Roctab and Yuden Index

[Clarissa Gallegos]

I'm looking for guidance on selecting a cutoff point for a variable estimated using probit. My variable of interest is OSIG (binary), and from the probit model, I've generated OSIG_HAT. Now, I need to create a dichotomous variable based on OSIG_HAT. I was advised to use a ROC table (roctab) and calculate the Youden index to select the cutoff point. First and foremost, I'd like to confirm if this method is suitable for my needs. Secondly, I'm seeking guidance on how to implement this in Stata.

Thank you all very much.

[Rich Goldstein]

because the Youden index is (sensitivity + specifity - 1) it only makes sense if the cost of each type of error (false positive, false negative) are equally bad; are they in your case?

I believe that selecting a cutoff (there is a lot of statistical literature on this issue but I am not a fan) should be based on substantive knowledge, including the prevalence of positives

there is at least one user-written routine that might help if you want this direaction: -cutpt-; use -search- to find and install

[Leonardo Guizzetti]

I agree with Rich, this sort of cut point selection is frequently done in medical literature and rarely are the costs (or harms) of a false-positive equal to that of a false-negative. It also forces one to think in binary after having thrown away useful (continuous or ordinal) information.