The perfect method for estimating parameters in Generalised Linear Models is Maximum Likelihood Estimation. However, except in special cases, the likelihood cannot be maximised analytically. Hence, a numerical approach is desired for the computer implementation. The Iteratively Re-Weighted Least Squares (IRWLS) algorithm was developed for such purpose, and is described below.
Let be a twice differential function and suppose we want to solve . Start with an initial solution and let be the current solution. The Newton-Raphson method is based on computing the updated solution given by:
or equivalently:
and this procedure is iterated until the changes in successive solutions are insignificant.
Fisher’s scoring method is applicable when is a random function representing the log-likelihood and we want to solve . Start with an initial estimate and let be the current estimate. Let and . Then the updated estimate is given by:
(5.1) |
This procedure is iterated until the changes in successive solutions are insignificant.