Yes. In such situation, you'vyou've to rethink about the choice of IV and DV. Since 'r' is very high, trying to remove multicolinearity may not yield good results and may take us away from the fundamental research objective. Hope I am not wrong. Best.
Dear Dr. Buddhadeb,
Thank you for your valuable comment. My dependent variable is CO2 emission. Do you suggest me to change it? My objective is to test the relationship between CO2 emissions and Trade Openness.
Mousumi is correct, however, the appropriate answer is that multicollinearity is not severe because one can still invert the Variance-covariance matrix, hence the variance inflation factors do not affect estimation.
Dear Udari,
Of course there is multicollinearity in the ARDL approach, but it is not important for your final results. You can skip it easily.
Yours.
My view is, that - hopefully - your dependent variable is correlated with the independent variables. Otherwise your Regression does not make sense. When the independent variables are correlated among them, and you really think you should use all of them (say, for theoretical reasons), why not test joint significance of these variables instead to significance of each of these variables? Or, you solve the "micronumerosity"...
First, run the regressions. Then ru.nl the correlation matrix. I generally use a simple correlation coefficient of 0.80 or greater as evidence of multi-colinearity. In fhis case drop, the weakest variables. Otherwise, you are measuring the same thing more than once and the coefficients of some of the independent variable are biased.
As Alfred Garloff says, you can only speak about multicollinearity when you find a strong correlation among independent variables.
As stated by many commentators above, the issue of multi colinearity arises when some of your RHS variables are linearly correlated among themselves. Check the correlation matrix for these variables only (don't include the dependent var here). And following Jan Reid's suggestions, you may drop one of them. However, if you think all these independent cars are important to you then look for transformation, eg log for real income. Again check the correlation matrix after that. If still not satisfactory, then you have to drop one of them. You may also like to check VIF procedure in Stata.
Lastly, don't get confused between multi colinearity with endogeneity problem.
Best.
One may expect a priori high correlations between LHS and RHS; problem arises when the RHS variables are linear combinations one of the other(s) and, of course; now this is a matter of degree.
Multicollinearity refers to the correlation among the independent variables.
I agree: Multicollinearity refers only to the correlation among the independent variables
Multicollinearity as per other comments only happens among independent variables. However, as noted by Gujarati et al. (2014), a correlation of 0.8 and above signals presence of Multicollinearity. Compare your magnitude of correlation (R) and the VIF results.
Hi Udari,
You have several excellent answers and I concur about multicollinearity being exclusively connected to the independent variables. As you continue with your research a few things might be helpful to consider.
Obviously, researchers would prefer models without issues like multicollinearity. When the real world presents us with challenges and we must make choices, then it helps to understand the consequences.
One consequence of multicollinearity is potential bias in the estimated coefficients. If this is the case, then the estimated impact of changing one of the independent variables is subject to that bias that makes estimating the impact on the dependent variable suspect at best.
On the other hand, if a researcher is after a model that reliably indicates the overall result and is not as concerned with the magnitude of each coefficient, then although multicollinearity is still an issue, its consequences are diminished somewhat. Regardless, it is still an issue and better research typically seeks better specifications of the model when possible. Others have already suggested some good ideas such as eliminating or transforming variables.
I hope this helps.
Rob
@Rob ; I like you but I have to correct a point in your advice. Under multicolliearity, estimates are still unbiased. The problem is that the standard errors tend to increase and the t-values get smaller.
When there are correlated independent variables (presence of multicolinearity). I suggest three alternatives:
1. Eliminate the smallest correlations variables (with the dependent variable).
2. Two-Stage Least Squares, with some of correlated variables can serve as an instrumental variable, maybe?.REG2( CO2 emission
when testing for multicollierity we are concerned with the relationship between independent variables only not the relationship between the dependend variable and the independent variable
I believe Mousumi is correct that multicollinearity is less of a problem in ARDL and we all know that we are talking about the regressors (r.h.s variables) and the standard error problem associated multicollinearity and how it causes the t-values to get small, when it shouldn't. What has not been talked about is multicollinearity and how it affect forecasting! In this case, it should not equally be a problem. In sum, then, if the variables are logged and the variance-covariance matrix can be inverted, collinearity is never a problem.
Yes may be important if there is limited variation in the data so it a matter of the degree of variation.
Sometimes collinearity is a problem of specification. CO2 emision and income? What is measuring "income" in such a relation?
High correlation between independent variables only has problem of multicollinearity.
not with dependent variable
Thank you all,
I have solved my question thankful to you all. Now I'm analyzing my results. Thank you again.
how to handle high correlation between independent and dependent variable in a regression model ?
I agree with: “Multicollinearity is problem only when there is high correlation between independent variables”
- Badges
- Science topic