Fixed Effects Regression

[Joachim Vanderheyden]

I have a regression model and I have to investigate the effect in different cases (with/without time trend, fixed effects).
I want to run 6 different cases:

	Time trend	Corporate fixed effects	Year fixed effects
Case 1	Yes	No	No
Case 2	No	No	No
Case 3	Yes	Yes	No
Case 4	Yes	No	Yes
Case 5	No	Yes	No
Case 6	No	No	Yes

I think that I have the following codes:
Case 1:
regress nettotoegevoegdewaarde schulden omzet arbeidsintensiteit aantalwerknemers BBPgroei loonkost

Case 3:
xtreg nettotoegevoegdewaarde schulden omzet arbeidsintensiteit aantalwerknemers BBPgroei loonkost, fe

Case 4:
xtreg nettotoegevoegdewaarde i.year schulden omzet arbeidsintensiteit aantalwerknemers BBPgroei loonkost

Are these codes correct?

Are these the 6 most important cases to investigate or should I add/delete something more?

The main question is: how can i run a model without a time trend?

Thanks in advance.

[Carlo Lazzaro]

Joachim:
I'm not clear with the core of your question.
If you're not interested in time trend, you can:
- simply omit -timevar- altogether;
- consider -i.time- as a predictor.
As an aside, time trend can well be not linear; hence, if interested, you should check whether evidence of non-linear (say, squared) relationship between time as a continuous predictor and regressand exists with your data:

c.time##c.time

Eventually, predictor that are usually time-invariant (such as -i.industry) are wiped out by the -fe- machinery.

[Joachim Vanderheyden]

I want it without the time trend to check if the results are robust. Therefore I need to find out how I can run fixed effects models without time trends.

[Carlo Lazzaro]

Joachim:
simply omit -timevar- altogether.
As an aside: robustness with respect to what? If time has actually a role as a predictor in data generating process, if you omit it, your results are biased.

[Joachim Vanderheyden]

Thank you. I have to check it from my supervisor to check if the variable debt is in all the cases significant.

[Carlo Lazzaro]

Joachim:
have you already checked (via -hausman-) that -fe- actually outperforms -re- specification for your data?

[Joachim Vanderheyden]

Yes I have run the hausman test and it gives me the result that I need to use fe instead of re.

[Carlo Lazzaro]

Joachim:
great!
The first half of the journey is made, then!

[Joachim Vanderheyden]

Are the codes below correct with the accompanying explanation?

1) regress y x1 x2 x3 x4 x5 x6

Time trend: No
Corporate FE: No
Year FE: No



2) xtreg y x1 x2 x3 x4 x5 x6, fe

Time trend: No
Corporate FE: Yes
Year FE: No


3) xtreg y x1 x2 x3 x4 x5 x6 c.year, fe

Time trend: Yes
Corporate FE: Yes
Year FE: No


4) xtreg y x1 x2 x3 x4 x5 x6 i.year, fe

Time trend: Yes
Corporate FE: Yes
Year FE: No


I also need a specification with year FE. Is is the following code to run that specification?
xtreg y x1 x2 x3 x4 x5 x6 c.year#i.name_n

I have a dataset of 2000 companies, do I have to add 2000 dummies to run that specification? If I have to do that, how can i add that quickly?

[Carlo Lazzaro]

Joachim:
code 1) Ok;
code 2) Ok if your data were

xtset firm year

code 3) Ok if your data were

xtset firm year

and you're interested in linear realtionship between time and regressand
code 4) No. Code 4) includes firm -fe- and year -fe- if your data were

xtset firm year

Your last code (that I fail to get) will not do what you think it will do, as it simply interact time as a continuous predictor with a categorical variable (and omits the so called conditional maion effect of both of them).

[Joachim Vanderheyden]

Carlo:
Thank you for clarifying this.
the next step I want to do is adding year FE.

Are the following codes correct?

1) xtreg x1 x2 x3 x4 x5 x6 c.year#i.firm

Time trend: Yes
Corporate FE: No
Year FE: Yes


2) xtreg x1 x2 x3 x4 x5 x6 c.year#i.firm, fe

Time trend: Yes
Corporate FE: Yes
Year FE: Yes

If i want ro tun these codes, then I get the next error:

matsize too small
You have attempted to create a matrix with too many rows or columns or attempted to fit a model with too many
variables. You need to increase matsize; it is currently 400. Use set matsize; see help matsize.

If you are using factor variables and included an interaction that has lots of missing cells, either increase
matsize or set emptycells drop to reduce the required matrix size; see help set emptycells.

If you are using factor variables, you might have accidentally treated a continuous variable as a categorical,
resulting in lots of categories. Use the c. operator on such variables.
r(908);

I think that the problem is that I use Stata version IC. Will it be possible to run these codes in Stata SE of MP. I have the data of 2000 companies over a period of 5 years.

[Carlo Lazzaro]

Joachim:

xtreg x1 x2 x3 x4 x5 x6 c.year##c.year

I would tweak your first code a bit, just to explore if the relationship between the regressand and time trend is linear or quadratic.
Your code 2) complicates things, as you have already included -firm- groupwise effect via -xtset-.
I still fail to get why you're changing from -fe- to -re- specification and back again.

[Joachim Vanderheyden]

Carlo:
1) I intend to investigate the linear trend with the 1set code. do I still have to change it?

2) I change from fe to re because I have to test different specifications from my supervisor. I have to check the different speciifcations to check if my results are robuust.
Is the 2nd code correct to have a model with a time trend, with corporate FE and with year FE? Or want should the code be to run that model?

[Carlo Lazzaro]

Joachim:
1)

xtreg x1 x2 x3 x4 x5 x6 c.year

for the linear time trend
2)

xtreg x1 x2 x3 x4 x5 x6 c.year i.firm i.year, fe

for something that I would not sponsor.

[Joachim Vanderheyden]

Carlo:
1) I have that code but can I add something to make that code also with year FE? So I mean that code " xtreg x1 x2 x3 x4 x5 x6 c.year " plus something so it will be also have year FE?

2) I can not run that code. I get error r(908).
Explanation:
matsize too small
You have attempted to create a matrix with too many rows or columns or attempted to fit a model with too many
variables. You need to increase matsize; it is currently 400. Use set matsize; see help matsize.

If you are using factor variables and included an interaction that has lots of missing cells, either increase
matsize or set emptycells drop to reduce the required matrix size; see help set emptycells.

If you are using factor variables, you might have accidentally treated a continuous variable as a categorical,
resulting in lots of categories. Use the c. operator on such variables.