Can I test the interaction effect while using the "teffects?"

[jieun kim]

I am a neophyte in using the Stata.

My question is that....While using the teffects, can I test the interaction effects?

If I elaborate what I wish to test using the stata example data,

webuse cattaneo2

If I wish to know the effects of mother's smoking on the birth weight, while matching the cases in terms of covariates, I would execute the line below.
teffects psmatch (bweight) (mbsmoke mmarried c.mage##c.mage fbaby medu)

My situation is that I should test whether there is any interaction effect of mbsmoke*alcohol(1 if alcohol consumed during pregnancy) while the cases are matched in terms of other variables.

Thank you for your time and effort.

[Benjamin Fisher]

I see that this post is old, but I'm wondering if you ever found a solution to this issue?

I am trying to run a similar model, but using a single covariate rather than an interaction. I have estimated the effect on a matched sample, but want to know how the effect changes when conducting the same analysis, but controlling for one theoretically relevant covariate.

Any advice would be much appreciated!

[Enrique Pinzon (StataCorp)]

Dear Benjamin and Jieun,

Notice that the result of your estimation is a treatment effect. This is based on calculations of probabilities using the covariates you are talking about. Therefore, the covariates affects the probabilities and then these probabilities affect the matching. You could potentially ask if the covariates have any effect in the probability, which can be done running margins after logit or probit. However, the question of how the covariates independently affect the treatment effect is not something that can be understood as a direct effect.

[Benjamin Fisher]

Dr. Pinzon,

Thank you for the reply. Is there another teffects command besides psmatch that would allow one to conduct the types of tests described above? Or does the same issue apply to all estimation models?

[Enrique Pinzon (StataCorp)]

Dear Benjamin,

The short answer to your question is that there is no way of doing this directly. The long answer makes some assumptions about what you want. So here goes the long answer:

* The treatment effect is what you want, so any other computation is auxiliary. Perhaps you are concerned about incorporating a variable in your model that is only providing noise or excluding a variable that is extremely important. Based on these two assumptions I suggest:

- Run all the -teffects- estimators available (ra, ipw, ipwra, aipw, psmatch, nnmatch). This implies that you will have to model both the treatment and the outcome. (In the case of psmatch you are modelling only the outcome). Verify what happens to the treatment effect across all estimators with or without the interaction term you are concerned about. If the treatment effect does not vary across estimators with or without the interaction, the model is robust. If the treatment effect changes when you include the interaction, you need to explore which model is correct.

-For the previous analysis look carefully at what happens with the aipw and ipwra estimators. They have a property called double robustness. This means that if you specify one (and only one) of the models (treatment or outcome) incorrectly you can still get correct treatment effects. This piece of information will give you further understanding of how robust your estimation is.

-Look at the balancing diagnostics and overidentification test after your psmatch estimator. In particular what happens to balancing of your covariates when you include or exclude the interaction term you mention. Verify that balancing is satisfied and that your results are robust to the inclusion or exclusion of the covariate you are worried about.

The long answer is that you can explore what is occurring when you exclude or exclude a regressor but that a formal test of the effect of a covariate on your specification is not available. Also, my belief is that ultimately what you would like to know is the effect of a treatment on an outcome.

[Enrique Pinzon (StataCorp)]

Erratum: I stated "(In the case of psmatch you are modelling only the outcome)" I meant to say the treatment instead of outcome.

[Benjamin Fisher]

Thanks for this reply - it is very helpful and I am going to work on these additional models. In terms of what I ultimately want to know, yes I am interested in a treatment effect on an outcome. However, I'm also interested in whether that treatment effect depends on the characteristics of the sample. In this case, my unit of analysis is high schools, and I want to know if the treatment effect depends on schools' racial composition, for example.

[Philip Jones]

-duplicate-, sorry