Statistics Glossary → Bootstrap likelihood ratio test (BLRT)
Bootstrap likelihood ratio test (BLRT)concept
A test in latent class or latent profile analysis of whether a K-class model fits significantly better than a K - 1 class model, with the distribution of the likelihood ratio generated by parametric bootstrap.
The usual way to compare nested models is the likelihood ratio test: -2(ln L_K-1 - ln L_K) follows a chi-square distribution. When comparing numbers of classes this rule does not hold, because the K - 1 class model is a case of the K-class model with one class proportion fixed at zero, on the boundary of the parameter space. BLRT solves this as follows: many artificial data sets are generated from the estimated parameters of the K - 1 class model, both the K - 1 and the K-class models are fitted to each, and their likelihood ratios computed. How extreme the observed ratio is in this bootstrap distribution gives the p value.
Simulation studies (Nylund and colleagues) found that BLRT generally performs better than information criteria and the Lo-Mendell-Rubin test in determining the number of classes. The price is computation: for each bootstrap sample two models must be fitted, often with several sets of starting values, and fits stuck at local maxima can distort the p value. In large samples BLRT can also find small, substantively unimportant classes significant, so it should be considered together with the interpretability of the classes and the size of the smallest class.
An everyday example
An education researcher looks for latent classes in 8 yes/no questions about students' study habits. Comparing the three-class model with the two-class model, 200 bootstrap data sets are generated, and none gives a likelihood ratio as large as the observed one (p < 0.005). Comparing four classes with three, p = 0.31, so three classes seem enough.
In MerQur
Available as 'BLRT bootstrap samples (0 = off)' in latent class analysis and 'BLRT bootstrap samples' in latent profile analysis; the K-class model is compared with the K - 1 class model.
Analyses where this concept appears
Latent Class Analysis (LCA)Latent Profile Analysis (LPA)