Does the Package Matter? A Comparison of Five Common Multilevel Modeling Software Packages. (October 2018)

• Record Type: Journal Article
• Authors: McCoach, D. Betsy; Rifenbark, Graham G.; Newton, Sarah D.; Li, Xiaoran; Kooken, Janice; Yomtov, Dani; Gambino, Anthony J.; Bellara, Aarti
• Is Part Of: Journal of educational and behavioral statistics. Volume 43:Number 5(2018)
• Page Start: 594
• Page End: 627
• Publication Date: 2018-10
• Subjects: multilevel modeling -- statistical software -- Monte Carlo simulation; Educational statistics -- Periodicals; Social sciences -- Statistical methods -- Periodicals; 370.2
• DOI: 10.3102/1076998618776348
• Languages: English
• ISSNs: 1076-9986

Abstract

This study compared five common multilevel software packages via Monte Carlo simulation: HLM 7, M plus 7.4, R (lme4 V1.1-12), Stata 14.1, and SAS 9.4 to determine how the programs differ in estimation accuracy and speed, as well as convergence, when modeling multiple randomly varying slopes of different magnitudes. Simulated data included population variance estimates, which were zero or near zero for two of the five random slopes. Generally, when yielding admissible solutions, all five software packages produced comparable and reasonably unbiased parameter estimates. However, noticeable differences among the five packages arose in terms of speed, convergence rates, and the production of standard errors for random effects, especially when the variances of these effects were zero in the population. The results of this study suggest that applied researchers should carefully consider which random effects they wish to include in their models. In addition, nonconvergence rates vary across packages, and models that fail to converge in one package may converge in another.