MerQuR+Statistics Glossary

Statistics Glossary → Overdispersion and the scale parameter

Overdispersion and the scale parameterconcept

The situation in which the observed variance of count or proportion data exceeds what the Poisson or binomial model predicts, and the scale (dispersion) parameter used to correct for it.

The Poisson model assumes that the variance equals the mean, and the binomial model that it equals np(1 − p). In real data the variance is often larger because of unobserved differences between units or clustering. Coefficient estimates usually remain unbiased, but the standard errors and p values come out too small. Overdispersion is usually assessed with the ratio φ = Pearson χ² / df (or deviance / df); a value clearly above 1 signals a problem.

The simple fix is to multiply the standard errors by √φ (the quasi-likelihood, or quasi-Poisson / quasi-binomial, approach). A more structural fix is a negative binomial model for counts or a mixed model with random effects. Overdispersion can also come from an omitted variable or from excess zeros, in which case the model itself should be revised. For ungrouped binary data made up of individual 0/1 observations, the Pearson ratio does not reliably reveal overdispersion; scale correction is more meaningful for grouped data in events/trials form.

An everyday example

An ornithologist counts bird nests in 50 plots. The mean is 4 nests per plot, but the variance is 12: some plots are near woodland and hold many nests, others are almost empty. A Poisson model assumes the variance is 4, so its standard errors are about √3 ≈ 1.7 times too small. A negative binomial model or a scale correction removes this error.

In MerQur

In logistic, ordinal and probit regression, the 'Overdispersion scale' option (None (scale = 1), Pearson, Deviance) scales the standard errors, and multinomial logistic regression has the same option. In Poisson / negative binomial regression, 'Extra output: Overdispersion tests' gives the Cameron-Trivedi tests and the Poisson-NB likelihood ratio test. In generalized linear mixed models, the Negative binomial NB2 choice of the 'Family' option models overdispersed counts, and the output also includes a conditional Pearson overdispersion check.

Analyses where this concept appears

Logistic RegressionPoisson / Negative Binomial Regression (Count Regression)Multinomial Logistic RegressionOrdinal Logistic RegressionGeneralized Linear Mixed Models (GLMM)Probit Regression

Related concepts

Distribution family (GLM)Robust (sandwich) standard errors (HC0-HC3)