Considerations around coding the membership probability function in a latent class analysis

Abstract

This technical note discusses how dummy and effects coding of categorical respondent characteristics in a class membership probability function should be interpreted by researchers employing a latent class analysis to explore preference heterogeneity in a discrete-choice experiment. Previous work highlighted issues arising from such coding when interpreting an alternative specific constant that represents an opt-out alternative or current situation in a discrete-choice experiment and did not fully address how this coding impacts the interpretation of parameters resulting from the membership probability function in a latent class analysis. Although latent class membership probability could be predicted separately for each respondent or subgroup of respondents, conclusions are often drawn directly from the model estimation using the full sample, which requires correctly interpreting the estimated parameters. In these cases, the misinterpretation that may arise if the problem is ignored could impact the policy conclusions and recommendations drawn based on the discrete-choice experiment results. This note provides an example comparing dummy and effects coding used to model respondent characteristics in the membership probability function in a discrete-choice experiment aimed to explore preferences for the treatment of chronic pain in the USA.