An Investigation of Factored Regression Models With Incomplete Binary Predictors (Poster 29)

• In Event: AERA Poster Session Thursday 8:00 a.m.
  In Poster Session: Advancements in Educational Data Analysis: Machine Learning, Multilevel Models, and Beyond

Abstract

The factored regression (FR) framework leverages the probability chain rule to decompose a joint distribution into a product of univariate conditional distributions. Recently, the missing data literature leveraged the FR to appropriately handle incomplete predictors that are nonlinearly related to an outcome. An open question in this literature is how well these models perform when incomplete predictors are misspecified. The present study used a Monte Carlo simulation to investigate the performance of incomplete binary predictors across a broad range of conditions and modeling approaches. Results indicate that under conditions of small sample sizes and asymmetric proportions of binary variables, estimations of interaction terms exhibit substantial bias. The sequencing of imputation model factorization also exerts an influence on outcomes under different specifications.

Authors

• Suyoung Kim, University of Chicago

• Jiwon Kim, Northwestern University