, K. J., Heagerty, P. J., Mock, C. N., Dettori, N. J., Comstock, B. A., Jarvik, J. G. (2016). Predictors of Persistent Disability and Back Pain in Older Adults with a New Episode of Care for Back Pain. Pain Medicine (Malden, Mass.) https://doi.org/10.1093/pm/pnw236 Saragiotto, B. T., Maher, C. G., Hancock, M. J., & Koes, B. W. (2017). Subgrouping Patients With Nonspecif ic Low Back Pain: Hope or Hype? The Journal of Orthopaedic and Sports Physical Therapy 47 (2), 44 48. https://doi.org/10.2519/jospt.2017.0602 Shi, L. (2012). The impact of primary care: a focused review. Scientifica 2012, 432892. https://doi.org/10.6064/2012/432892 Smith, M., Davis, M. A., Stano, M., & Whedon, J. M. (2013). Aging baby boomers and the rising cost of chronic back pain: secular trend analysis of longitudinal Medical Expenditures Panel Survey data for years 2000 to 2007. Journal of Manipulativ e and Physiological Therapeutics 36(1), 2 11. https://doi.org/10.1016/j.jmpt.2012.12.001 190

PAGE 191

Soley Bori, M., Soria Saucedo, R., Ryan, C. M., Schneider, J. C., Haynes, A. B., Gerrard, P., Kazis, L. E. (2015). Functional Status and Hospital Readmissions Usin g the Medical Expenditure Panel Survey. Journal of General Internal Medicine, 30(7), 965 972. https://doi.org/10.1007/s11606014 31709 Stanton, M., & Rutherford, M. (2005). The high concentration of U.S. health care expenditures. (Research in Action Issue 19 No. AHRQ Pub No. 060060). Rockville (MD): Agency for Healthcare Research and Quality. Sterne, J. A. C., White, I. R., Carlin, J. B., Spratt, M., Royston, P., Kenward, M. G., Carpenter, J. R. (2009). Multiple imputation for missing data in epidemiolo gical and clinical research: potential and pitfalls. BMJ (Clinical Research Ed.) 338, b2393. Stewart, W. F., Yan, X., Boscarino, J. A., Maeng, D. D., Mardekian, J., Sanchez, R. J., & Von Korff, M. R. (2015). Patterns of health care utilization for low bac k pain. Journal of Pain Research, 8 523 535. https://doi.org/10.2147/JPR.S83599 Stockbridge, E. L., Suzuki, S., & Pagn, J. A. (2015). Chronic pain and health care spending: an analysis of longitudinal data from the Medical Expenditure Panel Survey. Healt h Services Research, 50 (3), 847 870. https://doi.org/10.1111/1475 6773.12263 Tabachnick, B. G., & Fidell, L. S. (2006). Using Multivariate Statistics (5 edition). Boston: Pearson. Tang, N., Stein, J., Hsia, R. Y., Maselli, J. H., & Gonzales, R. (2010). Tre nds and characteristics of US emergency department visits, 19972007. JAMA, 304 (6), 664 670. https://doi.org/10.1001/jama.2010.1112 Thomas, E., Peat, G., Harris, L., Wilkie, R., & Croft, P. R. (2004). The prevalence of pain and pain interference in a gener al population of older adults: cross sectional findings from the North Staffordshire Osteoarthritis Project (NorStOP). Pain 110(1 2), 361 368. https://doi.org/10.1016/j.pain.2004.04.017 Toblin, R. L., Mack, K. A., Perveen, G., & Paulozzi, L. J. (2011). A populationbased survey of chronic pain and its treatment with prescription drugs. Pain 152(6), 1249 1255. https://doi.org/10.1016/j.pain.2010.12.036 Trescot, A. M., Helm, S., Hansen, H., Benyamin, R., Glaser, S. E., Adlaka, R., Manchikanti, L. (2008). Opioids in the management of chronic noncancer pain: an update of American Society of the Interventional Pain Physicians (ASIPP) Guidelines. Pain Physician, 11(2 Suppl), S5 S62. Turk, D. C., & Gatchel, R. J. (2002). Psychological Approaches to Pain Management, Second Edition: A Practitioners Handbook (2 edition). New York: The Guilford Press. 191

PAGE 192

Turk, D., & Monarch, E. (2002). Biopsychosocial perspective on chronic pain. In Psychological Approaches to Pain Management: A Practitioners Handbook New York: Guilford. United States Bone and Joint Initiative. (2014). The Burden of Musculoskeletal Diseases in the United States (BMUS) (No. Third Edition). Rosemont, IL. Retrieved from http://www.boneandjointburden.org/ Valderas, J. M., Sta rfield, B., Sibbald, B., Salisbury, C., & Roland, M. (2009). Defining comorbidity: implications for understanding health and health services. Annals of Family Medicine, 7 (4), 357 363. https://doi.org/10.1370/afm.983 van Dijk, G. M., Veenhof, C., Schellevis F., Hulsmans, H., Bakker, J. P., Arwert, H., Dekker, J. (2008). Comorbidity, limitations in activities and pain in patients with osteoarthritis of the hip or knee. BMC Musculoskeletal Disorders 9 95. https://doi.org/10.1186/1471 24749 95 Vermunt, J. K., & Magidson, J. (2003). Latent Class Models for Classification. Comput. Stat. Data Anal. 41 (3 4), 531 537. https://doi.org/10.1016/S01679473(02)001792 Vermunt, J. K., & Magidson, J. (2004). Latent class analysis. In M. Lewis Beck, A. Bryman, & T. Li ao (Eds.), The Sage Encyclopedia of Social Sciences Research Methods (pp. 549 553). Thousand Oaks, CA: Sage Publications. Vermunt, J. K., & Magidson, J. (2005). Latent GOLD 4.0 Users Guide. Belmont, Massachusetts: Statistical Innovations Inc. Vernon, H. (2008). The Neck Disability Index: stateof the art, 1991 2008. Journal of Manipulative and Physiological Therapeutics 31(7), 491 502. https://doi.org/10.1016/j.jmpt.2008.08.006 Vernon, H., & Mior, S. (1991). The Neck Disability Index: a study of reliability and validity. Journal of Manipulative and Physiological Therapeutics 14(7), 409 415. Vogeli, C., Shields, A. E., Lee, T. A., Gibson, T. B., Marder, W. D., Weiss, K. B., & Blumenthal, D. (2007). Multiple chronic conditions: prevalence, healt h consequences, and implications for quality, care management, and costs. Journal of General Internal Medicine, 22 Suppl 3, 391 395. https://doi.org/10.1007/s11606007 03221 Von Korff, M., Scher, A. I., Helmick, C., Carter Pokras, O., Dodick, D. W., Goule t, J., Mackey, S. (2016). United States National Pain Strategy for Population Research: Concepts, Definitions, and Pilot Data. The Journal of Pain: Official Journal of the American Pain Society 17(10), 1068 1080. https://doi.org/10.1016/j.jpain.2016.06. 009 192

PAGE 193

Ward, B. W., Schiller, J. S., & Goodman, R. A. (2014). Multiple chronic conditions among US adults: a 2012 update. Preventing Chronic Disease, 11, E62. https://doi.org/10.5888/pcd11.130389 Ward, M. M. (2002). Functional disability predicts total costs in patients with ankylosing spondylitis. Arthritis and Rheumatism 46 (1), 223 231. https://doi.org/10.1002/1529 0131(200201)46:1<223::AID ART498>3.0.CO;2 # Ware, J. E., Kosinski, M., Dewey, J. E., & Gandek, B. (2001). How to score and interpret singleitem health status measures: a manual for users of the SF 8TM Health Survey. Lincoln (RI): Quality Metric Incoporated. Ware, J., Kosinski, M., & Keller, S. D. (1996). A 12 Item Short Form Health Survey: construction of scales and preliminary tests of reliabili ty and validity. Medical Care 34 (3), 220 233. Weuve, J., Tchetgen Tchetgen, E. J., Glymour, M. M., Beck, T. L., Aggarwal, N. T., Wilson, R. S., Mendes de Leon, C. F. (2012). Accounting for bias due to selective attrition: the example of smoking and cogn itive decline. Epidemiology (Cambridge, Mass.) 23 (1), 119 128. https://doi.org/10.1097/EDE.0b013e318230e861 Whitson, H. E., Johnson, K. S., Sloane, R., Cigolle, C. T., Pieper, C. F., Landerman, L., & Hastings, S. N. (2016). Identifying Patterns of Multimorbidity in Older Americans: Application of Latent Class Analysis. Journal of the American Geriatrics Society 64(8), 1668 1673. https://doi.org/10.1111/jgs.14201 Wideman, T. H., Hill, J. C., Main, C. J., Lewis, M., Sullivan, M. J. L., & Hay, E. M. (2012). Comparing the responsiveness of a brief, multidimensional risk screening tool for back pain to its unidimensional reference standards: the whole is greater than the sum of its parts. Pain 153(11), 2182 2191. https://doi.org/10.1016/j.pain.2012.06.010 Wide man, T. H., Scott, W., Martel, M. O., & Sullivan, M. J. L. (2012). Recovery from depressive symptoms over the course of physical therapy: a prospective cohort study of individuals with work related orthopaedic injuries and symptoms of depression. The Journal of Orthopaedic and Sports Physical Therapy 42 (11), 957 967. https://doi.org/10.2519/jospt.2012.4182 Wideman, T. H., & Sullivan, M. J. L. (2011). Differential predictors of the long term levels of pain intensity, work disability, healthcare use, and medication use in a sample of workers compensation claimants. Pain 152(2), 376 383. https://doi.org/10.1016/j.pain.2010.10.044 Wodchis, W. P., Austin, P. C., & Henry, D. A. (2016). A 3 year study of high cost users of health care. CMAJ: Canadian Medical Ass ociation Journal = Journal de lAssociation Medicale Canadienne, 188(3), 182 188. https://doi.org/10.1503/cmaj.150064 193

PAGE 194

Wood, A. M., White, I. R., & Royston, P. (2008). How should variable selection be performed with multiply imputed data? Statistics in Medi cine 27(17), 3227 3246. https://doi.org/10.1002/sim.3177 Woolf, A. D., & Pfleger, B. (2003). Burden of major musculoskeletal conditions. Bulletin of the World Health Organization, 81(9), 646 656. Woolf, C. J. (2010). What is this thing called pain? The Jo urnal of Clinical Investigation, 120(11), 3742 3744. https://doi.org/10.1172/JCI45178 Zhang, J. X., Iwashyna, T. J., & Christakis, N. A. (1999). The performance of different lookback periods and sources of information for Charlson comorbidity adjustment in Medicare claims. Medical Care, 37(11), 1128 1139. Zhang, Y., & Jordan, J. M. (2010). Epidemiology of osteoarthritis. Clinics in Geriatric Medicine, 26(3), 355 369. https://doi.org/10.1016/j.cger.2010.03.001 194

PAGE 195

BIOGRAPHICAL SKETCH Trevor A. Lentz, PT, MPH, SCS received a Bachelor of Science (BS) in Exercise Science degree in 2004 and Master of Physical Therapy (MPT) degree in 2006 from the University of Florida. After completing a sports residency he spent 7 yea rs in clinical practice focused on post operative sports and orthopedics rehabilitation. In 2016, Mr. Lentz received his Master of Public Health with a concentration in health management and policy, followed by a PhD in Rehabilitation Science in 2017, both from the University of Florida. He has published on topics ranging from psychosocial factors following ACL reconstruction to modeling valuebased healthcare delivery in physical therapy. Mr. Lentz has been recognized for scholarly achievement with awards from the Foundation for Physical Therapy, Sports Physical Therapy Section of the American Physical Therapy Association, American Pain Society and University of Florida College of Public Health and Health Professions. 195