binomial fixed-effect p-values by simulation
On 24/08/2008, at 2:16 AM, Daniel Ezra Johnson wrote:
What would be the difference between simulating the z-statistic (if I'm getting this, it would be determining what proportion of the simulations have a z-statistic as large as the one from the observed data) versus doing the same thing with the difference of log-likelihoods)? One difference I see is that with the z-statistic approach there is no need to fit a null model, only the alternative model (to data generated by the null model)? D
They should produce similar results and which is better is probably of theoretical rather than practical interest. I will try this with a commercial program I have that does both. Ken