Sequential / Group Sequential Trial Design

There is no hard limit, but each additional look reduces the per-look boundary and increases design complexity. In practice, two to five looks are most common. More frequent monitoring — approaching continuous sequential testing — can be handled with alpha-spending functions that allow information-fraction-based timing rather than equally spaced looks.

O'Brien-Fleming boundaries are very conservative early in the trial (requiring very large Z values to stop early) and relax toward the end; they are preferred in most clinical trials because they preserve most of the alpha for the final analysis. Pocock boundaries are constant across looks and easier to cross early, but they require a stricter final boundary and may leave the trial with insufficient power at the end.

Yes. When a trial stops early because of a favourable result, the observed effect is drawn from the upper tail of the sampling distribution and is likely an overestimate of the true effect. Adjusted estimators (e.g. the conditional mean estimator or shrinkage methods) should be used in the final report, especially if the result will feed into a meta-analysis.

Not without invalidating the Type I error control. Any unplanned look consumes alpha that was not accounted for in the original spending function. If a truly unplanned look becomes necessary for safety reasons, the resulting p-values must be interpreted with extreme caution and regulatory guidance should be sought.