A Note on Likelihood Ratio Tests for Models with Latent Variables

Abstract

The likelihood ratio test (LRT) is widely used for comparing the relative fit of nested latent variable models. Following Wilks’ theorem, the LRT is conducted by comparing the LRT statistic with its asymptotic distribution under the restricted model, a $$\chi ^2$$ χ 2 distribution with degrees of freedom equal to the difference in the number of free parameters between the two nested models under comparison. For models with latent variables such as factor analysis, structural equation models and random effects models, however, it is often found that the $$\chi ^2$$ χ 2 approximation does not hold. In this note, we show how the regularity conditions of Wilks’ theorem may be violated using three examples of models with latent variables. In addition, a more general theory for LRT is given that provides the correct asymptotic theory for these LRTs. This general theory was first established in Chernoff (J R Stat Soc Ser B (Methodol) 45:404–413, 1954) and discussed in both van der Vaart (Asymptotic statistics, Cambridge, Cambridge University Press, 2000) and Drton (Ann Stat 37:979–1012, 2009), but it does not seem to have received enough attention. We illustrate this general theory with the three examples.

The Influence of Blue Light Exposure on Reconstructed 3-Dimensional Skin Model: Molecular Changes and Gene Expression Profile

“…After isolation of total RNA, we performed RNA-seq assay (Gene Expression Omnibus identification GSE190106). Data analysis was performed using likelihood ratio test statistical method ( Chen et al, 2020 ) in addition to Ingenuity Pathway Analysis software for pathway analysis. A representative image of differential gene expression in all digital light exposure models can be found in Figure 2 a.…”

The psychometric challenges of implementing wellbeing assessment into clinical research and practice: A commentary on “assessing mental wellbeing using the mental health continuum–short form: A systematic review and meta-analytic structural equation modeling”.

“…This may include nested model comparisons based on Δχ 2 and χ 2 -based fit indices. As Chen, et al (2020) noted, over-factoring is a likely consequence of comparisons that assume proper χ 2 distributions when regularity conditions are violated. Specifically, nested model comparisons require fixing a “parameter on the border of admissible parameter space … [that] may not yield proper χ 2 distributions” (Brown, 2006, pp.…”

A tensor factorization model of multilayer network interdependence

“…Important to note in defining a layer independence test is that the theory underlying the likelihood ratio test (LRT), Wilks' theorem [65], necessarily depends on the maximum likelihood being reached, something we cannot guarantee in our nonconvex problem context. Moreover, we propose this method for determining layer interdependence constrained to the classes of models for which the difference in the degrees of freedom d = (L − C)K 2 − LC > 0, which is not always true for certain values of L, K, and C. Alternative LRTs (including a bootstrapping method that avoids asymptotics altogether) for latent variable models have been proposed [13], and address this issue as well as other common issues that arise when using the LRT to compare latent variable models. Definition 4.1 (Layer Independence Test).…”