Frequentist inference

The Frequentist inference is a statistical framework that treats probability as the long-run frequency of events in repeated trials, rather than as a degree of belief. In the frequentist view, parameters are fixed but unknown constants of nature, and data are random samples from a population. The goal of inference is to estimate these parameters, construct confidence intervals, and test hypotheses using procedures whose properties are guaranteed over infinite repetitions.

The framework's founding documents are Ronald Fisher's work on maximum likelihood estimation and the analysis of variance, Jerzy Neyman and Egon Pearson's theory of hypothesis testing, and the subsequent development of confidence intervals as a method of interval estimation. The central criterion is the sampling distribution: the distribution of a statistic over hypothetical repetitions of the experiment. A 95% confidence interval does not mean there is a 95% probability that the true parameter lies within it; it means that 95% of intervals constructed by this procedure will contain the true parameter across repeated samples.

This distinction is not merely philosophical. It has practical consequences. The frequentist framework does not permit probability statements about hypotheses — only about procedures. You cannot say "the probability that this hypothesis is true is 0.05"; you can only say "this procedure rejects true hypotheses 5% of the time in the long run." This has been criticized as a bait-and-switch: the scientist wants to know what is true; the frequentist answers with what would happen in an infinite sequence of experiments that will never be run.

The frequentist framework's greatest strength is its objectivity — or rather, its procedural objectivity. Two scientists with the same data and the same procedure will reach the same conclusion, regardless of their prior beliefs. This is valuable in regulatory science, where decisions must be defensible to adversarial scrutiny. But the cost is severe: the framework cannot incorporate prior knowledge, cannot update beliefs in light of new evidence in a direct way, and cannot answer the question that most scientists actually care about.

The frequentist-Bayesian debate is not a debate about which framework is correct. It is a debate about which question is being asked. The frequentist asks: what procedure has good properties over repetitions? The Bayesian asks: what should I believe given what I have seen? These are different questions, and neither is a special case of the other. The claim that one is "more correct" is a category error.

The frequentist framework is not wrong. It is answering a question that few scientists care about, using methods that cannot answer the question they do care about, and declaring the result objective because it does not depend on anyone's beliefs. But the absence of beliefs is not objectivity. It is ignorance dressed in a confidence interval.