Fixed effects with multiple intersect categories

[Alex Beck]

Hi there,

I have a problem with a fixed effect regression in stata and no idea how to solve it. I have a sample with about 500 companies which are founded in different years and operate in different industries.

With a couple of variables, I want to measure the success of this companies (OLS). Of course, the founded date is from importance due to different economic environments. No problem with this, classical xtreg with year fe.

The problem: each company operates in multiple industries and has no unique category. For example, a company can operate in the field of health care as well as biotechnology and another company operates in health care and ecommerce (up to 5 different categories). In some cases, the industry is highly correlated with the success measurement (y) therefore, fixed effects would be quite useful!

Any idea how to implement fe in this case? Tried it with dummies and weighted dummies but it didn’t work out that well.
Any ideas?

Kind regards
Alex

[Carlo Lazzaro]

Alex:
welcome to this forum.
As it is always the case (see the FAQ), please post what you typed and what Stata gave you back, starting from the way you -xtset- your data. Thanks.
In addition, you can share an example/excerpt of your data via -dataex-.

[Alex Beck]

Hi Carlo,

honestly, the problem is, I have no idea what I should type :D

I have different companies with different industries (2-5). In total there are 40 different industries. The problem is the intersect between the industries. There is no unique industrie for the companie.

Maybe it is rather a statistic question than a specific stata question.
All in all I look for an idea to handle this kind of problem. Surprisingly enough, google wasnt much of an help, allthough it doesnt sound that uncommon for me.

Thats the code so far. Nothing special. And no industry fe.

 xtset foundeddate

 xtreg success x controllvar1 contrallvar2, fe

Kind regards
Alex

[Carlo Lazzaro]

Alex:
the first comment is that -panelid- is mandatory in -xtset- and -foundeddate- sounds as a -timevar- to me.
If the dependent variable of your regerssion is the successful performance of the company, you can simply add a predictor containing the number of industries each company is interested in (say: 1,2, ...5).:

xtset company foundeddate
xtreg success <depvar> controlvar1 controlvar2 i.industry, fe

[Alex Beck]

Succes (dummy)	indep var (dummy)	Biotech (dummy)	eCommerce (dummy)	Health (dummy)	foundeddate (year)
0	1	1	1	1	2015
1	0	1	0	0	2015
1	0	0	1	1	2017
0	0	1	1	0	2016

Hey Carlo,
Thanks for your quick response. You can imagine the sample like the table above (way bigger obviously).
The number of industries it not the interesting thing. Its more important that there is maybe a positiv effect of e.g. Biotech on the succes.

Kind regards Alex

[Carlo Lazzaro]

Alex:
thanks for providing further details.
I'm not sure that you actually have a panel dataset, as what I've perceived as a -timevar- is in fact the year of foundation.
Hence the question is: have you got panel (ie, the same company measured along years) or cross-sectional data?

[Alex Beck]

I admit this wasnt really clear sry. I have cross sectional data. The success measurement is fix. The company was successful or not, I dont aim to observe changes over time for a specific company. With the time fixed effects I want to exclude macro economicall impacts (as good as possible).

[Carlo Lazzaro]

Alex:
if identifying the predictors that made the company successful or not is the aim of you research, your dependent variable is not continuous but categorical (yes/no).
Hence, OLS is out debate here, and you should better switch to -logit or -logistic-.

[Alex Beck]

# Patents	indep var (dummy)	Biotech (dummy)	eCommerce (dummy)	Health (dummy)	foundeddate (year)
10	1	1	1	1	2015
1	0	1	0	0	2015
12	0	0	1	1	2017
3	0	1	1	0	2016

Sure probit or logit would be better in this case.
But back to the topic. If I would like to measure another dep var which is continious (e.g. Number of Patents).
How could I implement industry fe then?

[Carlo Lazzaro]

Alex:
-Patents- is not continuous, as it can take on integers only.
Hence, you should consider -poisson-:

poisson Patents controlvar1 controlvar2 i.industry

Beware of extradispersion (that in most cases reads overdispersion) in -poisson- (see -nbreg- entry in Stata .pdf manual).