How to choose control group=

[Mislav Sagovac]

I am trying to estimate the treatment effect of debt forgiveness on investment using difference-in-differences (DID) estimator. My code looks like this:

xtreg i treatment i.reportyear treatment##t, fe

where i is investment rate (outcome variable), report year = (2011, 2012, 2013, 2014), treatment is equal to 1 for firms that experience debt forgiveness in 2013 (treatment group) and 0 otherwise, t is equal to 1 for post-treatment period (2013, 2014) and 0 otherwise. I will also use control variables that I are usually used in investment equations but for know I abstract from them.

I am not sure which firms to include in control group? Since I try to estimate how improvement in financial situation of financially distressed firms effects investment, I was thinking to include firms that are illiquid (current ratio smaller then 0.6) or insolvent (tenure ratio bigger than 1) in 2012, but this criterion seems arbitrary so I would like to know what would be more systematic approach? My treatment group have 200 firms and control group can be chosen from all other firms in the country (sample of around 50.000 firms).

I suppose one approach is to use propensity score matching, but I am not sure how to apply it when I have fixed effects model (with multiple time periods).

[Carlo Lazzaro]

Mislav may be interested in the following paper http://www.nber.org/WNE/lect_10_diffindiffs.pdf.
That said, some remarks about your query:
- If you're going to use DID estimator, shouldn't you run an OLS on the differenced data (instead of -xtreg-)?;
- why repeating -treatment- when your interaction code will give you back both conditional main effect and interaction with t of that predictor?;
- about firms that are illiquid, a possible classification may probably be obtained by the literature in your research field.

[Mislav Sagovac]

Mislav may be interested in the following paper http://www.nber.org/WNE/lect_10_diffindiffs.pdf.
That said, some remarks about your query:
- If you're going to use DID estimator, shouldn't you run an OLS on the differenced data (instead of -xtreg-)?;
- why repeating -treatment- when your interaction code will give you back both conditional main effect and interaction with t of that predictor?;
- about firms that are illiquid, a possible classification may probably be obtained by the literature in your research field.

1) I thought I have to run -xtreg- command if I have panel data? Isn't first differenceing an option when try to estimate treatment effects (I can use levels also)?
2) You are right. I don't need treatment variable (actually it is removed because of collinearity)
3) I have chosen variables which can approximate liquidity but I was thinking about something else. It would be better maybe to apply matching procedure. In baseline period, I could use match companies from treatment group and comparison group. I know fo -teffects nnmatch- command, but I don't know how to save matched pairs when execute the command for one year?

[Carlo Lazzaro]

Mislav:
thid old Stata thread can be helpful http://www.stata.com/statalist/archi.../msg00738.html.
Another interesting entry (with Stata examples using -areg-) is the following one:
http://didattica.unibocconi.it/mypag...0221112713.pdf.

[Mislav Sagovac]

Carlo,

Would you suggest following approach: take mean of two per-treatment years (2011, 2012) and two post-treatment years (2013,2014), and than run the code:

reg i treatment#t

where i is outcome, treatment = 1 for treatment group and t =1 for post period? I have read pdf file you posted, and it seems to me -xtreg- and reg would give the same answer if there are two periods as in code above? And it is simpler.

Bsut I'm still not ure how to do matching in pre-period? How can I match groups in treatment and observations in comparison group via nearest neighbor match?

[Carlo Lazzaro]

Mislav:
as far as I can get your query, my first thought goes out to -areg-.
As far as the second part of your post is concerned, unfortunately I'm not that familiar with propensity score matching, hence I cannot advise on it any furher.

[amanuel lulu]

hello there i am new to use stata and i couldn't know to how perform fixed effect model after propensity score matching. thank you for your help
.

[amanuel lulu]

hello there,i have problem with comparing binary outcome on covariates using baseline year (reference year) on panel data. hope to hear from you soon

[Clyde Schechter]

Please read the FAQ for advice on how to post questions in ways that increase your likelihood of getting a helpful, timely response. In particular, your question is too broad. Make it more specific and explain in greater detail the nature of your task. As written, your question almost demands that somebody write a couple of chapters from a statistics textbook to answer you. Focus your question. And it would almost certainly help to show an example of your data, using the -dataex- command.