inteff versus margins?

[Catie Bailard]

Hello all,

I am running a random effects logit model clustering on the country-level for time-series cross-sectional data set consisting of specific groups (approximately 800 groups living across 120 countries) over a three-year time period.

I am also testing for interaction effects of two continuous variable, and I know that there are difficulties associated with interpreting interaction effects in a nonlinear model.

I was using factor variables to represent the interaction (i.e., c.interactionterm1##c.interactionterm2) in xtlogit and then using the margins and marginsplot commands to graph the marginal effects across combinations of the interaction terms at various levels to look for significant interaction effects.

However, I have been told that the inteff command may be preferable for calculating interaction effects in a logit model? Is there a general consensus on whether one is absolutely preferable over the other in this context?


And, one additional, less important question is why I keep getting this result when I attempt to install the inteff package to my Stata 13 software:

"package name: st0063_1.pkg
from: http://www.stata-journal.com/software/sj4-3/

checking st0063_1 consistency and verifying not already installed...
installing into c:\ado\plus\...
could not rename c:\ado\plus\stata.trk to c:\ado\plus\backup.trk

r(699);"


I only have Stata 11 on my computer, so I am using Stata 13 through my university's cloud. Perhaps accessing the software through a cloud is why I cannot install this package? Is there any work around? (I do need to use 12 or 13 for the rest of my analysis, so I unfortunately cannot use the version installed on my computer, where I was able to successfully install inteff.)

Any advice anyone has on either of these matters would be very appreciated. Thank you!!!

[Catie Bailard]

I just realized this may be a non-issue. While running a revised xtlogit model in Stata 11 just to get familiar with the inteff command, I received this error message:

"factor variables and time-series operators not allowed
r(101);"

Then I realized factor variables do not work with inteff, so I took them out and created the interaction term and wrote the code as specific by Norton and Ai. But, then I got this error message:

"last estimates not found
r(301);"

I also ran it with the , re specification and got this error message:
"option re not allowed
r(198);"

Based on these messages, particularly the one stating that time-series operators are not allowed, it appears then that inteff does not work anyway with xt models. Can anyone confirm if this is a fact? (Just want to be sure the result is not because of some sort of coding error on my part.)

Thanks!

[Richard Williams]

The inteff approach has never had much appeal to me, partly because I don't fully understand it and partly because it seems difficult to use in anything other than limited cases. Vince Wiggins had a lengthy discussion and justification for the way Stata handles margins and interactions:

http://www.stata.com/statalist/archi.../msg00293.html

I'd like to see a FAQ or FAQs on margins that included his points, among others.

[Catie Bailard]

Thank you Richard. I will check this discussion out.

I also appreciate your input on the inteff versus margins debate. Granted, I clearly have limited exposure to the inteff command, but I tend to concur with your sentiment.

[Dirk Enzmann]

Having installed the most recent version of inteff via

net install st0063_1, from("http://www.stata-journal.com/software/sj4-3")

and checking the program code using viewsource inteff.ado shows that as matter of fact inteff does not work with xt models but only as a post estimation comand after logit, logistic, or probit:

<snip>

quietly {
        * check for correct estimation results
        if ~inlist("`e(cmd)'","logit","logistic","probit") error 301

<snip>

[Catie Bailard]

Thanks for confirming, Dirk. I appreciate it!

[Rosa Blau]

Hello, thank you to all contributors for this helpful thread. I have a question about interpreting margins for interaction terms (as done on p. 319 of Williams 2012). Williams interpretation discusses the estimators in the Margin column as X% chance of having diabetes. Is the term 'chance' different from it being a probability?

[Richard Williams]

I just double-checked, and I am using the terms chance and probability interchangeably in that paper.

[Rosa Blau]

Best of thanks!