2
$\begingroup$

I have a simple logistic model, and I internally validated it using rms::validate in R. The estimated overoptimism for the intercept is -0.015 and for the slope is 0.05. How do I calculate the corrected predicted values? I know I have to multiply the coefficients by 0.95, but what should I do with the intercept? I guess I have to subtract -0.015, but I am not really sure.

$\endgroup$

1 Answer 1

2
$\begingroup$

First of all, note that recalibration of predictions is better than not recalibrating them, but you don't get a way to validate the recalibrated model. In your particular case where the calibration intercept and slope are -0.015 and 0.95 most statisticians would conclude that the model is OK as-is.

But all that depends on the calibration being linear. Show us the smooth estimated entire calibration curve and some indexes of miscalibration. One of my favorite indexes is the 0.9 quantile of absolute prediction error.

Iif you proceed with a linear recalibration then you would multiply $X\hat{\beta}$ by 0.95 and subtract 0.015.

See Ewout Steyerberg's book Clinical Prediction Models for much more on this subject.

$\endgroup$
2
  • $\begingroup$ Thank you, I was actually studying on Steyerberg's book and trying to grasp the concepts. Unfortunately the practical example provided is based on a Cox model so I was unsure about the intercept. $\endgroup$
    – Claudio
    Feb 10, 2022 at 15:05
  • $\begingroup$ Good point. It's easier with logistic models than with the Cox model which has to involve the underlying survival curve. $\endgroup$ Feb 10, 2022 at 16:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.