Fixed Effects with Propensity Score Matching

[Gerard Dericks]

Hi,
I want to run a fixed effects model with propensity score matching (PSM) on 470 companies observed over 2 periods. 70 companies are treated, 400 are not. My problem is that fixed effects (FE) models have no constant, and I could not find a PSM stata program which has an option for no constant (I have two time periods so I just first difference the data and run OLS - which is equivalent to FE). Does anyone have any ideas about how I could run FE with PSM automatically in Stata?

Cheers

[Gerard Dericks]

Still looking for a way to run PSM in Stata with no constant in the second stage. Any ideas?

[daniel klein]

How about you tell us a little bit more about (a) what your plan is, in terms of how exactly you want to combine fixed-effects regression with PSM, and (b) what you have tried already and what exactly did not work for you.

I have a rather clear idea about fixed-effects models and also basic knowledge about PSM, but I have not yet thought about how these two can be combined. As far as I understand the philosophy behind PSM is all about making sure you base your conclusions only on the data you have really seen (or in other words compare only what you have observed and can therefore compare). This needs only few assumptions. If you successfully match, a simple t-Test might provide a valid answer. Regression approaches on the other hand basically extrapolate beyond your specific sample, by assuming a specific functional relationship between an outcome and a set of predictors. This involves more assumptions, but on the other hand allows you to compare what you have observed with what you have not actually observed but can account for in you model.

This said, I have no clear idea about combining these methods and (therefore) no clear idea why exactly a missing constant in an FE regression is bothering you.

Best
Daniel

[Henrik Dalriksson]

Have someone figured this out?

I tried to combined fixed effects in PSM through this formula:

teffects psmatch (y) (t x1 x2 x3 x4 i.year i.company), gen(match) nn(5)

Where y represent the independent variable, t represent the treatment variable and x1, x2, x3 and x4 represent covariates. Variables such as i.year and i.industry are fixed effects used to rule out fundamental differences in treatment firms throughout years and between industries. The problem with this is that most SIC codes are dropped because they predict failure perfectly. Moreover, I receive this error code:

perfect predictors are not allowed; try estimating your logit model before calling teffects psmatch

Best,
Henrik

[amanuel lulu]

i have tried with that and it is also my question? what is the solution then

[daniel klein]

The problem in #4 has little to do with the initial question. A fundamental assumption of PSM is that each observation must have a probability not equal to 0 (or 1) to receive treatment (or not). Within this framework you cannot use observations for which the covariates predict treatment status perfectly.

Best
Daniel