Improving measurement invariance assessments in survey research with missing data by novel artificial neural networks

This study proposes the learning vector quantization estimated stratum weight (LVQ-ESW) method to interpolate missing group membership and weights in identifying the accuracy of measurement invariance (MI) in a stratified sampling survey. Survey data is rife with missing information, such as gender and race, which is critical for identifying MI, and in ensuring that conclusions from large-scale testing campaigns are accurate. In the current study, simulations were conducted to examine the accuracy and consistency of MI detection using multiple-group confirmatory factor analysis (MG-CFA) to compare different approaches for interpolating missing information. The results of the computerized simulations showed that the proposed method outperformed traditional methods, such as List-wise deletion, in terms of accurately and stably identifying MI. The implications for interpolating missing group membership and weights for survey research are discussed.

► Missing stratified memberships and weights in sampling surveys are solved by the LVQ-ESW.
► The study is verified in simulated designs of testing measurement invariance.
► The LVQ-ESW performs superior over list-wise deletion and weighting-adjustment class.