Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
A new paradigm to evaluate quality-adjusted life years (QALY) from secondary database: Transforming health status instrument scores to health preference
(USC Thesis Other)
A new paradigm to evaluate quality-adjusted life years (QALY) from secondary database: Transforming health status instrument scores to health preference
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
INFORMATION TO USERS This manuscript has been reproduced from the microfilm master. UM I films the text directly from the original or copy submitted. Thus, some thesis and dissertation copies are in typewriter face, while others may be from any type of computer printer. The quality of this reproduction is dependent upon the quality of the copy submitted. Broken or indistinct print, colored or poor quality illustrations and photographs, print bleedthrough, substandard margins, and improper alignment can adversely affect reproduction. In the unlikely event that the author did not send U M I a complete manuscript and there are missing pages, these w ill be noted. Also, if unauthorized copyright material had to be removed, a note w ill indicate the deletion. Oversize materials (e.g., maps, drawings, charts) are reproduced by sectioning the original, beginning at the upper left-hand comer and continuing from left to right in equal sections with small overlaps. Photographs included in the original manuscript have been reproduced xerographically in this copy. Higher quality 6" x 9* black and white photographic prints are available for any photographs or illustrations appearing in this copy for an additional charge. Contact UMI directly to order. ProQuest Information and Learning 300 North Zeeb Road. Ann Arbor, M l 48106-1346 USA 800-521-0600 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. A NEW PARADIGM TO EVALUATE QUALITY ADJUSTED LIFE YEARS (QALY) FROM SECONDARY DATABASE: TRANSFORMING HEALTH STATUS INSTRUMENT SCORES TO HEALTH PREFERENCE by Nishan Sengupta A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (PHARMACEUTICAL ECONOMICS AND POLICY) December 2000 Nishan Sengupta Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. UMI Number: 3041522 _ _ _® UMI UM I Microform 3041522 Copyright 2002 by ProQuest Information and Learning Company. A ll rights reserved. This microform edition is protected against unauthorized copying under Title 17, United States Code. ProQuest Information and Learning Company 300 North Zeeb Road P.O. Box 1346 Ann Arbor, Ml 48106-1346 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. UNIVERSITY OF SOUTHERN CALIFORNIA THE GRADUATE SCHOOL UNIVERSITY PARK LOS ANGELES. CALIFORNIA 90007 This dissertation, written by ..j/.lfJ.A .O ...................................................... under the direction of Dissertation Committee, and approved by all its members, has been presented to and accepted by The Graduate School in partial fulfillment of re quirements for the degree of DOCTOR OF PHILOSOPHY Dean of Graduate Studies Date DISSERTATION COMMITTEE Chairperson . / - X - y Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. ACKNOWLEDGEMENTS Dissertations need sagacious thinking and solitary effort, but this research could not have been finished without considerable guidance and assistance. My debts to these mentors and organizations are substantial. I sincerely hope that this acknowledgement will provide some small measure of my appreciation. First, thanks are due to University o f Southern California (USC) and the Department of Pharmaceutical Economics and Policy for the research support provided during all phases of this study. Thanks are also due to Kaiser Permanente, Southern California for the research grant. Through their funding initial survey database was generated. In addition, I would like to thank US Pharmacopeia, for funding fellowship grant during my final year. Second, I would like to thank my committee chair Dr. Michael B. Nichol, who has provided guidance and support throughout my academic career at USC. He prompted me to generate the research idea and also rekindled my interest in this particular topic. I tested his patience numerous times during the research and I never found it wanting. I am indebted to him in ways that can never be adequately repaid, and for that I am truly grateful. My greatest gratitude is to my parents, Ela Sengupta and Nirmalendu Bikash Sengupta. Both of them provided continual encouragement even while they wondered when this research project would reach completion. I definitely could not have embarked on this effort without their strong mental support. Their support after I made the decision to pursue on the doctoral ii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. degree was invaluable. My dissertation committee stimulated new approaches to the problems presented in the research and persuaded me to consider several alternative techniques on the topic. Thanks are due to all the committee members. In particular, Dr. Denise Globe extended her support by providing valuable suggestions on several quality of life and health utility issues. Professor Susan Groshen pointed out several robust statistical approaches to validate the hypothesis. Dr. Gerald Borok provided his precious experiences and input to interpret the results in clinical and managed care setting. Dr. Jeff McCombs and Dr. Joel Hay also provided their valuable time with patience. Both o f them were very open to meeting to discuss interesting issues related to the survey database, utilized in this dissertation. Despite the busy schedule all the eminent faculty members always put up with my frequent requests with customary indulgence and high spirit. Finally, I would like to thank all my friends who provided general encouragement. Specifically, colleagues and senior researchers in the department of Pharmaceutical Economics and Policy always provided assistance in attempting to run any o f the computational program. Some of them also reviewed part o f the study in its early stages and enlivened me with discussions of various quality of life issues and measurement techniques. Ail these suggestions helped me to steer me in improving research. While many people have generously contributed to this research, I alone, am responsible for all remaining errors. iii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table of Contents Acknowledgement List of Tables Abstract Chapter One—Introduction Measurement of Health State Preference Using Non Preference-Based Measure in Economic Evaluation Significance of this Study Organization of the Dissertation Chapter Two—Literature Review Development of Health State Preference Instruments Development of Non Preference-Based Health Status Measurement Instruments Developing a Standardized Algorithm to Estimate QALY from the SF-36 Developing Preference-Based Index from the SF-12 Conclusion Chapter Three—Methods Population Assessed Statistical Models Model Specification Test, Validation and Extension Chapter Four—Results Sociodemographic Assessment of Reliability Regression Model: Estimating HUI from the SF-36 Regression Model: Estimating HUI from the SF-12 Regression Model: Estimating LAS from the SF-12 and SF-36 Page ii-iii vi vii-viii 1 20 34 46 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table of Contents (Continued) Chapter Five—Conclusions 70 Contribution to the Mapping Technique Implications for Economic Evaluation from Clinical Research Limitations and Areas for Future Investigation Bibliography 78 V Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. List of Tables Page 1. Table 1. Descriptive Statistics o f Demographics 48 2. Table 2. Comparison of Health Status (SF-36) of USC/KP Study 48 Sample at Endpoint (t2> with the US Population 3. Table 3. Health Utility (HUI, LAS) and Health Status (SF-36) 49 4. Table 4a. Correlation Matrix: HUI-II, SF-36, CDS 51 5. Table 4b. Correlation Matrix: HUI-III, SF-36 at Study Endpoint (t2) 52 6. Table 5a. Estimated Effect of SF-36 Domains on HUI-II 55 7. Table 5b. Estimated Effect o f SF-36 Domains on HUI-III 56 8. Table 6. Results of Wu Test for Detecting Endogeneity in SF-36 56 9. Table 7. Test of Serial Correlation in 1 st Stage Equations 57 from Lagged SF-36 Domains 10. Table 8. Linear Regression Analysis of the Relationship between 61 Utilities (LAS and HUI) and Responses of the General Health Item o f the SF-12 11. Table 9a. Linear Regression Analysis o f the Relationship between 61 -63 LAS and 11 Items o f the SF-12 12. Table 9b. Linear Regression Analysis of the Relationship between 64-66 HUI-UI and 11 Items o f the SF-12 13. Table 10. Estimated Effect o f SF-36 Domains on LAS from 68 Cross-sectional Time Series Model 14. Table 11. Sample Size and 95% Confidence Intervals of Prediction 69 (HUI from SF-36) vi Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. ABSTRACT Assessing Health Related Quality of Life (HRQOL) through generic health status instruments has gained considerable attention in today’s clinical practices. Nevertheless, one of the drawbacks o f the generic HRQOL instruments (SF-36, SF-12 etc.) is that these are unsuitable for direct estimation o f health preference that can then be used to generate Quality Adjusted Life Years (QALY). QALY is a useful outcome measure that reflects both life expectancy and quality of life and is suitable for cost-effectiveness or cost-utility analyses. Objective: To estimate a summary health utility index (represented by the Health Utility Index (HUI) and Linear Analog Scale (LAS)) from the domains of the SF-36, and items of the SF-12. Methods: Data were collected from a longitudinal survey (three administrations) of a large cross-section of patients under managed care. The SF-36 was used to assess health status and the LAS and HUI were used to assess health state preference. The SF-12 items were also created from the SF-36 by using a published algorithm. The SF-36 and the LAS were administered at baseline, year 1 and year 2, but the HUI was administered only in the last cycle. A linear regression model was used to predict HUI at year 2 from the SF-36 domains at year 2 along with socioeconomic and disease covariates. The predictive power o f the SF-12 to estimate the HUI was tested using similar methods. A cross- sectional time series model was developed to estimate LAS from the SF-36 domains. Results: The SF-36 domains along with age and gender significantly vii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. accounted for 33% to 50% variations in HUI score depending on the model used. Additionally, the SF-36 domains were accounted for 47% variation in LAS scale score. The SF-12 items explained 46% variation in HUI score. Conclusion: This research provides substantial evidence that health state utility and QALYs can be estimated from secondary databases, which include the SF-36 or the SF-12. VUI Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. INTRODUCTION The focus of this dissertation is to develop a robust methodology to generate a mapping algorithm between preference-based health index and non preference-based health status measurements (HSMs) from a secondary database. Tne preference-based health status measurements are the basic components for evaluating Quality Adjusted Life Years (QALY), which is also helpful for economic evaluation of health care. However, unlike non preference-based generic HSM instruments (SF-36, SF-12 etc.) preference-based techniques are rarely used in most of the clinical studies. Thus, it is difficult to generate any QALY from these studies. To overcome the drawback o f the non preference-based HSM several mapping algorithms are developed in this research. Two different types of HSM instruments are utilized: non preference-based HSM (SF-36) and preference-based HSM (Health Utility Index, Linear Analog Scale). The purpose o f the study is to transform the SF-36 to one of the preference-based HSMs, which is able to generate QALY. Simultaneously, the study also extends the feasibility of the mapping technique and generates preference-based index from a shorter form of the SF-36 (SF-12). The rest of this chapter introduces a brief history o f the QALY concept as well as the need for it in economic evaluation. It elicits the differences within the concept of health state preference (utility vs. value) and describes the distinction I Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. between different HSM (preference and non preference-based) instruments. Limitations o f these instruments are discussed in depth. To overcome the limitations, this research highlighted the necessity of converting generic non preference-based instruments to preference-based HSM. Then the different techniques to integrate preference and non preference-based measurement are evaluated. Finally the significance o f this study along with the research questions and organization of the dissertation is detailed. QALYs are now widely used in medical decision making and health economics as an useful outcome measure that reflects both life expectancy and quality o f life (Bult, Hunink, Tsevat, Weinstein, 1998). Research conducted over the last twenty-five years has concluded that under certain set of assumptions (additive separability of individual utility function, constant rate of time preference, risk neutrality over the time period of life), QALYs are likely to offer a close approximation to a summary measure of health and patient preference in any cost- effectiveness analysis (Gold, Siegel, Russell, Weinstein, 1996). However, the development of a measure is merely the first step in determining its value. Establishing the ability of the measure to yield the same score over repeated administration (reliability) and the degree to which it measures the concepts it purports to measure (validity) is imperative in order to incorporate the QALY concept into pharmacoeconomic and clinical research. 2 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. To establish the validity o f QALY, Garber and Phelps (1997) discussed that if QALY adequately represents individual preference, then using it in any cost-effectiveness analysis as an effectiveness parameter is imperative to potential Pareto-improvement in resource allocation. Under Pareto-improvement (also called Kaldor-Hicks criterion) allocation individuals will optimally set priorities for health care expenditures by selecting those with cost/QALY ratios less than some threshold. This optimal cost/QALY threshold will be a function of an individual’s income and risk aversion parameter. Setting up this criterion will also be helpful to determine optimal coverage for an actuarially fair health insurance policy under perfect information (Garber and Phelps, 1997). However, the validity of QALY is restrictive. If QALYs do not represent patient preference or health state utility adequately, the usual cost-effectiveness approach (cost/QALY) can not produce a reliable guideline to Pareto-improvement of societal resource allocation. Unfortunately any of the three assumptions mentioned earlier are too restrictive in the real world. Thus it is not hard to think that QALYs are unlikely to serve as a comprehensive utility measurement. For example, individual may place a higher weight on surviving to a particular event (a non-constant rate of time preference) than other possible events. Considering this problem, several alternatives to QALYs are suggested. Among those alternatives the Healthy Year Equivalent (HYE) (Meherez and Gafhi, 1993; 3 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Gafhi and Birch 1995) and the Saved Young Life Equivalent (SAVE) (Nord, 1992a) are noteworthy. HYEs are calculated in two stages. First by measuring preference over complete life paths and then obtaining utility for each possible health path o f changing health states by the standard gamble method. Because of the two stage utility estimations, HYE is more robust than QALY. On the other hand, measurement strategy in SAVE approach is constructed to yield social values for health gains for individuals. The unit o f value is determined by the health gains afforded by the program (rather than health state of the individual) under study. The comparable standard measure is defined as the saving and restoring of full health of one young life. SAVE is distinct from other approaches (QALY or HYE) in a way that it calculates societal benefit rather than evaluating individual benefit. Despite the theoretical validity, both of these potential alternatives o f QALY are computationally difficult and in their early stages of development. Critics of HYE have raised concerns that the two-stage standard gamble method in HYE is unnecessarily complicated and there is no empirical evidence available to justify its usefulness over one stage standard gamble approach used in QALY (Gold, 1996). Because o f this conventional QALY still remains a dominant approach in any cost- effectiveness analysis (Gold, 1996). Nevertheless, the fundamentals o f QALY approach need proper understanding, including its ethical underpinnings. One such problem arises when 4 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. ‘utility maximization’ is used synonymously with that o f ‘QALY maximization’. To construe QALY just a mere measure o f utility is misleading. For example, it is possible that two individuals might enjoy the same health related quality of life and yet for a variety o f reasons (employment status, marital status, etc.) one might be happier or more fulfilled than the other (Mckie, Kuhse, Richardson and Singer, 1996). Thus, there is more to maximizing utility (in terms of human happiness or human fulfillment) than to QALYs. The scope o f QALYs is much more limited than ‘Utility’ towards health care resource allocation. Last but not least it is noteworthy to mention that the principle o f QALY-maximization is “egalitarian” within the health domains. It does not discriminate on the basis of wealth, social standing, race or anything irrelevant to health related quality of life (Mckie, Kuhse, Richardson and Singer, 1996). This egalitarian concept should not be thought as a commitment to the equal distribution of health care resources. According to this concept equal weight is given to the QALYs of all those potentially affected by an allocation decision. In cost-effectiveness analyses, QALYs are commonly calculated as the average number of additional years o f life gained from any intervention multiplied by population or individual preferences for obtaining this additional survival. Preference functions for alternative health outcomes may be ordinal or cardinal (Mas-Collel, Whinston and Green, 1995). However, cardinal preferences are used 5 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. as quality weights for QALYs, because of the interval scale property, which is appropriate for comparisons among different health outcomes. Measurement of Health State Preference Value and Utility are the two preference-based approaches used extensively to measure health state preference. Both approaches are founded on well- established, profound theory and measurement methodology. “Utilities” are cardinal preferences under uncertainty, whereas “Values” are preferences under certainty. Utility measurements are founded on von Neuman-Morgenstem (VN-M) (Von Neumann and Morgenstem, 1947) expected utility theory, which described preference maximization problem o f individual under uncertainty. In contrast “Values” are cardinal preferences, based on value theory and measured under certainty. The three major types o f instruments used to measure the preferences are the linear analog scale (LAS), the time trade-off (TTO) and the standard gamble (SG). The LAS also referred to in the literature as the visual analog scale (VAS) or the rating scale (RS) is simply a line, usually with well-defined endpoints, on which respondents can indicate their preferences. The most preferred health state is placed at one end o f the line and the least preferred at the other end. The remaining health states are placed on the line between these two, in order o f their preference. The interval or spacing between the placements corresponds to the difference in 6 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. preference as perceived by the subject (Torrance, 1986). It has been used in several instruments (EuroQol 5D (EQ-5D), Quality o f Well-being (QWB), Health Utility Index (HUI)) to measure health state preference and considered most feasible technique that demonstrates high response rate and high level of completion (Froberg and Kane, 1989; Kaplan et al., 1982; Gudex et al., 1996). However, the LAS measure does not present a choice between different health states to the responders. The lack of choice violates the axiom of VN-M expected utility theory and it also severely impairs LAS’s capability to measure preference on a cardinal scale. Because of this, critics raised their concern over the use of LAS in economic or decision theory (Richardson, 1994; Nord, 1991). The TTO technique has been customized as an alternative to SG to value health states by Torrance (1976) and used to develop EQ-5D and QWB preference measurement instruments. This technique offered individual patient two alternatives. The alternative 1 is a health state i for time t (life expectancy of an individual with the chronic condition) followed by death. The alternative 2 is another health state for time .t<r followed by death. Time x is varied until the respondent is indifferent between the two alternatives. At that point the respondent’s health state preference value is A , = x/t (Torrance, 1986). The TTO technique has not been related to any existing behavioral theory. Mehrez and Gafhi (1991) discussed the TTO in the context o f value function theory, because of the 7 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. identification of different trade-off combination of health and duration. Buckingham et al. (1996) compared the TTO method with the welfare economics approach of compensating variation, where welfare gain is measured by compensating loss of something else that is valuable so that the respondent is returned to their original level o f welfare. Nevertheless, applicability o f the TTO in preference measurement is questionable because the technique asks respondents to make choice between two certain health outcomes, when future health is uncertain (Mehrez and Gafhi, 1991). Critics also raised concerns on the assumption in the TTO method that individuals are prepared to trade off a constant proportion o f their remaining life years to improve their health status irrespective of the number o f years remain. However, it seems plausible that the valuation of a health state is influenced by the time an individual spends in that state (Sackett and Torrance, 1981). The SG approach is the only technique based on expected utility theory and the most widely used model of decision making under uncertainty. The SG offers individual patient two alternatives. The alternative 1 is a treatment with two possible outcomes: either the patient is returned to normal health and lives for an additional t year (probability P), or the patient dies immediately (probability l-P). The other alternative is no treatment with a certain outcome of chronic state i for life (t years). Probability is varied until the respondent is indifferent between the 8 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. two alternatives, at which point the required preference score for state i is P (i.e., hi -P ) (Torrance, 1986). Because of its theoretical underpinnings, SG is viewed as the classic method o f medical decision making and often referred as the gold standard for measuring health utility (Torrance, 1996; Gafhi, 1994). The preference-based instruments have been available for more than two decades yet they are not widely used in the health outcomes research. Drummond and Davies (1991) have identified three explanations for the reluctance to use preference-based technique among clinical decision-makers. One major concern is financial and administrative burdens of using these instruments in clinical trials. A second concern is that these preference elicitation techniques (SG and TTO) may be distressing to patients. It may also result in the dropout of patients from the trials and is questionable ethically. Final concern is preference-based measures suffer from being insensitive or even irrelevant for many health conditions. On the other hand, non preference-based health status measurement (HSM) instruments (will be described in detail in the next section) are far more widely used in clinical trials and health services research. Health Status Measurement Health o f an individual at a particular point of time is also represented as health status or health state. Health status is a multidimensional concept that measures impairments, perceptions, and social functions, caused by disease, injury, 9 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. and treatment from the individual’s perspective (Patrick and Erickson, 1993). Descriptive health status instruments and preference based instruments are the two main approaches used in assessing a patient’s Health Related Quality of Life (HRQOL). HRQOL is a way to assess the different dimensions or domains of health in terms of its relative desirability from an individual’s perspective at a particular point in time. Health status instruments describe functional status, well being, and other domains of health, whereas health utility and value measures capture a patient’s cardinal preferences for a health state in a single measure as mentioned in the earlier section (Torrance, 1986; Patrick and Erickson, 1993). The non preference-based descriptive health status measurement instruments (HSM) are now increasingly used in clinical trials to assess the efficacy and effectiveness of health care interventions concerning patient perceived health. These HSM instruments can be ‘generic’ and hence designed for use across all conditions or specifically designed for a particular disease (Brazier, Deverill, Green, Harper and Booth, 1999). Spilker et al. (1990) identify almost 200 different HSM instruments that vary widely in terms of content format and scaling. One of the widely accepted generic HSM instruments is the SF-36. It is a 36-item short form questionnaire to measure eight major health concepts; physical functioning; role limitation because of physical health problems; bodily pain; social functioning; general mental health (psychological distress and psychological well-being); role 10 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. limitations because of emotional problems; vitality (energy/fatigue); and general health perceptions. This questionnaire also generates summary physical and mental health measures from the eight health concepts described above (Ware, 1996; Ware and Sherboume, 1992). However, despite being widely used in the clinical trials and population survey HSM instrument has several limitations. One of the basic limitations is that it does not include mortality or survival (Feeny et al., 1989) and for economic evaluation as well as clinical decision making it is often necessary to combine survival with health status. Other major limitations of HSM are that it does not explicitly incorporate preferences (Williams, 1989; Johannesson et al., 1996). Secondly HSM instruments do not possess the interval scale properties required to undertake economic comparison (cost-effectiveness) across different treatment options. For example, patients may differ dramatically in how they view HRQOL associated with different health outcomes derived from the HSM instruments. Patients may also put different numerical weights across different questions in HSM instruments and there is no way of simple summing up numerical weights across questions on a single index (0 to 1 scale). Hunt et al. (1986) have also supported this by arguing that any simple addition of affirmative responses in HSM instruments gives misleading result because features of pain, social life, emotion and so on are qualitatively distinct which can not have a common denominators. 1 1 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Because of this, purpose of HSM instruments is restricted to derive a measure of health and not to derive QALY useful for economic evaluation. Using Non Preference-Based Health Measures in Economic Evaluation As mentioned earlier HSM are being widely used in clinical trial and studies are being published which detailed costs alongside quality of life measures (Brazier, Deverill, Green, Harper and Booth, 1999). The empirical evidence (which will be described in detail in next chapter) also suggests that HSMs are significantly correlated to preference-based measures. This indicates that HSM instruments should at least rank states in the same order as preference-based measures, provided there is no trade-off to be made between health dimensions. Nevertheless, this makes HSM instruments suitable for economic evaluation in only two clinical scenarios. First, a cost minimization analysis can be performed when HSM scores are same across all dimensions. In the second case, a cost comparison can be done against HSM scores of two alternatives, when the scores are better in at least one dimension and no worse on any other. However, HSM scores can be made useful for any economic analysis, if it is possible to generate a single preference-based summary index from HSM scores. Several studies were performed in past years to utilize potentially rich descriptive health status information from HSM instruments in economic evaluation. There are five methods for doing this: use arbitrary weights, map from 12 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. an HSM on to the classification o f a preference-based instrument, develop exchange rates between HSM scales and preference-based measures, value the items o f the HSM using preference-based methods, or use the descriptive data to derive health scenarios for valuation (Brazier, Deverill, Green, Harper and Booth, 1999). Arbitrary weights can be used to weight the different dimensions of HSM and generate an aggregate index of health status. For example, researchers aggregated the Nottingham Health Profile (Hunt, McKenna, McEwen et al., 1981) into a single index to estimate the QALYs gained from a heart transplant program (O’Brien et al., 1987). Despite its simplicity, this method cannot be used for economic evaluation because weights are not generated from people’s preference and there is no allowance for any possible interaction between the dimensions. Mapping non preference-based HSM questionnaire into the classification of one of the multi-attribute utility (MAU) instruments (HUI or QWB) is another possible way. The crucial difference between the MAUs and other measures of descriptive HSM questionnaires is that the weights used to score MAUs have been obtained using one o f the preference elicitation techniques (LAS, TTO or SG) described earlier. The MAUs produce a health preference index between 0 and 1 (0=dead and Imperfect health) and can be used in economic evaluation alongside clinical studies to value the benefit of health care in terms of QALYs. The validity 13 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. o f the mapping procedure could also be tested by administering the HSM and the MAU instrument on the same patients and examining the extent to which the procedure was able to correctly predict the patient position on the MAUs classification. Gudex (1986) most probably the first researcher mapped a group of scores from the Ruesch Social Disability Rating Scale (RDR) on to the Rosser matrix in a sample of patients receiving maintenance haemodialysis. Since, then several empirical studies have been done on this issue. However, information on the predictive accuracy of this procedure is limited because most of the studies lack external validation in separate sample. All o f these studies will be discussed in detail in the next chapter. There has been very little research done on estimating exchange rate between HSM and preference based instruments. Caims et al. (1991) attempted to explain the change in LAS score with that o f HSM, but a simple proportional relationship between condition specific HSM and LAS could not be established. Given the theoretical reasons also it is very unlikely that change in the HSM scores can be well explained by change in preference-based score. A more onerous approach is completely revaluing the content o f HSM using a preference elicitation technique. There are different strategies available for performing such a task (Froberg and Kane, 1989). One is the holistic approach where all health states defined by the HSM are valued directly. However, this is 14 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. not a feasible task for most of the HSMs which, measure several dimensions of health. For example, the 35 multilevel items of the SF-36 must all be valued and define a total of 2592* 1019 unique health states (calculated as the product o f the number of items in each dimension to the power of the number o f levels of that dimension) (Brazier, Deverill, Green, Harper and Booth, 1999). Thus, generating a summary index by using standard health state preference weighting (SG, TTO, MAU) from all these unique health states is not feasible (Hays and Sherboume, 1993). To substantially simplify the task, researchers used a part of the content of an HSM (using 14 of the 36 items o f the SF-36) and weighted those items with response choices (Brazier et al., 1998). This resulted in a six-dimension HSM instrument, which can generate a preference-based summary index. Nevertheless, the disadvantage o f this technique is that a substantial proportion of information about health status is lost and its implication on the sensitivity of the new instrument is not known yet. The last of these approaches is to construct special health state vignettes or whole health scenarios for specific treatment outcomes by using descriptive HSM data. This technique can also incorporate a time profile to elicit HYE. Despite its appealing aspect, there are almost no proper methodologies available for constructing the vignettes or scenarios from the large volume of descriptive HSM data. Vignettes have usually been constructed by informal methods using expert 15 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. opinion, or on qualitative interview o f patients, rather than the evidence produced in a large clinical trial (Cook et al., 1994) and it has limited reliability and validity. It is evident from the earlier discussion that developing a preference-based index from non preference-based HSM is tremendously beneficial. This can generate QALY from the HSM instruments and make them suitable for cost- effectiveness analysis. However, the scope for this development is limited. Crude aggregation is invalid, and the derivation o f exchange rate between HSM scales and preference-based techniques does not look promising since there does not appear to be any simple proportional relationship. Nevertheless, the two promising avenues to progress on this issue seem to be mapping non-preference based HSM to MAUs and revalue HSM using preference elicitation technique. The advantage of the mapping technique over revaluation is that it is robust to any kind of HSM instruments and there is almost no loss o f health status information. At the same time, feasibility of mapping approach is established by several empirical studies. Considering the theoretical validity and empirical support the present research uses this approach to evaluate a preference-based algorithm from the secondary database using the SF-36 and SF-I2. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Significance of this Study At the present time no published study has explored the mapping relationship between multi attribute health utility (HUI), linear analog scale (LAS) and health status measurement (SF-36 or SF-12) in a large US population. In this context, developing some mapping approaches between the SF-36 or the SF-12 and the Health Utility Index (HUI) as well as the SF-36 (SF-12) and the LAS will be helpful to generalize the predictive ability o f the SF-36 or the SF-12. This research is performed on a longitudinal database o f a large cross-section of managed care patients. It includes non preference-based HSM (SF-36, SF-12) and preference- based HSM (LAS, HUI) to generate some mapping algorithms between the different HSMs. The resulted mapping algorithm will provide health service researchers a robust methodology to estimate health utility as well as QALY from any type of retrospective study using the SF-36 or the SF-12. The result will also contribute significant addition to the available literatures on mapping techniques in this field. The specific research issues addressed in this dissertation are: 1. How can the preference-based HRQOL (Health Utility Index, Linear Analog Scale) be estimated from non preference-based generic HRQOL instrument like the SF-36 in a large US population? 2. Do the predictive power of the estimation models increase if adjusted for age, gender, other socio-demographic covariates and co-morbidities? 17 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 3. Is it possible to create a generalizable and robust algorithm to estimate health state utility from secondary databases using the SF-36? 4. How to develop an algorithm to generate preference-based index from the SF-12? Organization of the Dissertation: This study is organized into five chapters. Three major preference elicitation techniques were briefly discussed in the introductory chapter. The next or second chapter starts with the literatures on development of health state preference instruments. This chapter reviews the literature on the theories behind developing multi-attribute health utility instruments (HUI, QWB and EQ-5D), generic health status instrument (SF-36) and mapping methodology (from health status to health utility). Considerable attention is given to the other relevant transformation methodology (revaluing technique), although limited evidence is available regarding the empirical validity o f the revaluing technique. Chapter 3 details the methodology of model development for transforming the two health status instruments to health utility scores. This chapter includes a detailed description of all the independent variables, dependent variables, covariates and their data sources. Four steps are used to describe statistical models: First step is the assessment of reliability o f all health survey instruments used in the study. In the second step, 1 8 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. development o f predictive models and estimation algorithms (HUI from the SF-36, LAS from the SF-36, HUI from the SF-12, LAS from the SF-12) are elicited. The third step describes model specification, validation and extensions. The final step produces the predictive model and estimation algorithms to generate a preference score from the SF-36 or the SF-12. The fourth chapter describes the results obtained from all the models. This chapter also provides analysis of statistical tests and explanation for the outcomes. Validity as well as reliability of the models are also discussed in this chapter. Chapter 5 concludes with a discussion of the implications of the findings and the limitations. A delineation o f future research related to this topic is also supplemented this chapter. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. LITERATURE REVIEW Development of Health State Preference Instruments The three preference measurement techniques (SG, LAS, TTO) are also used to develop MAU theory based preference measurement instruments (HUI, QWB, EQ-5D). As noted previously, the SG is the preferred technique to elicit preference under uncertainty. Because of this, first we will review the development of HUI in detail, which was the only MAU instrument developed by using the SG technique (Torrance et al., 1982). Later on this section development of other MAU instruments (QWB, EQ-5D) will also be discussed. HUI (Mark II and III) are the two revised recent version o f the HUI series (Feeny et al., 1995; Torrance et al., 1996). The utility score derived from the HUI is anchored by ‘Dead=0’ and ‘Perfect H ealth=l\ The respondent completes a standard questionnaire constructed specifically for the HUI-II or the HUI-III. A multi-linear scoring algorithm applies utility weights to the respondent’s health status as reported in the standardized questionnaire. However, the construction of the HUI-n and the HUI-III is different. The HUI-II measures seven attributes of health status: sensation, mobility, emotion, cognition, self-care, pain and fertility and defines 24,000 health states. Each o f these attributes (except fertility, which is optional in the instrument) has three to five different preference levels. The HUI-III is an adaptation of the HUI-II and also based on MAU model. However, in this 20 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. instrument the number of attributes has been increased to eight. It included vision and hearing as separate attributes, along with speech, ambulation, dexterity, emotion, cognition, and pain while fertility was removed. The number of levels within each attribute has been increased to between five and six, and it is able to represent 972,000 unique health state. The preference scores for different levels within the attribute were generated from an interview of parents of childhood cancer patients and parents of similarly aged school children. The sample (n=352) was drawn from the Canadian population residing in Hamilton, Ontario. To generate preference scores, four SG and eight LAS instruments were used and later LAS responses were transformed into SG scores using a power transformation (Torrance et al., 1996). Preference or utility weights for attributes in the HUI-III were generated using both the SG and the LAS instruments from a randomized sample of adults (n=503) in the general population of Hamilton, Canada (Feeny et al., 1998). There has only been one test-retest reliability involving HUI-III in a general population survey (Boyle et al., 1995). Individual responses were found to be consistent between test for six attributes with the exception of speech and dexterity. The retest reliability (12-49 days apart) o f the overall HUI-III index score was found to be 0.77, well within accepted standards. Both the HUI-II and HUI-III are developed as a generic measure of health. However, some of the contents o f the 21 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. HUI-II are quite explicitly aimed at children (e.g. ‘ability to see, heard and speak normally for age’, ‘leams and remembers school work normally for age’). The developers of these instruments also argued for a ‘within skin’ definition o f health in HUI, which is only concerned with impairment and disability but not handicap. Thus, social and role activities were considered as a consequence o f people’s preference as well as over all choice set and being excluded from the questionnaire. Although, this concept is controversial because some attributes o f HUI-II (mobility, self-care, sensation, cognition) are not entirely ‘within skin’ and likely to be influenced by individual’s social setting (Brazier, Deverill, Green, Harper and Booth, 1999). Nevertheless, most of the attributes of HUI-II are reasonably succinct with the exception of emotion (Kanabar et al., 1995) and researchers suggested some simplification of that attribute. Regarding all of the concerns a number o f revisions were made to enhance the content and face validity o f the HUI-II. This development is reflected in the HUI-III, which is more relevant for adult population. The replacement o f ‘self-care’ attribute by ‘dexterity’ has improved its independence from the other attributes. At the same time creating visions, hearing and speech into separate dimensions make the HUI-III more comprehensive. As mentioned earlier, HUI-HI is able to generate a much larger but not overlapping health state classification. Construct validity of the HUI was performed in few researches. Barr et al. (1993) found leukemia patients (n=50) 22 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. with high risk had a lower HUI score than that o f low risk patients. Differences in the HUI scores have also been found in patients with childhood brain tumor (n=156) between those being treated and those no longer on treatment (Feenye/a/., 1993). The QWB, formerly the Index of Well-Being, is the oldest of the MAU instruments (Brazier et al., 1999). It consists of two main components: three multilevel dimensions relating to function (mobility, physical activity and social activity) and a list o f 27 symptom and problem complexes (e.g. general tiredness, weakness or weight loss). The functional dimensions and the symptoms combine to form 1170 health states. Preference weights in each functional dimension and symptoms were estimated by using the LAS from the general population (n=886) of San Diego, USA (Kaplan, 1982). The overall health state score (QWB Index) is calculated in two steps. First, by adding all the decrements in preference weight (negative utility) associated with the level of each dimension and the most highly weighted symptoms suffered by the patient. In the second step these decrements were aggregated by adding one. Unlike, HUI the QWB score allows negative values (“state worse than death”) but it is upper bounded by ‘ l=Perfect well being’. The QWB questionnaire is administered by trained interviewers in most of the cases and this make QWB is one o f the most expensive and time consuming among MAU instruments (Brazier, Deverill, Green, Harper and Booth, 1999). 23 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The QWB scale also has limited theoretical validity and reliability. Preference weights in the QWB were generated by the LAS technique. As discussed earlier, LAS is not a choice based technique and least preferred to generate preference weight. At the same time, the additive model in the QWB that generate the overall index has not been subject to any rigorous econometric testing. There is almost no published evidence of the inter-rate reliability of this scale (Brazier, Deverill, Green, Harper and Booth, 1999). The only published data on the retest reliability was an assessment of the inter-day reliability (Anderson et al., 1989) and it suggested a range of correlation between 0.78-0.99. Though, the results are questionable because the data were obtained retrospectively in one block rather than prospectively. Another MAU instrument EuroQol 5D (EQ-5D) is a five-item questionnaire, suitable for self-completion or interviewer administration. It was developed by a multidisciplinary group of researchers from seven centers across five European countries (EuroQoL Group, 1990). The five dimensions o f the EQ-5D are mobility, self-care, usual activities, pain or discomfort and anxiety or depression. Each of the dimensions has three level and together define 243 health states. Preference weights at each level of the five dimensions were generated by using the TTO and the LAS instruments from a representative (n=3,395) UK sample. The preference index is generated by an additive formula and similar to 24 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. the HUI it is also anchored by ‘0=Dead’ and ‘ l=Perfect Health’. Test-retest reliability o f the EQ-5D is found satisfactory, within a range o f 0.67 to 0.83 (Brazier et al., 1996; Hurst, 1994). However, content validity o f EQ-5D is restrictive compare to the HUI or the QWB because it is able to generate very limited unique health states (243 health states compare to 1170 by the QWB or 972,000 by the HUI). The three levels in each health dimension of the EQ-5D seem too crude to detect any small but clinically significant changes of health status. There are also little evidences available in the literature on the sensitivity of the EQ-5D. Although the TTO is an accepted method to elicit preference, it is theoretically inferior to the SG technique, as discussed earlier. Thus, use of the TTO in EQ-5D valuation is also a limitation to its construct validity. Review of the three major MAU instruments suggests that only the HUI is based on expected utility theory and theoretically most suitable for generating QALY. At the same time the HUI score’s ability to predict future health outcomes is supported by several studies (Torrance, 1996; Cohen, Barbano, Cox, et al., 1987). Nevertheless, practical application o f the HUI is limited to date specifically in the US population and it is very difficult to come to any conclusion about its empirical superiority over the EQ-5D or the QWB. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Development of Non Preference-Based Health Status Measurement Instrument Many outcome studies are now available where HRQOL is measured by descriptive or non preference-based HSM instruments; SF-36 (Hays, Sherboume, Mazel, 1993) or one of the similar, but shorter forms (i.e., SF-20, SF-12). The SF-36 is a 36 item self administered HSM survey that measures eight health domains: physical functioning (PF), social functioning (SF), role limitations due to physical health (RP), role limitations due to mental health (RE), general health (GH), emotional well-being (MH), vitality (VT) and bodily pain (BP) (Hays, Sherboume, Mazel, 1993). Each domain o f health contains several items or questions. The responses to these items are combined into dimension scores mainly using simple summation. The health domains described in the SF-36 ranges in score from 0-100, with the higher scores indicating higher level of functions or better health. The SF-36 scales can also be scored on a T-score metric (standardized (z) scores multiplied by 10 with 50 added), yielding mean of 50 and a standard deviation of 10. This linear transformation can facilitate interpretation of score, since, every scale is on a common metric (SF-36 Latest News, 1999). The reliability of the eight domains has been estimated using both internal consistency and test-retest methods. Review o f fifteen published studies revealed that reliability coefficients for each of the eight scale was equal or greater than 0.8 26 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. with the exception o f ‘Social functioning (SF)’, which had a median reliability of 0.76 (Ware, 1996). Two-week test-retest reliability exceeded 0.8 for PF, VT and GH domains but SF domain had a lower value of 0.6 (Brazier et al., 1999). Test- retest reliability for the domains after a delay of six months ranged between 0.6 and 0.9, except for the BP domain, which had a value of 0.43 (Nerenz, Repasky, Whitehouse et al., 1992). Criterion validity of the SF-36 is well established by McHomey et al. (1993). The domains discriminated between types and levels of disease and were also able to distinguish people with a chronic medical condition alone from those who had a medical disorder combined with a psychological one. Construct validity of the SF-36 was established by Ware et al. (1994) by comparing it with fifteen other HSM instruments. Correlations for the MH domain range from 0.51 to 0.81 with the corresponding scale in other leading measures; equivalent correlations were also found for other domains. Despite well-established reliability and validity, the SF-36 or any of its shorter forms was not designed for use in economic evaluation (discussed in the earlier chapter) and do not capture preference of an individual. Dimension scores of the SF-36 are not comparable among patients and currently there is no standard method to combine them into a single index. In a study on 1,862 advanced HIV- patients in the US, Bozzette et al., (1994) developed a perceived health index from the domains o f the SF-36. Their study suggested a unique way to summarize 27 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. descriptive health status instrument but did not address the issue of health state preferences. Neither it compared the relationship between perceived health index and any o f the health utility measurements. Due to this limitation, direct QALY computation from the SF-36 is not possible. This situation hampers researchers to derive a QALY approximation from most of the outcome studies using the SF-36. To overcome this problem, a limited set o f research has been reported to explore the potential for converting HSM into preference-based health assessment. In the next section, we reviewed all the relevant researches in this issue. First part of the review is dedicated to the empirical comparisons o f preference-based and non-preference based HSM instruments and it showed they are poor to moderately correlated. Next section reviewed the estimation algorithms developed to derive preference index from the SF-36. We found that most o f the studies with the exception of Brazier et al. (1998) used a mapping algorithm to estimate preference. However, Brazier et al. (1998) used the revaluing technique to estimate a preference-based index from the partial contents o f the SF-36. Considering the growing interest among clinicians for a shorter and precise HSM instrument, last section reviewed the development o f the SF-12 and the research done to estimate a preference-based index from it. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Empirical Comparisons of Preference and Non Preference-Based Health Status Measurements Several studies have been done to explore the relation between HSM and preference-based measures. Revicki and Kaplan (1993) has identified and published a review of fifteen relevant published studies between January 1985 and March 1993. Most of these studies employed QWB and TTO as one o f the preference-based instruments and compared them with HSM instruments like, SF- 36, General Health Rating Index, Kamofsky scale, Sickness Impact Profile etc. The Correlation coefficients between the TTO and the HSM (five studies) were estimated between poor (1%) to moderate (43%). The bivariate correlations between the LAS and the HSM instrument scores were varied between 0.17 and 0.46, with 3 to 21% of the variance in the LAS score being explained by individual HSM score but, multivariate analysis showed that 27-34% variance in the LAS score can be explained by the HSM score. Six studies used the QWB and found that 11-50% of variance was shared by HSM instruments and the QWB. One study (Brazier et al., 1993) compared the dimensions o f a MAU instrument (EQ-5D) with an HSM instrument (SF-36) and found bivariate correlations between 0.48 and 0.60. However, the SG scores (three studies) showed relatively unimpressive pair wise correlation (0.01 to 0.3) with the HSM scores. Multivariate analysis showed that only 1% to 25% variance o f the SG scores were accounted for the HSM scores. 29 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In conclusion, these studies suggested a moderate relation between preference and non preference-based HSM in various sample of patients. Developing a Standardized Algorithm to Estimate QALY from the SF-36 Fryback et al. (1996) developed a step-wise least square regression framework between the SF-36 and the Quality of well-being (QWB) scale on Beaver Dam Health Outcomes study (N= 1,430) and explored the prediction power of a health status measure (SF-36) to estimate the health value (QWB) of an individual. Results from this study showed that the SF-36 domains explained approximately 56% of the variance in the QWB score. The result was also cross validated on two different small cross-sectional samples (N=74 and N=57) of patients. In their study on 139 HIV infected patients and 124 primary care patients, Bult et al. (1998) correlated the SF-36 and the Time Tradeoff (TTO) utility assessment through latent class analysis. This analysis showed that 33% to 85% of variations in the TTO could be explained by the SF-36 scores, depending on the classification of patients in the study sample. On Dutch patients with intermittent claudication (n=76), Bosch and Hunink elicited a moderate correlation between the HUI (Mark 0 and the SF-36 (Adj. R2 =0.53) as well as the LAS and the SF-36 (Adj. R2=0.61). However, in that study the SF-36 scores did not explain well the variations in utility scores obtained from the Standard Gamble or Time Tradeoff 30 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. measurements. More recently in another study performed on a sample o f 165 people in UK, researchers developed a six item SF-36 questionnaire (Brazier, Usherwood, Harper, Thomas, 1998). These six items were able to predict 49% variations in the SG utility measurement and 68% variations in the LAS. A few other studies (Revicki, Brown, Henry, 1993; Nichol, Theil, Llewellyn-Thomas, et a l, 1992; Revicki, Weinstein, Alderman, e ta l, 1992) on small samples o f chronic disease patients have explored the relationship between the SG approach and various HSM. Health status measures (SF-36, SF-12, etc.) and the SG utility apparently shared only about 1% to 25% of common variance, depending on the measure of health states. However, before employing these methods for utility estimation, it is necessary to consider the limitations of the methods. Although the repeated measure research design where the same cross-section o f patients are observed over multiple observations are very common in clinical practice, the above mentioned studies only included cross-sectional databases. One of the drawbacks o f most of the studies mentioned earlier was the small sample size. All but the Beaver Dam Health Outcomes study were carried out on a sample size o f less than two hundred patients. Nevertheless, the predictive algorithm from the Beaver Dam study (Fryback et al., 1996) showed paradoxical relation between MH (of the SF-36), QWB score and BP (of the SF-36), QWB score. Patients with relatively good 31 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. mental health (MH) and low bodily pain (BP) were shown to have a lower QWB score, which is inconsistent with all the published empirical evidence discussed earlier. The preference-based six-item SF-36 was estimated from a small convenient UK sample. Out o f the 165 people only 55 peoples could be considered as outpatient and only half the total sample reported o f having any chronic diseases. None o f the studies on specific patient population (HIV, Primary care and Intermittent Claudication) attempted to adjust the models for socio-demographic characteristics of the patients. Another important disadvantage o f all the mapping algorithms was that the analyses were not performed on norm-based SF-36 scores and that made the results incomparable across studies. Thus, it is critical that the stability and consistency o f the mapping (transformation) approach be analyzed in different populations. Developing Preference-Based Index from the SF-12 Despite its extensive use in clinical trials, the SF-36 instrument has limitation. It is too long for inclusion in some large-scale health measurement and monitoring efforts (Ware, Kosinski, Keller, 1996). Recent researches have constructed SF-12 a shorter form of the SF-36 based on the physical and mental summary score of the SF-36. The SF-12 is evolved as a subset o f the SF-36 and its validity is well established by comparing result across numerous clinical studies that use the SF-36. Simultaneously test-retest reliability of the SF-12 is also 32 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. obtained from a general US population participating in the NSFHS survey (n=232) and a sample from the UK general population (n=187). Considering its ability to obtain sufficient health status information when constrained by questionnaire length or large sample size, the SF-12 has been translated and adapted over 30 language/country combinations (Ware, Kosinski, Keller, 1996). Moreover, in recent research on a large Swedish sample (n=5,404) Lundberg et al. (1999) developed a mapping algorithm from the items of the SF-12. Their study concluded that the SF-12 could account for 51% variation in the LAS and 23% in the TTO instruments. Conclusion Overall, it would appear that there are only moderate correlations between the preference measures and the SF-36, but this was not consistent between or within methods. Empirical evidences discussed in this section suggest that mapping methodology is more feasible than revaluing technique and both produces almost similar results. However, it is important to recognize that the preferred method would be to derive utility from community samples directly (Gold, 1996) by SG or TTO method but it is expensive and difficult to administer. At the same time, potential o f transforming HSM to a preference-based measurement is well acknowledged in the literature. Even so, more researches are necessary to suffice the potentials of current findings. 33 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. METHODS Populations Assessed The data set for this study included two populations: a random assignment set and a geographic set. In randomly assigned set, over 6,000 Kaiser Permanente members were randomly assigned to one of the three Pharmaceutical Care practice models (the Kaiser Permanente (KP) model, the State model or the Control group). These patients were surveyed at three separate times from 1992 to 1995. Seventy- five percent of the total 6,000 randomly assigned patients (4,500) completed the all three surveys. Additional sets of demonstration sites (large geographic sites) were set up in six counties in the Southern California area. This parallel survey (with random assignment sites) converted all pharmacies within large geographic sites to a single model o f care. Survey data were collected from 4,600 Kaiser Permanente patients residing in those six counties. Approximately sixty three percent of this sample (2,606) completed surveys at all three time points. The purpose of this original data collection design was to compare patient outcomes from different models of pharmacist consultations. Patients were also stratified in four sub sample based on their use o f prescription medications. Three mutually exclusive strata of high-risk patients (taking five or more medications, or target medications, or both) were specified to correspond to populations targeted for pharmacy consultation under KP model. A sample of normal patients who filled at least one 34 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. prescription during the prior year of the survey but did not meet any o f the high- risk criteria was also selected. Further details o f the design and purpose of the study are published elsewhere (McCombs, Kody, Besinique et al., 1995). Briefly, patients completed surveys including socio-demographic characteristics, the SF-36 and the LAS at initial baseline prior to pharmacist consultation (calendar year 1992). Follow-up surveys were completed after the first year of demonstration (April 1,1993 to March 30, 1994) and during the final 11 months o f demonstration (April 1, 1994 to February 28, 1995). The Health Utility Index questionnaire was administered during the final follow-up survey. The final dataset for this present research contained a total of 6,921 patients. From this data source, the following variables were created and are available for this research. Demographic Variables: Socio-demographic data including age, gender, ethnicity (asian black, latino, american indian, white), marital status (married, single, divorced, widowed), educational attainment (6 levels), employment (employed, unemployed, retired, disabled, working in the home or school) and income (8 levels) were included in the baseline survey. Health Status: Generic HRQOL was assessed through the SF-36 (version 1.0) and the LAS in the baseline and the follow-up surveys. However, for the purpose of this research the SF-12 items were also generated from existing SF-36 data in all the surveys. Survey responses in eight SF-36 domains were converted to 35 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. normalized t-scores. Imputed values were created for patients who failed to complete the survey questions required to calculate any one o f the eight domains of the SF-36. The imputed values were based on statistical models that regressed each domains o f health status on the remaining seven domains and other patient characteristics. A random component was also added to the imputed value to reflect the variation in the study population for the missing data. Patients with two or more missing values for SF-36 health domains were dropped from the database. Health Utility: Health utilities were assessed through the HUI-II and HUI-III standard questionnaire, which was completed during the third survey. Pharmacy Data: Pharmacy data including number and type o f prescriptions were abstracted from Kaiser’s automated database by retrieving a random 10% KP members in selected service area sites in the Southern California. It was broken down into twenty-eight dichotomous variables detailing the mix of drug classes used by the patient prior to the study period. Chronic Disease: The above mentioned drug class data were also used to create the Chronic Disease Score (CDS) (VonKorff, Wagner, Saunders, 1992). CDS is a measure of the number o f comorbidities scored from automated pharmacy data. The score ranges from 0-14, where higher scores indicate a greater number of chronic diseases. Using the CDS score instead o f the twenty-eight different drug classes (which are dichotomous variables) in analytical model as a regressor 36 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. preserves the degree of freedom in the analysis. It also simplifies the regression models which otherwise tended to be unstable. Statistical Models Assessment of Reliability of the Health Survey Instruments Bivariate Correlations of the attributes of the HUI-II, HUI-III, and eight domains o f the SF-36 (measured at the end of consultation study; t2 ) and the CDS (at the end of the study; t2 ) were estimated to determine the structural consistency o f the instrument in the study sample. Consistency of patient’s responses in the HUI and the SF-36 questionnaire were also assessed from this matrix. Reliability o f these instruments are already well established in several empirical study. However, to reconfirm the internal consistency reliability coefficients or Cronbach’s alpha (Cronbach, 1951) of these instruments were also derived in separate analysis. Because of its comparability across studies deriving alpha is considered over other reliability estimation techniques. Standardized Algorithm to Estimate HUI from the SF-36 As an econometric theoretical perspective, ordinary least squares (OLS) is the obvious method of choice in a cross-sectional study, because of its best linear unbiased estimation (BLUE) properties. A linear relationship among explanatory variables and the dependent variable (i.e., the composite score of an individual’s HUI) is assumed. 37 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Thus, the primary equation is stated as: 8 HUI = a + I PjtSF36(2)+Y Z + e (1) k = 1 where ‘HUI’ is the composite utility score of an individual patient at survey end point and obtained from HUI questionnaire. ‘SF36(2)’ is a vector of the eight domains of SF-36 measured at the consultation study end point fo). ‘Z’ is the vector of socio-demographic covariates and ‘e’ is the stochastic error term of the regression equation. All analyses were performed using the norm-based SF-36 T-scores by SAS® software. Since the USC/KP survey was a follow-up study on the same cross-section of patients, where the health state utility (HUI) was measured at the final stage and the SF-36 was measured at all three time periods, including the final stage we expected certain explanatory variables (such as the final stage SF-36 variables) to be endogenous (i.e., dependent on the past two SF-36 measures). Econometric literatures (Wu, 1973; Hausman, 1978) suggested the Wu-test could be used to determine the presence of endogenity. To control for endogenity, an instrumental variable method is introduced (Maddala, 1976) in addition to the simple OLS model. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. This model helped to mitigate endogenity bias o f the final stage SF-36 (2) variables by performing a series of OLS at 1 st stage equations, where the Ist stage equation is: i SF-36(2) = a ' + £ p SF36(0 + p CDSY2 + y ' Z + e, (2a) /=0 ‘SF36(t)’ are the SF-36 scores collected at baseline (to) and mid point (ti) of survey. The identification problem between the two structural equation models generated is minimized through the inclusion o f the CDS in the first of these two equations. As with the OLS model, ‘Z’ is the vector o f exogenous socio-demographic variables and ‘si ’ is the stochastic error term. However, each of the eight domains o f the SF-36 is distinct and measures different dimensions of health. Because of this each domain was estimated separately. These equations (2a) are based on the expectation that individual health status scores obtained from the SF-36 at the termination of study are dependent on the SF-36 scores o f the same individual at baseline, and midpoint, as well as their socio-demographic variables. The final stage predicted SF-36 variables are then used to obtain an unbiased estimation of the composite utility score. Thus the final 2SLS (two stage least square) model can be written as: UTILITY = a + p ffPT2 + p ^ P T 2 + P 3 GBT2 + P 4 MBT2 + p s 5FT2 + P 6 £ET2 + p 7 FTT2 + p g PFT2 + y Z + e2 (2b) 39 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Standardized Algorithm to Estimate LAS from the SF-36 The linear analog scale (LAS) and the SF-36 were simultaneously used in all the three surveys from baseline to final point. Considering this, a cross- sectional time series model is developed to estimate linear analog scale from the SF-36 domains. The single equation cross-sectional time series model is: 8 LAS,7 = a , + I p itkSF36//+ yZ,- + eit i= 1,2,.... ,N. r=0,l,2. (3) k- 1 where LAS,, is the linear analog scale of individual patients in three different time periods. SF36,-, is the eight domains of the SF-36 for each individual in different time periods. As this equation is developed to capture cross-sectional variations across time, coefficient o f the SF-36 is indexed for individual effect over three different time periods. (Z,) is the vector of socio-demographic variables that remain relatively static over time in this study. e„ represents the stochastic error term of this cross-sectional time series model. This model (eqn. 3) utilized the Yule-Walker (Yule, 1927; Walker, 1931) estimation technique. A second order autoregressive error method was assumed. The advantage o f this technique over the OLS estimation is that it helps to improve the model fit (R2 ) by computing a ‘Total R2’ statistics. This statistics is computed by using past error terms at baseline (to) and at midpoint (t|). 40 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Model Specification Test, Validation and Extension As data utilized for this study were generated not from a controlled experiment but from an observational study, particular attention is paid to specification and consistency o f models. This includes testing of model specification by Ramsey RESET test (Ramsey, 1969), reducing the potential inefficiency in parameter estimates due to heteroscedasticity and avoiding biases in estimation due to multicollinearity. Goodness-of-fit measures of the specified models are also compared with other general linear and non-linear models. All of the specification tests mentioned above are discussed briefly in this section. The RESET test (Ramsey, 1969) is conducted to check for model specification (for any omitted variable) because of its robustness and computational effectiveness. First, the predicted value of the dependent variable is raised to the second and third powers and then these newly created variables are added as regressors in the original OLS model. Next an F-test is performed separately on these two models against the null hypothesis that the basic model does not have any specification error due to omitted variables. Several approaches are taken to determine proper functional form (linearity vs. non-linearity) o f the models. Both the log-log model and the semilog model are compared with the linear model in terms of their goodness o f fits (R2). Simultaneously, a more formal test (Pg test, MacKinnon, White, Davidson, 1983) is carried out to obtain a significance test 41 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. against linear model. Pe test proposes two equations: hybrid (contains both linear and non-linear componnet) and linear. Then an OLS regression is performed on linear component and the differential prediction is obtained from the linear and the non-linear model. If the coefficient o f the differential term significantly contribute towards estimation, then the linear equation is rejected. Considering the presence of heteroscedasticity in regression model the t-values of parameter statistics are calculated from White’s corrected variance-covariance matrix (White, 1980b). Construct o f the SF-36 questionnaire suggested that domains of the SF-36 are correlated. Several econometric studies (Belsley, Kuh, Welsch, 1980) suggested this could generate multicollinearity (or near singularity in estimation matrix) and bias in estimation. Considering this problem, multicollinearity among health status variables (SF-36 components) in the models are minimized by not including interaction terms in the regression equations. Then a Condition Index/Variance Decomposition Proportion (Belsley, Kuh, Welsch, 1980) method is performed to detect bias (if any) in the estimation procedure attributable to multicollinearity. Point Estimate and Confidence Interval The appropriate use of all predictive algorithms in this study is to predict the mean HUI score for the sample population. One way of obtaining mean prediction is to apply the predictive equations at individual level and then average the resulting predicted HUI score. However, this approach is not useful for 42 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. researchers making secondary use of published SF-36 data. Raw scores o f the HSM instruments at individual level are rarely accessible from published data. Considering this aspect we choose an alternative approach to predict HUI, where the SF-36 data are averaged domain-wise and then the predictive equation is applied to the mean SF-36 domain scores for the sample. This method is compatible for any secondary HSM data with minimum reported information (for example, sample size and mean SF-36 scores). The basic assumption behind this alternative approach to generate the approximate confidence interval is that variance of the predicted mean HUI is obtained from two sources: 1) variance from making a regression prediction, and 2) variance due to the fact that the SF-36 means are sampled from a population. Estimation Model from the SF-I2 The SF-12 score is created from the SF-36 domains, available from the USC-KP consultation survey. This subset score is created by using the algorithm provided by the Medical Outcomes Trust (1994-95). At the same time, based on other studies1 by Medical Outcomes Trust on the SF-36 and the SF-12, we expect that creating embedded SF-12 would not underestimate or overestimate the HRQOL of the individual patient. The 12 items o f the SF-12 is categorized in the following way; physical functioning (PF), role limitation due to physical problem x SF-36 was originally embedded in a longer 149-item MOS Functional Health and Weil Being (FHWB) questionnaire and later administered unembedded. The SF12 was also originally administered as an embedded form o f the SF36 and later compared with unembedded form. 43 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (RP), role limitation due to emotional problem (RE) and mental health (MH) each has two items, social functioning (SF), general health (GH), bodily pain (BP) and vitality (VT) each consists of a single item. After creating the items o f the SF-12, only an OLS model is utilized in estimating HUI from the SF-12. Ordinary least square regressions (OLS) are used to determine the linear relationships of the HUI or the LAS with the SF-12 items, age, gender. O f the 12 items of the SF-12, 11 items represent different aspect o f health status, whereas the remaining item is a question about overall health status (poor health, fair health, good health, very good health or excellent health). Since this global or general health question is a measure of overall health status, it can be considered as a function of other 11 items of the SF-12. Considering this overlapping structure of the SF-12, two separate regression models are used to estimate HUI or LAS. In the first model only the general health item is entered to estimate HUI or LAS and in the second model the remaining 11 items o f the SF-12 are entered. In all of the regression models the SF-12 items are entered as categorical dummy variable with the worst health level as the baseline category (the first category of each SF-12 item). In this method, no assumptions were made about the size of the utility differences between individual response alternatives. As the baseline category (worst level of health) is the poorest health category, positive signs are expected for all SF-12 items used as regression coefficients. It is also hypothesized that the 44 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. magnitude of the regression coefficients would increase for each better response category. If this is not the case, the response category is dropped from the regression model so that an ascending sequence is obtained for an item. Both, the full regression model that included all items and all the response categories o f each item and the reduced model including only consistent regression coefficients are presented. Age and gender are included as explanatory variables in all regression models, since they may have effects on health-state utilities not captured by the SF-12. For instance, increasing age may lead to additional health problems that are not fully captured by the items o f the SF-12. The effect on health state utility of each item category is estimated as the effect compared to that of the baseline category. In short the basic estimation equations can be written as: UTILITY,— a , + p f GH .+ Z, + s, (4) it UTILITY,= a , + [3^SF12,+ y fZ ,+ e, (5) 45 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. RESULTS This chapter reports the results of the predictive models elicited for mapping HSM to the preference-based index. The chapter is organized around the five equations, which summarized the mapping algorithms between the SF-36, HUI, and SF-36, LAS. A final section summarizes the various findings and illuminates general patterns among parameter estimates in the equations. Sociodemographic The study sample (median age of 50.0 years at baseline (mean 49.8 years, range 25 to 95 years) was comparatively older than the national median (32.9 years for males and 35.2 years for females) (U.S. Bureau o f Census, 1995). More than sixty percent of the study population was female (Table 1). The ethnic diversity of the sample was similar to the distribution in the California population (State of California, Department of Finance, 1997) as fifty seven percent was White, 21.2% was Black, 14.1% was Latino, and 7% was Asian. Nearly half o f the study sample had some college education (42.6%) and almost 20% o f the sample graduated from college, which was consistent with the U.S. population in 1993. Almost sixty one percent of these managed care patients were employed and more than half o f the total sample (54%) had a self-reported income o f $50,000 or more per annum. Descriptive statistics o f non preference-based HRQOL (SF-36) and that for general US population are provided in table 2. The mean raw scores o f all the eight 46 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. domains o f the SF-36 between the study sample and the US general population are comparable and a t-test showed there is no statistically significant difference between the mean scores. There is a five-point clinical differences’ (Ware, 1996) between the mean scores in some o f the domains (RP, BP, PF, SF) which showed that study sample had limited activity for their physical problems compared to the general US population. Descriptive statistics provided at the table 3 illuminate a summary o f both the preference and the non preference-based HSM used in this study. Most of the participants in the study have a reasonably high health utility score in HUI. However, the LAS scores are relatively low compare to the HUI score. This supported the empirical evidences discussed earlier that the LAS underestimates the health utility of the individual because of the lack of choice options. On the other hand, unlike the previous table (Table 2), the SF-36 domain scales provided in this table are in norm-based T-score metric and this makes all mean domain scores cluster around fifty with a standard deviation of nine to eleven point. The scores were normalized by using general US population mean SF-36 scores. *A five-point difference is considered clinically significant in comparing standard scores. However, in this particular case comparisons were made between raw scores. ^ Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table I. Descriptive Statistics o f Demographics (N=6,921) Study Sample California Population' US Population2 Median age (years) 50.0 NA 35.2 Female (%) 63.9 NA 51.2 Ethnicity (%) Black 21.7 7.0 12.7 Asian 7.0 11.0 3.8 Latino 14.1 29.0 11.2* White 56.5 53.0 82.5 Education (%) College grad 19.8 NA 23.9 'State of California, Department of Finance, Race/Ethnic Population Estimates: Components of Change for California Counties, April 1990 to July 1997. http://www.dof.ca.gov/html/ 2 USA Statistics in Brief, 1999. http://www.census.gov/statab/ *Person of Latino or Hispanic origin in the US population may be of any race. Because of this the summation of the proportions exceeds one hundred percentage. Table 2. Comparison o f Health Status (SF-36) of USC/KP Study Sample at Endpoint (t?) with U.S. Population USC/KP Consultation Study Mean Raw Score (N=6,92l) US General Population Mean Raw Score (N=2,47l) SF-36 Domains General Health Perceptions (GH) 68.1 72.2 Role limitation due to Physical problem (RP) 70.7 81.2 Bodily Pain (BP) 66.3 75.5 Mental Health (MH) 72.9 74.8 Physical Functioning (PF) 77.6 84.5 Vitality (VT) 57.4 61.1 Role limitation due to Emotional problem (RE) 77.8 81.3 Social Functioning (SF) 78.4 83.6 f Source: Ware, Kosinski, Keller, 1994 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 3. Health Utility (HUI, LAS) and Health Status (SF-36) at Endpoint O2) (N=6,921) UCS/KP Consultation Study Mean (Std. Dev.) Median Mode Heath U tility Index (HUI-II) 0.80 (0.18) 0.85 0.92 Heath U tility Index (HUI-HI) 0.77 (0.21) 0.85 0.92 Linear A nalog Scale (L A S) 6 7 .9 3 (1 7 .6 1 ) 73 75 SF-36 Dom ains General Health Perceptions (GH) 46.98 (10.42) 48.90 54.85 R ole lim itation due to Physical problem (RP) 46.53 (11.83) 55.56 55.56 B odily Pain (BP) 46.35 (10.75) 48.51 60.40 M ental Health (M H ) 49.36 (9.97) 50.64 59.52 Physical Functioning (PF) 46.36 (11.30) 50.20 56.75 V itality (V T ) 48.30 (10.16) 4 9.49 59.07 R ole lim itation due to Emotional problem (RE) 48.94 (10.94) 55.66 55.66 Social Functioning (SF) 46.61 (11.33) 51.74 57.33 Assessment of Reliability The relation between the attributes of the HUI and the domains of the SF-36 was as expected (Table 4a, 4b). For example, mental health (MH) component of the SF-36 and emotional attribute of the HUI-II showed a high correlation of 0.70 but MH showed a relatively poor correlation o f 0.13 with mobility attribute o f the HUI-II. The correlation matrix shown on this table was also helpful to verify the construct validity of both the SF-36 and the HUI in this sample. Reliability coefficients or Cronbach’s alpha (Cronbach, 1951) were generated for the SF-36 49 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. domains and the HUI in this sample. Cronbach’s alpha of 0.86 to 0.88 and 0.65 for the SF-36 domains and the HUI were found from the analysis. Chronic disease score (CDS) were found negatively associated with the SF-36 domains, which is not surprising since the higher CDS represent poorer health. Similar associations were evident in the baseline and midpoint measures. As mentioned in the earlier chapter, the HUI-III is conceptually more appropriate for the adult population than the previous version o f the HUI. Considering this the relations between the HUI- III attributes and the SF-36 domains were also explored in another correlation matrix (Table 4b). For example, emotional attribute of the HUI-III is strongly correlated with all the three domains under mental component (MH, RE, VT) o f the SF-36 with correlation coefficient ranges between 0.64 to 0.42. Similarly, pain attribute of the HUI-III is strongly correlated with most of the domains under the SF-36 physical component but less strongly correlated with the mental health components. However, the newly created attributes vision, hearing and speech did not show strong correlation with any o f the domains of the SF-36, which may reflect the constructional differences between these two instruments. Internal consistency o f the HUI-m in the study sample is also obtained from a separate analysis and the cronbach alpha (0.65) showed no improvement over the HUI-II. 50 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 4a. Correlation Matrix HUI-II, SF-36, CDS at Endpoint fe) (N=6,921) HUI-II attributes at endnoiintitd CDS Sensation Mobility Emotion Cognition Self-care Pain UTILITY 0.48 0.42 0.55 0.72 0.28 0.66 -0.17 SF-36 domains at endnoint ft?l 0.23 0.32 0.38 0.28 0.21 0.40 -0.28 General health Role limitation due to physical problem 0.21 0.36 0.26 0.25 0.21 0.44 -0.22 Bodily pain 0.19 0.33 0.31 0.23 0.21 0.62 -0.16 Mental health 0.17 0.13 0.70 0.36 0.11 0.27 -0.04 Physical functioning 0.23 0.51 0.19 0.22 0.29 0.44 -0.28 Vitality 0.20 0.27 0.50 0.32 0.19 0.39 -0.14 Role limitation due to emotional problem 0.19 0.21 0.47 0.33 0.13 0.26 -0.10 Social functioning 0.17 0.31 0.43 0.26 0.24 0.38 -0.11 all correlation coefficien ts are significant at p > 0.0001 Regression Model: Estimating HUI from the SF-36 Results obtained from the OLS analysis (eqn.l) demonstrated a statistically significant association between all eight domains o f the SF-36 scores (measured at the final follow-up) and the HUI (R2 o f 0.51 and 0.49 (Table 5a, 5b)). Although the negative association between the patient’s age and utility was statistically significant, the magnitude was small. Gender and all other socio-demographic variables were not significantly associated with utility in this multivariate model. Considering the skewed distribution of some of the SF-36 domains, the linear model was compared with log-log and semi-log models. The Log-log 51 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 4b. Correlation Matrix HUI-III, SF-36 at Endpoint (t2> (N=6,921) HUI-III attributes at endnoint lt£ Vision Hearing Speech Emotion Cognition Pain Dexterity Ambu -lation UTILITY 0.39 0.47 0.36 0.51 0.66 0.57 0.36 0.48 SF-36 domains at endnoint ft 0.15 0.13 0.13 0.32 0.29 0.38 0.19 0.33 General health Role limitation due to physical problem 0.14 0.12 0.14 0.21 0.26 0.43 0.22 0.39 Bodily pain 0.10 0.11 0.11 0.25 0.22 0.63 0.25 0.37 Mental health 0.10 0.08 0.15 0.64 0.37 0.23 0.09 0.12 Physical function 0.17 0.13 0.14 0.14 0.21 0.42 0.27 0.55 Vitality 0.13 0.11 0.14 0.45 0.35 0.37 0.16 0.30 Role limitation due to emotional problem 0.12 0.09 0.16 0.42 0.34 0.23 0.14 0.21 Social function 0.10 0.08 0.14 0.37 0.28 0.33 0.14 0.31 all correlation coefficients are significant at p > 0.0001 model showed a gross misspecification by not having a full rank. Semi-log model showed poor R2 of 0.45 and Pg test (MacKinnon, White and Davidson, 1983) did not accept the hypothesis in favor o f the semi-log model. Model specification was tested by the RESET test (F= 4.403, p > 0.07) and it suggested absence of any omitted variable in the predictive model. At the same time, t-statistics (mentioned 52 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. in table 5a, 5b) obtained from White’s variance-covariance matrix (corrected for heteroscedasticity) showed no change in statistical significance of estimates obtained from the initial models (uncorrected for heteroscedasticity). The condition index/variance decomposition proportion method generated an index value of <30 for all the regressors used in the models, whereas literature (Welsh et al., 1980) suggests that an index > 30 can only generate estimation biases from multicollinearity. Thus any bias in estimation from multicollinearity in this study can be rejected. Derived Regression Equation to Correct Endogenity of the SF-36 Components As hypothesized and discussed earlier, the SF-36 domains at the final point of the survey were not exogeneous and rather dependent on the SF-36 domain scores from the 1 st two survey periods. Results of the Wu test (Wu, 1973) support the presence o f endogenity (Table 6). In this procedure, residual terms were obtained from two equations where mental health component score (MCS) and physical health component score (PCS) o f the SF-36 at final points of USC/KP survey were regressed separately on all true exogenous variables. High t -ratio statistics and low p value of these two residual terms (RESMCS, RESPCS) in table 6 suggest that the SF 36 components measured at final points are endogenous, i.e., dependent on baseline and midpoint SF-36 component scores. In all of the equations (2a) described earlier, the parameter estimates o f two lagged SF-36 53 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. components showed statistical significance (pO.OOOl) in determining the variation of their corresponding finai-stage SF-36 components. All equations also had a potential problem o f serial correlation between stochastic error terms and lagged SF-36 variables. To verify the presence of any Ist order serial correlation, a Durbin-Watson (DW) test and Durbin-h statistic (Durbin, 1970) was performed for all eight regression equations. It showed that amount of 1 st order serial correlation ( s 2) were almost non-significant (Table 7). Parameter estimates of seven domains of the predicted SF-36 scores showed statistical significance (eqn. 2) in determining the variation in utility score (R2 = 0.34 and 0.35, (Table 5a, 5b)) in the second stage. However, in the two-stage model in this study “Social functioning (SF)” showed statistical insignificance (Table 5a) and this might be due to the exclusion o f social health concept in the HUI questionnaire. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 5a. Estimated Effect of SF-36 Domains on HUI-II from OLS and 2SLS Models (N=6,921) OLS 2SLS Variable Parameter Estimate (t- value) Prob> |Ti Parameter Estimates (t-value) Prob> |T| INTERCEP 0 .0 451(2.74) 0.0061 -0 .12 2 (-4 .19) 0.0001 General health perceptions (G H ) 0.0 009(4.33) 0.0001 0.0009 (2.43) 0.0109 R ole lim itation due to physical problem (RP) 0.0 0046(2.30) 0.0213 0 .0 0 1 7 (3 .2 7 ) 0.0018 B od ily pain (BP) 0.0 0 4 3 (21.54) 0.0001 0 .0 0 4 7 (1 0 .3 3 ) 0.0001 M ental health (M H ) 0 .0 0 4 2 (1 8 .0 8 ) 0.0001 0.0042 (8.8 0 ) 0.0001 Physical functioning (PF) 0 .0 0 1 8 (9 .0 5 ) 0.0001 0.0025 (6.4 8 ) 0.0001 V itality (V T ) 0 .0 0 1 8 (7 .4 4 ) 0.0001 0.0023 (5.15) 0.0001 R ole lim itation due to em otional Problem (RP) 0 .0 0 1 5 4 (8.03) 0.0001 0 .0 0 1 9 (3 .3 5 ) 0.0007 Social functioning (SF) 0 .0 0 1 5 (8 .2 7 ) 0.0001 0 .0 0 1 3 (2 .0 4 ) 0.0644 Gender 0 .0 0 5 4 (1 .6 2 ) 0.1045 0 .0 0 6 8 (1 .5 3 ) 0.0959 A ge -0.0006 (-5.42) 0.0001 -0.0002 (- 1.57) 0.1390 M odel F = 268.92 (P > 0.0001) R‘= 0.51 Adj.R 2 = 0.50 F = 130.23 (P > 0.0001) R*= 0 .34 Adj.R2 = 0.33 55 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 5b. Estimated Effect of SF-36 Domains on HUI-III from OLS and 2SLS Models (N=6,921) OLS 2SLS Variable Parameter Estimate (t- value) Prob> |T| Parameter Estimates (t-value) Prob> |T| INTERCEP -0.104 (-6.59) 0.0001 -0 .3 5 3 (1 1 .5 4 ) 0 .0 0 0 1 General health perceptions (G H ) 0 .0 0 1 5 (5 .7 1 ) 0.0001 0 .0 0 1 5 (3 .3 9 ) 0.0007 R ole lim itation due to physical problem (RP) 0.0008 (3.65) 0.0003 0.0020 (3.44) 0.0006 B od ily pain (BP) 0 .0 0 4 3 (1 8 .2 5 ) 0.0001 0.0043 (8.22) 0.0001 M ental health (M H ) 0 .0 0 4 5 (1 6 .6 1 ) 0.0001 0 .0 0 4 6 (8 .1 3 ) 0.0001 Physical functioning (PF) 0 .0 0 2 8 (1 1 .6 8 ) 0.0001 0.0042 (9.40) 0.0001 V itality (V T ) 0.0021 (7.58) 0.0001 0.0028 (5.29) 0.0001 R ole lim itation due to em otional Problem (RP) 0 .0 0 1 6 (7 .2 9 ) 0.0001 0 .0 0 1 9 (2 .9 2 ) 0.0035 Social functioning (SF) 0 .0 0 1 6 (7 .2 5 ) 0.0001 0.0024 (3.06) 0.0022 Gender 0 .0 0 7 3 (1 .8 5 ) 0.0649 0.0107 (2.07) 0.0387 A ge -0.0012 (-8 .8 1) 0.0001 -0.0006 (-2.99) 0.0028 M odel F = 655.20 R2 = 0.50 F = 277.62 R2 = 0.35 (P > 0.0001) Adj.R2 = 0.49 (P > 0 .0 0 0 t) Adj.R2 = 0.34 Table 6. Results o f Wu Test for Detecting Endogenity in SF-36 Domains at Endpoint (t2 ) Variable Parameter Estimate t-ratio Prob>|Tl INTERCEP -0.29 -1.09 0.2772 M C SH A T 0 .0 1 1.55 0.1204 PC SH A T 0 .0 1 2.33 0.0197 RESM CS 0.01 52.07 0.0001 RESPCS 0.01 49.71 0.0001 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 7. Test o f Serial Correlations in 1st Stage Equations from Lagged SF-36 (N=6,921) Durbin W atson C oefficient t- ratio p-value Physical functioning 1.97 0 .29 23.55 0.0001 (PF) 0.53 45.38 0.0001 R ole lim itation due to 2.01 0.24 19.88 0.0001 physical problem (RP) 0 .39 32.12 0.0001 B od ily pain (B P) 2.01 0 .26 20.81 0.0001 0.43 34.79 0.0001 General health 1.95 0.28 23.37 0.0001 perceptions (GH) 0.51 41.30 0.0001 Vitality (V T ) 1.96 0.33 26.57 0.0001 0.46 37.02 0.0001 Social functioning (SF ) 2.00 0.25 18.71 0.0001 0.31 22.75 0.0001 Role lim itation due to 1.99 0.26 21.14 0.0001 em otional problem (R E ) 0.31 25.48 0.0001 Mental health (M H ) 1.95 0.31 24.65 0.0001 0 .44 36.27 0.0001 Regression Model: Estimating LAS and HUI from the SF-12 In table 8 the relationship between the LAS and the self-reported general health (GH) question from the SF-12 are shown. LAS was positively associated with health status, the LAS value increased as general health increased. Parameter estimates for all general health categories were significant. The LAS value also increased with age but the magnitude was negligible. Women appeared to have slightly lower LAS values, although the coefficient for gender was not significant. The regression equation predicted 50% o f the variance in the LAS. 57 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The predicted LAS values for the five health states for a 50 year old patient were: 0.38 (poor health), 0.51 (fair health), 0.66 (good health), 0.79 (very good health) and 0.90 (excellent health). The results of the linear regression analysis o f the relationship between the responses to the HUI-III and the general health (GH) item in the SF-12 are also shown in table 8. Similar to the LAS analysis the HUI score increased with increasing health status in GH, and all of the health states were highly significant. Unlike the LAS, the HUI score decreased significantly with increasing age, though the magnitude was small. The coefficient for gender was significant and women had slightly lower HUI scores. The regression equation explained only 23% of the variance in the HUI scores although the association was in the hypothesized direction (higher general health score was associated with higher HUI scores). The predicted HUI scores for a 50 year old male patient were 0.49 (poor health), 0.62 (fair health), 0.77 (good health), 0.85 (very good health), and 0.91 (excellent health). The relationship between the responses to the LAS and the eleven items of the SF-12 are shown in table 9a. This linear regression analysis was based on 6,778 observations. The LAS values decreased significantly with age. The LAS values were also slightly higher for women than for men. Most o f the items o f the SF-12 were significantly associated with the LAS values in the expected direction. 58 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Some o f the items in the mental health (MH) dimension (calm and peaceful, downhearted and blue) and social functioning (SF) dimension had an unexpected negative effect. As described in the methods section, the item was removed from the reduced model if it did not show consistency and statistical significance (p>0.0001 level). The reduced model, which is based on 6,778 observations, is also shown in table 9a. An F-test was performed to test the joint significance of all the response alternatives in the reduced model and it showed all the items were significant at the 1% level. The reduced regression model explained 31% of the variance in the LAS value. The reduced regression model predicted the LAS score for a 50 year old male patient between 0.41 (the worst level on all the SF-12 quality of life items) and 0.81 (the best level on all the SF-12 quality of life items). The results of the linear regression analysis of the relationship between the responses to the HUI and the items o f SF-12 are shown in table 9b. The HUI scores decreased significantly with an increase in age. The HUI scores were also significantly higher for women than men. Similar to the LAS analysis, most o f the SF-12 items were associated with the HUI values in the expected direction although several o f the 11 items had an unexpected negative association. For three items (two mental health items and the social functioning item), response categories were omitted in the reduced model to get a consistent estimation. The reduced model explained 45% o f the variation in the HUI score. An F-test was 59 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. performed to test the joint significance o f all response alternatives for items used in the reduced model. Most o f the items were significant at the 1% level. One o f the Vitality (VT) items and the first item in the Role Function-Emotional (RE) dimension was significant at the 5% level. The reduced regression model predicted the HUI score for a 50 year old male patient between 0.27 (the worst level on all the SF-12 quality of life items) and 0.95 (the best level on all the SF-12 quality of life items). The stability o f the estimated regression models was obtained from the model specification and other statistical diagnostic tests. Model specification ( P e test) (MacKinnon, White and Davidson, 1983) favored the linear model against non-linear one at the 5% level. The condition index/variance decomposition proportion method generated an index value o f <30 for all the regressors used in the models and thus any bias in estimation from multicollinearity can be rejected. All the models developed in the present study were internally validated by randomly splitting it into two subsamples. None of the regression coefficients differed significantly at the 5% level between the two subsamples. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 8. Linear Regression Analysis o f the Relationship between Utilities (LAS and HUI) and Responses of the General Health Item of the SF-12 LINEAR ANALOG SCALE (n=6,778) HEALTH UTILITY INDEX (n=6,858) Variable Parameter Estimate Parameter Estimate (SE)b (SE)b INTERCEP 0.37 (0.013) 0.59(0.020) General Pronortion health of Responses Poor* 0.02 Fair 0.18 0.13(0.010) 0.13(0.016) Good 0.41 0.28 (0.009) 0.28 (0.015) Very good 0.32 0.41 (0.010) 0.36(0.015) Excellent 0.07 0.52 (0.011) 0.42(0.017) Gender -0.003 (0.003)* -0.014(0.005) Age 0.0003 (0.0001) -0.002 (-9.821) Model F = 1140.39 (P>0.0001) F = 343.58 (P>0.0001) R2 = 0.51 R2 = 0.24 Adj.R2 = 0.50 Adj.R2 = 0.23 ♦Baseline category 'not significant at p>0.05 level, all other variables are significant at p>0.0001 level in the LAS model. b all the variables are significant at p>0.0001 level in the HUI model. Table 9a. Linear Regression Analysis o f the Relationship between LAS and 11 Items o f the SF-12 (N=6,778) Full Model Reduced Model Variable ProDortion of Parameter Estimate Parameter Estimate Responses (SE) (SE)“ INTERCEP 0.47 (0.021) 0.41 (0.017) Physical function Moderate activities Limited a lot* 0.11 ------- Limited a little 0.26 0.01 (0.006) 0.01 (0.007)** Not limited at all 0.63 0.03 (0.007) 0.03 (0.007) Climbing several stairs Limited a lot* 0.17 Limited a little 0.31 0.01 (0.005) 0.01 (0.006)** Not limited at all 0.52 0.05 (0.006) 0.05 (0.006) 61 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 9a. Contd. Bodily pain Pain interfered with work Extremely* Quite a bit Moderately A little bit Not at all 0.02 0.09 0.16 0.27 0.46 0.05(0.013) 0.07(0.013) 0.09 (0.013) 0.10(0.013) 0.04(0.013) 0.07(0.013) 0.09 (0.013) 0.12(0.013) Mental health Calm and peaceful None of the time* 0.04 A little of the time 0.11 -0.02 (0.008) *** Some of the time 0.21 -0.01 (0.007) *** A good bit of the 0.21 -0.01 (0.007) *** time Most of the time 0.37 -0.002 (0.006) *** All of the time 0.07 -0.001 (0.006) 0.01 (0.004) Downhearted and blue None of the time* 0.39 A little of the time 0.34 -0.06 (0.016) *** Some of the time 0.18 -0.04 (0.014) *** A good bit of the 0.05 -0.02 (0.0127) *** time Most of the time 0.02 -0.002 (0.0124) *** All of the time 0.01 0.001 (0.0124) 0.012(0.004) Role Function- Physical Physical health interfered with work Accomplished less than you would 0.37 -------- -------- like* Didn’t accomplish less 0.63 0.02 (0.005) 0.02 (0.005) Limited in the kind of work* 0.29 Wasn’t limited in the kind of work 0.71 0.02 (0.005) 0.01 (0.006) Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 9a. Contd. Vitality Enerev None of the time4 A little of the time Some of the time A good bit of the time Most of the time All of the time 0.07 0.15 0.25 0.22 0.27 0.04 0.02 (0.008) 0.04(0.007) 0.08 (0.008) 0.12(0.008) 0.15(0.012) 0.02 (0.008) 0.05 (0.007) 0.09 (0.008) 0.14(0.008) 0.17(0.012) Role Function- Emotional Emotional health interfered with work Accomplished less 0.28 -------- -------- than you would like4 Didn’t accomplish less 0.72 0.001 (0.005) 0.01 (0.005)44 Didn’t do work as carefully as usual4 0.17 ---------- Did work as carefully as usual 0.83 -0.01 (0.005) *** Social functioning Phvsical and emotional health interfered with work All of the time4 0.10 --------- A good bit of the time 0.06 -0.03 (0.010) *** Some of the time 0.17 -0.02 (0.008) *** A little of the time 0.20 -0.02 (0.007) *** None of the time 0.48 0.01 (0.006) 0.02 (0.004) Gender 0.012(2.791) 0.008 (0.004) Age -0.0012 (-7.938) 0.0003 (0.0001)44 Model F= 101.42 F = 171.18 (P>0.0001) (P>0.0001) R2 = 0.33 R2 = 0.31 Adj.R2 = 0.32 Adj.R2 = 0.30 ♦Baseline category. *all the variables are significant at p>0.0001 level in the reduced model, ♦♦not significant at p>0.05 level. ♦♦♦Item was removed because of the negative regression coefficients. 63 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 9b. Linear Regression Analysis of the Relationship between HUI-UI and 11 Items of the SF-12 (N=6,778) Full Model Reduced Model Variable Proportion Parameter Estimate Parameter Estimate of Responses (SE) (SE)a INTERCEP 0.41 (0.023) 0.31 (0.019) Physical function Moderate activities Limited a lot* 0.11 Limited a little 0.26 0.02 (0.007) 0.02 (0.007) Not limited at ail 0.63 0.03 (0.008) 0.04 (0.008) Climbing several stairs Limited a lot* Limited a little 0.17 0.31 0.04 (0.006) 0.04 (0.007) Not limited at all 0.52 0.05 (0.007) 0.05 (0.007) Bodily pain Pain interfered with work Extremely* Quite a bit 0.02 0.09 0.08 (0.015) 0.09(0.015) Moderately 0.16 0.16(0.015) 0.18(0.014) A little bit 0.27 0.19(0.015) 0.21 (0.014) Not at all 0.46 0.23 (0.015) 0.25(0.015) Mental health Calm and oeaceful None of the time* A little of the time 0.04 0.11 -0.04 (0.009) AA Some of the time 0.21 -0.01 (0.008) AA A good bit of the 0.21 0.001 (0.008) AA time Most of the time 0.37 0.01 (0.007) AA All of the time 0.07 0.02(0.005) 0.03 (0.005) Downhearted and blue None of the time* A little of the time 0.01 0.02 -0.11(0.018) AA Some of the time 0.05 -0.05 (0.016) AA A good bit of the 0.18 -0.003 (0.014) AA time Most of the time 0.34 0.03(0.013) 0.05 (0.006) All of the time 0.39 0.06 (0.013) 0.08 (0.006) 64 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 9b. Contd. Role Function- Physical Physical health interfered with work Accomplished less than you would like* Didn’t accomplish less Limited in the kind of work* Wasn’t limited in the kind of work 0.37 0.63 0.29 0.71 0.02 (0.006) 0.02 (0.006) 0.02 (0.006) 0.02 (0.006) Vitality Energy None of the time* 0.07 A little of the time 0.15 0.02 (0.009) 0.02 (0.009) Some of the time 0.25 0.05 (0.008) 0.06 (0.008) A good bit of the 0.22 0.07 (0.009) 0.08 (0.009) time Most of the time 0.27 0.08 (0.009) 0.09 (0.009) All of the time 0.04 0.11 (0.013) 0.12(0.013) Role Function- Emotional Emotional health interfered with work Accomplished less than you would like* 0.28 Didn’t accomplish 0.72 0.01 (0.006) 0.01 (0.006) less Didn’t do work as 0.17 carefully as usual* Did work as 0.83 0.04 (0.007) 0.05 (0.006) carefully as usual Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 9b. Contd. Social functioning Phvsical and emotional health interfered with work All of the time* A good bit of the time Some of the time A little of the time None of the time 0.10 0.06 0.17 0.20 0.48 -0.08 (0.011) -0.02 (0.009) -0.002 (0.008) 0.02 (0.007) AA AA AA 0.03 (0.005) Gender 0.012(0.004) 0.011 (0.004) Age - 0 . 0 0 1 (0.0001) -0.001(0.0001) Model F = 187.04 (P>0.0001) R2 = 0.47 Adj.R2 = 0.46 F= 234.28 (P>0.0001) R2 = 0.45 Adj.R2 = 0.44 *Baseline category. ^Combined with the following better response category. ■ * a ll the variables are significant at p>0.0001 level in the reduced model. Regression Model: Estimating LAS from the SF-36 As mentioned earlier, results o f this cross-sectional time series (CSTS) method (eqn. 3) was obtained from the Yule-Walker (Yule, 1927; Walker, 1931) estimation technique. The analysis on the estimation of the LAS score from the domains of the SF-36 demonstrated a statistically significant association between all eight domains of the SF-36 (except role limitation (RE) due to physical problem) and the LAS (Total R2 of 0.49 (Table 10)). However, this algorithm also showed a paradoxical result in estimating coefficient o f the RE (role limitation due to emotional problem) domain. The marginal effect o f the RE domain on estimating the LAS score turned out negative, which is inconsistent with findings 66 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. from the previous models. Considering the contradictory result, it can be hypothesized that the parameter estimate of RE from the CSTS technique was influenced by some random error. Another contradictory finding from the CSTS model is that positive association between the patient’s age and the LAS score. Though it was statistically insignificant in estimating the LAS. One explanation of this relation is that the LAS score does not actually capture the risk of choosing a health state. Gender and all other socio-demographic variables were not significantly associated with the LAS score in the multivariate model. Test of the model specification showed that biases in the estimation due to the autocorrelated error was negligible in the model (DW statistics =2), but potential efficiency of the model was suffered by the presence of heteroscedasticity. Several attempts (GARCH, IGARCH estimation) were made to mitigate this problem but the limited numbers of observed time periods did not permit us to resolve it. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 10. Estimated Effect of SF-36 Domains on LAS from Cross-sectional Time Series Model (N=6,921) Variable Parameter Estimate (t-value) Prob> |T| INTERCEP 3.84(6.22) 0.0001 General health perceptions (GH) 0.85 (80.92) 0.0001 Role limitation due to physical problem (RP) 0.02(1.89) 0.0586 Bodily pain (BP) 0.13(12.68) 0.0001 Mental health (MH) 0.03 (2.90) 0.0037 Physical functioning (PF) 0.09(9.34) 0.0001 Vitality (VT) 0.16(13.41) 0.0001 Role limitation due to emotional Problem (RE) -0.03 (-2.83) 0.0046 Social functioning (SF) 0.07(6.50) 0.0001 Age 0.11 (1.90) 0.0572 Model Total Ri = 0.49 Point Estimate and Confidence Interval As hypothesized earlier, empirical evidence from the present research showed that sample size and variability in the SF-36 domains are by far the largest influences on variability in prediction. Accordingly, a sampling experiment was performed from the total Kaiser sample. For sample sizes in the range o f n=25 to n=1000, an approximate standard error of the predicted HUI (from the SF-36) score was computed. These standard errors are reported in table 11, which also reports the incremental value used to compute 95% confidence intervals (1.96 times the approximate standard error of the mean). For example, if the sample size is 100, the approximate 95% confidence interval for the predicted HUI score is given by 68 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. subtracting and adding 0.09 to the predicted mean score computed from the OLS regression equation (with the interval truncated at 0 and 1.0, the bounds of the scale). However, this type o f mean prediction is suitable only from SF-36 domain score, because there is no way to generate any domain score from the itemized domains of the SF-12. Table 11. Sample Size and 95% Confidence Intervals of Mean Prediction (HUI from SF-36) Sample Size OLS 2SLS Approximate Standard Error of the Predicted Mean Increment for 95% Confidence Interval Approximate Standard Error of the Predicted Mean Increment for 95% Confidence Interval 25 0.065 0.127 0.067 0.131 50 0.058 0.II4 0.056 0.109 75 0.053 0.104 0.049 0.096 too 0.046 0.090 0.044 0.086 125 0.042 0.082 0.041 0.080 150 0.039 0.076 0.039 0.076 175 0.036 0.070 0.035 0.068 200 0.033 0.065 0.032 0.062 250 0.028 0.055 0.028 0.055 300 0.025 0.049 0.024 0.047 350 0.023 0.045 0.023 0.045 400 0.021 0.041 0.022 0.043 500 0.019 0.037 0.019 0.037 600 0.017 0.033 0.017 0.033 700 0.016 0.031 0.016 0.031 800 0.015 0.029 0.015 0.029 900 0.014 0.027 0.014 0.027 1000 0.013 0.025 0.013 0.025 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CONCLUSIONS The results of this study provide the answers o f the questions raised in the objective of the research. However, as with all the relevant studies in this field, this research has several limitations and it raises questions, which need further exploration. This chapter details a brief discussion of three issues. First, the contribution o f the present research to the mapping techniques (to generate a preference-based HSM from secondary database using the SF-36 or SF-12) is summarized. Second, the application o f this study to evaluate cost-effectiveness ratio from clinical research is outlined. Finally some of the limitations, which need further investigation are discussed. Contribution to the Mapping Technique Considerable work has explored the relationship between the SF-36 and preference-based techniques (Fryback etal., 1996; Bult et al., 1998). Although these studies emphasized the necessity to develop a preference-based HSM, they failed to develop any specific mapping algorithm that could generate a preference index from secondary database containing descriptive HSM like the SF-36. Fryback and co-authors carried out the first study in this field where a mapping algorithm was developed to transform the SF-36 to QWB score from a large US sample. Prior to this, most o f the research in this field emphasized the SF-36 and one o f the direct measures o f the health state preference technique from small 70 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. clinical trial samples. Because of this, algorithms developed from those studies are not generalizable and there is no evidence o f their empirical validity in a general US population. The present study attempts to broaden the research base in this area by developing several mapping algorithms that can convert the SF-36 domains to the HUI score. After Fryback et al, this is the only study that explores the empirical relation between a multi-attribute utility instrument and the SF-36 from a large heterogeneous US sample. This research sought to determine the proper estimation methodology depending on the nature of the database with extensive focus on the reliability and validity o f the methods. Further attention was given to emphasize the feasibility o f the mapping algorithm, so that researcher using any secondary database with the SF-36 can utilize it. Several areas of discovery should be emphasized. First, this research demonstrates that the linear combination of the eight domains of the SF-36 can predict almost 50% of the variations in the HUI score. This study also confirms that first order or second order interaction terms among the SF-36 domains do not increase the predictive power o f the model, but rather can generate severe multicollinearity and inconsistency in estimation. These findings also suggest that a linear relation between the SF-36 and the HUI is more suitable than a non-linear one in a large sample. 7 1 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Second, this study extends the applicability o f the mapping algorithm to the SF-12 items. Although the SF-12 is a shorter version of the SF-36 and more convenient to administer, studies suggest that it does not capture a substantial amount of health status information. However, our algorithm shows that the SF-12 items can explain almost 46% variation in the HUI score compared to 50% of variation predicted by the SF-36 domains. This is contrasted with the original work by Ware et al. (1996), which suggested that the SF-12 estimates were 10% less accurate than the SF-36. This finding points out the need for further study to reevaluate the construction of the SF-36 (Version 1.0) and the relationship with the SF-12. Recently, a newer and enhanced version of the SF-36 (Version 2.0) was developed. So it will be of immense interest to observe the predictive ability of this newer version to generate a preference index. Third, this research sheds some light on various techniques of handling secondary databases. It was confirmed that the unique nature of this longitudinal database generated endogenity bias in the dependent variables. Considering this several models are proposed to mitigate that bias. Finally, a two stage modeling approach is utilized to mitigate the bias. Simultaneously a cross-sectional time series approach is used to develop a stable algorithm from repeated measurements. These techniques are significant contributions to the research o f mapping. 72 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Implications for Economic Evaluation from Clinical Research There are many different types o f non preference-based HSMs available in the health outcomes research, but all of them share a common limitation that they have not been designed for economic evaluation. The scoring algorithms of those HSMs usually assume equal intervals between response choices, items and dimensions and hence are unlikely to reflect preferences. Some of the HSMs, which used more sophisticated techniques (multitrait-multimethod matrix, latent variable, or MIMIC modeling) of weighting a health index also have not incorporated any implicit economic theory o f choice. Thus, this health index is not likely to reflect health state preferences on an interval scale. The problem becomes more apparent for disease specific HSM instruments when profile or dimension scores are not even comparable across different instruments. Nevertheless, preference-based HSM instruments are also fraught with problems. There is no debate that these instruments are useful for economic evaluation and able to generate QALY. However, there is controversy about the definition of preference or utility scores. Utilities for some health states vary widely among individuals, as well as by the format o f health state descriptions, the way the outcomes are framed and the scaling task itself (Revicki, 1992). Other disadvantages include the cognitive complexity o f the measurement task, understandability of the scores, and time and labor intensiveness (Torrance, 1987; Feeny, 1989; Mulley, 1989). 73 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Several studies also pointed out that the utility scores are not as sensitive as non preference-based HSM instruments to subtle change in clinical status (Revicki, 1992). Considering the limitations discussed above, it seems the only solution is to use both preference and non preference-based HSM instruments. Despite that, very few clinical investigations use both the health status and health utility instruments. In summary, non preference-based HSM instruments are far more widely used in health service research for a number of reasons that we discussed in earlier chapters. In this context, the present research tries to determine whether it is possible to develop a link function between the two different HSM instruments. Findings o f this study suggest that it is possible to develop a mapping algorithm that can utilize the potentially rich source of descriptive health status information in an economic evaluation. It will eventually be possible to generate economic evaluation alongside any clinical trial using HRQOL measurements. Limitations and Areas for Future Investigation There are several limitations in this research, which need a thorough discussion. Some of these issues also deserve further investigation. The most obvious limitation o f the mapping technique is the prediction power. The models in this study can explain half o f the variation in the true utility score. 74 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. However, reported reliability o f the SF-36 (of all domains) and the HUI in this sample were 0.88 and 0.65. This implies that the variance that might be accounted for in the HUI (assuming a perfect relationship (R2 = l) with a single domain of the SF-36 whose measurement reliability was 0.8) is roughly (0.65)*(0.8)=0.52 or 52%. In the case of a multivariate setting (using more than one domain as predictors), the percentage o f accountable variance will increase but can never reach at 100%. Thus it seems plausible to conclude that one of the reasons behind the deviations between observed and predicted utility from our equation is the measurement error in the instruments (HUI and SF-36). Similar analogy also holds for the LAS and the SF-36. It should be recognized also that this study only included samples from Southern California, USA. However, all the patients in this study had similar scores on all domains of the SF-36 compared to the SF-36 scores obtained from the US general population and socio-demographically this sample is quite comparable to the US population. Simultaneously all the regression equations in this study are quite robust. We did not notice any paradoxical results unlike the Beaver Dam study (Fryback et al. 1997). All the regression coefficients of the SF-36 domains (except RE in the cross-sectional time series model) are positive and consistent with the pair-wise correlation matrix. Despite all the facts this study has no external validation and we suggest empirical specification of the present model 75 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. need to be validated by different sample to determine generalizability. One of the major empirical concerns o f this study is the domain score of the SF-36 and the items of the SF-12. The domain scores of the SF-36 were generated on the assumption that responses in the items of the instrument are of equal interval. However, this is a strong assumption and there is little empirical evidence to support or reject this. Later on, the SF-12 physical and mental component were also developed based on same assumption. However, Lundberg et al. (1999) and Brazier et al. (1998) criticized this concept and used a categorical variable approach for each response item in their algorithm. This area definitely needs further investigation, so that some empirical evidences can be generated for future research in this field. Another statistical concern is the modeling approach in cross-sectional time series algorithm. The present data sets have only three different time-periods and because of this, the scope of an efficient model is very limited. Although the estimates were adjusted for auto-correlated error terms, it could not achieve the efficiency for presence o f heteroscedasticity. Existing statistics or econometric techniques were also not helpful in resolving the problem. Thus, it would be useful to further delve in this issue and obtain some empirical solution. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Since main objective o f this study was to establish some mapping algorithm that can generate preference score from health status score, we did not explore other interesting and relevant issues. One such issue is the effect o f heterogeneity in estimation (due to socio-demographic and disease variables). Though all the predictive models in the study were controlled for these factors, our study was unable to capture any significant influence of heterogeneity among patients on their health state preferences. However, some literature has suggested that use of “Latent Class Approach” or “Finite Mixture Model” might be a more efficient way to handle this issue (Bult et al., 1998). To summarize, findings of this study are consistent with the relevant work that has been done in the last few years to transform SF-36 (SF-12) scores or items to the QWB, the TTO and the SG. Furthermore, we provide some insights about the effect o f endogenity in estimation procedures. The methodological framework presented in this study provides researchers with a tool to obtain an estimate of summary utility scores from secondary health status data using the SF-36 or the SF-12. This in turn can be employed to estimate the cost/QALY in cost- effectiveness analyses. Further research to validate this approach in other population and overcome its limitations, will benefit economists, health service researchers and policy makers alike. 77 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. BIBLIOGRAPHY Anderson JP, Kaplan RM, Berry CC, Bush JW, Rumbaut RG. Interday reliability o f function assessment for a health status measure-the quality o f well-being scale. Med Care 1989;27:1076-84. Barr RD, Furlong W, Dawson S, Whitton AC, Strautmanis I, Pai M, et al. An assessment of global health status in survivors of acute lymphoblastic leukaemia in childhood. Am J Pediatr Hematol Oncol 1993;15:284-290. Belsley DA, Kuh E, Welsch RE. Regression Diagnostics, New York, NY: John Wiley & Sons, Inc. 1980. Bombardier C, Wolfson AD, Sinclair AJ, McGeer A. Comparison of three preference measurement methodologies in the evaluation of a functional status index. In R. Deber & G. Thompson (eds.), Choices in health care: Decision making and evaluation o f effectiveness. Toronto: University of Toronto, 1982,145-59. Bosch JL, Hunink MGM. The Relationship between Descriptive and Valuational Quality-of-life Measures in Patient with Intermittent Claudication. Med Decis Making 1996;16:217-225. Boyle MH, Furlong W, Feeny D, Torrance GW, Hatcher J. Reliability o f the Health Utilities Index-Mark III used in the 1991 cycle 6 Canadian General Social Survey Health Questionnaire. Q Life Res 1995;4:249-57. Bozzette SA, Hays RD, Berry SH, Kanouse DE. A perceived health index for use in persons with advanced HIV disease: Derivation, Reliability, and Validity. Med Care 1994;32(7):716-731. Brazier J, Deverill M, Green C, Harper R, Booth A. A review o f the use o f health status measures in economic evaluation. Health Technology Assessment 1999;3(9): 1-161. Brazier J, Usherwood T, Harper R, Thomas K. Deriving a Preference Based Single Index from the UK SF-36 Health Survey. J Clin EpidemioI.1998;51(l 1): 1115-28. Brazier J, Walter SJ, Nicholl JP, Kohler B. Using the SF-36 and EuroQoL on an elderly population. Q Life Res 1996;5:195-204. 78 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Brazier J, Jones N, Kind P. Testing the validity of the EuroQoL and comparing it with the SF-36 health survey questionnaire. Q Life Res 1993;2:169-80. Buckinghum JK, Birdsall J, Douglas JG. Comparing three versions of the time tradeoff: time for a change? Med Decis Making 1996;16:335-47. Bult JR, Hunink MGM, Tsevat J, Weinstein MC. Heterogeneity in the relationship between the time trade-off and Short Form 36 for HIV infected and primary care patients. Med Care 1998;36(4):523-532. Caims J, Johnston K, McKenzie L. Developing QALYs from condition-specific outcomes measures. HERU discussion paper 14/1991. Aberdeen: University of Aberdeen. Cohen BB, Barbano HE, Cox CS, et al. Plan and operation of the NHANES 1 Epidemiologic Followup Study, 1982-84. Vital and health statistics, series 1, no.22 DHEW publication no.(PHS) 87-1324. Washington,DC: US Government Printing Office, 1987. Complete Guide to SAS Macro, Cary, NC: SAS Institute Inc. 1997. Cook J, Richardson J, Street A. A cost-utility analysis of treatment options for gallstone disease- methodological issues and results. Health Economics 1994;3:157-68. Cronbach Lf. Coefficient alpha and the internal structure o f tests. Psychometrika 1951;16:297-334. Dolan P, Sutton M. Mapping visual Analog scale health state valuation onto standard gamble and time-tradeoff values. Soc Sci Med 1997;44(10): 1519-1530. Drummond MF, Davies L. Economic analysis alongside clinical trials: revising the methodological issues. Int J Technol Assess Health Care 1991;7:561-573. Du an N, Manning WG, Morris CW, Newhouse JP. A comparison of alternative models for the demand for medical care. Journal of Business & Economic Statistics. 1993;1(2):115-126. Durbin J. Testing for serial correlation in least square regression when some o f the regressors are lagged dependent. Econometrica 1970;38:410-421. 79 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. EuroQoL Group. EuroQoL-a new facility for the measurement of health-related quality-of-life. Health Policy 1990;16:199-208. Feeny DH, Torrance GW. Incorporating utility based quality of life assessment in clinical trial. Med Care 1989;27:S 190-204. Feeny D, Leiper A, Barr RD, Furlong W, Torrance GW, Rosenbum P, Weitzman S. The comprehensive assessment o f health status in survivors of childhood cancer: application to high risk acute lymphoblastic leukaemia. Br J Cancer 1993;67:1047-52. Feeny D, Furlong W, Boyle M, Torrance GW. Multiattribute health status classification systems. Health Utility Index. PharmacoEconomics 1995;7:490-502. Feeny D, Furlong W, Barr RD. Multiattribute approach to the assessment o f health related quality of life: Health Utility Index. Medical and Pediatric Oncology 1998;Suppl 1:54-9. Froberg DG, Kane RL. Methodology for measuring health-state preferences-11: scaling methods. J Clin Epidemiol 1989a;42:459-71. Fryback DG, Lawrence WF, Martin PA, Klein R, Klein BEK. Predicting Quality of Well-being scores from the SF-36:Results from the Beaver Dam Health Outcomes Study. Med Decis Making. 1997;17:1-9. Gafni A. The standard gamble method: what is being measured and how it is interpreted. Health Serv Res 1994;29:207-24. Gafhi A, Birch S. Preferences for outcomes in economic evaluation: an economic approach to addressing economic problems. Soc. Sci. Med 1995;40:767-76. Garber AM. Phelps CE. Economic foundation of cost-effectiveness analysis. J Health Econ. 1997; 16:1-31. Gold MR, Siegel JE, Russell LB, Weinstein MC. Cost-Effectiveness in Health and Medicine. New York, NY: Oxford University Press, 1996. Gudex C, Dolan P, Kind P, Williams A. Health state valuations from the general public using the visual Analog scale. Q Life Res 1996;5:521-531. Gudex C. QALYs and their use by the health service. Discussion paper 20. York: Center for Health Economics. University o f York, 1986. 80 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Hausman J. Specification Tests in Econometrics. Econometrica. 1978;46:1251-71. Hays RD, Sherboume CD, Mazel RM: The RAND 36 item health survey 1.0 Health Economics. 1993;2:217-227. Hunt SM, McKenna SP, McEwen J, et al. The Nottingham Health Profile: subjective health status and medical consultations. Soc Sci Med 1981;15:221-229. Hunt SM, McEwen J, McKenna SP. Measuring health status. London: Croom Helm, 1986. Hurst NP, Jobanputra P, Hunter M, Lambert M, Lochhead A, Brown H. Validity o f EuroQoL-a generic health status instrument in patients with rheumatoid arthritis. Economics and Health Outcomes Research Group. Br J Rheumatol 1994;33:655-62. Johannesson M, Jonsson B, Karlson G. Outcome measurement in economic evaluation. Health Econ 1996;5:279-296. Kanabar DJ, Attard-Montalto S, Saha V, Kingston JE, Malpas JE, Eden OB. Quality of life in survivors of childhood cancer after megatherapy with autologus bone marrow rescue. Pediatric Hematology and Oncology 1995; 1 2( 1 ):29-36. Kaplan RM, Bush JW. Health-related quality o f life measurement for evaluation research and policy analysis. Health Psychol 1982;1:61-80. Loomes G. Disparities between health state measures: An explanation and some implications. York, UK: Department of Economics, University of York, 1988. Lundberg L, Johannesson M, Isacson DGL, Borgquist L. The relationship between health-state utilities and the SF-12 in a general population. Med Decis Making 1999;19:128-140. MacKinnon JG, White H, Davidson R. Teat for Model-Specification in the Presence o f Alternative Hypotheses. Journal o f Econometrics 1983;21:53-70. Maddala GS, Lee LF. Recursive models with qualitative endogenous variables. Ann Econ Soc Meas. l976;5(4):525-45. Mas-collel A, Whinston MD, Green JR. Microeconomic Theory, New York, NY: Oxford University Press, 1995. 81 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. McCombs JS, Cody M, Besinique K et al. Measuring the impact of patient counseling in the outpatient pharmacy setting: The research design of the Kaiser Permanente/USC patient consultation study. Clinical Therapeutics 1995;17(6):1188-1206. McHomey CA, Ware JE, Raczek AE. The MOS 36-item short-form health survey (SF-36):II. psychometric and clinical tests of validity in measuring physical and mental health constructs. Med Care 1993;31(3):247-63. Mckie J, Kuhse H, Richardson J, Singer P. Allocating healthcare by QALYs: the relevance of age. Cambridge Quarterly o f Healthcare Ethics 1996;5(4):534-45. Meherez A, Gafni A. Health-years equivalents versus quality-adjusted life years: in pursuit o f progress. Med Decis Makingl993;13:287-92. Meherez A, Gafhi A. The healthy-years equivalents: how to measure them using the standard gamble approach. Med Decis Making 1991; 11:140-46. Miller HW. Plan and operation of the Health and Nutrition Examination Survey. United States, 1971-73. Vital and health statistics, series 1, no. 10a, lOb.DHEW publication no.(PHS) 73-1310. Washington,DC: US Government Printing Office, 1977. Mulley AG Jr. Assessing patients’ utilities. Can the ends justify the means? Med Care 1989;27:S269-81. Nerenz DR, Repasky DP, Whitehouse FW, Kahkonen DM. Ongoing assessment of health status in patients with diabetes mellitus. Med Care 1992; 30(5 Suppl):MS 112-24. Nichol G, Theil E, Llewellyn-Thomas HA, et al. Do patients with angina report utilities that are consistent with their level of functional impairment on standard scales? Medical Decision Making. 1992; 12:352. Nord E. The validity of a visual analog scale in determining social utility weights for health states. Inti J of Health Planning Manage 1991 ;6:234-42. Nord E. Methods for quality adjustment o f life years. Soc Sci Med 1992a;34:559-569. 82 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. O’Brien BJ, Buxtonn MJ, Ferguson BA. Measuring the effectiveness of heart transplant programmes: quality of life data and their relationship to survival analysis. J Chronic Dis 1987;40(Suppl I): 137S-58S. O’Leary JF, Fairclough DL, Jankowski MK, Weeks JC. Comparison of Time- tradeoff utilities and rating scale values o f cancer patients and their relatives: Evidence for a possible plateau relationship. Med Decis Making 1995; 15:132-137. Patrick DL, Erickson P. Health Status and Health Policy: Allocating Resources to Health Care. New York, NY: Oxford University Press, 1993. Ramsey JB. Test for specification errors in classical linear least-squares regression analysis. Journal of the Royal Statistical Society. 1969;B31:350-371. Revicki DA, Brown RE, Henry D. Health Status Outcomes of Recombinant Human Erythropoietin Therapy for Predialysis Chronic Renal Disease Patients. Washington, DC: Battelle Medical Technology Assessment and Policy Research Center, 1993. Revicki DA, Kaplan RM. Relationship between psychometric and utility-based approaches to the measurement of health related quality of life. Quality of Life Research. 1993;2:477-487. Revicki DA, Weinstein M, Alderman M, et al. Health Utility and Health States Outcomes o f Antihypertensive Treatment. Washington, DC: Battelle Medical Technology Assessment and Policy Research Center, 1992. Richardson J. Cost utility analysis: what should be measured? Soc Sci Med 1994;18:227-8. Sackett DL, Torrance GW. The utility o f different health states as perceived by the general public. J Chronic Dis 1981;31:697-704. SF-36 Latest News 1999; http://www.sf-36.com/news. Spilker B, Molinek FR, Johnston KA et al. editors. Quality of life bibliography and indexes. Med Care 1990;28:DS1-DS77. State of California, Department o f Finance, Race/Ethnic Population Estimates: Components o f Change for California Counties, April 1990 to July 1997. http://www.dof.ca.gov/html/ 83 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Stiggelbout AM, Eijkemans MJC, Kiebert GM et al. The utility of the visual analog scale in medical decision making and technology assessment. Is it an alternative to the time trade-off? International Journal o f Technology Assessment in Health Care 1996; 12(2):291-298. Torrance GW. Social preferences for health states: An empirical evaluation o f three measurement techniques. Socio Economic Planning Science 1976;10:129-36. Torrance GW, Boyle MH, Horwood SP. Applications o f multi-attribute utility theory to measure social preferences for health states. Operations Res 1982;30:1043-69. Torrance GW. Measurement of health state utilities for economic appraisal. J Health Econ. 1986;5:1-30. Torrance GW, O’Brien B. An interview on utility measurement. J Rheumatol 1995;22:1200-2. Torrance GW, Erickson P, Patrick DL, et al. Technical notes.In: Erickson P, Wilson R, Shannon I. editors. Years o f healthy life:statistical notes, number 7. Hyattesville (MD): National Center for Health Statistics, 1996. Torrance GW, Feeny DH, Furlong WJ, Barr RD, Zhang Y, Wang Q. Multiattribute Utility Function for a Comprehensive Health Status Classification System. Health Utility Index Mark 2. Med Care 1996;34:702-22. U.S. Bureau o f Census, Current Population Reports Series pp.23-189. Population Profile of the United States. July, 1995. U.S. Government Printing Office, Washington, DC, 1995. Von Neumann J, Morgenstem O: Theory o f Games and Economic Behaviour, Princeton, NJ: Princeton University Press, 1947. VonKorfFM, Wagner EH, Saunders K. A Chronic Disease Score from the Automated Pharmacy Data. J o f Clin Epidemiol 1992;45:197-203. Walker G. On Periodicities in Series o f Related Terms. Procedings of the Royal Society 1931 ;A131:518- Ware JE. Patient based Assessment o f Health Outcomes. Paper presented at AHSR 13th Annual Meeting, Adanta, GA, 1996 84 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Ware JE, Kosinski M, Keller SD. A 12-item Short-Form Health Survey. Construction o f scales and preliminary test o f reliability and validity. Med Care 1996;34(3):220-233. Ware JE, Kosinski M, Keller SD. SF-36 physical and mental health summary scores: a user’s manual. Boston, MA: The Health Institute, New England Medical Center, 1994 Ware JE, Sherboume CD. The MOS 36-item Short-Form Health Survey (SF-36) Conceptual Framework and Item Selection. Med Care 1992;30:473-483. White H. A heteroscedasticity-consistent covariance matrix estimator and a direct test for heteroscedasticity. Econometrica 1980b;48:817-838. Williams A. ‘Should QALYs be programme specific?’ by Donaldson, Atkinson, Bond, Wright. J Health Econ 1989;8:485-7(discussion:489-91). Wu De-Min. Alternative Teats of Independence between Stochastic Regression and Disturbance. Econometrica I973;4I(3):733-50. Yule GU. On a Method of Investigating Periodicities in Disturbed Series, with Special Reference to Wolfer’s Sunspot Numbers. Philosophical Transactions 1927;A226:267. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Controlling for biases from measurement errors in health outcomes research: A structural equation modeling approach
PDF
A methodology to identify high -risk patients with diabetes in the California Medicaid populations (Medi -Cal)
PDF
Effects of a formulary expansion of the use of SSRIs and health care services by depressed patients in the California Medicaid program
PDF
Assessing the cost implications of combined pharmacotherapy in the long term management of asthma: Theory and application of methods to control selection bias
PDF
Compliance study of second-generation antipsychotics on patients with schizophrenia
PDF
The construct validity of the Health Utilities Index in patients with chronic respiratory disease in a managed care population
PDF
Assessment of prognostic comorbidity in hospital outcomes research: Is there a role for outpatient pharmacy data?
PDF
Biopharmaceutical strategic alliances: Interorganizational dynamics and factors influencing FDA regulatory outcomes
PDF
A study of employee health plan choice and medical cost: Panel data probit regression and sample selection model
PDF
An intervention and program evaluation to determine the effectiveness of public health reforms on primary prevention practices by chiropractic interns
PDF
Essays on organizational forms and performance in California hospitals
PDF
Dialectical feminism and contradictory economic systems: Campesina daily life from Somoza to Aleman
PDF
Cost analysis of three pharmacy counseling programs for diabetics in a health maintenance organization
PDF
Imputation methods for missing items in the Vitality scale of the MOS SF-36 Quality of Life (QOL) Questionnaire
PDF
Physician adherence to national hypertension guidelines in an elderly Medicaid population
PDF
Health care finance reform and market failure at King/Drew Medical Center, Los Angeles: An analysis of pediatric social outcomes in an urban public safety -net hospital
PDF
Computational prediction of peptide-MHC class I binding interactions
PDF
Health-related quality of life, occupation and prosthesis use in elderly people with lower extremity amputation in Taiwan
PDF
An analysis of health risk selection and quality of care under Medicare fee -for -service and Medicare managed care health care systems
PDF
Drug delivery to the posterior region of the eye: Characterization of organic cation transporters
Asset Metadata
Creator
Sengupta, Nishan
(author)
Core Title
A new paradigm to evaluate quality-adjusted life years (QALY) from secondary database: Transforming health status instrument scores to health preference
School
Graduate School
Degree
Doctor of Philosophy
Degree Program
Pharmaceutical Economics and Policy
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
health sciences, health care management,Health Sciences, Pharmacy,health sciences, public health,OAI-PMH Harvest,psychology, psychometrics
Language
English
Contributor
Digitized by ProQuest
(provenance)
Advisor
Nichol, Michael (
committee chair
), Globe, Denise (
committee member
), Groshen, Susan (
committee member
)
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c16-135526
Unique identifier
UC11328641
Identifier
3041522.pdf (filename),usctheses-c16-135526 (legacy record id)
Legacy Identifier
3041522-0.pdf
Dmrecord
135526
Document Type
Dissertation
Rights
Sengupta, Nishan
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Access Conditions
The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the au...
Repository Name
University of Southern California Digital Library
Repository Location
USC Digital Library, University of Southern California, University Park Campus, Los Angeles, California 90089, USA
Tags
health sciences, health care management
Health Sciences, Pharmacy
health sciences, public health
psychology, psychometrics