Abstract
Many important problems in psychology and biomedical studies require testing for overdispersion, correlation and heterogeneity in mixed effects and latent variable models, and score tests are particularly useful for this purpose. But the existing testing procedures depend on restrictive assumptions. In this paper we propose a class of test statistics based on a general mixed effects model to test the homogeneity hypothesis that all of the variance components are zero. Under some mild conditions, not only do we derive asymptotic distributions of the test statistics, but also propose a resampling procedure for approximating their asymptotic distributions conditional on the observed data. To overcome the technical challenge, we establish an invariance principle for random quadratic forms indexed by a parameter. A simulation study is conducted to investigate the empirical performance of the test statistics. A real data set is analyzed to illustrate the application of our theoretical results.
Introduction
In longitudinal studies, genetic pedigrees, and psychological assessments, researchers frequently employ mixed-effects models to account for the inherent correlations within data clusters. A fundamental question in these analyses is whether the random effects or latent variables are statistically necessary. If the variance of these effects is zero—a state known as homogeneity—the model can be simplified to a standard fixed-effects framework.
However, testing the hypothesis that a variance component is zero is notoriously difficult. Because variance cannot be negative, the null hypothesis ($H_0: \sigma^2 = 0$) lies on the boundary of the parameter space. This boundary condition violates the regularity assumptions required for standard likelihood ratio tests, often leading to incorrect p-values. This paper introduces a generalized class of score tests designed to overcome these limitations while maintaining flexibility across diverse model types.
Methodology
The Generalized Score Test
The authors propose a class of score test statistics that offer a significant computational advantage: they only require fitting the model under the null hypothesis (the simpler model without random effects). This avoids the often-unstable estimation of variance components when they are near zero.
To handle multi-dimensional random effects and complex correlation structures, the framework utilizes a Cholesky decomposition. A key innovation is the decomposition of the score statistic into two distinct, interpretable components:
- Pairwise Correlation ($S_P$): This term detects whether observations within a cluster are related to one another.
- Overdispersion ($S_O$): This term detects whether the observed data exhibits more variability than the simple model predicts.
Theoretical Innovations and Resampling
Proving the validity of these tests required the establishment of a new mathematical invariance principle for random quadratic forms indexed by a parameter. This provides the theoretical foundation for the test's asymptotic behavior.
Because the resulting theoretical distributions are complex Gaussian processes, the authors developed a multiplier resampling procedure. By generating independent random normal samples and applying them to the observed data, researchers can accurately approximate p-values. This method is robust, requiring only the specification of the first and second moments (mean and variance) of the latent variables, rather than their full distribution.
Empirical Validation
Simulation Studies
The performance of the proposed test ($S_S$) was evaluated using Logistic Mixed Models for binary outcomes and Linear Mixed Models for continuous outcomes. The results demonstrated that the $S_S$ test consistently outperformed existing benchmarks, such as the Lin (1997) test, particularly in scenarios involving complex correlation structures. Furthermore, the resampling procedure effectively controlled Type I error rates, ensuring that the test does not produce excessive false positives even in the presence of nuisance overdispersion.
Application: Yale Family Study
The methodology was applied to the Yale Family Study of Comorbidity of Alcoholism and Anxiety (YFSCAA), which includes data from 222 probands and 1,194 adult first-degree relatives. The goal was to test for familial aggregation of alcoholism.
The generalized score test found overwhelming evidence for familial clustering ($p < 0.0001$). Interestingly, the diagnostic decomposition showed that while the correlation term ($S_P$) was highly significant, the overdispersion term ($S_O$) was not. This suggests that the observed heterogeneity in the data is driven specifically by familial relationships rather than general unexplained variance.
Conclusion
This paper provides a powerful and flexible toolkit for testing homogeneity in complex datasets. By addressing the boundary problem and providing a robust resampling method, the authors offer a solution that is both theoretically sound and practically applicable. While the resampling procedure is more computationally intensive than closed-form alternatives, its ability to provide diagnostic insights into the source of heterogeneity—whether correlation or overdispersion—makes it a valuable asset for researchers in the social and biomedical sciences. Future work may further refine these functional central limit theorems to optimize the efficiency of the tests in even broader contexts.
Summary updated