The positive association between shark-attack incidence and ice-cream sales is one of the most-cited textbook examples of confounded correlation in observational data. To our knowledge, the pair has never been empirically decomposed in the peer-reviewed literature. We pre-register peak-phase offset analysis - the lag between annual peak weeks of two suspected shared-confound variables - as a visualization-based diagnostic for shared-driver hypotheses and characterize its operating envelope across an eight-probe calibration suite spanning three pre-registered probe classes and two temporal resolutions. We report two findings. First, at weekly resolution with a state-matched reference, the diagnostic discriminates phase structure cleanly: three reference-ablation tests pass, weekly dispositions are leave-one-year-out stable, and the shark probe is reference-source robust under a Google Trends Wikimedia pageviews swap. Second, at monthly resolution with a nationally-pooled reference- a configuration forced by the NCHS Mortality file’s state-of-occurrence suppression policy- the diagnostic fails on negative-controls under a strict pre-registered gate. We localize this failure under a fourth pre-registration lock, advance a reference-stability hypothesis, then refute our own hypothesis under a fifth lock. We position peak-phase offset analysis as a single-figure teaching tool for undergraduate methodology courses, offered in the spirit of the GAISE College Report’s multivariate-thinking recommendation.
Nathan Humphrey (Sat,) studied this question.