Matching Techniques in Stated Preferences for Health
Author(s)
Turner AJ1, Vass C*2
1The University of Manchester, Manchester, UK, 2RTI Health Solutions, Manchester, UK
OBJECTIVES: There is a growing interest in using stated preference methods to quantify preferences, and the degree of heterogeneity in these preferences. A popular preliminary investigation into preference heterogeneity involves a split-sample analysis to make comparisons across subgroups. However, subgroups may differ in other observed characteristics which may bias comparisons if these characteristics are associated with preferences. This study aimed to explore the use of different matching approaches to identify true differences in preferences across subgroups.
METHODS: We used simulated data to compare the preferences of a sample of patients and the public for a hypothetical healthcare intervention. Simulations assumed patients were older and received higher benefit income. The utility function for the sample of patients and the public was specified to be identical (implying preference homogeneity) and utility was assumed to increase with health and life-years, and reduce with risk and cost. Utility for cost was specified as a function of benefit income and age. We estimated preference weights using a conditional logit regression with three samples: (1) an unmatched sample; (2) a propensity-score-matched sample; and (3) a sample weighted using entropy balancing. Approaches were compared based on their ability to reduce imbalance in characteristics and to detect preference homogeneity.
RESULTS: Due to differences in age and benefit income, unmatched analysis detected statistically significant differences in the preference for cost between the public and patients. Both propensity score matching and entropy balancing reduced imbalance in characteristics across subgroups, although the reduction was greater when using entropy balancing. Following matching or weighting, there were no significant differences in the preferences for any attributes.
CONCLUSIONS: Our results show simple subgroup analyses may produce erroneous conclusions regarding heterogeneity in preferences. Matching methods may be useful for preference researchers seeking to compare preferences for health and health care.
METHODS: We used simulated data to compare the preferences of a sample of patients and the public for a hypothetical healthcare intervention. Simulations assumed patients were older and received higher benefit income. The utility function for the sample of patients and the public was specified to be identical (implying preference homogeneity) and utility was assumed to increase with health and life-years, and reduce with risk and cost. Utility for cost was specified as a function of benefit income and age. We estimated preference weights using a conditional logit regression with three samples: (1) an unmatched sample; (2) a propensity-score-matched sample; and (3) a sample weighted using entropy balancing. Approaches were compared based on their ability to reduce imbalance in characteristics and to detect preference homogeneity.
RESULTS: Due to differences in age and benefit income, unmatched analysis detected statistically significant differences in the preference for cost between the public and patients. Both propensity score matching and entropy balancing reduced imbalance in characteristics across subgroups, although the reduction was greater when using entropy balancing. Following matching or weighting, there were no significant differences in the preferences for any attributes.
CONCLUSIONS: Our results show simple subgroup analyses may produce erroneous conclusions regarding heterogeneity in preferences. Matching methods may be useful for preference researchers seeking to compare preferences for health and health care.
Conference/Value in Health Info
2020-11, ISPOR Europe 2020, Milan, Italy
Value in Health, Volume 23, Issue S2 (December 2020)
Code
SP1
Topic
Patient-Centered Research
Topic Subcategory
Stated Preference & Patient Satisfaction
Disease
No Specific Disease