Can run interaction term in Heckman treatment effect model and instrumental variable effect model?

[Chen Huang]

Hello everyone,

I am using Heckman two-stage treatment effect model and Instrumental variable model. I want to ask if I can use two-way interaction term in those two models, where the interaction code in stata would be A#B (where A and B are dummy variables).

I did that in Stata and it actually gives me some results. However, I noticed that the combination in the results only show as follow:

A#B
0 1 | 5.453087 2.730711 2.00 0.046 .1009911 10.80518
1 0 | 0 (omitted)
1 1 | 0 (omitted)

You can see there are two omitted. However, the last one with "1 1" is what I need the most for the regression. Anyone can help me with this issue? Thank you very much!


Chen

[Maarten Buis]

The problem is probably not with Heckman or IV. So we need to know more about your problem to identify the problem.

[Chen Huang]

The problem is probably not with Heckman or IV. So we need to know more about your problem to identify the problem.

Hello

I can not get good results in OLS with interaction terms. That's why I want to try Heckman or IV. I have instrumental variables and endogenous test rejected Ho hypothesis. But I have not seen that published paper who used this way which used interaction term in the paper. So my question is whether this way of doing interaction term is reseasonable. Many thanks.

[Maarten Buis]

What do you mean with "good results"? What did you type exactly in Stata, and what did Stata tell you in return?

[Chen Huang]

What do you mean with "good results"? What did you type exactly in Stata, and what did Stata tell you in return?

I mean to make the predictor variable to be significant. If I run interaction term in Heckman ot IV, it gave me some significatn results, but I think no one used this in the published paper,though i have not figured out why....but still Stata allows you to do that...

[Maarten Buis]

I suppose you know that significant in statistics is not a synomym for good or important.

I still don't know what you typed in Stata and what Stata told you in return....

[Chen Huang]

I suppose you know that significant in statistics is not a synomym for good or important.

I still don't know what you typed in Stata and what Stata told you in return....

Hi

I typed as such " etregress A B C D c.A#B, treat (E= F) two step first" In this example, I treat E and I use interaction A and B. And Stata returned me with significant results.

To be honest I am a beginner. but I guess sometimes if you really put two unrelated things in regression and both things are actually not intuitively related but still you get significant results, then it doesn't mean good results, am I right?

[Maarten Buis]

Significant is neither good nor bad, it just means that you reject the null-hypothosis (whatever that may be, you did not tell us) at some arbitrary level (often 5%, but you did not tell us what level you chose). However, fishing for significant results using ever more complex methods is definately bad. You need to have a good reason for using more complex methods. If you don't have those you should leave them alone.

[Carlo Lazzaro]

Chen:
with a quite large sample even a sneeze -/+ between two groups may be statistically significant, although negligible in practice.
As Maarten suggested, probably the best approach is to keep things simple(r) and mimick what others did in the past when dealing with the same research topic.

[Chen Huang]

Significant is neither good nor bad, it just means that you reject the null-hypothosis (whatever that may be, you did not tell us) at some arbitrary level (often 5%, but you did not tell us what level you chose). However, fishing for significant results using ever more complex methods is definately bad. You need to have a good reason for using more complex methods. If you don't have those you should leave them alone.

Hi

Yes, of course, statistically we say that it is significant at some point. 1%,5% or 10%. In my paper I choose 95% confidence and now I got result with statistically significant at 1% or 5%. Sure, I need some reason for using more complicated methods. Thanks for your help.