0
$\begingroup$

I wanted to find the ability of a numeric continuous variable to predict mortality (dead/alive), and what cut off value to take in the continuous variable.

I wanted to construct an ROC curve and find the cut off value to achieve this.

I did a logistic regression, taking mortality (m) to be the dependent variable, and the continuous variable (v) to be the predictor.

The p value of the coefficient I got for v was 0.19. The study take p value of <0.05 as significant.

Given that the coefficient is not significant, can this model be used for prediction, and ROC curve constructed using this model?

$\endgroup$
0
$\begingroup$

It can be used as a predictor if you feel that the coefficient you obtained is clinically important, you don't mind a p-value of 0.19, and the sensitivity or specificity is particularly favorable for your diagnostic or prognostic problem.

However, a coefficient with a p-value of 0.19 is unlikely to correctly classify dead/alive status.

$\endgroup$
0
$\begingroup$

The question of how well something predicts something else is about effect size, not statistical significance. If N is small, then a fairly good predictor might be non-significant. If N is very large, then a very poor predictor might be significant.

However, with small N, our estimates of parameters are usually not very precise, so you will have a lot of doubt about any predictions you make.

$\endgroup$
2
  • $\begingroup$ Thank you, the size in question here is 63 data points. Would you suggest checking AIC for whether to use this model or not? $\endgroup$ – Sudhanva Rao Nov 10 '19 at 12:48
  • $\begingroup$ AIC is only useful for comparing one model to another. I suggest using a measure of effect size. $\endgroup$ – Peter Flom Nov 10 '19 at 16:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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