On Imputing Binary Data Without Making Any Loglinear Assumptions via Pairwise Associationsand Corresponding Conditional Probabilities

• Activity Number: 72
• Type: Contributed
• Date/Time: Sunday, July 29, 2012 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Irene Helenowski*+ and Hakan Demirtas
• Companies: Northwestern University and University of Illinois at Chicago
• Keywords: multiple imputation ; binary data

Abstract

In this work, we present a method for imputing binary data without making any multinomial or loglinear model assumptions. Our approach employs principles of generating binary data from multivariate normally distributed values, as discussed in Emrich and Piedmonte (1991). Specifically, a multivariate normally distributed data set is generated using a covariate matrix associated with pairwise tetrachoric correlations derived from phi coefficients and specified proportions, the same fraction of missing information is introduced in the generated data as is in the original binary data, multiple imputation is applied to the generated values via joint modeling under the normality assumption, and imputed values are dichotomized by quantiles corresponding to the original specified proportions. Simulation studies and real data applications indicate this new method as a promising approach in imputing binary data while relaxing multinomal and loglinear model assumptions.