<%BANNER%>

Assessment of Response Shift in Patients Treated with a Calcium Antagonist or an Atenolol-Led Hypertension Strategy in t...

Permanent Link: http://ufdc.ufl.edu/UFE0042532/00001

Material Information

Title: Assessment of Response Shift in Patients Treated with a Calcium Antagonist or an Atenolol-Led Hypertension Strategy in the International Verapamil-Trandolapril Study
Physical Description: 1 online resource (137 p.)
Language: english
Creator: Gandhi, Pranav
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: Pharmaceutical Outcomes and Policy -- Dissertations, Academic -- UF
Genre: Pharmaceutical Sciences thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Numerous reliable and valid measures for HRQoL assessments are available but these assessments may be subject to change over time or after an intervention from factors separate from the effects of the intervention itself. These changes are referred to as response shifts. We investigated the occurrence of a response shift using short form-36 (SF-36) HRQoL instrument in hypertensive patients with coronary artery disease (CAD). First, two structural equation modeling (SEM) techniques were compared and examined for convergence to examine plausible occurrence of response shift over two time points (i.e., at baseline and at one year). We provide alternate explanations for the divergent results we found compared to the Ahmed et al. study. Our study identified recalibration for the SF-36 physical function (PF) scale using both Oort and Schmitt approaches. Based on our results, the differences between the Oort and Schmitt approaches in our study may be due to variation in the method and not the sample used. Method differences may result from divergence in defining the type of response shift linked to changes in various parameters of the SEM measurement model. Our finding of response shift may be explained by a single or a combination of factors including the disease condition, assignment of antihypertensive treatment strategies, age, gender, and/or the presence of depressive symptoms. Second, we investigated measurement bias and response shift jointly with use of SEM over a one year period. We found that gender caused response shift in measurement. Women have a higher incidence of depression, and in combination with lower PF scores would worsen their quality of life (QOL). Thus, it is imperative to improve depression-caused reductions in PF and maintain an individual?s ability and willingness to perform daily activities which may help reduce institutionalization. Our results suggest that the SF-36 PF scale may be susceptible to response shift ? this deserves further research. By looking more closely at the scores for SF-36 PF domain in this study population will enable us to provide nuanced attention and direct treatment for the most impaired aspects of QOL.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Pranav Gandhi.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: Ried, Lyle D.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2011-12-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0042532:00001

Permanent Link: http://ufdc.ufl.edu/UFE0042532/00001

Material Information

Title: Assessment of Response Shift in Patients Treated with a Calcium Antagonist or an Atenolol-Led Hypertension Strategy in the International Verapamil-Trandolapril Study
Physical Description: 1 online resource (137 p.)
Language: english
Creator: Gandhi, Pranav
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: Pharmaceutical Outcomes and Policy -- Dissertations, Academic -- UF
Genre: Pharmaceutical Sciences thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Numerous reliable and valid measures for HRQoL assessments are available but these assessments may be subject to change over time or after an intervention from factors separate from the effects of the intervention itself. These changes are referred to as response shifts. We investigated the occurrence of a response shift using short form-36 (SF-36) HRQoL instrument in hypertensive patients with coronary artery disease (CAD). First, two structural equation modeling (SEM) techniques were compared and examined for convergence to examine plausible occurrence of response shift over two time points (i.e., at baseline and at one year). We provide alternate explanations for the divergent results we found compared to the Ahmed et al. study. Our study identified recalibration for the SF-36 physical function (PF) scale using both Oort and Schmitt approaches. Based on our results, the differences between the Oort and Schmitt approaches in our study may be due to variation in the method and not the sample used. Method differences may result from divergence in defining the type of response shift linked to changes in various parameters of the SEM measurement model. Our finding of response shift may be explained by a single or a combination of factors including the disease condition, assignment of antihypertensive treatment strategies, age, gender, and/or the presence of depressive symptoms. Second, we investigated measurement bias and response shift jointly with use of SEM over a one year period. We found that gender caused response shift in measurement. Women have a higher incidence of depression, and in combination with lower PF scores would worsen their quality of life (QOL). Thus, it is imperative to improve depression-caused reductions in PF and maintain an individual?s ability and willingness to perform daily activities which may help reduce institutionalization. Our results suggest that the SF-36 PF scale may be susceptible to response shift ? this deserves further research. By looking more closely at the scores for SF-36 PF domain in this study population will enable us to provide nuanced attention and direct treatment for the most impaired aspects of QOL.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Pranav Gandhi.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: Ried, Lyle D.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2011-12-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0042532:00001


This item has the following downloads:


Full Text

PAGE 1

1 ASSESSMENT OF RESPONSE SHIFT IN PATIENTS TREATED WITH A CALCIUM ANTAGONIST OR AN ATENOLOL LED HYPERTENSION STRATEGY IN THE INTERNATIONAL VERAPAMIL TRANDOLAPRIL STUDY B y PRANAV KIRIT GANDHI A DISSERTATION PRESENTED TO THE G RADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2010

PAGE 2

2 2010 Pranav Kirit Gandhi

PAGE 3

3 To my mom and d ad

PAGE 4

4 ACKNOWLED GMENTS Time spent as a graduate student at the University of Florida constituted some of the most important and in many ways the life shaping years of my life. I would like to thank my Ph.D. advisor Dr. L. Douglas Ried for his constant guidance and support in my work. His commitment to work was always inspiring and I have learned a lot under his guidance. I would also like to thank my Ph.D. committee consisting of Dr. Carole Kimberlin, Dr. Teresa Kauf, and Dr. I Chan Huang for providing expertise and positi ve, constructive feedback. I am also thankful to Dr. Richard Segal who provided me constant support guidance and appreciation for the work I did. I believe unwavering support and encouragement from my family has been the key ingredient in all the successe s throughout my life. I am and always shall remain short of words to thank my parents for all their love, care, and personal sacrifices. They have always supported and believed in me through everything. I would like to thank my younger sister and her husba nd for their constant support and a source of inspiration in my life. In the end I would like to thank my wife Poonam for her unconditional support, understanding, and for putting faith in me. She completes me and has always supported and believed in me.

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 8 LIST OF FIGURES ................................ ................................ ................................ ........ 10 LIST OF ABBREVIATIONS ................................ ................................ ........................... 11 ABSTRACT ................................ ................................ ................................ ................... 14 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .... 16 Background ................................ ................................ ................................ ............. 17 Types of Response Shift ................................ ................................ .................. 17 Importance of Response Shift in Clinical Trials ................................ ................ 18 Hypertension and Coronary Artery Disease as Chronic Conditions ........................ 19 Structural Equation Modeling and Response Shift ................................ .................. 20 Measurement Perspective of Response Shift ................................ ......................... 21 Importance of Response Shift Assessment in This Study ................................ ....... 23 Study Objectives ................................ ................................ ................................ ..... 25 Study Significance ................................ ................................ ................................ .. 26 Research Questions ................................ ................................ ............................... 27 2 LITERATURE REVIEW ................................ ................................ .......................... 29 Unification of Incoherent Findings in Quality of Life and Emergence of Response Shift ................................ ................................ ................................ .... 29 Summary of Studies that Identify Response Shift ................................ ................... 30 Importance of Response Shift Assessment among Hypertensive CAD patients .... 31 Impact of Hypertension and CAD on QOL ................................ ........................ 33 Impact of Antihypertensive Treatment Strategies on QOL and Need to Examine Response Shift ................................ ................................ ............... 33 .......... 35 ................................ ................................ ................ 36 Oor ................................ ................................ ..................... 37 .............................. 38 Measurement Bias and Response Shift ................................ ................................ .. 39 3 METHODS ................................ ................................ ................................ .............. 42 Subject and Data Collection Procedure ................................ ................................ .. 42 Source of Data ................................ ................................ ................................ 42

PAGE 6

6 SF 36 HRQoL ................................ ................................ ................................ ... 43 Statistical Approach for Research Question 1 ................................ ........................ 44 Oort ................................ ................................ ....................... 44 Procedure ................................ ................................ ................................ .. 44 Model 1: Establishing an Appropriate Measurement Model ....................... 44 Model 2: No Response Shift Model ................................ ............................ 44 Model 3: Testing for Response Shift ................................ .......................... 45 Model 4: Identification of True Change and Effect Size Calculation ........... 45 Controlling for Type I Error ................................ ................................ ............... 47 Model Evaluation ................................ ................................ .............................. 47 Sample Size Consideration ................................ ................................ .............. 50 ................................ ................................ .................. 50 Missing Data Evaluation ................................ ................................ ................... 51 Statistical Analysis for Research Question 2 ................................ .......................... 52 Variables ................................ ................................ ................................ .......... 52 Hypothesis#1 ................................ ................................ ............................. 53 Hypothesis#2 ................................ ................................ ............................. 53 Hypothesis#3 ................................ ................................ ............................. 53 Hypothesis#4 ................................ ................................ ............................. 53 Explanatory variables (E) ................................ ................................ ................. 53 Potential violator variables (V) ................................ ................................ .......... 54 Step 1: Establish ing a Measurement Model ................................ ............... 55 Step 2: Detecting Measurement Bias and Response Shift in Measurement ................................ ................................ .......................... 56 Missing Data Evaluation ................................ ................................ ................... 57 4 RESULTS ................................ ................................ ................................ ............... 61 Question 1) ................................ ................................ ................................ .......... 61 Description of the Population ................................ ................................ ............ 61 Comparison of Treatment Groups at Baseline ................................ ................. 61 Measurement Model ................................ ................................ ......................... 61 ................................ ..................... 62 Detection of Response Shift and True Change ................................ ................ 62 Step 1: Measurement Model ................................ ................................ ...... 63 Step 2: No Response Shift Model ................................ .............................. 63 Step 3: Identification of Resp onse Shift ................................ ..................... 63 Step 4: Final Model ................................ ................................ .................... 64 Evaluation of Response Shifts and True Change ................................ ............. 64 Recalibration Response Shift ................................ ................................ ..... 64 Contributions of Response Shifts and True Change to Change in the Observed Variables ................................ ................................ ................ 65 Impact of Response Shifts on the Measurement of True Change .............. 65 ................................ ................ 65 Missing Data Evaluation and Imputation ................................ ................................ 66 Measurement Bias and Response Shift (Research Question 2) ............................. 68

PAGE 7

7 Description of the popula tion ................................ ................................ ............ 68 Comparison of Treatment Groups at Baseline ................................ ................. 68 Step 1: Measurement Model ................................ ................................ ...... 69 Step 2: Measurement Bias and Response Shift ................................ ......... 69 True Change ................................ ................................ .............................. 72 Contributions of Response Shifts and True Chang e to Change in the Observed Variables ................................ ................................ ................ 73 Missing Data Evaluation and Imputation ................................ ................................ 73 5 DISCUSSION ................................ ................................ ................................ ....... 104 Difference in Results between Oort and Schmitt Procedures in this Study ........... 105 Difference between the Present Study and Ahmed et al. Study ........................... 107 Validating the Presence of Response Shift in Hypertensive CAD Patients ........... 108 Identification of Response Shift for the PF Domain ................................ .............. 112 Implications ................................ ................................ ................................ ........... 114 Limitations ................................ ................................ ................................ ............. 117 For Research Question# 2 ................................ ................................ .................... 119 Implications ................................ ................................ ................................ ........... 122 Limitations ................................ ................................ ................................ ............. 124 6 CONCLUSION ................................ ................................ ................................ ...... 126 LIST OF REFERENCES ................................ ................................ ............................. 129 BIOGRAPHICAL SKETCH ................................ ................................ .......................... 137

PAGE 8

8 LIST OF TABLES Table page 1 1 ........................ 28 3 1 SF 36 HRQoL measure ................................ ................................ ...................... 59 3 2 Weight s assigned to different comorbid states in the modified Charlson Index .. 59 3 3 Analytical approach ................................ ................................ ............................ 60 4 1 Comparison of patients at baseline assigned to the verapamil SR strategy and the atenolol strategy among the SADD Sx patients returning both baseline and one year surveys (N = 909) ................................ ........................... 78 4 2 Testing for normality in HR QoL indices ................................ .............................. 78 4 3 Means, standard deviations, and effect sizes for SF 36 scales at baseline and one year (n = 909) ................................ ................................ ....................... 79 4 4 Oort m 2 differences between models without controlling for Type I error (n = 909) ................................ ................................ ... 80 4 5 2 differences between models when controlling f or Type I error (n = 909) ................................ ................................ ... 81 4 6 Parameter estimates in the final model (Model 4, Table 4 5) (n = 909) .............. 82 4 7 Significance tests of response shifts, and effect sizes of observed change, response shift, and true change in the final model (From tables 4 5 and 4 6) .... 83 4 8 2 differences between models (n = 909) 84 4 9 ................................ ......... 85 4 10 Comparison of patients at baseline assigned to the Verapamil SR strategy and Atenolol led treatment strategy among SADD Sx patients returning both baseline and one year surveys (n = 788) ................................ ............................ 86 4 11 Means, standard deviations, and effect sizes for SF 36 scales (n = 788) .......... 87 4 12 Goodness of fit of models in measurement bias and response shift detection procedure (when age was used as a continuous variable) (n = 788) ................. 88 4 13 Goodness of fit of models in measurement bias and response shift detection procedure (when age was used as a dichotomous variable) (n = 788) .............. 89 4 14 Parameter estimates in the final model (Model 2, Table 4 13) (n = 788) ............ 90

PAGE 9

9 4 15 Direct effects indicating measurement bias and response shift in measurement (based on Table 4 13) ................................ ................................ 91 4 16 Significance tests of response shifts, effect sizes of observed change, response shift, and true change in the final model (Tab le 4 13 and Table 4 14) ................................ ................................ ................................ ...................... 91 4 17 Oort method: Imputation using approach 1 (PF and MH present and other domain scores imputed using MCMC) (n=988) ................................ .................. 92 4 18 Oort method: Imputation using approach 2 (PF imputed with IADL scores and remaining domain scores imputed using MCMC) (n=1057) ................................ 93 4 19 Oort method: Imputation using app roach 3 (PF imputed with MCMC approach and remaining domain scores with IADL scores) (n=1037) ................ 94 4 20 Oort method: Imputation using approach 4 (N = 2317) ................................ ...... 95 4 21 Schmitt method: Imputation using approach 1 (PF and MH present and other domain scores imputed using MCMC) (n=988) ................................ .................. 96 4 22 Schmitt method: Imputation using ap proach 2 (PF imputed with IADL scores and remaining domain scores imputed using MCMC) (n=1057) ......................... 97 4 23 Schmitt method: Imputation using approach 3 (PF imputed with MCMC approach and remaining domain scores with IADL scores) (n=1037) ................ 98 4 24 Schmitt method: Imputation using approach 4 (N = 2317) ................................ .. 99 4 25 Measurement bias and response shift detection (when age was used as a dichotomous variable) using imputation approach 1 (n = 988) ......................... 100 4 26 Measurement bias and response shift detection (when age was used as a dichotomous variable) using imputation approach 2 (n = 1057) ....................... 101 4 27 Measurement bias and response shift detection (when age was used as a dichotomous variable) using imputation approach 3 ( n = 1037) ....................... 102 4 28 Measurement bias and response shift detection (when age was used as a dichotomous variable) using imputation approach 4 (n = 2317) ....................... 103

PAGE 10

10 LIST OF FIGURES Figure page 1 1 Graphical representation of measurement bias and response shift .................... 22 2 1 Graphical represent ation of measurement bias and response shift .................... 40 4 1 Sx substudy of the International Verapamil SR Trandolapril Study (n = 2317). ................................ 75 4 2 The measurement model used in response shift detection ................................ 76 4 3 Sx substudy of the International Verapamil SR Trandolapril Study (n = 2317). ................................ 77

PAGE 11

11 LIST OF ABBREVIATION S A Attribute variables BP Bodily Pain CABG Coronary artery bypass graft CAD Coronary artery disease CES D Center for Epidemiologic Stu dies Depression CFA Confirmatory factor analysis CFI Comparative fit index CHF Congestive heart failure CI Confidence Interval COPD Chronic Obstructive Pulmonary Disease d Effect size DBP Diastolic blood pressure df Degrees of freedom Diag ( ) Residual fac tor variances E Explanatory variables GH General Health H a Alternative hypothesis H 0 Null hypothesis HRQoL Health related quality of life IADL Instrumental activities of daily living ICF International Classification of Function INVEST International Verapa mil Trandolapril Study JNC VI Joint National Committee on Prevention, Detection, Evaluation, and Treatment of High Blood Pressure LISREL Linear Structural Relationship

PAGE 12

12 MCMC Markov Chain Monte Carlo MENT Mental Health construct MH Mental Health MI Myocardia l infarction MLE Maximum likelihood estimation n Sample size NFI Normed fit index NHP Nottingham Health Profile NNFI Non normed fit index p Significance value P Number of indicator variables Patt ( ) Common factor loadings PF Physical Functioning PHYS Phys ical Health construct q Number of parameters to be estimated QOL Quality of life RMSEA Root mean square error of approximation RP Role Physical SADD Sx Study of Antihypertensive Drugs and Depressive Symptoms SBP Systolic blood pressure SD Standard deviati on SEM Structural equation modeling SF Social Functioning SF 36 Short form 36 SRMR Standardized root mean square residual

PAGE 13

13 Tau ( Intercepts UF University of Florida V Potential violator variables VT Vitality X Domain scores z scores Standardized scores 1 Common factor means at baseline 2 Common factor means at one year blocker Beta blocker 2 1 Observed change 2 Chi squar e

PAGE 14

14 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy ASSESSMENT OF RESPONSE SHIFT IN PATIENTS TREATED WITH A CALCIUM ANT AGONIST OR AN ATENOLOL LED HYPERTENSION STRATEGY IN THE INTERNATIONAL VERAPAMIL TRANDOLAPRIL STUDY By Pranav Kirit Gandhi December 2010 Chair: L. Douglas Ried Major: Pharmaceutical Sciences P harmaceutical Outcomes and Policy N umerous reliable and v alid measures for HRQoL assessment s are available but these assessments may be subject to change over time or after an intervention from factors separate from the effects of the intervention itself. These changes are referred to as response shifts. We inve stigated the occurrence of a response shift using short form 36 ( SF 36 ) HRQoL instrument in hypertensive patients with coronary artery disease (CAD) First, two structural equation modeling (SEM) techniques were compared and examined for convergence to exa mine plausible occu rrence of response shift over two time points (i.e. at baseline and at one year). We provide alternate explanations for the divergent results we found comp ared to the Ahmed et al. study. Our study identified recalibration for the SF 36 physical function (PF) scale using both Oort and Schmitt approaches. Based on our results, the differences between the Oort and Schmitt approaches in our study may be due to variation in the method and not the sample used. Method differences may result fro m divergence in defining the type of response shift linked to changes in various parameters of the SEM measurement model. Our find ing of response shift may be explained by a single or a

PAGE 15

15 combination of factors including the disease condition, assignment of antihypertensive treatment strategies age, gender, and /or the presence of depressive symptoms Second, we investigated measurement bias and response shift jointly with use of SEM over a one year period. We found that gender caused response shift in measur ement. Women have a higher incidence of depression, and in combination with lower PF scores would worsen their quality of life ( QOL ) Thus, it is imperative to improve depression willingness to perform daily activities which may help reduce institutionalization. Our results suggest that the SF 36 PF scale may be susceptible to response shift this deserves further research. By looking more closely at the scores for SF 36 PF domain in this stu dy population will enable us to provide nuanced attention and direct treatment for the most impaired aspects of QOL.

PAGE 16

16 CHAPTER 1 INTRODUCTION Health related quality of life (HRQoL) is self perceived and can change over time. e assigned to duration of life as modified by the impairments, functional states, perceptions and social opportunities influenced by report measures of HRQoL are increasingly becoming part of the assessment p rofile for impact of drug interventions in randomized trials. Although numerous reliable and valid measures for HRQoL assessments are available, these assessments may be subject to change over time or after an intervention from factors separate from the ef fects of the intervention itself [2]. In other words, individuals experiencing change in health status owing to improvement or deterioration may adopt a different frame of reference over time and reevaluate the importance of fundamental elements of HRQoL [ 2]. These changes are referred to as response shifts. Response shifts refers to a change in meaning of quality of life (QOL) over time. Consequently, self report assessments may over or underrate the true effects of the intervention [3]. The identificatio n of response shift may enable the assessment of relative importance of HRQoL outcomes from the point of view of patients and offer valuable insights into cost effectiveness analysis of the various intervention strategies. The w ork here investigated the oc currence of a response shift in hypertensive patients with coronary artery disease (CAD) treated with antihypertensive treatment strategies. First, two structural equation modeling (SEM) techniques were compared and examined for convergence to examine plau sible occurrence of response shift in the study population over two time points (i.e. at baseline and at one year). Second, we

PAGE 17

17 investigated measurement bias and response shift jointly with use of SEM in the study population over a one year period. The pre sent study utilized short form 36 (SF 36) measurement instrument for HRQoL to achieve both objectives aforesaid. Background Types of Response Shift In the beginning, response shift research was based in evaluation of educational training, organizational ch ange, and management science. More recently, the concept has been utilized in HRQoL research [4 evaluation of a target construct as a result of (a) a change in l standards of measurement (i.e. scale recalibration); (b) a the importance of component domains constituting the target construct, reprioritization) ; or (c) a redefinitio n of the target construct (i.e. reconceptu Reconceptualization. With reconceptualization, the patient revises the meaning of the item content. For instance, during post hip replacement surgery, an individual may consider the ability to work as important to his HRQoL. Prior to the injury, however, reconceptualization [4,5]. Recalibration. With recalibration, the patient item response values [4,5]. For instance, in the immediate period post angioplasty, even though an individual is in pain and unable to engage in vigorous activity, he/she may judge current health as good. The same person who at 3 months post angioplasty is able to perform vigorous activities may now look back and judge his/her initial health as

PAGE 18

18 poor (scale recalibration). Recalibration can be uniform or non uniform. Non uniform recalibration occurs when a portion of the m easurement scale is recalibrated (e.g. only some response options are associated with greater pain than before). Conversely, uniform recalibration occurs when the entire scale is recalibrated (e.g. all points of the response scale are associated with great er pain than before) [4,5]. Reprioritization. With reprioritization, an individual, for example, might initially value the HRQoL concepts of social network, work and family as important [4,5]. After a health scare (i.e. change in health status), the same individual may find the same concepts important, but the order of importance (i.e. change in values) may change to family first, social network second, and lastly, work, indicating reprioritization. Importance of Response Shift in Clinical Trials Randomize d trials offer the benefit of balancing measured and unmeasured variables, which may include conceptualization of HRQoL and internal standards [2]. In the present study, in addition to hypertension, CAD and elderly age, most patients (89%) had one or more associated conditions (diabetes, dyslipidemia, cerebral or peripheral vascular disease, etc.) contributing to increased risk for adverse outcome [7,8]. Moreover, patients who deal with treatment strategies and chronic disease are faced with the necessity t o cope with the consequences of the treatment modalities and/or the illness. As a result, a response shift may have occurred in these patients. Response shift in measurement of HRQoL is defined as an adaptation to changing health [9]. It may be a beneficia l process for patients because it can help in adapting to a new situation. To measure change in HRQoL, the baseline (pretest) score is usually subtracted from the score after implementation of the intervention (posttest) [9]. The supposition that the admin istration of the same HRQoL instrument at two points in time

PAGE 19

19 or between two different groups is, in fact, measuring the same phenomenon may be their way of interpreting it h as not changed over the relevant time frame [10]. Sprangers et al. [4] state that the variability observed due to response shift may reflect shifts in an in addition to changes in actual health state. As a result, it becomes difficult to separate which component of change is due to response shift as opposed to true change in the HRQoL construct [2]. Thus, HRQoL research must take into account the importance of assessm ent of response shift in longitudinal/clinical trial studies. Hypertension and Coronary Artery Disease as Chronic Conditions Experiencing a response shift is seen as a natural respon se to change in health status. Ahmed et al. [2] suggest that response shif t may attenuate or exaggerate findings from clinical trials that incorporate HRQoL to evaluate treatment interventions. Response shift is the result of mechanisms used to accept or adapt to changes in physical, emotional, social health status [9] and natur e of impairments and activity limitations that many hypertensive CAD patients may experience. Hypertension and CAD are chronic health conditions that impact all aspects of function, perception, cogn ition, mood, and QOL [11 13]. Hypertension and CAD are hea lth conditions where symptoms can persist over a long period of time [11 13]. Literature renders support for the contention that hypertension and CAD are health conditions with changes in physical, pain, and social health status [14]; a likely scenario in which response shift may occur. To obtain patient preferences and their individual adjustment between benefits and side effects of various treatment options, estimation of response shift can be useful. For example, before

PAGE 20

20 receiving the medication for hyper tension the subjects may not have associated QOL become a major problem [15]. Hypertensive medications from different pharmacologic classes sometimes have varying side studies report that the choice of antihypertensive treatment may influence risk for depression and consequently clinical and health outcomes [11,16]. Depression is highly prevalent in h ypertensive CAD patients and is associated with poorer outcomes [11,17 19]. However, literature regarding the potential relationship between depression and blocker (beta blocker) and calcium channel antagonist) h ave inconsistent findings regarding whether medications from one or both pharmacologic categories are associated with depression [19,20]. Some studies report blockers than for calcium antagonists [21]. These conclu sions are based on findings that antidepressant prescriptions were more blocker treatment [22,23]. In other instances, these findings have not been replicated [24,25]. However, other studies report an association of calc ium channel antagonists with depression [26]. Response shift is important to consider in treatment evaluations, especially as it may serve to attenuate or to exaggerate estimates of treatment effects as patients adapt to treatment modalities or disease pro gression over time. Structural Equation Modeling and Response Shift Confirmatory factor analysis (CFA) is a type of SEM that deals specificall y with measurement models, i.e. the relationships between observed measures or indicators (e.g. test items, test scores) and latent variables or factors, i.e. variables which cannot

PAGE 21

21 be measured directly due to presence of measurement error. Schmitt has described the analysis of covariance structures approach using SEM to identify both reconceptualization and recalib ration by evaluating changes in factor structure and factor loading over time, respectively [2,27 30]. An alternate method of determining response shift described by Oort uses SEM to EM approach provides a direct measure of true change and identifies the presence of all three types of response shift, i.e. reconceptualization, reprioritization and recalibration. Oort operationalizes various types of response shift (Table 1 1) as below. Reconceptu alization Response shift is suggested by changes in factor loading patterns over time, with an observed variable loading onto one latent variable at one point in time, and at a follow up time loading onto a different latent variable. Reprioriti zation. Reprioritization is demonstrated when the factor loading of an observed variable may become stronger or weaker over time, reflecting changes in values or priorities. Recalibration. Uniform recalibration is determined by differences between intercep ts across occasions whereas non uniform recalibration is determined by differences between error variances across occasions. Measurement Perspective of Response Shift If scale scores of a test instrument are not fully determined by the common HRQoL factors (i.e. SF 36 physical and mental constructs), then it is called measurement bias [31,32]. In other words, when differences between the observed scale scores cannot be fully explained by true differences between respondents in the common HRQoL factors, it is termed as measurement bias [31,32]. In longitudinal research, when the relationships are not consistent across measurement occasions, i.e. over time, the measurement bias is considered to be response shift [31,32]. In

PAGE 22

22 longitudinal analysis, measurement bias is investigated by testing the invariance of factor loadings and intercepts [31,32]. Fig ure 1 1 Graphical representation of measurement bias and response shift All circles in Figure 1 1 represent sets of variables: A are the attributes o f interest; X are measurements of A; E are explanations of A; and V are all other variables. One sided arrows represent causal relationships and double sided arrows represent mere associations. Dashed arrows indicate measurement bias. Example: A = SF 36 ph ysical health and mental health constructs; X = domains of SF 36 self report health questionnaire; V = sex, age, race, baseline co morbid cond itions; E = change in systolic blood pressure, etc To illustrate Figure 1 1 with an example, bias is introduced from the measurement perspective when the relationship between the measurement instrument (i.e. observed variables such as SF 36 domains, denoted by X) and other variables (potential violator variables such as age, sex, race, denoted by V) cannot be full y explained by the relationships between explanatory var iable such as change in systolic blood pressure (denoted as E) to V, E to common HRQoL factors (denoted by A), A V and A X. A V X E

PAGE 23

23 Similarly, if E X relationships cannot be fully explained b y the E V, E A, A V and A X relationships, then there is measurement bias with respect to variables E. Importance of Response Shift Assessment in This Study As aforementioned, in the present study, in addition to hypertension, CA D and elderly age, most patients (89%) had one or more associated conditions (diabetes, dyslipidemia, etc.) contributing to increased risk for adverse outcome [7,8]. These multiple medical conditions place patients at high risk of depressive symptoms. Bush et al. [33] report higher mortality at levels of depressive symptoms not generally considered clinically significant and below levels usually considered predictive of increased postacute myocardial infarction ( MI ) mortality [17 ]. We presumed that as time evaluate how they were in the past, subsequently leading to occurrence of a response shift. In the absence of response shift evaluation, it cannot be determined wheth er temporal change observed in the HRQoL construct is due to true change or to changes in the structure or measurement of the HRQoL construct [34]. The effect size of HRQoL change can be estimated with the method described by Oort [10], taking response shi ft into account, to suggest the clinical importance of the response shift that occurs [34]. Some individuals may have made actual changes in HRQoL. Conversely, others may have undergone response shift, thereby influencing the effects of an intervention ove r time. Barclay Goddard et al. [34] report that presence of response shift may over or under estimate true change leading to biased estimates of the magnitude of change. Moreover, it is important to measure response shift in clinical trial research, as the

PAGE 24

24 estimates of the treatment effects may be underestimated and therefore inaccurate, possibly leading to a conclusion that represents a false negative [34]. Methods that assess response shift are not only necessary in valuating treatment effects, but also p rovide assistance for distinguishing the impact of disease over time approaches to detect response shift in chronic obstructive pulmonary disease (COPD) patients. The authors report Oort procedure to be more sensitive in detecting a response shift compared to the Schmitt procedure. However, the Ahmed et al. study [35] did not control for Type I suggeste d to increase chance of a Type I error [10]. As a result, their study may have wrongly identified a response shift when differences may be attributed to other types of changes. Contrary to the Oort procedure adopted by Ahmed et al. [35] in their study where they chose to release c onstraints on residuals, followed by intercepts, an d then factor loadings, we conduct the Oort approach as recommended and tested by several other authors including Oort [10,36,37] Donaldson [38] and Barclay Goddard [ 39 ]. Similar to the approach t ested by these authors, we test sequentially for invariance of factor loadings, intercepts, and error variance. The divergence in the Oort procedure conflicting results. Thus, our fi approaches to test if conclu sions regarding response shift we re supported in our entire dataset. A self report instrument measuring change in health status over time highlights the concern that respondents may change their frame of reference when answering the

PAGE 25

25 questionnaire or item. This may result in incomparable scores from separate measurement occasions. P revious studies [2,27,36,37,39 ] report paradoxical and inconsistent findings when using the SF 36 measur ement instrument to examine response shift. The presence or absence of response shift in one or more of the SF 36 domains may thus be a reflection of an explanatory or violator variable explained previously (shown in Figure 1 1). Our second goal was to hig hlight the differences between the measurement perspective about bias and response shift in HRQoL data from hypertensive CAD patients using SEM approach suggested by Oort [31]. When investigating from the measurement perspective, the SF 36 scales should me asure the physical and mental constructs exclusively. If, however, differences between respondents in SF 36 scales scores cannot be fully explained by true differences between respondents in the physical and mental constructs, it indicates measurement bias [31,32]. In other words, the scale scores will not only be indicative of the common HRQoL factors [31,32] but also of some other variable such as treatment strategies or comorbid conditions. The SF 36 instrument may thus include domain/s to which a variab le responds differently to. This study is the first to illustrate the detection of bias and response shift from the measurement perspective in hypertensive CAD 36. Study Objectives To determine convergent validity among Oo examining response shift in the entire sample of hy pertensive CAD patients over a one year period. To investigate measurement bias and response shift in measurement jointly in HRQoL data from hypertensive CAD patients usi ng SEM over a one year period.

PAGE 26

26 Study Significance To our knowledge, this is the first study to assess the occurrence of response shift in hypertensive CAD patients assigned to antihypertensive treatment strategies. A previous study that demonstrated the co approaches to identify response shift did not account for Type I error [35]. In addition, they chose to release constraints on residuals, followed by intercepts, and then factor loadings. This approach is divergent compar ed to the Oort approach recommended and tested by several other authors including Oort [10] Barclay Goddard [39] and Donaldson [38] We thus chose to compare the two statistical techniques to test if conclu sions regarding response shift we re supported in our dataset, controlling for Type I error and using the order of testing for response shift as suggested by Oort which will strengthen the acceptability of the results. The comparison of the said statistical approaches will assist in developing a proposed set of HRQoL measurement recommendations under circumstances where response shift is expected to occur in a specific patient popula tion. Previous studies [2,27,36,37,39 ] report paradoxical and inconsistent findings when using the SF 36 measurement instrum ent to examine response shift. Though both the antihypertensive treatment strategies in the International Verapamil Trandolapril Study (INVEST) trial were shown to be clinically p harmacological properties and side effects, occ urrence of stroke or MI over a one year period, depression scores, self reported prior history of depression, or the presence of comorbid conditions leading to varying effects on HRQoL (as shown in Figure 1 1) Barclay Goddard suggests that researchers need to come to a generally accepted definition of response shift, using prior work as a basis. Based on these

PAGE 27

27 bias and response s observed scale scores cannot be fully explained by true changes in the latent construct that we want to measure. In other words, the SF 36 scale scores may not only be indicative of the latent c onstruct but also of the variables such a s treatment strategies, age, or sex among others Research Questions Was convergent validity observed when investigating response shift between pertensive CAD p atients over a one year period? Was measurement bias and response shift observed using SEM in hy pertensive CAD patients over a one year period?

PAGE 28

28 Table 1 1. Matrix Parameter Type of change Response shift Common factor loadings Reconceptualization Common factor loadings Reprioritization Intercepts Recalibration (uniform) Residual factor variances Recalibration (non uniform)

PAGE 29

29 CHAPTER 2 LITERATURE REVIEW Unification of Incoherent Findings in Quality of Life and Emergence of Response Shift Howard et al. [40 ] introduced the term response shift when investigating educational training interventions. Prior to their study, applicati on of response shift was found i n organiza tional change within the field of management scienc es, where Golembiewski et al [41 ] discussed three types of observed changes in self report of some existential state, given a constantly calibrated measuring instrument related to a constant conceptual domain [41 change, typically measured with pretest posttest study designs [41 involves a variation in the leve l of some existential state, complicated by the fact that some intervals of the measurement continuum associated with a constant conceptual d omain have been recalibrated [41 reconceptualization of some domain, a major change in the perspective or frame of reference within which phenomena are perceived and classified, in what is taken to be relev ant in some slice of reality [41 Sprangers and Schwartz [4,5] applied the idea of response shift to the field of HRQoL research. They combined and extended the previous definitions of response shift and proposed the following definition: response shift refers to a change in the evaluation of a target construct as a result of: (1) recalibration, that importance of component domains constituting the target construct); or (3)

PAGE 30

30 reconceptualization that is, a redefinition of the target construct. A change in values (termed as reprioritization) was not described in the alpha, beta, gamma framework, but was added by Schwartz and Sprangers in 1999 [4,5]. Over the years, contradictory findings about wh at QOL measures and how scores obtained from these measures need to be interpreted have raised several questions. For example, people with severe chronic illnesses report QOL equal or superior to less se verely ill or healthy people [42 44 ] and inconsistenc ies persist between clinical measures of heal th and patients' self report [42,45,46 ]. Some studies on HRQoL found that patients score better than healthy people do, or that significantly disabled or terminally ill people report similar or higher levels of QOL after deteriorating health. These counterintuitive findings may reflect individual differences and intra individual changes in internal standards, values, and meaning of QOL due to response shift. QOL research has a lot to gain by using methods that in clude the response shift phenomena. Theoretical and empirical evidence on response shifts in QOL support the concept that differences in evaluation enter into all self report assessments of QOL. Next, we summarize the QOL studies that have identified respo nse shift followed by studies that merit response shift assessment in our study population. We then demonstrate the literature illustrating use of Schmitt and Oort SEM techniques to detect response shift. Lastly, we elaborate on the issue of measurement bi as and response shift in presence of explanatory and violator variables. Summary of Studies that Identify Response Shift Several studies [47 53 ] document the presence and importance of response shift in both treatment outcome research and naturalistic long itudinal observations of QOL. However, only a few studies are cited here. Researchers are recommended to refer to

PAGE 31

31 Rapkin and Schwartz [42 ] for other studies that examine the presence of response shi ft. Sprangers and colleagues [47 ] report occurrence of rec alibration for fatigue in two subgroups of cancer patients undergoing radiotherapy: in patients experiencing diminishing levels of fatigue and in patients facing early stages of adaptation to increased levels of fati gue. Schwartz and colleagues [52 ] report a harmful QOL effect of a psychosocial intervention was due to recalibration and reconceptual ization. Rees and colleagues [53 ] report a 10% increase in QOL with addition of recalibration in prostate cancer patients. Several researchers have argued that be ta change potentially threatens the internal validity of judgments about change based on self reports [54 58 ]. King et al. [59 ] compared overall HRQoL of stroke survivors to that of normative controls and found that HRQoL scores were comparable. A comparis on of studies evaluating the outcome of stroke found that persons with major stroke reported higher levels of HRQoL than indi viduals with moderate stroke [59,60 ]. In one other study, patients with cancer report better levels of HRQoL than the general popul ation [59,61 ]. People with spinal cord injuries were found to report only slightly worse levels o f HRQoL compared to controls [59,62 ]. Primarily, these findings provide substantiation that people's goals and values continue to evolve during progression of disease, thus warranting assessment of response shift in longitudinal research and clinical trials. Importance of Response Shift Assessment among Hypertensive CAD patients Hypertension and CAD are chronic health conditions that impact all aspects of functi on, perception, cognition, mood, quality of life [11 13]. In a study by Mitchell et al. [14], poorer mean scores were reported in hypertensive CAD patients when using the Nottingham Health Profile (NHP) questionnaire. The study was conducted among

PAGE 32

32 elderly subjects that had differing cardiovascular status (cardiovascular normal, being hypertensive, having isolated CAD, or both being hypertensive and having CAD). The NHP questionnaire incorporates questions related to activity limitations and impairments simi International Clas sification of Function (ICF) [63 ]. Under the ICF model, difficulties with psychological functioning include impairments in cognitive or emotional functioning. Impairmen ts such as decreased strength cause limitations in activities, such as walking, which eventually may lead to participation restrictions, i.e. restricted ability to carry out usual activities in the community, such as through work or volunteering. Response shift can result due to mechanisms used to accept or adapt to changes in health status and nature of impairments and activity limitations that many hypertensive CAD individuals are likely to experience. In addition to hypertension, CAD and elderly age, mos t INVEST trial patients (89%) had one or more associated conditions (diabetes, myocardial infarction, abnormal coronary angiogram, etc.) contributing to increased risk for adverse outcome [7,8]. These patients with comorbidities place them at high risk of depressive symptoms. An essential outcome of hypertension and CAD, and their treatment is mental depression [17]. Literature supports the notion that depression is a risk factor for death and myocardial infarction among patients with CAD [17,33]. Moreover, higher mortality has been observed at levels of depressive symptoms not generally considered clinically significant and below levels usually considered predictive of increased postacute MI mortality [17,33].

PAGE 33

33 Impact of Hypertension and CAD on QOL QOL in p atients with hypertension has been extensively evaluated in recent y ears. A study by Wang et al. [64 ] report hypertensive patients to score lower in five SF 36 domains compared to normotensives. The authors observed hypertensive subjects with comorbidities to score lower SF 36 scores compared to those without comorbidities. Boini et al. [65 ] report that CAD patients were more likely to score lower in four of the eight SF 36 domains. The authors agree that the study of HRQoL in CAD patients may be used to in form treatment decisions to practitioners. Cost effectiveness studies of antihypertensive modalities often utilize HRQoL outcomes, thereby suggesting the need exist a pos Therefore, facilitating response shift assessment is particularly important in hypertensive CAD patients when impairments and activity limitations are not expected to recover fully, but improved HRQoL is a goal. Impact of Antihypertensive Treatment Strategies on QOL and Need to Examine Response Shift Antihypertensive treatment focuses on individual attributes/domains of HRQoL which are important to the individual hypertensive CAD pat ient. Croog et al. [13] have demonstrated that effects on QOL vary among classes of antihypertensive agents: captopril improved QOL, whereas methyldopa and propranolol worsened it. In a separ ate study, Testa et al. [66 ] conclude that two angiotensin conver ting enzyme inhibitors, captopril and enalapril, the same according to clinical assessments of efficacy and safety, had varying effects on QOL. Assessment of HRQoL in cardiovascular clinical trials is important to provide a more complete understanding of t reatment effects.

PAGE 34

34 Hypertensive medications from different pharmacologic classes sometimes have studies report that the choice of antihypertensive treatment may influence ris k for depression and consequently clinical and health outcomes [11,16]. Response shift is thus important to consider in treatment evaluations, especially insofar as it may serve to attenuate or to exaggerate estimates of treatment effects as patients adapt to treatment individual adjustment between benefits and side effects of various treatment options, estimation of response shift can be useful. Based on INVEST trial [8] results patients in the verapamil SR group reported constipation more frequently than in the atenolol led g roup. A study by Wald et al. [67 ] reports the impact of constipation on QOL in different cultural and national settings. HRQoL was assessed with th e SF 36 questionnaire. There were significant differences in HRQoL between constipated and non constipated individuals and a significant, negative correlation between the number of symptoms and complaints and SF 36 scores. The authors observed that constip ated individuals consistently reported low scores in the psychological components of the SF 36 questionnaire. On the other hand, atenolol led treatment strategy has been reported to show adverse somatic side effects such as decreased energy and increased f atigue [17]. The side effects associated with verapamil and atenolol may thus cause a change in HRQoL over time in the study population. Ried et al. [17] d emonstrate a significant improvement in depressive symptoms for those assigned to the verapami SR str ategy; whereas there was little change for those

PAGE 35

35 assigned to the atenolol strategy among a subgroup of INVEST patients. They explain that changes in depressive symptoms at one year in the verapamil SR group could be related e ffects. Analogous to their [17] findings regarding depressive status changes [17], Ried and colleagues [68 ] in a separate study report that HRQoL improved among those assigned to the verapamil SR led strategy, but found no significant change in HRQoL for those as signed to atenolol led strategy. A study by Bar On and Amir [15] identified the presence of response shift in hypertensive patients. Bar On and Amir [15] examined beta change (i.e. recalibration) over the period of one year in hypertensive and normotensive male subjects randomly assigned to methyldopa, isradipine or placebo treatments. Among the 297 subjects in their study, 11% performed one of the two scales recalibrations. Nonetheless, there were no significant differences in those who showed recalibratio n between normotensives and hypertensives. When, however, the subjects who showed severe scale recalibration were excluded from the comparison, the difference between the normotensives and hypertensives became significant. To summarize, several studies dem onstrate that taking antihypertensive medications over a one year period is potent enough to affect the QOL assessments. Nonetheless, there is modicum of evidence that response shift may be detected in these patients over time. Assessment of Response Shift There are two broader alternatives for detecting response shift, the first is to address response shift phenomenon from imposing study design changes; second alternative is to address response shift from a statistic al framework [5]. The study design methods are considered time consuming and burdensome f or the individual

PAGE 36

36 involved [5,69 ]. Statistical methods are often used where it is impractical to incorporate response shift evaluation into the study design or when do ing secondary data analysis. The analysis of covariance structures approach using CFA, a statistical framework approach [28] identified both reconceptualiz ation and recalibration by evaluating changes in factor structure and factor loading over time, respectively. Ahmed et al. 36 measurement model over time among 238 individuals with stroke a nd 392 controls, separately. Response shift was not identified between one and six months even though it was suggested using individualized methods [2]. It was later argued that a reason for not finding response shift with SEM model in Ahmed et al. study [ 2] was related to the measurement model used in their study. demonstrate the presence of gamma change (i .e. reconceptualization) with use of gamma and beta change by analyzing data collected from 116 individuals who had lost their job, before and after they had secured sub sequent employment. In one other study [29], three methods to evaluate response shift were compared coefficient of congruence (assessment of the similarity between the factor structures of before and after measures), then approach. Coefficient of reconceptualization response shift. In the same study, the then

PAGE 37

37 technique both showed indications of recalibration; the coefficient of congruenc e method did not. The authors recommend the use of then compared to the coefficient of congruence technique to identify response shift. A more contemporary work by Oort and colleagues [10] attempt to rectif y the issues inherent in previous research using covariance analysis. Oort [10] investigated all three types of response shift by considering variation in the factor loading patterns to reflect reconceptualization, significant differences in the values of the factor loadings to reflect reprioritization, and significant differences in the specific factor mean to reflect recalibration. patients identify the presence of response sh ift for five SF 36 scales: reconceptualization of general health (GH) reprioritization of social functioning (SF) and recalibration of role physical (RP) bodily pain (BP) and vitality (VT) In one other study, Visser et al. [37] demonstrate the converg ent validity of the then test, anchor SF 36 and a Multidimensional Fatigue Inventory. Results showed agreement between the then test and SEM approach on the absence (si x scales) and presenc e (two scales, SF 36 BP and RP ) of response shift in eight of the nine scale s. For the n inth scale (SF 36 GH ) both methods detected response shift, but in opposite directions. However, the anchor recalibration task agreed with the othe r approaches on only the absence of EM approach, Barclay Goddard [39 ] identify the presence of response shift among 678 individuals at 1, 3, 6, and 12 months post stroke in mental healt h (MH) construct using multiple

PAGE 38

38 measures. Uniform r ecalibration was identified at one y ear in the SF 36 role emotional (RE) and preference based stroke index self esteem. In addition, non uniform recalibration was identified at 6 months in SF 36 MH and at one year in EQ 5D anxiety /depression, SF 36 MH and Stroke Impact Scale emotional. In recent past, although the Schmitt and Oort SEM approaches were the most often used statistical techniques to ident ify response shift, only Ahmed et al. [35] demonstrate the convergence among the said techniques. The authors found Oort procedure to be more sensitive in detecting a response shift compared to the Schmitt procedure in chronic obstructive pulmonary disease patients. However, the Ahmed et al. [35] study did not control for Type I been suggested t o increase the chances of Type I error. As a result, their study may have wrongly identified a response shift when diff erences may be attributed to other types of changes. In addition, similar to Nolte et al. [70] Ahmed et al. [35] chose to release constraints on residuals, followed by intercepts, and then factor loadings. This approach wa s divergent compared to the Oort approach recommended and tested by several other authors including Oort [10] Barclay Goddard [39] and Donaldson [38] Moreover, Ahmed et al. [35] did not illustrate any approach to assess validation of the mode l. As Barclay Goddard et al. [39 ] point out in their framework article, SEM requires validation of the model to ensure that it is truly a good model for the population and not just the sample. We thus chose to compare the two statistical techniques to test if conclusions regarding response shift are supported in our dataset, controlling for Type I error, releasing constraints as tested by several studies, and illustrating an approach to assess apparent validation which will strengthen the acceptability of the results. The

PAGE 39

39 next paragraph briefly exemp approaches in assessment of response shift over time. The SEM method described by Schmitt does not identify reprioritization (since the covariance analysis used was incapable of identifying changes in d efinition from change in values) or non uniform recalibration and it interprets changes in factor loadings differently from the approach described by Oort. In the approach described by Schmitt, constraints are added to the model to identify response shift; in the approach described by Oort, constraints are removed to identify response shift. The constraints added or removed are equality constraints, where a parameter estimate at one time is made to be equal to another time. Given the inconsistent reports fr om individual studies using comparing the said statistical methods justify the need to compare these methods to identify response shift in our patient population. Measure ment Bias and Response Shift Previous sections in this chapter summarize HRQoL studies that report patients score better than healthy people. On the other hand, some HRQoL st udies [31,43,71 ] report that patients may score better after worsening health sta te. The self report and multidimensional nature of HRQoL thus necessitates the idea of measurement bias and response shift particularly relevant to the evaluation of the said concepts over time. Patients undergoing a change in health status over time may h ave different frames of reference when answering a questionnaire or item [31]. Consequently, the measurement may be biased. As Oort et al. [31] explains, when observed differences 36 domain scores reflect something other than true d ifferences in the latent construct (i.e. HRQoL), it indicates measurement bias. Oort and

PAGE 40

40 King Kallimanis [31,32] in separate studies explain response shift as a special case of measurement bias which occurs when investigating the change in latent construct With other than true changes in the latent construct that we want to measure. King data with SF 36 to investigate measurement bias and response shift with a simplified model (Figure 2 1), i.e. with use of potential violator variables (V). Six measurement biases were found, five of which were considered response shift. The GH scale appeared more s usceptible to response shift in their study. Moreover, they also found patients to have reconceptualized their perception of BP and GH scales. However, to detect measurement bias and response shift using multiple explanatory and violator variables, we chos e to empirically investigate measurement bias and response shift using a joint model illustrated in Figure 1 1. To the best of our knowledge, ours is the first study to empirically investigate the presence of measurement bias and response s model as illustrated in Figure 1 1. Figure 2 1 Graphical representation of measurement bias and response shift To con clude, several studies [2,27,36,37,39 ] cited in preceding sections report paradoxical and inconsistent findings when using SF 36 measurement instrument to V V A A X X

PAGE 41

41 examine response shift. Though both the antihypertensive treatment strategies in the INVEST trial were shown to be clinically equal [8], the presence of response shift on one or more of the SF 36 domains may reflect medicati properties and side effects, presence of comorbid conditions, or the disease condition itself leading to varying effects on HRQoL. The SF 36 instrument may demonstrate a domain to which a variable (e.g. treatment strategies, age, comorbid conditions, etc.) may respond differently to. In other words, the SF 36 scale scores may not only be indicative of the latent construct but also of the variable, thus warranting the need to assess measurement bias and response shift in our s tudy population.

PAGE 42

42 CHAPTER 3 METHODS Subject and Data Collection Procedure Source of Dat a A description of the Study of Antihypertensive Drugs and Depressive Symptoms (SADD Sx) is reported elsewhere [17]. Briefly, SADD Sx was a substudy of INVEST INVEST w as a randomized, open label, blinded end point study of 22,576 hypertensive patients with CAD aged >50 years conducted from September 1997 to February 2003 [7,8]. Patients were randomized to antihypertensive treatment with either a verapamil SR or atenolo l based strateg y to achieve blood pressure control according to the sixth report of the Joint National Committee on Prevention, Detection, Evaluation, and Treatment of High Blood Pressure (JNC VI) [72 ]. SADD Sx patients residing in the United States were m ailed surveys between April 1, 1999, and October 31, 1999 (N = 2317). SF 36 HRQoL and demographic characteristics such as age, gender and race were included in the SADD Sx survey and mailed to each study subject. Patients were mailed baseline surveys the d ay after randomization and follow up surveys at six months and at one year. If surveys were not returned within 10 working days, they were mailed a second survey. Approximately 2 weeks before follow up surveys were mailed, a letter asking for continued sup port was mailed to each patient to enhance response rate. If patients failed to respond to the second survey, no added attempts were made to contact them for the purposes of SADD Sx. Only patients with complete survey responses from both the baseline and o ne year follow

PAGE 43

43 SADD Sx was conducted according to the principles of the Declaration of Helsinki. The University of Florida (UF) Institutional Review Board approved the study protocol. SF 3 6 HRQoL The SF 36 is a generic measure (Table 3 1) of perceived HRQoL health status that incorporates behavioral functioning, subjective well being and perceptions of health by assessing eight health concepts: Physical Functioning (PF) (10 items), Role Phy sical (RP), limitations in role activities due to physical health problems (4 items), Bodily Pain (BP) (2 items), General Health (GH) (5 items), Vitality (VT) (energy & fatigue) (4 items), Social Functioning (SF) (2 items), Mental Health (MH) (5 items), Ro le Emotional (RE), limitations in usual role activities due to emotional problems (3 items) [73,74 ]. All items are measured using Likert type scales. Response choices varied and range from two to six levels. The transformed scores on all of the health conc ept scales range from 0 to 100 with higher sco res indicating better health [73,74 ]. The SF 36 scales formed the basis of developing the measurement models for each construct. The eight SF 36 domains used in the measurement model were assessed for normality 2 test of overall of goodness of fit, and standard errors for all parameter estimates when data is normally distributed, using the un transformed scores of the SF 36. On the other hand, if the data wa s not normally distributed, the robust MLE method was used to account for non normality in data. The Satorra 2 is the commonly used fit statistic with the robust MLE method. When data are continuous or ordered categorical non normal data, the S atorra 2 2 from non robust methods [75 ].

PAGE 44

44 Statistical Approach for Research Question 1 Relationship (LISREL) ver sion 8.8 [76 ]. The latent factors in the model were the physical health (PHYS HRQoL) and mental health (MENT HRQoL) constructs. The eight scales of the SF 36 were the observed variables explained by the latent variables. Procedure The framework here was built on the meth odology illustrated by Oort [10]. Specification of response shift detection and true change was done in four steps [10]: (1) establishment of an appropriate measurement model, (2) fitting a model of no response shifts, (3) detection of response shifts, and (4) assessment of true change. Model 1: Establishing an Appropriate Measurement Model Based on the content of the items, results of an exploratory factor analysis and based on the published results of principal components analyses of the SF 36, we created a measurement model. Model 1 had no across occasion constraints. Model 2: No Response Shift Model Building Model 2 included constraining the values of common factor loadings (reprioritization), the intercepts (uniform recalibration), and residual variance s (nonuniform recalibration) across time. Two models were compared, the free model in which parameter estimates (paths, error variances, and intercepts) were all freely estimated, and the fully constrained model in which all of the paths, error variances, and intercepts were made to be equal across time. 2 difference test evaluated the change in overall fit of the free and fully constrained (no response shift) models. If the difference in fit between models 1 and 2 was not significant, we concluded that there

PAGE 45

45 were no response shifts, and skip model 3. 2 difference test was statistically significant, the null hypothesis that the two models fit equally well was rejected, providing evidence that response shift was present in the data. Model 3: Testing for Response Shift In Model 3 all measured r esponse shifts were accounted for. The specification search of Model 3 started with Model 2, and was be guided by modification indices and standardized residuals. Inspection of modification indices and standardized residuals indicated which of the equality con straints were not feasible. Each modification was 2 difference test. First, factor loading constraints were released. This was followed by testing of differences between intercepts (uniform recalibration) while still constraining common fa ctor loadings across time. Lastly, error variances (non uniform recalibration) were released keeping equal intercepts and common factor loadings across time. Across occasion differences between common factor loadings (in the operationalization of response shifts. That is, common factor loadings that changed from zero into non zero (or vice versa) indicated reconceptualization, other changes in common factor loadings ind icated reprioritization, changes in intercepts indicated uniform recalibration, and changes in residual variances indicated non uniform recalibration. Model 4: Identification of True Change and Effect Size Calculation To investigate change in the means, va riances, and correlations of the common factors, we fitted additional models with Model 3 as the starting point. The across occasion invariance of parameters were tested step by step, maintaining all equality constraints that proved tenable. After establis hing (partial) invariance of factor loadings,

PAGE 46

46 intercepts, and residual variances, we tested other types of invariance, including the equality of common factor means, variances, and correlations. Identification of all model parameters, scales and origins of the common factors were established by fixing the means at zero and t he variances at one. In Steps 2 4 of the detection procedure, only the baseline factor means and variances were fixed; second occasion means and variances were then identified by constra ining intercepts and factor loadings to be equal across occasions. target construct. In case of HRQoL, positive true change refers to better health or improved HRQoL. The c ommon factors were seen as operationalization of target operationalization of true change. The statistical significance of true change was tested 2 difference test. If the null 1 = 2 = 0) was rejected, then the across occasion difference between the common factor 2 1 2 were taken as a measure of true change (since the first occasion 1 was fixed at zero). E 2 difference tests were used to evaluate the statistical significance of response shifts and true change. It was also informative to evaluate the sizes of response shift and true change effects on observed chan ge. According to Oort, 2 1 is the model for observed change, and we assumed 1 = 0 (for Models 3 and 4). Observed change was decomposed into three components, 2 1 2 1 2 1 2 1 2 2 1 ), constituted th e 2 1 2 constituted the

PAGE 47

47 1 2 constituted the contribution of true change to observed change. The parameter estim ates of Model 4 (or, if preferred, those of Model 3), were used to estimate the components of observed change. The parameter estimates of the final model, Model 4, were used to calculate effect size indices for true change, as well as for the contributions of response shifts and true change to observed change. Division by the estimated standard deviation of observed change yielded effect size indices d. respectively. The response shift effect on the estimation of true change was investigated by comparing estimates of true change from a model in which response shifts were accounted for (Model 4) with estimates from a model in which response shifts were not accounted for (Model 2). Controlling for Type I Error The probabili ty of making at least one T ype I error among a family of tests (i.e., familywise error rate) increases when multiple tests are conducted [39 ]. A family can be defined to include all of the inferences that are made in a study; in the present study, all the constraints freed in the model constituted a family. The Bonferroni correction is one well known approach to control the familywise error rate to the nomi nal level of ]. Under this method, the statistical significance of each result is p is the number of tests (i.e. hypotheses) being conducted. Model Evaluation The acceptability of the fitted CFA solution was evaluated on the basis of three major aspects: (1) overall goodness of fit; (2) the presence or absence of localized

PAGE 48

48 areas of strain in the solution (i.e. specific points of ill fit); and (3) the interpretability, es. Two statistics frequently used to identify focal areas of misfit in a CFA solution are residu als and modification indices [75 ] were also applied in our study. Overall goodness of fit. The f 2 2 supports the alternative hypothesis, meaning that the model estimates do not sufficiently reproduce the sample variances and covaria n ces (i.e. the model does not fit the data well). A non 2 means there is little difference between the models, suggesting a good fit. 2 2 statisti c to its degrees of freedom decreases the effect of the 2 ; values up to 5.0 su ggest a reasonable model fit [75 ]. Fit indices can be broadly characterized as falling under three main categories: absolute fit, adjusting for model parsimony, and comparative or incremental fit. We reported at le ast one index from each category when evaluating the fit of their models. Standardized root mean square residual (SRMR) falls in the category of absolute fit. The SRMR can take a range of values between 0.0 and 1.0, with 0.0 indicating a perfect fit (i.e. the smaller the SRMR, the better the model fit). A widely used index from the category of parsimony correction is the root mean square error of approximation (RMSEA). According to a generally accepted rule of thumb, an RMSEA value below ]. As with the SRMR, RMSEA values of 0 indicate perfect fit and values very close to 0 suggest good model fit. In addition, a 90% interval is typically used to obtain confidence intervals for RMSEA.

PAGE 49

49 The confid ence interval indicates the precision of the RMSEA point estimate. Comparative fit index (CFI) is an example of comparative fit indices. The CFI has a range of possible values of 0.0 to 1.0 with values closer to 1.0 implying good model fit [75 ]. Similarly, normed fit index (NFI) and non normed fit index (NNFI) have a range of possible values of 0.0 to 1.0 with values closer to 1.0 implying good model fit [75]. Res iduals. Standardized residuals were computed u sing the LISREL 8.8 software [75 ]. Large, positiv e standardized residuals indicated that additional parameters are needed in the model to better account for the covariance between the indicators. overestimate the relations hip between two indicators to some extent. Since standardized residuals can be roughly interpreted as standardized scores ( z scores ) the z score values that correspond to conventional statistical significance levels were employed as practical cutoffs. For instance, we scanned for standardized residuals that were equal to or greater than the absolute value of 1.96 because this value corresponds larger standardized residuals because the size of the standard errors of the fitted residuals is inversely related to sample size. For this reason, we looked for larger cutof f values (i.e. 2.58) [75 ]. Mo dification Indices. Modification indices were computed for each fixed par ameter (e.g. parameters that are fixed to zero such as indicator cross loadings and error covariances) and constrained parameter in the model. The modification index reflects an approximatio 2 would decrease if the fixed or constrained parameter was freely estimated. A good fitting model should produce

PAGE 50

50 modification indices that are small in magnitude. Modification indices of 3.84 or greater suggest that the ove rall fit of the model could be significantly improved (p < 0 .05) if the fixed or constrained pa rameter was freely estimated [75 ]. Sample Size Consideration The measurement model that was used in the response shift detection procedure assumed the extreme va l ue of RMSEA = 0.08 which indicates the fit of the model is at 0.80 and assuming the degrees of freedom (d f ) to be > 20 would require a m inimum sample size of 421 [75]. Degre es of freedom (d f ) = p (p + 1)/ 2 q, w here, p = numb er of indicator variables (i.e. observed scale variables) and q = number of p arameters to be estimated (i.e. factor loadings, error variances, error covariances, etc.). We also reported the power of the overall test of response shifts, which guard against capitalization on the chance of Type I error, and the power of the tests of the particular response shifts that we found in our s tudy. All power calculations [77 79 ] were based on a 5% level of signific ance. This procedure increased our confidence that we found a correct measurement model, and that response shifts if detected were indeed present. var iance covariance matrix between time 1 and time 2. If a significant difference was found, then the subsequent steps of the analysis were to evaluate if these differences were because of reconceptualization, configural invariance, which is invariance of the factor structure (the a priori pattern of factor loadings imposed on the scales). If configural invariance did not hold, then the observed scores represents different

PAGE 51

51 constructs assessed from one time to the next, and it would not make sense to test furth er differences over time. If there was configural invariance, then the subsequent step was to assess if there was scale recalibration. This was evaluated by testing the equivalence of the factor variances and the factor loadings over time as they represent the scale metric or units of measurement. Finally, change in uniqueness (error variances) was tested to evaluate if changes were because of change in measurement error over time. The acceptability of the fitted SEM solution was evaluated on the basis of o verall goodness of approach. Missing Data Evaluation Useable survey responses were obtained for the baseline survey of study patients. We compared missingness in the patients who responded to the ba seline SF 36 surveys with gender, race, age, history of diabetes, angina, abnormal coronary angiogram, myocardial infarction, stroke, coronary bypass, parkinsons, cancer, peripheral vascular disease, left ventricular hypertrophy, congestive heart failure, arrhythmias, alzheimers, transient ischemic attack, renal insufficiency, h ypercholesteremia, and smoking If any of the above said variables differed based on return of baseline surveys, the groups were compared within the two antihypertensive treatment st rategies to determine whether randomization took place appropriately. Validation is the process of determining that a model is good. In our study, we demonstrate an approach to assess sensitivity analysis, in which model fit was assessed for four imputed d ata sets using MLE. The four imputed data sets were compared to the data set with complete cases (i.e. complete baseline and one year cases) based on presence of response shift and model fit statistics. For the first

PAGE 52

52 imputation approach, we looked at patie nts who scored on the PF and MH scales at both baseline and one year time points, and imputed the remaining domains using markov chain monte carlo (MCMC) approach. For the second imputation approach, we chose to impute PF domain scores with instrumental ac tivities of daily living (IADL) complete scores and impute the remaining seven doma ins at both time points using MCMC approach. For the third approach, we analyzed cases with complete PF scores at both time points and imputed the seven domain scores with c omplete IADL scores. For the fourth approach, we chose to impute all missing cases using MCMC approach. Statistical Analysis for Research Question 2 Variables We distinguished between outcome variables measuring HRQoL, change in systolic BP (SBP) over one year, change in diastolic BP (DBP) over one year, non fatal stroke within one year of enrollment, non fatal MI withi n one year of enrollment, baseline CES D scores, one year CES D scores, and self reported prior history of depression were included as expla natory variables (E). Antihypertensive treatment strategies, baseline comorbid conditions, age, sex, race, depression diagnosis by the physician at one year, living status, and educational level were included as potential violator variables (V). From here on, the SF 36 physical and mental latent constructs are denoted by A and the eight SF 36 domains are denoted by X. All variables were collected at baselin e. However, SBP, DBP, CES D scores and SF 36 scale scores were ex, age, race, living status, and educational level were obtained by the survey. Baseline comorbid conditions were measured through

PAGE 53

53 self report. We described latent constructs (A) for the first research question and thus will not be reiterated here. Hypoth esis#1 H 0 : The relationships between explanatory variables (E) and SF 36 scales (X) will be fully explained by the relationships between explanatory variables and violator variables (V), explanatory variables and latent constructs (A), latent constructs an d violator variables, and latent constructs and SF 36 scales, no measurement bias will be found with respect to the explanatory variables. H a : The E X relationships will not be fully explai ned by the E V, E A, A V and A X relationships, measurement bias will be found with respect to variables E. Hypothesis#2 H 0 : The V X relationships will be fully explained by the E V, E A, A V and A X relationships, no measurement bias will be found with respect to variables V. H a : The V X relationships will not be fully explain ed by the E V, E A, A V and A X relationships, measurement bias will be found with respect to variables V. Hypothesis#3 H 0 : The E X relationships will be consistent across measurement occasions (i.e. over two time points), no response shift will be observed. H a : The E X relationships will not be consistent across measurement occasions (i.e. over two time points), res ponse shift will be observed. Hypothesis#4 H 0 : The V X relationships will be consistent across measurement occasions (i.e. over two time points), no response shift will be observed. H a : The V X relationships will not be consistent across measur ement occasions (i.e. over two time points), response shift will be observed. Explanatory variables (E) Change in systolic blood pressure and diastolic blood pressure over one year : Change in SBP and change in DBP over one year reported for patients assign ed to verapamil SR and atenolol led treatment strategies were taken from the INVEST trial dataset.

PAGE 54

54 Stroke or Myocardial infarction within one year of enrollment : Patients who had s troke or MI within one year of enrollment (i.e. after allocation to treatmen t strategies) were coded as (1) or (0) if they did not have an occurrence of stroke or MI within the same time period. These two variables will be combined to form a single variable with patients who had a stroke or MI will be coded as (1) or (0) if the y d id not have the same within one year of enrollment. Baseline and one year CES D scores : as sessed using the CES D scale [80 ] included in the baseline mail questionnaire. The CES D is a 20 item self reported scale that i s considered a reli able 0.91) and valid instrument. CES D scores range from 0 to 60, with higher scores indicating more depressive symptoms. Scores > 16 are generally consistent with depressive symptoms of clinically depressed patients. People with m ajor chronic medical conditions are most likely to score in the high depressive symptoms range, so a higher threshold of 23 is recommended for studies of older persons with chronic illnesses [11,17]. Self reported prior history of depression : f reported prior history of depression was coded as (1) or (0) if they did not report depression. Potential violator variables (V) Pharmacologic Hypertension Treatment Strategy: I NVEST patients were randomly assigned to either the atenolol led or the verapamil led treatment strategy [8]. Patients assigned to the atenolol led strategy were coded as zero (0) and those assigned to the verapamil SR strategy were coded as one (1). Basel ine comorbid conditions : History of coronary artery bypass graft, angina, arrhythmia, abnormal coronary angiogram, Parkinson disease, peripheral vascular disease, myocardial infarction, stroke, alzheimers, cancer, left ventricular hypertrophy, hypercholest erolemia, congestive heart failure, transient ischemic attack, renal insufficiency, smoking, and diabetes were noted at baseline (coded conditions that are different to the on es present in this study. A modified Charlson Index (which removes the points for the CAD complications of myocardial infarction and heart failure from the original index) (Table 3 2 ) has been applied in patients with CAD [81,82 ]. Each of the indicated dia gnoses is assigned a weight Age : Patients age was dichotomized as (1) for > 75 years old and (0) for those < 74 years old. Gender : Patients who were femal es were coded zero (0) and those who were males were coded one (1). Data regarding gender were obtained during the baseline INVEST visit.

PAGE 55

55 Race : Patients who were Caucasians were coded zero (0) and those who were Non Caucasians were coded one (1). Data rega rding race were obtained during the baseline INVEST visit. medical doctor or psychiatrist has told you that you were depressed during the as (1) if they reported yes or (0) if they Education level : Patients who completed a high school graduation were coded as (1) or (0) if they did not complete a high school graduation. Living status: Patients who were living alone were coded as (0) or (1) if they were living with someone. A two step procedure was used to investigate measurement bias and response shift in the HRQoL data gathered at baseline and at one year period. In each of the two steps, SEM described by Oor t et al. [31,32] was applied. Step 1: Establishing a Measurement Model SEM was used to fit a confirmatory factor model to the 16 x 16 variance covariance matrix of the eight HRQoL scales measured at two occasions. We used the MLE method to fit a model with both measurement occasions and two common factors, PHYS HRQoL and MENT HRQoL, with a pattern of factor loadings that was similar to that developed for the first research question. Overall goodness of fit was evaluated 2 test of exact fit and the RMSEA as a measure of approximate fit. A 2 test indicated good model fit. An RMSEA value of less than 0.08 suggests satisfactory fit, and a value of less than 0.05 suggests close fit [75 ]. In addition to overall goodness of fit, component fit was evaluated through inspection of modification indices and standardized residuals [73].

PAGE 56

56 Step 2: Detecting Measurement Bias and Response Shift in Measurement In the second step, we extended the four factor model to include the following: change in SB P and DBP, stroke a nd MI within one year of enroll ment, baseline CES D scores, one year CES D scores, self reported prior history of depression, depression diagnosis by the physician at one year, antihypertensive treatment strategies, age, sex, race, educa tion level, living status, and baseline comor bid conditions. There were six explanat ory variables and eight potential violator variables that were included in the model as exogenous variables, with residual variances fixed at zero. All violator variables w ere correlated with each other, with the explanatory variables and with the common HRQoL factors. Measurement bias was indicated by significant modification indices for direct effects of the six ex p lanatory variables and the eight violator variables on the eight SF 36 scales and for across occasion constraints on factor loadings and intercepts [31,32]. In all, there were 264 modificat ion indices to consider ([16 x 6 direct effects fixed at zero] + [16 x 8 direct effects fixed at zero] + [8 x 4 across occasi on constraints on factor loadings] + [8 across occasion constraints on intercepts]), however, ei ght modification indices were subtracted because of impossible effects of the second occasion health state on first occa sion SF 36 scales. This left 2 56 modific ation indices for cons trained parameters that were interpreted as measurement bias if unconstrained. As there were a large number of tests, to maintain a family wise Type I error rate of 5%, a Bonfer roni adjusted critical value [78,79 ] of 12.9 (associated with a probability of 0.05/ 2 56 ) was used. The Step 2 model was modified by changing one parameter at a time and constantly checking the estimates and results to ensure that the changes were meaningful and interpretable. This process was continued until the larges t modification

PAGE 57

57 index was less than 12.9 We should note that the modification index often 2 difference). Therefore, we also checked for parameter changes that may be associated with m odification indices less than 12.9 2 difference. Missing Data Evaluation For research question#2, data was analyzed to determine the type of missingness. Useable survey responses were obtained for the baseline survey of study patients. We compare d missingness in the patients who responded to the baseline SF 36 surveys with gender, race, age, history of diabetes, angina, abnormal coronary angiogram, myocardial infarction, stroke, coronary bypass, parkinsons, cancer, peripheral vascular disease, lef t ventricular hypertrophy, congestive heart failure, arrhythmias, alzheimers, transient ischemic attack, renal insufficiency, hypercholesteremia, and smoking. If any of the above said variables differed based on return of baseline surveys, they were compar ed within the two antihypertensive treatment strategies to determine whether randomization took place appropriately. Validation is the process of determining that a model is good. In our study, we demonstrate d an approach to assess sensitivity analysis, in which model fit was assessed for four imputed data sets using MLE. The four imputed data sets were compared to the data set with complete cases (i.e. complete baseline and one year surveys) based on presence of response shift and model fit statistics. For the first imputation approach, we looked at patients who scored on the PF and MH scales at both baseline and one year time points, and imputed the remaining domains and variables using MCMC approach.

PAGE 58

58 For the second imputation approach, we chose to impute PF domain scores with IADL complete scores and impute the remaining seven domains at both time points and other variables using MCMC approach. For the third approach, we chose cases with complete PF scores at both time points and imputed the seven domain s cores with complete IADL scores and remaining variables using MCMC. For the fourth approach, we chose to impute all cas es at both time points using MCMC approach

PAGE 59

59 Table 3 1. SF 36 HRQoL measure Measure Short Form 36 Type General health profile Purpose Evaluative Domains Eight Physical functioning, role limitations due to physical, bodily pain, vitality, social functioning, role limitations due to emotional, mental health, general health Number of items 36 How administered Interview in person or pho ne Scaling 2 points to 6 points depending on item Scoring Domains 0 100 Higher = Better Physical and mental health components 0 100 with mean = 50, SD = 10 Algorithm required Used in response shift studies Yes Table 3 2. Weights assigned to different comorbid states in the modified Charlson Index Points Condition 0 Coronary artery bypass graft, angina myocardial infarction, congestive heart failure 1 Peripheral vascular disease, arrhythmia, abnormal coronary angiogram, alzhemier, gastrointestinal bleed, hypercholesteremia, left ventricular hypertrophy, Parkinson, transient ischemic attack 2 Stroke, cancer, renal insufficiency, diabetes

PAGE 60

60 Table 3 3. Analytical approach Task Measure Software Characterizing study participants Descriptive statist ics, cross sectional at two points in time, missing data excluded SAS 9.14 Characterizing variables Univariate tests for normality of the subscales for HRQoL indices SAS 9.14 Missing data evaluation Chi square comparison of specific outcomes and missin gness to determine missing data pattern SAS 9.14 Model building Based on Medical Outcomes Study Framework of Health Indicators and incorporating modification indices LISREL 8.8 Identification of response shift For research question#1 SEM approach descr ibed by Oort [10] to identify reconceptualization, recalibration, or reprioritization response shift SEM approach described by Schmitt [28] For research question#2 SEM approach described by Oort [31] and King Kallimanis [32] LISREL 8.8 LISREL 8.8 Mode l estimation Variance covariance data analyzed LISREL 8.8 Model evaluation Overall goodness of fit 2 (df, p value) Multiple fit indices (SRMR, RMSEA, CFI NFI, NNFI ) Localized areas of ill fit (modification indices, standardized residuals) LISREL 8.8

PAGE 61

61 CHAPTER 4 RESULTS Question 1) Description of the Population The initial SADD Sx sample consisted of 2,317 INVEST patients assigned to either the verapamil SR based (n = 1184) or atenolol based (n = 1133) treatment strategy. Complete survey respon ses from both the baseline and one year follow up surveys (Figure 4 1) ple of patients was male and the majority were Caucasian (80.7%). Comparison of Treatment Groups at Baseline At baseline, patients were randomized to treatment group. The final sample of patients in the two treatment groups was similar on their sociodemogr aphic characteristics and other variables (Table 4 1) indicating success of the randomization procedures. Measurement Model Below we first explain the measurement model that was used in the response shift identification procedure. The same measurement mode l was used for both the Oort and Schmitt SEM procedures to detect response shift. We present the results of response shift detection and true change evaluation conducted with the Oort procedure and we conclude with an evaluation of the size of response shi fts and true change. Lastly, we present the results of response shift detection with the Schmitt procedure. Table 4 3 gives baseline and one year means scores and standard deviations for all SF 36 scales. The last column of Table 4 3 presents the standardi zed difference

PAGE 62

62 between baseline and one year mean scores (i.e. index) that is the change in the observed scale mean scores without accounting for potential response shifts. Conventional t tests indicate improvement in RP, BP, GH, VT, MH and RE, and no change in PF and SF scale mean scores. Results from exploratory factor analyses gave rise to the measurement model displayed in Figure 4 2. Two latent variables are the latent factors PHYS HRQoL and MENT HRQoL. PHYS HRQoL is measured by PF, RP, BP, GH, VT, and SF, MENT HRQoL is measured by MH, RE, and again SF. Other latent variables are the residual factors ResPF, ResRP, ResBP, etc. The residual factors represent all that is specific to individual scales (i.e. PF, RP, BP, etc) in each latent facto r plus random error variation. The measurement model portrayed in Figure 4 2 resembles the principal components model of the SF 36 scales described by Ware et al. [ 74 ]. The general ith largely the same indicators. Similar to the Oort et al. study, we found that the wording of the SF items combines physical and mental aspects, causing SF to load on both PHYS HRQoL and MENT HRQoL. We allowed the residual factors for RP and RE to co var y (0.22 correlation) since the covariance between RP and RE was not adequately explained by the correlation between the two latent factors. Detection of Response Shift and True Change Fit results for the four mo dels that resulted from carrying out the four step procedure are given in Table 4 4 and Table 4 5.

PAGE 63

63 Step 1: Measurement Model The measurement model of Figure 4 2 was the basis of Model 1, a SEM for measurements at two occasions, but without any across occas io 2 2 (84) = 301.284), but the RMSEA measure indicated reasonable fit (RMSEA = 0.054 (90% CI = 0.048, 0.061), Table 4 4). The CFI was 0.991; a value of >.90 suggest a reasonably good fit. Other fit indi ces, such as the N FI, NNFI and SRMR had values 0.99, 0.99, and 0.03, respectively; all indicating good fit (i.e. the data fit the proposed model well). Step 2: No Response Shift Model In Model 2, all response shift parameters were held invariant across o ccasions. This means that all across occasion invariance constraints on parameter estimates (factor loadings, intercepts, and residual variances) were imposed. Two models were compared, the model in which parameter estimates were all freely estimated (i.e. Model 1), and the fully constrained model (i.e. Model 2) in which all of the parameter 2 difference test evaluates the change in overall fit of the free and fully constrained (no response shift) models. The fit of Model 2, although still satisfactory (RMSEA = 0.051, Table 4 4), was significantly 2 difference 2 (25) = 64.128, p < 0.0001). Therefore, we reject the null hypoth esis that the two models fit equally well and conclude that there may be evidence of response shift present in the data. Step 3: Identification of Response Shift Inspection of modification indices and standardized residuals indicated which of the equality constraints were not tenable. Step by step adjustment of Model 2 yielded

PAGE 64

64 Model 3, which showed three cases of response shift. The fit of Model 3 was good (RMSEA = 0.049 (90% CI = 0.043 0.055)), and significantly better than the fit of Model 2 (3) = 34.950), Table 4 4. The next step was to address the possibility of making a Type I error with multiple we identified two cases of response shift. The fit of the new Model 3 was good (RMSEA = 0.049 (90% CI = 0.043 2 (2) = 30.892), Table 4 5. Step 4: Final Model To investigate change in the means, variances, and correlations of the common f actors, we fitted additional models with the new Model 3 as the starting point. The across occasion invariance of parameters was tested step by step, maintaining all equality constraints that proved tenable. This procedure finally yielded Model 4, which fi 2 (111) = 340.681, RMSEA = 0.048 (90% CI = 0.042 0.054)), Table 4 5. Evaluation of Response Shifts and True Change Recalibration Response Shift abou t uniform and non uniform recalibration, respectively. For PF, we found differences between first and second occasion intercepts, indicating uniform recalibration for the PF scale. According to Oort, uniform recalibration is reflected by change in the mean s of the observed variables that cannot be attributed to change in the common factor means. We also found a change in the variance of the residual factor PF, indicating non uniform recalibration is

PAGE 65

65 reflected by the change in the variances of the observed variables that cannot be attributed to change in the common factor variances. Contributions of Response Shifts and True Change to Change in the Observed Variables In addition to significance test results, Table 4 7 provides effect sizes for observed change, and the response shift and true change contributions to observed change, as implied by the parameter estimates of Model 4 (in Table 4 5 and Table 4 6). From Table 4 7 it appears that the respon se shift effects on ob served change were only small: 0.118 for the uniform recalibration for PF scale and zero for the non uniform recalibration for PF scale. The effects of true change were smaller for PF. For PF the effects of response shifts and true c hange were in opposite directions. Impact of Response Shifts on the Measurement of True Change We found that most Model 4 parameters were invariant across occasions, except for the latent means that did change over time (Table 4 6). Common factor variances and common factor correlations did not change across occasions, but the common factor means did. Common factor means were fixed at zero for the first occasion (because of identification requirements) so that the second occasion estimates were a d irect rep resentation of change. The Schmitt approach started with the same baseline longitudinal measurement model as the Oort method. As the longitudinal model fit well, there was no reconceptualization, configural v ariance over time, which is invariance of factor structure (the a priori pattern of factor loadings imposed on the scales). Since configural invariance was absent, the subsequent step was to assess if there was scale

PAGE 66

66 recalibration. This was evaluated by te sting the equivalence of the factor variances and the factor loadings (metric invariance) over time as they represent the scale metric or units of measurement. When constraints were applied and added one at a time, variances and factor loadings (recalibrat ion), we found one significant change between subsequent models (Table 4 8). The fit of Model 4, was satisfactory (RMSEA = 0.050 (90% CI, 0.044 0.057), Table 4 8). However, we found that freeing the factor loading for PF significantly improved the fit of 2 2 (1) = 3.906, p = 0.048). We also found an improvement to the model when PF error variances were 2 2 (1) = 8.446 p = 0.0 04 ), which according to the definition, a significant decrease in the variance accounted for by the models signals a difference with respect to the reliability of measurement before and after assignment to treatment strategies. Missing D ata Evaluation and Imputation Useable survey responses were obtained for the baseline survey from 68.1% (n = 1578) of study patients. Gender ( 2 2 = 69.19, p<0.0001), 2 = 10.10, p = .001), history of 2 = 11.04, p = 0.001) 2 = 7.81, 2 = 4.50, p = 0.03) history of left vent 2 = 5.64, p = 0.02), history of 2 = 5.65, p = 0.02) were the least likely to respond to the baseline survey. Within gender, race, history of angina, history of abnormal coronary angiogram, history of coronary by pass, history of cancer, history of left ventricular hypertrophy, and history of smoking, baseline return rates were similar for the two antihypertensive treatment strategies. Figure 4 1 details reasons for not completing the baseline survey.

PAGE 67

67 In this study we chose missing data imputation using sensitivity analysis with four different approaches; model fit and response shift was assessed across the four approaches had a reason able fit and showed pre s ence of response shift (Table 17 Table 24 ). Oort SEM procedure. When using the first imputation approach (Table 17 ) we found presence of uniform recalibration for BP in addition to the presence of uniform recalibration and non unif orm recalibration for PF. It seems that for some patients (using the first imputation approach) the meaning of the response scale anchors for the BP scale changed since assignment to treatment strategy. Similar to our first imputation approach, we found pr esence of uniform recalibration for BP using the second imputation approach (Table 18 ) However, we did not identify non uniform recalibration for PF compared to our original results. When using the third imputation approach (Table 19 ) w e found reprioriti zation and non uniform recalibration for VT in addition to the presence of uniform and non uniform recalibration for PF. On the other hand, when using the fourth imputation approach, we found several instance s of response shift as shown in Table 20 Schmit When using the first imputation approach (Table 21 ) we found random error for BP and VT in addition to the presence of recalibration for PF and random error for PF. Comparatively, when using the seco nd imputation approach (Table 22 ); we found that the variances of the physical function construct changed over time indicating presence of recalibration We also found evidence of random error for PF, BP, and VT. Its seems that using imputation approach 2, some respondents are

PAGE 68

68 perceiving more or less difference in the relevant constructs due to the disease condition or treatment strategy assigned. When using the third imputation approach (Table 23 ) we found random error for VT in addition to the response shift obtained for our original analys es. On the other hand, when using the fourth imputation approach (Table 24 ) we found several instance s of response shift. It seems that each imputation approach led to slightly different results compared to our original analyses. Since there is no best si ngle imputation technique available, these results should be interpreted with caution. This deserves further investigation. Measurement Bias and Response Shift (Research Question 2) Description of the population The initial SADD Sx sample consisted of 2,31 7 INVEST patients assigned to either the verapamil SR based (n = 1184) or atenolol based (n = 1133) treatment strategy. Complete survey respon ses from both the baseline and one year follow up surveys we re obtained from 34.01% (n = 788 al sample of patients (Figure 4 3) Nearly 57% (n = 451 he majority were Caucasian (82.9 %). Comparison of Treatment Groups at Baseline At baseline, patients were randomized to treatment group. The fina l sample of patients in the two treatment groups was similar on their sociodemographic characteristics and other variables (Table 4 10 ) indicating success of the randomization procedures. Below we first explain the measurement model that was used in the re sponse shift identification procedure. We present the results of response shift detection and true

PAGE 69

69 change evaluation conducted with the Oort procedure and we conclude with an evaluation of the size of response shifts and true change. Step 1: Measurement Mo del The measurement model for this research question was analogous to the first research question, F igure 4 2. Briefly, PHYS HRQoL wa s measured by PF, RP, BP, GH, VT, and SF, MENT HRQoL wa s measured by MH, RE, an d again SF. Table 4 11 gives baseline and on e year means scores and standard deviations for all SF 36 scale s. The last column of Table 4 11 presents the standardized difference between baseline and one year mean scores (i.e. index) that is the change in the observed scale mean scores with out accounting for potential response shifts. Conventional t tests indicate improvement in RP, BP, GH, VT, MH and RE, and no change in PF and SF scale mean scores. In Model 1, factor loadings and intercepts were not constrained to be equal across occasions 2 test 2 (84) = 287. 554 ), but the RMSEA measure indicat ed reasonable fit (RMSEA = 0.055 (90% CI = 0.048, 0.062 ), Table 4 12 ). We concluded that this mo del was satisfactory and applied this model for the investigation of measurement bias and response shift in measurement. Step 2: Measurement Bias and Response Shift In the second step, all factor loadings and intercepts were constrained to be equal across occasions. Explanatory variables were included occurrence of non fatal s troke or non fatal MI, baseline CES D scores one year CES D scores, change in SBP, change in DBP, and prior history of depression with direct effects on the latent factors, PHYS HRQ oL and MENT HRQoL. Potential violators of measurement invariance were

PAGE 70

70 included as exogenous variables age, gender, race, treatment strategy, education, living alone, depression diagnosis by physician at one year, and modified comorbidity score The exogenous variables were allowed to correlate with the explanatory variables and with the common HRQoL factors, but not directly affect the observed HRQoL subscale scores. Two models were compared, the mode l in which parameter estimates we re all free 2 test of exact fit 2 (273) = 744.881 ) and the RMSEA measure indicated reasonable fit ( RMSEA = 0.045 (90% CI = 0.041, 0.049 ), Table 4 12 ). The fit of Model 2, although s till satisfactory, was significantly worse than the fit of Model 1, indicating the presence of 2 2 ( 189) = 457.327 p < 0.0001). Modification indices revea led that the fit of the model could be further improved by accoun ting for instances of measurement bias. We found one intercept that was not equal across measurement occasions, which indicated recalibration response shift. For PF, we found differences between first and second occasion intercepts, indicating uniform reca libration for the PF scale. The relationship between the potential violator variables and the observed variables should be explained via their relationships with the common HRQoL factors. In addition, the relationships between the explanatory variables and the observed variables should be explained via their relationships with the common HRQoL factors. When this does not occur, measurement bias has been found. The results indicated four insta nces of measurement bias which could be considered as response shi ft in measurement. For the model that i dentified four instances of measurement bias, age

PAGE 71

71 was used as a continuous variable. However, when age was dichot omized ( < 74 years and > 75 years), we identified only two instances of measurement bias/response shift. Below we elaborate and provide explanations on both these approaches. Subsequently, we demonstrate which of these two approaches provide a better explanation of the presence of measurement bias and/or response shift in our study population. For the first approach when age was used as a continuous variable, we found the relationship between age and PF was not fully determined by their relationship with the PHYS HRQoL latent factor. This indicated that PF was not just indicative of PHYS HRQoL but also of age Therefore, a direct relationship between age and PF was included. The violation of measurement invariance was not consistent acros s occasions (estimated at 0.177 at baseline and 0.199 at one year) which indicated that older patients reported worse PF t han younger patients, even if their PHYS HRQoL was similar. We also found that the relationship between age and RP was not fully determined by their relationships with the PHYS HRQoL factor. This indicated that RP was not just indicative of PHYS HRQoL but also of age. Therefore, a direct relationship between age and RP was included. The violation of measurement invariance was not consistent across occasions (estimated at 0 .140 at baseline and 0.119 at one year) and indicated that older patients reported w orse RP than younger patients, even if their PHYS HRQoL was similar. Lastly, a direct relationship between gender and PF was included. The violation of measurement invariance was not consistent across occasions (estimated at 0.100 at baseline and 0.131 at one year ) which indicated that male patients reported better PF than female patients, even if their PHYS HRQoL was similar. After accounting for the three response shifts stated above, the modified model

PAGE 72

72 showed i m provement and close fit ( RMSEA = 0.039, see Model 2, Table 4 12 ). The largest modification indices were wel l below the critical value of 12.9 Contrary to our first approach, when age was dichotomized, we did not find any relationship between age and PF or age and RP. On the other hand, similar to the first approach, we found a relationship between gender and PF that was not fully determined by their relationships with the common PHYS HRQoL factor. This indicated that PF was not just indicative of PHYS HRQoL but also of gender. Therefore, a direct r elationship between gender and PF was included. The violation of measurement invariance was not consistent across occasions (estimated at 0.113 at baseline and 0.150 at one year, Table 4 14) which indicated that male patients reported better PF than female patients, even if their PHYS HRQoL was similar. After accounting for response shifts, the 2 = 638.562, RMSEA = 0.041 (90% CI = 0.036 0.045) Table 4 13 ). The largest modification indices were wel l below the critical value of 12.9 Based on the findings in our study, the second approach resulted in another model, albeit more parsimonious. Similar to the approach adopted by King Kallimanis [ 32 ], here, we chose the model (second approach when age was dichoto mized) with measurement bias and response shift in the smallest number of scales, providing a more parsimonious model. True Change Common factor means were fixed at zero for the first occasion (because of identification requirements) so that the second occ asion estimates were a direct representation of change. In model 2, the estimates of the common factor means indicated that after assignment to treatment strategies patients reported worsened PHYS HRQoL and improved MENT HRQoL

PAGE 73

73 Contributions of Response Sh ifts and True Change to Change in the Observed Variables In addition to signi ficance test results, Table 4 16 provides effect sizes for observed change, and the response shift and true change contributions to observed change, as implied by the parameter es timate s of Model 2 shown in Table 4 13. From Table 4 16 it appears that the response sh ift effects on observed change we re only small: 0.097 for the uniform recalibration for PF sca le. The effects of true change we re smaller for PF. For PF the effects of r esponse shifts and true change we re in opposite directions. Missing Data Evaluation and Imputation Useable survey responses were obtained for the baseline survey from 68.1% (n = 157 2 = 7.06, p = 2 = 69.19, p<0.0001), 2 = 10.10, p=.001), history of abno 2 = 11.04, p = 0.001) 2 = 7.81, p = 0.005), history of cancer ( 2 = 4.50, p = 0.03) history of 2 = 5.64, p = 0.02), history of 2 = 5.65, p = 0.02) were the least likely to respond to the baseline survey. Within gender, race, history of angina, history of abnormal coronary angio gram, history of coronary bypass, history of cancer, history of left ventricular hypertrophy, and history of smoking, baseline return rates were similar for the two antihypertensive treatment strategies. Figure 4 3 details reasons for not completing the ba seline survey. In this study, we chose missing data imputation using sensitivity analysis with four different approaches; model fit and response shift was assessed across the four imputed datasets using MLE. All longitudinal models using Oort SEM approach had a reasonable fit and showed presence of resp onse shift ( Table 25 Table 28 ).

PAGE 74

74 When using the first imputation approach (Table 25 ) we found presence of uniform recalibration for BP in addition to the presence of uniform recalibration and non uniform re calibration for PF. It seems that for some patients (using the first imputation approach) the meaning of the response scale anchors for the BP scale changed since assignment to treatment strategy. We also found that high school education had a direct effec t on BP at the second occasion. The effect of high school education on BP at the second measureme nt occasion was negative suggesting that at one year patients with high school education reported worse BP than those with no high school education even if th eir true HRQoL was similar. Similar to the first imputation approach, we found high school education to have a direct effect on BP using imputation approach 2 (Table 26 ) However, we found n o effect of gender on PF, but, found age to have a direct effect o n PF at the second occasion. Imputation approach 3 (Table 27 ) gave us similar results to those using imputation approach 2. Using imputation approach 4 (Table 28 ), we found uniform recalibration for PF and BP scales, gender to have a direct effect on PF, a nd education to have a direct effect on baseline PF among several others

PAGE 75

75 Figure 4 1. Sx substudy of the International Verapamil SR Trandolapril Study (n = 2317). Abbreviation: SADD Sx Study of Antihypertensive Drugs and Depressive Symptoms

PAGE 76

76 Figure 4 2. The measurement model used in response shift detection Notes: Circles represent latent variables (common and residual factors) and squares represent observed variables (SF 36 scales). Abbreviations: PF physical functioning; RP role physical; BP bodily pain; GH general health; VT vitality; SF social functioning; RE role emotional; MH mental health.

PAGE 77

77 Figure 4 3. S x substudy of the International Verapamil SR Trandolapril Study (n = 2317). SADD Sx = Study of Antihypertensive Drugs and Depressive Symptoms.

PAGE 78

78 Table 4 1. Comparison of patients at baseline assigned to the verapamil SR strategy and the atenolol strategy among the SADD Sx patient s returning both baseline and one year surveys (N = 909) Variable Atenolol Strategy (N = 429) Verapamil SR strategy (N = 465) p value Demographic characteristics Male 61.3% 55.7% 0.09 Caucasian 82.5% 82.8% 0.91 Patient aver age age at entry years (SD) 67.4 (9.0) 67.5 (9.4) 0.91 a Medical History at Baseline Coronary Artery Bypass Graft (CABG) 30.3% 27.7% 0.23 Congestive Heart Failure (CHF) 7.7% 5.6% 0.21 Angina 45.7% 45.0% 0.82 Peripheral vascular disease 13.5% 14.8% 0.57 Arrhythmia 7.5% 6.9% 0.74 Abnormal coronary angiogram 65.5% 65.2% 0.92 Myocardial Infarction (MI) 42.7% 46.5% 0.25 Stroke 5.1% 6.5% 0.40 Alzheimer b 0.5% 0.4% 0.38 Cancer 5.6% 4.3% 0.37 Hypercholesteremia 68.1% 65.6% 0.43 Left Ventricular Hype rtrophy 19.1% 17.4% 0.51 Diabetes 29.6% 25.8% 0.21 Parkinson b 0.0% 0.2% 0.52 Transient Ischemic Attack 4.9% 4.3% 0.67 Renal Insufficiency 3.0% 3.7% 0.60 Smoking 55.7% 54.8% 0.79 SADD Sx = Study of Antihypertensive Drugs and Depressive Sy mptoms; SD = Standard Deviation a Independent t test T able 4 2. Testing for normality in HRQoL indices Variable Normality Subscale scores SF 36 physical health, role physical, role emotional, social function, mental health, pain, vitalit y, general health Moderately non normal to normally distributed Moderately non normal = skewness <2, kurtosis<7, severely non normal = skewness>2, kurtosis>7

PAGE 79

79 Table 4 3. Means, standard deviations, and effect sizes for SF 36 scales at baseline and one year (n = 909) Scale Before assignment to treatment strategy After assignment to treatment strategy Post Pre d Index # Mean Standard deviation Mean Standard deviation PF 55.142 29.013 54.526 29.030 0.021 RP 47.992 40.866 51.375 41.201 0.082* BP 62. 162 29.428 66.425 30.967 0.138** GH 52.031 21.851 54.334 21.994 0.105** VT 46.368 22.473 48.398 21.861 0.090** SF 70.421 23.826 71.741 23.101 0.055 RE 62.871 41.071 66.593 40.228 0.091** MH 72.095 18.772 73.627 18.762 0.082** N = 909; # Standardized differences; *p < 0.05, **p < 0.01 in paired t test. Abbreviations: PF physical functioning; RP role physical; BP bodily pain; GH general health; VT vitality; SF social functioning; RE role emotional; MH mental health.

PAGE 80

80 Table 4 4. 2 differences between models without controlling for Type I error (n = 909) Model Df CHISQ RMSEA SRMR CFI NFI NNFI Model 1 84 301.284 0.054 (0.048 0.0 61) 0.034 0.991 0.988 0.987 Model 2 109 365.412 0.051 (0.045 0.057) 0.038 0.989 0.985 0.989 Model 3 Uniform recalibration (PF) 108 341.168 0.049 (0.043 0.055) 0.037 0.991 0.986 0.989 Uniform recalibration (BP) 107 337.118 0.049 (0.043 0.055) 0.037 0.991 0.986 0.989 Non uniform recalibration (PF) 106 330.462 0.049 (0.043 0.055) 0.037 0.991 0.987 0.989 Model 4 Final Model (all tenable constraints imposed) 110 336.630 0.048 (0.042 0.054) 0.040 0.991 0.986 0.989 Abbreviations: RMSEA root mean s quare error of approximation, SRMR standardized root mean square residual, CFI comparative fit index, NFI normed fit index, NNFI non normed fit index. PF physical functioning, BP bodily pain. RMSEA <= 0.05 close fit; 0.05 to 0.08 reasonable fit ; >= 0.10 poor fit CFI > 0.90 reasonably good fit, SRMR < 0.10 is favorable

PAGE 81

81 Table 4 5. Oort method 2 differences between models when controlling for Type I error (n = 909) Model Df CHISQ RMSEA SRMR CFI NFI NNFI Model 1 84 301.284 0.054 (0.048 0.061) 0.034 0.991 0.988 0.987 Model 2 109 365.412 0.051 (0.045 0.057) 0.038 0.98 9 0.985 0.989 Model 3 Uniform recalibration (PF) 108 341.168 0.049 (0.043 0.055) 0.037 0.991 0.986 0.989 Non uniform recalibration (PF) 107 334.520 0.049 (0.043 0.055) 0.038 0.991 0.986 0.989 Model 4 Final Model (all tenable constraints imposed) 111 3 40.681 0.048 (0.042 0.054) 0.040 0.991 0.986 0.989 Abbreviations: RMSEA root mean square error of approximation, SRMR standardized root mean square residual, CFI comparative fit index, NFI normed fit index, NNFI non normed fit index. PF phys ic al functioning RMSEA <= 0.05 close fit; 0.05 to 0.08 reasonable fit; >= 0.10 poor fit CFI > 0.90 reasonably good fit, SRMR < 0.10 is favorable

PAGE 82

8 2 Table 4 6. Parameter estimates in the final model (Model 4, Table 4 5) (n = 909) Before assignment to treat ment strategy After assignment to treatment strategy PHYS HRQo L1 MENT HRQo L1 PHYS HRQo L2 MENT HRQo L2 Factor loadings PF1 0.742 PF2 0.769 RP1 0.748 RP2 0.748 BP1 0.661 BP2 0.661 GH1 0.775 GH2 0.775 VT1 0.761 VT2 0.761 SF1 0.348 0.450 SF2 0.348 0.450 RE1 0.722 RE2 0.722 MH1 0.766 MH2 0.766 55.142 51.709 47.660 47.660 62.973 62.973 52.081 52.081 46.276 46.276 69.991 69.991 63.187 63.187 72.115 72.115 Residual variances 0.449 0.409 0.440 0.440 0.563 0.563 0.399 0.399 0.421 0.42 1 0.428 0.428 0.478 0.478 0.414 0.414 PHYS HRQo L1 MENT HRQo L1 PHYS HRQo L2 MENT HRQo L2 Common factor correlations PHYS HRQo L1 1 MENT HRQo L1 0.795 1 PHYS HRQo L2 0.795 0.643 1 PHYS HRQo L2 0.643 0.795 0.795 1 Common factor means 0.000 0.000 0.131 0.104 G 2 (111) = 340.681, RMSEA = 0.048 (0.042 0.054), SRMR = 0.040, CFI = 0.991, NFI = 0.986, NNFI = 0.989. Abbreviations: PF physical functioning; RP role physica l; BP bodily pain; GH general health; VT vitality; SF social functioning; RE role emotional; MH mental health.

PAGE 83

83 Table 4 7. Significance tests of response shifts, and effect sizes of observed change, response shift, and true change in the final model (From tables 4 5 and 4 6) Scale Response shift Significance test Effect sizes* 2 (df =1) p value # Observed change Response shift contribution True change contribution PF Uniform recalibration 24.244 p < 0.0001 0.021 0.118 0.097 Non uniform recalibration 6.648 p = 0.01 0.000** RP 0.082 0.082 BP 0.138 0.138 GH 0.1 05 0.105 VT 0.090 0.090 SF 0.055 0.055 RE 0.091 0.091 MH 0.082 0.082 *effect size has been described as small (0.2), medium (0.5), and large (0.8). **non uniform recalibration was identified, but non u niform recalibration is not estimated in the effect size calculation Abbreviations: PF physical functioning; RP role physical; BP bodily pain; GH general health; VT vitality; SF social functioning; RE role emotional; MH mental health.

PAGE 84

84 Tabl e 4 8. 2 differences between models (n = 909) Model Df CHISQ RMSEA SRMR CFI NFI NNFI Model 1 84 301.284 0.054 (0.048 0.061) 0.034 0.991 0.988 0.987 Model 2 Covariances constrained (reconceptualization) 87 305.641 0.053 ( 0.047 0.059) 0.038 0.991 0.988 0.988 Model 3 Factor covariances and variances constrained (recalibration) 91 310.750 0.052 (0.046 0.058) 0.040 0.991 0.987 0.988 Model 4 Factor covariances, variances, and loadings constrained (recalibration) 98 321.021 0 .050 (0.044 0.057) 0.041 0.991 0.987 0.989 PF 97 317.115 0.050 (0.044 0.057) 0.041 0.991 0.987 0.989 Model 5 Factor covariances, variances, loadings, and uniqueness constrained 106 338.198 0.049 (0.043 0.055) 0.039 0.991 0.986 0.989 PF 105 329.7 52 0.049 (0.043 0.055) 0.039 0.991 0.987 0.989 Abbreviations: RMSEA root mean square error of approximation, SRMR standardized root mean square residual, CFI comparative fit index, NFI normed fit index, NNFI non normed fit index. RMSEA <= 0.05 close fit; 0.05 to 0.08 reasonable fit; >= 0.10 poor fit CFI NFI, and NNFI > 0.90 reasonably good fit, SRMR < 0.10 is favorable

PAGE 85

85 Table 4 Recalibration Two t ypes uniform and non uniform recalibration Uniform recalibration is indicated by change in intercepts, non uniform recalibration is demonstrated by change in residual variances e.g., Perhaps new experiences after the first occasion have changed their ide a of how much health should limit physical activity. Indicated by change in factor loadings and factor variances are equal across measurements e.g., an increase or decrease in the factor variances would indicate that the respondents perceive more or less d ifference in the relevant constructs as a result of the intervention. Reprioritization Shown by change in factor loadings. Variable has become more indicative or less indicative of the concept involved Does not identify reprioritization Reconceptualizati on Shown by change in factor patterns over time. If a domain does not belong to the set of variables defining the concept of a latent construct at the first occasion, whereas at the second occasion it does, reconceptualization has occurred. Factor patterns and factor covariances e.g., no reconceptualization means that the relationship among factors with similar intercorrelations should be the same before and after intervention.

PAGE 86

86 Table 4 10 Comparison of patients at baseline assigned to the Verapamil SR s trategy and Atenolol led treatment strategy among SADD Sx patients returning both baseli ne and one year surveys (n = 788 ) Variable Atenolol strategy Verapamil SR strategy P value Stroke or MI within one year of enrollment 1.6 % 1.9 % 0.706 Baseline CES D s core b,c 13.837 13.338 0.523 Baseline SBP b,c 147 988 146 608 0. 282 Baseline DBP b,c 8 2.765 82. 450 0.6 81 Self reported prior history of depression b 18.0 % 17.5 % 0.849 Age ( > 75 years) a 21.8 % 20.4 % 0.652 Male a 59.4 % 55.2 % 0.236 Non Caucasian a 17.5 % 16.8 % 0.789 High school graduate b 69.5 % 71.5 % 0.531 Living status alone or someone b 75.9 % 75.9 % 0.987 comorbidity index c 2.584 2.628 0.719 SD = standard deviation aInformation regarding these variables was obtained from INVEST data. bInfo rmation regarding these variables was obtained from the SADD Sx baseline mail survey from those patients who returned both the baseline and one year follow up surveys. cIndependent t test.

PAGE 87

87 Table 4 11 Means, standard deviations, and effect sizes for SF 36 scales (n = 788 ) Scale Before assignment to treatment strategy After assignment to treatment strategy Post Pre d Index # Mean Standard deviation Mean Standard deviation PF 55.557 29.260 55.065 29.284 0.017 RP 49.228 40.780 51.840 41.250 0.064 BP 63.047 29.648 66.284 3 1.195 0.109 ** GH 52.834 22.041 55.051 22.143 0.101 ** VT 46.937 22.736 48.617 22.052 0.074* SF 70.963 24.063 72.303 23.071 0.056 RE 64.150 40.786 67.936 39.895 0.093 ** MH 72.475 18.744 74.132 18.920 0.088 ** N = 788 ; # Standardiz differences; *p < 0.05, **p < 0.01 in paired t test. Abbreviations: PF physical functioning; RP role physical; BP bodily pain; GH general health; VT vitality; SF soci al functioning; RE role emotional; MH mental health.

PAGE 88

88 Table 4 12 Goodness of fit of models in measurement bias and response shift detection procedure (when age was used as a continuous variable) (n = 788 ) Step Model Df CHISQ RMSEA (90% CI) SRMR CFI NFI NNFI Step 1: Model 1: Measured Model 84 287.554 0.055 (0.048 0.062) 0.035 0.991 0.987 0.987 Step 2: Detect measurement bias Model 2: First model, without accounting for measurement bias 273 744.881 0.045 (0.041 0.049) 0.038 0.986 0. 978 0.978 Uniform recalibration for PF 272 729.136 0.045 (0.040 0.049) 0.037 0.987 0.979 0.979 Age TPF 271 703.021 0.043 (0.047 0.048) 0.037 0.987 0.980 0.980 Age BPF 270 669.555 0.042 (0.038 0.046) 0.035 0.988 0.981 0.981 Age TRP 26 9 658.643 0.042 (0.037 0.046) 0.035 0.989 0.981 0.981 Age BRP 268 631.316 0.040 (0.036 0.045) 0.033 0.989 0.982 0.983 Gender TPF 267 614.376 0.040 (0.036 0.044) 0.033 0.990 0.982 0.983 Gender BPF (Final model) 266 600.530 0.039 (0.035 0.043) 0.032 0.990 0.983 0.984 Abbreviations: PF, physical functioning; RP, role limitations because of physical health; BPF, baseline physical functioning; TPF, one year physical functioning; BRP, baseline role limitations because of physical health; TRP, one year role limitations because of physical health. RMSEA root mean square error of approximation, SRMR standardized root mean square residual, CFI comparative fit index, NFI normed fit index, NNFI non normed fit index. RMSEA <= 0.05 close fit; 0.05 to 0.08 reasonable fit; >= 0.10 poor fit CFI > 0.90 reasonably good fit, SRMR < 0.10 is favorable

PAGE 89

89 Table 4 13 Goodness of fit of models in measurement bias and response shift detection procedure (when age was used as a dichotomous variable) (n = 788 ) Step Model Df CHISQ RMSEA (90% CI) SRM R CFI NFI NNFI Step 1: Model 1: Measured Model 84 287.554 0.055 (0.048 0.062 ) 0.035 0.99 1 0.98 7 0.98 7 Step 2 : Detect measure ment bias Model 2: First model, without accounting for measurement bias 27 3 692.494 0.043 (0.039 0.047) 0.036 0.98 8 0.98 0 0.98 0 Uniform recalibration for PF 27 2 676.847 0.042 (0.038 0.046) 0.035 0.98 8 0.98 0 0.98 1 Gender TPF 27 1 657.167 0.042 (0.037 0.046) 0.035 0.98 9 0.98 1 0.98 2 Gender BPF (Final model) 27 0 638.562 0.041 (0.036 0.045) 0.034 0.98 9 0.98 1 0.98 2 Abbreviations: PF, physical functioning; RP, role limitations because of physical health; BPF, baseline physical functioning; TPF, one year physical functioning RMSEA root mean square error of approximation, SRMR standardized root mean square residual, CFI comparative fit index, NFI normed fit index, NNFI non normed fit index. RMSEA <= 0.05 close fit; 0.05 to 0.08 reasonable fit; >= 0.10 poor fit CFI > 0.90 reasonably good fit, SRMR < 0 .10 is favorable.

PAGE 90

90 Table 4 14 Parameter estimates in the final mode l (Model 2, Table 4 13 ) (n = 788 ) Before assignment to treatment strategy After assignment to treatment strategy PHYS HRQoL1 MENT HRQoL1 PHYS HRQoL2 MENT HRQoL2 Factor loadi ngs PF1 0.73 4 PF2 0.752 RP1 0.727 RP2 0.742 BP1 0.664 BP2 0.646 GH1 0.781 GH2 0.787 VT1 0.749 VT2 0.778 SF1 0.404 0.372 SF2 0.424 0.383 RE1 0.668 RE2 0.681 MH1 0.838 MH2 0.823 a 55.557 52.711 48.895 48.895 63.520 63.520 53.004 53.004 46.819 46.819 70.561 70.561 64.542 64.542 72.457 72.457 Common factor variances PHYS HRQoL1 MENT HRQoL1 PHYS HRQoL2 MENT HRQoL2 1.000 1.000 1.000 1.000 Common f actor correlations PHYS HRQoL1 1 MENT HRQoL1 0.762 1 PHYS HRQoL2 0.797 0.609 1 MENT HRQoL2 0.563 0.738 0.750 1 Abbreviations: PF, physical functioning; RP, role limitations because of physical health; BP, bodily pain; SF, social functioning; MH, mental health; RE, role limitations because of emotional problems; VT, vitality; GH, general health. Notes: (a) across time differences in bold between factor loadings indicate (recalibration) response shift.

PAGE 91

91 Table 4 15 Direct effects i ndicating measurement bias and response shift in measurement (based on Table 4 13 ) Measurement bias Baseline One year Gender PF 0.113 0.150 Abbreviations: PF, physical functioning Notes: across time differences in measurement bias indicate response sh ift in measurement. Table 4 16 Significanc e tests of response shifts, effect sizes of observed change, response shift, and true chang e in the final model (Table 4 13 and Table 4 14 ) Scale Response shift Significance test Effect sizes* 2 (df =1) p value # Observed change Response shift contribution True change contribut ion PF Uniform recalibration 15.647 p < 0.0001 0.017 0.097 0.080 RP 0.064 0.064 BP 0.109 0.109 GH 0.101 0.101 VT 0.074 0.074 SF 0.056 0.056 RE 0.093 0 .093 MH 0.088 0.088 *effect size has been described as small (0.2), medium (0.5), and large (0.8). Abbreviations: PF physical functioning; RP role physical; BP bodily pain; GH general health; VT vitality; SF social functioning; RE role emotional; MH mental health.

PAGE 92

92 Table 4 1 7 Oort method: Imputation using approach 1 (PF and MH present and other domain scores imputed using MCMC) (n=988) Model Df CHISQ RMSEA SRMR CFI NFI NNFI Model 1 84 322.618 0.054 (0.048 0.060 ) 0.034 0.99 1 0.988 0.987 Model 2 109 406.214 0.053 (0.047 0.058) 0.038 0.989 0.985 0.988 Model 3 Uniform recalibration (PF) 108 373.719 0.050 (0.045 0.056) 0.038 0.990 0.986 0.989 Uniform recalibration (BP) 107 366.927 0.050 (0.044 0.055) 0.037 0.990 0.986 0. 989 Non uniform recalibration (PF) 106 358.522 0.050 (0.044 0.055) 0.038 0.990 0.986 0.989 Model 4 Final Model (all tenable constraints imposed) 110 366.436 0.049 (0.044 0.054) 0.040 0.990 0.986 0.989 Abbreviations: RMSEA root mean square error of approximation, SRMR standardized root mean square residual, CFI comparative fit index, NFI normed fit index, NNFI non normed fit index. PF physical functioning, BP bodily pain. RMSEA <= 0.05 close fit; 0.05 to 0.08 reasonable fit; >= 0.10 poor fit CFI > 0.90 reasonably good fit, SRMR < 0.10 is favorable

PAGE 93

93 Table 4 18 Oort method: Imputation using approach 2 (PF imputed with IADL scores and remaining domain scores imputed using MCMC) (n=1057) Model Df CHISQ RMSEA SRMR CFI NFI NNFI Model 1 8 4 354.338 0.056 (0.050 0.062 ) 0.034 0.990 0.987 0.986 Model 2 109 421.830 0.052 (0.047 0.058) 0.038 0.989 0.985 0.988 Model 3 Uniform recalibration (PF) 108 397.402 0.051 (0.045 0.056) 0.037 0.990 0.986 0.988 Uniform recalibration (BP) 107 391.856 0.051 (0.045 0.056) 0.037 0.990 0.986 0.989 Model 4 Final Model (all tenable constraints imposed) 111 399.377 0.050 (0.045 0.055) 0.040 0.990 0.986 0.989 Abbreviations: RMSEA root mean square error of approximation, SRMR standardized root mean sq uare residual, CFI comparative fit index, NFI normed fit index, NNFI non normed fit index. PF physical functioning, BP bodily pain. RMSEA <= 0.05 close fit; 0.05 to 0.08 reasonable fit; >= 0.10 poor fit CFI > 0.90 reasonably good fit, SRMR < 0.10 is favorable

PAGE 94

94 Table 4 19 Oort method: Imputation using approach 3 (PF imputed with MCMC approach and remaining domain scores wit h IADL scores) (n=1037 ) Model Df CHISQ RMSEA SRMR CFI NFI NNFI Model 1 84 346.451 0.055 (0.049 0.061) 0.034 0.990 0.987 0.986 Model 2 109 420.019 0.052 (0.047 0.057) 0.038 0.988 0.984 0.987 Model 3 Reprioritization (VT) 108 413.882 0.052 (0.046 0.057) 0.038 0.988 0.984 0.987 Uniform recalibration (PF) 107 392.839 0.050 (0.045 0.056) 0.037 0.989 0.985 0.988 Non uni form recalibration (PF) 106 385.458 0.050 (0.045 0.055) 0.037 0.989 0.985 0.988 Non uniform recalibration (VT) 105 374.760 0.049 (0.044 0.055) 0.037 0.990 0.986 0.988 Model 4 Final Model (all tenable constraints imposed) 109 380.925 0.049 (0.043 0 .054) 0.039 0.990 0.986 0.989 Abbreviations: RMSEA root mean square error of approximation, SRMR standardized root mean square residual, CFI comparative fit index, NFI normed fit index, NNFI non normed fit index. PF physical functioning, BP bodily pain. RMSEA <= 0.05 close fit; 0.05 to 0.08 reasonable fit; >= 0.10 poor fit CFI > 0.90 reasonably good fit, SRMR < 0.10 is favorable

PAGE 95

95 Table 4 20 Oort method: Imputation using approach 4 (N = 2317) Model Df CHISQ RMSEA (90% CI) SRMR CFI NFI NNFI Model 1 84 775.190 0.069 (0.064 0.073) 0.037 0.988 0.986 0.983 Model 2 109 1574.624 0.076 (0.073 0.080) 0.051 0.980 0.979 0.978 Model 3 Reprioritization (VT) 108 1545.216 0.076 (0.073 0.079) 0.048 0.980 0.979 0.978 Reprioritization (SF) 107 1532.692 0.076 (0.073 0.079) 0.046 0.981 0.979 0.978 Uniform recalibration (PF) 106 1380.884 0.072 (0.069 0.076) 0.044 0.983 0.981 0.980 Uniform recalibration (BP) 105 1365.455 0.072 (0.069 0.076) 0.044 0.983 0.981 0.980 Non uniform recalibrati on (PF) 104 1290.516 0.071 (0.067 0.074) 0.043 0.984 0.982 0.981 Non uniform recalibration (RP) 102 1217.155 0.069 (0.065 0.072) 0.043 0.985 0.983 0.982 Non uniform recalibration (GH) 102 1190.308 0.068 (0.065 0.072) 0.042 0.985 0.984 0.983 Non uniform recalibration (VT) 101 1136.287 0.067 (0.064 0.071) 0.042 0.986 0.985 0.983 Non uniform recalibration (SF) 100 1087.854 0.066 (0.063 0.070) 0.041 0.987 0.985 0.984 Non uniform recalibration (RE) 99 1061.802 0.066 (0.063 0.069) 0.040 0.98 7 0.986 0.984 Non uniform recalibration (MH) 98 1027.187 0.065 (0.061 0.068) 0.038 0.987 0.986 0.984 Model 4 Final Model (all tenable constraints imposed) 102 1095.932 0.066 (0.062 0.069) 0.045 0.986 0.985 0.984 Abbreviations: RMSEA roo t mean square error of approximation, SRMR standardized root mean square residual, CFI comparative fit index, NFI normed fit index, NNFI non normed fit index. PF physical functioning, BP bodily pain. RMSEA <= 0.05 close fit; 0.05 to 0.08 reason able fit; >= 0.10 poor fit CFI > 0.90 reasonably good fit, SRMR < 0.10 is favorable

PAGE 96

96 Table 4 21 Schmitt method: Imputation using approach 1 (PF and MH present and oth er domain scores imputed using M CMC ) (n=988) Model Df CHISQ RMSEA SRM R CFI NFI NN FI M odel 1 84 322.618 0.054 (0.048 0.060) 0.034 0.991 0.988 0.98 7 Model 2 Covariances constrained (reconceptualization) 87 326.875 0.053 (0.047 0.059) 0.038 0.991 0.988 0.98 7 Model 3 Factor covariances and variances constrained (recalibration) 91 333.80 6 0.052 (0.046 0.058) 0.040 0.991 0.987 0.98 8 Model 4 Factor covariances, variances, and loadings constrained (recalibration) 98 345.920 0.051 (0.045 0.057) 0.041 0.991 0.987 0.98 8 PF (factor loading) 97 339.732 0.051 (0.045 0.057) 0.040 0.991 0.987 0.98 9 Model 5 Factor covariances, variances, loadings, and uniqueness constrained 107 370.565 0.050 (0.045 0.056) 0.041 0.990 0.986 0.98 9 PF 106 359.650 0.050 (0.044 0.055) 0.041 0.990 0.986 0.98 9 BP 105 354.438 0.049 (0.044 0.055) 0.040 0.991 0.987 0.98 9 VT 104 349.062 0.049 (0.044 0.055) 0.040 0.991 0.987 0.98 9 Abbreviations: RMSEA root mean square error of approximation, SRMR standardized root mean square residual, CFI comparative fit index, NFI normed fit index, NNFI non norme d fit index. RMSEA <= 0.05 close fit; 0.05 to 0.08 reasonable fit; >= 0.10 poor fit CFI > 0.90 reasonably good fit, SRMR < 0.10 is favorable

PAGE 97

97 Table 4 22 Schmitt method: Imputation using approach 2 (PF imputed with IADL scores and remaini ng domain score s imputed using M CMC ) (n=1057) Model Df CHISQ RMSEA SRM R CFI NFI NNFI Model 1 84 354.338 0.056 (0.050 0.062) 0.034 0.990 0.987 0.98 6 Model 2 Covariances constrained (reconceptualizatio n) 87 358.203 0.055 (0.049 0.060) 0.037 0.990 0.987 0.98 7 Model 3 Factor covariances and variances constrained (recalibration) 91 365.847 0.054 (0.048 0.060) 0.039 0.990 0.987 0.98 7 Factor variance 90 361.756 0.054 (0.048 0.060) 0.038 0.990 0.987 0.98 7 Model 4 Factor covariances, variances, and loadings constrain ed (recalibration) 97 370.365 0.052 (0.047 0.058) 0.039 0.990 0.987 0.98 8 Model 5 Factor covariances, variances, loadings, and uniqueness constrained 107 393.537 0.051 (0.045 0.056) 0.040 0.990 0.986 0.98 8 PF 106 389.523 0.051 (0.045 0.056) 0.040 0.990 0.986 0.98 8 BP 105 384.445 0.050 (0.045 0.056) 0.039 0.990 0.986 0.98 9 VT 104 378.731 0.050 (0.045 0.056) 0.039 0.990 0.986 0.98 9 Abbreviations: RMSEA root mean square error of approximation, SRMR standardized root mean square residual, CFI comparative fit index, NFI normed fit index, NNFI non normed fit index. RMSEA <= 0.05 close fit; 0.05 to 0.08 reasonable fit; >= 0.10 poor fit CFI > 0.90 reasonably good fit, SRMR < 0.10 is favorable

PAGE 98

98 Table 4 23 Schmitt method: Imputation usi ng approach 3 (PF imputed with M CMC approach and remaining domain scores with IADL scores) (n=1037 ) Model Df CHISQ RMSEA SRM R CFI NFI NN FI Model 1 84 346.451 0.055 (0.049 0.061) 0.034 0.990 0.987 0.98 6 Model 2 Covariances constrained (reconceptualiza tion) 87 351.215 0.054 (0.048 0.060) 0.038 0.990 0.987 0.98 6 Model 3 Factor covariances and variances constrained (recalibration) 91 355.871 0.053 (0.047 0.059) 0.039 0.990 0.987 0.98 7 Model 4 Factor covariances, variances, and loadings constrained (re calibration) 98 368.402 0.051 (0.046 0.057) 0.041 0.990 0.986 0.98 7 PF 97 361.927 0.051 (0.046 0.057) 0.040 0.990 0.986 0.98 8 Model 5 Factor covariances, variances, loadings, and uniqueness constrained 107 396.937 0.051 (0.045 0.056) 0.041 0.989 0 .985 0.98 8 PF 106 387.352 0.050 (0.045 0.056) 0.041 0.989 0.985 0.98 8 VT 105 374.368 0.050 (0.044 0.055) 0.040 0.990 0.986 0.98 8 Abbreviations: RMSEA root mean square error of approximation, SRMR standardized root mean square residual, CFI c omparative fit index, NFI normed fit index, NNFI non normed fit index. RMSEA <= 0.05 close fit; 0.05 to 0.08 reasonable fit; >= 0.10 poor fit CFI > 0.90 reasonably good fit, SRMR < 0.10 is favorable

PAGE 99

99 Table 4 24 Schmitt method: Imput ation using appr oach 4 (N = 2317 ) Model Df CHISQ RMSEA SRMR CFI NFI NNFI Model 1 84 775.190 0.069 (0.064 0.073) 0.037 0.988 0.986 0.983 Model 2 Covariances constrained (reconceptualization) 87 978.013 0.067 (0.064 0.071) 0.038 0.988 0.987 0.983 Model 3 Factor co variances and variances constrained (recalibration) 91 1051.605 0.068 (0.065 0.072) 0.045 0.987 0.986 0.983 90 1041.211 0.069 (0.065 0.072) 0.043 0.987 0.986 0.983 89 1015.815 0.068 (0.065 0.072) 0.041 0.987 0.986 0.983 88 984.569 0.067 (0.0 64 0.071) 0.039 0.988 0.987 0.983 Model 4 Factor covariances, variances, and loadings constrained (recalibration) 95 1021.200 0.066 (0.062 0.070) 0.041 0.987 0.986 0.984 PF 94 1010.544 0.066 (0.062 0.069) 0.041 0.988 0.986 0.984 VT 93 989.275 0. 065 (0.062 0.069) 0.039 0.988 0.987 0.984 Model 5 Factor covariances, variances, loadings, and uniqueness constrained 103 1347.050 0.072 (0.069 0.076) 0.043 0.983 0.982 0.980 PF 102 1264.030 0.070 (0.067 0.074) 0.042 0.984 0.983 0.981 RP 101 119 6.048 0.069 (0.065 0.049) 0.042 0.985 0.984 0.982 GH 100 1171.043 0.068 (0.065 0.072) 0.042 0.985 0.984 0.982 VT 99 1118.205 0.067 (0.064 0.071) 0.042 0.986 0.985 0.983 SF 98 1068.300 0.066 (0.063 0.070) 0.040 0.987 0.985 0.984 RE 97 1048.9 51 0.066 (0.062 0.070) 0.040 0.987 0.986 0.984 MH 96 1020.859 0.065 (0.062 0.069) 0.040 0.987 0.986 0.984

PAGE 100

100 Table 4 2 5 M easurement bias and re sponse shift detection (when age was used as a dichotomous variable) using imputation approach 1 (n = 988) Step Model Df CHISQ RMSEA (90% CI) SRM R CFI NFI NN FI Step 1: Model 1: Measured Model 84 322.12 5 0.054 (0.048 0.060) 0.03 4 0.99 1 0.98 8 0.9 87 Step 2: Detect measuremen t bias Model 2: First model, without accounting for measurem ent bias 27 3 765.70 8 0.046 (0.042 0.050) 0.04 0 0.98 2 0.97 4 0.9 73 Uniform recalibratio n for PF 27 2 732.42 3 0.045 (0.041 0.048) 0.04 0 0.98 4 0.97 5 0.9 75 Uniform recalibratio n for BP 27 1 725.63 1 0.044 (0.041 0.048) 0.04 0 0.98 4 0.97 6 0.9 75 Gender TPF 27 0 689.58 1 0.043 (0.039 0.047) 0.03 9 0.98 5 0.97 7 0.9 77 Gender BPF 26 9 665.85 7 0.042 (0.038 0.046) 0.03 9 0.98 6 0.97 8 0.9 78 Educ BPF 26 8 650.76 3 0.041 (0.037 0.045) 0.03 9 0.98 6 0.97 8 0.9 78

PAGE 101

101 Table 4 26 M easurement bias and re sponse shift dete ction (when age was used as a dichotomous variable) using imputation approach 2 (n = 1057 ) Step Model Df CHISQ RMSEA (90% CI) SRM R CFI NFI NN FI Step 1: Model 1: Measured Model 84 360.04 4 0.056 (0.050 0.062) 0.03 4 0.99 0 0.98 7 0.98 6 Step 2: Detect measuremen t bias Model 2: First model, without accounting for measurem ent bias 27 3 827.93 0 0.047 (0.043 0.051) 0.04 1 0.98 1 0.97 4 0.97 2 Uniform recalibratio n for PF 27 2 802.37 2 0.046 (0.043 0.050) 0.04 1 0.98 2 0.97 5 0.97 3 Uniform recalibrat io n for BP 27 1 795.96 7 0.046 (0.043 0.050) 0.04 1 0.98 2 0.97 5 0.97 3 Age TPF 27 0 749.50 5 0.044 (0.040 0.048) 0.04 1 0.98 4 0.97 6 0.97 5 Educ BPF 26 9 732.85 9 0.043 (0.040 0.047) 0.04 1 0.98 4 0.97 7 0.97 6

PAGE 102

102 Table 4 27 M easurement bias and re spon se shift detection (when age was used as a dichotomous variable) using imputation approach 3 (n = 1037 ) Step Model Df CHISQ RMSEA (90% CI) SRM R CFI NFI NN FI Step 1: Model 1: Measured Model 84 346.451 0.055 (0.049 0.061) 0.03 4 0.99 0 0.98 7 0.98 6 Step 2: Detect measuremen t bias Model 2: First model, without accounting for measurem ent bias 27 3 941.096 0.052 (0.048 0.055) 0.04 1 0.97 5 0.96 6 0.96 2 Uniform recalibratio n for PF 27 2 894.588 0.050 (0.047 0.054) 0.04 1 0.97 6 0.96 8 0.96 4 Unifo rm recalibratio n for BP 27 1 875.707 0.050 (0.046 0.053) 0.04 1 0.97 7 0.96 9 0.96 5 Age TPF 27 0 831.231 0.048 (0.044 0.051) 0.04 1 0.97 9 0.97 0 0.96 7 Educ BPF 26 9 816.748 0.047 (0.044 0.051) 0.04 1 0.97 9 0.97 1 0.96 8

PAGE 103

103 Table 4 28 M easurement bi as and re sponse shift detection (when age was used as a dichotomous variable) using imputation approach 4 (n = 2317) Step Model Df CHISQ RMSEA (90% CI) SRMR CFI NFI NNFI Step 1: Model 1: Measured Model 84 775.190 0.069 (0.064 0.073) 0.037 0.988 0.986 0.983 Step 2: Detect measurement bias Model 2: First model, without accounting for measurement bias 273 2459.390 0.062 (0.059 0.064) 0.046 0.973 0.970 0.959 Reprioritization for PF 272 2446.359 0.062 (0.059 0.064) 0.045 0.973 0.970 0.9 59 Reprioritization for RP 271 2439.532 0.062 (0.059 0.064) 0.045 0.973 0.970 0.959 Reprioritization for BP 270 2434.171 0.062 (0.060 0.064) 0.045 0.973 0.970 0.959 Reprioritization for GH 269 2420.597 0.062 (0.059 0.064) 0.045 0.973 0.971 0 .959 Uniform recalibration for PF 268 2323.380 0.060 (0.058 0.063) 0.045 0.975 0.972 0.961 Uniform recalibration for BP 267 2306.358 0.060 (0.058 0.063) 0.045 0.975 0.972 0.961 Gender TPF 266 2221.249 0.059 (0.057 0.062) 0.044 0.976 0.973 0.962 Gender BPF 265 2174.330 0.059 (0.057 0.061) 0.044 0.976 0.974 0.963 Age TPF 264 2077.683 0.057 (0.055 0.060) 0.043 0.978 0.975 0.965 Age BPF 263 2045.344 0.057 (0.054 0.059) 0.043 0.978 0.975 0.965 Educ BPF 262 1989.029 0.05 6 (0.054 0.058) 0.043 0.979 0.976 0.966 Gender TRP 261 1953.894 0.056 (0.053 0.058) 0.042 0.979 0.976 0.967 Age BRP 260 1921.734 0.055 (0.053 0.058) 0.042 0.979 0.977 0.967 Age TRP 259 1884.471 0.055 (0.052 0.057) 0.041 0.980 0.977 0.967

PAGE 104

104 CHAPTER 5 DISCUSSION T he Oort procedure showed evidence of uniform recalibration. Apparently, the meaning of the response scale anchors for the PF scale changed since assignment to treatment strategy. The Oort approach also showed evidence of no n u niform recalibration for PF scale. possibly the meaning of some of the anchors for PF response scale has changed in a group of hypertensive CAD patients. The Oort approach also identified uniform recalibration for the BP scale. uniform recalibration for the PF scale was identified. All response shifts occurred within PHYS HRQoL construct. We found that most Model 4 parameters were invariant across oc casions, except for the latent means that did change over time (Table 4 6). Common factor variances and common factor correlations did not change across occasions, but the common factor means did. We thus did not find any evidence of reconceptualization of a possible higher order factor representing HRQoL, or of reprioritization of its components. Common factor means were fixed at zero for the first occasion (because of identification requirements) so that the second occasion estimates were a direct represe ntation of change. In addition to significance test results, Table 4 7 provides effect sizes for observed change, and the response shift and true change contributions to observed change, as implied by the parameter estimates of Model 4 (in Table 4 5). From Table 4 7 it appears that the response shift effects on observed change were only small: 0.118 for the uniform recalibrat ion for PF scale The effects of true change were smaller for PF. For PF the effects of response shifts and true change were in oppos ite directions. In

PAGE 105

105 had a marginal impact on the estimation of true change in the means of the physical factor. Although minimal, we did identify response shift among the se patients which warrant the assessment of response shift. Difference in Results between Oort and Schmitt Procedures in this Study We now elaborate on alternate interpretations for the disparate results among the Oort and Schmitt procedures in our study. With Schmitt procedure, constraints we re added to the model to identify a response shift. Contrary to the Schmitt procedure, constraints we re removed to identify a response shift with the Oort procedure. The constraints added or removed are equality constr aints, where a parameter estimate at one time is made to be equal to another time. All types of response shift (reconceptualization, reprioritization, and uniform and non uniform recalibration) can be detected using the Oort procedure. On the contrary, the Schmitt procedure does not identify reprioritization (since the covariance analysis used is incapable of identifying changes in definition from change in values) or non uniform recalibration and it interprets changes in factor loadings differently from th e approach described by Oort. Comparable to the Oort approach identifying a response shift in this study, the Schmitt procedure was also supportive of a response shift. Both the Oort and Schmitt SEM procedures showed evidence of a recalibration for PF scal e. Based on our results, the differences between the Oort and Schmitt approaches in our study are because of variation in the method and not the sample used. Method differences may result from divergence in defining the type of response shift linked to cha nges in various parameters of the SEM measurement model. Although both the statistical approaches seem apt to identify response shift, researchers ar e recommended to clearly

PAGE 106

106 demonstrate the approach used and limitations for the said approaches. Bonferroni correction is strongly recommen ded when using the Oort procedure to detect response shift If the order of testing does not matter, w e recommend use of both sequences (1) release constraints on residuals, followed by inter cepts, and then factor loadings; and (2) release constraints on factor loadings, intercepts, and then residuals. T he Oort procedure operationalizes various forms of response shift differently using the S chmitt approach, our conclusions about the type of a response shift are different between the two approaches. Oort defines uniform recalibration as a change in means of the observed variables over time compared to Schmitt who defines recalibration as a cha nge in factor loadings or factor variances over time. According to Oort, respondents change their interpretation of the response scale options (values, labels, or both). If this change affects all response options in the same direction and to the same exte nt, then the recalib ration is explained as uniform. One explanation for this change may be that new experiences after the assignment of treatment strategies may have changed the idea of how much does health limit those performing moderate or vigorous activ ities, lifting objects, walking or climbing stairs. An alternate explanation may be if patients have learned how to cope with their illness, so that their functioning improved more, or deteriorated less, than would be expected on account of their physical health. In addition to the identification of uniform recalibration, non uniform recalibration definition, if only some points of a response scale are associated with change in t he

PAGE 107

107 same direction and to the same extent, then the shift may be non uniform. Conversely, the Schmitt technique showed evidence of a change in r andom error over time for the PF scale. According to Schmitt, the test for recalibration concerns the equality of scaling units in the factors (when the measures are congeneric, i.e. have the same factor loadings). In other words, absence of recalibration indicates that the elements of factor loadings are equal ac ross measurements. Difference between the Present Stu dy and Ahmed et al. Study Ahmed et al. [ 35 in the same sample. Converse to the results of our study, Ahmed et al. [ 35 ] found that the Oort procedure detected a response shift after a chronic o bstructive pulmonary disease self management program not detected by the Schmitt pro cedure. The authors report the Oort procedure to be more sensitive in detecting a response shift compared to the Schmitt procedure. We elaborate and provide alternate expla nations on the [35] First, our study applied the Bonferroni correction to avoid making a Type I error when identifying a response shift with the Oort procedure. The Ahmed et al. [ 35 ] study did n ot control for Type I error even though the Oort procedure [ 10 ] has been suggested t o increase the chances of Type I error. As a result, their study may have wrongly identified a response shift when differences may be attributed to other types of changes. Second, similar to the approach taken by Nolte and Osborne [ 70 ], Ahmed et al. [ 35 ] chose to release constraints on residuals, followed by intercepts, and then factor [35] we conducted the Oort SEM approach as recommended by several other authors. As reflected in the Oort study and the order tested by the authors, and also recognized by

PAGE 108

108 Donaldson [38] and Barclay Goddard [ 39 ], we chose to test sequentially for invariance of factor loadings, intercepts, and error variance. Ahmed et al. [35] may have found divergent r esults if they would have tested for the presence of response shift by releasing parameter constraints on factor loadings, followed by intercepts, and then residuals Ahmed et al. [ 35 ] advocate that the order in which the constraints are released make a difference as to whether or not subsequent parameters would be significant. However, it has been questioned that if the sequence of testing does not alter the end result, order of t esting may not matter. Therefore, we tested for pres ence of response shift using a similar approach to Ahmed et al. [35] and Nolte and Osborne [70] where we chose to release constraints on residuals, followed by intercepts, and then factor loadings. Using either of the two sequence of te sting did not alter our results; order of testing did not matter in this study. Based on our results, the differences between our study and Ahmed et al. study are because of variation in the method and the sample used. Meth od differences resulted due to use of Bonferroni and order of testing for response shift as suggested by previous authors. Sample differences may have resulted due to a larger proportion of the study sample experiencing a response shift compared with indiv iduals with COPD, making response shift among hypertensive CAD patients more detectable. Validating the Presence of Response Shift in Hypertensive CAD Patients One explanation of the finding of response shift for the SF 36 PF scale may be demonstrated as a consequence of the disease condition. Hypertension and CAD are chronic health conditions that impact all aspects of function, perception, cognition, mood, QOL [11 13]. All of the negative components of health (impairments, activity

PAGE 109

109 limitations, and partic ipation restrictions) are grouped under the term disability. Under the ICF model [63] impairment, such as decreased strength or coordination of the limbs, causes limitations in activities, such as walking, which eventually leads to participation restricti ons. The aspects of functioning that are impacted by ischemic heart disease (which commonly refers to CAD) [63] have been agreed upon by a consensus process conducted by the ICF Research Branch of the WHO Family of International Classifications Collaborati ng Center. Examples of the content areas included in the ICF Core Set that are likely to have a d irect impact on HRQoL include: reduced walking distances, shortness of breath, energy and drive functions, lifting and carrying objects, moving around, and doi ng housework. The SF 36 PF consists of items describing walking distances, vigorous activities, moderate activities, lifting and carrying objects, climbing stairs, among others. We suggest that change in interpretation may have affected the response scale options (values, labels, or both) for the PF scale in the same direction and to the same extent for a group of patients due to aspects of functioning impacted by the disease condition. We also found presence of non uniform recalibration for the PF scale. W e recommend that evaluation of specific items in the PF scale may explain why non uniform response shift occur red in PF WHO ICF recommend s a spects that describe difficulty in PF include impairments (such as decreased strength) or activity limitations (dif ficulty walking, lifting and carrying objects). Analogous to the aspects that describe di fficulty in PF we believe that items describing vigorous activities such as running, lifting heavy objects; lifting or carrying groceries, and walking more than a mil e or several blocks may explain why non uniform response shift occurred for PF. The items

PAGE 110

110 describing aspect undergo change in interpretation over time. Perhaps new experiences after th e assignment to treatment strategy have changed their idea of how much does health limit these patients in their physical activities. Presence of response shift may be explained due to mechanisms used to accept or adapt to changes in health status and natu re of impairments and activity limitations that many individuals with hypertension and CAD are likely to experience. Some studies have shown lower HRQoL in individuals with CAD compared to those without CAD. Mitchell et al. [ 14 ] conducted a study among eld erly subjects with differing cardiovascular status (cardiovascular normal, being hypertensive, having isolated CAD, or both being hypertensive and having CAD). The NHP questionnaire was administered to these participants. The NHP questionnaire incorporates questions related to activity limitations and impairments similar to the classificati on provided by the ]. Impairments such as decreased strength cause limitations in activities, such as walking, which eventually may lead to participation res trictions, i.e. restricted ability to carry out usual activities in the community, such as through work or volunteering. Mitchell et al. [ 14 ] observed lower mean scores on the physical ability dimension in hypertensive CAD patients. In a separate study, Me yer Rosberg et al. [ 83 ] have shown that the physical ability scale of NHP is substantially correlated ( 0.79) to the SF 36 PF scale. Vascular disease such as CAD mainly influences ph ysical health [84] Sevinc and Akyol [ 85 ] have found that QOL scores were lowest in PF domain for CAD patients compared to emotional and social domain scores. By analogy, several

PAGE 111

111 studies report that hypertensive CAD patients feel limited in the physical functioning domain. An alternate explanation for the presence of response sh ift may be due to the assignment of antihypertensive treatment strategies. Antihypertensive treatment focuses on individual attributes/domains of HRQoL which are potentially important to the hypertensive CAD patients. Bass et al. [ 86 ] conducted a double bl ind prospective randomized trial on patients awaiting coronary artery bypass graft surgery assigned to atenolol or placebo treatment. Atenolol was reported to show adverse somatic side effects such as feelings of pressure and tightness in the head or body, dizziness and faintness, fatigue, and muscular weakness. In a separate study, Ried et al. [ 17 ] report that patients assigned to atenolol led treatment strategy are more likely to have decreased energy and increas ed fatigue. These aspects are analogous to the content areas included in the ICF Core Set that describe difficulty in physical functioning including impairments (such as decreased strength) or activity limitations (difficulty walking, lifting and carrying objects). These aspects may partly explain why uniform and non uniform response shift occurred for PF. The items describing various activities Response shift is thus important to consider in treatment evaluations, es pecially insofar as it may serve to attenuate or to exaggerate estimates of treatment effects as patients adapt to tr eatment side effects or disease progression over time. In addition to the alternative explanations stated above, several studies suggest a relationship between gender and PF and age and PF among CAD patients. De Graff et al. [84] found that gender had a significant impact on the SF 36 PF domain among CAD

PAGE 112

112 patients. The authors also report a small but significant inverse relation between age an d SF 36 PF domain. Sevinc and Akyol [85] have found better QOL scores among male CAD patients compared to female CAD patients. The authors also found that the scores for the SF 36 PF domain for CAD patients in the 70 79 year age group were significantly lo wer than those of the other age groups. In a s eparate study, Norris et al. [87 ] report, when compared to men, women with CAD had lower scores in all QOL domains. Few studies [88,89 ] report of a stereotype implying that men are more physically oriented and invested to do heavier physical labor or strenuous activities. The authors of these studies allege that if PF (e.g. walking long distances, heavy lifting) image, then it would not be surprising to observe m en underreporting their factual physical limitations (i.e. higher PF scores) to preserve their self image [88, 8 9 ]. In other words, men seem to be over reporting their PF to maintain their perceptions of their masculinity and to protect their egos by reduci ng gender role conflict. The identification of response shift may also be due to different age groups or gender roles which operate diffe rently for males and females [88, 8 9 ]. Identification of Response Shift for the PF Domain None of the previous studies h ave found response shift for the SF 36 PF scale. Oort et al. [ 36 ] suggest that SF 36 PF items are less prone to recalibration because the [36] after allowing for possible recalibration for BP items, most of their significant modification indices disappeared. In our study, we checked for alternative options in specification searches, but no other selection provided a better model fit. We thus chose to free the intercept constraint for PF scale, indicat ing evidence of recalibration.

PAGE 113

113 Barclay Goddard et al. [39 ] identified the presence of response shift post stroke in PF construct over time. Based on their results, Barclay Goddard et al. [ 39 ] suggest that that are not considered to be true change may be interpreted as response shift in evaluation methods, while it may be measurement error in performance based (such as a timed valuation of difficulty in their study partly explained why response shift occurred in the model of PF construct. They, however, acknowledge that it was not the only explanation; two of the four measured variables for PF construct in which response shift w as suggested contained the evaluation of difficulty. We propose that presence of response shift for the SF 36 PF scale in our study populati on may be explained due to a single or a combination of the factors including the dis ease condition, treatment strat egies assigned presence of depressive symptoms age, and/or gender [39] This concept deserves further evaluation. Barclay Goddard et al. [39] point o ut in their framework article that SEM requires validation of the model We thus chose to conduct a sensi tivity analysis comparing the complete cases data set (n=909) with the four different imputed data sets as an example of validation. Model fit and presence of response shift was compared across the imputed datasets using MLE. When comparing the results fro m complete cases and those with different imputation approaches we found somewhat similar, but not identical identification of the location of response shift in our study. Table 1 7 Table 2 4 demonstrate s assessment of response shift using Oort and Schmitt a pproaches using

PAGE 114

114 the four imputation approaches. The uncertainty with results obtained after imputation cannot be ignored and should be interpreted with caution. Implications The findings in this study can serve to inform future research of response shift f or PF in hypertensive CAD patients. Barclay Goddard et al. [ 39 ] suggest that response shift presence in PF will affect the way in which one evaluates PF in the clinic and in research studies, the type of measures that we use, and the way in which change in physical function is analyzed. Schwartz and Rapkin [90 ] describe the difference between performance based measures (such as a timed walking test), perception based measures (such as an individual describing how many times a day t hey walk), and evaluation based/ self perceived measures (i.e. how difficult is it to walk?). They describe changes over time that is not true change may be: measurement error in performance based measures, response bias in perception based measures, or response shift in evaluation based methods [90 ] We assert that t 36 dif ficulty in PF may be susceptible to response shift. A compariso n of models of PF with only pe rformance based measures and models with self perceived measures would be of interest in CAD and other chronic conditions. Barclay Goddard et al. [ 39 f measures of PF demonstrate response shift, choice of self perceived or performance bas ed measures need to be carefully chosen by the clinician to avoid b ias in change estimates Identification of response shift in SF 36 PF in our study and PF construct in Barclay Goddard study [39]

PAGE 115

115 warrant assessment of response shift when using self perce ived measures in similar populations and other chronic populations. Measurement of change in HRQoL will be inaccurate if response shift is present, but not accounted for. The effect size of HRQoL change can be estimated, taking response shift into account, after evaluation with the Oort SEM technique. In the present study, HRQoL change was small after accounting for response shi ft. This was demonstrated in PF where observed change (true change plus response shift) in physical activity had a small effect si ze which increased slightly when response shift was removed. Moreover, for PF, the effects of response shifts and true change were in opposite directions. In other words, after accounting for response shift, we observe a reversal in sign from negative (for observed score) to positive (for true change). According to Rosenberg [ 9 1 ] a distorter variable reveals that the correct interpretation is precisely the reverse of that suggested by the original data. Analogous to report [ 9 1 ] response shift a cts a distorter variable, which, when controlled, causes the relationship between the independent and dependent variables to change direction. In our case, it changed the direction from negative sign for the observed scores to pos itive sign for true change scor es. The literature on the SF 36 health survey in a variety of disease states has suggested a minimal clinical ly meaningful difference to be 7 16 points on the SF 36 PF scale as they reflect a magnitude of change perceptible to patients [92 ] After ac counting for response shift with the Oort procedure, the mean PF score at one year was estimated to be 58, resulting in a mean change of ~3 points. Our finding of a n average 3 point change was a little less than one half to one fifth of the 7 16 point

PAGE 116

116 diff erence deemed clinically significant. Previous research has shown that a lower SF 36 Physical Component Summary score (note: physical component score consists of physical functionin g) may be associated with an increased risk of stroke and predicts higher m ortality [ 93 ]. Although there was an improvement in mean scores for the PF scale (3 points), presence of response shift resulted in slight worsenin g of PF scale scores over time. I nitially we found a decrease in change in scores from pre to post test. Ho wever, this relationship did not take into account the effects of response shift (acting as a distorter) After controlling for response shift, we found that the change from pre to post test was positive, i.e., there was an impr ovement in PF scores over t ime. It should be noted that the true change was marginal, warranting the need for future studies to investigate the role of response shift as a distorter variable. According to Wyrwich et al. [94 ], measuring clinically meaningful differences will likely f acilitate the clinical interpretation of health status changes experienced by patients with heart disease and improve clinical decision making. The authors suggest that HRQoL measures provide valuable evidence on changes in emotional, social and physical f highlight different perspectives concerning treatment options that can be balanced against changes in ]. However, the authors agre e that clinically m eaningful differences are not sufficient to drive subsequent clinical, treatment, and reimbursement decisions. They recommend use of clinically meaningful differences in research settings to demonstrate the clinically important effects of interventions on HRQoL [94 ] Future studies measuring clinical importance has

PAGE 117

117 implications for interpreting the measu rement of change in both clinical and research endeavors. Barclay Goddard et al. [ 34,39 ] recommend that clinically, if recalibration response shift is suspe cted in a particular construct of HRQ oL (as demonstrated in PF in our study), a design method such as a then test could be used to capture response shift over time during treatment. They further propose to measure response shift in clinical trial research, as the estimates of the treatment effects may be underestimated and therefore inaccurate, possibly leading to a conclusion that is false negative. In studies or settings where response shift is suspected and design methods are not feasible to assess the a mount of response shift, statistical methods should be applied such as those applied here R esponse shift assessment is particularly important in hypertensive CAD patients when impairments and activity limitations may never recover fully, but improved HRQo L is a go al. If poor PF observed in patients with hypertension and CAD is associated with poor QOL, this information may be used by health professionals to target recovery interventions to those who might be most responsive. Increased knowledge of response shift will therefore affect the way in which HRQoL measures are used in clinical, research, and policy decisions. Limitations L imitations should be noted. Ahmed et al. [35] report that SEM approaches detect response shift at a group level. O nly if a subst antial number of participants undergo a response shift, will it be detected using either of the Oort [ 10 ] or Schmitt SEM [ 28 ] procedure. Thus, if barely a few individuals undergo a response shift on the SF 36 domains, it may not be detected using either of the two statistical approaches. Barclay

PAGE 118

118 Goddard et al. [ 34,39 ] suggest, not all individuals would experience response shift. However, those individuals who do experience response shift may not experience it at the same time. This may influence study resul ts when using group level analyses similar to our study. Converse to the group level analyses, Mayo et al. [ 9 5,96 ] propose a method that uses the residuals generated from a growth model where health behaviors measured or functions were used to predict perc eived health. Ahmed et al. [35] suggest that if the re are enough time points (i.e. greater than three time points), analytical approaches that examine response shift at an individual level, such as growth curve modeling [ 97 ], and latent trajectory analysis described by Mayo et al. [ 95,96 ] will be more helpful to identify subgroups of patients who experience a response shift. Future comparisons with individual level methods are warranted. The use of residuals to identify those who potentially underwent respo nse shift is an intriguing area of study. This would enable multi group analysis in SEM with, for example, those who appear to have undergone response shift based on residuals, and those who do not appear to have undergone response shift. This is a promisi ng area of study with secondary data analysis. Cognitive interviews, used in previous response shift studies [ 9 8 ] are shown to provide insight into whether the statistical, design, and individualized approaches are a true reflection of response shift. We i dentified response shift among hypertensive CAD patients assigned to antihypertensive treatment strategies. However, availability of a control group not expected to experience a response shift would provide support for our results. Power and sample size in large longitudinal models needs to be considered. In our study the

PAGE 119

119 sample size was 909 with more than 100 parameters estimated. It has been suggested that the number of cases to free parameters estimated be at least ten to one. For the unconstrained long itudinal model, a sample size of 909 might be considered som ewhat small. The measurement model that we used in the response shift detection procedure using both Oort and Schmitt procedures fitted the data closely (RMSEA = 0.05). With a sample size of 909, the statistical power to reject such a model if it had been unreasonable was 95% [78]. We did also take a look at presence of response shift when comparing complete cases data set with imputed data sets. We found similar, but not identical identification o f the location of response shift in our study when using imputed data sets. This deserves future study. For Research Question# 2 Based on the findings in this study, we found that patients recalibrated their perception of PF after assignmen t to treatment s trategies, i.e. at one year. When investigating from the measurement perspective, we found that the relationship between the SF 36 scale (PF) and gender cannot be explained via their relationship with PHYS HRQoL. We chose to allow gender to have a direct e ffect on PF indicating that the measu rement bias is in the PF scale. The violation of measurement invariance was not consistent acro ss occasions (estimated at 0.113 at baseline and 0.150 at one year, Table 4 15 ) which indicated that male patients reported better PF than female patients, even if their PHYS HRQoL was similar. King Kallimanis et al. [32] recommend that the PF scale contains specific questions which can be answered almost objectively. According to the authors, it is difficult for individualized interpretation of the PF items. They believe that other SF 36 scales measuri ng PHYS HRQoL are more prone to subjective interpretations. However, to obtain a more parsimonious model with the

PAGE 120

120 same fit, they allowed for a direct effect of age on PF scale in their study. By analogy to their study, we chose the final model based on the presence of measurement bias and response shift in the smallest number of scales, providing a more parsimonious model. The order and absolute value of modification indices assist in subjective decision making of the SEM model. In addition to the modification indices in our study, past research played a role in deciding the best model fit when we allowed for a direct effect of gender on PF scale. De Graff et al. [84] found a relati on ship between gender and the SF 36 domains for PF RP SF and RE among CAD patients. In agreement to our study results, Sevinc and Akyol [85] found better QOL scores among male CAD patients compared to female CAD patients. In a separate study, Norris et al. [87 ] report, when compared to men, women with CAD had lower scores in all QOL domains. Few studies [88,89 ] report of a stereotype implying that men are more physically oriented and invested to do heavier physical lab or or strenuous activities. The auth ors of these studies allege that if PF (e.g. walking long distances, heavy lifting) is an integral part of image, then it would not be surprising to observe men underreporting their factual physical limitations (i.e. higher PF sco res) to preserve their self image [88,89 ]. In other words, men seem to be over reporting their PF to maintain their perceptions of their masculinity and to protect their egos by reducing gender role conflict. Our study results are indings that gender roles operate differ ently for males and females [ 88,89 ]. Further assessment into the reasons for the apparent gender difference is warranted. For exploratory purposes, we operationalized presence of depressive symptoms as change in CES D scores over time (note: for the original analysis we used baseline

PAGE 121

121 and one year CES D scores as separate variables). We found that change in CES D scores had a direct effect on PF at the second occasion. The effect of change in CES D scores on PF at the second measurement occasion was positive (0.086), suggesting that at one year depressed patients reported better PF than less depressed patients, even if their true HRQoL was similar. However, use of change in CES D scores as a covariate led to problems wi th interpretation since it is not known if the direct effect is an indication of measurement bias and/or presence of response shift. The differences in results warrants further investigation on the operationalization of depression as a variable to detect r esponse shift. Future studies should define depression in different ways to enable proper interpretation of the results. We found that depressed patients were not likely to undergo response shift. F ew studies corroborate our study findings where they sugge st that depressed individuals will not undergo response shift. Sprangers and Schwartz [4] claim that depression might impede the occurrence of response shift. A separate study by Bar On et al. [15] found that depressed individuals were less likely to under go recalibration response shift. It events such as side ef fects, therefore, these patients are less likely to undergo response shift. However, contrary to our findings, L enert et al. [9 9 ] suggest depressive patients may have systematic differences in how they report their val ues in psychometric testing. Gibbons et al. [ 100 ] report that depressed patients may report higher QOL since they may re evaluate their baseline statu s and decide that they were much worse in the past compared to now (i.e. negative response shift) resulting in response shift. The authors

PAGE 122

122 suggest downward comparison among persons who are depressed and individuals who are ill and have a pessimistic progn osis i.e. those expecting to get worse may result in a negative response shift. They explain that learning of others whose deterioration is more advanced or more evident can remind individuals of their own decline. They suggest that a pessimistic shift in focus is more likely among persons who are depressed to begin with. Moreover, they recommend that by preventing depression one may moderate the type of response shift experienced. In the framework, validation of the full model only is suggested. There is an opportunity to assess validation through the process of the identification of location of response shift. We used MCMC and found similar, but not identical identification of the location of response shift in our study which deserves future study. Impli cations Our findings have several clinical and public policy implications Women have a higher incidence of depression, and in combination with lower PF scores would worsen their QOL. Ameliorating depression caused reductions in PF and maintaining an indiv need for earlier institutionalization and in crease independent living time. Boini et al. [65] advocate that impairment in a particular HRQoL dimension should be followed u p by the implementation of specific strategies. For instance, Boini et al. [65] suggest that when impairment is observed on mental dimensions, psychologist when impairment is observed on physical dimensions, cardiac rehabilitation programs medical management. Patients should be educated that self management must take

PAGE 123

123 into account not only cl assical aspects such as diet low in transfat, saturated fat, cholesterol weight, and exercise programs, but also adaptation to the illness and its treatment, that is strategies such as stress or anger management for self car e of the disease [101 ]. In addit ion, cardiac rehabilitation which includes a structured program of education and activity guided toward lifestyle modification, increasing functional capabilities, improving QOL, and emotional support should be made available to improve impact of CAD on pa tient daily life [101 ]. Comparing the influence of different diseases on HRQoL is relevant, since the attitude of the doctor (and society) should be influenced by the impact of a disease on for each domain of the SF 36, it is possible to give specific attention to and direct treatment for the most impaired aspects of QOL of patients with CAD and those vulnerable to response shift; our findings suggest that SF 36 PF domain was vulnerable to bi as in this population. Sevinc and Akyol [85] advocate that nurses dealing with CAD patients should constantly requirements into their clinic and nursing plans. Understanding ho w changes in magnitude and direction of the response shift affects perception of health can provide valuable information when comparing recovery among CAD populations. Assessment of response shift thus is pivotal and warrants further understanding of how d ifferent group of patients perceive their QOL over time. The findings from our study also have implications for cost effectiveness studies. By analogy to the findings from Lenert et al. [9 9 ] our results suggest that bias in self reports of HRQoL could res ult in inadvertent discrimination against patients with poor

PAGE 124

124 health. Assuming other domain scores are similar across males and females, if females were less optimistic in their PF rating scale it would lead to over evaluation of difference in utility bet ween males and females. In agreement to Lenert et a l. [9 9 ] this would result in a n increase in the cost effectiveness of treatments for the females and potentia lly direct health resources away from the males. A and treatment B have similar utilities at baseline and at one year resulting in similar cost effectiveness ratios for both treatment strategies. If however, after accounting for response shift, we may find that change in utilities for treatment B was sm aller than previously estimated. Assuming costs to be constant, we observe that there is an increase in cost effectiveness for assigned to treatment A have utility values of 0.5 and 0.7 at baseline and at one year, respectively. Comparatively, patients assigned to treatment B have utility values of 0.3 and 0.4 at baseline and at one year, respectively. We may conclude that treatment A is more cost effective compared to trea tment B. However, after accounting for response shift we may find that patients assigned to treatment B had a higher change in utility equal to 0.3 instead of 0.1 as shown above. In this case, contrary to our previous results, we observe that treatment B i s more cost effective. With our study results and the two examples stated above we propose the need to demonstrate the effect of response shift on the estimation of utilities and cost effectiveness studies. Limitations The limitations of the study should b e noted whe n interpreting and applying findings to other populations. First, the CES D score is not a diagnosis of depression, but a measure of depressive symptoms experienced in the previous two weeks.

PAGE 125

125 Second, missing data may introduce potential bias. Re spondents who completed both baseline and one year surveys may differ from non respondents. Third, the relationship between depression and physical functioning is likely reciprocal. The model presented in this study assumes that the causal path goes from d epressive symptoms to physical functioning and not vice versa. The recursive nature of this model should be tested. It should be n oted that the sample size of 788 patients might be considered small relative to the number of parameters estimated and statist ical tests. However, we have almost 100% power to reject the hypothesis that the model does not fit our data (according to a power analysis based on RMSEA values of 0.05 and 0.10). By analogy to the King Kallimanis et al. study [32] an issue specific to t he model we selected for the measurement perspective is that the distinction between explanatory variables and violating variables may not always be clear. We distinguished between explanatory variables and potential violating variables based upon prior kn owledge and handled variables during model fitting. However, we agree with King Kallimanis et al. [32] to avoid unjust data exploration and chance capitalization, researchers should have a well defined and clear understanding of the research questions and hypotheses.

PAGE 126

126 CHAPTER 6 CONCLUSION The overall objective of this thesis was to assess response shift in hypertensive CAD patients. For the first objective, we chose to compare Oort [10] and Schmitt [28] SEM approaches to identify response shift in hyperte nsive CAD patients. Both the Oort and Schmitt approaches identified recalibration for the SF 36 PF scale. We found divergent results in our study compared to the Ahmed et al. study when identifying response shift using Oort and Schmitt procedures. First, o ur study applied the Bonferroni correction to avoid making a Type I error when identifying a response shift with the Oort procedure. The Ahmed et al. [ 35 ] study did not control for Type I error even though the Oort procedure [ 10 ] has been suggested t o incr ease the chances of Type I error. As a result, their study may have wrongly identified a response shift when differences may be attributed to other types of changes. Second, contrary to the Oort procedure adopted by Ahmed et al. [35] in their study where t hey chose to release constraints on residuals, followed by intercepts, and then factor loadings, we conducted the Oort approach as recommended and tested by several other authors including Oort [10] Donaldson [38] and Barclay Goddard [ 39 ]. Similar to the approach tested by these authors, we chose to test sequentially for invariance of factor loadings, intercepts, and error variance. However, even when using the Oort procedure adopted by Ahmed et al. [35] we found simila r results. Using either of the two d ifferent sequence of testing did not alter our results; order o f testing did not matter in our study. Our results were robust to either of the two order of testing, justifying the presence of response shift in our study population. Ahmed et al. [35] may ha ve found divergent r esults if they would have tested for the presence of response shift

PAGE 127

127 by releasing parameter constraints on factor loadings, followed by intercepts, and then residuals. Based on our results, the differences between our study and Ahmed et al. study appear due to variation in the method and the sample used. Method differences resulted due to use of Bonferroni and order of testing for response shift as suggested by previous authors. Sample differences may have resulted due to a larger proport ion of the study sample experiencing a response shift compared with individuals with COPD, making response shift among hypertensive CAD patients more detectable. To the best of our knowledge this is the second study to identify response shift in PF Noneth eless, ours is the only study to identify response shift for SF 36 PF scale. The issue of res ponse shift in PF among hypertensive CAD patients and gender role differences deserves further evaluation. The findings in this study can serve to inform future re search of response shift in PF in hypertensive CAD patients. According to Barclay Goddard et al. [34,39] response shift presence in PF will affect the way in which we evaluate PF in the clinic and in research studies, the type of measures that we use, and the way in which change in PF is analyzed. Facilitating response shift can be particularly important when impairments and activity limitations are not expected to recover fully in hypertensive CAD patients, but improved HRQoL is a goal. Given the diverg en t results from Ahmed et al. [35 ] and our study comparing the Schmitt and Oort SEM approaches to detect response shift justify the need to compare these methods to identify response shift in future studies. Groups of individuals with different health condit ions and catalysts for change should be evaluated with the Oort and Schmitt SEM procedures to allow further understanding of response shift over time

PAGE 128

128 in various health conditions Response shift may influence the measurement characteristics of HRQoL measur ement instruments, so it is important to measure response shift to assist in understanding assessment of HRQoL. Our findings suggest that the SF 36 PF scale was vulnerable to bias in this study population. By looking more closely at the scores for SF 36 PF domain in this study population will enable us to provide nuanced attention and direct treatment for the most impaired aspects of QOL. We also found that gender caused response shift in measurement. F emales were less optimistic in their PF rating scale wh ich would lead to over evaluation of difference in utility between males and females. This would result in an increase in the cost effectiveness of treatments for the females and potentially direct health resources away from the males. Women have a higher incidence of depression, and in combination with lower PF scores would worsen their QOL. Thus, it is imperative to improve depression willingness to perform daily activities which may help re duce institutionalization. Knowledge of multiple comorbid conditions present in an individual may help us obtain a better understanding of how one may approach the issues that come along with these conditions, specifically in elderly patients. Assessment o f response shift is pivotal and warrants further understanding of how different group of patients perceive their QOL over time.

PAGE 129

129 LIST OF REFERENCES 1. Rabins PV, Black BS. Measuring quality of life in dementia: purposes, goals, challen ges, and progress. Int P sychogeriatr 2007;19:401 07. 2. Ahmed S, Mayo N, Wood Dauph inee S, Hanley J, Cohen S. The structural e qua tion modeling technique did not show a response shift, contrary to the results of the then t est and the In dividualized approaches. J Clin Epi demiol 2005 ;58:1125 33. 3. Ahmed S, Mayo N, Wood Dauphinee S, Han ley J, Cohen S. Response Shift i nflue nced estimates of c hange in Health Related Quality of Life Poststroke. J C li n Epidemiol 2004;57:561 70. 4. Sprangers MA, Schwartz CE. Integrating Response Shift into Healt h Related Quality of Life Research: A Theoretical Model. Soc Sci Med 1999;48:1507 15. 5. Schwartz CE, Sprangers MA. Methodological Approaches for Assessing Response Shift in Longitudinal Health Related Quality of Life Research. Soc Sci Med 1999;48:1531 48. 6. Wi lson IB. Clinical Understanding and Clinical Implications of Response Sh ift. Soc Sci Med 1999;48:1577 88. 7. Pepine CJ, Handberg Thurmond E, Marks RG, Conlon M, Cooper DeHoff R, et al. Rationale and design of the International Verapamil SR/Trandolapril Study (INVEST): an Internet based randomized trial in coronary artery disease patients with hypertension. J Am Coll Cardiol 1998;32:1228 37. 8. Pepine CJ, Handberg EM, Cooper DeHoff RM, Marks RG, Kowey P, Messerli FH, et al. A calcium antagonist vs a non calcium antagonist hypertension treatment strategy for patients with coronary artery disease. The International Verapamil Trandolapril Study (INVEST): a Randomized Controlled T rial. JAMA 2003;290:2805 16. 9. Korfage IJ, de Koning HJ, Essink Bot ML. Response shift due to diagnosis and primary treatment of localized prostate cancer: a Then te st and a vignette study. Qual Life Res 2007;10,1627 34. 10. Oort FJ. Using structural equation modeling to detect respons e shifts and true change. Qual Life Res 2005;14:587 98. 11. Ried LD, Tueth MJ, Taylor MD, Sauer BC, Lopez LM, Pepine CJ. Depressive symptoms in coronary artery disease patients after hypertensive treatment. Ann Pharmacother 2006;40:597 604 12. Aggarwal A, Ades PA. Exercise rehabilitation of older patients with cardiovascular disease. Car diol Clin 2001;19:525 36.

PAGE 130

130 13. Croog SH, Levine S, Testa MA, Brwon B, Bulpittt CJ, Jenkins CD, et al. The effects of antihypertensive therapy on the quality of life. N Engl J Med 1986;314 (26) :1657 64. 14. Mitchell RA, Imperial E, Kelleher P, Brunker P, Gass G. Perceived health problems in subjects with varying cardiovascular diagnose s. J Behav Med 1991;14:505 12. 15. Bar On D, Lazar A, Amir M. Quantitative assessment of respon se shift in QOL research. Soc Indic Res 2000;49:37 49. 16. Chobanian AV, Bakris GL, Bla ck HR, Cushman WC, Green LA, Izzo JL Jr, for the National Heart, Lung, and Blood Institute Joint National Committee on Prevention, Detection, Evaluation, and Treatment of High Blood Pressure; National High Blood Pressure Education Program Coordinating Comm ittee. The seventh report of the Joint National Committee on Prevention, Detection, Evaluation, and Treatment of High Blood Pressure: the JNC 7 report. JAMA 2003;289:2560 72. 17. Ried LD, Tueth MJ, Handberg E, Kupfer S, Pepine CJ, and the INVEST study group. A Study of Antihypertensive Drugs and Depressive Symptoms (SADD Sx) in Patients Treated With a Calcium Antagonist Versus an Atenolol Hypertension Treatment Strategy in the International Verapamil SR Trandolapril Study (INVEST). Psychosom Med 2005;67:398 406 18. Ried LD, Johnson RE, McFarland BH, Brody KK. Antihypertensive drug use and the risk of depression among older hypertensives in an HMO. J Pharm acoepi 2000;8:1 25. 19. Ried LD, McFarland BH, Johnson RE, Brody KK. Betablockers and depression: the more the murki er? Ann Pharmacother 1998;32:699 708. 20. Ko DT, Hebert PR, Coffey CS, Sedrakyan A, Curtis JP, Krumholz HM. Beta blocker therapy and symptoms of depression, fatigue, and sexual dysfunction JAMA 2002;288:351 57. 21. Lacy CF, Armstrong LL, Goldman MP, Lance LL. Dru g Information Handbook 8t h ed. Hudson: OH ; 2000 01 22. Avorn J, Everitt DE, Weiss S. Increased antidepressant use in patients prescribed beta blockers. J AMA 1986; 255:357 60. 23. Thiessen BQ, Wallace SM, Blackburn JL, Wilson TW, Bergman U. Increased prescribing of antidepressants subsequent to beta bloc ker therapy. Arch Int ern Med 1990;150:2286 90. 24. Bright RA, Everitt, DE. Beta blockers and depression. Evidence against an association. J AMA 1992;267:1783 87.

PAGE 131

131 25. Gerstman BB, Jolson HM, Bauer M, Cho P, Livingston JM, Pl att R. The incidence of depression in new users of beta blockers and selected antih ypertensives. J Clin Epi demiol 1996;49:809 15. 26. Lindberg G, Bingefors K, Ranstam J, Rastam L, Melander A. Use of calcium channel blockers and risk of suicide: ecological find ings confirmed in population based cohort study. Br Med J 1998;316:741 5. 27. Ahmed S, Mayo N, Corbiere M, Wood Dauphinee S, Hanley J, Cohen R. Change in Quality of Life of People with Stroke Over Time: True Change or Response Shift? Qual Life Res 2005;14:611 27. 28. Schmitt N. The use of analysis of covariance structures to assess beta and gamma change. Multivar iate Behav Res 1982;17:343 58. 29. Schmitt N, Pulakos ED, Lieblein A. Comparison of three techniques to assess group level beta and gamma change Appl Psych ol Meas 1984; 8: 249 60. 30. Schaubroeck J, Green SG. Confirmatory factor analytic procedures for assessing change during organizational entry. J Appl Psychol 1989;74(6):892 900. 31. Oort FJ, Visser MRM, Sprangers MAG. Formal definitions of measurement bias and expla nation bias clarify measurement and conceptual perspectives on res ponse shift. J Clin Epidemiol 2009;62:1126 37. 32. King Kallimanis BL, Oort FJ, Visser MRM, Sprangers MAG. Structural equation modeling of health related quality of life data illustrates the mea surement and conceptual perspectives on response shift. J Clin Epidemiol 2009;62:1157 64. 33. Bush DE, Ziegelstein RC, Tayback M, Richter D, Stevens S, Zahalsky H, et al Even minimal symptoms of depression increase mortality risk after acute myocardial infarc tion. Am J Cardiol 2001;88:337 41. 34. Barclay Goddard R, Epstein JD, Mayo NE. Response shift: a brief overview and proposed research priorities. Qual Life Res 2009;18:335 46. 35. Ahmed S, Bourbeau J, Maltais F, Mansour A. The Oort structural equation modeling app roach detected a response shift after a COPD self management program not detected by the Schmitt technique. J Clin Epidemiol 2009;62:1165 72. 36. Oort FJ, Visser MRM, Sprangers MAG. An application of structural equation modeling to detect response shifts and t rue change in quality of life data from cancer patients undergoing invas ive surgery. Qual Life Res 2005; 14:599 609. 37. Visser M, Oort F, Sprangers M. Methods to Detect Response Shift in Quality of Life Data: A C onvergent Validity Study. Qual Life Res 2005;14: 629 39.

PAGE 132

132 38. Donaldson GW. Structural equation models for quality of life response shifts: Promises and pitfalls. Qual Life Res 2005; 14:2345 51. 39. Barclay Goddard R, Lix LM, Tate R, Weinberg L, Mayo NE. Response shift was identified over multiple occasions with a structural equation modeling framework. J Clin Epidemiol 2009;62:1181 88. 40. Howard GS, Ralph KM, Gulanick NA, Maxwell SE, Nance SW, Gerber SK. Internal invalidity in pretest posttest self report evaluations and a reevaluation of retrospective prete sts. Appl Psychol Meas 1979;3 :1 23. 41. Golembiewski R, Billingsley K, Yeager S. Measuring Change and Persistence in Human Affairs: Types of Change Generated by OD Desi gns. J Appl Behav Sci 1976;12 :133 57. 42. Rapkin BD, Schwartz CE. Toward a theoretical model of quality o f life appraisal: Implications of findings from studie s of response shift. Health Qual Life Outcomes 2004;2:14. 43. Breetvelt IS, van Dam FSAM. Underreporting by cancer patients: the case of response shift. Soc Sci Med 1991; 32: 981 87. 44. Groenvold M, Fayers PM, S prangers MAG, Bjorner JB, Klee MC, Aaronson NK, et al. Anxiety and depression in breast cancer patients at low risk of recurrence compared with the general population unexpected findings. J Clin Epidemiol 1999, 52 (6) : 523 30. 45. Padilla GV, Mishel MH, Grant M M. Uncertainty, appraisal and quality of life. Qual Life Res 1992; 1: 155 65. 46. Wilson IB, Cleary PD. Linking clinical variables with health related quality of life: a conceptual model of patient outcomes. JAMA 1995; 273: 59 65. 47. Sprangers MAG, V an Dam FSAM, Broe rsen J, Lodder L, Wever L, Oosterveld P, et al. Revealing response shift in longitudinal research on fatigue: The use of the then test approach. Acta Oncol 1999; 38: 709 18. 48. Jansen SJ, Stiggelbout AM, Nooij MA, Noordijk EM, Kievit J. Response shift in qualit y of life measurement in early stage breast cancer patients undergoing radiotherapy. Qual Life Res 2000; 9: 603 15. 49. Adang EMM, Kootstra G, Engel GL, van Hooff JP, Merckelback HLGJ. Do retrospective and prospective quality of life assessments differ for pancr eas kidney transplant recipients? Transpl Int 1998; 11: 11 5. 50. Schwartz CE, Coulthard Morris L, Cole B, Vollmer T. The quality of life effects of Interferon Beta 1b in multiple sclerosis: An Extended Q TWiST analysis. Arch Neurol 1997; 54: 1475 80.

PAGE 133

133 51. Hagedoorn M, Sneeuw KC, Aaronson NK. Changes in physical functioning and quality of life in patients with cancer: Response shift and relative evaluation of one's condition. J Clin Epidemiol 2002; 55: 176 83. 52. Schwartz CE, Feinberg RG, Jilinskaia E, Applegate JC. An evalu ation of a psychosocial intervention for survivors of childhood cancer: Paradoxical effects of response shift over time. Psychoo ncology 1999; 8: 344 54. 53. Rees JE, Waldron D, O'Boyle CA, MacDonagh RP. Response shift in individualized quality of life in patient s with advanced prostate cancer [abstract]. Clin T her 2002; 24 Suppl B : 33 4. 54. Howard GS. Response shift bias: A problem in evaluating interventions with pre/post self reports. Eval Rev 1980;4 (1):93 106. 55. Howard GS, Dailey PR. Response shift bias: A source of contamination of self report measures. J Appl P sychol 1979;64: 144 50. 56. Howard GS, Ralph KM, Gulanick NA, Maxwell SE, Nance DW, Gerber SK. Internal invalidity in pretest posttest self report evaluations and a reevaluation of retrospective pretests. Appl Psy chol Meas 1979;3: 1 23. 57. Howard GS, Schmeck RR, Bray JH. Internal invalidity in studies employing self report instruments: A suggested remedy. JEM 1979;16 :129 35. 58. Terborg JR, Howard GS, Maxwell SE. Evaluating planned organizational change: A method for asses sing alpha, beta, and gamma change. Acad Manage Rev 1980;5 :109 21. 59. King RB. Quality of life after s troke. Stroke 1996;27(9):1467 72. 60. engs TO, Yu M, Luistro E. Health related quality of life after stroke A comprehensive review. Stroke 2001;32(4):964 71. 61. B reetvelt IS, Van Dam FS. Underreporting by cancer patients: The case of response shi ft. Soc Sci Med 1991;32(9):981 87. 62. Decker SD, Schulz R. Correlates of life satisfaction and depression in middle aged and elderly spinal cord injured persons. A m J Occup Th er 1985 ;39(11):740 5. 63. World Health Organization [Internet]. International Classification of Functioning, Disability, and H ealth. [updated 2001; cited 2010 Oct 5]. Available from : http://books.google.com/books?id=pwb9ywSVKxwC&dq=ICF,+WHO&printsec=f rontcover&source=bn&hl=en&ei=2wD9SaTHNMWEtweZ05nFCg&sa=X&oi=boo k_result&ct=result&resnum=4

PAGE 134

134 64. Wang R, Zhao Y, He X, Ma X, Yan X, et al. Impact of hypertension on health related quality of life in a population based study in Shanghai, China. Public Health 2009;13(8):534 9. 65. Boini S, Briancon S, Guillemin F, Galan P, Hercberg S. Occurrence of coronary artery disease has an adverse impact on health related quality of life: A longitudinal controlled study. Int J Card iol 2006;113(2):215 22. 66. Testa MA, Anderson RB, Nackley JF, Hollenberg NK. Quality of life and antihypertensive therapy in men. A comparison of captopril and enalapril. The quality of life hypertension study group. N Engl J Med 1993;328(13):907 13. 67. Wald A, Scarpignato C, Kamm MA, Mueller Lissner S, Helfrich I, Schuijt C, et al. The burden of constipation on quality of life: results of a multina tional survey. Aliment Pharmacol Ther 2007;26:227 36. 68. Ried LD, Gandhi PK Predictors of health status change among patients treated with a calcium antagonist or an atenolol led hypertension strategy in the Internation Verapamil SR Trandolapril Study (INVES T) [abstract]. Value Health 2008; 11(6):A411. 69. Ahmed S, Mayo N, Corbiere M, Wood Dauphinee S, Hanley J, Cohen R. Using the Patient Generated Index to Evaluate Res ponse Shift Post Stroke. Qual Life Res 2005;14:2247 57. 70. Nolte S, Elsworth GR, Sinclair AJ, Osbor ne RH. Tests of measurement t demiol 2009;62:1173 80. 71. Osborne RH, Elsworth GR, Sprangers MAG, Oort FJ, Hopper JL. The value of the Hospital Anxiety and Depression Scale (HADS) for com paring women with early onset breast cancer with population based reference women. Qual Life Res 2004;13:191 206. 72. Joint National Committee on Prevention Detection Evaluation and Treatment of High Blood Pressure. The sixth report of the Joint National Commi ttee on prevention, detection, evaluation, and treatment of high blood pressure. Arch Intern Med 1997;157:2413 46. 73. Saris Baglama RN, Dewey CJ, Chisholm GB, Kosinski M, Bjorner JB, Ware JE. SF 36 Health Outcomes Scoring Software. 2004 : 1 157. 74. Ware JE, Snow K K, Kosinski M, Gandek B. SF 36 Health Survey: Manual and Interpretation Guide. Boston, MA: The Health Instit ute, New England Medical Center; 1993. 75. Brown TA. Confirmatory factor analysis for applied research. G uilford Press. New York: London, Guilford Press ; 2006.

PAGE 135

135 76. Jreskog K, Srbom D. LISREL 8.8: Structural Equation Modeling with the SIMPLIS Command Language Lincolnwood, Il: Scientific Software Internatonal Inc.; 1993. 77. Ludbrook J. Statistics in Physiology and Pharmacology: A Slow and Errat ic Learning Curve Clin Exp Pharmacol Physiol 2001;28:488 92. 78. MacCallum RC, Browne MW, Sugawara HM. Power analysis and determination of sample size for covariance structure modeling. Psychol Methods 1996;1(2):130 49 79. Hochberg Y, Benjamini Y. More Powerful Procedures for Mul tiple S ignificance Testing. Stat Med 1990;9:811 8. 80. Radloff LS. The CES D scale: a self report depression scale for research in the general population. Appl Psychol Meas 1977;1:385 401. 81. Sachdev M, Sun JL, Tsiatis AA, Nelson CL, Mark DB, Jollis JG. The progn ostic importance of comorbidity for mortality in patients with stable coronary artery disease. J Am Coll Cardiol 2004;l43:576 82. 82. Chirinos JA, Veerani A, Zambrano JP, Schob A, Perez G, Mendez AJ, et al. Evaluation of comorbidity scores to predict all cause mortality in patients with established co ronary artery disease. Int J Cardiol 2007;117:97 102. 83. Meyer Rosberg K, Burckhardt CS, Huizar K, Kvarnstrom A, Nordfors L, Kristofferson A A comparison of the SF 36 and Nottingham Health Profile in patients with ch ronic neuropathic pain. E ur J Pain 2001;5:391 403. 84. de Graff JC, Ubbink DTh, Kools EJC, Chamuleau SAJ, Jacobs MJHM. The impact of peripheral and coronary artery disease on health related quality of life. Ann Vasc Surg 2002;16:495 500. 85. Sevinc S, Akyol AD. C ardiac risk factors and quality of life in patients with coro nary artery disease. J Clin Nurs 2010;19:1315 25. 86. Bass C, Akhras F, Upward J, Keates J, Lowe D, Harry J, et al. Does atenolol improve physical and psychological function after coronary artery byp ass surgery: a controlled study. J Psychosom Res 1987;31:521 9. 87. Norris CM, Ghali WA, Galbraith PD, Graham MM, Jensen LA, Knudtson ML for the APPROACH investigators. Women with coronary artery disease report worse health related quality of life outcomes com pared to men. Health Qual Life Outcomes 2004;2:21. 88. standing illness. Soc Sci Med 1999;48:89 98.

PAGE 136

136 89. McCreary DR, N ewcomb MD, Sadava SW. Dimensions of the male gender role: A confirmatory analysis in men and women. Sex Roles 1998;39:81 95. 90. Schwartz CE Rapkin BD Reconsidering the psychometrics of quality of life assessment in light of response shift and appraisal. Hea lth Qual Life Outcomes 2004;2(1):16. 91. Rosenberg MJ. The Logic of Survey Analysis. New York:Basic Books; 1968. 92. Lauridsen HH, Hartvigsen J, Manniche C, Korsholm L, Grunnet Nilsson N. Responsiveness and minimal clinically important difference for pain and disa bility instruments in low back pain patients. BM C Musculoskelet Disord 2006;7:83 93. Myint PK, Surtees PG, Wainwright NWJ, Luben RN, Welch AA, Bingham SA, et al. Physical health related quality of life predicts stroke in the EPIC Norfolk. Neurology 2007;69(24 ):2243 8. 94. Wyrwich KW, Spertus JA, Kroenke K. Tierney WM, Babu AN, Wolinsky FD. Clinically important differences in health status for patients with heart disease: An expert consensus panel report. Am Heart J 2004;147(4):615 22. 95. Mayo NE, Scott SC, Ahmed S. C ase management poststroke did not induce response shift: the value of residuals. J Clin Epi demiol 2009;62:1148 56. 96. Mayo NE, Scott SC, Dendukuri N, Ahmed S, Wood Dauphinee S. Identifying response shift statistically at the individual level. Qual Life Res 20 08;17:627 39. 97. Brossart D, Clay D, Willson V. Methodological and statistical considerations for threats to internal validity in pediatric outcome data: response shift in se lf report outcomes. J Pediatr P s ychol 2002;27:97 107. 98. Bloem EF, van Zuuren FJ, Koenem an MA, Rapkin BD, Visser MRM, Koning CCE, et al. Clarifying quality of life assessment: do theoretical models capture the underlying cognitive processes? Qual Life Res 2008;17:1093 1102. 99. Lenert LA, Treadwell JR, Schwartz CE. Associations between health sta tus and utilities implications for policy. Med Care 1999;37:479 489. 100. Gibbon FX. Social comparison as a mediator of response shift. Soc Sci Med 1999;48:1517 30. 101. Mayo Clinic.com [Internet]. Mayo Foundation for Medical Education and Research. [updated 2010 Ju ly 2; cited 2010 Oct 5]. Available from: http://www.mayoclinic.com/health/coronary artery disease/DS00064/DSECTION=lifestyle and home remed ies

PAGE 137

137 BIOGRAPHIC AL SKETCH Pranav Kirit Gandhi was born in 1980 at Mumbai, India to Kirit and Kalpana Gandhi. He g rew up there with his parents. His younger sister was born in 1983. Pranav did all his schooling in Mumbai. He attended the Mumbai Educational Trust Institute of Pharmacy, Mumbai ciences in 2002. He then attended graduate school at University of the Science in P hiladelphia and dministration in 2006. He then attended graduate school at the University of Florida (UF), Gainesville Florida and earned a doctorate in pharmaceutical sciences pharmaceutical outcomes and p olicy in 2010. While at UF, he got married to his love interest, Poonam Sanjanwala in 2009.