MerQuR+Statistics Glossary

Statistics Glossary → Likelihood ratio test (LRT)

Likelihood ratio test (LRT)concept

A test that compares the maximized likelihoods of two nested models to decide whether the added parameters improve the model significantly.

The restricted (simpler) model is obtained by fixing some parameters of the general model at zero or another given value. The test statistic is LR = -2 (log L_restricted - log L_general), and under the null hypothesis it approximately follows a chi-square distribution with degrees of freedom equal to the difference in the number of parameters. The 'model chi-square' in logistic regression is an LRT computed as a difference in -2LL.

The models must be nested and fitted to the same data; non-nested models are compared with AIC or BIC. When the parameter lies on the boundary, as when testing whether a variance component is zero, the chi-square approximation is conservative and a 50:50 mixture of chi-square distributions is used. The Wald and score tests address the same hypothesis; in small samples the LRT is usually more reliable than the Wald test.

An everyday example

Someone wants to add 'has a garden' to a house-price model. Two models are fitted, with and without the garden variable, and the likelihood measures how well each explains the data. If the single added variable lowers -2 log likelihood by 9 units, this corresponds to p ≈ 0.003 on a chi-square distribution with 1 degree of freedom, so the variable contributes significantly.

In MerQur

The 'Omnibus + Type III (LR/Wald)' output of logistic regression, 'Omnibus + Type III tests' in Poisson/negative binomial regression, 'Likelihood-ratio tests per effect' in multinomial logistic regression, 'Likelihood-ratio tests' in linear mixed models and the 'Random-effect LRT' in GLMM. Model comparisons in variance components analysis, item response theory and path analysis also rely on this test. In the chi-square independence and goodness-of-fit tests the 'G-test' checkbox makes the likelihood-ratio G² the main statistic; in Latent Class and Latent Profile Analysis, 'BLRT bootstrap samples' compares the K-class model with the K - 1 class model by a parametric bootstrap likelihood ratio test; 'Variable selection' (Forward LR, Backward LR) in Cox regression also relies on this test.

Analyses where this concept appears

Chi-Square IndependenceChi-Square Goodness of FitCrosstabLogistic RegressionPoisson / Negative Binomial Regression (Count Regression)Multinomial Logistic RegressionOrdinal Logistic RegressionLinear Mixed Models (LMM)Generalized Linear Mixed Models (GLMM)Confusion Matrix MetricsLatent Class Analysis (LCA)Latent Profile Analysis (LPA)Item Response Theory (IRT / Rasch)Variance Components Analysis (VARCOMP)Probit RegressionSAR (Spatial Autoregressive Lag)Path AnalysisCox Regression (Proportional Hazards Model)Competing RisksCluster-Robust Cox (Marginal and Frailty Models)Sample Selection and Truncated Regression (Heckman)

Related concepts

Wald test and Wald confidence intervalScore (Lagrange multiplier) testMaximum likelihood (ML) estimation