3

Question

So first this question asks "how I would discriminate between model(1) and model(2)".

It appears that both models are non-nested so I would come up with a hybrid model consisting of Xt,Zt and Qt regressors. I know I need to test for Zt=0 and Qt=0 but I'm not sure if I need to test for Xt=0. My logic tells me no as the point is to compare model(1) and model(2), since they both contain Xt as regressor this won't tell me if model(1) is better than model(2) and vice-versa. But someone please correct me if I'm wrong.

For the second part where: Given the available information, would it be desirable for you to come up with a third and dominant model?

First a third model would be desirable only if model(1) and model(2) are not "good" or we would want a better model. But from the way the question is formulated it cannot be the second one, so I believe I should test each individual models to see if they are "good".

For this part the only thing I can think about is doing the following for each equation: t-test(Individual Significance Test) F-test(Joint Significance Test)

Apart from these two tests there aren't really any other information that can be used to tests the models.

So if model(1) and model(2) fail any of these tests then a third and dominant model would be desirable. Is my understanding correct?

ankc
  • 799
  • 2
  • 8
  • 21

1 Answers1

4

I would look also at the efficiency of the estimators, i.e. at their variance-covariance matrix. Efficiency is important since it measures the uncertainty characterizing the point estimates.

The two models self-select with respect to a single regressor. So it is a duel between the two regressors. Put them in the arena together, and see who remains alive. That's the "hybrid" model you mention - used here as a "testing device" rather than as a nesting model. If one of the two regressors becomes statistically insignificant in the presence of the other, I would go with the model that contains the second -it is more strongly engaged with the dependent variable.

Since this enhanced model nests the initial two, if you are prepared to make some distributional assumption on the error terms, you can conduct straightforward Likelihood Ratio Tests, pairwise.

Also, there are formal procedures to compare "overlapping" models like yours (strictly speaking, they are midway between "nested" and "non-nested" models, since they have an explanatory variable in common), but require distributional assumptions for the distribution of the error terms in the two models (conditional on the regressors). You can consult Vuong 1989. Also, this paper by Lewis et al (2011) seems to be more focused on practical implementation of Vuong's tests.

Model selection is an open-ended task - I would not be surprised if some other answer proposes a completely different approach.

Alecos Papadopoulos
  • 52,923
  • 5
  • 131
  • 241
  • so I only need to test for Zt=0 and Qt=0 right?There is no need to test for Xt=0? "The two models self-discriminate with respect to a single regressor" What does this mean? Can you check if I'm right for the second part? – ankc May 18 '14 at 07:09
  • In the context of the enhanced specification containing all three variables, you should also verify that $X_t remains statistically significant. About the sentence: it should be "self-select", I changed it. About the "second part": Ιf in the enhanced model, all variables remain statistically significant, and things like estimator variance, or adjusted R-square do not get worse, I would go with it rather than the reduced models. All three variables can be reasoned to affect FTSE returns, and going form two to three regressors does not really create issues of parsimony. – Alecos Papadopoulos May 18 '14 at 10:03
  • But given the available information I won't be able to determine if the enchanced model is better. The question does state that I must reach a conclusion with the available information. @Alecos – ankc May 18 '14 at 10:27
  • Indeed, but what is the "available information"? The estimated coefficients, the SE, and the R-squared only? – Alecos Papadopoulos May 18 '14 at 10:45
  • Yes, available information is everything given above. There is nothing else. The only thing possible are t-tests and F-tests. Is that right? @Alecos – ankc May 18 '14 at 17:38
  • To perform an F-test, you need the _unadjusted_ $R^2$ (the coefficient of determination) to express the F-statistic. To obtain the coefficient of determination through that adjusted $R^2$, which is what you are given, you need to know the size of the sample. Is the sample size part of the "available information"? If yes, you can also perform F-test, and not only t-tests. If for both models the test rejects the null of zero coefficient, you should find the p-values of the statistic in order to be able to say something discriminating – Alecos Papadopoulos May 18 '14 at 18:00
  • Well the questions says it's monthly returns and from Feb 1995 to Nov 2012, so there are 214 observations. If I carry out an F-test, is it worth carrying out t-tests? – ankc May 18 '14 at 18:42
  • Yes it is, for a more complete picture. – Alecos Papadopoulos May 18 '14 at 19:15