coldiag2 interpretation of output

[ulas alk]

Hi,

I am running coldiag2 command to detect multicollinearity in my data. I understand that each condition index is the ratio of the square root of the largest eigenvalue divided by the corresponding eigenvalue in the row. However, I am having difficulty understanding the variance decomposition entries. You can find a sample code and output below:

sysuse auto.dta
coldiag2 weight length displacement trunk, noconstant

Condition	number using scaled variables = 30.47
Condition	Indexes and Variance-Decomposition Proportions
condition
index	weight length displacement trunk
1 1.00	0.00 0.00 0.00 0.00
2 7.58	0.00 0.03 0.34 0.11
3 11.44	0.02 0.12 0.02 0.89
4 30.47	0.98 0.85 0.64 0.00

Here, each column sums up to 1. Since we are trying to understand how much a variable contributes to the variance of an estimated coefficient (betas), shouldn't we have rows sum up to 1 instead of columns? Can anyone explain what these numbers represent? I looked at the help file but couldn't find a detailed discussion. Thank you.

Ulas

[ulas alk]

Anyone have any idea?

[Rich Goldstein]

I'm not sure I'm reading your output correctly (please read the FAQ on how to post so your post is legible), but (1) the rows should not add to 1 - they show the dependencies; (2) you only need to look at rows where the condition number is "high" - different authors have different criteria for "high" but a very rough rule-of-thumb would say 10 is marginal and above 30 is high (though you should probably look at Belsley, DA, Kuh, E and Welsch, RE (1980), Regression Diagnostics, Wiley (or Belsley, DA (1991), Conditioning Diagnostics, Wiley) for more on this issue); within such a row, if there is an issue there will be at least 2 values that are high (e.g., over .5); if there are not two, there is no problem; if there are at least 2, then those variables are involved in the issue; note that (1) collinearity is not always a problem and even when it is harmful, it may be minor and (2) the problem, if any, can be solved by acquiring additional data (if possible)