In addition, multicollinearity was not present as evidenced by the Tolerance and VIF values for the predictor variables.
In addition, examination of the Tolerance and VIF values for the predictor variables revealed that multicollinearity was not present.
The VIM from RF we used is the mean decrease in accuracy generated by this predictor whenever it was selected for splitting, compared to when the values for the predictor was permuted.
For instance, the values for the predictor unit sum ranged from 3 to 17 (e.g., for 28 + 49, the sum of the unit digits would be larger: 8 + 9 = 17, while for 31 + 52, the sum of the unit digits would be smaller: 1 + 2 = 3).
Regression analyses were performed using the brglm package (Kosmidis 2013), which addresses issues of near perfect separation in logistic regressions (i.e., when there is perfect correspondence of the response variable for most values of the predictors, but not for all) (Heinze and Schemper 2002).
For the same predictor variables as in Strategy 1, but with binary Y = 1 if TC/HDL-C ≥ 5.0, Y = 0 otherwise, additive logistic models log[ p/(1 - p)] = b0 + b1 X1 + b2 X2 + ··· + b k X k + e, where p = probability that Y = 1 for given values of the predictors and e = binomial error term, were fitted.
For logistic regression, this corresponds to the proportional odds assumption which states that the cumulative odds ratio for any two values of the predictors is constant across response categories.
23 Given a time after diagnosis for values of the predictor variables, the model produces a survival function for the probability that the binary event of interest (eg, death or survival at the endpoint) occurs.
Multicollinear relationships were observed between the variable group of creatinine, cystatin C and MMAS, and when controlling for height, the R values for these predictors were similar (Table S2).
