Unitless Index of Partial Predictive Ability

I am trying to quantify the proportion of total predictive information contributed by each predictor in a model with a binary outcome fit in a large sample. I’m reading page 206 in Regression Modelling Strategies covering this topic.

The helpful plot(anova(my.model)) in the rms package provides Wald chi2 - df statistics for each predictor to approximate the penalized LR for each predictor.

Is it at least approximately valid to divide the Wald Chi2-df for the predictor of interest by the sum of Wald Chi2 - df for all the predictors to get this “partial” index?

I have bootstrapped the process to provide medians and 2.5%-97.5% percentiles of the estimates.

Not the sum of Wald \chi^2 but rather the total \chi^2 for the global null hypothesis. Better would be to go to extra trouble and do this on the basis of likelihood ratio \chi^2 statistics. But see other quicker methods at https://fharrell.com/post/addvalue.

That’s a nice post I had forgotten about - thank you!