Abstract
Confidence intervals based on the central limit theorem (CLT) are acornerstone of classical statistics. Despite being only asymptotically valid,they are ubiquitous because they permit statistical inference under very weakassumptions, and can often be applied to problems even when nonasymptoticinference is impossible. This paper introduces time-uniform analogues of suchasymptotic confidence intervals. To elaborate, our methods take the form ofconfidence sequences (CS) -- sequences of confidence intervals that areuniformly valid over time. CSs provide valid inference at arbitrary stoppingtimes, incurring no penalties for "peeking" at the data, unlike classicalconfidence intervals which require the sample size to be fixed in advance.Existing CSs in the literature are nonasymptotic, and hence do not enjoy theaforementioned broad applicability of asymptotic confidence intervals. Our workbridges the gap by giving a definition for "asymptotic CSs", and deriving auniversal asymptotic CS that requires only weak CLT-like assumptions. While theCLT approximates the distribution of a sample average by that of a Gaussian ata fixed sample size, we use strong invariance principles (stemming from theseminal 1970s work of Komlos, Major, and Tusnady) to uniformly approximate theentire sample average process by an implicit Gaussian process. We demonstratetheir utility by deriving nonparametric asymptotic CSs for the averagetreatment effect based on doubly robust estimators in observational studies,for which no nonasymptotic methods can exist even in the fixed-time regime (dueto confounding bias). These enable doubly robust causal inference that can becontinuously monitored and adaptively stopped.