I have a repeated measures study design with two separate conditions in a memory task where one must recognize previously seen items in a list. The conditions are concrete vs. abstract. I'd like to generate a confidence interval of the effect of concreteness on the probability of correct recognition.
I had considered bootstrapping or back transform of logistic regression estimates but it seems to me that stratified bootstrapping would not weight proportions near 1. Further, I'd like the CI coverage to be of the effect in proportion rather than the back transformed effect.