Some Comments on the Causal Analysis of Surveys with Log-Linear Models

Recent developments in analysis of cross-classified data have given social scientists powerful tools for investigating relationships among qualitative variables. Since investigators can analyze cross-classifications with concepts similar to those used in regression and path analysis, they may feel that the level of measurement problem has been solved. This study shows, however, that at least one approach, log-linear analysis, does not by itself solve every measurement problem. Unless the technique is applied carefully, it may produce highly misleading results. The conditions under which log-linear models can be misleading are explored within the context of causal analysis of surveys.