Pursuing weak paths: a performance evaluation of multiple imputation in PLS

Missing data is a common issue in structural equation modeling, including partial least squares structural equation modeling, which can lead to biased results and weaker statistical conclusions. Despite the popularity of partial least squares structural equation modeling in multivariate analysis, there is a lack of rigorous handling and reporting of missing data in the literature. Most studies rely on complete case analysis and mean imputation, overlooking modern techniques. Recent works have applied and evaluated modern approaches but lack insights into parameters that influence performance. This paper addresses this gap by evaluating the performance of various imputation methods, including expectation maximization and multiple imputation, through the analysis of simulated datasets with varying sizes and levels of missingness. In contrast to previous works, this study considers the effects of factors such as multivariate distribution and model complexity. The findings provide evidence for the superior performance of donor-based multiple imputation techniques and the impact of properties of data and model. In addition, we present recommendations for researchers on how to effectively manage missing data in partial least squares structural equation modeling, enhancing the reliability and validity of their results.