ZiF Cooperation Group
Statistical Models for Psychological and Linguistic Data
August 2019 - Juli 2021
Convenors: Reinhold Kliegl (Potsdam, GER), Harald Baayen (Tübingen, GER), Douglas Bates (Madison, USA)
The goal of the cooperation group is to investigate and further develop a series of statistical methods that are now available for (a) the analysis of experimental and psychometric data from psychology and psycholinguistics, (b) the modeling of linguistic distributional data, and, possibly going beyond these domains, (c) the analyses of genome-wide associations. The methods in focus are (generalized) linear mixed models [(G)LMMs], generalized additive (mixed) models [GA(M)Ms], and multivariate (generalized) mixed models [MV(G)MMs]. These statistical methods deal with inferential statistical problems that arise from dependencies between, for example, measures on the same subjects or the same items in psycholinguistic experiments or, again as an example beyond the core domains, the same nucleotides in the genome (i.e., within-unit correlations). The research uses the R Computational Environment and the Julia Programming Language.
Topics of Research Workshops
- Case studies that demonstrate the advantages of Julia-based MixedModels with respect to data size and model complexity. The case studies show that users can seamlessly switch between the R and the Julia language. For example, they may want to use the familiar R Computational Environment for data preparation and visualization and use Julia MixedModels only for model fitting.
- Work on methodological issues of (G)LMMs relating to (a) reliability and precision of parameter estimates, (b) model identification based on principal components of random-effect structure, and (c) model selection within the spectrum of identified models. In perspective, this work promotes much needed sensitivity to the relevance of variance and correlation parameters with respect to the interpretation of fixed effects.
- Work on new GA(M)M-related extensions to handle non-linear functional relations such as quantile GAMs and piecewise exponential additive mixed models. Both make it possible to examine in great detail whether predictors have effects that are constant or variable across the distribution.
- Work on new MV(G)MMs to model simultaneously more than one dependent variable. The goal is to coordinate two lines of research on multivariate regression, that is (a) work with multiple regression with linear mappings and work linking two LMMs with different responses in a nonlinear mixed model.