Overview

Regression models can be used to predict outcomes from predictors. A linear predictor term \(\eta\) based on the predictors is passed through an inverse link function, \(g^{-1}\), and then used in the observation model, \(f\). The observation model may have other parameters \(\theta\) which are not necessarily determined through the predictors.

\[ \begin{align} y_i &\sim f(g^{-1}(\eta), \theta) \\ \eta_i &= \alpha + \beta x_i \end{align} \]

Depending on the choice of \(f\), different link functions \(g\) are used. For example, if \(f\) is the \(\text{normal}\) distribution, \(g\) is the identity function. If \(f\) is the \(\text{Bernoulli}\) distribution, \(g\) is most commonly the logit function.

Notation

Symbol	Explanation
\(y_i\)	Observed outcome
\(x_i\)	Observed predictors
\(\alpha\)	Intercept
\(\beta\)	Regression coefficients (predictor weights)