Interaction lsens

[Wesley Mokkink]

Dear all,

I want to minimize total errors in my prediction following from a logit model. Hence, I reckoned that the intersection of the sensitivity and specificity curves may be a good starting point. However, I have not found a way on either STATALIST or -help lsens- to find the cutoff value corresponding with the intersection. Does anyone know how to determine the intersection point between the two curves?

Kind regards,

Wesley

[Clyde Schechter]

Well, you could just plot the sensitivity and specificity against a range of cutoffs and visually estimate where the graphs cross.

But I think you are working from a false premise here. The intersection of the sensitivity and specificity curves does not give you the minimum total errors. The minimum total errors also depends on the proportions of actual success and failures in your outcome data. Let's suppose the probability of a success (1) outcome in your data is p. Then the probability of a prediction error will be p*(1-sens) + (1-p)*(1-spec). That is the quantity you are trying to minimize, and the minimizing cutoff will depend on p.

[Steve Samuels]

I agree with Clyde. For more possibilities, see the posts at http://www.statalist.org/forums/foru...t-optimization