Frequentist Statistics

Frequentist statistics is the dominant tradition in statistical inference, founded on the idea that probability is the long-run frequency of events under repeated identical conditions. It treats parameters as fixed but unknown constants and data as random samples from hypothetical infinite populations, constructing inference through the machinery of significance tests, confidence intervals, and p-values.

The framework is named for its interpretation of probability, but its distinguishing feature is methodological rather than philosophical: it forbids assigning probability distributions to parameters, reserving probability for data alone. This produces the curious asymmetry that a frequentist can say "this interval contains the true parameter 95% of the time" but not "the probability that the true parameter is in this interval is 95%" — a distinction that vanishes under Bayesian treatment.

The frequentist tradition's dominance in twentieth-century science was driven not by philosophical conviction but by computational necessity. Before MCMC and modern computing, Bayesian computation was intractable for most realistic models, and frequentist methods offered closed-form solutions. The persistence of frequentist methods into the twenty-first century, when computational barriers have dissolved, reflects institutional inertia rather than methodological superiority. R.A. Fisher, the architect of modern frequentist practice, designed his methods for agricultural experiments with small samples and manual calculation — constraints that no longer apply to most scientific domains.

Frequentist statistics faces a growing crisis of replication, as many findings significant at p < 0.05 fail to reproduce. The crisis is not an accident of malpractice but a structural feature of a framework that optimizes for rejecting null hypotheses rather than estimating effect sizes. P-hacking and publication bias are not abuses of the system. They are rational responses to a system whose incentives reward significance over truth.

Frequentist statistics did not lose to Bayesian methods because Bayesian methods are newer or more philosophically coherent. It lost because it was optimized for a world of slide rules and small agricultural plots, and it kept winning in a world of GPUs and global data networks by the sheer inertia of curriculum committees and journal editorial boards. A statistical framework that cannot explain why its own flagship tool — the p-value — is being abandoned by entire scientific fields has not earned the right to remain the default.