MerQuR+Statistics Glossary

Statistics Glossary → Link function

Link functionconcept

In generalized linear models, the function that connects the expected value of the response to a linear combination of the predictors: g(μ) = Xβ.

In linear regression the mean of the response is set directly equal to Xβ (the identity link). Probabilities, however, must lie between 0 and 1 and counts must be positive, whereas Xβ can take any value. The link function reconciles these scales. For binary responses the logit link uses log(p / (1 - p)) = Xβ and the probit link Φ⁻¹(p) = Xβ; for counts the log link uses log(μ) = Xβ. Each distribution family has a canonical link (logit for the binomial, log for the Poisson, identity for the normal), but other links can be chosen.

The choice of link determines how the coefficients are read: with the logit, exp(β) is an odds ratio; with the log link, it is a ratio (risk or rate ratio); with the probit, coefficients are in standard-deviation units of a latent normal variable. The complementary log-log link is asymmetric and suits rare events or time-based data, while the cauchit assumes a heavy-tailed latent distribution. Logit and probit give similar probabilities in most applications, with coefficients differing by a factor of roughly 1.6 to 1.8. A link function is not the same as transforming the response: a transformation changes the data, whereas the link transforms the expected value.

An everyday example

You want to model the probability that customers accept an offer as a function of age. A straight line could predict meaningless probabilities such as 120% or -15% for very young or very old customers. The logit link squeezes the output of the linear model into the 0-1 range along an S-shaped curve, so predicted probabilities stay meaningful at every age.

In MerQur

It is chosen with the 'Link function (Cumulative Logit, Cumulative Probit, Complementary Log-Log, Negative Log-Log, Cauchit)' parameter in Ordinal Regression, 'Link function' in GEE, 'Link function (binomial)' in GLMM and 'Link (distribution) (Probit (normal), Logit (logistic))' in Probit Regression. In the Generalized Additive Model the link (Gaussian identity, binomial logit, Poisson and gamma log) is set together with the response distribution.

Analyses where this concept appears

Ordinal Logistic RegressionGeneralized Estimating Equations (GEE)Generalized Linear Mixed Models (GLMM)Generalized Additive Model (GAM, Nonlinear Regression)Probit RegressionSample Selection and Truncated Regression (Heckman)

Related concepts

Distribution family (GLM)Generalized linear model (GLM)