To pool or not to pool: Accounting for task non-attendance in subgroup analysis

Abstract

Pooling data from different subgroups offers advantages of shrinking standard errors and simplifying characterization of the data structure. The ability to pool data also facilitates meta-analysis to evaluate consensus among multiple studies and to inform benefit transfer to new choice settings. Testing for poolability requires accounting for differences in response variance or scale among subgroups. This is commonly done by assuming a single scale factor within each subgroup of interest. This assumption may not hold for many subgroups, especially when task non-attendance is present. We use data from a prior DCE study to show that task non-attendance, and by extension the assumption of a single scale factor across subgroups, can lead to inaccurate conclusions when determining poolability. To address this concern, we propose a latent-class/random-parameters Logit (LCRP) model specification that accommodates task non-attendance or other causes of scale differences within subgroups and directly tests for poolability.