I'm running a predictive model using the logistic model in SAS and, currently, I'm trying to perform some diagnostics about the collinearity issue in the estimated model.
To do that, I followed step-by-step what the SAS support suggested (look at here).
Shortly, I run the following code to compute the diagonal weight matrix:
ODS SELECT CORRB PARAMETERESTIMATES; PROC GENMOD DATA = BETTING.TRAINING_DUMMIES4 DESCENDING; /* to analyze the event Z1 = 1 */ CLASS Z1 VAR; MODEL Z1 = VAR /LINK = LOGIT /* to use the logit function */ DIST = BINOMIAL /* to use the Binomial distribution */ SCORING = 50 /* ensure using the Fisher's Scoring */ CORRB; /* correlation matrix estimation */ OUTPUT OUT = EST_PARAM HESSWGT = W; /* compute the diagonal weight matrix */ RUN;
in order to use that in a linear regression model with weights $w$ and to be be able detecting the presence of collinearity in the explanatory variables, as follows:
ODS SELECT COLLINDIAG COLLINDIAGNOINT; PROC REG DATA = EST_PARAM; WEIGHT W; /* weights computed using the HESSWGT= option in the model above */ MODEL Z1 = VAR /COLLIN COLLINOINT; RUN; QUIT;
The output is around 5 in the last condition index value for the variable.
According to what SAS suggests in the Support Blog, where one detects collinearity between the intercept and an explanatory variable because of an index condition equal to 315, the model should be not affected by collinearity.
Can you confirm that?
Does it exist some cutoff points in such kind of analysis?