Statistics Glossary → Lo-Mendell-Rubin (LMR) test
Lo-Mendell-Rubin (LMR) testconcept
An adjusted likelihood-ratio test in latent class and latent profile analysis of whether a model with K classes fits the data significantly better than one with K − 1 classes.
The ordinary likelihood-ratio test cannot be used to compare numbers of classes: the K − 1 class model corresponds to a boundary case of the K class model in which one class proportion is zero, and under that condition the statistic does not follow a chi-square distribution. Building on Vuong's approach to comparing non-nested models, Lo, Mendell and Rubin proposed an approximate distribution for the statistic and an adjusted LR value.
A significant p value (for example p < 0.05) suggests that the K class model may be preferred; a non-significant one suggests that one class fewer is enough. The theoretical basis of the test has been questioned, and in simulation studies the bootstrap likelihood-ratio test (BLRT) usually performs better. The number of classes should therefore be decided by looking at the LMR, the BLRT, information criteria such as BIC, entropy and the interpretability of the classes together.
An everyday example
A researcher studies students' study habits with latent class analysis. The 3-class model against the 2-class model gives LMR p = 0.01, and the 4-class model against the 3-class model p = 0.38. BIC also reaches its minimum at 3 classes, and the three classes (steady, last-minute, avoidant) are meaningful, so the 3-class solution is chosen.
In MerQur
'LMR test (always reported)' in Latent Class Analysis and Latent Profile Analysis; it compares the K class (profile) model with the K − 1 class model using an adjusted likelihood ratio.
Analyses where this concept appears
Latent Class Analysis (LCA)Latent Profile Analysis (LPA)
Related concepts
Bootstrap likelihood ratio test (BLRT)Information criteria (AIC, AICc, BIC)