Omitted variables in pooled OLS model with interaction terms

[Lena Funke]

Dear all,

I apply a pooled OLS model to investigate the relationship between teleworking and job satisfaction with different interaction terms (all consist of dummy variables) using STATA 13.0.

What could be the reasons why some interactions are omitted?
If I add only one of the interactions this does not happen.

My interaction terms are e.g.:
Teleworking*Female
Teleworking*Male
Teleworking*Children
Teleworking*no children


Is there a difference whether I (1) generate the interactions myself using product terms and then add them to the model or (2) have stata generate them:

(1) gen Teleworking_Female = Teleworking*Female
(2) Teleworking##Female

I am very happy about help!
Thank you in advance.


Kind regards,
Lena Funke

[Marcos Almeida]

To start, it is much better to leave this job to Stata, I mean, estimating interactions.

Second, it is paradoxical, to say the least, to create two interactions with the same meaning. Instead, you could create and interaction between Telework and Sex, plus an interaction between Telework and HasChildren.

In short, saying "yes" for "no children" and "no" for Children is exactly the same thing.

To end, you may use # for interaction or, bettter yet, ## if you wish the so-called "main effects" as well.

Hopefully that helps.

[Carlo Lazzaro]

Lena:
as an aside to MArcos' excellent advice, please note that, as its face-value (ie, judging from low R-squared and wide 95% CI for some of your predictors) you need to check whether your model specification can be improved.

[Lena Funke]

Thank you a lot for your help Marcos Almeida !

Second, it is paradoxical, to say the least, to create to interactions with the same meaning. Instead, you could create and intercation between Telework and Sex, plus an interaction between Telework and HasChildren.

I understand your point! My intention was that I would like to have the coefficient for male teleworkers and female teleworkers in one model.

Otherwise, if I e.g. only include the interaction teleworking##female in the model, the Teleworking coefficient tells me the effect for the reference group = men. For women, I would have to add the Telworking coefficient + Teleworking*Female coefficients + Female coefficients, right? Or alternatively calculate it with marings.

Should I include all interactions (Teleworking##Female, Teleworking##Children etc. ) in one model?

Thank you!!


Best regards,
Lena

[Lena Funke]

Carlo, thank you for your advise!
How can I improve my model specification?
I unfortunately don't know what to do.

[Carlo Lazzaro]

Lena:
I would start from the usual -regression postestimation- tests, such as:
-estat ovtest-;
-estat hettest-;
-estat vif-.

[Marcos Almeida]

I believe my remarks were clear enough, but your reply in #4 demonstrates I need to be much clearer.

Taking the example of two variables, say, female (1 for female) and male (1 for male) , the combined result in a model is preposterous, you just need one variable, say, sex, and select 1 or 0 for one of the genders. The same for children.

I hope to have clarified your doubts.