max Youden index

[Davide Barbieri]

Dear experts
I am performing a ROC analysis.
Is there a STATA function which calculates the cut-off value corresponding to the maximum Youden index (sensitivity + specificity -1)?
Thanks

[Joseph Coveney]

If there isn't a user-written command, then you can use a combination of minbound and estat classification, but it might be easier just to

sysuse auto
logit foreign c.mpg, nolog
lsens , gensens(sensitivity) genspec(specificity)
generate float youden_index = sensitivity + specificity - 1
egen float max = max(youden_index)
list mpg if youden_index == max

I think that some try to interpolate between the maximum and next-highest, but that gives the whole exercise an aura of legitimacy that seems undeserved.

[Steve Samuels]

Phil Clayton's cutpt command (SSC) may improve on Joseph's procedure (one I've used myself) when the classification variable is continuous. See also this thread,

[Fredrik Hjelm]

Hi,

thanks for great advice, how do I get the probability cutoff value where the two lines intersect?

Thank you in advance,

Fredrik

[Joseph Coveney]

how do I get the probability cutoff value where the two lines intersect?

sysuse auto
logit foreign c.mpg, nolog
lsens , genprob(cutoff_p) gensens(sensitivity) genspec(specificity)
generate float abs_delta = abs(sensitivity - specificity)
egen float min = min(abs_delta)
list cutoff_p if abs_delta == min

What am I missing with all this?

[Marcin Kubeczko]

If there isn't a user-written command, then you can use a combination of minbound and estat classification, but it might be easier just to

sysuse auto
logit foreign c.mpg, nolog
lsens , gensens(sensitivity) genspec(specificity)
generate float youden_index = sensitivity + specificity - 1
egen float max = max(youden_index)
list mpg if youden_index == max

I think that some try to interpolate between the maximum and next-highest, but that gives the whole exercise an aura of legitimacy that seems undeserved.

Joseph, thank you so much! I was trying to find how to calculate cut-off in Stata for weeks now. Your solution gave me exact results as in R. Finally, thanks a lot!