I have recently estimated a model using fixed effect ordered logit with blow up and cluster (buc) estimator. As this method reports pseudolikelihood, I could not perform a likelihood ratio test with other comparable models that report log likelihood. Wald chi square was substantially lower in buc estimator than ologit while Pseudo R2 was also lower in buc. This confuses me as I thought buc should have a better goodness of fit.
Any thoughts on the above is highly appreciated.
This seems reasonable to me
when you take account of clustering the effective sample size will be smaller as you do not have as many observations as you think to had; the standard error has previously been underestimated and you were potentially committing a Type 1 error.
Think about the problem of how many blonde people there are in the world- if you pick 1 hair from 5000 heads it has a lot more information than 5000 hairs from 1 head - the cluster estimate would increase the standard error of the estimate of the blondness estimate.
- Badges
- Science topic