Explain how you would use logistic regression in an A/B testing framework.

Official answer

For a binary outcome such as purchase within seven days, fit a prespecified model such as logit P(Y=1 | T,X)=α+βT+γX, where T is randomized assignment and X contains selected pretreatment covariates. Preserve the experiment's assignment unit and intention-to-treat population; avoid adjusting for treatment-induced variables such as post-assignment engagement.

In a model without treatment interactions, exp(β) is a conditional odds ratio at fixed X. It is not a risk ratio or a percentage-point lift. For example, doubling odds from a baseline probability of 0.20 changes probability to 1/3, not 0.40.

If the decision concerns the average conversion difference, predict each target unit's probability under T=1 and T=0 and average the differences, with uncertainty that accounts for model fitting and the design. State the standardization population and modeling assumptions; adjusted and unadjusted odds ratios need not agree even without confounding.

Check missing outcomes, separation, functional form, interactions and cluster dependence. Randomization supplies the basis for causal comparison; logistic regression alone does not explain the mechanism, repair broken assignment or guarantee transport to new populations. Report the simpler randomized contrast as a useful reference alongside the planned adjusted analysis.