Generalized Linear Models
Best for: Structured statistical modeling
How it works
$$g(\mu)=\beta^\top x$$Extends linear regression to any exponential-family response via a link function $g$ relating the mean to a linear predictor, $g(\mu)=\beta^\top x$, so $\mu=g^{-1}(\beta^\top x)$. Standard choices are the identity link (Gaussian/linear regression), the logit link (Binomial/logistic) and the log link (Poisson counts). Coefficients are estimated by maximum likelihood via IRWLS, reweighting observations by the variance function at each iteration.
Common fields
Biostatistics · actuarial science · econometrics