Hi,
I am running a 3 level logistic model (level3:houseID, level2:dogID, level1:isolateID) and was wondering if there is a way to examine the 1st level (isolateID) residuals?
Also, are there any postestimation commands available?
Thanks!
Erin
Residuals & post estimation
-
GeorgeLeckie
- Site Admin
- Posts: 432
- Joined: Fri Apr 01, 2011 2:14 pm
Re: Residuals & post estimation
Hi Erin,
Currently you can't use runmlwin to calculate level-1 residuals for discrete response models.
You could caculate the predicted probability for each observation and then calculate the level-1 residuals as the difference between the observed binary response and the predicted probability.
I have to admit, I don't find level-1 residuals are not ususally very helpful quantities to look at in discrete response models and this is probably why we haven't implemented anything.
George
Currently you can't use runmlwin to calculate level-1 residuals for discrete response models.
You could caculate the predicted probability for each observation and then calculate the level-1 residuals as the difference between the observed binary response and the predicted probability.
I have to admit, I don't find level-1 residuals are not ususally very helpful quantities to look at in discrete response models and this is probably why we haven't implemented anything.
George
Re: Residuals & post estimation
Hi George,
Thank you for your quick response. I will have to look into calculating the level one residuals.
Also, are there any post-estimation commands available for runmlwin that would be similar to the <lrtest>? Just looking for something to compare full and reduced models while model building.
Thanks again!
Erin
Thank you for your quick response. I will have to look into calculating the level one residuals.
Also, are there any post-estimation commands available for runmlwin that would be similar to the <lrtest>? Just looking for something to compare full and reduced models while model building.
Thanks again!
Erin
-
GeorgeLeckie
- Site Admin
- Posts: 432
- Joined: Fri Apr 01, 2011 2:14 pm
Re: Residuals & post estimation
Hi Erin,
MLwiN uses maximum likelihood estimation to fit continuous response multilevel models. You can therefore use Stata's lrtest command post-estimation for these models.
MLwiN uses quasi-likelihood estimation to fit discrete response multilevel models. This is only an approximate method. We therefore do not present the likelihood for discrete response models and it is not possible to use Stata's lrtest command post-estimation for discrete response models. In terms of adding fixed effects to the model you can use the Wald tests (Stata's test command)
As quasi-likelihood estimation is only approximate, we recommend MLwiN users (and therefore runmlwin users) to fit their final discrete response multilevle models using MCMC estimation. Users can then use the Bayesian DIC statistic (similar to the AIC and BIC statistics) for model comparison purposes. See the MLwiN MCMC manual for more details.
Best wishes
George
MLwiN uses maximum likelihood estimation to fit continuous response multilevel models. You can therefore use Stata's lrtest command post-estimation for these models.
MLwiN uses quasi-likelihood estimation to fit discrete response multilevel models. This is only an approximate method. We therefore do not present the likelihood for discrete response models and it is not possible to use Stata's lrtest command post-estimation for discrete response models. In terms of adding fixed effects to the model you can use the Wald tests (Stata's test command)
As quasi-likelihood estimation is only approximate, we recommend MLwiN users (and therefore runmlwin users) to fit their final discrete response multilevle models using MCMC estimation. Users can then use the Bayesian DIC statistic (similar to the AIC and BIC statistics) for model comparison purposes. See the MLwiN MCMC manual for more details.
Best wishes
George