Missing data for non-treated group in logistic regression interaction

[Roger Clements]

Hi all,

I am trying to figure out how to handle missing observations for the non-affected firms in an interaction using logistic regression. I have two outcomes for my dependent variable: companies that get targeted by a journalist (1) and companies that do not get targeted by a journalist (0). I can run a main effect, where I look at the effect of CEO gender on activism. Straightforward.

Now I want to test an interaction whether I interact CEO gender X journalist experience to see whether journalists' experience moderates the relationship between CEO gender and probability of targeting by a journalist. However, the issue I am facing is that I only have journalist experience data for companies that were actually targeted by a journalist (1), and nothing for companies that do not get targeted (0).

I see two ways to handle this:
1. I code journalist experience as 0 for all companies that do not get targeted (0)
2. I only analyze the the subsample of companies that get targeted by a journalist (1) and look at the moderating effects of journalist experience within that group. However, I do not see how I can do this since the outcome (targeted) never varies (i.e., it is always 1).

Any advice on how to handle this dilemma would be appreciated.

Roger

[Phil Bromiley]

I would suggest your follow the FAQ on asking questions.

Including the journalist on the rhs seems tricky - it inherently depends on the dv (can only exist if dv=1). Indeed, all the journalists exist in any given year, so your journalist experience is endogenous in a complex way.

If you have global data on journalists, you might be able to do something . Maybe there is a way to include journalist in calculating a dv? Maybe a heckman selection model? But, this is reaching.

[Roger Clements]

Thanks Phil. I'll explore the heckman option. I agree that it might cause more issues than it would solve. I appreciate your thoughts. I also stumbled upon the firthlogit command today, so I tried it out in Stata and it gave me results (whereas xtlogit wouldn't even run). I need to dig further though to see what is happening.

I will also read that FAQ on question asking.