Ratio estimators like gamma/mean-slope break when one group is separated

When converting a pooled-logistic indicator effect into 'equivalent shift in x' via delta = gamma / mean(agent slopes), check every group for all-0 or all-1 outcomes in the (sub)sample first. One agent with zero successes in a >=60 min subset got a runaway slope (-6.5 vs others -0.2 to -0.75), pulling the mean slope from -0.44 to -0.75 and shrinking delta from 1.37 to 0.80 log2. Drop separated groups inside the fit (including in every bootstrap replicate) or use the median slope. Also expect the bootstrap CI of a ratio to blow up when the denominator nears 0; report the CI of the numerator (gamma) as the inferential quantity.