Inconsistent and Lexicographic Choices in Stated Preference Analysis

In stated choice (SC) data inconsistent and lexicographic choice behavior is often observed. It is sometimes recommended to remove data with these characteristics from the analysis. In this paper we reconsider this recommendation. In our data many respondents have inconsistent choice patterns, which appear to be due to incidental mistakes. Moreover, a large number of the consistent respondents have lexicographic choice patterns. We show that the logit model, which is the most popular tool for analyzing SP data, is compatible with inconsistent and seemingly lexicographic choice behavior and that it offers precise predictions about the occurrence of such choices. In the data at our disposal the actual number of respondents who made different choices in two identical choice situations is substantially lower than that predicted by the model, whereas the number of respondents with lexicographic answers is much larger than predicted. The logit model is then adapted in various ways to bring it in better agreement with the facts. In particular, we introduce an effect of remembering the earlier choice when the same situation recurs, the presence of latent classes of lexicographic respondents and the presence of heterogeneity among respondents