Measures for Identifying Highly Associated Categorical Variables in Survey Data

• Activity Number: 252 - SPEED:Improving Survey Data Quality with Multiple Data Sources, Administrative Data, and Nonresponse Bias Control, Part 2
• Type: Contributed
• Date/Time: Monday, July 29, 2019 : 2:00 PM to 2:45 PM
• Sponsor: Survey Research Methods Section
• Author(s): Natalia Weil* and Ismael Flores Cervantes
• Companies: Westat and Westat
• Keywords: complex survey data; auxiliary categorical data; Goodman and Kruskal’s tau; measures of association for categorical variables ; nonresponse adjustments; Cramer’s V

Abstract

In some population surveys that collect data at different stages (e.g., screener and person questionnaires), auxiliary categorical variables are commonly used for the development of nonresponse adjustments. Having fewer variables in the nonresponse adjustment model makes it easier to interpret the model; therefore, it is important to identify and remove predictor variables that are highly associated among each other before proceeding with analysis. Although well-established methods for identifying correlated continuous variables exist, standard methods for highly associated categorical variables are not well-developed nor well-suited for survey data. In this paper, we develop measures of association that account for complex sample design from existing measures, such as Goodman and Kruskal’s tau and Cramer’s V. We also evaluate their performance and impact on weighting adjustments when using these weighted measures to select categorical auxiliary variables for nonresponse adjustments as compared to a manual selection of variables.