Goodness-of-fit tests for functional form of Linear Mixed effects Models

TL;DR

This paper introduces a novel graphical and statistical approach to assess the correctness of the functional form in linear mixed effects models, enhancing model validation and potential improvements.

Contribution

It proposes a new residual-based process for goodness-of-fit testing of LMMs' functional form, including a method for p-value estimation via sign-flipping, bootstrap, or simulation.

Findings

Plots of the residual processes effectively identify model mis-specification.

The p-value estimation method is validated through theoretical analysis and extensive simulations.

The approach is applicable to complex LMMs with multi-level or crossed random effects.

Abstract

Linear mixed effects models (LMMs) are a popular and powerful tool for analyzing clustered or repeated observations for numeric outcomes. LMMs consist of a fixed and a random component, specified in the model through their respective design matrices. Checking if the two design matrices are correctly specified is crucial since mis-specifying them can affect the validity and efficiency of the analysis. We show how to use random processes defined as cumulative sums of appropriately ordered model's residuals to test if the functional form of the fitted LMM is correctly specified. We show how these processes can be used to test goodness-of-fit of the functional form of the entire model, or only its fixed and/or random component. Inspecting plots of the proposed processes is shown to be highly informative about the potential mis-specification of the functional form of the model, providing clues for potential improvement of the model's fit. We show how the visual inspection can be objectified by using a novel procedure for estimating $p$-values which can be based on sign-flipping/bootstrap or simulations and show its validity by using theoretical results and a large Monte Carlo simulation study. The proposed methodology can be used with LMMs with multi-level or crossed random effects.% and could also be extended to generalized LMMs.