Ramsey reset test

[Bernt Evensen]

Hello again.

I am studying the Growth rate of Health expenditures for all the OECD- countries in the period 1970- 2015.

I am building a Reference model. Where I include 3 explanatory variables: BNP, o65 and BD.
I run the regression 3 times. First I use pooled OLS. Then I include country- fixed effects. And in the lat regression I include time and country- specific effects.

I attach the regression results.


These are the corresponding Ramsey- reset test values.

Ramsey reset test, H0:No omitted variable bias

F(3, 1039) = 24.45

F(3, 1005) = 27.11

F(3, 961) = 21.84

As you can see I also ran a Ramsey reset test for each regression. The corresponding F- values are high which indicates wrong functional form. To get this F- value Down I had to among other Things include BNP^5 in the regression. From a theoretical perspective it makes so little sence, and it doesent really add very much to my model except lowering the value of BNP.
I also re- estimated this model using the average Growth rate for 5- year periods. (I did this by first typing: gen period5 = 5 * floor(year/5), and then: collapse hce BNP W o65 BD, by(land period5)). When I did this the R^2 went up and the F- values for the Ramsey reset test went Down.

My question is what I should do (if anything) about this hight Ramsey- value in the Reference model of my paper? Should I start to include squared terms to lower it, or should I just ignore it?

Also, which other test do you miss in this model? What is absolutely necessaryt to test for in a model like this?

Answers are very much appriciated!

[Carlo Lazzaro]

Bernt:
Ramsey test is telling you that the specification of your model can be improved. The core issue, however, is why -reset- is telling you so.
You may have an improvable specification form: usually, adding squared and, less frequently, cubic terms, you can fix the problem. (I would not reccomend to add higher order terms, as this would make your result hard to explain).
Sometimes you may want to consider to add interactions (see -help fvvarlist- on that topic).
In other instances, logging part of the variables can be wise.
I would also test your model for heteroskedasticity via -estat hettest-; -estat vif- is for many authoritative contributors on this list not that informative (and the hint is to look at the 95%CI of your predictors and see whether they make sense).
As an aside, please read the FAQ about (not) attaching files, but provide an excerpt/example of your dataset via -search dataex-.

[Bernt Evensen]

Thank you very much, that helps me a lot.

Another question, my dependant variable and most important variable is not normally Distributed. Is this something I should pursue further? Or should just ignore it? I have read some other papers who use similar data as I, and they don`t mention it.

[Carlo Lazzaro]

Bernt:
in OLS normality may play some role in residual distribution only.

[Joao Santos Silva]

Dear Bernt,

Just to add to Carlo's excellent advice, I believe that the RESET test is not being computed using a robust covariance matrix, as it probably should be; I suggest you do it "manually" using only squares or squares and cubes.

Also, with you sample size, not even the errors have to be normal so you can ignore that issue.

Best wishes,

Joao

[Bernt Evensen]

@Joao. Thank you! But do you know a paper or a book were they state that non- normal residuals is not a problem with large N?

[Joao Santos Silva]

Dear Bernt,

Any econometrics textbook should say that; the normality assumption is used for pedagogical reasons, but soon dropped. See, for example, Jeff Wooldridge's books.

Best wishes,

Joao

[Bernt Evensen]

When Ramsey- Reset test tells me that I should include one of my explanatory variables to the 5th power as a regressor, could this be related to to outliers and extreme observations?

[Joao Santos Silva]

Dear Bernt,

The test never tells you that, and I do not recommend performing the test with such powers.

Best wishes,

Joao

[Bernt Evensen]

I mean that the H(0) (No misspecification) in Ramsey reset can no longer be rejected when I include BNP^5 in my regression. So my question is if this could be related to extreme values and outliers?

[Carlo Lazzaro]

Bernt:
the main issue is that, assuming that -BNP^5- causes -estat ovtest- not to reject the null, how will you do disseminate that coefficent?
Usually, if squared terms do not do the trick, you should look for a different/better specification of your regression model (and the literature in your research field can be a precious source in that respect).
As an aside, there's a sort of (undeserved) magic sorrounding Ramsey's test and it probably starts from its name (omitted variables bias, that sounds like that test is able to tell you which predictors you omitted in your right-hand side of the equation. I must confess that sometimes I wished it could be so helpful...).

[karthick veerapandian]

Hello Statalist!

Can you help me to estimate Ramsey REST post probit regression?. What command should I use and how to proceed after nonlinear regression?

[Dung Le]

Hi Karathink,

I am not sure where Stata has a built-in command to perform Ramsey's RESET test. However, you can do such test manually. The logic of that test is to regress the dependent variable on its predicted value and powers of its predicted value, so it quite straight forward:

*** Run probit model, get predicted value and generate polynomial terms of the predicted value
probit y x1 x2 x3
predict yhat1 if e(sample), xb
generate yhat2 = yhat1^2
generate yhat3 = yhat1^3
generate yhat4 = yhat1^4

*** Ramsey's RESET test
probit y yhat1 yhat2 yhat3 yhat4
test yhat2 yhat3 yhat4

As an aside, you also can use Pregibon's link test for model specification check and Stata provides a postestimation to do so

probit y x1 x2 x3
linktest

Hope this helps

Dung Le

[karthick veerapandian]

Thank You Mr.Dung Lee. It is helped me a lot.

[Javier Rua Montes]

Hello Mr Joao, I am trying to generate a gravity model for soybean trade, and I have raised the ramsey test and it does not pass it or presents a result p> chi = 0.003, which makes me reject the null hypothesis. My question is that for this test I use a fit2 but if I apply a fit ^ 2 and a fit ^ 3, my model passes the test, however I would like to know the difference of applying only fit ^ 2 and applying fit^2 and fit ^ 3.

Thanks
Javier