What is the best estimation methodology in a penal data when few of the explanatory variables are also likely to have multicollinearity? Is VAR the right choice in such a model? The model has macro economic variables, like productivity, interest rate, exchange rate, investment rate, etc.
To check multicollinearity in your data you have to do is a correlation matrix between your variables and if you detect any extreme correlations between the variables then you must drop one of them out of your regression. This is an ex-ante method to detect multicollinearity. An ex-post way is to run/estimate your model and then test for VIF. if your VIF value is greater than 10 then u probably have a multicollinearity problem.
Thanks.
But I already know that there is multicollinearity and do not wish to drop the variable, then what to do ? what estimation method to use?
by using more than one high frequency regressor oftens leads to multicollinearity issues and, in the case of the non-linear weighting, increases the complexity of estimation dramatically. An alternative approach suggested
by Andreou, et al. (2013) would be to estimate several univariate models and then
use forecast combination to produce a final forecast.
Can refer to a paper
Energy and Food Commodity Prices Linkage:
An Examination with
Mixed
-
Frequency Data
It will help us to know how you plan to use the information. Our answers might differ if your primary objective is forecasting compared with isolating the impact of a change in interest rates, for example.
Moussumi is right, but I think that it is not necessary to estimate univariate models If you have n explaining variables and variables i and j are correlated, then I would propose to do two regressions with n-1 variables: one excluding variable i and one excluding variable j. Then you form a combined variable k according to the coefficients of the two regressions and estimate again taking k instead of i and j. This procedure gives you an additional check: The coefficients of the other variables should be similaf in all estimates. But even this does not mean that the coefficients of the correlated variables are "correct", but I think the result is better than if one did not care about.
Rob, idea is to isolate the impact of explanatory variables, interest rate, investment rate, etc.
Multicollinearity is something to worry about even before you collect your data, rather than during analysis or after analysis. If certain variables are expected to be highly correlated, you may think of dropping one of those variables. But, if there is sound theoretical reason to have both in the model so that you cannot drop one of them, you should organise your data collection differently e.g., using principal component analysis or cluster analysis for data reduction, which will group similar data. That way you will be able to retain your variable of interest & yet avoid multicollinearity. Also, an easy way out - if the variables exhibiting multicollinearity are not the central focus of your research problem, please do not much bother about multicollinearity & proceed. Hope it helps.
- Badges
- Science topic
- Similar topics
- Books