A Penalized h-Likelihood Variable Selection Algorithm for Generalized Linear Regression Models with Random Effects
Algorithm 1
[(Step 1)] (initialization).
(i)
Assume a partial linear model excluding variable selection. Express in a parametric way. For example, a cubic regression spline can be expressed by using the truncated power basis:
where the 5 knots are percentiles of , are the associated coefficients, and , are the numbers corresponding to the cubic regression spline representation.
(ii)
Initialize the fixed effects , where is the h-likelihood estimates by treating in a parametrical way. Then, we have
(iii)
Denote the estimates by :
where are the h-likelihood estimates.
(iv)
Determine initial value for random effects using
with and .
[(Step 2)] (loop).
(i)
Use and to get
(ii)
For the iteration, set the estimator from the iteration and update by
(iii)
For , set if , for a small cutoff value c.
(iv)
Compute and compare to a small predetermined value . If is smaller than , stop the loop.