Statistics Glossary → Distribution family (GLM)
Distribution family (GLM)concept
The choice in a generalized linear model that sets the conditional distribution of the response and therefore the relationship between its mean and variance.
Generalized linear models do not require a normally distributed response; the distribution is chosen from the exponential family. The Gaussian (normal) family is for continuous responses with constant variance, the binomial for 0/1 or proportion data (variance = μ(1 − μ)), the Poisson for counts (variance = μ), the gamma for positive, right-skewed continuous responses (variance ∝ μ²) and the negative binomial for overdispersed counts (variance = μ + μ²/θ). Together with a link function (for example log for Poisson, logit for binomial), the family defines the model.
The choice should rest on the nature of the response (continuous, count, binary, positive) and on the observed mean-variance relationship. If the variance exceeds what the family implies (overdispersion), the standard errors come out too small; quasi-likelihood, the negative binomial or robust standard errors can then be used. Even with the wrong family, the coefficients can be consistent if the mean model is right, but inference will mislead.
An everyday example
A city council models monthly accident counts at junctions using traffic volume. Accident counts cannot be negative, are whole numbers, and both the mean and the spread rise at busy junctions, so the Poisson family is chosen instead of normal regression. When the variance turns out to be three times the mean, it switches to the negative binomial family because of overdispersion.
In MerQur
'Distribution family' in Poisson / Negative Binomial Regression; 'Family' in GEE and Generalized Linear Mixed Models; 'Family' in the Generalized Additive Model; 'Target type' / 'Target distribution' in Ridge, Lasso and Regularized Regression; and the 'Analysis model' (linear, logistic, Poisson) in Multiple Imputation.
Analyses where this concept appears
Poisson / Negative Binomial Regression (Count Regression)Ridge RegressionLasso RegressionGeneralized Estimating Equations (GEE)Generalized Linear Mixed Models (GLMM)Generalized Additive Model (GAM, Nonlinear Regression)Regularized Regression (LASSO / Ridge / ElasticNet)Multiple ImputationLOESS (Local Regression)
Related concepts
Link functionGeneralized linear model (GLM)Overdispersion and the scale parameter