Multicollinearity test for ratio scores

[Paul Olz]

Dear all,

I am currently checking my unbalanced panel data for various aspects and have some trouble to put the puzzle together. My n = 311 and T= 12.

I started to check first for multicollinearity. Due to the fact that I am looking at four independent, but related, variables (score 1, score 1.1, score 1.2, score 1.3) and its impact on one dependent variable, I get high values at the vif. Since the three last mentioned scores are parts of the first score, this might cause high multicollinearity, right? All the other values are below the threshold of 5 (or 10).

From a scientific perspective, should I conduct the following analysis all one by one (see below) or should I try to tackle the multicollinearity somehow?

xtreg DV score1 'controls

xtreg DV score1_1 'controls

xtreg DV score1_2 'controls

Any thoughts will be appreciated!!

Paul

[Carlo Lazzaro]

Paul:
if the 4 variables are correlated, high -vif- has to be expected.
Provieded that I do not know whether the following advice make sense in yiour case, I would consider two different approaches to avoid running four different panel data regression:
1) using -pca- to reduce score variables to a single predictor;
2) creating a new predictor by summing up the scores of the 4 scores independent variables.

[Jeff Wooldridge]

Paul: If score1 is comprised of score1._1, score1_2, and score1_3 then you would never control for all of them. Is the test score the key variable of interest? If so, I would try with just score1. Then, try with all three composite scores, score1._1, score1_2, and score1_3.

Why are you reporting vifs without showing us the coefficient estimates and standard errors? The vifs are factored into the standard errors, so if your estimates are sufficiently precise, you don't really care about the vifs. Are you using fixed effects estimation? It seems like you aren't. That's one issue that needs to be resolved.

xtreg DV score1 controls, fe vce(cluster id)
xtreg DV score1_1 score1_2 score1_3 controls, fe vce(cluster id)

If you show the output from these, I might be able to say more.

[Paul Olz]

Dear Carlo, dear Jeff,

Thank you very much for your thoughts.

@Carlo: From the context that I am focussing on a reduction to one single predictor or the creation of a new predictor is contrary to what I would like to look at.
@Jeff: Your first line is the input I needed and is totally logical! Based on the differentiation, I might look at score1 on the first level, and then on score 1_1, score 1_2, and score 1_3 on the second level.

Do you have any proper literature that suggests a similar approach to separation?