Statistics Glossary → Complete and quasi-complete separation
Complete and quasi-complete separationconcept
In logistic-type models, a situation in which a predictor or combination of predictors separates the outcome perfectly or almost perfectly, so the maximum likelihood estimates run off to infinity.
Under complete separation, a linear combination of predictors divides cases with and without the event without error; for example, everyone above a certain value had the event and no one below it did. Under quasi-complete separation the division is perfect apart from a few ties at the boundary. In both cases the likelihood keeps increasing as the coefficient grows, and no finite maximum likelihood estimate exists.
The symptoms are huge coefficients and odds ratios, extremely large standard errors, convergence warnings and meaningless Wald tests. Common causes are a small sample, a rare event and a category with no events at all (a zero cell in the crosstab). Remedies include Firth's penalized likelihood correction, exact logistic regression, Bayesian estimation with informative priors or merging categories; dropping the problematic variable often means dropping the strongest predictor.
An everyday example
A study of 40 patients records whether each smokes and whether each develops a certain disease. No non-smoker in the sample developed the disease. Because the model tries to conclude that 'risk is zero for non-smokers', it inflates the odds ratio without limit and cannot give a meaningful estimate.
In MerQur
In Logistic Regression, the 'Firth correction' applies the penalized likelihood with the Jeffreys prior and gives finite estimates under separation; a warning is written under quasi-separation. In multinomial logistic regression MerQur warns and shows no estimates under separation; in Ordinal Logistic Regression a warning is written under separation (very large standard errors).
Analyses where this concept appears
Logistic RegressionMultinomial Logistic RegressionOrdinal Logistic Regression