Clustering or Fixed Effects on firm & industry levels

[Panagiotis Vorgias]

Dear Statalisters,

I face the following issue:

I have a cross-sectional dataset where the observation unit is a pair of firms and look at the probability of a pair been "connected" (i.e. connected is the dependent variable and is binary). So, for each firm (firm A) I have several pairs with other firms (firm B) and firm B's are not the same for every firm A.
For each firm I have the industry code. I am thinking that for my analysis is relevant to control for unobserved correlation within industries and account for the non iid errors of pairs that refer to the same firm A.

So, I thought that industry fixed effects and clustered SEs at firm-level would be the appropriate specification. I tried this with xtreg (setting industry as panelvar) and returned the error "panels are not nested within clusters" obviously. I tried also with reghdfe command (absorb(firm industry)) which worked but given that I consider to perform logit regressions I intended to go with xtlogit.

And the questions are:

- Has anyone any comment on what the appropriate specification would be in such a setting? (clustering, FEs, or both and at what level(s)?)

- Is there a way for reghdfe with logit regression or any alternative to this?


Thanks in advance.

[Phil Bromiley]

You'll increase your chances of a helpful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output, and sample data using dataex. I don't see how you do this analysis without making each pair a separate observation (which sort of looks like a panel).

This sounds like a network problem - you should look at the network estimation literature. Actually, I think Stata has a seminar or webinar on network estimation. Networks present complex covariance/endogeneity problems.

If you're only going fixed effects, then at least for the linear case you can include i.var in the regressors. I don't know if this works for logit (and I think "fixed effects" logit is a different beast the fixed effects regression). Depending on your discipline, many like the linear probability model which is what reghdfe gives you.

[Panagiotis Vorgias]

Thank you Phil. I apologise for the incomplete post. Apparently I have lived enough into this problem that believed it could be seen clear without the proper presentation. My bad.
Thanks for your advice. I will try towards your direction and some ideas I have and may come back again later.

[David Benson]

I assume you are looking at some variable whether the firms decide to form an alliance or some other partnership. If the two firms are in different industries, which industry are you wanting to include as your fixed effect?

If firms are involved in multiple partnerships over time, you might, as Phil mentioned, just include i.industry in your model and cluster on firmA (as you have done). I assume that the industry fixed-effect is to control for the fact that firms in some industries are more prone to partner than firms in other industries.