Link function

A function used to transform the data from a non-continuous distribution into "real space" so it can be used for regression analysis.

Real space means there are no bounds to the numbers - they range from −∞ to +∞.

It is denoted with g(x).

Observation Model Observation model Typical Link Function
Gaussian Observation Model $$p(y_i\textbar x_i)={1\over\sqrt{2\pi\sigma^2}}e^{-(y_i-\mu_i)^2/\sigma^2}$$ None (already in real space)
Poisson Observation Model $$p(y_i\textbar E_i\lambda_i)={(E_i\lambda_i)^{y_i}e^{E_i\lambda_i}\over y_i!}$$ $$g(\lambda_i)=\log(\lambda_i)$$
Bernoulli observation model $$p(y_i \textbar \theta_i)=\theta_i^{y_i}(1-\theta_i)^{1-y_i}$$ $$g(\theta_i)=\log\left( {\theta_i\over1-\theta_i}\right)$$
Binomial observation model $$p(y_i \textbar \theta_i)={n_i \choose y_i}\theta_i^{y_i}(1-\theta_i)^{n_i-y_i}$$ $$g(\theta_i)=\log\left( {\theta_i\over1-\theta_i}\right)$$