Panel Data with Time invariant variables

[Nicole Kristine]

Hi everyone,

I am new to Stata and am currently studying the effect of CSR on daily returns during a stock market crash. My dataset consists of 50 firms (I assume that my number of firms is still too small and I will need to increase my sample) over the time period of 35 days. This means that my independent variable changes daily but CSR doesn't and also some firm sepcific accounting data, which are used as control variables, are time invariant. I also wanted to include industry dummies to control for effects in different industries during the crash.

Now, I know that I cannot use xtreg, fe since all time invariant variables will be ommitted. Does this mean that my only solution is to apply a random effects model? Can I still incorporate industy dummies in this model? If so, how?

Thank you in advance, I really appreciate your help.

[Carlo Lazzaro]

Nicole:
welcome to this forum.
As you correctly stated, the -fe- machinery wipes out all time-invariant predictors (and variables at large).
If company do not change -industry- as time goes by, the related coefficient will not be estimated (as time-invariant) under -fe- specification.
A different approach might be -re- (provided that a panel-wise effect is evident) or the Mundlak correction (see the community-contributed programme -mundlak-).

[Nicole Kristine]

Thank you Carlo for your fast reply!

I am currently looking at the following paper by Lins et al (2017): https://papers.ssrn.com/sol3/papers....act_id=2555863
Here,they are also testing the impact of CSR on returns during a stock market crash.

I am wondering how they were able to incorporate industry dummies in their baseline regression (p. 1798) and what regression model exactly they were using. In my understanding they were also using panel data on firm level, that is what is confusing me.

[Carlo Lazzaro]

Nicole:
as far as I can get from a quick glance given to regression tables, it seems that Authors used an OLS with -robust- standar errors to take heteroskedastcity into account.
Industry is simply -i.industry- when translated into Stata coding.
Your confusion is probably caused by the meaning of Panel (A and B) reported in the paper, that does not mean panel in econometric sense, but is simply of way for indicating different tables in a paper (especially when Instructions to Authors impose constraints on the number of tables/figures, one of the most frequent trick of the trade is to name two different tables/figures as Table/Figure 1 Panel A and Table/Figure 1 Panel B).

[Nicole Kristine]

Thank you Carlo for taking the time to look at the results in the paper! It really helped.

I actually had to go a step back in my analysis and I am looking for linearity in my data. I tried to see the linear relationship, which I assumed would be there, with the help of scatter plots. As an example of my scatter plots I attached one of them. It shows my independent time variant variable returns and dependent time variant variable LN Market Capitalization and looks like the following. That's clearly not a linear relationship .I am confused whether the command scatter x y is correct for panel data or whether I need to cluster them somehow by firm or something like this? And how is it even supposed to work with time- invariant vairables. I searched the forum and Stata but couldn't find anything helpful.
Maybe I'm misunderstanding a crucial point. Related literature also showed a linear relationship.

[Carlo Lazzaro]

Nicole:
please remeber that linearity in OLS relates to coefficients, not variables.
Hence, it is totally legal to have a squared relationship between the regressand and one (or more) predictor(s).
Despite your graph was left unattached in your last post, it is often the case that regressand and time (when considered as a continuous predictor) show non-linear relationship.
Again, I do not think that authors used a panel regression related command, such as -xtreg- in Stata, but ran in fact an OLS.

[Nicole Kristine]

Oh there must have gone something wrong. I can see the picture but I will post it here again.

[Carlo Lazzaro]

Nicole:
have you already tried to include:

c.year##c.year

among your predictors?

What does Stata give you back? (Please post both code and outcome table within CODE delimiters. Thanks).

[Cristina Leone]

Hello everyone,

I'm new in Statalist and I posted one question yesterday but I didn't receive any reply. Maybe replying to a post already existing in the forum can be useful to receive help in the resolution of the issue. Here is my post: https://www.statalist.org/forums/for...ith-panel-data. I would appreciate so much any type of suggestion. Thanks!
Best,
Cristina