Is testing for molel specification important?

Dear Saadia Irfan,

Most modelers are typically careful in specifying their model assumptions but they rarely report results from sensitivity analyses. I may be wrong or right, but if by “testing the model specification formally” it means performing a sensitivity analysis, yes, you require to conduct an analysis to ascertain the robustness of your model.

Good luck.

by model specification I meant 'testing that the model does not have ommitted variables and also does not have extra variables' Usually this is tested by 'linktest' and ovtest' in stata...

Thanks for the clarification. Still, I think you need to test. There would be no harm in doing so. In fact, you will then know that you have the right number of variables, right?

Yes the model specification command is not difficult to conduct. However I have used a lot of sample data from internet and conducted this test and the test says the model is not correctly specified. I was wondering if there is any model which satisfies this test. One of my PhD fellows told me that if your model is coming from theory it does not matter if your model is not correctly specified (through this test). So I posted this question to check if this thing is true...

Can the PhD fellow collaborate with similar views from the literature what he/she told you. If yes, then you are ready to go. Otherwise, you need to research more.

I will also research on the same & get back to you. Good evening.

Well nobody states in the literature review explicitly thathe they have not performed this test. But I don't see this test performed anywhere....so I think the researchers don't care about model specification as long as their variables are coming from theory...but I'm not sure if I'm right...

Take it easy & also rest.

You should be sure, otherwise, you are giving people room to laugh at you; although sometimes it does not matter if they laugh or not.

Meanwhile, read:

Jan F. Kiviet & Garry D. A. Phillips (1986) Testing Strategies for Model Specification

Goodnight

Also:

J. A. HAUSMAN (1978) Specification tests in econometrics, Journal of the Econometric Society.

He says "SPECIFICATION TESTS FORM ONE of the most important areas for research in econometrics...."

Sadia,

There is a bulk of literature about this. A good starting point surely is: http://people.stern.nyu.edu/tcooley/papers/ath_mac_a_cri.pdf. Then, you can use it to search up (using google scholar citations) and down (using this paper bibliography) to improve your knowledge about this controversy.

Personally, I think it is necessary to combine both, theoretical and empirical insights. Candidate variables (potential covariates) should be selected from alternative theories (in your case, the underlying theory). But final model specification should not be selected without some data-validation.

For this reason we develop an automatic model selection technique (in Stata) called: GSREG (http://www.stata-journal.com/sjpdf.html?articlenum=st0383).

You can download it here: https://ideas.repec.org/c/boc/bocode/s457737.html or just typing ssc install gsreg in STATA.

Hope it helps!!

I am quite old. In the ancient time, we used the term "structural model" would that be of some help ? 

I agree you Demian T P

Dear Saadia,

I don't fully understand what do you mean by testing the model specification formally.

i meant 'checking if the model is fully specified-meaning it has no ommitted variables and no extra variables. normally this is tested using ovtest and link test in stata

Tests for heteroskedasticity, autocorrelation, omitted variables parameter constancy etc.  tests should always be used in econometrics. They can be regarded as both tests of a specific problem but also as general tests of the validity of the model. There are very few models based on economic theory that can be accepted without testing. If the data do not fit the model either your model or your data are wrong 

Well said John C.F.. Now Saadia knows that she has to do the tests and I hope she has gotten what she was looking for.