I am presently using the mediation package in R-statistics and am trying to fit the mediator using an ordered probit model. The problem is that I return the following error: "insufficient variation on mediator". I suspect this is due to ~90% of zero values (sample size ~11,000).
I have also found the following paper, which provides a solution in SAS, although I do not have this software:
Wang, W., & Albert, J. M. (2012). Estimation of mediation effects for zero-inflated regression models. Statistics in Medicine, 31(26), 3118–3132. doi:10.1002/sim.5380
Any thoughts or suggestions?
If you have access to Stata, I suggest you try the user-written program khb. I believe it will handle your overdispersion issue.
I am also providing a link to a paper I recently wrote on various mediation models. KHB was one of the models that was tested, and it performed quite well on variety of different simulation models.
Article Using mediation analysis to identify causal mechanisms in di...
Dear Ariel, thank you so much for your reply! This is very helpful. Just for the record, here is a reply to an email I sent to the authors of the mediation package in R-statistics:
On 20/03/2015 3:44 pm, "Teppei Yamamoto" wrote:
Dear Andrew,
Ordered probit would probably not work well in your case, because you
likely have too few observations in some of the mediator levels.
The mediate function in our package doesn't support negative binomial or
zero-inflated poisson regressions out of the box. The algorithm behind
it can in theory, however. So you could try and implement the algorithm
on your own, either modifying the mediate function or writing one on
your own. The best reference for that would be this paper:
http://imai.princeton.edu/research/BaronKenny.html
An easier alternative would be to use a poisson model. This would "work"
in the sense that you will get estimates from the function, but results
may not be reliable because of the poor fit of the mediator model. You
may want to run some model diagnostics on the mediator model and see if
the fit may perhaps be adequate.
You could also use ordered probit if you coarsen the mediator such that
it takes on several levels like "zero", "low", "medium" and "high" for
example, if you can set reasonable thresholds. Of course, this would
through away information, but you can at least compare the inferences
from that model and a zero-inflated negbin model and see if results are
roughly the same.
In a nutshell, the question seems to boil down to whether you can find
an alternative model for the mediator that fits reasonably well by
standard metrics (which you can do outside the mediation package).
Hope this helps!
Best,
Teppei
Yes indeed. I spent quite a bit of time engaging the developers of the "mediate" approach before and during the writing of that paper I sent you. I think the approach is a good one, but I like the straightforwardness and general applicability of khb.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling