Comparing Four Different Statistical Packages for Hierarchical Linear Regression: GENMOD, HLM, ML2, and VARCL

Abstract

An overview is given of the available statistical theory and software for analyzing hierarchically nested data. Programs are evaluated, and general techniques are proposed to analyze data from several domains. This research is part of a larger project to evaluate elementary education in the Netherlands. The models discussed are the random coefficient models, the hierarchical mixed linear models, and the multilevel linear models. The abstract characteristics of the three classes of models and the systematic treatment of random and non-random parts of each class are described. Transformation of the models and the likelihood function are considered. The following four computer programs, using various types of algorithms, are discussed: (1) GENMOD; (2) HLM; (3) ML2; and (4) VARCL. Each is compared for design, implementation, performance and results, and ease of use. To overcome some of the disadvantages of these techniques, a new program, MULTIPATH, is proposed for a more general approach to the analysis of data from different domains. Thirteen data tables and a 61-item list of references are included. (SLD)