Using QIC for optimal model estimation of GEE models (problems with interaction terms)

[Eeshan Khandekar]

I've been using a general estimating equation (with a negative binomial family, exchangeable correlation structure, and log link) to model response time in a surveillance data-set. Essentially, I am trying to model the change in response time over time to reported malaria cases. I addition to the actual response time and time variables, I have additional control variables (i.e. the district a case is reported in, number of reported cases, the household size - which is an important metric for workload burden in the surveillance system, and Fourier terms to control for periodicity).

Here the full model I am attempting to run:

time_diff_days DayRank i.district##c.NewCase_Count hhld_size sin_DayRank cos_DayRank if time_diff_days>0 & DayRank>35, fam(nb) link(log) i(dmso) corr(exc) eform

time_diff_days is the response time in terms of days
DayRank represents time
i.district is the district the case is reported in
NewCase_Count represents the number of cases
hhld_size represents the number of individuals in a household
sin_DayRank & cos_DayRank represent Fourier terms to control for periodicity.

I have been able to use QIC to assess if the addition of variables (excluding the interaction term) have made a better model. Thus far, I have been able to conclude that the following model has the smallest QIC value. I should note here that, instead of merely using qic command before the model, I have had to add the prefix "xi:" to make it work.

time_diff_days DayRank i.district NewCase_Count sin_DayRank cos_DayRank if time_diff_days>0 & DayRank>35, fam(nb) link(log) i(dmso) corr(exc) eform

However, when I attempt to estimate the QIC value of the model with the interaction term, I get the following error: "interaction terms not allowed"

Is there a way around this, or a method in which to force STATA to estimate the QIC?

I hope I explained this right. Please let me know if anyone has any questions.

Thanks,
Eeshan

[Andrew Lover]

If I remember correctly, QIC (from SSC as per the FAQ) is an older .ado, and uses the deprecated -xi- prefix syntax (see -help xi-). This may help.
Also, depending on how many levels your -district- and -NewCase_Count- vars have, you might be hitting a limit on how many combined levels QIC can handle (just a guess).