Matching and clogit function

[Jane Allen]

I am trying to run a binary logistic regression to predict REIT takeover targets. I collected data of REITs that have been acquired from 2000-2017 (target = 1) and also all other non-target REITs in the same period (target = 0). However, because the non-target REITs data set is far larger than the target REITs data set, I would need to do some matching before I can run the regression.

I am not quite sure how I can do the matching with stata. I am thinking of matching by year and total assets or reit_id (subsector).

I've read another forum that says, but I'm not sure what to put in the group parenthesis, and if there are 2 different variables that I would like to match by:

egen match = group(XXX)

After matching, I also understand that I can then add this new variable into the clogit function as follows:

clogit depvar [indepvars] [if] [in] [weight] , group(varname) [options]

I've tried something like that, but I think the matched variable is wrong, and the results are not conclusive.

clogit target i.reit_id roaa roae noi_ta market_cap market_book pe_ratio cash_ta, group(year)

Would really appreciate if anyone can advise me on the matching process and the clogit process.

Thank you very much.

[Mike Lacy]

Why are you committed to 1-1 matching? I'd have two reasons for not thinking it necessary:
1) -clogit- does not require equal numbers within group;
2) It's not clear that matching is necessary (or even desirable) at all: You have a case-control (or "outcome selective") design. It's a common misconception that matching of event and non-event observations is a necessary (or even a good) strategy for controlling confounding. (See e.g.Rothman, K.J., Greenland, S. and Lash, T.L., 2008. Modern epidemiology. 3rd. Philadephia: Lippincott Williams & Wilkins.) Especially with just 2 variables to match by, why not just treat them as covariates in a regular binary logit model?

[Mike Lacy]

Whoops, I see that you apparently collected data on *all* members of the population at risk, so you don't have an outcome-selective design. Nevertheless, I don't see why you're committed to 1:1 matching.

[Jane Allen]

Hi Mike,

thanks for your response. Indeed, I am not committed to a 1:1 matching. I am not sure if a 1:n (about 4-5) matching is necessary.

If you recommend that i do not conduct the matching and just proceed with the binary logit model, do I use the logit function? Or still clogit function?

With the clogit function, does that mean I would not need to specify a group(variable)?

thanks so much for your response in advance.

the main problem I have is that my r2 appears to be too Low no matter what I run (about 0.15) and I thought matching would help in making the model a better fit.

[Mike Lacy]

Presuming assets to be measured as a continuous variable, and that you have an outcome variable = 1 if a firm WasAcquired and 0 if not, you could try:

logit WasAcquired c.assets i.year

You might also use a -clogit- model specified as:

clogit WasAcquired assets, group(year)

Note that the later precludes estimating the relation of year to WasAcquired.

Pseudo-R2 values for logit models are often low, and cannot be interpreted like a conventional R2, and are not in an ordinary way a reason that would motivate someone to choose a matched or fixed-effects model.

More generally: The questions that you ask here make me think that consulting a conventional textbook regarding these statistical models would help, as I think you have some misunderstandings that go beyond what we could address here. It sounds like you have picked up some "fancier" ideas but could use help with some more basic ideas.