REST test OVTEST

[Dawud Sikander]

So I am running a Ramesy test on stata to show OLS is the best conditionally linear unbiased estimator and to to examine the impact of key omitted variables on our regression. So I am new to stata and I have done the tests (attached) but I am unsure how to interpret the OVTEST, I checked online but could not find any material to help me.
Any help will be appreciated

[Dung Le]

Hi Dawud,

My understanding is that in the Ramsey's RESET test, your null hypothesis is the model you are using is correctly specified. That is your pure hypothesis, then you want to test it using -ovtest- command. In the 1st picture, the results of -ovtest- command show that F statistic is 1.26 (3 and 1,035 degrees of freedom), indicating statistically insignificant. That said, this result supports your null hypothesis: the model you are using is correctly specified.

By contrast, in the 2nd picture, the model you are using is not correctly specified based on the results of -ovtest- command reported.

Hope that helps.

[Dawud Sikander]

Hi Dung Le,

That you for your response it is very appreciated. For the 2nd picture how would I fix that or is there an alternative test I can use instead of the ovtest.

[Dawud Sikander]

Hi Dawud,

My understanding is that in the Ramsey's RESET test, your null hypothesis is the model you are using is correctly specified. That is your pure hypothesis, then you want to test it using -ovtest- command. In the 1st picture, the results of -ovtest- command show that F statistic is 1.26 (3 and 1,035 degrees of freedom), indicating statistically insignificant. That said, this result supports your null hypothesis: the model you are using is correctly specified.

By contrast, in the 2nd picture, the model you are using is not correctly specified based on the results of -ovtest- command reported.

Hope that helps.

Hi Dung Le,

That you for your response it is very appreciated. For the 2nd picture how would I fix that or is there an alternative test I can use instead of the ovtest.

[Carlo Lazzaro]

Dawud:
a bit of confusion here due to posting screenshots instead of reporting what you typed and what Stata gave you back via CODE delimiters.
The first picture Dun Le (whose -estat ovtest- interpretations I share) refers to is the first from the bottom: with a Prob>F=0.2874 the null is not rejected: hence, there's no evidence of model misspecification.
The second picture Dun Le refers to is actually the first from above: with a with a Prob>F<0.0001 the null is baldly rejected: hence, there's evidence of model misspecification.
No test can tell you the opposite.
In my opinion, here you have too many observations and a handful of predictors (or you missed some interactions between the existing predictors and/or experience may have a non-linear relationship with the regressand).

[Dung Le]

Hi Dung Le,

That you for your response it is very appreciated. For the 2nd picture how would I fix that or is there an alternative test I can use instead of the ovtest.

Ramsey's RESET test is used to examine whether the independent variables are correctly specified, conditional on the specification of the dependent variable, so you may want to try to use interaction or quadratic terms of your independent variables. However, you may also need a theoretical framework behind interaction or quadratic terms or you may have strong arguments on why you should include interaction or quadratic terms in your models.

-ovtest- is a good option, no need other commands. But, if you want, you can compute Ramsey's RESET test by hand.

[Phil Bromiley]

In structural equation modeling, there is a well-known phenomena that one often can reject a model with reasonable fit if one has a sufficiently large sample size.

I wonder if this also applies to specification tests. If you look at the two posted models, their r-squares are precisely the same. It looks as if the second one is simply some kind of sampling from the larger database. But, with almost 30 times as many observations, the test may be able to reject much smaller deviations in the larger sample.

I'm not sure if this is correct – I'm just throwing out the question.

[Dawud Sikander]

Dawud:
a bit of confusion here due to posting screenshots instead of reporting what you typed and what Stata gave you back via CODE delimiters.
The first picture Dun Le (whose -estat ovtest- interpretations I share) refers to is the first from the bottom: with a Prob>F=0.2874 the null is not rejected: hence, there's no evidence of model misspecification.
The second picture Dun Le refers to is actually the first from above: with a with a Prob>F<0.0001 the null is baldly rejected: hence, there's evidence of model misspecification.
No test can tell you the opposite.
In my opinion, here you have too many observations and a handful of predictors (or you missed some interactions between the existing predictors and/or experience may have a non-linear relationship with the regressand).

Hello,
Thank you very much for your reply.Yes the one with a prob>=F=0.2874 is measuring the return to education to ethnic minorities in the UK with gross weekly pay as the dependent variables. The second one Prob>F<0.0001 with the more observations is measuring the returns to education of white ethnicity in the UK also with gross weekly pay as the dependant variable. All observations come from the same dataset. I will look through the data set and see if I can find anymore predictors.
I am attempting to ensure the regression results are complete and robust, and test for misspecification.

[Dawud Sikander]

In structural equation modeling, there is a well-known phenomena that one often can reject a model with reasonable fit if one has a sufficiently large sample size.

I wonder if this also applies to specification tests. If you look at the two posted models, their r-squares are precisely the same. It looks as if the second one is simply some kind of sampling from the larger database. But, with almost 30 times as many observations, the test may be able to reject much smaller deviations in the larger sample.

I'm not sure if this is correct – I'm just throwing out the question.

Hello,
Thank you for your reply. Both come from the same dataset. However, the smaller one measures the return to education for ethnic minorities in the UK and the larger one measures it for white ethnicity in the UK which has a many more observations. I am trying to run tests to show that the data is robust and complete with no misspecifitons. What would you suggest for me to do?

[Dawud Sikander]

Ramsey's RESET test is used to examine whether the independent variables are correctly specified, conditional on the specification of the dependent variable, so you may want to try to use interaction or quadratic terms of your independent variables. However, you may also need a theoretical framework behind interaction or quadratic terms or you may have strong arguments on why you should include interaction or quadratic terms in your models.

-ovtest- is a good option, no need other commands. But, if you want, you can compute Ramsey's RESET test by hand.

Hello,
If I were to make some of my independent terms quadratic would I need to do this on both of my models. The smaller one measures return to education for ethnic minorities in the UK whereas the larger one measures returns for education for white ethnicity in the UK.Both come from the same data-set. Also, what do you mean by using interaction?

[Dung Le]

Hello,
If I were to make some of my independent terms quadratic would I need to do this on both of my models. The smaller one measures return to education for ethnic minorities in the UK whereas the larger one measures returns for education for white ethnicity in the UK.Both come from the same data-set. Also, what do you mean by using interaction?

Hi Dawud,

or you missed some interactions between the existing predictors and/or experience may have a non-linear relationship with the regressand, quoted by Carlo in #5

It means that there could be cases that two (or more, I am not sure) variables may have simultaneous impacts on your outcome variable. For example, let's consider a simple model, where you have a wage variable - the outcome and two predictors: age and gender. You may think that age may have an impact on wage because the older the age, the higher the wage (due to accumulative experiences) at certain point (however it could also be cases that individual's wage will start decreasing when it reaches its peak - then you may need to add a age-squared variable in your model, but we do not consider that case here for simplicity). Also, male, on average, have higher wage than female. You now can see that both variables have impacts on wage. Thus, interaction means that you may want to create an interaction variable (c.age##i.gender - I assume that age is a continuous variable) between the two and add that interacted variable into your model.

Hope that helps.

[Dawud Sikander]

Hi Dawud,


It means that there could be cases that two (or more, I am not sure) variables may have simultaneous impacts on your outcome variable. For example, let's consider a simple model, where you have a wage variable - the outcome and two predictors: age and gender. You may think that age may have an impact on wage because the older the age, the higher the wage (due to accumulative experiences) at certain point (however it could also be cases that individual's wage will start decreasing when it reaches its peak - then you may need to add a age-squared variable in your model, but we do not consider that case here for simplicity). Also, male, on average, have higher wage than female. You now can see that both variables have impacts on wage. Thus, interaction means that you may want to create an interaction variable (c.age##i.gender - I assume that age is a continuous variable) between the two and add that interacted variable into your model.

Hope that helps.

Thank you very much. Your explanation has given me a clear understanding. I will create some interaction variables and see what the result is. I did add another variable in it did change the results a bit (attached) but I assume the results still say my null hypothesis is rejected,

[Dung Le]

Hi Dawud,

You are right about the results presented in #12.