Abstract

Latent profile analysis (LPA) is a method commonly used in psychology to identify subgroups of individuals who share common characteristics. To apply LPA on data with missing values, full information maximum likelihood (FIML) and multiple imputation (MI) are commonly recommended. In this study, we propose k-nearest neighbor (kNN) imputation, as an efficient alternative to handle missing data in LPA and examined its potential using simulated datasets. Datasets were generated with varying conditions: missing value generation mechanisms, missing rates, distances between subgroups, and sample sizes. Complete data were generated by kNN imputation from the simulated datasets and were used in LPA. Results were compared to the results from FIML in terms of the number of estimated subgroups, the accuracy of mean profiles, and the quality of classification. The accuracy of the number of subgroups from kNN imputation was comparable to the results from FIML in most conditions, and kNN imputation performed better in some conditions Neither method consistently performed better in terms of the accuracy of mean profiles. The quality of classification from kNN imputation was better in all conditions, and was closer to the results from complete data analyses. From the results, we suggest kNN imputation as an alternative to FIML to handle missing data in LPA, especially in conditions wherein FIML often fails. We also suggest using kNN imputation as well as FIML to compare results to check the stability of parameter estimates.