Is there any way of testing which the omitted variables are?

[Eva Vanat]

Hi there,

I am quite new to both STATA and statistics. I have an assignment in which I need to build a regression model and perform the diagnostic tests. The ovtest showed that I had omitted variables, which I suspected anyway. However, I am really not sure which variables that could be and the dataset is big(ish). Is there any way of getting STATA to look for which the omitted variables are? Or is the only solution that I look for them myself? Thank you!

[Richard Williams]

One obvious possibility is to think about whether or not some squared terms are warranted. Other than that I would say you are on your own. Use theory to guide you. There are things like stepwise selection procedures but they are generally looked down upon.

[daniel klein]

The ovtest showed that I had omitted variables, [...] I am really not sure which variables that could be and the dataset is big(ish).

Just so there is no misunderstanding, ovtest does not know anything about the variables in the dataset that were not included in the model. The test is based solely on powers of fitted values from the model (or optional the powers of the predictors in the model). Thus, this test cannot tell you anything about which additional variables in your dataset to include.

Personally, I find the name omitted-variable test very misleading and would prefer calling this a test of misspecification.

Best
Daniel

[Carlo Lazzaro]

Eva:
I share both Richard and Daniel's stances, in that the result of -ovtest- rarely gives you some substatntive clue.
The best approach is to rely on the theory of your research field and skim the existing literature in search of what other researchers did in the past in dealing with your same research topic?
As an aside: did you perform -hettest- as well?

[Eva Vanat]

Thanks everyone! Yes, I performed the -hettest- as well and it showed heteroscedasticity. After that I added the -robust- command. I guess the only solution now is to manually look for more variables that I should include based on literature? The assignment is from a subject area I am not very familiar with, but it shouldn't be a big problem to find more variables to include. I was just wondering whether there might have been a quicker solution by asking STATA to look for them for me, but life is not so easy, I guess! I was told stepwise was not really what was looked for in the assignment, so I don't think I will use that. Thank you very much for your help! Very useful

[Carlo Lazzaro]

Eva:
Stata (not STATA, please) cannot figure out what's other reserchers did in the past as far as the "right" set of predictorsi is concerned (as Daniel explain).
In my humble opinion, your goal should be to look for the most common set of predictors reported in other researches on the same topic.
As Richard highlighted, you might be missing some qudratic term among your predictors and this would cause -ovtest- result to be statistical significant:
Hence, I would consider if there is any turning point in your predictor. If a missing quadtratic term is actually the culprit, I would recommend you to plug it in with the linear one as well via - fvvarlist-.
By the way, -help fvvarlist- provides the link to some interesting video tutorials.
As a closing-out remark, stepwise is not an advisable way to go, because you usually ends up with a model that has lot to do with your dataset but might be miles away from the real data generating process.

[Jeff Wooldridge]

I wish Stata would stop presenting RESET as a test for omitted variables. Contrary to what was claimed by Ramsey and others, the test never was that and never will be that. It is a functional form test. Can it have some power for omitted variables? Sure, it's incidental. If you have no information on the omitted variables, or proxies for them, you can't test for them. I made this case in an obscure chapter published in 2001: Wooldridge, J.M. (2001), “Diagnostic Testing,” in Companion to Theoretical Econometrics. B.H. Baltagi (ed.), 180-200. Oxford: Blackwell.

The argument is simple once one realizes the fixed-in-repeated samples paradigm for the regressors is a non-starter when evaluating RESET or any other specification test.

Eva: the ovtest does not and cannot tell you there are omitted variables. If it was significant you can re-specify your model with quadratics and interactions, as Richard said.

[Carlo Lazzaro]

I do agree with Daniel and Jeff.
It seems that Ramsey made a long-lasting advertising stunt in favour of -ovtest- by labelling it as an omitted variable test.
When my side-interest in econometrics started to grow and I knew almost nothing about that stuff, the idea we could rely on a clever test that works for us in pointing out what is missing among our predictors was very appealing. Later I discovered, at my own expenses, that it was only maths (no white magic, unfortunately...)

[Richard Williams]

A lot of tests don't necessarily show you what they purport to show you. For example, a significant value on a hetero test may be due to the fact that you omitted an interaction term or a squared term or something else. When I teach hetero, I tell my students it could be a mistake to immediately jump to robust standard errors or weighted least squares or whatever. Consider whether some other model assumption is being violated (e.g. vars are omitted) and if so fix that first.

[Carlo Lazzaro]

I agree with Richard:
when constrasted against omitted variabe bias, heteroskedasticity bias looks like a misdemeanor.
Hence, the first problem should be dealt with first before invoking -vce(robust)-

[Rich Goldstein]

as the original author of both ovtest and hettest, I think they do what they say but that many people want more (reasonably enough); ovtest, under certain assumptions, can be interpreted as meaning that a failure is do to a mis-specified model in the sense of at least one omitted variable; that is a reasonable thing to want and have; it does not purport, as several people above have already said, to tell you what is omitted -- nor could it do so since that is both context dependent (i.e., theory) and the omitted variable may not be one in the current data set; these tests are very limited - but they do provide something of value even so (in my opinion)

[Carlo Lazzaro]

Rich is right.
The devil is probably in too much expectation from -ovtest. when naively approaching the regression machinery.