Creating 95% CI around Sensitivity/Specificity values

[SAS_ter]

Hi,

Does anyone know how I can use an existing dataset to output 95% confidence intervals around sensitivity/specificity values? I have created a cross-tab so that I can derive my sens/spec values by hand. However, I'd like SAS to run the 95% CI for me without me having to input the data values each time.

Secondly, what test can I do in SAS to compare two sensitivity values to determine if they are significantly different from each other? i.e. Algorithm vs,Test A gives me a sens. value of 0.87 and Algorithm vs.Test B gives me a sens. value of 0.89, this is done on the same population. What test/code do I use in SAS to determine this?

Thanks in advance.

[HWSteinberg]

Hi SAS'ter,

What do you need? Sensitivity and Specificity are polygons running from (low,1) to (high,0) and from (low,0) to (high,1) resp. or vice versa when the measured data values range from low to high, and combined as ROC curve, forgetting the data values, from (0,0) to (1,1) where the x-axis is 1-specificity and the y-axis is sensitivity.

So do you need confidence bands? Then I'm sorry I can't help you. Or do you rely to specific parameters like the AUC (Area Under roc Curve) or to the Youden index, the maximum of all possible (sensitivity + specificity -1) values or, you speaking of sensitivity values, just those sensitivity values from the Youden index? The values you mention, 0.87 and 0.89, could be all 3 of them.

The first, since SAS9.2, can be done with PROC LOGISTIC and the ROC statement, the second and third one only, as far as I know, with bootstrap resamples of your data. Eventually this last task can be done with PROC SURVEYSELECT, but still I didn't apply that. Otherwise it's an easy task to create 1000 or 10000 resamples of your data and take the limits of the inner 95% of the values in question as the confidence limits

Hope that helps a bit.

Hans