"Stated Preference Theory (Conjoint Analysis) is the Best Way to Assess Health-Related Quality of Life for Economic Assessment of Drugs and Medical Interventions: An Application in Alzheimer Disease" Joel Hay, PhD Department of Pharmaceutical Economics and Policy University of Southern California April 21 2006 Presented at ISPOR Student Teleconference Research supported by the State of California Alzheimer Disease Research Centers Network JHay@usc.edu 1 Overview Measuring Health Related QOL for Cost Effectiveness Analysis Problems with CE ratios in making decisions Justifications for Net Monetary Benefits based on willingness to pay JHay@usc.edu 2 Overview Introduction to Conjoint Analysis Alzheimer Conjoint Experiment Design Results for the Alzheimer CA Experiment Conclusions & Future Directions JHay@usc.edu 3 Stated Preference Elicitation Methods Important to the design of successful drugs and other interventions Can reduce R&D costs enormously by focusing on treatment and disease symptoms that are most important Can be used to weigh utility of outcomes in cost-effectiveness analysis JHay@usc.edu 4 Stated Preference Methods Willingness to Pay/ Stated Value Standard Gamble / Time Trade-off Global Assessment / Visual Analog Health Status Assessment / Multiattribute Scales Conjoint Analysis/ Discrete Choice Experiments JHay@usc.edu 5 Stated Preference Methods Stated preference is not revealed preference All stated preference methods ask the subject to make hypothetical choices All stated preference methods require the subject (patient) to consider or imagine the value of health states they don’t actually experience JHay@usc.edu 6 Denominator of CEA Ratio [Cost(A) - Cost(B)] / [QALY(A) - QALY(B)] How are QALYs determined? Patient or subject survey of health states EQ-5D, HUI, QWB, SF-## Results converted into QOL-weighted time Uses someone’s utility weighting algorithm Utility translation imprecision is generally ignored JHay@usc.edu 7 Problems with QALYs Doesn’t accurately reflect individual utilities Implies that people have constant preference for healthy years Ignores risky behaviors Ignores life expectancy constraints Ignores alternative consumption opportunties Creates serious analytic problems in a CE ratio 8 JHay@usc.edu STOCHASTIC UNCERTAINTY / VARIABILITY 2000 2000 1500 1500 1000 Difference in costs Difference in costs • Feiller’s Theorem Probability Ellipses 500 0 -500 -1000 -1500 -2000 -0.25 1000 500 0 -500 -1000 -1500 -0.15 -0.05 0.05 0.15 0.25 -2000 -0.25 Difference in effectiveness -0.15 -0.05 0.05 0.15 0.25 Difference in effectiveness JHay@usc.edu 9 Analyzing CEA Ratio Results Computing 95% CI around mean ratio is plagued with many problems ΔC ΔE JHay@usc.edu 10 Analyzing CEA Ratio Results Computing 95% CI around mean ratio is plagued with many problems ΔC ΔE • CI can include undefined values • ΔC / ΔE • What if ΔE = 0? • CI = (LL, ∞) U (∞, UL) JHay@usc.edu 11 Analyzing CEA Ratio Results Computing 95% CI around mean ratio is plagued with many problems • CI can be undefined JHay@usc.edu 12 Analyzing CEA Ratio Results Computing 95% CI around mean ratio is plagued with many problems ΔC • The same value for a CER can have 2 completely different meanings b1 a2 ΔE a1 • a1 = a2 but many prefer a2 • b1 = b2 but in reality, b1 << b2 b2 JHay@usc.edu 13 Analyzing CEA Ratio Results Computing 95% CI around mean ratio is plagued with many problems ΔC a • Not properly ordered outside trade-off quadrants • b better than c • c better than a • a and b equal when in ΔE reality b >> a c b JHay@usc.edu 14 Analyzing CEA Ratio Results Acceptability Curves Incremental costs positive Quadrant 4 Inferior Quadrant 1 Trade off R Incremental effects negative Incremental effects positive Quadrant 3 Trade off Quadrant 2 Superior Incremental costs negative JHay@usc.edu 15 Acceptability Curves Cost Effectiveness Acceptability Curves Shows the probability that the new therapy will be cost effective as a function of the societal willingness-to pay (for a QALY) threshold Gets around the problems of CEA confidence intervals Main advantage: Adopts the natural perspective / interpretation of decision makers: “how likely is it that the intervention will be cost effective” JHay@usc.edu 16 Acceptability Curves P (Intervention is Cost effective) 1 Curve tend to 1-“p” value (one-sided) for effectiveness difference One-sided “p” value for cost (savings) difference Proportion of cases In which intervention is more effective (Quadrants 2+1) Proportion of cases in which the intervention is cost-savings (Quadrants 2+3) 0 $0 $25,000 $50,000 $75,000 Cost-effectiveness Threshold R JHay@usc.edu $100,000 $ 17 Acceptability Curves Analysts have had the natural tendency to interpret the results as the probability that the cost-effectiveness ratio is lower or equal than a specific threshold (given the data) AC are not always monotonically increasing! Their shape depends on the data – they can go up and down! 95% intervals cannot always be defined JHay@usc.edu 18 Using Net Monetary Benefits to Make Decisions The incr net benefit (INB) was introduced to estimate whether a treatment is cost effective INB = IC – λ *IE λ is decisionmaker willingness to pay for QALY A therapy, for which the INB is lower than zero, may be considered cost-effective However, even when acceptability curve indicates that new treatment is cost-effective at a λ the decision maker is willing to pay, there is still a probability that the decision to reimburse the new treatment is wrong JHay@usc.edu 19 Using Net Monetary Benefits to Make Decisions Incr Net Ben = Incr Cost – λ * Incr Effect λ reflects the decision maker’s willingness to pay for new therapy Why are we using subjects to generate HRQOL and QALYs but not using subjects to generate λ (willingness to pay for QALYs)? The CEA researcher is throwing away the opportunity to capture subject willingness to pay FOR NO REASON!! JHay@usc.edu 20 EXPECTED VALUE OF PERFECT INFORMATION Is quantifying uncertainty useful to decision makers? Focus should be on determining the value of collecting new information (= cost of uncertainty) New information valued in terms of expected reductions in decision errors (NMB loss in money) See Claxton K. The irrelevance of inference: a decision making approach to the stochastic evaluation of health care technologies Journal of Health Economics, 1999;18:341-64 JHay@usc.edu 21 Why not Capture Willingness to Pay Directly? Gives a more accurate and precise measure of subject trade-off between money and health states Doesn’t force health utilities into an artificial QOL framework Allows direct assessment of Net Benefits rather than artificial CEA ratios JHay@usc.edu 22 Why not Capture Willingness to Pay Directly? Even if data come from current HRQOL measures it’s possible to develop WTP conversion scales Need to subject existing HRQOL instruments to WTP assessment Conjoint Analysis is the way to do this JHay@usc.edu 23 Advantages of Conjoint Analysis Rigorous behavioral choice model: Random Utility Theory Actually estimates utility parameters Experimental design efficiently maps choice preferences in multi-attribute space Subjects can’t “game” responses JHay@usc.edu 24 Random Utility Maximization Uiq = Viq + iq Uiq > Ujq for all j i element of A Pi = Pr[(Viq - Vjq) > (jq - iq )] JHay@usc.edu 27 Stated Preference Elicitation in Alzheimer Disease AD treatments involve trade-offs across different functional domains Important, since AD patients have difficulty in making & expressing choices Relies on proxy responders, usually AD caregivers JHay@usc.edu 31 Stated Preference Elicitation in Alzheimer Disease Develop a CA experiment to elicit preferences across function domains for hypothetical AD patients Include treatment costs to estimate willingness-to-pay for improvements/decreases in daily function JHay@usc.edu 32 Stated Preference Elicitation in Alzheimer Disease Validate the CA design in a sample of pharmacy students Apply the CA experiment to a sample of AD caregivers Compare results JHay@usc.edu 33 CA Experiment Attributes & Levels from Health Utilities Instrument-- Mark 3 This instrument is widely used and validated It was developed to map into a standard gamble utility scale valid for cost effectiveness analysis Utility scores from the HUI can be compared to RUT scores JHay@usc.edu 34 JHay@usc.edu 35 Survey PILOT STUDY The CA design was validated in a sample of pharmacy students: -HUI functional domain attributes were shown to be significant and consistent predictors of choice - Total HUI utility score is much better predictor of choice than individual components JHay@usc.edu 36 Survey •DEMOGRAPHICS •Survey was administered to 74 AD caregivers enrolled in the California AD Research Centers. •STATED PREFERENCE SCENARIOS •HUI-3 for the AD PATIENT, as PROXIED by the CAREGIVER •HUI-3 for the CAREGIVER •SF-36 for the CAREGIVER JHay@usc.edu 37 Survey STATED PREFERENCE SCENARIOS Fractional Factorial Experimental Design Attributes and Levels taken from the Health Utilities Instrument - Mark 3 best – 3rd – worst 243 combinations of comparisons that were not dominant or dominated choices 25 x 10 choice sets each 61 x 4 choice sets each Cost of each choice = $100 / $50 / $0 JHay@usc.edu 38 JHay@usc.edu 39 JHay@usc.edu 40 Speech Ambulation Dexterity JHay@usc.edu 41 JHay@usc.edu 42 JHay@usc.edu 43 JHay@usc.edu 44 Scoring the HUI • Patient HUI as proxied by the CG • Caregiver HUI • HUI of each choice set Multi-attribute utility function: u = 0.371(b1*b2*b3*b4*b5*b6*b7*b8)-0.371 JHay@usc.edu 45 Analysis Plan • Demographics • MNL for HUI, cost in predicting choice • Strength of each attribute in determining subject choice • Subject WTP for various choice attributes – determining strength of cost attribute • Estimate subject utility - Health utility of AD patients - Mapping of the HUI-3 to the SF-36 for AD JHay@usc.edu 46 Results: Descriptives Mean Patient HUI Caregiver HUI SF-36 Physical Function SF-36 Physical Role Function SF-36 Bodily Pain SF-36 General Health SF-36 Vitality SF-36 Social Functioning SF-36 Emotional Role Function SF-36 Mental Health Caregiver Age Caregiver Years of Education Caregiver duration of caregiving since disease onset (months) Caregiver hours per week Caregiver is Spouse of Pt Caregiver is Daughter of Pt Caregiver is Son of Pt Working Full Time Retired Primary Caregiver Married White Hispanic Asian Black Gender (1=male) JHay@usc.edu 0.35 0.88 82.00 76.47 74.09 76.32 63.34 80.69 73.49 76.68 60.79 15.31 82.70 Std. Deviation 0.38 0.13 20.68 33.84 22.59 16.54 18.76 20.87 34.78 15.37 14.45 3.68 136.53 23.76 0.40 0.28 0.17 0.42 0.30 0.76 0.73 0.45 0.31 0.11 0.09 0.35 32.21 0.49 0.45 0.38 0.49 0.46 0.43 0.44 0.50 0.46 0.32 0.29 0.48 47 Results: Correlations Patient HUI Patient HUI Pearson Correlation Caregiver HUI 1 Sig. (2-tailed) Caregiver HUI Pearson Correlation Sig. (2-tailed) Caregiver SF-36 Physical Function Pearson Correlation Sig. (2-tailed) Caregiver SF-36 General Health Pearson Correlation Sig. (2-tailed) Caregiver SF-36 Mental Health Pearson Correlation Sig. (2-tailed) 0.303 Caregiver SF-36 Physical Function Caregiver SF36 General Health 0.303 0.097 -0.003 0.115 0.00 0.07 0.96 0.03 1 0.528 0.356 0.335 0.00 0.00 0.00 1 0.401 0.216 0.00 0.00 1 0.511 0.00 0.097 0.528 0.07 0.00 -0.003 0.356 0.401 0.96 0.00 0.00 0.115 0.335 0.216 0.511 0.03 0.00 0.00 0.00 JHay@usc.edu Caregiver SF-36 Mental Health 0.00 1 48 Dependent Variable: Patient HUI Results: Regression (Constant) Caregiver SF-36 Physical Function Caregiver SF-36 Physical Role Function Caregiver SF-36 Bodily Pain Caregiver SF-36 General Health Caregiver SF-36 Vitality Caregiver SF-36 Social Functioning Caregiver SF-36 Emotional Role Function Caregiver SF-36 Mental Health Age Years of Education CG duration of caregiving since disease onset (months) CG hours per week Caregiver is Spouse of Pt Caregiver is Daughter of Pt Caregiver is Son of Pt Working Full Time Retired Primary_CG Married White Hispanic Asian Black Gender Caregiver HUI JHay@usc.edu Unstandardized Coefficients B Std. Error 0.226 0.197 -0.005 0.001 -0.004 0.001 Sig. 0.253 0.000 0.000 0.005 -0.005 0.003 0.003 0.001 0.001 0.001 0.002 0.001 0.001 0.000 0.000 0.070 0.015 0.056 -0.005 -0.002 -0.004 0.000 0.002 0.002 0.005 0.000 0.002 0.344 0.504 0.528 -0.003 0.176 0.033 0.197 0.012 0.129 0.095 -0.234 -0.119 -0.395 -0.004 -0.395 -0.078 1.261 0.001 0.083 0.064 0.086 0.050 0.059 0.054 0.057 0.090 0.090 0.101 0.111 0.058 0.180 0.000 0.034 0.605 0.022 0.812 0.029 0.078 0.000 0.185 0.000 0.966 0.000 0.181 0.000 49 Results •In a multinomial logistic regression, treatment choice was positively related to HUI score for the chosen intervention and negatively related to treatment costs (P < 0.01). •At the margin, caregivers would be willing to spend an additional $5-$7 per month for any AD intervention that increased patient HUI utility scores by 1%. •The strongest rankings include improvements in ambulation, emotion and cognition. JHay@usc.edu 50 Results •While HUI scores are related to SF-36 scores, the magnitude of response is fairly small. •Neither HUI score is related to SF-36 Vitality and the Caregiver HUI score is not significantly related to Physical Role Function. •Regression based on the HUI multi-attribute score was superior to any single-attribute model and non-inferior to models allowing unconstrained parameter weights for each HUI attribute domain. JHay@usc.edu 51 Conclusions •Conjoint Analysis is a useful method for benchmarking the potential values for AD treatments trade-offs in terms of their costs and impacts on patient functioning. •As found in prior studies, HUI is a useful scale for characterizing proxy patient utility levels. JHay@usc.edu 52 Conclusions •Consistent with utility theory, results show a methodologically independent and innovative validation of the HUI utility scale as a strong predictor of subject health state preferences. •Demonstrates that we can convert HUI, SF-36 and other HRQOL scales directly into willingness to pay values for alternative health states •Do not need to increase imprecision and difficulty by using Cost Effectiveness Ratios JHay@usc.edu 53 Conclusions •These methods will become even more relevant as we increasingly utilize brain scanning methods to map utility of health states and utility of money directly •Functional Magnetic Resonance Imaging •Positron Emission Tomography JHay@usc.edu 54 SUGGESTED READINGS Handling Variability and Uncertainty (1) Briggs AH, Fenn P. Confidence intervals or sufaces? Uncertainty on the costeffectiveness plane. Health Econ 1998; 7:723-740. Briggs AH. Handling uncertainty in cost-effectiveness models. Pharmacoeconomics 2000; 17(5):479-500. Clark DE. Computational methods for probabilistic decision trees. Comput Biomed Res 1997; 30(1):19-33. Critchfield GC, Willard KE, Connelly DP. Probabilistic sensitivity analysis methods for general decision models. Comput Biomed Res 1986; 19(3):254265. Dittus RS, Roberts SD, Wilson JR. Quantifying uncertainty in medical decisions. J Am Coll Cardiol 1989; 14(3 Suppl A):23A-28A. Doubilet P, Begg CB, Weinstein MC, Braun P, McNeil BJ. Probabilistic sensitivity analysis using Monte Carlo simulation. A practical approach. Med Decis Making 1985; 5(2):157-177. Glick HA, Briggs AH, Polsky D. Quantifying stochastic uncertainty and presenting results of cost-effectiveness analyses. Expert Rev Pharmacoeconomics Outcomes Res 2001; 1(1):25-36. 55 JHay@usc.edu SUGGESTED READINGS Handling Variability and Uncertainty (2) Halpern EF, Weinstein MC, Hunink MG, Gazelle GS. Representing both first- and second-order uncertainties by Monte Carlo simulation for groups of patients. Med Decis Making 2000; 20(3):314-322. Shaw JW, Zachry WM. Application of probabilistic sensitivity analysis in decision analytic modeling. Fomulary 2002; 37:32-40. Stinnett AA, Paltiel AD. Estimating CE ratios under second-order uncertainty: the mean ratio versus the ratio of means. Med Decis Making 1997; 17(4):483-489. Stinnett AA, Mullahy J. Net health benefits: a new framework for the analysis of uncertainty in costeffectiveness analysis. Med Decis Making 1998; 18(2 Suppl):S68-S80. Whang W, Sisk JE, Heitjan DF, Moskowitz AJ. Probabilistic sensitivity analysis in cost-effectiveness. An application from a study of vaccination against pneumococcal bacteremia in the elderly. Int J Technol Assess Health Care 1999; 15(3):563-572. Fenwick E, O'Brien BJ, Briggs A. Cost-effectiveness acceptability curves - facts, fallacies and frequently asked questions. Health Econ. 2004 May;13(5):405-15. Recommended Books Hunink M, Glasziou P, Siegel J, et al (2001). Decision Making in Health and Medicine. Integrating Evidence and Values. Cambridge, UK: Cambridge University Press. Drumond M, McGuire A (Eds) (2001). Economic Evaluation in Health Care: Merging Theory with Practice. New York, NY: Oxford University Press. 56 JHay@usc.edu SUGGESTED READINGS Expected Value of Perfect Information Claxton, K., Sculpher, M. and Drummond, M. (2002) A rational framework for decision making by the National Institute For Clinical Excellence (NICE) . Lancet 360, 711-716. Claxton, K.C., Neuman, P.J., Araki, S.S. and Weinstein, M.C. (2001) The value of information: an application to a policy model of Alzheimers disease. International Journal of Technology Assessment in Health Care 17 (1): 38-55. Claxton, K. (1999a) Bayesian approaches to the value of information: Implications for the regulation of new pharmaceuticals. Health Economics 8, 269-274. Conjoint Analysis Hay JW: Conjoint Analysis in Pharmaceutical Research. J Managed Care Pharmacy 8:206-9, 2002. Chiou C-F; Hay J; Wallace J, Bloom B; Neumann P, Sullivan S, Yu H-T, Keeler E, Henning J, Ofman J. Development and Validation of A Grading System for the Quality of Cost-Effectiveness Studies. Medical Care 2003; 41:32–44. Dwight-Johnson M, Largomarsino I, Aisenberg E, Hay J. “Understanding Depression Treatment Preferences among Low-Income Latinos using Conjoint Analysis”. Psychiatric Services 2004;55(8):934-37. Johnson FR. Einstein on Willingness to Pay per QALY:Is There a Better Way? Med Decis Mak 2005; 607-8. Recommended CA Book Louviere JJ, Hensher DA, JD Swait: Stated choice methods : analysis and applications. Cambridge, UK, New York, Cambridge University Press, 2000. JHay@usc.edu 57