Generalizability of Patient Preference Study Results: A Weight and Test Approach
Author(s)
Divya Mohan, PhD1, Marco Boeri, BSc, MSc, PhD1, Joshua Coulter, MA2, Byron Jones, PhD3, Brett Hauber, PhD4.
1OPEN Health, London, United Kingdom, 2Pfizer, Inc, Grand Rapids, MI, USA, 3Novartis, Basel, Switzerland, 4Pfizer, New York, NY, USA.
1OPEN Health, London, United Kingdom, 2Pfizer, Inc, Grand Rapids, MI, USA, 3Novartis, Basel, Switzerland, 4Pfizer, New York, NY, USA.
OBJECTIVES: Despite approaches to ensure the representativeness of samples in patient preference studies (PPS), a common limitation is that results may not be generalizable. This study used data from an existing PPS to assess the feasibility of weighting a study sample such that the results are applicable to a different target population.
METHODS: Data from a previous PPS in osteoarthritis with six attributes (symptom control, 3 adverse events, mode/ frequency of administration and out-of-pocket cost), were weighted using two methods - post stratification and iterative proportional fitting. Weights were calculated based on 4 sample characteristics - age, gender, race and time since osteoarthritis diagnosis. Weights were applied in three scenarios - the US and UK samples were each weighted to reflect a clinical trial cohort as the target population and the US sample was weighted to reflect the UK sample. Weighting was applied (a) to the sample before estimating a mixed logit model, and (b) after estimation using individual conditional parameters generated by the mixed logit model. Model coefficients, relative attribute importance, and trade-off estimates from the baseline, pre-weighted, and post-weighted mixed logit models were compared for each scenario.
RESULTS: The sample and target distributions of the four sample characteristics used in this study were similar in gender and age, and different in race and time since diagnosis. The means of the results from both the pre- and post-estimation weighted models were numerically slightly different from those of the unweighted models in all three scenarios. However, these differences were numerically small and statistically insignificant (p>0.05), with the 95% confidence interval overlapping for all estimates.
CONCLUSIONS: The findings suggest that sample weighting to apply the results from a sample to a different target population is operationally feasible. Additional research is needed to confirm that sample weighting yields results that are truly generalizable.
METHODS: Data from a previous PPS in osteoarthritis with six attributes (symptom control, 3 adverse events, mode/ frequency of administration and out-of-pocket cost), were weighted using two methods - post stratification and iterative proportional fitting. Weights were calculated based on 4 sample characteristics - age, gender, race and time since osteoarthritis diagnosis. Weights were applied in three scenarios - the US and UK samples were each weighted to reflect a clinical trial cohort as the target population and the US sample was weighted to reflect the UK sample. Weighting was applied (a) to the sample before estimating a mixed logit model, and (b) after estimation using individual conditional parameters generated by the mixed logit model. Model coefficients, relative attribute importance, and trade-off estimates from the baseline, pre-weighted, and post-weighted mixed logit models were compared for each scenario.
RESULTS: The sample and target distributions of the four sample characteristics used in this study were similar in gender and age, and different in race and time since diagnosis. The means of the results from both the pre- and post-estimation weighted models were numerically slightly different from those of the unweighted models in all three scenarios. However, these differences were numerically small and statistically insignificant (p>0.05), with the 95% confidence interval overlapping for all estimates.
CONCLUSIONS: The findings suggest that sample weighting to apply the results from a sample to a different target population is operationally feasible. Additional research is needed to confirm that sample weighting yields results that are truly generalizable.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
PCR103
Topic
Economic Evaluation, Methodological & Statistical Research, Patient-Centered Research
Disease
Musculoskeletal Disorders (Arthritis, Bone Disorders, Osteoporosis, Other Musculoskeletal)