Panel data - questions regarding suitable estimator

[Bart Do]

Hi all,

I have panel data on a company level between the years 2003 and 2023. For each year, we know (a) who the CEO was, and (b) certain properties of the CEO such as CEO enthusiasm.
Our goal is to explain what factors influence CEO enthusiasm, and perhaps some other CEO-level properties.

I have two questions regarding this:

(a) In principe the panel is based on company and year level. However, since we want to explain a CEO property, I reckon the best way is to use:

xtset ceo_ID year ?

(b) I am wondering what a suitable estimator would be for this situation. There are various considerations.
- The property of the CEO we try to explain is autoregressive. So a good estimator of CEO enthusiasm in year t is CEO enthusiasm in year (t-1)
- CEO's are nested in companies, and a company can have multiple CEO's on the time period 2003-2023.

Suppose I would therefore be interested in examining whether company financial distress influences CEO enthusiasm. Would a model such as the following be suitable:

xthybrid CEOEnthusiasm CompanyDistress, clusterid(ceo_ID) full

My guess would be that a standard FE model would not be suited as (i) it ignored between variance and (ii) has issues with the autoregressive component.

Thanks!

[Carlo Lazzaro]

Bart:
welcome to this forum.
1) given your research question, your -xtset- specification seems correct;
2) as far as the CEO's enthusiasm is concerned, I would simply cluster the standard errors on -ceo_ID- variable;
3) if your time-varying variables show little within-time variation, -fe- won't do and -re- should be considered;
4) -xthybrid- is also an option if you want to estimate a coefficient for a time-invariant variable. See also https://blog.stata.com/2015/10/29/fi...dlak-approach/

[Bart Do]

Bart:
welcome to this forum.
1) given your research question, your -xtset- specification seems correct;
2) as far as the CEO's enthusiasm is concerned, I would simply cluster the standard errors on -ceo_ID- variable;
3) if your time-varying variables show little within-time variation, -fe- won't do and -re- should be considered;
4) -xthybrid- is also an option if you want to estimate a coefficient for a time-invariant variable. See also https://blog.stata.com/2015/10/29/fi...dlak-approach/

Dear Carlo. Thanks for the warm welcome. Your suggestions are well taken, and I am going to look into it. Best, Bart