www.cmm.bristol.ac.uk/forum

Dear all,

I am using a multilevel and multiprocess event history model with MCMC estimation. In order to present my results in a clearer fashion, I have computed the average marginal effects following one of your answers on the runmlwin forum https://www.cmm.bristol.ac.uk/forum/viewtopic.php?t=877. However, I obtain really low predicted probabilities.

One possible explanation is that the data is regrouped in six months intervals to reduce the data size (this is done following the article by Steele, Goldstein and Brown 2004 in Statistical Modeling) . However, that I'm not sure how that would affect the predicted probability given that I'm predicting for a specific duration. I've compared my coefficients with similar models (for instance the Steele et al. 2004 paper again) and they seem similar in the sense that the constant is very large and negative, which overall pulls down the AMEs.

Do you have any experience with this issue? If it is the case that the AMEs are predicting the risk only in the following six months, do you know whether would be possible to adjust them (multiplying them for the median duration?)

Thank you in advance for your time.

Have a nice day.
Roberta

Hi Roberta,

The topic viewtopic.php?t=877 shows you how to compute average marginal effects for a single-level logistic regression model. The presented formula was simply the difference in two predicted probabilities where we changed the.

If you instead have a multilevel model you have the additional problem of what to do with the random effects when calculating these predicted probabilities. If you calculate predicted probabilities setting the random effects to zero you will not obtain population-averaged probabilities. Rather you will obtain population-median probabilities. Population-median probabilites are more extreme (further away from 0.5) than population-averaged probabilities. Hence, I am guessing, your lower than expected predicted probabilities. To obtain the population-averaged probabilities, you need to integrate out the random effects. You can approximate the intractable integral via a simulation approach.

See Module 7 (Concepts and then Stata practical) of our free online multilevel modelling course (LEMMA) for further details including Stata syntax.

Best wishes

George