[Carlo Lazzaro]

Dominik:
the results of test Jeff suggested (which is similar to -linktest-) is telling you that your model is misspecified (and you have heteroskedasticity, too).
Misspecification is the relevant issue (whereas heteroskedastcity can be accomodated with -robust- standard errors), as it may mean different things:
- you simply forgot to plug in one or more predictors (and/or interactions) that complete the data generating process;
- linked to the previous point, you may have endogeneity.
You may want to log the regressand and see if any improvement comeas alive.

[Dominik Miksch]

Thank you very much for your reply Carlo Lazzaro .
I already had the regressand in log form for the test.
But does the result of the test Jeff suggest also tell me nonlinearity is an issue in the model I tested? As Jeff said 'If you're really interested in seeing whether nonlinearity is present, use a heteroskedasticity-robust RESET, or something like it. You can do it "by hand:

And then I did the test Jeff suggested

What I got from all the helpful input is "The appearance seems consistent with the residuals not being a nonlinear function of the fitted values. But statistical testing let us reject the null hypothesis at the one percent level what provides evidence that nonlinearity is present in thre model I tested"

Kind regards
Dominik

[Jeff Wooldridge]

If you are estimating a treatment effect, the test only makes sense if you perform it separately for the treatment and control groups. Otherwise, you may simply pick up different slopes on the covariates for the control and treatment groups.

Generally, you need to understand that rejecting the particular linear regression that you've tried does not mean you should reject regression adjstument. What happens when you include squares and interactions of the covariates? Maybe you'll pass the test. At a minimum, you should do RESET on the control and treatment groups separately.

Why do you care about heteroskedasticity? That is not an issue that affects consistency of any of the treatment effect estimators. Heteroskedasticity is not functional form misspecification. You already have evidence you need to do better with functional form.

[Dominik Miksch]

Jeff Wooldridge Thank you very much for your reply. Your input helped me a lot.
According to what you said, I performed the test for both groups- treatment and control.
I also adjusted the model meaning I included squares and interactions of the covariates. Every P-Value was significant at the 1% level for the treatment and the control group respectively.
Therefore I argue: statistical testing let us reject the null hypothesis at the one percent level what indicates that nonlinearity is present for both, treatment and control group with the five models I tested. I try to include the tables and graphs I generated, I hope it does not mess it up here. Otherwise I use edit and keep them out.

	Sales	LogSales	LogSales	LogSales	LogSales
Dum_ICT	4.49873e+09	0.0205	0.106	0.543**	0.630**
	(4.17424e+09)	(0.0846)	(0.0853)	(0.272)	(0.254)
Dum_Year	4.38173e+09	-0.271***	-0.285***	-0.187***	-0.205***
	(2.98351e+09)	(0.0590)	(0.0596)	(0.0452)	(0.0461)
FirmSize	192390.2
	(681561.1)
LogFirmSize		0.924***		1.089***
		(0.0242)		(0.0195)
FirmSizeSqurt			0.119***		0.146***
			(0.00332)		(0.00279)
PowerOutage	2.47067e+09	0.531***	0.543***	0.249***	0.261***
	(3.98525e+09)	(0.0786)	(0.0794)	(0.0605)	(0.0617)
FirmAge	58144378.1
	(99705194.3)
LogFirmAge		-0.0321		0.306***
		(0.0657)		(0.0482)
FirmAgeSqurt			-0.00565		0.0549***
			(0.0101)		(0.00754)
ForeignOwnshp	46519330.5	0.0147***	0.0154***	0.00895***	0.00931***
	(52568144.3)	(0.00105)	(0.00106)	(0.000750)	(0.000765)
Country	-438737249.3**	-0.0186***	-0.0182***
	(197841324.4)	(0.00387)	(0.00392)
Dum_ICTxYear2				0.444***	0.459***
				(0.125)	(0.128)
Dum_ICTxLogFirmSize				-0.125***
				(0.0409)
Dum_ICTxPower				-0.360**	-0.312*
				(0.172)	(0.175)
N	7030	7720	7720	7720	7720
RESET Prob > F	0.0000	0.0034	0.0021	0.0000	0.0028
RESET Prob > F (control group)	0.0000	0.0000	0.0067	0.0000	0.0036

Kind regards
Dominik

[Jeff Wooldridge]

Dominik: My suggestion is to not be dogmatic about particular estimators. Use several that are thought to have good properties. Stata's teffects has most of them.

Also, note that the coefficient on your treatment variable, Dum_ICT, varies from columns (2) and (3) especially. This is likely because you haven't centered the variables that you use in the interaction terms. Equivalently, use the margins option. Or just just teffects with the ra option.