Back to Overview of Sessions

Session Details

In social research, many theoretical and empirical arguments for the analysis of multivariate associations exist. One of the special cases of multivariate associations is the interaction of features, i.e., a combination of categories of variables, which determines some phenomenon. Beginning from the 1960s, such interaction effects have been in the focus of researchers' attention due to their wide prevalence in the survey data: since then, methods for automatic interaction detection have been developed, and the results of these methods have been included in the predictive models. However, although the idea of finding interactions was initially focused on survey data, in fact, this practice is not widely used in empirical research. Partially, it is explained due to the lack of a universal method of searching for the necessary interactions. In the case of categorical variables, there are many such methods (for example, log-linear analysis, decision trees, multiple correspondence analysis, and so on), and their implementation differs significantly, and methods of searching for interactions of continuous variables, on the contrary, are practically unknown. On the other hand, difficulties arise in obtaining final predictive models: there is a risk of parameter estimates bias due to multicollinearity, and the interpretation of the effects becomes complicated. At the same time, the researcher who is limited to “traditional” modeling and examines only the two-dimensional associations could face the Simpson's paradox. This phenomenon, which is widely known in statistics, occurs when a multivariate association “disappears” or changes its direction if a researcher analyzes only aggregated (two-dimensional) data. This problem leads to incomplete or incorrect conclusions about social reality. The question addressed to the participants of the session is: how to strike a balance between the possibilities of improving the explanatory or predictive abilities of models by considering interaction effects and limitations that arise in the process of selecting the desired effects and building the final model? Approximate themes of reports of the session are devoted to a comparison of methods of searching for interaction effects, ways of overcoming the identified limitations, the best practices in building predictive models with the interaction effects, and a comparison of alternative ways of building predictive models with multivariate associations – for example, using multilevel regression.