Pca

[Jihane bou zakhem]

Hello everybody,

I have two variables with high vif and with high correlation. I need them both in my panel data model , and I cannot drop one of them.

Could I use the pca analysis like follow:

pca depvar indvar1 indvar2
screeplot
predict pc1, score

Than, replace both the independent variables by pc1 in the model? And if so, I still consider it as panel data and run all the tests (heteroskedasticity , endogeneity, autocorrelation, Hausman,...)

Regards,

Jihane

[Nick Cox]

You can do that, but it's really not a good idea. You're mushing in the outcome or response (you say "dependent variable") in what you're going to use in the model as a predictor. The outcome is out of place in the PCA; Although I can work up no enthusiasm for it, the principal components approach would be to put two predictors in a PCA and then use the first PC.

If you think you need both of them, then try that. What do you mean by a high correlation? If it is really almost 1, then you don't need both of them: just select one on substantive grounds.

Why does the high correlation arise any way? We need to see a scatter plot to advise and also to know more about the meanings of the two predictors concerned.