[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: st: possible problem with GLM command
Just a quick thought as to what may be the problem. Using the log link with the bernoulli family (binomial with a response of 1 or 0) results in fitted values that may lie outside the 1/0 bounds, which is an assumption of the Bernoulli parameterization of the binomial family. Because of this, and depending on the specific nature of the data, convergence problems may occur. You may find that using the -irls- option clears up the problem.
James Hardin, who co-authored the glm command with me, wrote a program for binomial models with non-standard links some years back. It was published in the STB. Fitted values that exceeded 1 or 0 were internally truncated at each iteration. I forget the name of the program, but you can probably find it by using the -lookup binomial- command in Stata. It should work.
> Dear all
> Is anyone aware of a problem with the GLM command?
> I have recommended to a number of my collegues that they fit a
> GLM with a log link and binomial error term when analysing cross-
> sectional data with a dichotomous outcome. The benefit of this is
> that the exponental of the paramater estimate is equivalent to the
> relative prevalence. [I'm not looking for a debate on the wisdom of
> A number of them have reported a problem with their models when
> they have several independent variables. For example I have a
> model with three categorical predictors: agegroup (3 levels), sex (2
> levels) and housing (2 levels).
> The problem is that I get a message indicating that the log
> likelihood function is not concave. Although this is possible I am
> able to run the model without any problems is SAS. Furthermore,
> there doesn't seem to be anything in the data that suggest there
> would be a problem, i.e. all covariate patterns have a reasonable
> number of observations for each level of the outcome of interest.
> Surprisingly to me, the problem goes away if I add an interaction
> term for sex and housing, although the interaction term is
> statistically significant and the data suggest it should be.
> Any suggestions are greatly appreciated.
* For searches and help try: