Some Common Pitfalls in Causal Analysis of Categorical Data

Abstract

Examples of some common pitfalls in the analysis of categorical data are discussed in the context of causal interpretation of the results. Though no statistical technique can replace theory, the author shows that log-linear modeling and chi square automatic interaction detection can provide researchers with powerful tools for gaining valuable causal insights into their data. Examples include the biasing effects of omitted variables, omitted interactions, improper contrast coding, and misspecification of the structure of an hypothesized interaction.