1 IMPACT OF MEDICAID/SCHIP DISENROLLMENT ON HEALTH CARE USE AND EXPENDITURES AMONG CHILDREN IN THE UNITED STATES By JINGBO YU A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2008
2 2008 Jingbo Yu
3 To my family
4 ACKNOWLEDGMENTS I am extremely fortunate to have obtained help and support from many people during my doctoral study. I would like to express my gratitude to my supervisory committee chair, Dr. R. Paul Duncan, for his invaluable advice and supp ort all the time. His belief in my abilities fortified my resolve to do my be st in my study. I give my sincere thanks to Dr. Jeffery Harman. I thank him for his encouragemen t and instruction in research methodology development and for always being there to help me. I would like to pay my sincere thanks to Dr. Alyson Hall, who has been a constant source of invaluable advice a nd encouragement from the very beginning of my study and work here. I would like to thank Dr. Babette Brumback whose sharp questions on my methodology and advice help me improve my st udy. I am indebted to Zhou, my professor and dear friend. I am very grateful for ha ving her around for valuable insights. My sincere thanks go to the Florida Center for Medicaid and the Uninsured for providing me research assistant opportunity and supportin g my study financially. Thanks go to my dear colleagues and friends there: Lorn a, Heather, Jianyi, Michele, A ndrea, and Zoe, for their love and support to me. I feel lucky working together with you folks. Thanks go to Lorna for always being a patient lis tener and providing great advice. I would like to mention a special thank you to my dear friends, Kezia and Valerie. I thank Kezia for helping me go through those stressful moments and offering valuable insights. I can never thank Valerie enough for her efforts, patience and help to me ever since I came to this country. I am very fortunate to have them as my friends. I would like to thank my friends and classmates, Jesse, Alex, Swathy, Cameron, Erik, and a long list of others. I appreciate your friendly help and company in the process of get ting our PhDs. Best wishes to everyone of you in your journey to success.
5 I would like thank my parents and my sister. I th ank them for their love to me ever. I thank them for being my cheer leaders all the time. I thank my parents for letting me go across the ocean to pursue my dream. I thank my sister a nd my brother-in-law for taking good care of our parents. I thank my little nephew for bringing all the laughter to the family and keeping grandma from worrying about her far-away daughter.
6 TABLE OF CONTENTS page ACKNOWLEDGMENTS...............................................................................................................4 LIST OF TABLES................................................................................................................. ..........8 LIST OF FIGURES................................................................................................................ .......10 LIST OF ABBREVIATIONS........................................................................................................11 ABSTRACT....................................................................................................................... ............12 CHAPTER 1 INTRODUCTION..................................................................................................................14 Overview....................................................................................................................... ..........14 Study Objectives............................................................................................................... ......15 2 LITERATURE REVIEW.......................................................................................................16 Health Insurance Coverage of Children in the United States.................................................16 Importance of Continuous Coverage......................................................................................20 Retention/Drop-out in Medicaid/SCHIP................................................................................23 Behavioral Model of Health Services Use..............................................................................25 Existing Evidence on Consequence of Medicaid/SCHIP Disenrollment...............................27 Medicaid versus Uninsured.............................................................................................27 Medicaid/SCHIP versus Private Insurance.....................................................................33 Gaps in Literature............................................................................................................37 Features of this Study......................................................................................................... ....38 3 METHODOLOGY.................................................................................................................41 Research Questions and Hypotheses......................................................................................41 Data Source.................................................................................................................... .........42 MEPS-HC Data...............................................................................................................42 Sampling design.......................................................................................................42 Procedures for data collection..................................................................................43 Scope of MEPS-HC.................................................................................................43 Sampling weight specification.................................................................................44 Rationale of Using MEPS-HC Data for the Study..........................................................45 Why Use Round as Observation Period Unit?................................................................45 Sample Inclusion Criteria for this Study.........................................................................46 Description of Outcome Measures and Independent Variables.............................................46 Outcome Measures..........................................................................................................46 Expenditure measures..............................................................................................47
7 Utilization measures.................................................................................................48 Major Independent Variables..........................................................................................49 Control Variables.............................................................................................................50 Study Design and Statistical Analysis....................................................................................53 Descriptive Analysis........................................................................................................54 Multivariate Analysis......................................................................................................54 Difference-in-difference analysis.............................................................................55 Three-part DiD model..............................................................................................57 Sensitivity analysis...................................................................................................64 4 RESULTS........................................................................................................................ .......72 Overview....................................................................................................................... ..........72 Description of the Sample......................................................................................................72 Medicaid/SCHIP Disenrollment and Change in Total Expenditures.....................................75 Medicaid/SCHIP Disenrollment and Change in Well-child Care Use...................................80 Medicaid/SCHIP Disenrollment and Change in Physician Visits..........................................83 Medicaid/SCHIP Disenrollment and Change in ER Visits....................................................85 Medicaid/SCHIP Disenrollment and Cha nge in Number of Hospitalizations.......................87 Medicaid/SCHIP Disenrollment and Change in Prescription Drug Use................................89 5 DISCUSSION..................................................................................................................... ..134 Overview....................................................................................................................... ........134 Summary and Interpretation of Findings..............................................................................134 Who Disenrolled from Medicaid/SCHIP?.....................................................................134 What are the Consequences of Dropout?......................................................................135 Expenditures...........................................................................................................135 Preventive care use.................................................................................................138 Ambulatory care use...............................................................................................139 Emergency care use................................................................................................140 Inpatient care..........................................................................................................141 Prescription drug use..............................................................................................141 Policy Implication.........................................................................................................142 Limitations of the Study.......................................................................................................145 Conclusion..................................................................................................................... .......146 LIST OF REFERENCES.............................................................................................................147 BIOGRAPHICAL SKETCH.......................................................................................................156
8 LIST OF TABLES Table page 3-1 Summary of outcome measures.........................................................................................69 3-2 Sources of explanatory variab les (point-in-time measures)...............................................70 3-3 Description of covariate vari ables in multivariate analysis...............................................71 4-1 Study sample by insurance group......................................................................................92 4-2 Individual and family characterist ics of study sample by insurance group.......................93 4-3 Type of change in total expenditures by insurance group.................................................95 4-4 Descriptive statistics on amount of total expenditures by insurance group.......................95 4-5 Multinomial logit regression predicting pr obability of experiencing each type of change in total expenditures...............................................................................................96 4-6 Ordinary least square (OLS) regressi on predicting amount of change in total expenditures given negative or positive change................................................................98 4-7 Generalized linear regression (GLM) predicting amount of change in total expenditures given negative or positive change..............................................................100 4-8 Bootstrapping prediction of expenditure change afte r three-part model (OLS)..............102 4-9 Bootstrapping prediction of expenditure change afte r three-part model (GLM).............102 4-10 Type of change in number of well-child visits by insurance group.................................103 4-11 Descriptive statistics on nu mber of well-child visits.......................................................103 4-12 Multinomial logit regression predicting proba bility of each type of change in wellchild visits................................................................................................................... .....104 4-13 GLM regression predicting amount of change in well-child visits given negative or positive change................................................................................................................ .106 4-14 Bootstrapping prediction of change in well-child visits after 3-part model....................107 4-15 Type of change in number of physician visits by insurance type....................................108 4-16 Descriptive statistics on number of physician visits........................................................108 4-17 Multinomial logit regressi on predicting probability of experiencing each type of change in physician visits................................................................................................109
9 4-18 GLM regression predicting amount of change in number of phys ician visits given negative or positive change..............................................................................................111 4-19 Bootstrapping prediction of change in physician visits after 3-part model.....................112 4-20 Type of change in number of ER visits by insurance group............................................113 4-21 Descriptive statistics on number of ER visits..................................................................113 4-22 Multinomial logit regressi on predicting probability of experiencing each type of change in ER visits...........................................................................................................114 4-23 GLM regression predicting amount of change in number of ER visits given negative or positive change............................................................................................................116 4-24 Bootstrapping prediction of change in number of ER visits after 3-part model..............117 4-25 Type of change in number of hospitalizations by insurance group.................................118 4-26 Descriptive statistics on number of hospitalizations........................................................118 4-27 Multinomial logit regressi on predicting probability of experiencing each type of change in number of hospitalizations..............................................................................119 4-28 GLM regression predicting amount of cha nge in hospitalizations given negative or positive change................................................................................................................ .121 4-29 Bootstrapping prediction of change in number of hospitalization after 3-part model.....122 4-30 Type of change in prescription drug use by insurance group..........................................123 4-32 Multinomial logit regression predicting proba bility of having each type of change in prescription drug use........................................................................................................124 4-33 GLM regressions predicting amount of change in prescription drug use given negative or positive change..............................................................................................126 4-34 Bootstrapping pred iction of change in prescription drug use..........................................127
10 LIST OF FIGURES Figure page 2-1 Andersen Behavioral Model of Health Services Use (1995).............................................40 3-1 Panel design of MEPS (2003-2004)..................................................................................66 3-2 Sample inclusion criteria.................................................................................................. ..66 3-3 Measuring change of insurance status...............................................................................67 3-4 Study design............................................................................................................... ........67 3-5 Models for predicting amount of utiliz ation/expenditure change under the DiD specification.................................................................................................................. .....68 4-1 Kernel density plot afte r OLS regression on absolute amount of negative change in total expenditures.............................................................................................................128 4-2 P-P plot after OLS regression on absolu te amount of negative change in total expenditures................................................................................................................... ..128 4-3 Q-Q plot after OLS regression on absolu te amount of negative change in total expenditures................................................................................................................... ..129 4-4 Residual-fitted plot after OLS regressi on on absolute amount of negative change in total expenditures.............................................................................................................129 4-5 Kernel density plot after OLS regres sion on amount of positive change in total expenditures................................................................................................................... ..130 4-6 P-P plot after OLS regression on amount of positive change in total expenditures........130 4-8 Residual-fitted plot after OLS regres sion on amount of positive change in total expenditures................................................................................................................... ..131 4-9 Residual-fitted plot afte r GLM regression on absolute am ount of negative change in total expenditures.............................................................................................................132 4-10 Residual-fitted plot afte r GLM regression for amount of positive change in total expenditures................................................................................................................... ..132 4-11 Residual-fitted plot afte r GLM regression on absolute am ount of negative change in number of well-child visits..............................................................................................133 4-12 Residual-fitted plot after GLM regression on amount of pos itive change in number of well-child visits.............................................................................................................. ..133
11 LIST OF ABBREVIATIONS AAP American Academy of Pediatrics ACSC Ambulatory care sensitive conditions AHRQ Agency for Healthcare Research and Quality AIC Akaikes information criterion CI Confidence interval DiD Difference in difference ER Emergency room FPL Federal poverty level GLM Generalized linear model H-L Hosmer-Lemeshow MEPS Medical Expenditure Panel Survey MM Stay in Medicaid/SCHIP during the study time periods MP Disenroll from Medicaid/SCHIP a nd transition to private insurance MU Disenroll from Medicaid/SCHIP and become uninsured NHIS National Health Interview Survey OLS Ordinary least square OR Odds ratio RRR Relative risk ratio SCHIP State Childrens H ealth Insurance Program
12 Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy IMPACT OF MEDICAID/SCHIP DISENROLLMENT ON HEALTH CARE USE AND EXPENDITURES AMONG CHILDREN IN THE UNITED STATES By Jingbo Yu May 2008 Chair: R. Paul Duncan Major: Health Services Research Most children who disenroll from Medicaid/S CHIP lose coverage and become uninsured. Although it is widely believed that disenrolling from Medicaid/SCHIP will influence health care utilization for children, available data confirmi ng and quantifying this impact are limited. This study examines the impact of Medicaid/SCH IP disenrollment on health care use and expenditures among children in the United States. With a retrospective quasi-experimental study design, data from 1996 Medical Expenditure Panel Survey (MEPS) were used to examine utilization and expenditures preand post-disenrollment. I measured how and how much health care expenditures and utilization changed after Medicaid/SCHIP disenrollment for two groups, children who became uninsured (MU) and those who transitioned to private insurance (MP), relative to a control group, those who stayed in Medicaid/SCHIP (MM). In multiva riate analysis, a modification of the two-part model was employed within a general differencein-difference (DiD) spec ification. The outcome measures were total expenditures and five types of health care utilization: well-child visits, physician visits, emergency room (ER) visits hospitalizations, and prescription drugs. The dependent variables were the changes in outcome measures over two rounds (~5 months/round). The first equation of the modified two-part model was a multinomial logit regression modeling
13 the probability of having negative, zero, and pos itive changes in expenditures and utilization over time. The second and third equations were li near regressions (OLS or GLM) modeling the amount of decrease or in crease in expenditures and utilization given decreas e or increase in the corresponding outcome measure. The childs and pa rents socio-demographi c characteristics and childs health condition change we re controlled in the analysis. The 95% confidence intervals of predictions were obt ained by bootstrapping. Compared with children who stayed in Medicaid/SCHIP, those children who lost Medicaid/SCHIP coverage and became uninsured experienced a decrease in expenditures, wellchild visits, physician visits, and prescription drug use, and no si gnificant change in ER visits and hospitalizations, controlling for other factors. Compared w ith children who stayed in Medicaid/SCHIP, those children who transitione d from Medicaid/SCHIP to private insurance experienced a decrease in well-child visits, a nd no significant change observed in physician visits, ER visits, prescription drug use, hospitalizations, and total expenditures. This study indicates that losing Medicaid/SCHI P coverage resulted in decreased preventive care utilization among children. In addition, reducti ons in physician visits and prescription drug use were observed among those children who became uninsured.
14 CHAPTER 1 INTRODUCTION Overview The health and welfare of children has been a longstanding social concern for policymakers in the United States. Health insuran ce coverage, public or priv ate, is an important element in enabling access to care and utilization of health se rvices for children. Two major public insurance programs, Medicaid and the State Childrens Health Insurance Program (SCHIP), play significant roles in providing co verage and improving access to care for lowincome children. Stable insuranc e coverage, which is an importa nt determinant of continuous care, is particularly critical for children during key periods of growth and development. However, the dropout rates of Medicaid and SC HIP have been high. While some children who disenroll from Medicaid/SCHIP obtain private insurance, most lose coverage and become uninsured (Sommers 2005a). How does the Medicaid /SCHIP disenrollment influence the use and expenditures of care for those children? Although it is widely believed that disenr olling from Medicaid/SCHIP will influence health care utilization for children, available da ta confirming and quantifying this impact are limited. In this study, these issues were pursued by comparing th e change in utilization and expenditures for care over time for children w ho disenrolled and those who stayed in Medicaid/SCHIP. Specifically, the research ques tion was examined using a quasi-experimental study design (before-and-after de sign with an untreated comparison group) and DiD analysis approach, with a nationally representative sample and longitudinal data. Changes in utilization and expenditures after disenrollment from Me dicaid/SCHIP for two groupsthose disenrolled from Medicaid/SCHIP to become uninsured or tr ansitioned to private insurancerelative to a control group, those who stayed in Medicaid/SCH IP, were measured. In multivariate regression
15 analysis, a three-part model was employed w ithin the general DiD specification. The study provides scientific evidence on the consequences of disenrollment from Medicaid/SCHIP on childrens health care utilization, and informs policymaker s of the importance of keeping Medicaid/SCHIP children enrolled stably. Study Objectives The main purpose of this study was to examine the impact of Medicaid/SCHIP disenrollment on health care utilization and expenditures among children in the United States. The specific objectives of the study were Objective 1 : To quantify the effect of two type s of Medicaid/SCHIP disenrollment on health care use and expenditures among children. Sp ecifically, the two types of disenrollment of interest are disenrolling from Medicaid/SCHIP to become unin sured and transitioning from Medicaid/SCHIP to private insurance. Objective 2 : To measure the impact of Medicaid/ SCHIP disenrollment on utilization of various kinds of health care, including preventive care, outpatient care, inpatient care, emergency care, and prescription drugs. Objective 3 : To examine how and how much expend itures change when children disenroll from Medicaid/SCHIP.
16 CHAPTER 2 LITERATURE REVIEW Health Insurance Coverage of Ch ildren in the United States Among the 77.9 million children in the United States, 56% are covered by employersponsored insurance, 4% by individual private insurance, 28% by public insurance, and 12% (9 million) are uninsured. The large number of children who lack health insurance or have unstable coverage has been a source of concern for public officials and policymakers. Children are at greater risk than adults of bei ng uninsured because as a group th ey are more likely to be poor. While about a third of adults under age 65 come from low-income families (incomes less than twice the poverty level, whic h in 2005 was about $40,000 for a family of four), over 40% of children come from low-income families. Despite this higher risk of uninsurance, children are much less likely than adults in fact to be unins ured (12% vs. 21% in 2005 ) because of two public insurance programs: Medicaid and the SCHIP. Medicaid and SCHIP cover children whose family incomes are below levels set by individual states that are at or above federal minimum requirements (Schwartz, Hoffman, and Cook 2007). Medicaid, SCHIP, and other state programs for childrens coverage play an important role in providing health insurance for children. T hose publicly funded program s are the only source of coverage available to millions of low-income children who have no access to affordable jobbased insurance. Medicaid cove rs 28 million low-income children, and SCHIP covers about 6 million low-income children who do not qualify for Medicaid but fall below specified income thresholds in 2005 (KFF 2007). The programs cover about 60% of all poor children (family incomes less than the poverty level) and about 40% of near-poor children (those with family incomes between one and two times the povert y level) (Schwartz, Hoffman, and Cook 2007). Including higher income children with disabi lities who qualify, these public programs insure
17 more than a quarter of all childr en in the United States. Because of the expansion of the two big public programs, Medicaid and SCHIP, insurance c overage of children has increased since late 1990s. Background of Medicaid and SCHIP Programs for Children Title XIX of the Social Security Act, also known as Medicaid, was established in 1965 as a joint federal-state program. Medicaid provides medical assistance to certain individuals and families (including children) with low incomes and resources. The 1997 Balanced Budget Act created Title XXI of the Social Security Act, also known as SCHIP, providing states with capped federal subsidies to insure children in families with too much income to qualify for Medicaid and too little to afford private insurance. Eligibility There are two primary pathways for children to be eligible for Medicaid: (1) membership in a family with an income below specified fe deral poverty level (FPL) thresholds ($20,650 for a family of four in 2007) and (2) membership in a one-parent (and in some cases two-parent) family with dependent children with income a nd resources sufficiently low to meet the 1996 Aid to Families with Dependent Children (AFDC) st andards in the state of residence. Federal mandatory income thresholds for children vary by age. The minimum income standard for children up to (not including) age six is 133 percent of the FPL For children age six through 18, the minimum income standard is 100 percent of the FPL. States have the flexibility to modify their income standards for children age six and over (Schneider et al. 2002). SCHIP represents a large expa nsion in eligibility for public coverage. Eligibility for SCHIP is based on a mix of federal targets and state flexibility. The statute allows federal funding for health assistance for children under age 19 in families with income below 200
18 percent of the FPL. For states that had already expanded coverage to child ren as of enactment, this eligibility limit could be as much as 50 pe rcentage points higher than their Medicaid limit. Within these and other federal parameters, states may set their own elig ibility rules. By 2006, 42 states covered children with fam ily incomes of 200% of FPL, in cluding 7 states (in which the cost of living is particularly hi gh) that set income thresholds for SCHIP eligibility at 300% of FPL. It is estimated that 91% of children c overed by SCHIP come from families with incomes below 200% of FPL (Iglehart 2007; KFF 2007). Structure States have the option of pr oviding child health assistance in SCHIP through Medicaid, a separate program, or a combination of the two. States opting for Medicaid extend their existing benefits and delivery systems to children in families with higher income. Because Medicaid guarantees coverage to children w ho are eligible, children in states that use Medicaid for SCHIP would remain eligible for covera ge even if SCHIP funding runs out States opting for separate programs may adopt different bene fits packages and delivery systems for targeted, low-income children, subject to standards (described below). Children in separa te programs are not entitled to coverage; states may impose waiting lists if fe deral SCHIP funding is no longer available. In 2006, 11 states and the District of Columbia had Medicaid SCHIP programs, 18 states had separate programs, and 21 states had a combination of the two (Lambrew 2007). Benefits Generally, Medicaid and SCHIP pay for a comp rehensive set of serv ices for children, including physician and hospital vi sits, screening and treatment, well-child care, vision care, and dental services (KFF 2007b). Under Early and Periodic Sc reening, Diagnostic, and Treatment
19 (EPSDT), SCHIP must cover any medi cal service that a child is found to need as long as it is the kind of service that Medicaid covers. While SCHIP offers flexibility in benefit design in separate, non-Medicaid programs, benefits must meet certain standards. States can offer ( 1) benchmark coverage, a package substantially equivalent to the Federal Empl oyee Health Benefits Program (FEHBP)s Blue Cross/Blue Shield Standard Op tion, a state health employees pl an, or the most popular HMO in the state; (2) benchmark-equivalent coverage, meaning a plan with an aggregate actuarial value no less than a benchmark plan; (3) existing comprehensive coverage, the option in place in states that had expanded coverage prior to SCHI P; or (4) Secretary-approved coverage, which could include the Medicaid package. Certain se rvices are required in states that opt for benchmark-equivalent coverage (e.g., well-child care and certain levels of drug coverage). All alternative benefits packages must limit cost -sharing and premiums. Sp ecifically, cost-sharing for children in families with income below 150 percent of the FPL must be nominal, while costsharing (including premiums) cannot exceed more than 5 percent of income for other families. In 2003, 21 states had Secretary-approved plans (i ncluding Medicaid packages), 14 states used benchmark plans based on state employee plans, f our used the largest HMO, and four used the federal employees plan. As of July 2006, 35 stat es charged premiums or enrollment fees for childrens coverage, and 22 charged co-payme nts for services, predominantly in SCHIP programs. Financing The payments for child health assistan ce under both Medicaid and SCHIP qualify for federal matching payments. Each states enha nced federal matching rates are set at 30 percentage points above 70 per cent of their Medicaid matching rate, with an upper limit of 85
20 percent. On average, the federal government pays 70 percent and state governments pay 30 percent of program costs. This is higher than the federal matching rate for Medicaid, which averages 57 percent. Unlike Me dicaid, SCHIP matching payments are subject to annual statebased caps. The law included specific limits for aggregate annual federal program payments for FY 1998 through FY 2007, cumulating in $39 billion over this period. The annual federal limit is divided into state allotments. The allotments limit all federal matching payments for SCHIP expenditures. For FY 2007, the tota l amount available for allocati on to states is $5 billion Importance of Continuous Coverage Health insurance provides an important gatewa y to the health care system. Research shows that children with insurance coverage receive more medical care services than uninsured children (Ku, Lin, and Broaddus 2007). It is not simply being insured at a point in time, but a continuous relationship with an insurer that facilita tes finding and keeping a medical home. For children to benefit fully from health care coverage, it must be stable and ongoing. Stable insurance coverage and continuous care can be particularly critical for children during key periods of growth and development. They require a series of primary care services to ensure the healthiest childhood possible. Moreover, many childhood and adolescent problems, such as developmental delays, risk-taking behaviors, and nutritional concerns, require regular supervision for early identification and optimal treatment (AAP 2000). Continuous coverage whether public or private, conveys important protections and is superior to intermittent coverage or no c overage at all (Olson, Tang, and Newacheck 2005). Continuity of coverage is strongly associated with access to care and continuous use of care (Honberg et al. 2005). Stable insura nce coverage is likely to be an important prerequisite to having a usual source of care, and having a us ual source of care is the foundation for the pediatric concept of a medical home (AAP 2002) The American Academy of Pediatrics (AAP)
21 recommends that all children and adolescen ts have a primary care professional (or a multidisciplinary team for children with severe chronic illnesses) whose practice serves as a medical home to help ensure that needed services are accessible, family-centered, continuous, comprehensive, coordinated, compa ssionate, and culturally effective The need for an ongoing source of health careideally a medical homefor all children has been identified as a priority for children health policy reform. The US De partment of Health and Human Services Healthy People 2010 goals and objectives state that "all children with special health care needs will receive regular ongoing comprehensive care within a medical home" (U.S. Department of Health and Human Services 1999) and multiple federal programs require that all children have access to an ongoing source of health care. Research has shown that having a usual sour ce of care and a medical home are important elements in childrens access to care. A stable patient-provider relationship is expected to improve the timeliness of preventive care, compli ance with care regimens, and coordination with specialists, as well as family satisfaction with care For example, for children with chronic conditions, such as asthma or diabetes, c ontinuous coverage is critical to ensure compliance with care regimens (Diette et al. 2001; Institute of Me dicine 2001). Previous research suggests that children who have continuity with a regular pract itioner are more likely to adhere to prescribed medications; receive preventive care and well-co ordinated, resource-efficient, and familycentered care; and less likely to visit the emerge ncy department and be hospitalized; in addition, their practitioner is more likely to recogni ze their problems and track their information (Christakis et al. 2000; Christak is et al. 2001; Christakis et al. 2003; Saultz and Lochner 2005; Starfield and Shi 2004)
22 Discontinuous coverage defined as losing coverage, only to re-establish it within a few months, or gaps in coveragedoes not promote th ese favorable patterns of care (Irvin et al. 2001). Gaps in coverage represent barriers to adequate care (Aiken, Freed, and Davis 2004). Unstable insurance coverage makes it more di fficult to establish a me dical home and obtain continuous care. Gaps in cove rage can disrupt patient-provi der relationships and affect appropriate utilization for childr en. Research shows that lack of insurance or having a gap in health insurance coverage is an important determ inant for not having a regular source of care for children (Kogan et al. 1995), and lack of a usual source of care is associ ated with decreased use of preventive care and increased use of emer gency departments for nonemergency conditions (Christakis et al. 1999; Summer and Mann 2006). Research has demonstrated that the prevalence of fragmented care for children with gaps in c overage was similar to those who were uninsured for a full year. Moreover, instability and chur ningthat is, in, out, in, out of the programs can result in a subs tantial amount of wasted ti me and spending for the administrators. Medicaid and SCHIP officials, as well as health plan s and providers that se rve Medicaid and SCHIP enrollees, report significan t costs related to churning (Fairbroth er et al. 2004). More importantly, lack of insurance or gaps or insufficiency in c overage may result in inferior health outcomes for children. Recent data have shown that there are negative health consequences to the churning (Summer and Mann 2006). Medicaid and SCHIP help ensure that children have the financial coverage necessary to obtain and retain a medical home. Children in Medicaid and SCHI P are far more likely to have a usual health care source than uninsured childre n, and about as likely to have a usual source of care as privately insured children. The extent to which children covered by Medicaid or SCHIP have a medical home has improved over the past decade. Between 1993 and 2004, the
23 percentage of children covered by Medicaid or SCHI P with a usual source of health care rose to a level similar to that of privately insured chil dren. In contrast, uninsur ed childrens access to a usual source of health care has worsened. Phys icians willingness to provide charity care has dwindled in recent years (Ku, Lin, and Broa ddus 2007). Thus, it has become increasingly important for children to have continuous insura nce coverage in order to get medical care. Given the importance of stable insuran ce coverage, policymakers have focused increasingly on the problem of discontinuous coverage in Medi caid and SCHIP, the two key programs providing health insurance c overage to low-income children. Retention/Drop-out in Medicaid/SCHIP Currently, Medicaid/SCHIP has low retenti on rates. Nationally, 27.7 percent of the children enrolled in Medicaid or SCHIP were no longer enrolled 12 months later between 1998 and 2001, according to analysis of Current Popula tion Survey March Supplement data (Sommers 2005a). Dropouts varied significan tly across states. Below are fi ndings from some studies: in four states (Alabama, Colorado, Michigan, and North Carolina), 52 to 74 percent of children dropped out at the time of renewal (Hill and Lutz ky 2003); in New York, half of children lost coverage in a year despite most remaining techni cally eligible (Lipson et al. 2003); in ten states (including California) 40 percen t of SCHIP children dropped out in one year and 50 percent after 15 months (Wooldridge et al. 2005 ); and 42 to 84 percent of ch ildren were out of the SCHIP coverage after two years in another four-state (Florida, Kansas, New York, and Oregon) study (Dick et al. 2002). The change in Medicaid/SCHIP coverage for children has a churning pattern: half of children lapsed at renewal but one-fourth returned in two months in the four-state study of SCHIP retentio n (Dick et al. 2002). Low retention in health insurance programs hamp ers efforts to increase insurance coverage for children. Some of the disenrollment is explained by acquisition of private insurance.
24 However, 45.4 percent of those ch ildren dropped out despite apparently remaining eligible and having no other insurancecorresponding to 3.0 million children annually between 1998 and 2001 nationally (Sommers 2005a). Approximately six million children were eligible for Medicaid or SCHIP but remained uninsured at least part of th e year in 2004 and 2005 (Dubay et al. 2007; Hudson and Selden 2007). Even though some children returned to the programs later, churning in Medicaid and SCHIP is problematic in maintaining continuity of care for children. A series of factors contribute to dropout or failure to re-enrol l problems in Medicaid/SCHIP. Reasons for Low Retention in Medicaid/SCHIP First, the frequency and complexity of re -enrolling and eligibility determination in Medicaid are important determinants of disenrol lment. Requiring re-enrollment more than once a year, requiring a face-to-face interview (rather th an phone or mail applications), or requiring verification of citizen status makes it less likely for families to re-apply (Sommers 2005b). Second, there are problem of access to heal th care providers under Medicaid. Many providers refuse to treat Medicaid patients becau se of the much lower reimbursement rate than private insurance. If Medicaid does not provide adequate access to a broad choice of physicians to enrollees, the benefit of rema ining in the program may not be large enough for some to justify re-enrolling (Miller and Phillips 2002). Third, receipt of welfare benefits is often associated with stigma that may reduce the attractiveness of Medicaid. This stigma may be social, bure aucratic, or self-generated. Fourth, previous studies have identified some groups of child ren as more likely to leave public insurance programs than their cohorts. Th ey are adolescents, African American children, children with single parents, healthier children, those children with families paying premiums, and children living in regions with less physician access (Sommers 2006).
25 Improving insurance retention is a cost-effective, though under-appreciated, way to increase the number of insured in dividuals. If every child with pub lic or private health coverage at the beginning of a given year retained it through the next 12 months, the number of uninsured low-income children would decline by nearly 40 percent (Ku and Ross 2002). Behavioral Model of Health Services Use The behavioral model of health services us e proposed by Andersen and colleagues is the theoretical framework for this study. The Andersen model reflect s a social psychology/sociology perspective and has been widely applied to the study of health care se rvices utilization. The original model, also the most parsimonious model, suggests that a persons use of health services is a function of a set of predisposing, enabling, and need factors (Andersen 1968). Over the past decades, the behavioral model of health serv ices use has undergone several revisions and expansions along with the improve ment of knowledge in health services research. The most recent version of this model captures the dynamic and feedba ck relationships among four constructs: environment characteri stics, population characteristics, health behavior, and health outcomes (Andersen 1995). The new model will be used as the theoretical framework in hypothesizing the relationship between insurance st atus change and health services use and as the basis for selecting variables for the multivariate analysis (Figure 2-1). According to the Andersen model, the releva nt population characteri stics encompass three groups: predisposing, enabli ng, and need factors. Predisposing factors are those biological (demographic characteristics), sociological (socia l structure) and psychol ogical (health beliefs) attributes that exist before the disease. Dem ographic characteristics of children include age, gender, race, and geographic area in this study. Social structure measur es factors such as education, occupation, and social interactions, whic h influence a persons status in a society. For children, the family structure influences health ca re use in a different way compared with adults.
26 Children cannot make decisions for their own h ealth care and depend on their family, usually parents, for those decisions and financial s upport (Chen and Escarce 2006). Therefore, family structure factors (number of pare nts living with and parent edu cational attainment) are family predisposing factors to be included as control va riables in this study. Health beliefs refer to an individuals values and attitudes toward the benefit of health care and knowledge of health care information. Predisposing charac teristics are useful in understanding the degree to which an individual is prone to us e health care services. Enabling characteristics measure an individuals ab ility to realize the use of health care. Enabling resources facilitate or impede the use of services. In this study, childrens health insurance coverage, family income, parents employment status and insurance coverage will be included as enabling factors. Heal th insurance coverage, as an im portant enabler, will influence the access and use of health care by improving affordability of care (Penchansky and Thomas 1981) and reducing the financial barriers to receiving care. Need factors include self-perceived imperatives and the professional assessments about the need for health care. Both self-evaluated health status and professional dia gnosis of health conditions were used to measure health status in this study. Environmental characteristics can be summari zed by two components: external environment and health care system. External en vironment refers to t hose physical and social environmental factors that directly impact an in dividuals health behavior and health outcomes. These factors include, but are not limited to, the local economic climate, relative wealth, community politics, level of violence, prevailing social nor ms, etc. Health care system encompasses those factors within the health care system that influence the availability, accessibility, acceptability, and accommodation of health care. Examples of these factors include
27 provider to population ratio, locatio n and type of providers, and number of hospitals. In this study, one variable, time, will be included to control for the change in environmental characteristics, since there is no other e nvironmental factors available in the data. Existing Evidence on Consequence of Medicaid/SCHIP Disenrollment What are the consequences of Medicaid/SCHIP disenrollment on hea lth care utilization and expenditures for those children ? Previous studies indicate th at children losing coverage may miss the opportunity to receive needed health care (Olson, Tang, and Newacheck 2005). While there is little evidence regarding longitudinal changes in utilization and expenditures after disenrolling from Medicaid/SCHIP, some studies have compared utilization and expenditures among children of different insurance coverage and can provide relevant evidence for this question. Medicaid versus Uninsured Generally, compared to childre n with Medicaid/SCHIP covera ge, uninsured children have worse access to health care, less use of preventive care, fewer outpatient visits, more or similar ER visits, less prescription drug use, and more avoidable hospita lizations (Currie and Thomas 1995; Duderstadt et al. 2006; Lave et al. 1998; Newacheck et al. 1998; Short and Lefkowitz 1992; Stoddard, St. Peter, and Newacheck 1994c; Szilagyi et al. 2004). According to the Andersen model, health in surance coverage can improve access to care, and th erefore facilitate utilization of care, by reducing the financia l barriers to receiv ing care (Andersen, 1995). Continuous Medicaid/SCHIP coverage can prom ote more timely use of medical care for children. The type and continuity of insurance coverage could alte r patterns of use and overall expenses. Preventive care
28 The importance of preventive care for chil dren has long been recognized in federal legislation such as Title V (Maternal and Child Health Services Bloc k Grant), which provides preventive care for children with special heal th-care needs; Title XIX (Medicaid), whose EPSDT program provides preventive care for Medicaid-eligible children; and Title XXI (SCHIP), which expands access to preventive care for low-inco me children (Green and Palfrey 2002). The AAP recommends that children obtain regular preventive health care, or well-child visits. At such visits, children receive preventive health services (such as immunizations), are screened for signs of developmental or medical problems that could pose a long-term risk to their health or wellbeing, have their vision and hear ing checked, and receive health education and counseling about healthy behaviors. Well-child care plays an important role in the provision of quality h ealth care for children. Well-child visits help promote timely immunizat ions and screening for health conditions and normal development. They also offer an opportun ity for providers to an swer parents healthrelated questions and provide anticipatory guidance. Receiving the recommended number of preventive visits in early childhood may reduce em ergency department visits and hospitalizations (Hakim and Ronsaville 2002; Ha kim and Bye 2001). Measures of well-child visit have been widely used as indicators of preventive care utilization among pediatric population. Both the AAP and the Maternal and Child Health Bureau (MCHB) recommend at least 6 well-child visits in the first year, 3 in the second, and 14 from ages 2 (AAP 2000). However, less than half of the children and adolescents in the United States are meeting recommendations for well-child visit frequency (C hung et al. 2006). Well-child visits are core elements of the health services offered to ch ildren by Medicaid and SCHIP. Some studies have shown that,
29 compared with children covered by Medicaid/SCHIP, uninsured chil dren are less likely to have access to and receive appropriate prev entive care (KFF 2002; Yu et al. 2002). In addition, less utilization of preventive care are associated with these factors: African American and Hispanic children, a parent withou t college education, nonc itizen, living in the West or the South area, poor heal th status, and from a low-income family (Yu et al. 2002; Selden 2006). Ambulatory care In the case of ambulatory care, both theory and previous empirical research (e.g., the RAND Health Insurance Experiment) suggest a positive effect of insurance on utilization (Newhouse and the Insurance Expe riment Group 1994). One of the mo st widely used measures of ambulatory care is physician visits. It is es pecially important for children, since physician visits result not only in the treatment of acute illnesses but also in the greater receipt of preventive care (e.g., immunizations). Previous research has shown that compared with insured children, uninsured children are less likely to receive medical care from a physician (Stodda rd, St. Peter, and Newacheck 1994b). Aiken and colleagues found that children who tr ansitioned from public insurance to uninsured (currently) were less likely to have any physician visits th an children covered by public insurance continuously (Aiken, Freed, and Davi s 2004). It is estimated that making a child eligible for Medicaid lo wers the probability of going without a visit in past one year by 12.8 percent (Currie and Gruber 1996) Another study, which focused on children who report chronic conditions indicating a need for care, found that insurance covera ge increases those childrens physician visits by 21%% (Stoddard, St. Peter, and Newacheck 1994a).
30 The effect of insurance on ambulatory care use was shown to be more significant among white children (than African American children), and children from higher income families (than children from lower income families) (Currie and Thomas 1995; Short and Lefkowitz 1992). Emergency care The Emergency Medical Treatment and Active Labor Act mandates that all individuals presenting to ERs in the United States receive st abilizing care for their condition, regardless of insurance status or ability to pay. Among all of the age groups, ch ildren have the highest ER visit rate and account for nearly 25% of overall ER visits (Hostetler et al 2007). Research has shown that children who utilize the ER as a primary source of care have a smaller chance of receiving appropriate health supervision and treatment gui dance because ER providers generally address immediate problems and are unlikely to have the time, expertise, and access to medical records to provide pediatric preventiv e care (Johnson and Rimsza 2004). The evidence on ER use among children of different insurance coverage is mixed. Some studies reported that after chil dren enrolled in state public in surance programs, the percent of children using the ER as their pr imary source of care fell dramatically (Children's Defense Fund Texas 2006; Florida Healthy Kids Corporat ion 1997). Johnsons study of a community population found uninsured children were nearly 4 times more likely to use the ER than insured children; however, an important confounding factor in the rela tionship between insurance and ER use, health status, is not controlled in their multivariate analysis (Johnson and Rimsza 2004). Some other studies give a different picture: after adjust ing for individual characteristics (including health status) the unins ured children dont have more ER visits than Medicaid/SCHIP. In Luos study using 1997 MEPS data, Medicaid/SCHI P children were more likely to have an ER visit during the year than uninsured children (OR: 1.46), but the difference became
31 insignificant after controlling for individual characteristics (Luo et al. 2003). Similarly, in Cunninghams study using 2000 and 2003 Community Tracking Study (CTS) data, there is no significant difference in ER use between chil dren covered by Medicaid /SCHIP and uninsured children controlling for health status and other individual char acteristics. In the same study, though, uninsured children are more likely to rece ive services through an emergency care setting rather than through an outpatient setting re lative to insured children (Cunningham 2006). The two studies suggest that although the uninsured are dependent on ERs for their care to a much greater degree than low-income insured ch ildren, they still use ERs no more than Medicaid/SCHIP enrollees as a result of fewer health problems, an overall lack of financial access to medical care, and parents concer ns about incurring high medical bills. Besides insurance status, some other factor s associated with hi gher ER utilization among children include African American ethnicity, fewer financial resources, and living in a rural area (Wilson and Klein 2000). Inpatient care In contrast to the case of ambulatory care, be cause of a lower elastic ity of demand, there is less reason to expect strong positive insurance eff ects for inpatient utilization. Moreover, because the lack of insurance may cause inefficient use of medical carethat is, inadequate prevention and excessive reliance on emergency rooms which may in turn lead to avoidable hospitalizations, it is possible that losing insuranc e coverage will increase some types of hospital utilization (avoidable hospita lization, or hospitalization fo r ambulatory care sensitive conditions) in the long term. At the same ti me, Medicaid/SCHIP coverage can reduce the financial burden when children need hospitalizat ion service which is usually too expensive for the family to pay out-of-pocket.
32 Research shows that childr en covered by Medicaid/SCHIP are more likely to be hospitalized than uninsured child ren. Currie and Gruber found that Me dicaid eligibility raises the probability of being hospitalized in one year by 4% (Currie and Gruber 1996). Another study by Dafny and Gruber found that Medicaid eligibility reduces the rate of avoidable hospitalizations by 3.4% while increasing the overall hospitalizat ion rate (Dafny and Gruber 2000). Regarding variation among sub-populations, Currie and Gruber (1996) found a slightly smaller effect for native-born children and no significant effect for immigrant children. Prescription drug use Although prescription drugs are an optional service in Me dicaid, all 50 states and the District of Columbia provide prescription drug benefit to beneficiaries (KFF 2006). Among them, 42 states have different levels of copay requirements, although the copays are minimal. Children served by public programs are about three times as likely to have a medical problem that needs regular treatment with medication (treatment for three or more months) than uninsured children (Bloom et al. 2003). H ealth status is an important confounding factor that should be controlled for when examining the relations hip between insurance and medication use. Relatively few studies have examined the re lationship between insurance status and prescription drug use in the pe diatric population. One study found th at uninsured children are less likely than children in Medicaid/SCHIP to use any prescription drug even after adjusting for health status and other indi vidual characteristics (Chen a nd Chang 2002). Comparison between subgroups showed that African Am erican, Asian, and Hispanic ch ildren had lower prescription drug expenditures than white children and childr en in near-poor families had lower prescription drug expenditures than those in hi gh-income families. Expenditures
33 Total health care expenditures can reflect no t only the cost of medical care, but overall health care utilization. The differe nt utilization patterns of health care services between children covered by Medicaid/SCHIP and uninsured childr en can be reflected in the amount of total health care expenditures. It has been argued that the Medicaid /SCHIP expansions lead to a more efficient use of health care resources by im proving access to and increa sing utilization of preventive and primary care for poor children, and the increased ut ilization of preventive and primary care by children in the public progr ams helps reduce unnecessary and expensive hospitalizations. One study found that cost savi ngs from reducing avoidable hospitalizations more than offset the cost of additional physicia n visits generated by expanding Medicaid in the long term (Dafny and Gruber 2000). Previous studies suggest that uninsured child ren had lower expenses than children covered by public insurance (Newacheck et al. 2003) One existing study found that the total expenditures for children with transitions (any types of change in insurance status) were significantly higher than for conti nuously insured (public or priv ate insurance) children (Davis and Bruckman 2003). This indicates that the discontinuity of cove rage may be associated with higher health care expenditures. Medicaid/SCHIP versus Private Insurance Historically, low reimbursement rates, admini strative hassles, and residential segregation between providers and patients were cited as factors that contribute to access problems for children covered by Medicaid (Fossett 1992; Mitchell 1991; Sloan, Mitchell, and Cromwell 1978). After Medicaid expansion an d establishment of SCHIP, th ere has been some significant improvement in the benefits and program stru cture of public programs for children. Currently, the Medicaid benefit package is comprehensive, in cluding services such as well-child and dental visits under the EPSDT program, and requires little if any, cost sharing. In comparison, private
34 insurance benefits vary widely and are typically less comprehensive, while generally higher cost sharing. Some low-cost private plans do not even offer basic services like prescription drugs or preventive care. Economic theory and the RAND health insurance study suggest that higher cost sharing, via raising the price of se rvices patients pay out-of-pocket might result in a reduction in using health care (Freeman and Corey 1993; Pauly 1968; Pauly 2004). In terms of individual charac teristics, children covered by Medicaid/SCHIP and those with private insurance are quite different. Children co vered by Medicaid/SCHIP are more likely to be in fair or poor health, living in a lower-i ncome household and with unemployed parents, compared with privatel y insured children. In general, recent research suggests that Me dicaid-covered low-income children are more likely to receive services and to have more visits when they receive care than those low-income children with private insurance (Dubay and Kenne y 2001). This suggests that the higher costsharing and more limited benefits of private in surance than Medicaid/S CHIP could hinder the use of health services by low-income children. Therefore, it is likely that the utilization of preventive care and other medical care will decr ease after children transit from Medicaid/SCHIP to privately insurance. Preventive care Recent studies have been consistent in th e findings about whether Medicaid/SCHIP or private insurance serves children better in te rms of preventive care: children covered by Medicaid and SCHIP have a highe r probability of having any well-child visit and more visits given any visit than those with private insurance. And this diffe rence is still significant after more rigorously controlling for differences in the health, socioeconomic and demographic characteristics of children on Medicaid, and t hose with private health insurance (Dubay and
35 Kenney 2001; Duderstadt et al. 2006; Ku, Li n, and Broaddus 2007; Olson, Tang, and Newacheck 2005). These findings indicate that public coverage helps children get preventive health care and may even be more effective than priv ate health insuran ce for this purpose. Physician visits Cunninghams study (2006) using CTS data indicat es that low-income children covered by Medicaid/SCHIP on average have more physician vi sits than low-income children with private insurance, but the difference becomes insignifican t after adjusting for in dividual characteristics (including health status). Another study by Aiken and co lleagues (2004) compared the probability of having any physician visits betw een children who transitioned from public insurance to private insurance and children covered by private insura nce continuously, and no significant difference was found. Emergency care Children enrolled in Medicaid or SCHIP are reportedly more likely to use emergency rooms than children with private insurance (Ku, Lin, and Broaddus 2007; Luo et al. 2003; Johnson and Rimsza 2004). In Cunninghams study using CTS data, Medicaid/SCHIP children were 38% more likely to use ER services and had almost twice the number of ER visits than children who were privately insure d. It has been argue d that children enrolle d in public insurance programs are often in poorer health than pr ivately insured children and sometimes have difficulties getting a medical appointment quickly ; therefore, they use emergency care more often (Ku, Lin, and Broaddus 2007). This is supp orted by multivariate analysis: the difference became insignificant after controlling for health status and other individual characteristics (Luo et al. 2003; Cunningham 2006). Inpatient care
36 Few studies examined the difference in inpatie nt care use between children covered by Medicaid/SCHIP and those with private insurance. One existi ng study by Todd and colleagues (2006) found that compared with pr ivately insured children, children with public insurance have significantly higher rates of hos pital admission between 1995 and 2003. But this conclusion is based on the multivariate analysis that only adjust ed for age, race, and chronic conditions. They also examined the hospitalization rate among ch ildren with some chronic conditions and found a significantly higher hos pitalization rates for Medicaid/SCHIP children who have ambulatory care-sensitive conditions (ACSC), such as asthma, vaccine-preventable disease, psychiatric conditions, etc. And the insurance-associated differences were identified only for these ACSC but not other chronic conditions which are unlikely to be altered by primary care (e.g., appendectomy, childhood cancer, and trauma) (Todd et al. 2 006). Thus, the higher number of hospitalizations among children with public insura nce coverage should be attributed to higher hospitalization rates for ACSC, wh ich are avoidable if children received good, timely primary care. This suggests that the higher hospitalizations among child ren covered by Medicaid/SCHIP may be caused by failure to receive the timely primary care that they need. Todds study identified some ot her factors that affect inpa tient care utilization: the hospitalization rates are highest in young children (<5 years of age), decrease in school-aged children, and increase in olde r teenagers; Hispanic childr en had significantly higher hospitalization rates. Prescription drugs Few studies have examined the relationship between insurance coverage and prescription drug use in the pediatric population. One av ailable study, which analyzed 1996 MEPS data, indicated that children in Medica id have similar prescription dr ug use (probability of using any
37 prescription drug and number of prescriptions) with privately in sured children (Chen and Chang 2002). Expenditures It was found that privately insured adolescents had higher expenses than children covered by public insurance (Newacheck et al. 2003). No existing eviden ce on the difference in total expenditures between younger children covered by public insurance and those covered by private insurance is found. Gaps in Literature The current accepted relationship of Medica id/SCHIP disenrollment and changes in utilization and expenditures is based on crosssectional comparisons of the utilization and expenditure between children c overed by Medicaid/SCHIP and ch ildren without insurance, or children covered by Medicaid/SCHIP and those with private insurance. These studies themselves did not overcome some estimation difficulties. On e difficulty is the possibility of confounding differences among children of different insurance coverage. Individuals w ith different health insurance coverage are likely to differ in ways that are related to their utiliz ation of care but that are not readily observed by resear chers (i.e., unobserved heterogene ity). As a result, differences in utilization for children with and without Medi caid coverage will reflect the combination of a causal effect of insurance and th e effect of unmeasured characteri stics that are correlated with insurance coverage and/or use of care. Another estimation limitation is that point-i n-time measures of health insurance and utilization may not provide a comprehensive pict ure of the dynamic between insurance coverage and health care use. An analysis of monthly insurance coverage data from the MEPS found that 14% of U.S. children in 1999 experienced insu rance transitions that involved periods of uninsurance (Tang, Olson, and Yudkowsky 2003). Se veral state-level studies measured the
38 change in access to care after children enrolled in public insurance progra ms with multiple pointin-time measures of health in surance and access (Holl et al. 2000; Kempe et al. 2005). But no study has examined the relationships between Medicaid/SCHIP disenrollment and change in childrens utilization and expenditu res using measures over time. Some simulations of the effect of disenrol lment based on these cross-sectional comparison results may not be accurate Two studies have simulated th e effect of disenrollment on utilization. In both of the studies the authors assumed a certain disenrollment rate, 10% and 25% respectively, among current Medicaid/SCHIP children enrollees, and it was assumed that all of those dropouts became uninsured and had exactly the same utilization pa ttern with uninsured children in cross-sectional analysis (Cunni ngham 2006; Rimsza, Butler, and Johnson 2007). These assumptions are unlikely to be tr ue. Those children who disenrolled from Medicaid/SCHIP to become uninsured may be different from those who are continuously uninsured in some unmeasured individual char acteristics which are a ssociated with their utilization pattern. Therefore, there is a need to conduct a longitudinal study to answer the question, How does utilization and expenditures change when ch ildren disenroll from public insurance? Features of this Study This study will fill in this information gap w ith an attempt to overcome the limitations of previous studies. In this study, health care utilization and expe nditure changes resulting when children disenroll from Medica id/SCHIPand then become uninsured or transit to private insurancewill be measured using a national longitudinal data, MEPS. There are several important features of this study. First, a quasi-experimental study design will be used. Those children who disenrolled from Medicaid/SCHIP, and became uninsured or trans ited to private insuran ce, are two treatment
39 groups of interest. Those childre n who stayed in Medicaid/SCHIP during the same time period will serve as the control group. Use and expenditu res, insurance status, and covariates are measured both before and after the disenrollment. The design of using the same individuals in the pretest and posttest periods is more power ful than the untreated comparison group design with separate pretest and posttest samples in previous studies. Second, a DiD approach will be used in multivariate regressions. DiD analysis can control for time-invariant heterogeneity that may be correlated with insurance status change and/or utilization of care among children. Third, various types of health care utilization will be examined. It includes preventive care, outpatient care, inpatient care, emergency care, and prescription drug use. Fourth, the study uses a large longitudinal data, MEPS, and will give nationa l-level estimation on the research question.
40 Figure 2-1. Andersen Behavioral Model of Health Services Use (1995). Source: Andersen R.M. (1995). Revisiting the behavioral model and access to medical care: Does it matter? J Health Soc Behav 36: 1. Personal health practice Use of health services Predisposing Enabling Need Characteristics Resources Health Care System External Enviornment
41 CHAPTER 3 METHODOLOGY Research Questions and Hypotheses This study answered two research questions. 1. How does losing Medicaid/SCHIP coverage in fluence health care use and expenditures among children who became uninsured afterwards? 2. How does the transition from Medicaid/SCHIP coverage to private insurance influence health care use and expenditures among children? Research Question 1 Hypothesis 1a : After disenrolling from Medicai d/SCHIP and becoming uninsured, children use less preventive care, outpatient care inpatient care, and fewer prescription drugs than children continuously en rolled in Medicaid/SCHIP. Hypothesis 1b : After disenrolling from Medicai d/SCHIP and becoming uninsured, children use a similar amount of or more emerge ncy care than children co ntinuously enrolled in Medicaid/SCHIP. Hypothesis 1c : After disenrolling from Medicai d/SCHIP and becoming uninsured, children have lower total health care expenditu res than children continuously enrolled in Medicaid/SCHIP. Research Question 2 Hypothesis 2a : After transitioning from Medicaid/SCHIP to private insurance, children use less preventive care (well-child visit) th an children continu ously enrolled in Medicaid/SCHIP. Hypothesis 2b : After transitioning from Medicaid/SCHIP to private insurance, children have similar utilization in medical care, including outpatient care, inpatient care, emergency care, and prescription drug use, compared with chil dren continuously enrolle d in Medicaid/SCHIP.
42 Hypothesis 2c : After transitioning from Medicaid/SCHIP to private insurance, children have similar total health care expenditures co mpared to children con tinuously enrolled in Medicaid/SCHIP. Data Source MEPS household component (HC) 1996-2005 data will be used for analysis in this study. This section focuses on the design and scope of the MEPS data, the rationale for using the MEPS data, and limitations of MEPS data for this study. MEPS-HC Data The MEPS-HC is a national probability survey of the noninstitutionalized civilian population of the United States, conducted by the Agency for Healthcare Research and Quality (AHRQ). It is designed to produce national and regional estimates of the health care use, expenditures, sources of payment, and in surance coverage of the U.S. civilian noninstitutionalized population. The MEPS includ es surveys of medical providers (MPC), employers, and other health insurance providers to supplement the data provided by household respondents. The MEPS design permits both personbased and family-level estimates. In this study, the household components of MEPS data will be used to get individual-level estimates. Sampling design The MEPS-HC has an overlapping panel design. Information is collected from each household to cover a 2-year period. The first nine MEPS panels span 1996, 1997, 1998, 1999, 2001, 2002, 2003, and 2004, respectively. The MEPS-HC survey uses the National Health Interview Survey (NHIS) as its sampling frame. The NHIS sampling plan follows a stratified multistage area probability design. The complete NHIS sample consists of 358 primary sampling units, or PSUs, which are counties or groups of contiguous counties. The sample PSUs are stratified by geogra phic area, metropolitan
43 status, and socio-demographic measures. Each year a new MEPS-HC panel is established, drawing about one-quarter subsample from th e previous years NHIS sample. The MEPS-HC surveys carry over an over-sam pling of Hispanics and African Americans from NHIS. From 2002 on, the MEPS sample design over-samples Asia ns, persons predicted to have incomes less than 200 percent of the poverty level, and additional oversam pled African Americans in 2004 (AHRQ 2007; Cohen et al. 2000). Procedures for data collection Five in-person interviews are conducted w ith each MEPS panel over an approximately 30month field period to collect two full years of data. All interviews are conducted in person, using a computer-assisted personal inte rview as the principal data collection mode. Each interview takes on average 90 minutes to conduct. A knowle dgeable adult, typically a parent, answered questions for children under 17 years of ag e; 17-year-olds answered for themselves. Data collection is conducted over five rounds, with Round 3 spanning both calendar years. Each round of interviews asks about the period from last interview to the date of that interview. Round 1 asks about the period from January 1 of the calendar year to the date of that interview, and Round 2 asks about the time from the Round 1 interview through the date of the Round 2 interview. For each panel, Rounds 1, 2, and part of Round 3 typically contain data from calendar year 1; the remaining part of Round 3, and Rounds 4 and 5 cover calendar year 2 for a panel, as illustrated in Figure 3-1. Individuals data can be linked across rounds to build a longitudinal round-by-round data file spanning the 2 year s of survey participation. Scope of MEPS-HC The MEPS-HC collects detailed self-reported da ta, including demographic characteristics, family structure, household income, health and functional status, health insurance coverage, access to care, health care use, and expenditures.
44 A sample of medical care pr oviders identified by respondents is contacted to supplement and validate the information on medical events reported in the MEPS-HC about diagnosis, charges, payments, and specific services provided (MPC supplements). Data are collected on medical and financial characte ristics of medical and pharmacy events reported by HC respondents, including diagnoses, procedures, inpatient stays clas sified by DRG, prescriptions (medication names, strength, and quantity dispensed), charges, and payments. The MPC is conducted through telephone interv iew and mailed survey materials. Insurance coverage is verified by checking insurance cards (MEPS-IC). Weighted sequential hot-deck imputation is used to estimate missing data on the basis of responses from similar respondents. The combined response rates for the 1996 MEPS-HC range from 64.5% to 67.4% for the full-year files. Sampling weight specification Because of the complex design of the MEPS HC, the MEPS sample data must be weighted to obtain unbiased national es timates. The sampling weights reflect the disproportionate sampling adopted in NHIS to oversample minority populations. They also reflect adjustments for household nonresponse (MEPS Round 1), attr ition of persons (subsequent Rounds), poststratification (Censu s population estimates), and trimming of extreme weights. The final person weight contains a small number of cases assigned zero weight (AHRQ 2007; Cohen, DiGaetano, and Goksel 1999). There are two groups of weighting variables: person-level weights in the annual data file and longitudinal weights in panel files. The long itudinal weight variable s should be used when the sample includes persons partic ipating in both years of one pa nel to make national estimates of person-level changes in selected variables (e.g., health insurance, health st atus, utilization and expenditures), which is the case in this study
45 Rationale of Using MEPS-HC Data for the Study Administrative data (like cl aims data) usually only contain observations of certain insurance coverage, such as Medi caid, SCHIP, or private insuranc e. Utilization and expenditures information are not available for those children w ho are uninsured and those services that are not covered by the insurance plan. Therefore, it is im possible to use administrative data to examine the utilization/expenditures after childr en disenroll from Medicaid/SCHIP. The MEPS data has several main advantages th at make it a good data source to answer the research questions. First, it contai ns detailed demographic characteri stics, health status, insurance status, and utilization and expendi tures of care. Second, it is repr esentative of the U.S. population, including those without insuranc e coverage. Third, it allows for longitudinal analysis. By linking individuals data across rounds to build a longitudinal round-by-r ound data file, the association of changes in insurance with changes in hea lth care use and expenditures can be examined. Fourth, it allows for combining da ta of several years to gain a large sample size. The major survey questionnaires are relatively unchanged since 1996. It is feasible to combine data of nine years (1996) to increase sta tistical power of analysis. Why Use Round as Observation Period Unit? There are three options for th e observation time unit using MEPS data for the study: year, round, and month. First, yearly health insurance coverage va riables using monthly insurance variables in MEPS-HC data set were construc ted. If the insurance coverage is defined by covered/not-covered for the full year (12 months), the number of children who disenrolled from Medicaid/SCHIP is close to zer o. This is because by this defin ition the insuran ce changed right between 12/31 of year 1 and 1/1 of year 2, which is a rare case. Defining insurance coverage in the year by at least 10 month covered or not-cover ed was also attempted. Still, the sample size remained very small, which is consistent wi th the finding from a previous study examining
46 insurance transition using MEPS (Davis, 2003). If insu rance coverage in the year is defined by at least 6 months covered or not-covered, the meas urement of insurance coverage is subject to measurement error. In MEPS-HC data, there are a group of monthly insurance coverage variables. However, one important confounding factor in the relations hip between insurance ch ange and change in utilization/expenditureschange in health st atus over monthsis unobserved. This makes the monthly analysis model subject to omitted variable bias. When round is used as the unit for observati on time period, MEPS can provide a large enough sample size for the multivariate analysis. A nd, both health insurance and health status are measured in each round. Sample Inclusion Criteria for this Study Only observations that meet these crit eria were included in the analysis: 1) The child was born and younger than 19 years old fo r at least four cons ecutive rounds in one panel. That means those children w ho were born at the start of round a and those who were less than 19 years old at the end of round d will be included in the study. (Rounds a, b, c, and d are four consecutive r ounds in one panel. Round a can be any of round 1. Figure 32). 2) Children were covered by Medicaid/SCHIP fo r at least two round before the point of change ( a and b ). And, after the point of change, all of those children were under the same insurance coverage (Medicaid/SCHIP, uni nsured, or private insu rance) for next two interviews or rounds ( c and d ). 3) Children whose longitudinal weights were positive. 4) If a child had multiple observations (change of insurance), only the first observation was included. Description of Outcome Measures and Independent Variables Outcome Measures Two groups of outcome measures were examin ed, use of care and expenditures of care (Table 3-1). The observation unit time period wa s round. Considering that the length of rounds
47 were not the same for each round, the amount of utilization and expenditures given nonzero utilization and expenditures during the round were ca lculated into standardized variables to make the utilization/expenditures comp arable between person-rounds. The specific steps were as below: first, divide the amount of utilization/expenditures gi ven nonzero in the round by the number of days in the round to obtain the averag e utilization/expenditures per day; then, multiply the average utilization/expend itures per day with the averag e length of round (145.8 days). round one in days of number average round the in days of number Y Yoriginal dardized s tan Expenditure measures Total health care expenditure is a measure of all-purpose health care utilization and total expense of care. Expenditures in the MEPS are de fined as the sum of payments made, including those made out of pocket and by third-party paye rs (Zuvekas and Cohen 2002). It is important to distinguish between expenditu res and charges when measuring health care expenses. Charges are the fees billed to patients and insurance compan ies by providers and health care organizations for health services they provided (Berk and Monheit 2001). Expenditu res or expenses are direct payments to providers and health care organizatio ns by individuals or health plans for health services provided (AHRQ 2002). Expenditures do not include charge s associated with bad debts, uncollected liability, and charitable care for wh ich no payment is made. Furthermore, in MEPS the payments of some medical events reported we re zero, when the care was covered under a flat fee arrangement beginning earlier, or when follo w-up visits were provide d without a separate charge (e.g., after a surg ical procedure). Prescrip tion drug file does not have a charge variable in MEPS. Amount of total expenditures during the round was the expenditure measure in the study. The expenditure variables were constructed from the original event files by summing all of the
48 total expenditure for each event by person-round. Health care e xpenditures are a better measure of what is paid because insurance companies an d other health care purc hasers often negotiate discounts on charges that substa ntially reduce the amount that th ey actually pay for health services. Expenditure is a better measure for the amount of expense actually incurred for care used. The number of events under flat fee usually is very small. Therefore total amount of expenditures during the round was used to measure the total expense of al l of the care. All the expenditure data were conve rted to 2005 U.S. dollar us ing the Medical Care CPI (U.S. Department of Labor 2008). Utilization measures Five types of health care use were measure d. They were preventive care, outpatient care, inpatient care, emergency care, and prescription drug use. The measures were the quantity of each type of care in the round. Those variables were constructed from MEPS Events Files by counting the number of events with in the round for each type of care. Well-child visits. The specific type of preventive care measured in this study was wellchild visitnumber of well-child visit in the r ound. Well-child visit includ ed well-child visits, general check-ups (general checkups when the ch ild is not sick or injured) and a small number of visits for immunizations or shots. Well-child care was iden tified using visit-level information on office-based and hospital outpa tient visits. The MEPS asks respondents the primary reason for the visit and prompts responden ts with a list of possi ble reasons. For this analysis, visits were coded as well-child care if the prim ary reason given was well-child examination, general checkup, or immunization or shots. Physician visits. The measurement of outpatient care in this study was physician visit the number of physician visits in the round. Phys ician visit is a measure in common use in previously published studies of health care use a nd its relationship to heal th insurance coverage.
49 Physician visits were identified using visit-le vel information on office and hospital outpatient visits. In the MEPS survey, the ques tion is asked if the patient talke d to MD at this visit. If the answer is yes, the visit was coded as a physician visit. Hospitalizations. The measure of inpatient care was th e number of hosp italization in the round. The variable came from Ho spital Inpatient Stay file. The hospitalization was excluded if the reason entered hospital was to be born. ER visits. The emergency care utilization was measured with the number of ER visit in the round. The MEPS contains an event file that has information on each ER visit. All of the ER visits within the round were coun ted in the number of ER visits. Prescription drug use. Prescription drug use was meas ured with the number of prescription drug used in the round. The variab le came from the MEPS prescription drug file. The count excluded diabetic equi pment, supplies, and insulin. Major Independent Variables Change of insurance status was the major i ndependent variable in the study. Specifically, there were 3 types of change of insurance coverage of interest in this study: 1) Stayed in Medicaid/SCHIP through the two roundsMM; 2) Disenrolled from Medicaid/SCHIP (round b ) to become uninsured (round c )MU; 3) Transitioned from Medicaid/SCHIP (round b ) to private insurance (round c )MP. Medicaid/SCHIP coverage refers to whether the child was covered by Medicaid or SCHIP. Private insurance included employer-sponsored a nd nongroup policies. The child was considered uninsured if not covered by any of the listed insurances source in MEPS, including Medicaid, SCHIP, TRICARE, Medicare, or other public hospital/physician or pr ivate hospital/physician insurance. Children covered only by state-spec ific programs that provide non-comprehensive coverage and those without hospita l/physician benefits (for example, private insurance for dental or vision care only) were not consid ered to be insured in this study.
50 If the child was covered by one type of insu rance at two adjacent interviews, he/she was categorized as having the coverage for the full round between the two inte rview dates. It was assumed that he/she did not change insurance co verage between two interview dates (about five months). For example, if a child was covered both at interview a and interview b then he/she is covered by Medicaid/SCHIP for the full round b (Figure 3-3). In MEPS-HC data file, there are a series of variables about the health insura nce status in each round (at interview date). MCAID31X, MCAID42X, and MCAID53X were us ed to identify thos e children who were covered by Medicaid/SCHIP in each round. PRIV 31, PRIV42, and PRIV53 were used to identify those children who were covered by private in surance in that round. INS31X, INS42X, and INS53X were used to identify those child ren who were uninsured in that round. Control Variables The observed differences in change of utilization and expenditure as insurance coverage changed could be associated with differences in in dividual characteristics (such as health status). Therefore, those factors were controlled fo r in the analysis. The control variables, a comprehensive range of factors known or likely to affect health care ut ilization, included the childs age, gender, race, geogr aphic area, rural/urban, family structure, family income, and health status. Table 3-3 descri bes those control variables orga nized by the Andersen model. Demographic characteristics Age was grouped on the basis of the categories used in the eligibility criteria of Medicaid/SCHIP. There were fi ve age groups: 0yr, 2yr, 6yr, and 13yr. Gender is a dichotomous variable: female and male. The race and ethnicity was categorized into four groups: non-Hispanic wh ite, non-Hispanic African American, Hispanic, and other. The region of the childs residence was grouped into Northeast, South, Midwest, and West. Rural versus urban residence was determined by whether the respondent lived in a metropolitan statistical area.
51 Family structure characteristics included number of pa rents, moms age, parents highest education, employment status of parents, and insurance coverage of parents. Number of parents was a 3-category variable: no parents, single parent, and two parents in the household. Moms age was a 2-category variable: and and over. All of the childrens mothers were 24 and older in this study. Pa rents education was the highest education attainment of the childs parentsmother or father, whoever ha s the higher educationwhe n first entered MEPS. It was a 3-category variable: le ss than high school (HS), comple ted HS, and any college. The other two family characteristics variables, parents employment status and insurance status, were measured in each round and the changes over the two rounds ( b and d ) were used in multivariate analysis. Parents employment status has two categor ies: employed, defined as at least one parent working, and unemployed, defined as no parent working in th e round. Based on this, change in parents employment status over the two rounds was define d as a four-category variable: always employed, employed to unemployed, unemployed to employed, and always employed. Insurance coverage of pare nts was grouped into four categories: both parents were covered by public insurance, bot h parents had private insurance, mother and father had different types of insurance, and both parents were not insured. Based on this measurement, the change in parents insurance status variable was defi ned as: changed, when the parents insurance status measures in round b and round d were not the same, and no change, when the parents insu rance status measures in round b and round d were different. Family income was defined as income from all rela ted family members. The sources of income included annual earnings, unemployment and workers compensation, interest and dividends, child support, TANF, etc. The origin al variable in MEPS, POVCATyy, was used in the analysis. POVCATyy was constructed by divi ding family income by the applicable poverty
52 line (based on family size and composition), w ith the resulting percentages grouped into five categories: negative or poor (less than 100%), near poor (100% to less than 125%), low income (125% to less than 200%), middle income (200% to less than 400%), and high income (greater than or equal to 400%). The FPL for a house hold of four was $19,350 in 2005 (U.S. Department of Health and Human Services 1999). The fam ily income was measured once each year in MEPS. Because round b always falls in the first year and round d always in the second year of the panel, the yearly income variables were us ed as approximations of round income measures. The change in family income over the two round s, a 3-category variable, was defined as the change in income from year 1 to year 2: increased, no change, and decreased. Health status was measured with two variables. On e was the self-evaluated (or parentevaluated) health, an ordinal measurement. In MEPS, the respondents were asked in general, would you say your health is with a 5-catego ry Likert response scale of Excellent, Very Good, Good, Fair, and Poor. This question was as ked in each round. Based on this variable, the change in general health status a 3-category ordinal variable, was developed: better, no change, and worse. The other measure was th e number of medical conditions (0, 1, 2, ) in the round. It was obtained from the medical cond ition file. The difference in number of medical conditions between round d and round b was the change variable used in multivariate analysis: decrease, no change, and increase. In addition, one variable time indicating which years the child was surveyed was included as control variable. Sin ce the environmental characteristic s were not controlled in this study, the time fixed effect was in tended to filter out any secula r trend due to changes in the economy or health care market.
53 Study Design and Statistical Analysis Longitudinal data of preand post-disenrollm ent was analyzed using a retrospective quasiexperimental study design, that is, a comparison group pre-test/pos t-test (before/after) design. This comparison group pre-test/post-test design is th e same as the classic controlled experimental design except that the subjects cannot be randomly a ssigned to either the treatment or the control group, or the researcher cannot control which gr oup will get the treatment (Campbell and Stanley 1966; Cook and Campbell 1979). Both of the pre/ post comparison and the treatment/control group features are essential to drawing appropriate infere nces about the effects of Medicaid/SCHIP disenrollment in the absence of a randomized c ontrolled trial. This method of analysis is preferable when multiple data poi nts are available over time and improves upon crosssectional analysis in that the history of the sample is availa ble and therefore provides more information to make estimation. The definitions of treatment and comparison groups are demonstrated in Figure 3-3 and Figure 3-4. All subjects were in itially enrolled in Medicaid/S CHIP for at least one round (baseline round, b ). Children who disenrolled from Medicaid/SCHIP and then became uninsured (MU) or transitioned to private insurance (MP) were the two treatment groups of interest. Children who stayed in Medicaid/SCHIP over the two rounds during the same year (MM) were considered as comparison group. The change of in surance status happened at some point in round c Round b was the pre-change period (baseline round) and round d was the post-change period (following round). That is, for outcome m easures, the pre-change measures were the utilization/expend itures during round b and the post-change measures were the utilization/expend itures during round d (as shown in Figure 3.4). Each round covered the period from last round interview date to current round in terview date. For each child, the changes in the amounts of utilization/e xpenditure between the baseline r ound and the following round were
54 calculated. The changes were then compared be tween the treatment groups and the comparison group. Bivariate analysis and multivariate analysis were conducted to examine the relationship between Medicaid/SCHIP disenrollment a nd utilization/expenditure changes. Descriptive Analysis Bivariate analyses were perfor med to describe the characteri stics of study sample and the relationship between insurance ch ange and outcome measures. Multivariate Analysis There are some challenges to identifying the effect of Medicaid/SCHIP disenrollment on use and expenditures. Children were not randomly assigned to the three insurance states, and therefore may not have the same individual and fa mily characteristics, which can influence the use and expenditures of care. Those observabl e differences should be controlled in the multivariate analysis. In addition, there may be unobserved heterogeneity (omitted variables) correlated with both the likelihood of di senrolling from Medicaid/SCHIP and the utilization/expenditures of health care. Fo r example, children who disenrolled from Medicaid/SCHIP may have less family/commun ity support to get Medicaid/SCHIP renewed and medical care when needed. The existence of omit ted variables could threaten internal validity. When there are omitted variables, any predictors that are correlated with them will end up proxying for them, and the predictors estimated coe fficient cannot be interpre ted as the effect of that predictor per se, since it also captures part of the effect of the omitted variables (Ettner 2004). In this case, it means that the estimated effect of insurance ch ange on utilization and expenditures will be biased. This problem was addressed and the impact of Medicaid/SCHIP disenrollment on utilization and expenditures was identified using a DiD approach (also called double difference).
55 Difference-in-difference analysis Specifically, health care util ization/expenditures changes (from pre-measure and postmeasure) between the treatment groups and the comparison group were compared. The first difference, which is the change over time (befor e and after disenrollment) within individuals, eliminates the influence of time-invariant unobserved individual heterogeneity. The second difference, between a treatment group (MU or MP) and a comparison group (MM), eliminates the influence of difference in public insuranc e policy and other health system environment factors. The general DiD analysis specification was as below: Y kd Ykb = + 1 MU + 2 MP + 3i Covariates + (3-1) where Y kd is the kth outcome measure of round d and Y kb is the kth outcome measure of round b The unit of analysis is person-round, is the grand intercept term, and MU and MP are two dummy variables indicat ing what type of change in insu rance status the child had over two rounds. MU =1, if the child dise nrolled from Medicaid/SCHIP (round b ) and were uninsured (round d ); MU =0, if the child stayed in Medicaid/SCHIP in the two rounds, and 1 is the coefficient to be estimated for the impact of MU. MP =1, if the child switched from Medicaid/SCHIP (round b ) to private insurance (round d ); MP =0, if the child stayed in the Medicaid/SCHIP in the two rounds, and 2 is its coefficient. Covari ates are a vector of timevarying control variables including family inco me and health status, plus a vector of timeinvariant variables including age, gender, race, rural/urban, region, family structure variables, and time (Table 3-3). Change of family income, pa rents employment status and insurance status, and health status were obtained by subtracting the measures in round b from the measures in round d Timet is the vector of year dummy variables. 3i is a vector of coefficients to be estimated and is error term.
56 The validity of the DiD method relies on two assumptions: first, th at individual-level unobserved heterogeneity is constant over time and, second, that time effects such as changes in health system environment are id entical for the treatment and c ontrol groups (Blundell and Dias 2000; Wooldridge 2005). Within the general DiD specification, the e ffect of Medicaid/SCHIP disenrollment on change in amount of utilization/expenditure was estimated with multivariate regression models. Estimating amount of change in utilization/expenditures Regression models within DiD model (equ ation 3-1) were selected based on the distribution of data as displayed in Figure 3-5. Common features of utilization and expenditures data. The distribution of health care utilization and expenditures data usually are le ft censored, with a spike of zero values and skewness. These properties make untransformed ordinary least squares (OLS) estimation inefficient. One popular method for handling these non-normal distribution problems with utilization/expenditure data is to run a two-part model, which models the probability of nonzero values separately from their level conditional on nonzero. Usually the first equation is a logistic regression for a binary outcome variable (use/no use) and the second is an OLS for those with nonzero (positive) use. The dependent variable is commonly log-transformed before OLS estimation to accommodate skewness (Duan 1983; Duan et al. 1984). In this study, as specified in equation 3-1, the dependent variables were the changes in outcome measures over two time periods ( Y kd Y kb). There were many zeros in ( Y kd Y kb) since many children didnt use any servic es or used the exact same am ount of services in the two rounds. In this case, a conventiona l two-part model will not fit th e data mainly for two reasons.
57 One is that the dependent variables have both positive and negative values. The negative values cannot be transformed (log or square root) to meet OLS assumptions if needed for the second linear regression. Neither can they be modele d directly with gamma regression (gamma distribution is only fo r nonnegative values). Therefore, when ( Y kd Y kb) dont have a normal distribution, those negative values cannot be modeled efficiently together with positive values using transformed OLS or gamma regression. Second, it is likely that the effect of insurance change on utilization/expenditures (coefficients ) is different among those who had increased utilization/expenditures and thos e who had decreased utilization/e xpenditures. In such situation, estimations made by using one single linear regre ssion for the two types of change (decreased and increased) are biased. A three-part model, in stead of the two-part model, was used to solve those potential issues. Three-part DiD model Given the general DiD specification in equati on 3-1, the impact of Medicaid/SCHIP disenrollment was predicted with a three-part multiple regression model (see Figure 3-5). The first equation modeled wh ether the utilization/expenditures decreased Dk=1 ( Y kd Y kb <0); did not change, Dk=2 ( Y kd Ykb =0); or increased, Dk=3 ( Y kd Y kb >0), with multinomial logit regression. The second equation modeled how much the amount of utilization or expenditures decreased conditional on that the utilizatio n/expenditures decreased (D=1), that is, E(Y|D=1). The third equation modeled how much the utiliz ation or expenditures increased conditional on that the utilization/expenditure s increased (D=3), that is, E(Y|D=3). The second and the third equation were modeled with OLS regressi on and/or generalized linear model (GLM). Thus, the responses were decomposed into three estimation equations that dealt with specific parts of the distribution. After estimating each of these equations, the results we re then combined to get the full effect of disenrollment on use of cer tain type care or total expenditures.
58 Part I: Multinomial logit regression for how use/expenditures changed Part I of the model predicted the probability that a child had decreased, not changed, or increased utilization/expenditures when insuranc e status changed. A multi nomial logit regression was employed to model how certain type of util ization or total expenditures changed over two rounds. The dependent variables (Dk) for the regressions were multiple categorical variables coded based on the difference in amount of utiliza tion/expenditures in the two rounds as below: Dk=1, if Y kd Y kb <0; Dk=2, if Y kd Y kb =0; Dk=3, if Y kd Y kb >0. where Ykd is the standardized amount of u tilization or expenditures in round d and Ykb is the standardized amount of utiliz ation or expenditures in round b There are three types of regression which can be used to model this type of dependent variable: ordered logit, multinomial logit, and multinomial probit. Ordered logit assumes that the coefficients are constant across categories ( no change versus decreased and increased versus no change), which is not true in the study. Multivariate logit does not make that assumption. While multinomial probit is also free of that assumption, the procedure is too computationally complex to be practical for ap plication. The goodness of fit tests showed that the multinomial probit results were equivalent to the results from the multinomial logit. Therefore, multinomial logit was used in this study. Part II and III: Linear regression for how much use/expenditures changed The amounts of change in use/expenditures (giv en any change) are a group of continuous variables Two linear regressions were run to model the relationshi p between insurance change and use/expenditures change. One linear regres sion was used for the change in amount of
59 use/expenditures given negative change (Dk=1), and another for the change in amount of use/expenditure given positive change (Dk=3). For those observations with negativ e change in outcome measures (Dk=1) ( Part II ): E[ Y kd Y kb | ( Y kd Y kb)<0] = + 1 MU + 2 MP + 3i Covariates + (32) If transformation (log, square ro ot, etc) of dependent variables is needed to meet OLS assumptions or identify GLM link, multiply ( Y kd Y kb) with -1 before running regressions. For those observations with positiv e change in outcome measures (Dk=3) ( Part III ): E[ Y kd Y kb | ( Y kd Y kb)>0] = + 1 MU + 2 MP + 3i Covariates + (33) The linear regressions (Part II and Part III) were OLS and/or GLM regressions, depending on which model fit the data the best. Some researchers have co mpared log-transformed OLS and a range of GLM alternatives under a variety of da ta conditions that researchers often encounter in health care use and cost study, and they c oncluded that no single m odel is best under all circumstances examined (Basu and Ra thouz 2005; Manning and Mullahy 2001). OLS regressions By far the most prevalent modeling approach us ed in health economics and health services research is to use OLS or a least-squares variant with ln(y ) as the dependent variable. OLS regression analysis is a method for linear regression that determines the parameters in a statistical model by minimizing the sum of squared residuals. The OLS approach to regression analysis has been shown to be optimal when it satisfies th e Gauss-Markov theorem. In other words, the assumptions of OLS need to be met to obtain unb iased, consistent and efficient OLS estimators.
60 Normality of error term The distribution of Y conditional on X should be normally distributed. If there are outliers (skewed), the OLS estimates are inefficient. The utilization and expenditure data may not be norma lly distributed, as they tend to have a distribution with a long right tail (Diehr et al 1999). Jarque-Bera test (skewness and kurtosis), kernel density plot (kdensity command in Stata), st andardized normal probability pl ot (pnorm), and quantiles plot (qnorm) were used to check the normality of the residuals. If the error terms were nonnormal, the dependent variables were transfor med. The log transformation can pull in the upper tail of the distribution, as does the square root transformation to a lesser extent (Mullahy 1998). Log-transformation, square root, cubed root fourth root, and other transformation were examined for the dependent variables to determine the best fit. The distribution of the residuals from the transformed models were plotted to de termine the suitable transformation. In this study, the log transformation of expe nditure change is the clos est approximation of a normal distribution and was used in expenditure anal ysis. Subgroup smearing retransformation was used in prediction after the log-tran sformed OLS regression. For various utilization variables, none of those transformations (log and roots) produced an approximation of a normal distribution and therefore GLM regressions were used instead of OLS regression. Linearity The relationship between dependent vari ables and independent variables should be linear. When the relationship is not linear (inc orrectly specified), the coefficients are biased and the standard errors are incorrect (specifi cation errors). The Hosmer-Lemeshow test was conducted to examine the linearity assumption. Error terms are independent and identically distributed. The 's are independent and identically distributed which implies there s hould be no heterogeneity of variance and no autocorrelation among the residuals.
61 Homoskedasticity. The variance of the error term, given Xi, is constant. If the variances of error terms vary with Xi, there is heteroskedasticity. Hete roskedasticity leads to invalid estimation of standard errors of coefficients. Th is does not bias actual OLS coefficient estimates, but can cause invalid statistical inferences of coefficient significance. The Breusch-Pagan tests and Whites test were performed to test hete roskedasticity. Because in Stata robust option cannot be used after the svy: regress command, Whites heteroskedasticity-consistent covariance matrix (1980) cannot be used to correct for heteroskedasticity. When heteroskedasticity was detected, subgroup smear ing estimators were calculated for each insurance change group in the pr ediction after OLS regression. No autocorrelation Error terms of observations are not correlated. The OLS estimators are still unbiased but are not efficient when there is autocorrelation. The presence of autocorrelation causes the variance of er ror terms to be underestimated, wh ich leads to false positives and smaller confidence intervals (J ohnston and DiNardo 1997). In th is study, if more than one observation from one child is included in the sa mple, there will be autocorrelation. Therefore, every child was included only once in the analysis. Independence of the xs and This is also called zero c onditional mean assumption. When the assumption holds, we can say we have exogenous explanatory variables. This is a critical assumption. But we never know for sure whether is unrelated to xs. Including a lagged dependent variable at the righ t-hand-side (RHS) of the equa tion provides a simple way to account for historical factors that cause current differences in th e dependent variable that are difficult to account for in other ways (Wooldr idge 2005). The previous utilization/expenditures can affect the amount of utilization/expenditure s change with insurance status change. Some children tend to use more health care than othe rs with the same health care need. Including Ykb
62 can aid in getting a better estimat e of the effect of insurance ch ange on utilization/expenditures. When the model satisfies the key zero conditiona l mean assumption, the OLS estimators are still consistent. For the equation yt = 0 + yt1+ ut, the zero conditional mean assumption is E( ut| yt1) =0 (Wooldridge 2005). For the OL S regressions which include Ykb at the RHS of the equation, it is assumed that this zero conditional mean assumption is satisfied, that is, Ykb is not correlated with the error term. In the sensitivity anal ysis, the performance of models including Ykb at the RHS and not including Ykb at the RHS were compared. GLM Regressions The GLM is a useful generalization of OLS regression. There are no distributional assumptions required for GLM. It relates the random distribution of the measured variable of the experiment (the distribution function) to the systematic (non-random) portion of the experiment (the linear predictor) through a function called the link function. To build a GLM, a family and a link function must be identified based on the data. The family specifies the distribution that reflects the mean-variance relationship. For exam ple, in Gaussian (normal) distribution the variance is constant, and in ga mma distribution variance is pr oportional to square of mean (Fahrmeir and Tutz 2001). The link function, e.g ., identity and log, spec ifies the relationship between the covariates and th e mean. GLMs are attractive becau se the link function directly characterizes how the mean on the raw untransfo rmed scale is related to the predictors. Therefore, a GLM with log link has the advantage of the log models, but can give the prediction directly and does not require any smearing correction as transformed OLS does. One disadvantage of GLM is its inefficiency with he avy-tailed data. If the log-scale residuals are heavy-tailed, i.e., log-scale resi duals have high kurtosis (>3), GLM estimators are still unbiased but can suffer from substantial precis ion loss (Manning, Basu, and Mullahy 2005).
63 The generalized Gamma model (GGM) with a lo g link has been shown to be a suitable model for risk adjustment of skewed outcomes data in health services re search (Blough, Madden, and Hornbrook 1999; Dodd et al. 2006). The GGM is based on the generalized Gamma distribution, which is skewed, takes only positiv e values, and has one scale parameter and two shape parameters (three parameters). The GGM w ith log link can be estimated with maximum likelihood. It has been shown that given the poten tial for bias and inefficiency in standard approaches (including OLS with a normal error, OLS for the log-normal), the GGM is a more flexible alternative. If none of the standard approaches is appropriate for the data, the GGM provides a more efficient estimator because it bette r approximates the distribution than the more restrictive alternatives (Manning, Basu, and Mullahy 2005). In the ca se of the very heavy tailed data, the GGM with log link is not as efficient as transformed OL S, but can perform better than most other GLMs and can estimate wit hout a noticeable loss in precision In the analysis of utilization outcome measures, GL M regressions were used. Modifi ed Park tests were conducted to determine family and linktest was used to check on the link. Predictions based on the three equations The predictions from the three equations were then combined to obtain the estimates of adjusted change in utilization/expenditures per pe rson for each insurance change group (3-4, 3-5 and 3-6). E(Yk |MM) = P(Dk=0 |MM) E(Yk |MM, Dk=0) *(-1) + P(Dk=2 |MM) E(Yk |MM, Dk=2) (3-4) E(Yk |MU) = P(Dk=0 |MU) E(Yk |MU, Dk=0) *(-1) + P(Dk=2 |MU) E(Yk |MU, Dk=2) (3-5) E(Yk |MP) = P(Dk=0 |MP) E(Yk | MP, Dk=0) *(-1) + P(Dk=2 |MP) E(Yk |MP, Dk=2) (3-6) where Yk = Y kd Y kb.
64 The confidence intervals (CI) of the pr edictions were obtained by bootstrapping. Bootstrapping is a non-parametric method for tests of means (Kennedy 1998). If the 95% CI contains zero then the difference is not significant at the 0.05 level. In the expenditure analysis, the dependent variable Y was re-transformed to obtain estimates on the original scale before the pred iction, since it was transformed to satisfy OLS assumptions. Because heteroskedasticity was det ected, subgroup-specific smearing factors were used to obtain unbiased estimates. Mannings subgroupj-specific smearing factor (3-7) and the retransformation formula (3-8) are as below (Manning 1998): j j ju n S 1, j =0, 1, 2 (3-7) E(Yj) = e E(X ) Sj (3-8) Sensitivity analysis Sensitivity analysis was performed to identify the best model for multivariate analysis. Alternative: Include Ykb (baseline outcome variables) in the covariate. There is the argument that a lagged dependent variable on the right-hand side of an OLS regression is too problematic for use in most situations. More specifically, if the lagged dependent variable ( Ykb) is correlated with a residual, the coefficient estimators will be inefficient. At the other side of the argument, some research has shown that while the lagge d dependent variable is inappropriate in some circumstances, it remains an appropriate model for the dynamic theories often tested by applied analysts (Keele and Kelly 2006). Regressions with Ykb (baseline outcome variables) in the model as a covariate were run and the performances of the models with Ykb and without Ykb were compared.
65 All estimates were produced using survey modules of Stata 10.0 (Stata Corp., College Station, TX). To account for the complex sampling design and overlapping panel nature of MEPS, all statistics analysis were adjusted with related weighting variables in the dataset. All estimates, tests, and differences were statis tically significant at the 5% level or better.
66 Figure 3-1. Panel desi gn of MEPS (2003). Source: MEPS HC-89: 2004 Full Year Consolidated Data File. 2006. Agency for Healthcare Research and Quality. Rockville, MD. Figure 3-2. Sample inclusion criteria 0
67 Figure 3-3. Measuring cha nge of insurance status Figure 3-4. Study design Insurance a-b Insurance c-d Comparison group: Medicaid/SCHIP-----------------Medicaid/SCHIP (MM) Treatment group: Medicaid/SCHIP-----------------Uninsured (MU) Treatment group: Medicaid/SCHIP-----------------Private Insurance: (MP) (Following round) (Baseline round) Pre-change Measure Post-change Measure Health a Income a Health c Income c Round a Round b Round d Round c ( a is round 1 or round 2 in one panel ) Disenrolled Change Round a Round b Round d Round c ( a is round 1 or round 2 in one panel ) Insurance a = Insuranceb (Medicaid/SCHIP) Insurance c = Insurance d Medicaid/SCHIP: MM Uninsured: MU Private insurance: MP
68 Figure 3-5. Models for predicting amount of utilization/expenditure change under the DiD specification The general DiD specification: Y kd Y kb = + 1 MU + 2 MP + 3i Covariates + (3-1) No Yes (Y kd Ykb): Many zero? Three-Part Model I: Multinomial Logit Get: P( y kd y kb<0) P( y kd y kb=0) P( y kd y kb>0) II: Linear regression for negative values *multiply ( y kd y kb) with -1 be f o r e m ode lin g III: Linear regression for positive values OLS OLS assumptions GLM Family: Gaussian, Gamma Link: identity, logit, log Re-transformation Smearing estimator, or subgroup smearing estimators if log-transformed OLS Predictions GLM Predictions One-Part Model OLS OLS assumptions GLM Family: Gaussian Link: identity Evaluate and Choose Absolute prediction errors
69 Table 3-1. Summary of outcome measures Measures Types Utilization Number of well-child visit in the round (standardized) Continuous Number of physician visit in the round (standard ized) Continuous Number of ER visit in the round (standardized) Continuous Number of hospitalization in the round (standardized) Continuous Number of prescription filled in the round (standardized) Continuous Expenditure Total expenditure in the round (standardized) Continuous
70 Table 3-2. Sources of explanatory va riables (point-in-time measures) Variable Types Categories Original variables in MEPS Predisposing Factors Age Categorical 0yr 2yr 6yr 13yr AGErrX Gender Categorical Male, female SEX Race/Ethinity Categorical Non-Hispanic white Non-Hispanic African American Hispanic Other RACEX, RACETHNX Rural/Urban Categorical Urban, Rural MSArr Region Categorical Northeast, West, Midwest, South REGIONrr Number of parents Categorical 0, 1, 2 *Created Parents age Categorical Under 24 yr 24 yr 35 yr or over AGEyy Highest education of parents Categorical Less than HS Completed HS or has GED Any College HIDEGYR Enabling Factor Family Income Ordinal categorical =<100%FPL 101%FPL 201%FPL 301%FPL >400% FPL POVCATyy Health Insurance Categorical Medicaid/SCHIP Uninsured Private insurance MCAIDrrX; INSrrX; PRIVrr At least one parent work Ca tegorical Yes, No EMPSTrr Insurance coverage of parents Categorical Medicaid Private insurance Mixed types of ins Uninsured MCAIDrrX; INSrrX; PRIVrr Need Factors General health status Ordinal categorical Excellent Very Good Good Fair Poor RTHLTHrr # medical conditions Continuous 0, 1, 2, 3, *Created Time Dummy Dummy for each panel
71 Table 3-3. Description of covariate variables in multivariate analysis Variable Covariates in DD analysis Coding of covariates Age Measure in the baseline round (b) 0=0yr 1=2yr 2=6yr 3=13yr Gender Measure in the baseline round (b) 0=male 1=female Race Measure in the baseline round (b) 0=Non-Hispanic white 1= Non-Hispanic African American 2=Hispanic 3=Other Rural/urban Measure in the base line round (b) 0=rural, 1=urban Region Measure in the baseline round (b) 0=Northeast 1=West 2=Midwest 3=South Number of parents living with child Measure in the baseline round (b) 0, 1, 2 Moms age Measure in the baseline round (b) 0=Under 24 yr 1=24 yr 3=35 yr or over Highest education of parents Measure in the baseline round (b) 0=Less than HS 1=Completed HS or has GED 2=Any college Family Income Change from round a to round c 0=Decreased 1=No change 2=Increased At least one parent work Change from round a to round c 0=No change 1=Change Insurance coverage of parents Change from round a to round c 0=No change 1=Changed General health status Change fr om round a to round c 0=Decreased 1=No change 2=Increased Number of medical conditions Change from round a to round c 0=Decreased 1=No change 2=Increased Timet Dummy for each panel Dummy
72 CHAPTER 4 RESULTS Overview The results of the study are presented in two sections. First, the demographic and socioeconomic characteristics of children and their parents in each insurance change group are described. Percentage and Chi-square test re sults are included. Sec ond, the results from the bivariate analysis and multivariate analysis of the relationship between Medicaid/SCHIP disenrollment and change in outcome measures are presented. This section is organized by outcome measures in the sequence of total expend itures, well-child visits, physician visits, ER visits, hospitalization, an d prescription drugs. Description of the Sample Table 4-1 presents the number of children e ligible for the study. The final study sample consisted of 932 children in the MU group, 889 children in the MP group, and 10,865 children in the MM group (Table 4-1). Among the study sample, 7.35% were MU and 7.01% were MP. The unweighted percentages may not reflect th e real Medicaid/SCHIP disenrollment rate. One reason is that it is not weighted. Another is that the study includes only those children who stayed in Medicaid/SCHIP for at least two rounds before the insu rance change and remained in the same insurance status for at least two rounds. More than half of the children who disenrolled from Medicaid/SCHIP were exclude d because they didnt stay in the same insurance status for more than one round. Individual and Family Characterist ics of Children in the Study In Table 4-2, the characteristics of children in three insurance change groups (MM, MU, and MP) were compared. Chi-square tests were conducted to determine wh ether children of the
73 three insurance groups had differe nt individual characteristic s. Significant differences among children of the three groups were found with respect to individua l and family characteristics. Predisposing factors The age distributions of the children was si gnificantly different among the three insurance groups (p=0.007). Older children (1 3 age group) were more likely, and younger children (0 5 age group) were less likely, to lose Medi caid/SCHIP coverage and become uninsured. Children younger than 5 year-old were slightly more likely to transi tion from Medicaid/SCHIP to private insurance. The children of the three in surance groups were similar in term of gender distribution. There was obvious difference in racial components among the three groups. Relatively, African American children were mo re likely to stay in Medicaid/SCHIP, while Hispanic children were more likely to lose Medicaid/SCHIP coverage and become uninsured, and white children were more likely to transition from Medicaid/SCHIP to private insurance. In terms of geographic area, the MP group had a higher percentage of rural children than the MM and MU groups. Children in the South and the West were more likely to lose Medicaid/SCHIP coverage and become uninsur ed, and children in the Midwest were more likely to transition from Medicaid/SCHIP to private insurance. Children in a single parent or no parent family were more likely to stay in Medicaid/SCHIP, while children in a two-parent family were more likely to drop out from Medicaid/SCHIP. Children whose mom was 35 year-o ld or older were more likely to lose Medicaid/SCHIP coverage and become unins ured. When the childs parents had more education, the child was more likely to disenrol l from Medicaid/SCHIP, and the MP childrens parents had the highest edu cation among the three groups. Change in enabling factors.
74 MM children were more likely to be in househol ds with stable family income, while MU children and MP children were more likely to be in households with increasing income. Generally most of the childrens parents were always employed in all three groups. Relatively, the MM children were more likely to be in hous eholds with always-unemployed parents, while MU and MP children were more likely to be in households with always-employed parents. The insurance status of MU and MP childrens parent s were more likely to have experienced change during the study time period. Change in need factors. Surprisingly, no significant di fference was found in change of general health status among children of three insurance groups. MU and MM children were very close in the percentages of experiencing worse, no cha nge, and better general health status. Relatively, the general health stat us of MP children was slightly less likely to get worse, and more likely to get better (MP versus MM: p=0.069). Another measure of health change, how th e number of medical conditions changed, was significantly different among the th ree groups. MU children tended to experience no change in or a drop in number of medical conditions, whil e MP children tended to experience either an increase or a drop in numb er of medical conditions. Environmental factors. Between 1996 and 2000, children were more likel y to disenroll from Medicaid/SCHIP. Relatively, between 2000 and 2005, children we re less likely to disenroll from Medicaid/SCHIP. In summary children of those characteristics were more likely to lose Medicaid/SCHIP coverage and become uninsured: older (13 ye ar old) children, Hisp anic children, children
75 living in the south or the west, children livi ng with both parents, ch ildren whose mom was 35 year-old or older, children w hose parents highest education was less than high school or any college, children whose parents insuran ce status was changing, ch ildren whose parents were always employed and/or family income in creases, or children w hose health condition was not getting worse. Children of those characteristics were mo re likely to transition from Medicaid/SCHIP to private in surance: young (0) children, non-Hispanic white children, children living in rural and/or the Midwest area, child ren living with both parents, children whose parents had at least high school level educ ation, children whose pare nts insurance status changed, children whose parents were always employed and/or family income increases, or children whose health condition changed (better/worse). Genera lly, compared with the year 1996, children were less likely to disenroll from Medicaid/SCHIP during the year 2000 2005. The findings above can help to identi fy children who have higher risk of Medicaid/SCHIP disenrollment. Medicaid/SCHIP Disenrollment and Change in Total Expenditures Bivariate Analysis The weighted percentages of children expe riencing each type of change in total expenditures for the three insurance groups are pr esented in Table 4-3. Overall, the percentages of children who experience decreased, no cha nge, and increased tota l expenditures after insurance status change are 43.21%, 25.69%, an d 31.10%, respectively. Compared with the change among MM children, the total expenditu res of MU children were more likely to decrease or to not change, and the expenditure s of MP children were more likely to change (decrease or increase) after disenr ollment from Medicaid/SCHIP. The amount of expenditures in the pre-cha nge round and the post-change round and the change over the two rounds for each insurance group is given in Table 4-4. On average,
76 compared with the expenditures in the pre-ch ange round, the expenditure s in the post-change round were $58.37 less among MM children, $496.88 less (decrease d 80%) among MU children, and $89.22 more among MP children. Multivariate Analysis The adjusted relationship between Medicaid/S CHIP disenrollment and change in total expenditures was obtained by running a three-part model within the general DiD specification. Three regressions were run to predict the change in total expenditures for each insurance group: a multinomial logit regression was used to predict the direction of change, followed by two linear regression equations to predict the amount of change (given corresponding type of change) for each insurance group. The results of the regressions are presented in Table 4-5 and Table 4-6. Part I The first part is a multinomial logit regr ession predicting the pr obability of having negative change, zero change, or positive change in expenditures for each insurance group. The relative risk ratio (RRR) values their linearized standard erro rs, and p-values of significance tests (t-test) from the regressi on are given in Table 4-5. Each type of change (negative or positive) has a separate set of coefficients. Given that the other variables in the model were held constant, compared with MM children, MU children had half the risk of experiencing increasing expenditures (p<0.001) and about 90 % of the risk of experiencing decreasing expenditures (p=0.474); MP children had about 1 .2 times of risk of experiencing change (increase or decrease) in e xpenditures but the differences were not significant. Significant results in the covariates were hol ding other factors cons tant, older children relatively had lower risk of experiencing change (decrease or increase) in expenditures; nonwhite children were less likely to experience change in expenditures compared with white
77 children; the risk of experiencing increasing ex penditures for rural children was 1.28 times of urban children; relative to child ren in northeast area, children in other areas (especially in the west and Midwest) were less likely to experience change in expenditures. Model fit test The commands for goodness of fit assessments, such as likelihood ratio test, Hosmer-Lemeshow (H-L) test, R2, and Akai kes Information Crite rion (AIC), cannot be obtained with svy analysis in Stata. Del measure developed by Hildebrand and colleagues (Hildebrand, Laing, and Rosenthal 1977) was cal culated. Del measure for the multinomial logit regression is 0.20813, which was very close to the Del measures for multinomial probit (0.207987) and higher than ordered logit regressi on (0.087). Considering the time needed (more than two and half months) to run multinomial probit for this study, multinomial logit was the best method for the study and was used in the final prediction. Part II The second part was a linear model pred icting the amount of negative change in expenditures given negative change for each in surance group. Both log-transformed OLS and GLM were conducted for this part. The results from the two methods were very close. Log-transformed OLS regression The coefficients, their linear ized standard errors, and pvalues of significance tests (t -test) from the OLS regression are given in Table 4-6. The dependent variable was multiplied with -1 a nd then log transformed to meet the normality distribution requirement. OLS re gression results showed that, given negative change in expenditures, MU children had smaller scale of change than MM children but the relationship was not significant (p=0.955) and MP children ha d a larger scale of ch ange than MM children (p=0.006). OLS assumption and model fit tests After log-transformation, the kurtosis was 3.942 and the skewness was 0.345. Figure 4-1 is the kernel de nsity plot. It was very close to the normal
78 distribution line. Figure 4-2 is the standardized normal pr obability (P-P) plot (pnorm command). It is sensitive to deviation from norma lity in the middle range of data. Figure 4-3 is the quantiles of residuals against quantiles of normal distribution pl ot (qnorm co mmand). It is sensitive to non-normality near the tails. Those fi gures show that the re siduals are close to a normal distribution. The H-L test of lineari ty indicates good calibration (p=0.6607). Whites test and Parks test detected heteroskedastici ty. But the robust option cannot be used with svy command in Stata to correct for heterosk edasticity. Therefore, subgroup smear estimators for each insurance group were calculated after OLS regression. The residual-prediction plot (Figure 4-4) shows that the model fit well. M odel goodness-of-fit statistics such as the HosmerLemeshow test and AIC, cannot be obtained after a svy regression command in Stata. Part III The third part was another linear model predicting the amount of positive change in expenditures given positive change for each insurance group. Both log-transformed OLS and GLM regressions were conducted for this part as well. The results from the two methods were very close. Log-transformed OLS regression The coefficients, their linear ized standard errors, and pvalues of significance tests (t-t est) from the OLS regression are presented in columns 5-7 of Table 4-6. The dependent variable was log tr ansformed. OLS regression shows that, given positive change in expenditures, MU children had a slightly smaller scale of increase than the change among MM children but the relationship was not significant (p=0 .935) and MP children had larger scale of increase in e xpenditures than MM children (p=0.001). OLS assumption and model fit tests After log-transformation, the kurtosis was 4.022 and the skewness was 0.207. The Kernel density plot (Figure 4-5), pnorm (Figure 4-6) and qnorm (Figure 4-7) show that the residuals are close to nor mal distribution. The linearity
79 assumption was met according to H-L test of linearity (p=0.06). Heteroskedasticity was detected and therefore, subgr oup smear estimators for each insurance group were calculated after log-transformed OLS regression. The residu al-fitted plot (Figure 4-8) shows that the model was well calibrated to the data. Bootstrapping predication Using equation 3-4, 3-5, and 36, all the predictions from the above three parts were combined together to obtain the predicted amount of change in expenditures for each insurance group. These estimates were then bootstrapped (1000 iterations) to obtain the 95% c onfidence intervals to determine whether the difference was statistically significant. The bootstrap results are presented in Table 4-8. Generally, health care expend itures decreased significantl y among children in the MM group and children in the MU group, while no si gnificant change was seen in the MP group. Compared with MM children, the expenditures for MU children were averagely $227.74 less in the round being uninsured after Medicaid/SCHIP disenrollment than th e expenditures in the round in Medicaid/SCHIP before disenrollmen t (p<0.05). Compared with MM children, the expenditures for MP children were averagely $71.79 more in the round w ith private insurance after Medicaid/SCHIP disenrollment than the expenditures in the r ound in Medicaid/SCHIP before disenrollment, but the difference was not significant. Compare OLS and GLM The significance levels of coefficients for insurance variables in GLM regression results were similar with those in OLS regression. That is, holding other factor s constant, MP children were experiencing significantly larg er scale of change (increase or decrease) than MM children, given any change in expenditures; MU children experienced larger scal e of drop or smaller scale of increase in expenditures than MM childre n (p>0.05), given the type of change (Table
80 4-7). The modified Park test shows that gamma distribution fit the data the best and the Link test shows the linear relationship (log link) is correct. The residual-fitted plot (Figure 4-9 and Figure 4-10) shows that the model fit most of the data well with a few extreme values. The extreme values can be explained by the high h ealth care expenditures occurring among children with serious disease, information which cannot be captured in the model. Those plots show that log-transformed OLS fit the da ta better than GLM with ga mma family and log link. The bootstrapping means and S.D. of the 3-part m odel (GLM) are smaller than those from OLS 3part model but the significance is the sa me between OLS results and GLM results. Sensitivity test: Including Ykb at the RHS of the equations. When the lagged outcome measures, Yb, was included as an inde pendent variable in the regression analysis, there were some extreme values (as large as 7.90*e+13 for the expenditure change in MU group) in the predictions ( both OLS and GLM). When a log-transformed Ykb was used instead of Ykb at RHS, there were less extreme values in the predictions, but the residuals of predictions were still much larger than those obtained without including any form of Ykb at RHS. Therefore, the original general model specification (equa tion 3-1) was used in the study. Summary : When controlling for other factors, th e effect of MU disenrollment on total expenditures was significant. When children left Medicaid/SCHIP and became uninsured, their total health care expenditures dropped. This indi cates that losing Medi caid/SCHIP coverage may lead to less health care utilization for those children. Medicaid/SCHIP Disenrollment and Change in Well-child Care Use Bivariate Analysis Generally, the percentages of children who had negative change, no change, and positive change in the number of wellchild visits after an insura nce status change were 19.33%, 69.43%, and 11.24%, respectively (Table 4-10). Compared with MM children, MU children
81 were more likely to have negative change, a nd MP children were more likely to have any change (either decrease or increase) in the num ber of well-child visits after Medicaid/SCHIP disenrollment. The amount of change in the number of wellchild visits for each insurance group is presented in Table 4-11. On average, compared with the measure in th e pre-change round, the number of well-child visits in the postchange round was 0.06 le ss (decreased 31%) among MM children, 0.17 less (decreased 68%) among MU children, and 0.18 less (decreased 55%) among MP children. Multivariate Analysis The adjusted relationship between Medicaid/S CHIP disenrollment and the change in the number of well-child visits was obtained by r unning a three-part model within the general DiD specification. The first part was a multinomial logit regression, the same as with the expenditure analysis. The second and third parts are GLM regr essions, instead of OLS regressions, because the normality assumption of OLS cannot be met afte r transformations of the dependent variable. The results of the regressions are presented in Table 4-12 Part I The first part was a multinomial logit regression predicting the probability of having each type of change in number of well-ch ild visit for each insurance group. The results from the regression are given in Table 4-12. Hold ing other variables constant, compared with MM children, MU children had 1.26 times of the risk of having fewer well-child visits (p=0.132) and 73.7% of the risk of having more well-child visits (p=0.063); MP children had about 1.45 times of risk of having fewer well-chil d visits (p=0.004) and 1 .32 times of risk of having more well-child visits (p=0.111).
82 Model fit test Del measure for the multinomial logit regression was 0.135257, which was very close to the Del measures for multinomial probit (0.135756) and higher than ordered logit regression (-0.00022). Therefore, multinomial logit was a better fit for the data and was used in the final prediction. Part II The amount of negative change in number of well-child visit was predicted with a GLM regression of gamma family and log link. Th e results are presented in the column 2-4 of Table 4-13. Given negative change s, the absolute amount of cha nge in number of well-child visit among MU children (p=0.062) and MP children (p<0.001) wa s larger than the change among MM children of the same individual and family characteristics. That is, when childrens well-child visits decreased, MU and MP childre n had larger decrease after Medicaid/SCHIP disenrollment, relative to MM children. GLM model fit tests The Modified Park test was used to determine the family distribution of the data and it shows that gamma distribution fit the data the best. Th e Link test showed the linear relationship (log link) wa s correct. The residual-fitted plot showed a cloud around zero residual (Figure 4-11). The model fitted well with the data. Part III The amount of positive change in the nu mber of well-child visit was predicted with another GLM regression of gamma family and log link. The results are presented in the column 5-7 of Table 4-13. Given positive changes, the absolute amount of change in number of well-child visit among MU childre n (p=0.316) and MP children (p =0.097) was smaller than the change among MM children of the same indivi dual and family characteristics. That is, if childrens well-child visits increased, MU a nd MP children had smaller (though not significant) increase after Medicaid/SCHIP disenrollment, relative to MM children.
83 GLM model fit tests The Modified Park test showed that gamma distribution fitted the data the best. The Link test showed the linear relationship (log link) was correct. The residualfitted plot showed a cloud around zero residual (F igure 4-12). The model fitted well with the data. Bootstrapping predication The bootstrapping results ar e presented in Table 4-14. Generally, well-child visits decr eased significantly among children in all of the three insurance groups: 0.040 among MM group, 0.160 among MU group, and 0.154 among MP group. The decrease in well-child care use generally may be explained by the age effect. Older children have less well-child care need. The differences between MM children and MU children or MP children were significant. Compared with MM children, on average MU children had 0.120 fewer visits and MP children had 0.114 fewer visits during the post-change round than during the pre-change round. Summary. After children leave Medicaid/SCHI P, they had less well-child care utilization. The multivariate anal ysis showed that, when controlling for other individual and family factors, the effect of Medicaid disenro llment on well-child care use was significant no matter whether those children obtain ed private insurance or not. Medicaid/SCHIP Disenrollment and Change in Physician Visits Bivariate Analysis Generally, the percentages of children w ho had a negative change, no change, and a positive change in the number of physician visi ts after insurance status change were 32.51%, 43.44%, and 24.05%, respectively (Table 4-15). Compared with MM children, MU children were more likely to have no change, and MP chil dren are more likely to have changes (either decrease or increase) in the num ber of physician visits after Medi caid/SCHIP disenrollment.
84 The amount of change in number of physician visits for each insurance group is presented in Table 4-16. On average, compared with the measure in the pre-cha nge round, the number of physician visits in post-change round was 0.03 less (decreased 4%) among MM children, 0.31 less (decreased 41%) among MU children, and 0.22 less (decreased 16% ) among MP children. Multivariate Analysis Part I The first part was a multinomial logit regression predicting the probability of having each type of change in the number of phy sician visits for each insurance group (Table 417). Holding other variables constant, the risk of experiencing decreasi ng physician visits for MU children was 77.1% of the risk for MM child ren (p=0.07), and the chance for MU children to experience increasing physician visits wa s only 45% of that for MM children (p<0.001). Given other factors equal, MP children had about 1.43 times of risk of experiencing decreasing physician visits (p=0.003) and 1.27 times of risk of having more physician visits (p=0.07). That is, compared with MM children, MU children were less likely to experience a change and MP children were more likely to experience a change in the number of physician visits. Model fit test. Del measure for the multinomial logit regression was 0.219215, which was very close to the Del measures for multinomial probit (0.218574) and higher than ordered logit regression (0.093). Therefore, multinom ial logit was a better fit for the data and was used in the final prediction. Part II The amount of negative change in the nu mber of physician visit was predicted with a GLM regression of gamma family and log link (Table 4-18). Given negative change, MU children and MP children tended to experience a larger scale of a dr op in the number of physician visits (p<0.05) than MM children. All of the goodness-of-fit tests used in above well-
85 child visit analysis were perfor med, and those tests showed that the model fitted well with the data. Part III The amount of positive change in the nu mber of physician visit was predicted with another GLM regression with inverse Gaussian family and log link (Table 4-18). Given positive change, MU children experienced a smaller scale of increase in physician visits than MM children (p=0.820); MP children e xperienced a larger scale of increase in physician visits than MM children (p=0.008), holding other fact ors constant. The goodnessof-fit tests showed that the model fitted well with the data. Bootstrapping predication The bootstrapping results ar e presented in Table 4-19. Compared with those children staying in Medi caid/SCHIP, children who lost Medicaid/SCHIP coverage and became uninsured averagely expe rienced on average 0.277 fewer physician visits in the round after disenrollment than in the round before disenrollment. The difference in physician visits between children transitioned to private insurance and those staying in Medicaid/SCHIP was not significant. Summary After adjusting for other individual and family characteristics, losing Medicaid/SCHIP coverage without any insura nce led to a decrease in ambulatory care use (measured with physician visits) among children. Medicaid/SCHIP Disenrollment and Change in ER Visits Bivariate Analysis Generally, the percentages of children who experienced a decrease, no change, and an increase in ER visits after an insuran ce status change were 7.23%, 87.28%, and 5.49%, respectively (Table 4-20). Compared with MM child ren, MU children were slightly more likely to experience no change or an increase, and MP children were more likely to experience an increase, in ER visits after Me dicaid/SCHIP disenrollment.
86 The amount of change in the number of ER vis its for each insurance group is presented in Table 4-21. On average, compared with the meas ure in the pre-change round, the number of ER visits in the post-change round was 0.001 le ss among MM children, 0.028 less (decreased 27%) among MU children, and 0.030 more (inc reased 30%) among MP children. Multivariate Analysis The results of the regression analysis are pr esented in Table 4-22 (Part I) and Table 4-23 (Part II and Part III). Part I Multinomial logit regression (Table 422) shows that, holdi ng other variables constant, the risk of experiencing decreasing ER visits for MU children was 84.1% of the risk for MM children (p=0.411), and th e chance for MU children to e xperience increasing ER visits was 1.13 times of that for MM children (p=0.593) Given other factors equal, MP children had about 87.2% of risk of experien cing decreasing ER visits (p=0.472) and 1.44 times of risk of having more ER visits (p=0.042). Model fit tests. The fitted probabilities all fell in no change category which means the model fit was poor. Multinomial probit and ordere d logit had the same goodness-of-fit. It may be caused by the small number of negative change and positive change in the sample. In the MU group, there were only 51 children in negative change and 39 children in positive change category. Part II The amount of negative change in ER vi sit was predicted w ith a GLM regression of gamma family and log link (Table 4-23). Given negative change, MU children and MP children tended to experience a larger scale of a drop in the number of ER visits (p<0.05) than MM children. The goodness-of-fit test s showed that the model fitted well with the data.
87 Part III The amount of positive change in the numb er of ER visits was predicted with another GLM regression with gamma family and log link (Table 4-23). Given positive change, MU children experienced a smalle r scale of increase in ER vi sits than MM children (p=0.121); MP children experienced a larger scale of increase in ER vi sits than MM children (p=0.888), holding other factors constant. The goodness-of-fit tests showed that the model fitted well with the data. Bootstrapping predication The bootstrapping results are presented in Table 4-24. MP children experienced more ER visits after dise nrolling from Medicaid/SCHIP, but this trend was not significant compar ed with MM children. Summary After adjusting for other individual and family characteristics, losing Medicaid/SCHIP coverage and going without any insurance did not influence the ER care use significantly among children. Medicaid/SCHIP Disenrollment and Change in Number of Hospitalizations Bivariate Analysis Generally, the percentages of children who experienced a decrease, no change, and an increase in hospitalizations after an insurance status ch ange were 1.89%, 97.17%, and 0.94%, respectively (Table 4-25). Compared with MM child ren, MU children were slightly more likely to experience no change or a decrease, and MP children were more likely to experience change (increase or decrease) in hos pitalizations after Medicaid/ SCHIP disenrollment. The amount of change in numb er of hospitalizations for each insurance group is presented in Table 4-26. Compared with the measure in the pre-change round, the average number of hospitalizations in the post-change round wa s 0.005 less (decreased 28%) among MM children, 0.022 less (decreased 79%) among MU children, and 0.017 less (decreased 55%) among MP children.
88 Multivariate Analysis The results of the regression analysis are pr esented in Table 4-27 (Part I) and Table 4-28 (Part II and Part III). Part I Multinomial logit regression (Table 427) shows that, holdi ng other variables constant, the risk of experiencing decreasing ho spitalizations for MU ch ildren was 1.22 times of the risk for MM children (p=0.605), and the chan ce for MU children to experience increasing hospitalizations was only 42.2% of that for MM children (p=0.129). Given other factors equal, the risk of experiencing decr easing hospitalizations for MP children was 1.32 times (p=0.418), and the risk of experiencing increasing hos pitalizations for MP children was 1.07 times (p=0.879) of the risk for MM children. None of a bove differences was statistically significant. Model fit tests. The fitted probabilities all fell in no change category which means the model fit was poor. Multinomial probit and ordere d logit had the same goodness-of-fit. It may be caused by the small number of negative change and positive change in the sample. In the MU group, there were only 17 children in the negative change and 4 children in the positive change category. Part II The amount of negative change in hospi talizations was pred icted with a GLM regression of gamma family and log link (Tab le 4-28). Given negative change, MU children and MP children tended to experience larger scale of drops in hospita lizations (p>0.05) than MM children. The goodness-of-fit test s showed that the model fitted well with the data. Part III The amount of positive change in number of hospitalizations was predicted with another GLM regression with gamma family and log link (Table 4-28). Given positive change, MU children and MP children experienced larger scale of increases in hospitalizations than
89 MM (p>0.05), holding other fact ors constant. The goodness-of-fit tests showed that the model fitted well with the data. Bootstrapping predication The bootstrapping results are pr esented in Table 4-29. None of the differences between in surance groups was significant. Summary After adjusting for other individual and family characteristics, losing Medicaid/SCHIP coverage and going without any insurance did not have an observed effect on inpatient care utilization among ch ildren during the study time period. Medicaid/SCHIP Disenrollment and Ch ange in Prescription Drug Use Bivariate Analysis Generally, the percentages of children who had negative change, no change, and positive change in the number of prescr iption drugs used after an insu rance status change were 27.62%, 52.54%, and 19.83%, respectively (Table 4-30). Compared with MM children, MU children were more likely to have no change, and MP chil dren were more likely to experience changes (especially an increase) in the number of prescription drugs used after Medicaid/SCHIP disenrollment. The amount of change in the number of prescription drugs used for each insurance group is presented in Table 4-31. Compared with the measure in the pre-cha nge round, the number of prescriptions in the post-ch ange round was 0.03 less (decr eased 4%) among MM children, 0.327 less (decreased 42%) among MU children, and 0.361 more (increased 20%) among MP children. Multivariate Analysis The results of the regression analysis are pr esented in Table 4-32 (Part I) and Table 4-33 (Part II and Part III).
90 Part I The multinomial logit regres sion (Table 4-32) shows that holding other variables constant, for MU children the risk of experien cing decreasing prescr iption drug use was 70.0% (p=0.011), and the chance for them to experi ence increasing prescription drug use was 55.2% (p<0.001) of that for MM children. Given othe r factors equal, MP children had about 1.065 times of risk of experiencing decreasing pres cription drug use (p=0.617) and 1.435 times of risk of having more prescription drug use (p=0.003) Compared with MM children, MU children were less likely to experience a change and MP children were more likely to experience an increase in prescription drug use. Model fit test Del measure for the multinomial logit regression was 0.235763, which was very close to the Del measures for multinomial probit (0.235911) and higher than ordered logit regression (0.017997). Therefore, multinomial logit wa s a better fit for the da ta and was used in the final prediction. Part II The amount of negative change in pres cription drug use was predicted with a GLM regression of gamma family and log link (T able 4-33). Given negative change, both MU children (p=0.119) and MP children (p<0.001) experi enced a larger scale of drop in prescription drug use than MM children. The goodne ss-of-fit tests showed that th e model fitted well with the data. Part III The amount of positive change in prescription drug use was predicted with another GLM regression with inverse Gaussian family and log link (Table 4-33). Given positive change, MU children experienced a smaller scal e of increase in prescr iption drug use than MM children (p=0.216); MP children expe rienced a larger scale of incr ease in prescription drug use than MM children (p=0.007), holding other fact ors constant. The goodnessof-fit tests showed that the model fitted well with the data.
91 Bootstrapping predication The bootstrapping results are pr esented in Table 4-34. After leaving Medicaid/SCHIP, those children who had no insurance experienced significantly less prescription drug use (0.235 less on average) and those children who obtained private insurance experienced significan tly more prescription drug use (0.330 more on average), given other individual and family factors equal. Comp ared with MM children, the drop in prescription drug use among MU children was still significan t, and the increase in prescription drug use among MP children was not significant a ny more, controlling for other factors. Summary After adjusting for other individual and family characteristics, losing Medicaid/SCHIP coverage and becoming uninsur ed resulted in less prescription drug use among children; transitioning from Medicaid/S CHIP to private insurance led to more prescription drug use but the change was not si gnificant when compared with those children staying in Medicaid/SCHIP.
92 Table 4-1. Study sample by insurance group Number of ChildrenPercentage (%) MM 10,86585.65 MU 9327.35 MP 8897.01 Total 12,686100.00
93 Table 4-2. Individual and family character istics of study sample by insurance group Variables MM (N=10865) % MU (N=932) % MP (N=889) % Total (N=12686) % Chi2 test p-value Predisposing Factors Age 0 11.249.9812.2611.22 2 26.5223.3928.0126.39 6 40.8540.1338.4740.63 13 21.4026.5021.2621.76 0.007 Gender Female 50.1550.1147.1349.94 Male 49.8549.8952.8750.06 Race 0.222 Non-Hispanic white 24.3724.2540.7225.51 Non-Hispanic African American 28.6216.7421.3727.24 Hispanic 42.4155.9032.6242.72 Other 4.593.115.294.53 <0.001 MSA Rural 22.1322.4229.9222.69 Urban 77.8777.5870.0877.31 <0.001 Region Northeast 16.6110.9417.3216.25 Midwest 14.2211.1628.8015.02 South 38.2944.4230.3738.18 West 30.8833.4823.5130.55 <0.001 # Parents living with 0 8.566.447.098.30 1 45.8235.7335.7744.37 2 45.6257.8357.1447.33 <0.001 Age of mom 24 57.1553.3358.4956.96 35 or over 42.8546.6741.5143.04 0.049 Parents education Less than HS 37.0838.5218.3435.87 Completed HS 52.9647.7562.5453.25 Any college 9.9613.7319.1210.88 <0.001 Enabling Factors Change in parents insurance coverage No change 92.0478.0075.3789.84 Change 7.9622.0024.6310.16 <0.001 Change in parents employment status Employed to unemployed 2.262.902.592.33 Always unemployed 23.318.155.5120.95 Unemployed to employed 184.108.40.2065.34 Always employed 69.2082.8385.9471.38 <0.001
94 Table 4-2. Continued Variables MM (N=10865) % MU (N=932) % MP (N=889) (%) Total (N=12686) (%) Chi2 test p-value Change in family income Decrease 17.6517.2718.6717.70 No change 60.5250.7549.8359.05 Increase 21.8331.9731.5023.25 <0.001 Need Factors Change in general health status Worse 25.8425.3222.3825.56 No change 49.1949.0350.8449.30 Better 24.9725.6426.7725.15 0.242 (m/p: 0.069) Change in # medical conditions Decrease 38.7839.7044.0939.22 No change 40.2843.3534.2040.08 Increase 20.9516.9521.7120.71 <0.001 Environmental Factors Time 1996 80.9212.067.02100.00 1997 81.0610.928.02100.00 1998 79.209.8810.91100.00 1999 81.8210.847.34100.00 2000 86.557.765.69100.00 2001 87.995.026.99100.00 2002 87.005.007.00100.00 2003 88.245.706.05100.00 2004 88.994.936.07100.00 <0.001 Note: Unweighted. Percentages are column percen tage except time variab le (row percentage).
95 Table 4-3. Type of change in total expenditures by insurance group (weighted) MM (%) MU (%) MP (%) Total (%) Decreased 42.6247.6145.16 43.21 No change 25.8232.0118.88 25.69 Increased 31.5620.3835.96 31.10 Total 100.00100.00100.00 100.00 Table 4-4. Descriptive statisti cs on amount of total expenditu res by insurance group (weighted, 2005 U.S. $/round) MM MU MP Baseline round 360.63 (41.82)625.23 (310.10)684.77 (141.84) Following round 302.27 (37.05)128.35 (23.27)773.99 (162.31) Change over two rounds -58.37 (49.46)-496.88 (307.10)89.22 (210.73) Mean (S.D.)
96 Table 4-5. Multinomial logit re gression predicting probability of experiencing each type of change in total expenditures (Part I) Relative risk of having negative change vs. no change Relative risk of having positive change vs. no change RRR s.e. p-value RRR s.e. p-value MM Reference MU 0.9040.1270.4740.506 0.073 0 MP 1.2190.1690.1541.262 0.169 0.082 Age 0 Reference 2 0.3810.03900.498 0.049 0 6 0.2680.02900.369 0.039 0 13 0.2630.03400.303 0.040 0 Gender Male Reference Female 1.0200.0610.7451.087 0.069 0.188 Race Non-Hispanic white Reference Non-Hispanic African American 0.5730.05400.460 0.050 0 Hispanic 0.7490.0710.0020.679 0.073 0 Other 0.7430.1130.0520.693 0.134 0.058 MSA Urban Reference Rural 0.9850.0880.8661.280 0.124 0.011 Region Northeast Reference Midwest 0.8340.1040.1450.764 0.098 0.036 South 0.9030.1010.3620.857 0.099 0.182 West 0.7260.0890.0090.675 0.080 0.001 # Parents 2 Reference 1 1.1750.1040.0691.258 0.117 0.014 0 1.2570.1750.1011.030 0.150 0.842 Age of mom 24-34 Reference 35 or over 1.0170.0930.8571.187 0.109 0.062 Education of parents Less than HS Reference Completed HS 1.0470.0820.561.123 0.094 0.163 Any college 1.0300.1430.831.312 0.183 0.052 Family income change Increase Reference No change 1.0410.0960.661.096 0.102 0.322 Decrease 1.0530.1140.631.038 0.119 0.746
97 Table 4-5. Continued Relative risk of having negative change vs. no change Relative risk of having positive change vs. no change RRR s.e. p-value RRR s.e. p-value Parent employment status change Always employed Reference Employed to unemployed 1.4400.3540.1380.966 0.2450.892 Unemployed to employed 0.7470.1120.0510.944 0.1660.743 Always unemployed 0.8780.0790.150.913 0.0900.355 Parent insurance coverage change No change Reference Change 0.9090.1200.4691.053 0.1460.708 Change in general health status Better Reference No change 0.8820.0700.1120.927 0.0800.38 Worse 0.9520.0940.6180.881 0.0890.211 Change in # medical conditions Decrease Reference No change 0.2070.01500.334 0.0270 Increase 0.2550.02300.612 0.0560 Year 1996 Reference 1997 1.2910.2270.1471.371 0.2380.069 1998 1.4540.2550.0331.107 0.2100.593 1999 1.2270.2150.2441.552 0.2760.014 2000 0.8650.1600.4351.028 0.1940.884 2001 0.9830.1380.9020.993 0.1440.961 2002 0.9090.1310.5080.878 0.1310.382 2003 0.9320.1260.6040.854 0.1190.257 2004 0.7660.1060.0550.833 0.1250.224
98 Table 4-6. Ordinary least square (OLS) regr ession predicting amount of change in total expenditures given negative or positiv e change (Part II and Part III) Amount of negative change (n=5295) Amount of positive change (n=3818) Coefficient (s.e.) p-value Coefficient (s.e.) p-value MM Reference MU -0.006 (0.114)0.955-0.011 (0.134) 0.935 MP 0.322 (0.116)0.0060.426 (0.127) 0.001 Age 0 Reference 2 -0.157 (0.090)0.0810.093 (0.101) 0.357 6 -0.108 (0.079)0.1740.271 (0.101) 0.007 13 0.134 (0.099)0.1750.504 (0.120) <0.001 Gender Male Reference Female 0.058 (0.052)0.264-0.120 (0.064) 0.060 Race Non-Hispanic white Reference Non-Hispanic African American -0.234 (0.080)0.003-0.397 (0.085) <0.001 Hispanic -0.241 (0.085)0.005-0.388 (0.075) <0.001 Other -0.204 (0.118)0.082-0.144 (0.167) 0.391 MSA Urban Reference Rural -0.043 (0.083)0.604-0.038 (0.085) 0.656 Region Northeast Reference Midwest -0.074 (0.109)0.493-0.074 (0.120) 0.540 South -0.116 (0.082)0.1580.039 (0.102) 0.702 West -0.027 (0.098)0.783-0.013 (0.101) 0.895 # Parents 2 Reference 1 0.106 (0.076)0.165-0.003 (0.076) 0.971 0 0.265 (0.107)0.0130.029 (0.128) 0.822 Age of mom 24 Reference 35 or over 0.073 (0.076)0.3380.094 (0.071) 0.188 Education of parents Less than HS Reference Completed HS 0.018 (0.068)0.7960.008 (0.070) 0.913 Any college 0.378 (0.103)<0.0010.167 (0.114) 0.142 Family income change Increase Reference No change -0.113 (0.083)0.1760.027 (0.079) 0.736 Decrease -0.055 (0.092)0.553-0.010 (0.107) 0.929 Parent employment status change Always employed Reference
99 Table 4-6. Continued Amount of negative change (n=5295) Amount of positive change (n=3818) Coefficient (s.e.) p-value Coefficient (s.e.) p-value Employed to unemployed 0.409 (0.279)0.1430.102 (0.249) 0.683 Unemployed to employed 0.158 (0.145)0.276-0.052 (0.147) 0.724 Always unemployed 0.042 (0.091)0.6420.111 (0.095) 0.243 Parent insurance coverage change No change Reference Change -0.175 (0.098)0.0750.072 (0.137) 0.600 Change in general health status Better Reference No change -0.113 (0.067)0.089-0.153 (0.077) 0.047 Worse 0.099 (0.074)0.1830.036 (0.096) 0.711 Change in # medical conditions Decrease Reference No change -0.471 (0.066)<0.001-0.273 (0.075) <0.001 Increase -0.370 (0.073)<0.001-0.140 (0.088) 0.113 Year 1996 Reference 1997 0.418 (0.133)0.0020.333 (0.150) 0.026 1998 0.527 (0.130)0.0000.528 (0.188) 0.005 1999 0.307 (0.179)0.0870.581 (0.155) 0.000 2000 -0.111 (0.149)0.4570.301 (0.163) 0.065 2001 0.213 (0.113)0.0610.384 (0.151) 0.011 2002 0.251 (0.124)0.0420.339 (0.148) 0.022 2003 0.259 (0.115)0.0250.192 (0.144) 0.182 2004 0.090 (0.118)0.4460.469 (0.132) <0.001
100 Table 4-7. Generalized linear regression (G LM) predicting amount of change in total expenditures given negative or positiv e change (Part II and Part III) Amount of negative Change (n=5295) Amount of positive change (n=3818) Coef. s.e. p-value Coef. s.e. p-value MM Reference MU 0.4130.2760.135-0.293 0.188 0.118 MP 0.4590.1520.0030.609 0.210 0.004 Age 0 Reference 2 -0.3120.1440.031-0.262 0.156 0.095 6 -0.6050.1340.000-0.063 0.169 0.710 13 -0.5340.1600.0010.292 0.195 0.135 Gender Male Reference Female 0.0830.0930.372-0.058 0.097 0.548 Race Non-Hispanic white Reference Non-Hispanic African American -0.1570.1240.203-0.426 0.119 0.000 Hispanic -0.2110.1210.081-0.117 0.137 0.393 Other -0.4850.1500.0010.055 0.219 0.802 MSA Urban Reference Rural 0.0260.1180.8260.242 0.132 0.067 Region Northeast Reference Midwest -0.2540.1700.135-0.093 0.183 0.613 South -0.3210.1480.030-0.053 0.155 0.731 West -0.2280.1600.155-0.249 0.158 0.115 # Parents 2 Reference 1 -0.0320.1170.786-0.163 0.113 0.150 0 0.2350.1700.167-0.144 0.187 0.443 Age of mom 24 Reference 35 or over 0.5520.1390.0000.210 0.124 0.091 Education of parents Less than HS Reference Completed HS -0.0940.1090.3850.085 0.106 0.419 Any college 0.3600.2000.0720.335 0.174 0.054 Family income change Increase Reference No change 0.0760.1160.5120.057 0.104 0.583 Decrease -0.0220.1380.8720.071 0.153 0.642 Parent employment status change Always employed Reference
101 Table 4-7. Continued Amount of negative Change (n=5295) Amount of positive change (n=3818) Coef. s.e. p-value Coef. s.e. p-value Employed to unemployed 0.9520.4120.0210.278 0.281 0.322 Unemployed to employed 0.7980.2970.0070.012 0.205 0.955 Always unemployed 0.2280.1320.0860.228 0.125 0.069 Parent insurance coverage change No change Reference Change -0.3620.1640.0270.129 0.224 0.566 Change in general health status Better Reference No change -0.2270.1380.099-0.178 0.107 0.095 Worse -0.1420.1430.3230.164 0.132 0.213 Change in # medical conditions Decrease Reference No change -0.9370.1060.000-0.571 0.121 0.000 Increase -0.8540.1100.000-0.429 0.126 0.001 Year 1996 Reference 1997 0.1640.2270.4710.265 0.207 0.200 1998 0.1800.2070.3861.153 0.325 0.000 1999 0.0890.2350.7070.784 0.205 0.000 2000 0.0710.2590.7850.038 0.201 0.850 2001 0.3250.2470.1880.774 0.233 0.001 2002 0.1280.2210.5620.320 0.196 0.103 2003 -0.0160.1980.936-0.067 0.172 0.696 2004 0.0350.2110.8680.315 0.185 0.088 Note: GLM with gamma family and log link
102 Table 4-8. Bootstrapping pred iction of expenditure change after three-part model (OLS) N Mean S.D. 95%CI MM 1000-59.37**20.38(-101.00, -19.57) MU 1000-287.11**145.74(-639.36, -85.03) MP 100012.42125.14(-238.12, 263.21) MU-MM 1000-227.74**146.66(-573.71, -18.47) MP-MM 100071.79127.34(-188.03, 333.47) **Significant at 0.05 level Table 4-9. Bootstrapping pred iction of expenditure change after three-part model (GLM) N Mean S.D. 95%CI MM 1000-59.14**27.50(-113.70, -6.14) MU 1000-324.86**137.25(-605.46, -102.27) MP 1000-20.29113.12(-232.43, 216.03) MU-MM 1000-265.72**133.98(-540.87, -49.79) MP-MM 100038.85115.61(-171.29, 279.95) **Significant at 0.05 level
103 Table 4-10. Type of change in number of well-child visits by in surance group (weighted) MM (%) MU (%) MP (%) Total (%) Decreased 18.5321.5725.2319.33 No change 70.2470.4160.6569.43 Increased 11.238.0214.1211.24 Total 100.00100.00100.00100.00 Table 4-11. Descriptive statisti cs on number of well-child vi sits (visits/round) (weighted) MM MU MP Baseline round 0.180 (0.008) 0.242 (0.032)0.3341 (0.032) Following round 0.196 (0.009)0.128 (0.018)0.264 (0.033) Change over two rounds -0.056 (0.013)-0.165 (0.039)-0.184 (0.043) Mean (S.D.)
104 Table 4-12. Multinomial logit regression predicting pr obability of each type of change in wellchild visits (Part I) Relative risk of having negative change vs. no change Relative risk of having positive change vs. no change RRR s.e. p-value RRR s.e. p-value MM Reference MU 1.261 0.194 0.1320.737 0.121 0.063 MP 1.454 0.187 0.0041.322 0.231 0.111 Age 0 Reference 2 0.279 0.026 00.399 0.043 0 6 0.138 0.013 00.238 0.032 0 13 0.124 0.015 00.252 0.042 0 Gender Male Reference Female 1.096 0.070 0.1511.063 0.085 0.441 Race Non-Hispanic white Reference Non-Hispanic African American 0.937 0.093 0.5140.881 0.112 0.318 Hispanic 1.008 0.093 0.931.001 0.114 0.996 Other 0.790 0.143 0.1910.846 0.194 0.467 MSA Urban Reference Rural 1.082 0.099 0.3890.977 0.118 0.846 Region Northeast Reference Midwest 0.623 0.075 00.554 0.082 0 South 0.582 0.061 00.648 0.090 0.002 West 0.493 0.056 00.550 0.074 0 # Parents 2 Reference 1 1.111 0.106 0.2690.914 0.103 0.426 0 1.181 0.148 0.1820.874 0.156 0.453 Age of mom 24 Reference 35 or over 0.987 0.085 0.8751.090 0.125 0.45 Education of parents Less than HS Reference Completed HS 0.956 0.085 0.6140.990 0.104 0.922 Any college 0.896 0.115 0.3941.308 0.178 0.049 Family Income Increase Reference No change 0.950 0.089 0.5851.020 0.113 0.858 Decrease 0.959 0.110 0.7121.045 0.146 0.755
105 Table 4-12. Continued Relative risk of having negative change vs. no change Relative risk of having positive change vs. no change RRR s.e. p-value RRR s.e. p-value Parent employment status change Always employed Reference Employed to unemployed 1.627 0.422 0.0611.240 0.407 0.512 Unemployed to employed 0.869 0.150 0.4170.916 0.181 0.656 Always unemployed 1.029 0.110 0.7921.091 0.128 0.458 Parent insurance coverage change No change Reference Change 1.029 0.137 0.8331.084 0.163 0.593 Change in general health status Better Reference No change 0.937 0.072 0.4011.091 0.112 0.395 Worse 1.027 0.096 0.7720.879 0.107 0.288 Change in # medical conditions Decrease Reference No change 0.702 0.056 00.729 0.064 0 Increase 0.708 0.067 00.947 0.103 0.615 Year 1996 Reference 1997 1.491 0.224 0.0081.807 0.323 0.001 1998 1.570 0.261 0.0071.602 0.361 0.036 1999 1.077 0.189 0.6741.925 0.379 0.001 2000 1.038 0.191 0.8381.169 0.251 0.466 2001 1.240 0.176 0.1291.563 0.273 0.011 2002 0.958 0.134 0.7581.330 0.234 0.105 2003 1.016 0.143 0.9081.281 0.229 0.165 2004 0.808 0.120 0.1511.132 0.195 0.473
106 Table 4-13. GLM regression predic ting amount of change in well-ch ild visits given negative or positive change (Part II and Part III) Amount of negative change (n=2297) Amount of positive change (n=1370) Coef. s.e. p Coef. s.e. p MM Reference MU 0.121 0.065 0.062 -0.092 0.091 0.316 MP 0.253 0.062 0.000 -0.136 0.082 0.097 Age 0 Reference 2 -0.213 0.051 0.000 0.030 0.059 0.606 6 -0.181 0.058 0.002 0.147 0.071 0.038 13 -0.223 0.076 0.003 0.143 0.086 0.096 Gender Male Reference Female 0.062 0.035 0.080 -0.002 0.044 0.971 Race Non-Hispanic white Reference Non-Hispanic African American 0.006 0.055 0.909 0.018 0.070 0.800 Hispanic -0.020 0.049 0.684 -0.135 0.058 0.019 Other -0.077 0.107 0.475 -0.181 0.072 0.012 MSA Urban Reference Rural 0.057 0.049 0.242 -0.038 0.061 0.536 Region Northeast Reference Midwest 0.032 0.066 0.630 0.056 0.092 0.543 South 0.009 0.057 0.871 -0.006 0.065 0.930 West -0.042 0.062 0.502 0.036 0.063 0.574 # Parents 2 Reference 1 0.037 0.046 0.421 -0.045 0.056 0.424 0 0.119 0.071 0.092 -0.073 0.105 0.485 Age of mom 24 Reference 35 or over 0.006 0.051 0.909 0.066 0.055 0.229 Education of parents Less than HS Reference Completed HS 0.085 0.044 0.053 -0.031 0.055 0.566 Any college 0.103 0.067 0.125 0.084 0.087 0.333 Family income change Increase Reference No change -0.016 0.050 0.749 0.044 0.063 0.484 Decrease -0.027 0.060 0.653 -0.060 0.069 0.384
107 Table 4-13. Continued Amount of negative change (n=2297) Amount of positive change (n=1370) Coef. s.e. p Coef. s.e. p Parent employment status change Always employed Reference Employed to unemployed 0.088 0.089 0.322 0.227 0.165 0.168 Unemployed to employed -0.163 0.092 0.076 -0.113 0.077 0.142 Always unemployed -0.037 0.054 0.493 0.044 0.058 0.444 Parent insurance coverage change No change Reference Change -0.134 0.065 0.039 0.054 0.090 0.546 Change in general health status Better Reference No change -0.081 0.047 0.089 0.038 0.056 0.492 Worse -0.099 0.053 0.065 0.020 0.074 0.783 Change in # medical conditions Decrease Reference No change -0.098 0.043 0.021 0.034 0.051 0.501 Increase -0.096 0.056 0.087 0.050 0.060 0.412 Year 1996 Reference 1997 0.196 0.083 0.019 0.184 0.097 0.057 1998 0.375 0.079 0.000 -0.033 0.127 0.796 1999 0.222 0.079 0.005 0.120 0.095 0.209 2000 -0.043 0.091 0.640 -0.081 0.095 0.393 2001 -0.041 0.068 0.548 -0.043 0.082 0.599 2002 -0.161 0.071 0.023 0.121 0.086 0.161 2003 0.025 0.068 0.714 0.037 0.094 0.695 2004 -0.120 0.076 0.114 -0.015 0.079 0.851 Note: Part II is a GLM with Gamma family a nd log link), and Part I II is a GLM with Gamma family and log link). Table 4-14. Bootstrapping predic tion of change in well-child visits after 3-part model N Mean S.D. 95%CI MM 1000 -0.040***0.009(-0.058, -0.022) MU 1000 -0.160***0.032(-0.225, -0.101) MP 1000 -0.154***0.036(-0.223, -0.085) MU-MM 1000 -0.120***0.033(-0.186, -0.058) MP-MM 1000 -0.114***0.038(-0.189, -0.043) **Significant at 0.05 level. *** Significant at 0.01 level.
108 Table 4-15. Type of change in number of phys ician visits by insurance type (weighted) MM (%) MU (%) MP (%) Total (%) Decreased 31.9131.8639.0332.51 No change 43.5553.8033.2343.44 Increased 24.5414.3427.7424.05 Total 100.00100.00100.00100.00 Table 4-16. Descriptive statistics on nu mber of physician visits (weighted) MM MU MP Baseline round 0.719 (0.027)0.759 (0.102)1.386 (0.139) Following round 0.693 (0.027)0.450 (0.118)1.165 (0.137) Change over two rounds -0.026 (0.034)-0.309 (0.078)-0.220 (0.190) Mean (S.D.)
109 Table 4-17. Multinomial logit regression predicti ng probability of experiencing each type of change in physician visits (Part I) Relative risk of having negative change vs. no change Relative risk of having positive change vs. no change RRR s.e. p-value RRR s.e. p-value MM Reference MU 0.771 0.111 0.070.450 0.072 0 MP 1.427 0.171 0.0031.270 0.167 0.07 Age 0 Reference 2 0.330 0.033 00.449 0.045 0 6 0.193 0.021 00.247 0.026 0 13 0.193 0.024 00.205 0.027 0 Gender Male Reference Female 1.119 0.062 0.0441.128 0.069 0.049 Race Non-Hispanic white Reference Non-Hispanic African American 0.588 0.054 00.509 0.052 0 Hispanic 0.910 0.083 0.3020.895 0.088 0.261 Other 0.809 0.108 0.1150.550 0.105 0.002 MSA Urban Reference Rural 0.907 0.075 0.2381.025 0.095 0.788 Region Northeast Reference Midwest 0.917 0.109 0.4680.744 0.090 0.015 South 0.931 0.099 0.5020.817 0.090 0.067 West 0.750 0.085 0.0110.605 0.067 0 # Parents 2 Reference 1 1.090 0.099 0.3421.028 0.093 0.757 0 1.389 0.175 0.0090.980 0.140 0.889 Age of mom 24 Reference 35 or over 0.930 0.072 0.3481.189 0.099 0.039 Education of parents Less than HS Reference Completed HS 1.067 0.081 0.391.054 0.093 0.555 Any college 1.085 0.129 0.4921.174 0.150 0.208 Family income change Increase Reference No change 1.012 0.088 0.8951.162 0.103 0.089 Decrease 0.997 0.103 0.9751.246 0.138 0.047
110 Table 4-17. Continued Relative risk of having negative change vs. no change Relative risk of having positive change vs. no change RRR s.e. p-value RRR s.e. p-value Parent employment status change Always employed Reference Employed to unemployed 1.183 0.306 0.5160.891 0.211 0.626 Unemployed to employed 0.875 0.129 0.3651.106 0.193 0.563 Always unemployed 0.997 0.091 0.9771.104 0.109 0.316 Parent insurance coverage change No change Reference Change 0.930 0.111 0.5461.205 0.151 0.138 Change in general health status Better Reference No change 0.965 0.070 0.620.913 0.073 0.251 Worse 1.219 0.104 0.0210.979 0.094 0.825 Change in # medical conditions Decrease Reference No change 0.281 0.019 00.437 0.030 0 Increase 0.311 0.027 00.710 0.062 0 Year 1996 Reference 1997 1.417 0.214 0.0211.831 0.288 0 1998 1.619 0.251 0.0021.422 0.243 0.04 1999 1.135 0.194 0.4581.445 0.236 0.024 2000 1.095 0.174 0.5691.257 0.249 0.247 2001 0.994 0.134 0.9651.210 0.166 0.165 2002 0.849 0.117 0.2361.093 0.165 0.553 2003 0.982 0.123 0.8851.102 0.158 0.495 2004 0.777 0.103 0.0581.028 0.143 0.842
111 Table 4-18. GLM regression predic ting amount of change in number of physician visits given negative or positive change (Part II and III) Amount of Negative Change (n=3956) Amount of Positive Change (n=2981) Coef. s.e. p-value Coef. s.e. p-value MM Reference MU 0.1570.0780.045-0.035 0.155 0.820 MP 0.4730.1030.0000.286 0.107 0.008 Age 0 Reference 2 -0.1070.0620.0850.038 0.063 0.543 6 -0.1850.0590.0020.109 0.074 0.138 13 -0.1400.0850.1000.223 0.104 0.031 Gender Male Reference Female 0.0370.0450.406-0.063 0.046 0.175 Race Non-Hispanic white Reference Non-Hispanic African American -0.1690.0610.006-0.151 0.077 0.051 Hispanic -0.1390.0630.029-0.225 0.073 0.002 Other 0.0240.1310.855-0.352 0.088 0.000 MSA Urban Reference Rural -0.0230.0670.7280.018 0.070 0.793 Region Northeast Reference Midwest -0.0460.0860.5890.144 0.100 0.150 South -0.1260.0690.0650.124 0.095 0.191 West -0.0890.0750.235-0.015 0.086 0.860 # Parents 2 Reference 1 -0.0390.0690.572-0.021 0.068 0.752 0 0.2790.0790.000-0.048 0.102 0.638 Age of mom 24 Reference 35 or over 0.1670.0650.0100.189 0.071 0.008 Education of parents Less than HS Reference Completed HS 0.0780.0550.151-0.069 0.059 0.239 Any college 0.1110.0830.183-0.141 0.087 0.107 Family income change Increase Reference No change 0.0240.0570.677-0.078 0.065 0.225 Decrease 0.1410.0740.055-0.066 0.090 0.462 Parent employment status change Always employed Reference
112 Table 4-18. Continued Amount of Negative Change (n=3956) Amount of Positive Change (n=2981) Coef. s.e. p-value Coef. s.e. p-value Employed to unemployed 0.3770.2510.1340.133 0.273 0.626 Unemployed to employed 0.0200.0960.8310.014 0.094 0.885 Always unemployed 0.0260.0730.7180.102 0.066 0.120 Parent insurance coverage change No change Reference Change -0.0430.0790.582-0.029 0.095 0.759 Change in general health status Better Reference No change -0.0880.0560.116-0.025 0.061 0.680 Worse -0.0580.0640.366-0.006 0.073 0.935 Change in # medical conditions Decrease Reference No change -0.2920.0500.000-0.017 0.058 0.762 Increase -0.1980.0550.0000.188 0.070 0.007 Year 1996 Reference 1997 0.1790.0860.039-0.003 0.105 0.979 1998 0.4050.0980.000-0.019 0.125 0.877 1999 0.2090.1060.0490.142 0.123 0.248 2000 -0.0600.1200.6170.045 0.124 0.716 2001 -0.0480.0840.5700.119 0.118 0.316 2002 -0.0640.0990.5150.171 0.119 0.150 2003 -0.1150.0900.202-0.155 0.103 0.133 2004 -0.1600.0910.080-0.219 0.109 0.043 Note: Part II is a GLM with Gamma family a nd log link, and Part III is a GLM with inverse Gaussian family and log link. Table 4-19. Bootstrapping pred iction of change in physician visits after 3-part model N Mean S.D. 95%CI MM 10000.0060.022(-0.041, 0.050) MU 1000-0.271**0.023(-0.402, -0.138) MP 1000-0.1740.065(-0.420, 0.077) MU-MM 1000-0.277**0.127(-0.408, -0.141) MP-MM 1000-0.1800.068(-0.428, 0.080) **Significant at 0.05 level.
113 Table 4-20. Type of change in number of ER visits by insurance group (weighted) MM (%) MU (%) MP (%) Total (%) Decreased 7.326.087.427.23 No change 87.4988.4984.1187.28 Increased 5.205.438.475.49 Total 100.00100.00100.00100.00 Table 4-21. Descriptive statistics on number of ER visits (weighted) MM MU MP Baseline round 0.081 (0.005)0.103 (0.023)0.100 (0.017) Following round 0.080 (0.005)0.074 (0.017)0.130 (0.019) Change over two rounds -0.001 (0.006)-0.028 (0.028)0.030 (0.024) Mean (S.D.)
114 Table 4-22. Multinomial logit regression predicti ng probability of experiencing each type of change in ER visits (Part I) Relative risk of having negative change vs. no change Relative risk of having positive change vs. no change RRR s.e. p-value RRR s.e. p-value MM Reference MU 0.841 0.177 0.411 1.130 0.258 0.593 MP 0.872 0.166 0.472 1.437 0.256 0.042 Age 0 Reference 2 0.631 0.086 0.001 0.455 0.068 0.000 6 0.409 0.056 0 0.349 0.051 0.000 13 0.515 0.092 0 0.384 0.073 0.000 Gender Male Reference Female 0.968 0.085 0.716 0.819 0.090 0.068 Race Non-Hispanic white Reference Non-Hispanic African American 0.836 0.112 0.181 0.815 0.120 0.165 Hispanic 0.850 0.113 0.223 0.693 0.104 0.015 Other 0.830 0.225 0.49 0.566 0.177 0.068 MSA Urban Reference Rural 0.937 0.114 0.59 1.364 0.182 0.020 Region Northeast Reference Midwest 1.075 0.191 0.685 1.132 0.201 0.485 South 1.110 0.175 0.507 0.862 0.140 0.360 West 0.780 0.135 0.15 0.638 0.111 0.010 # Parents 2 Reference 1 1.236 0.154 0.09 1.062 0.144 0.657 0 1.320 0.249 0.142 0.971 0.190 0.880 Age of mom 24 Reference 35 or over 0.854 0.098 0.166 1.084 0.143 0.539 Education of parents Less than HS Reference Completed HS 1.117 0.128 0.334 0.871 0.122 0.326 Any college 1.541 0.259 0.01 0.920 0.194 0.694 Family income change Increase Reference No change 1.011 0.118 0.928 1.162 0.163 0.284 Decrease 1.077 0.154 0.602 1.196 0.233 0.357
115 Table 4-22. Continued Relative risk of having negative change vs. no change Relative risk of having positive change vs. no change RRR s.e. p-value RRR s.e. p-value Parent employment status change Always employed Reference Employed to unemployed 1.135 0.304 0.636 1.511 0.437 0.153 Unemployed to employed 1.673 0.325 0.008 1.601 0.374 0.044 Always unemployed 1.077 0.148 0.588 1.288 0.181 0.073 Parent insurance coverage change No change Reference Change 1.247 0.221 0.213 1.172 0.219 0.395 Change in general health status Better Reference No change 0.804 0.094 0.063 1.088 0.139 0.509 Worse 0.924 0.114 0.52 1.062 0.172 0.712 Change in # medical conditions Decrease Reference No change 0.270 0.032 0 0.666 0.086 0.002 Increase 0.288 0.043 0 0.964 0.137 0.795 Year 1996 Reference 1997 0.831 0.166 0.354 1.449 0.331 0.104 1998 1.101 0.275 0.699 1.018 0.287 0.950 1999 0.898 0.191 0.613 1.478 0.322 0.073 2000 0.998 0.247 0.994 1.173 0.307 0.542 2001 0.943 0.187 0.768 1.247 0.266 0.301 2002 1.050 0.197 0.796 1.294 0.301 0.268 2003 0.865 0.165 0.449 0.962 0.224 0.869 2004 0.873 0.170 0.485 1.436 0.326 0.111
116 Table 4-23. GLM regression predicting amount of change in number of ER visits given negative or positive change (Part II and Part III) Amount of Negative Change (n=836) Amount of Positive Change (n=621) Coef. s.e. p-value Coef. s.e. p-value MM Reference MU 0.4540.1530.003-0.168 0.1080.121 MP 0.1620.0680.0180.011 0.0770.888 Age 0 Reference 2 -0.0010.0530.9830.025 0.0770.745 6 -0.0530.0510.300-0.035 0.0720.624 13 -0.0420.0870.630-0.052 0.0950.587 Gender Male Reference Female 0.0010.0460.9880.028 0.0520.589 Race Non-Hispanic white Reference Non-Hispanic African American -0.1430.0610.0200.166 0.0710.020 Hispanic -0.0320.0670.630-0.044 0.0780.579 Other 0.0330.0930.7190.059 0.1450.684 MSA Urban Reference Rural -0.0060.0550.9170.073 0.0610.225 Region Northeast Reference Midwest 0.0130.0850.881-0.010 0.0960.917 South -0.0670.0710.3460.084 0.0690.229 West -0.1200.0820.1420.096 0.0690.162 # Parents 2 Reference 1 0.0890.0620.149-0.006 0.0810.945 0 0.2370.0870.007-0.158 0.1080.144 Age of mom 24 Reference 35 or over 0.0440.0560.4340.113 0.0620.067 Education of parents Less than HS Reference Completed HS -0.0400.0630.5210.010 0.0700.883 Any college -0.0630.0830.4420.084 0.0820.307 Family income change Increase Reference No change -0.0430.0580.4580.093 0.0580.109 Decrease 0.1680.0770.0290.235 0.0820.004 Parent employment status change Always employed Reference
117 Table 4-23. Continued Amount of Negative Change (n=836) Amount of Positive Change (n=621) Coef. s.e. p-value Coef. s.e. p-value Employed to unemployed -0.0170.1060.8710.179 0.1780.316 Unemployed to employed 0.0420.0830.6160.006 0.1120.957 Always unemployed -0.0150.0660.822-0.007 0.0710.923 Parent insurance coverage change No change Reference Change -0.1570.0790.046-0.076 0.0700.281 Change in general health status Better Reference No change -0.2070.0570.000-0.075 0.0550.173 Worse -0.1990.0640.002-0.015 0.0740.842 Change in # medical conditions Decrease Reference No change -0.0820.0570.1490.002 0.0630.970 Increase -0.2450.0600.0000.051 0.0690.461 Year 1996 Reference 1997 -0.0530.0990.5960.084 0.1030.414 1998 0.3130.1040.003-0.304 0.1690.072 1999 0.0470.1030.647-0.055 0.0910.547 2000 -0.1420.0980.150-0.076 0.0960.431 2001 0.0070.1220.9530.025 0.0910.783 2002 -0.1150.0930.217-0.009 0.0790.906 2003 -0.1200.0890.181-0.182 0.0950.056 2004 -0.1190.1120.288-0.027 0.0730.717 Note: Both Part II and Part III are GLM with Gamma family and log link. Table 4-24. Bootstrapping predic tion of change in number of ER visits after 3-part model N Mean S.D. 95%CI MM 10000.0220.025(-0.007, 0.072) MU 1000-0.0060.034(-0.068, 0.070) MP 10000.072**0.054(0.002, 0.203) MU-MM 1000-0.0280.030(-0.089, 0.027) MP-MM 10000.0500.038(-0.006, 0.142) **Significant at 0.05 level
118 Table 4-25. Type of change in number of hospitalizations by insu rance group (weighted) MM (%) MU (%) MP (%) Total (%) Decreased 1.832.002.371.89 No change 97.2197.5996.4597.17 Increased 0.960.401.180.94 Total 100.00100.00100.00100.00 Table 4-26. Descriptive statistics on number of hospitaliz ations (weighted) MM MU MP Baseline round 0.018 (0.002)0.028 (0.014)0.031 (0.010) Following round 0.013 (0.002)0.006 (0.004)0.014 (0.005) Change over two rounds -0.005 (0.002)-0.022 (0.015)-0.017 (0.011) Mean (S.D.)
119 Table 4-27. Multinomial logit regression predicting probability of experiencing each type of change in number of hospitalizations (Part I) Relative risk of having negative change vs. no change Relative risk of having positive change vs. no change RRR s.e. p-value RRR s.e. p-value MM Reference MU 1.219 0.468 0.6050.422 0.239 0.129 MP 1.317 0.448 0.4181.065 0.436 0.879 Age 0 Reference 2 0.258 0.054 00.416 0.177 0.039 6 0.114 0.030 00.440 0.149 0.015 13 0.355 0.083 00.544 0.204 0.105 Gender Male Reference Female 0.944 0.159 0.730.929 0.212 0.746 Race Non-Hispanic white Reference Non-Hispanic African American 1.272 0.266 0.250.952 0.262 0.859 Hispanic 0.832 0.178 0.391.269 0.346 0.381 Other 0.365 0.217 0.0910.641 0.414 0.49 MSA Urban Reference Rural 1.351 0.270 0.1332.016 0.521 0.007 Region Northeast Reference Midwest 1.276 0.383 0.4170.581 0.223 0.158 South 1.137 0.296 0.6220.666 0.218 0.214 West 1.146 0.315 0.620.840 0.310 0.637 # Parents 2 Reference 1 1.346 0.293 0.1731.036 0.283 0.898 0 1.117 0.336 0.7121.336 0.472 0.411 Age of mom 24 Reference 35 or over 0.838 0.152 0.3311.820 0.477 0.023 Education of parents Less than HS Reference Completed HS 0.652 0.124 0.0241.326 0.378 0.322 Any college 0.993 0.273 0.9781.271 0.477 0.524 Family income change Increase Reference No change 1.051 0.225 0.8181.479 0.431 0.179 Decrease 1.159 0.338 0.6121.407 0.540 0.373
120 Table 4-27. Continued Relative risk of having negative change vs. no change Relative risk of having positive change vs. no change RRR s.e. p-value RRR s.e. p-value Parent employment status change Always employed Reference Employed to unemployed 1.302 0.616 0.5782.941 2.122 0.135 Unemployed to employed 0.925 0.372 0.8471.326 0.627 0.55 Always unemployed 1.586 0.350 0.0370.927 0.303 0.817 Parent insurance coverage change No change Reference Change 0.817 0.279 0.5541.182 0.579 0.733 Change in general health status Better Reference No change 0.740 0.149 0.1350.644 0.174 0.103 Worse 1.041 0.227 0.8550.748 0.243 0.372 Change in # medical conditions Decrease Reference No change 0.283 0.058 00.589 0.165 0.059 Increase 0.293 0.072 00.681 0.199 0.189 Year 1996 Reference 1997 0.918 0.307 0.7972.950 1.453 0.028 1998 0.648 0.254 0.2682.517 1.527 0.129 1999 0.688 0.236 0.2755.142 2.464 0.001 2000 0.829 0.328 0.6362.512 1.412 0.101 2001 1.124 0.330 0.693.808 1.827 0.005 2002 0.535 0.185 0.0712.341 1.175 0.091 2003 0.624 0.207 0.1561.090 0.605 0.876 2004 0.424 0.133 0.0061.722 0.895 0.296
121 Table 4-28. GLM regression predic ting amount of change in hosp italizations given negative or positive change (Part II and III) Amount of negative change (n=231) Amount of positive change (n=112) Coef. s.e. p Coef. s.e. p MM Reference MU 0.2660.2680.3210.5740.322 0.075 MP 0.1630.1360.230-0.0620.131 0.638 Age 0 Reference 2 0.1790.1080.0980.0690.155 0.657 6 0.2370.1510.1170.2290.185 0.217 13 0.0120.1770.9470.1540.162 0.342 Gender Male Reference Female 0.1990.0940.0340.1190.096 0.214 Race Non-Hispanic white Reference Non-Hispanic African American 0.0270.1510.8590.3890.158 0.014 Hispanic 0.1060.1460.4660.1870.205 0.363 Other -0.0680.2580.7920.5960.294 0.043 MSA Urban Reference Rural 0.0700.1180.5550.1320.148 0.375 Region Northeast Reference Midwest 0.0260.1470.8590.1000.179 0.577 South -0.1090.1450.454-0.3360.120 0.005 West -0.2020.1680.229-0.2390.159 0.135 # Parents 2 Reference 1 -0.0110.1170.924-0.1290.096 0.180 0 0.4570.1570.004-0.0040.188 0.983 Age of mom 24 Reference 35 or over 0.2210.1250.0790.0160.121 0.896 Education of parents Less than HS Reference Completed HS 0.0850.0910.3540.0700.142 0.624 Any college 0.0880.1850.634-0.0240.153 0.873 Family income change Increase Reference No change -0.0500.1050.633-0.0470.101 0.640 Decrease 0.0310.1300.813-0.2200.231 0.341 Parent employment status change Always employed Reference
122 Table 4-28. Continued Amount of Negative Change (n=231) Amount of Positive Change (n=112) Coef. s.e. p Coef. s.e. p Employed to unemployed 0.4570.2440.062-0.9210.232 0.000 Unemployed to employed 0.0540.1820.7680.0920.193 0.634 Always unemployed -0.1580.1050.1340.0400.178 0.822 Parent insurance coverage change No change Reference Change -0.0580.1300.655-0.0640.169 0.707 Change in general health status Better Reference No change 0.0030.1310.982-0.0340.134 0.801 Worse -0.0200.1510.8940.0650.179 0.717 Change in # medical conditions Decrease Reference No change -0.2570.1140.024-0.2920.202 0.149 Increase -0.0600.1290.641-0.1030.128 0.425 Year 1996 Reference 1997 0.1680.1590.293-0.1540.182 0.399 1998 0.0940.1710.580-0.0690.316 0.827 1999 0.0150.1670.929-0.1830.227 0.420 2000 -0.4280.1760.016-0.0330.243 0.892 2001 -0.0220.1710.898-0.0760.268 0.777 2002 -0.0480.1630.7690.1670.200 0.404 2003 -0.1830.1280.1520.3580.323 0.269 2004 -0.1460.2070.480-0.1690.240 0.481 Note: Part II and III are GLM w ith Gamma family and log link. Table 4-29. Bootstrapping predictio n of change in number of hosp italization after 3-part model N Mean S.D. 95%CI MM 1000-0.0030.004(-0.008, 0.006) MU 1000-0.0090.033(-0.040, 0.045) MP 1000-0.0100.012(-0.032, 0.012) MU-MM 1000-0.0070.032(-0.039, 0.042) MP-MM 1000-0.0080.012(-0.032, 0.014) **Significant at 0.05 level
123 Table 4-30. Type of change in prescr iption drug use by insurance group (weighted) MM (%) MU (%) MP (%) Total (%) Decreased 27.7724.7828.6727.62 No change 52.5861.9443.9352.54 Increased 19.6513.2827.4019.83 Total 100.00100.00100.00100.00 Table 4-31. Descriptive statistics on nu mber of prescription drug use (weighted) MM MU MP Baseline round 0.853 (0.037) 0.784 (0.124) 1.779 (0.226) Following round 0.823 (0.045) 0.457 (0.066) 2.140 (0.484) Change over two rounds -0.030 ( 0.045) -0.327 (0.108) 0.361 (0.356) Mean (S.D.)
124 Table 4-32. Multinomial logit regression predicting proba bility of having each type of change in prescription drug use (Part I) Relative risk of having negative change vs. no change Relative risk of having positive change vs. no change RRR s.e. p-value RRR s.e. p-value MM Reference MU 0.700 0.098 0.0110.552 0.086 0 MP 1.065 0.135 0.6171.435 0.173 0.003 Age 0 Reference 2 0.697 0.066 00.649 0.065 0 6 0.457 0.045 00.472 0.048 0 13 0.522 0.065 00.416 0.052 0 Gender Male Reference Female 1.006 0.062 0.920.902 0.053 0.078 Race Non-Hispanic white Reference Non-Hispanic African American 0.512 0.048 00.516 0.052 0 Hispanic 0.797 0.071 0.0110.759 0.069 0.003 Other 0.771 0.148 0.1750.650 0.124 0.024 MSA Urban Reference Rural 1.225 0.102 0.0151.588 0.143 0 Region Northeast Reference Midwest 1.086 0.148 0.5480.964 0.131 0.788 South 1.290 0.137 0.0171.027 0.119 0.822 West 1.016 0.110 0.8860.797 0.096 0.061 # Parents 2 Reference 1 1.121 0.101 0.2061.012 0.081 0.886 0 1.222 0.151 0.1050.877 0.113 0.308 Age of mom 24 Reference 35 or over 0.943 0.076 0.4661.157 0.092 0.068 Education of parents Less than HS Reference Completed HS 1.067 0.079 0.3790.989 0.084 0.893 Any college 1.031 0.117 0.7891.081 0.131 0.52 Family income change Increase Reference No change 0.973 0.084 0.7531.150 0.105 0.127 Decrease 1.089 0.122 0.4461.154 0.140 0.238
125 Table 4-32. Continued Relative risk of having negative change vs. no change Relative risk of having positive change vs. no change RRR s.e. p-value RRR s.e. p-value Parent employment status change Always employed Reference Employed to unemployed 0.853 0.212 0.5230.748 0.172 0.207 Unemployed to employed 1.003 0.145 0.9861.066 0.182 0.708 Always unemployed 1.017 0.091 0.8521.100 0.105 0.323 Parent insurance coverage change No change Reference Change 0.921 0.117 0.518 0.903 0.118 0.435 Change in general health status Better Reference No change 0.726 0.058 00.746 0.062 0 Worse 0.936 0.083 0.4520.850 0.081 0.087 Change in # medical conditions Decrease Reference No change 0.166 0.012 00.367 0.029 0 Increase 0.197 0.018 00.634 0.059 0 Year 1996 Reference 1997 1.127 0.172 0.4331.200 0.169 0.195 1998 1.025 0.146 0.8630.909 0.151 0.567 1999 1.056 0.181 0.7520.941 0.168 0.734 2000 0.923 0.155 0.6330.955 0.166 0.79 2001 0.812 0.104 0.1030.819 0.109 0.135 2002 0.863 0.112 0.2590.866 0.109 0.252 2003 0.805 0.105 0.0970.665 0.083 0.001 2004 0.748 0.102 0.0330.642 0.090 0.002
126 Table 4-33. GLM regressions predicting amount of change in prescription drug use given negative or positive change (Part II and Part III) Amount of negative change (n=3415) Amount of positive change (n=2415) Coef. s.e. p s.e. Coef. p MM Reference MU 0.1890.1210.119-0.123 0.0990.216 MP 0.3070.0800.0000.362 0.1350.007 Age 0 Reference 2 -0.0800.0750.2900.064 0.0770.411 6 -0.0210.0720.7740.277 0.0880.002 13 -0.0500.1010.6170.295 0.1220.015 Gender Male Reference Female -0.1210.0550.028-0.161 0.0650.014 Race Non-Hispanic white Reference Non-Hispanic African American -0.1350.0800.092-0.065 0.0910.471 Hispanic -0.1310.0670.0490.022 0.0850.800 Other -0.0660.1360.630-0.349 0.1090.001 MSA Urban Reference Rural -0.1130.0640.0770.113 0.0790.154 Region Northeast Reference Midwest 0.0590.0950.532-0.172 0.1340.201 South -0.0250.0910.782-0.033 0.1250.791 West -0.1480.0900.100-0.300 0.1210.013 # Parents 2 Reference 1 0.0920.0790.243-0.114 0.0720.117 0 0.1810.1020.076-0.035 0.1340.797 Age of mom 24 Reference 35 or over 0.2010.0730.0060.106 0.0710.132 Education of parents Less than HS Reference Completed HS 0.0430.0580.453-0.067 0.0800.403 Any college 0.2690.1390.053-0.112 0.1110.311 Family income change Increase Reference No change -0.0360.0660.589-0.186 0.0930.045 Decrease 0.0650.0820.429-0.135 0.1150.241 Parent employment status change Always employed Reference
127 Table 4-33. Continued Amount of negative change (n=3415) Amount of positive change (n=2415) Coef. s.e. Coef. s.e. Coef. s.e. Employed to unemployed 0.5670.2310.014-0.110 0.136 0.419 Unemployed to employed 0.1860.1210.1240.085 0.122 0.488 Always unemployed -0.0160.0750.8350.199 0.087 0.023 Parent insurance coverage change No change Reference Change -0.071 0.0900.428-0.053 0.093 0.568 Change in general health status Better Reference No change -0.014 0.0610.825-0.089 0.079 0.262 Worse 0.001 0.0680.988-0.016 0.092 0.859 Change in # medical conditions Decrease Reference No change -0.3030.0640.000-0.221 0.076 0.004 Increase -0.1780.0730.0150.016 0.091 0.860 Year 1996 Reference 1997 0.0570.0880.5210.095 0.110 0.391 1998 0.4510.1110.0000.112 0.146 0.445 1999 0.2560.1090.0190.193 0.101 0.057 2000 -0.2490.1290.054-0.026 0.113 0.817 2001 0.0160.1130.8860.320 0.134 0.017 2002 -0.0860.0870.3270.246 0.115 0.033 2003 0.0470.1040.6540.251 0.119 0.036 2004 -0.1850.0990.0610.197 0.113 0.083 Note: Part II is a GLM with Gamma family and log link. Part III is a GLM with inverse Gaussian family and log link. Table 4-34. Bootstrapping prediction of change in prescription drug use N Mean S.D. 95%CI MM 1000 0.0160.042(-0.058, 0.094) MU 1000 -0.235**0.093(-0.418, -0.050) MP 1000 0.330**0.189(0.002, 0.734) MU-MM 1000 -0.251**0.095(-0.436, -0.069) MP-MM 1000 0.3140.183(-0.019, 0.702) **Significant at 0.05 level
128 0 .05 .1 .15 .2 .25 Density -10 -5 0 5 10 Residuals Kernel density estimate Normal densitykernel = epanechnikov, bandwidth = .15Kernel density estimate Figure 4-1. Kernel density plot after OLS regression on absolute amount of negative change in total expenditures 0.00 0.25 0.50 0.75 1.00 Normal F[(negresid-m)/s] 0.00 0.25 0.50 0.75 1.00 Empirical P[i] = i/(N+1) Figure 4-2. P-P plot after OLS regression on ab solute amount of negative change in total expenditures
129 -10 -5 0 5 10 Residuals -10 -5 0 5 10 Inverse Normal Figure 4-3. Q-Q plot after OLS regression on absolute amount of negative change in total expenditures 9.62639 + | | | **** | ** ************ ***** | ************************************** R | ************************************************* e | ************************************************** s | *************************************************** i | ********************************************************* d | ********************************************************* u | ****************************************************** a | ****************************************************** l | ************************************************ s | ****************************************** | ** ******************************** ** | ** **** ********* ** ***** *** | *********** * | ** * ** | -10.9885 + +----------------------------------------------------------------+ 3.58581 Fitted values 6.8125 Figure 4-4. Residual-fitted plot after OLS regression on absolute amount of negative change in total expenditures
130 0 .05 .1 .15 .2 .25 Density -10 -5 0 5 10 Residuals Kernel density estimate Normal densitykernel = epanechnikov, bandwidth = .15Kernel density estimate Figure 4-5. Kernel density plot after OLS regression on amount of positive change in total expenditures 0.00 0.25 0.50 0.75 1.00 Normal F[(posresid-m)/s] 0.00 0.25 0.50 0.75 1.00 Empirical P[i] = i/(N+1) Figure 4-6. P-P plot after OLS regression on amount of positive change in total expenditures
131 -10 -5 0 5 10 Residuals -10 -5 0 5 10 Inverse Normal Figure 4-7. Q-Q plot after OLS regression on amount of positive change in total expenditures 9.5644 + | | | ** **** ** | ******************************** | ******************************************* R | **************************************************** e | *************************************************** *** s | ******************************************************* i | ******************************************************** ** d | ********************************************************* ** u | ********************************************************** a | ****************************************************** l | ** *********************************************** s | ** **************************************** | ** ******************************* | ** ******** ** ** ** | * *** ** *** | ** | -11.5997 + +----------------------------------------------------------------+ 3.88551 Fitted values 6.60799 Figure 4-8. Residual-fitted plot after OLS regr ession on amount of positive change in total expenditures
132 151712 + | | | | | | r | e | s | n | e | g | | | ** | | | ** | ** ** * | **************** ****** *** -4919.56 + ********************************************* ** ** +----------------------------------------------------------------+ 85.8909 predicted mean txpdinv 4933.94 Figure 4-9. Residual-fitted plot after GLM regression on absolute amount of negative change in total expenditures 82094.2 + | | | | | | r | e | s | p | o | s | | * | | *** | *** * | ************ | ************ ** | ************************ *** -7273 + ******* **** * +----------------------------------------------------------------+ 98.2934 predicted mean txpd 7717.61 Figure 4-10. Residual-fitt ed plot after GLM regression for am ount of positive change in total expenditures
133 8.84927 + | | | | | | r | e | s | ** n | * e | * ** g | ** *** * | **** *** ***** ****** | ** ******** *********** *** | *************** ***** ********** *** | ***************** ******** **** *** | ********************************* | ************************************* | ******************************* **** ** ** -1.77576 + *** ** ** * +----------------------------------------------------------------+ .585066 predicted mean wvdinv 2.64688 Figure 4-11. Residual-fitt ed plot after GLM regression on absolute amount of negative change in number of well-child visits 6.98492 + | | | | | | r | e | * s | ** p | * o | * *** ** ** s | ***** ***** * | ** ** ** *** ** *** **** | ***** ** ********** *** | ******* ********************** ** *** | ************************************* ** ** | **** ******************************************* | ** ********************************** ***** ****** | *** ****************** ** *** -1.88257 + ** * +----------------------------------------------------------------+ .79879 predicted mean wvd 2.26676 Figure 4-12. Residual-fitt ed plot after GLM regression on amount of positive change in number of well-child visits
134 CHAPTER 5 DISCUSSION Overview Compared with children who stayed in Medicaid/SCHIP, children who lost Medicaid/SCHIP coverage and became uninsured experienced reductions in expenditures, wellchild visits, physician visits, and prescription drug use, and no si gnificant change observed in ER visits and hospitalizations, contro lling for other factors. Compared with children who stayed in Medicaid/SCHIP, those children who transitione d from Medicaid/SCHIP to private insurance experienced a drop in well-child visits, and no si gnificant change observe d in physician visits, ER visits, prescription drug use, hosp italizations, and total expenditures. This study indicates that losing Medicaid/SCHI P coverage resulted in decreased preventive care utilization among children, and decreased p hysician visits and prescription drug use among those who became uninsured. Policies to achieve enrollment stability are important in ensuring access to important health care for children. Summary and Interpretation of Findings Who Disenrolled from Medicaid/SCHIP? The study supports the notion that many ch ildren dropping out of Medicaid/SCHIP become uninsured. However, the disenrollment rate in the study is not representative of the general disenrollment rate since the study only includes t hose children who were in Medicaid/SCHIP for at least two rounds and remained in the same insurance status after their insurance status change. The actual disenr ollment rates should be higher. The study disenrollment sample decreased to about half when setting this criteria instead of including every disenrollment. This study didnt examine the exact trend but the re sults implied that the churning
135 problem in Medicaid/SCHIP is substantial. This trend may offset the efforts of expanding public insurance programs and decreasing the number of uninsured children. The study identified some individual and family characteristics of children who have a higher risk of disenrolling from Medicaid/SCHI P. Those children who left Medicaid/SCHIP without any insurance were more likely to be older children (aged 13), Hispanic children, children living in the South or th e West, children whose health stat us was not changing or getting better, children whose mom was older than 35, ch ildren whose family income increased, children whose parents were always employed, and childre n whose parents insurance status changed. Those predictors are quite consistent with pr evious studies. Future policy aimed at improving Medicaid/SCHIP retention rates sh ould pay special attention to children with those features. What are the Consequences of Dropout? Expenditures It was hypothesized that Medicaid/SCHIP di senrollment results in decreased total expenditures among children who end up with no insurance, and does not have significant influence on the expenditures among children who obtain private insurance. The analysis results support the hypothesis generated from the theoretical model and cros s-sectional stud y evidences. Those children who disenrolled from Medica id/SCHIP had much hi gher expenditures in the round before disenrollment than those stay -in children. After disenrollment, those who became uninsured had a large drop in expenditure s. The above pattern of use and expenditures may be related with the churning pattern in Medicaid/SCHIP enrollment. It has been reported that churning is common among Medicaid/SCHIP enrollees: 40% of those with any time on Medicaid/SCHIP left and later re-enrolled in one of these programs (Short 2003). Those dropout children are very likely to go on and off Medicai d/SCHIP programs. Previous studies show that at the initial stage of enrollment in Medica id/SCHIP, people tend to experience a rise in
136 expenditures. This rise can be explained by th at many of those new enrollees were enrolled through emergency room or hospitals because they were in need of expensive health care (e.g. hospitalizations). Later on, when they drop out of Medicaid/SCHI P and become uninsured, they tend to experience a big drop in expenditures because their medical conditions were relieved. Another explanation is that there are some unobser ved difference in the pattern of health care utilization between those stay-in children and those dropout children. Such as, the parents of those children who stayed in Medicaid/SCH IP may be more proactive in term of Medicaid/SCHIP enrollment and health care utilization. A large and significant drop (79%) in expend itures was observed in the period following Medicaid/SCHIP disenrollment among children who became uninsur ed afterwards. According to the results of multinomia l logit regression, those children were less likely to experience any change (increase or decrease) in expenditures and the scales of absolute change given any were smaller than those stay-in-Medicaid/SCHIP child ren. There are several possible explanations for the decrease in expenditures. One is that those ch ildren had difficulties in getting health care that they needed. Without insurance coverage, their pa rents had to pay out of pocket for any health care expense that occurred for those children. The health care expenses for those uninsured children increased the financial burden for those low-income families and may prevent those children from getting necessary health care in a timely manner. This may result in increased severity of potentially avoidabl e illnesses. This explanation is confirmed by the analysis of utilization measures in the study. Another reason is that those ch ildrens actual need of care decreased. Even though change in health status (g eneral health status and number of medical conditions) were controlled in the multivariate anal ysis, there still might be other need factors, such as the seriousness of disease and type of dis eases, not captured in the analysis. It is possible
137 that the diseases that the dropout children had were less serious and less costly than those that stay-in children had. However, the change in total health care e xpenditures cannot reflect the financial burden on childrens families. Financial burdenthe proportion of family income spent on out-ofpocket health care expenditures (including prem iums)on the patients families may increase after Medicaid/SCHIP disenrollment. Public insurance coverage can mitigate the financial burden for low-income families. Cross-sectional comparisons show that public coverage provides significantly greater pr otection from financial burden than private coverage or no coverage for low-income families. The financial burden for low-income families with public insurance is about half of the burden for lo w-income families whose children are without insurance and less than one third of the burden for low-income families whose children have private insurance, adjusting for age, race and se x (Wong et al. 2005). It is unclear whether those families with children who disenrolled from Medicaid/SCHIP experienced increased financial burden. This study cannot provide information on th is aspect. In the future study, the impact of Medicaid/SCHIP disenrollment on financial burde n of childrens families should be examined. In the perspective of societ y, the decreased health care expenditures observed among those children who lost Medicaid/SCHIP coverage and became uninsured doesnt mean that the total health care costs for those children are lower. It has been reported that uninsured children are more likely to show up in emergency room or be hospitalized for ACSC, which results in higher costs in the long term. In addition, an increase in uninsured children may lead to increased costs that are associated with uncompensated care most of which is paid by government dollars. A previous analysis of Medicaid disenrollment in a community suggests that Medicaid/SCHIP disenrollment would increase the number of uni nsured children and increase the community's
138 health care costs as a result of a shift in sites of care from less expens ive ambulatory office sites to more expensive ERs and increased hospitali zations for ACSC (Johnson and Rimsza 2004). Those children who transitioned to private insurance experienced a slight, but not significant, increase (13%) in expenditures dur ing the time period following Medicaid/SCHIP disenrollment, relative to thos e children who stayed in Medicaid /SCHIP. They were more likely to experience changes (increase or decrease) in expenditures than those stay-in-Medicaid/SCHIP children, and given the change, th ey experienced a larger scale of decrease or increase in expenditures. The increase in expenditures experienced by those children who transitioned to private insurance can be partly explained by th e differences in provide r payment rates between private insurance and public insurance. Private insurance has higher payment rates to providers than Medicaid and SCHIP, and therefore private insurance costs more than public insurance for the same services to children. In term of methodology, both log-transformed OLS and GLM with gamma family and log link were used for the linear regression parts. Th e predictions obtained from the two methods are close. The residuals from GLM re gressions are relatively larger th an those from OLS regressions, which supports that OLS is more efficient than GLM. Preventive care use The results for well-child visits show that Medicaid/SCHIP disenrollment led to a drop in well-child visits, no matter whether the child ob tained private insurance or became uninsured. And the scales of decreases in well-child visits were similar between those who obtained private insurance and those who became uninsured. The fi nding is consistent with Dubay and Kenneys (2001) and Yu and colleagues (2002) studies co mparing the well-child ca re use among children of different insurance coverage; that is, children covered by Me dicaid/SCHIP have better access
139 to preventive care than similar low-income childre n with private insuran ce and children without insurance. The results regarding those uninsured childre n are not surprising. It has been reported widely that uninsured children use less preven tive care and ambulatory care because of cost barriers. This highlights the im portance of keeping eligible low-income children enrolled in Medicaid and SCHIP continuously. The decreased number of well-child visits afte r children transitioned to private insurance indicates that Medicaid and SCHIP are relativ ely successful at prom oting and financing preventive care. Public plans use strong standa rds to ensure that a ppropriate and recommended services, especially preventive car e, are available to children. Relatively, state insurance rules which govern private insurance do not have as stri ct requirement on coverage of preventive care (such as pediatric vaccines) for children. So me low-cost private plans do not even offer preventive care. Ambulatory care use The multivariate analysis of physician visits demonstrates a large and significant decrease in physician visits by children who became in sured upon leaving Medicaid/SCHIP. The number of physician visits after disenro llment was only about half of th e visits that they had while covered by Medicaid/SCHIP. It is consistent with previous studies. This indicates that those uninsured children may lose their usual source of care and experience a discontinuity in care. Lower continuity of primary care was reported to be associated with higher risk of ER utilization and hospitalization (Christakis et al. 1999). Efforts to improve and maintain continuous coverage and continuous care should be warranted. Without adjusting for individual charact eristics, children who transitioned from Medicaid/SCHIP to private insurance experienced a 16% drop in the number of physician visits.
140 After adjusting for individual char acteristics, the change in physician visits among those children became insignificant. Those results agree with fi ndings from a previous cross-sectional study in which low-income children covered by Medicaid /SCHIP had more physician visits than lowincome children with private insurance, and the difference was not significant when controlling for health status and other individual f actors (Cunningham 2006) Those children who transitioned from Medicaid/SCHIP to private insurance had a highe r level of physician visits in both preand post-disenrollment time period than the other two groups of children, and a higher level of use in other types of care as well. This suggests that Medicaid/SCHIP children who had a higher level of utilization were more likely to tr ansition to private insurance, probably in search of a better choice of providers. However, according to this study, they did not actually have more physician visits after getting pr ivate insurance, which might be related with the higher costsharing (cost barriers) in private insurance plans. Private insurance does not always mean a better choice for children, especially those from low-income families. Emergency care use The descriptive data show that, compared with stay-in-Medicaid/SCHIP children, disenrollees had slightly more ER visits on average when they were covered by Medicaid/SCHIP. After disenrollment, those ch ildren who became uninsured experienced a drop (26%) in ER visits while those who obtained private insurance e xperienced an increase (30%) in average number of ER visits. Af ter adjusting for other individual characteristics, the changes in ER visits were not significant. This is consistent with the cr oss-sectional studies by Cunningham (2006) and Luo (2003) in which unins ured children dont have more ER visits than children with Medicaid/SCHIP coverage, adjus ting for socioeconomic status and health status. The results suggest that the differences in ER use change among insurance groups may be caused by other factors rather than insurance change.
141 However, the ER visit analysis has some me thodology limitations. One is that the first part of the model (multinomial logit) did not fit well with the data. It may be because of the small number of observations with a non-zero cha nge in ER visits during the study time period. Another is that the self-re ported ER visits may be underreported, especially among the uninsured. Usually the uninsured ar e more likely to not report thei r ER visits in survey data collection than those insured. The uninsured often go to the ER to obtain health care which they are unable to afford in physicians offices because they know they cannot be turned away at the ER. When they are asked about the ER visits, th ey are more likely to hide the experience than those insured. Therefore, the ER visit anal ysis results should be used carefully. Inpatient care As hypothesized, no significant changes in hospitalizations were detected in the multivariate analysis. One reason is that the sta tistical power of the analysis may not be large enough to detect the difference. There were onl y a few children who had hospitalizations among the disenrollment groups. The multinomial logit regres sion (Part I) fitted poorly. It may indicate that there was no significant change in hospitalization in the short-te rm (about half of year in the study) after Medicaid/SCHIP dise nrollment. Usually it is believe d that the uninsured children, especially those with ambulat ory care sensitive conditions, ar e less likely to use expensive inpatient care than children with public insu rance. Another existing argument is that the insufficient physician visits among uninsured childr en may result in hospitalizations later for ACSC (such as asthma). In the future, the long -term effect of Medicaid /SCHIP disenrollment on hospitalizations should be examined with data of a larger sample for extended study observation time period. Prescription drug use
142 Medicaid/SCHIP disenrollment was expected to lead to decreased prescription drug use among those children who became uninsured. That wa s, in fact, found to be the case. In the study, those children who lost Medicaid/SCHIP coverage and became uninsured experienced a large and significant drop (42%) in prescription drug use relative to their cohort who remained in Medicaid/SCHIP all the time. This suggests that those newly uninsured children were vulnerable to the cost barriers in access to prescription drugs. Children who transitioned to private insurance later used more than twice of prescription drugs before disenrollment than those child ren who stayed in Medicaid/SCHIP, without adjusting for other factors. After disenrollm ent, they experienced an increase (20%) in prescription drug use relative to those children who stay in Medicaid/SCHIP, but the difference is not significant after adjusting for other factors. It indicates that those children who transitioned to private insurance may be a group of children who needed more prescription drugs. And they got no worse access to prescription drugs after transitioning to private insurance. This finding is consistent with previous literat ure. It appears that the higher cost-sharing in private insurance plan did not prevent those child ren from using prescription drugs. Policy Implication Overall, this study shows that losing Medicaid/SCHIP covera ge can result in significant drops in health care utili zation and expenditures. Speci fically, this study found that Medicaid/SCHIP disenrollment leads to worse a ccess to important health care for children, including preventive car e, physician visits, and prescription drugs. Strategies to promote stable insurance cove rage merit serious consideration for a number of reasons. First, childre n with gaps in coverage may experi ence financial barriers in access to care and in maintaining a medical home. This lack of a medical home and instability in coverage may result in inferior health outcomes for ch ildren. According to previous studies, uninsured
143 children are more likely to use expensive ER care and inpatient care for ACSC relative to insured children (source) Second, from the perspective of hospitals and other providers, an increase in uninsured children may lead to increased costs as sociated with uncompensated care and resulting pressure on the health care safety net (Johns on and Rimsza 2004). This becomes even more serious an issue since the federal funding on th e safety net per uninsured person has been declining over recent years (KFF, 2007a). Third, those children who di senrolled from Medicaid/SC HIP generally had more utilization and expenditures befo re disenrollment than those c ontinuously enrolled children, and they experienced a drop in utilization and expe nditures when losing Medicaid/SCHIP coverage. They are very likely to re-enroll in Medica id/SCHIP later on since churning is a common phenomenon among Medicaid/SCHIP population (Short 2003). Upon re-enrollment, they may experience a significant rise in health care expe nditures based on the resu lts from this study and previous studies on Medicaid population with so me medical conditions (Harman et al. 2007). The churning may lead to fragmented care for th ose children. It will also result in higher administrative cost for public programs as Me dicaid and SCHIP official s and providers report significant costs related to chur ning (Fairbrother et al. 2004). Providing health coverage for low-income unins ured children continues to challenge the nation. There are signs that the progress in child rens health coverage is eroding. A number of states have adopted policies that will reduce childrens en rollment in Medicaid or SCHIP because of state budget pressures. The reaut horization can only ensure SCHIP funding until 2009. The future of SCHIP is uncertain in the curre nt political environment. In anticipation of the rapid policy change and unstable financing situation for public insurance programs of children, which will most likely occur in the upcoming years to cause instability in public
144 insurance coverage, quantifying the effect of Medicaid/SCHIP disenrollment on health care utilization and expenditure in th e pediatric population is of utmost importance. The results based on longitudinal analysis in this study provide important data fo r predicting and simulating the influence of Medicaid and/or SCHIP policy chan ge on health care utiliz ation and expenditures among children. Changes in public programs can substantia lly reduce disenrollment. The causes of dropouts have been examined over the past seve ral years and a range of practices that can contribute to coverage stabil ity has been identified: reduc e renewal problems by developing simplified forms, streamlining verification re quirements, and providi ng renewal assistance; reduce the frequency of renewals; and eliminate re newal requests when needed information is already on hand. This study also identified some indi vidual and family char acteristics associated with higher risk of Medicaid/ SCHIP disenrollment: older children (aged 13-18), Hispanic children, children whose family income increase d, and children whose parents insurance status changed. Future policy aimed at improving Medica id/SCHIP retention rates should pay specific attention to these groups of children. In addition, strategies should be developed that ensure cont inuous medical care even when children leave Medicaid/SCHIP and become uni nsured. Some existing programs have provided good resources for those uninsured children. Medi cal home projects and disease management programs are two examples Medical Home proj ects sponsored by AAP (spell out who is this) provide free laboratory tests and medications, and physician visits of low cost to uninsured children who do not qualify for Medicaid or SCHI P (AAP 2008). Disease management programs can be provided for children w ith chronic conditions, such as asthma. Disease management
145 programs in some states have been reported to be effective in improving access to care and increasing enrollment in Medicaid (Burton 2001). Limitations of the Study There are some limitatio ns to this study. First, the study cannot capture the long-term effect of insurance change. The use of survey data for assessment of insurance disenrollment and health care use/expenditures among children is limited because of the inability to track a singl e child for more than two consecutive years. The effect of insurance change on uti lization of certain types of care, such as ER and inpatient care, usually takes a longer time to show up. Second, Medicaid/SCHIP coverage and privat e insurance coverage were treated as endogenous variables in this study. Th e characteristics of those plans, such as benefit structures and cost-sharing requirements, were not include d in the multivariate analysis model. As mentioned earlier, the Medicaid benefit packag e is comprehensive and requires little cost sharing. Those types of care examinedwell-chil d visits, physician visits, inpatient care, ER care, and prescription drugsare ba sic services usually included in benefit packages of private health plans. It was assumed that all of the private insurance plan s cover those services. Relative to Medicaid/SCHIP, private plans have higher cost-sharing. Higher cost sharing may reduce the likelihood that children from low-income fam ilies with private coverage receive recommended care (Newhouse and the Insurance Experiment Group 1994). However, there was no specific cost-sharing information available in this study. Third, the state where the child lives cannot be identified us ing the public-use MEPS data file. Medicaid/SCHIP eligibility criteria in term of income level and benefit structure may vary across states. Considering that all of the state pr ograms are required to provide certain service
146 package and cost sharing is to be minimal, not controlling the program variation among states should not bias the estimation. Fourth, Medicaid and SCHIP cannot be distinguished. Th e MEPS does not distinguish between Medicaid and SCHIP and simply combines them in the survey. In roughly one third of states, Medicaid and SCHIP actually are run as a single combined program, with the remaining states administering separate pr ograms. Previous research sugge sts that the choice of program structure can affect program dr opout and there may be differences in dropout between Medicaid and SCHIP (Sommers 2005b) but the MEPS data does not allow for analysis of the differentiate effects between Medicaid and SCHIP. Considering that the two programs have similar benefits, this study explored the impact of disenrollmen t from childrens public insurance programs in general, comprising both Me dicaid and SCHIP. Conclusion When children left Medicaid/SCHIP programs and became uninsured, they experienced a significant drop in health care expenditures and us e, including preventive care, ambulatory care, and prescription drug use. When ch ildren transitioned from Medicaid/SCHIP to private insurance, they experienced a drop in well-c hild visits and no significant change in other type of care observed. It indicates that losi ng Medicaid/SCHIP coverage resulte d in a decrease in utilization of important health services among children. Po licies to achieve enro llment stability are important in ensuring access to impo rtant preventive care for children.
147 LIST OF REFERENCES AAP. 2000. Recommendations for Preventive Pedi atric Health Care (RE9939) Elk Grove Village, IL: American Academy of Pediatrics. AAP. 2002. "The Medical Home." Pediatrics 110 (1): 184-186. AAP. 2008. Medical Home Concepts. [Accessed on April 14, 2008.] Available at http://www.medicalhomeinfo.org/ AHRQ. 2002. National Healthcare Disparities Report. Rockville, MD: Agency for Healthcare Research and Quality. AHRQ. 2007. Medical Expenditure Panel Su rvey: Household Component Weighting and Estimation Issues. Rockville, MD: Agency for Healthcare Research and Quality. Aiken, K.D., G.L. Freed, and M.M. Davis. 2004. "When Insura nce Status Is Not Static: Insurance Transitions of Low-Income Childre n and Implications for Health and Health Care." Ambulatory Pediatrics 4 (3): 237-243. Andersen, R.M. 1968. A Behavioral Model of Families' Use of Health Services Chicago, IL. Center for Health Administration Studies, University of Chicago, Andersen, RM. 1995. "Revising the Behavioral Model and Access to Medical Care: Does it Matter?" Journal of Health Social Behavior 36: 1. Basu, A. and P.J. Rathouz. 2005. "Estimating Ma rginal and Incremental Effects on Health Outcomes Using Flexible Link and Variance Function Models." Biostatistics 6 (1): 93109. Berk, M.L. and A.C. Monheit. 2001. "The Con centration of Health Care Expenditures, Revisited." Health Affairs 20 (2): 9-18. Bloom, B., R.A. Cohen, J.L. Vickerie, and E.A. Wondimu. 2003. Summary Health Statistics for U.S. Children: National Health Interview Survey 2001 Hyattsville, MD: National Center for Health Statistics. Blough, D.K., C.W. Madden, and M.C. Hornbr ook. 1999. "Modeling Risk Using Generalized Linear Models." Journal of Health Economics 18 (2): 153-171. Blundell, R. and M.C. Dias. 2000. "Evaluati on Methods for Non-experimental Data." Fiscal Studies 21 (4): 427-468. Burton J. 2001. Careand Disease Management for Uninsured Individuals with Congestive Heart Failure. Academy Health Services Research Health Policy Meeting 18: 16. Campbell, D. and J. Stanley. 1966. Experimental and Quasi-experi mental Designs for Research Chicago: Rand McNally.
148 Chen, A.Y. and R.K. Chang. 2002. "Factors Associ ated with Prescription Drug Expenditures among Children: an Analysis of the Medical Expenditure Panel Survey." Pediatrics 109 (5): 728-732. Chen, A.Y. and J.J. Escarce. 2006. "Effects of Fa mily Structure on Children's Use of Ambulatory Visits and Prescription Medications." Health Services Research 41 (5): 1895-1914. Children's Defense Fund Texas. 2006. CHIP: Saving Texans Money, Keeping Families First. [Accessed on March 2, 2008]. Available at http://www.cdftexas.org/attached/CHIPCostSavings.pdf Christakis, D.A., L. Mell, J.A. Wright, R. Da vis, and F.A. Connell. 2000. "The Association between Greater Continuity of Care and Ti mely Measles-Mumps-Rubella Vaccination." American Journal of Public Health 90 (6): 962-965. Christakis, D.A., L. Mell, T.D. Koepsell, F.J. Zimmerman, and F.A. Connell. 2001. "Association of Lower Continuity of Care with Greate r Risk of Emergency Department Use and Hospitalization in Children." Pediatrics 107 (3): 524-529. Christakis, D.A., J.A. Wright, T.D. Koepsell, S. Emerson, and F.A. Connell. 1999. "Is Greater Continuity of Care Associated with Less Emergency Department Utilization?" Pediatrics 103 (4): 738-742. Christakis, D.A., J.A. Wright, F.J. Zimme rman, A.L. Bassett, and F.A. Connell. 2003. "Continuity of Care is Associat ed with Well-Coordinated Care." Ambulatory Pediatrics 3 (2): 82-86. Chung, P., T. Lee, J. Morrison, and Schuster M. A. 2006. "Preventive Care for Children in the United States: Quality and Barriers." Annual Review of Public Health 27: 491-515. Cohen, J.W., S.R. Machlin, S.H. Zuvekas, M.N. Stagnetti, and J. Thorpe. 2000. Health Care Expenses in the United States, 1996. Report No. 01-0009. Rockville, MD: Agency for Healthcare Research and Quality. Cohen, S.B., R. DiGaetano, and H. Goks el. 1999. Methodology Report #5: Estimation Procedures in the 1996 Medi cal Expenditure Panel Surv ey Household Component. Rockville, MD: Agency for Healthcare Research and Quality. Cook, T. and D. Campbell. 1979. Quasi Experiment ation: Design and Analys is Issues for Field Settings. Chicago, IL: Rand McNally. Cunningham, P.J. 2006. "Medicaid/SCHIP Cuts and Hospital Emergency Department Use." Health Affairs 25 (1): 237-247. Currie, J. and J. Gruber. 1996. "Health Insuranc e Eligibility, Utilization of Medical Care, and Child Health." The Quarterly Journal of Economics 111 (2): 431-466.
149 Currie, J. and D. Thomas. 1995. "Medical Care for Children: Public Insurance, Private Insurance, and Racial Di fferences in Utilization." The Journal of Human Resources 30 (1): 135-162. Dafny, L. and J. Gruber. 2000. Does Public Insura nce Improve the Efficiency of Medical Care? Medicaid Expansions and Child Hospitalizations. Working Paper no. 7555. Cambridge, MA: National Bureau of Economic Research. Davis, M., and D. Bruckman. 2003. Health Care Expenditures and Utilization for Children Who Experience Insurance Transitions. Abstract AcademyHealth Meeting 20 (abstract #298). Dick, A., R. Allison, S. Haber, C. Brach, a nd E. Shenkman. 2002. "Consequences of States' Policies for SCHIP Disenrollment." Health Care Finance Review. 23: 65-88. Diehr, P., D. Yanez, A. Ash, M. Hornbrook, and D. Lin. 1999. "Methods for Analyzing Health Care Utilization and Costs." Annual Review of Public Health 20 125-144. Diette, G.B., E.A. Skinner, L.E. Markson, P. Algatt-Bergstrom, T.T.H. Nguyen, R.D. Clark, and A.W. Wu. 2001. "Consistency of Care with National Guidelines for Children with Asthma in Managed Care." The Journal of Pediatrics 138 (1): 59-64. Dodd, S., A. Bassi, K. Bodger, and P. W illiamson. 2006. "A Comparison of Multivariable Regression Models to Analyze Cost Data." Journal of Evaluation in Clinical Practice 12 (1): 76-86. Duan, N. 1983. "Smearing Estimate: A N onparametric Retransformation Method." Journal of the American Statistical Association 78 (383): 605-610. Duan, N., W.G. Manning, Jr., C.N. Morris, and J.P. Newhouse. 1984. "Choosing between the Sample-Selection Model and the Multi-Part Model." Journal of Business & Economic Statistics 2 (3): 283-289. Dubay, L., J. Guyer, C. Mann, and M. Odeh. 2007. "Medicaid at the TenYear Anniversary of SCHIP: Looking back and Moving forward." Health Affairs 26 (2): 370-381. Dubay, L. and G.M. Kenney. 2001. "Health Care Access and Use among Low-Income Children: Who Fares Best?" Health Affairs 20 (1): 112-121. Duderstadt, K., D. Hughes, M. Soobader, a nd P. Newacheck. 2006. "The Impact of Public Insurance Expansions on Children's Access and Use of Care." Pediatrics 118 (4): 16761682. Ettner, S.L. 2004. "Methods for Addressing Selec tion Bias in Observatio nal Studies." National Research Service Award Trainees Research Conference. Agency for Healthcare Research and Quality. Rockville, MD. Available at http://www.ahrq.gov/fund/training/ettnertxt.htm
150 Fahrmeir, L. and G. Tutz. 2001. Multivariate Statistical Modelli ng Based on Generalized Linear Models. Second Edition: Springer. Fairbrother, G., M.J. Dutton, D. Bachrach, K.A. Newell, P. Boozang, and R. Cooper. 2004. "Costs of Enrolling Children in Medicaid and SCHIP." Health Affairs 23 (1): 237-243. Florida Healthy Kids Corporation. 1997. "Flori da Healthy Kids Annual Report." [Accessed October 1, 2007]. Available at http://www.amsa.org/cph/CHIPfact.cfm Fossett, J.W. 1992. Medicaid and Access to Child Health Care in Chicago. Journal of Health Politics, Policy and Law 17(2): 273-298. Freeman, H. and C.R. Corey. 1993. "Insurance Status and Access to Health Services among Poor Persons." Health Services Research 28 (5): 531-542. Green, M. and J. Palfrey. 2002. B right Futures: Guidelines for He alth Supervision of Infants, Children, and Adolescents 2nd edition. Arlington, VA: Nati onal Center for Education in Maternal and Child Health. Hakim, RB. and DS. Ronsavil le. 2002. "Effect of Compliance with Health Supervision Guidelines among US Infants on Em ergency Department Visits." Archives of Pediatrics Adolescent Medicine 156 (10): 1015-1020. Hakim, R.B. and B.V. Bye. 2001. "Effectiveness of Compliance with Pediatric Preventive Care Guidelines among Medicaid Beneficiaries." Pediatrics 108 (1): 90-97. Harman, J.S., A.G. Hall, and J. Zhang. 2007. Change s in health care use and costs after a break in Medicaid coverage among persons with depression. Psychiatric Services 58(1): 4954. Hildebrand, D.K., J.D. Laing, and H. Rosenthal. 1977. Prediction Analysis of Cross Classifications New York: John Wiley and Sons, Inc. Hill, I. and A. Lutzky. 2003. Is There a Hole in the Bucket ? Understanding SCHIP Retention Washington, DC: Urban Institute. Holl, J.L., P.G. Szilagyi, L.E. Rodewald, L.P. S hone, J. Zwanziger, D.B. Mukamel, S. Trafton, A.W. Dick, R. Barth, and R.F. Raubertas. 2000. "Evaluation of New York State's Child Health Plus: Access, Utilization, Quality of Health Care, and Health Status." Pediatrics 105 (3): 711-718. Honberg, L., M. McPherson, B. Strickland, J. C. Gage, and P.W. Newacheck. 2005. "Assuring Adequate Health Insurance: Results of th e National Survey of Children with Special Health Care Needs." Pediatrics 115 (5): 1233-1239. Hostetler, M., S. Mace, K. Br own, J. Finkler, D. Hernandez, S. Krug, and N. Schamban. 2007. "Emergency Department Overcrowding and Children." Pediatric Emergency Care 23 (7): 507-515.
151 Hudson, J.L. and T.M. Selden. 2007. "Children's Eligibility and Coverage : Recent Trends and a Look ahead." Health Affairs 26 (5): w618-w619. Institute of Medicine. 2001. Crossing the Quality Chasm: A New Health System for the 21st Century. Washington, DC: National Academies Press Iglehart, J.K. 2007. "The Battle over SCHIP." The New England Journal of Medicine 357 (10): 957-960. Irvin, C., D. Peikes, C. Trenholm, and N. Khan. 2002. Continuous Coverage: Removing Barriers to Children's Health Care. Princeton, NJ: Mathematica Policy Research, Inc. Irvin, C., D. Peikes, C. Trenholm, and N. Khan. 2001. Discontinuous Coverage in Medicaid and the Implications of 12-Month Continuous Coverage for Children (Final Report). PR0180. Cambridge, MA: Mathematica Policy Research, Inc. Johnson, W.G. and M.E. Rimsza. 2004. "The Effects of Access to Pediatric Care and Insurance Coverage on Emergency Department Utilization." Pediatrics 113 (3): 483-487. Johnston, J. and J. DiNardo. 1997. Econometric Methods. 4th edition. New York, NY: The McGraw-Hill Companies, Inc. Keele, L. and N.J. Kelly. 2006. "Dynamic Models for Dynamic Theories: The Ins and Outs of Lagged Dependent Variables." Political Analysis 14 (2): 186-205. Kempe, A., B.L. Beaty, L.A. Crane, J. Stoksta d, J. Barrow, S. Belman, and J.F. Steiner. 2005. "Changes in Access, Utilization, and Quality of Care after Enrollment into a State Child Health Insurance Plan." Pediatrics 115 (2): 364-371. Kennedy, P. 1998. Guide to Econometrics 4th Ed. Cambridge, MA: The MIT Press. KFF. 2002. Children's Health: Why Health Insurance Matters? Washington, DC: Kaiser Family Foundation. KFF. 2006. Medicaid Benefits: Online Database Washington, DC: Kaiser Family Foundation. KFF. 2007a. Covering the Uninsured: Growing Need, Strained Resources. Washington, DC: Kaiser Family Foundation. KFF. 2007b. State Children's Health Insurance Program (SCHIP) at a Glance. Washington, DC: Kaiser Family Foundation. Kogan, M.D., G.R. Alexander, M.A. Teitelbaum, B.W. Jack, M. Kotelchuck, and G. Pappas. 1995. "The effect of gaps in health insuran ce on continuity of a regular source of care among preschool-aged children in the United States." The Journal of the American Medical Association 274 (18): 1429-1435.
152 Ku, L., M. Lin, and M. Broaddus. 2007. Improving Children's Health: A Chartbook about the Roles of Medicaid and SCHIP. Washington, DC: Center on Budget and Policy Priorities. Ku, L. and D.C. Ross. 2002. Staying Covered: The Importance of Retaining Health Insurance for Low-Income Families New York, NY: The Commonwealth Fund. Lambrew, J.M. 2007. The State Children's Health Insurance Program: Past, Present, and Future New York, NY: The Commonwealth Fund. Lave, J.R., C.R. Keane, C.J. Lin, E.M. Ricci, G. Amersbach, and C.P. LaVallee. 1998. "Impact of a Children's Health Insurance Program on Newly Enrolled Children." The Journal of the American Medical Association 279 (22): 1820-1825. Lipson, K., E. Fishman, P. Boozang, D. and D. Bachrach. 2003. Rethinking Recertification: Keeping Eligible Individuals Enrolled in New York's Public Health Insurance Programs. New York, NY: The Commonwealth Fund. Luo, X., G. Liu, K. Frush, and L.A. Hey. 2003. "Children's Health Insurance Status and Emergency Department Utilization in the United States." Pediatrics 112 (2): 314-319. Manning, W.G. 1998. "The logged dependent va riable, heteroscedasticity, and the retransformation problem." Journal of Health Economics 17 (3): 283-295. Manning, W.G., A. Basu, and J. Mullahy. 2005. "G eneralized modeling approaches to risk adjustment of skewed outcomes data." Journal of Health Economics 24 (3): 465-488. Manning, W.G. and J. Mullahy. 2001. "Estimating log models: to transform or not to transform?" Journal of Health Economics 20 (4): 461-494. Miller, J. and J. Phillips. 2002. Does Context Affect SCHIP Disenrollment? Findings from a Multilevel Analysis Chicago, IL: Joint Center for Poverty Research. Mitchell, J. 1991. Participa tion in Medicaid Revisited. Medical Care 29 (7): 645-653. Mullahy, J. 1998. "Much Ado about Two: Reconsid ering Retransformation and the Two-Part Model in Health Econometrics." Journal of Health Economics 17 (3): 247-281. Newacheck, P.W., J.J. Stoddard, D.C. Hughes, and M. Pearl. 1998. "Health Insurance and Access to Primary Care for Children." New England Journal of Medicine 338 (8): 513519. Newacheck, P.W., S.T. Wong, A. A. Galbraith, and Y.Y. Hung. 2003. "Adolescent Health Care Expenditures: a Descriptive Profile." Journal of Adolescent Health 32 (6, Supplement 1): 3-11. Newhouse, J.P,, and the Insurance Experiment Group. 1994. Free for All? Lessons from the RAND Health Insurance Experiment Cambridge, MA: Harvard University Press.
153 Olson, L.M., S.F. Tang, and P.W. Newacheck. 2005. "Children in the United States with Discontinuous Health Insurance Coverage." The New England Journal of Medicine 353 (4): 382-391. Pauly, M.V. 1968. The Economics of Moral Hazard: Comment. American Economics Review 58 (pt 1): 531-537. Pauly, M.V. 2004. Medicare Drug C overage and Moral Hazard. Health Affairs 23 (1):113-122. Penchansky, R. and J.W. Thomas. 1981. "The Con cept of Access: Definition and Relationship to Consumer Satisfaction." Medical Care 19 (2): 127-140. Rimsza, M.E., R.J. Butler, and W.G. Johnson. 2007. "Impact of Medicaid Disenrollment on Health Care Use and Cost." Pediatrics 119 (5): e1026-e1032. Saultz, J.W. and J. Lochner. 2005. "Interpersona l Continuity of Care and Care Outcomes: A Critical Review." Annals of Family Medicine 3 (2): 159-166. Schneider,A., R.Elias, R.Garfield, D.Roussea u, and V.Wachino. 2002. "The Medicaid Resource Book." Washington, DC: The Kaiser Comm ission on Medicaid and the Uninsured. Schwartz, K., C. Hoffman, and A. Cook. 2007. H ealth Insurance Coverage of Americans' Children. Washington, DC: Kaiser Commission on Medica id and the Uninsured. Selden, T.M. 2006. "Compliance with Well-chil d Visit Recommendations: Evidence from the Medical Expenditure Panel Survey, 2000-2002." Pediatrics 118 (6): E1766-E1778. Short, P. and D. Lefkowitz. 1992. "Encouraging Pr eventive Services for Low-Income Children the Effect of Expanding Medicaid." Medical Care 30 (9): 766-780. Sloan, F., R. Mitchell, and J. Cromwell. 1978. P hysician Participation in State Medicaid Programs. Journal of Human Resources 13 (Suppl.): 211-245. Sommers, B.D. 2005a. "From Medicaid to Unin sured: Drop-Out among Children in Public Insurance Programs." Health Services Research 40 (1): 59-78. Sommers, B.D. 2005b. "The Impact of Program Structure on Children's Disenrollment from Medicaid and SCHIP." Health Affairs 24 (6): 1611-1618. Sommers, B.D. 2006. "Insuring Children or Insuring Families: Do Parental and Sibling Coverage Lead to Improved Retention of Children in Medicaid and CHIP?" Journal of Health Economics 25 (6): 1154-1169. Starfield, B. and L. Shi. 2004. "The Medical Home Access to Care, and In surance: A Review of Evidence." Pediatrics 113 (5): 1493-1498.
154 Stoddard, J.J., R.F. St. Peter, and P.W. Ne wacheck. 1994a. "Health Insurance Status and Ambulatory Care for Children." The New England Journal of Medicine 330 (20): 14211425. Stoddard, J.J., R.F. St. Peter, and P.W. Newacheck. 1994b. "Health Insurance Status and Ambulatory Care for Children." The New England Journal of Medicine 330 (20): 14211425. Stoddard, J.J, R. F. St. Peter, and P.W. Newacheck. 1994c. "Health Insurance Status and Ambulatory Care for Children." The New England Journal of Medicine 330 (20): 14211425. Summer, L. and C. Mann. 2006. Instability of Public Health Insurance Coverage for Children and Their Families: Causes, Consequences, and Remedies. Report #935. New York, NY: The Commonwealth Fund. Szilagyi, P.G., A.W. Dick, J.D. Klein, L.P. Shone, J. Zwanziger, and T. McInerny. 2004. "Improved Access and Quality of Care after Enrollment in the New York State Children's Health Insurance Program (SCHIP)." Pediatrics 113 (5): e395-e404. Tang, S.F., L.M. Olson, and B.K. Yudkowsky. 2003. "Uninsured Childre n: How We Count Matters." Pediatrics 112 (2): e168-e173. Todd, J., C. Armon, A. Griggs, S. Poole, and S. Berman. 2006. "Increased Rates of Morbidity, Mortality, and Charges for Hospitalized Children with Public or No Health Insurance as Compared with Children with Private Insura nce in Colorado and the United States." Pediatrics 118 (2): 577-585. U.S. Department of Health and Human Services Health Resources and Services Administration. 1999. Measuring Success for Healthy People 2 010: National Agenda for Children with Special Health Care Needs Washington, DC. U.S. Department of Labor. 2008. Consumer Price Index Data. [Accessed on November 20, 2007.] Available at http://www.bls.gov/data/home.htm Wilson, K.M. and J.D. Klein. 2000. "Adolescents Who Use the Emergency Department as Their Usual Source of Care." Archives of Pediatrics Adolescent Medicine 154 (4): 361-365. Wong, S.T., A. Galbraith, S. Kim, and P.W. Ne wacheck. 2005. Dispariti es in the Financial Burden of Childrens Healthcare Expenditures. Archives of Pediatrics Adolescent Medicine 159 (11): 1008-1013. Wooldridge, J., G. Kenney, C. Trenholm, L. Duba y, I. Hill, M. Kim, L. Moreno, A. Sommers, and S. Zuckerman. 2005. Congressionally Mandated Evalua tion of the State Children's Health Insurance Program: Final Report to Congress Washington, DC: U.S. Department of Health and Human Services.
155 Wooldridge, J.M. 2005. Introductory Econometrics: A Modern Approach Third Edition. Mason, OH: Thomson South-Western. Yu, S.M., H.A. Bellamy, M.D. Kogan, J.L. Dunbar, R.H. Schwalberg, and M.A. Schuster. 2002. "Factors That Influence Receipt of Recommende d Preventive Pediatric Health and Dental Care." Pediatrics 110 (6): e73 Zuvekas, S.H. and J.W. Cohen. 2002. "A Guide to Comparing Health Care Expenditures in the 1996 MEPS to the 1987 NMES." Inquiry 39 (1): 76-86.
156 BIOGRAPHICAL SKETCH Jingbo Yu was born in Weihai of China, one of the most beautiful places in the world. Her parents always support her to pursue as much edu cation as she wants. She wanted to become a doctor and studied medicine in college. While do ing her internship in hospital, she witnessed some seriously sick patients having to end treatmen t because they were too poor to afford it. She realized, sadly, that solely advance in medi cal techniques and tec hnologies cannot ensure improvement of health. She developed an intere st in health insuran ce policy and obtained a masters degree in healthcare administration from Shandong Universit y. After obtaining her master degree, she decided to become a health services researcher and get more training in research methods. In 2004, she came to the United States to study for a PhD degree in health services research at the University of Florida. At UF, she received training in various health services research methods and developed her inte rest in health insura nce area. Meantime, she worked at Florida Center for Medicaid a nd the Uninsured. She graduated in Spring 2008.