Streamlining Missing Data Analysis by Aggregating Multiple Imputations at the Data Level: A Monte Carlo Simulation to Assess the Tenability of the SuperMatrix Approach

Abstract

A Monte Carlo Simulation Study was conducted to assess the tenability of a novel treatment of missing data. Through aggregating multiply-imputed data sets prior to model estimation, the proposed technique allows researchers to reap the benefits of a principled missing data tool (i.e., multiple imputation), while maintaining the simplicity of complete case analysis. In terms of the accuracy of model fit indices derived from confirmatory factor analyses, the proposed technique was found to perform universally better than a naive ad hoc technique consisting of averaging the multiple estimates of model fit derived from a traditionally conceived implementation of multiple imputation. However, the proposed technique performed considerably worse in this task than did full information maximum likelihood (FIML) estimation. Absolute fit indices and residual based fit indices derived from the proposed technique demonstrated an unacceptable degree of bias in assessing direct model fit, but incremental fit indices led to acceptable conclusions regarding model fit. Chi-squared difference values derived from the proposed technique were unbiased across all study conditions (except for those with very poor parameterizations) and were consistently more accurate than such values derived from the ad hoc comparison condition. It was also found that Chi-squared difference values derived from FIML-based models were negatively biased to an unacceptable degree in any conditions with greater than 10% missing. Implications, limitations and future directions of the current work are discussed.