Jerzy Neyman (27 July 1894 – 16 April 1981) was a Polish-American mathematician and statistician whose work helped establish the foundations of modern statistics. With Egon Pearson, he developed a theory of statistical hypothesis testing based on controlling error probabilities and comparing the power of alternative tests. He also gave confidence intervals their systematic frequentist formulation and made influential contributions to survey sampling, agricultural experimentation, and causal inference. At the University of California, Berkeley, he built a major center for statistical research and training.
Life and academic career
Neyman was born in Bendery, then part of the Russian Empire and now generally known as Bender, in present-day Moldova, into a Polish family. He studied mathematics at Kharkov University, where his interests increasingly turned toward probability and its applications.
After moving to Poland in 1921, he worked on statistical problems associated with agricultural experiments and obtained his doctorate at the University of Warsaw in 1924. Research visits to London and Paris brought him into contact with leading European statisticians and mathematicians. His collaboration with Egon Pearson began in the 1920s and produced a series of papers that reshaped statistical testing.
In 1934, Neyman joined University College London. In 1938, he moved to the University of California, Berkeley, where he established the Statistical Laboratory. Berkeley created a separate Department of Statistics in 1955, with Neyman as its first chair. The Berkeley Symposia on Mathematical Statistics and Probability, which he initiated, became important international meetings for theoretical and applied statistics.
Neyman received the United States National Medal of Science in 1968. He remained active in research after his formal retirement and died in 1981.
The Neyman–Pearson theory of testing
Errors and power
Neyman and Pearson developed a framework in which a statistical test is a rule for deciding, from observed data, whether to reject a specified null hypothesis in favor of an alternative hypothesis.
Their theory distinguishes two kinds of error:
- A Type I error occurs when a true null hypothesis is rejected.
- A Type II error occurs when the null hypothesis is not rejected even though the specified alternative is true.
A test's significance level bounds its Type I error probability. Its power is the probability of rejecting the null hypothesis under an alternative. The central design problem is therefore not simply to make rejection unlikely under the null, but to achieve high power against relevant alternatives while respecting the chosen error bound.
For a rejection region , a simple null hypothesis , and a simple alternative , this problem can be expressed as
Here, denotes the data and is the maximum permitted Type I error probability. For a composite null hypothesis, the error bound must hold across all parameter values included in that hypothesis.
The Neyman–Pearson lemma
Their landmark 1933 paper, On the Problem of the Most Efficient Tests of Statistical Hypotheses, established the result now called the Neyman–Pearson lemma.
For testing one fully specified distribution against another, the lemma shows that a most powerful test rejects the null when the ratio
is sufficiently large, with possible randomization at the boundary to achieve an exact error probability. The functions and are the probability densities or probability mass functions under the two hypotheses. The ratio compares their likelihoods for the observed data.
The lemma provides a precise optimality result for simple hypotheses. It does not establish that a single best test exists for every problem involving unknown parameters. With composite alternatives, different tests may be most powerful against different alternatives, and additional restrictions or optimality criteria may be required.
Interpretation
Neyman emphasized evaluating procedures through their performance over hypothetical repetitions. A test is justified by its error rates under specified models, not by treating its output as a probability that a hypothesis is true.
This perspective differs from interpreting a p-value as a measure of incompatibility between observations and a null model. Although modern practice often combines significance levels, p-values, and power calculations, these elements arose within partly distinct approaches to inference.
Confidence intervals
In his 1937 paper Outline of a Theory of Statistical Estimation Based on the Classical Theory of Probability, Neyman presented a systematic theory of confidence intervals.
A confidence procedure assigns an interval to each possible dataset. A confidence coefficient of at least requires
for every parameter value covered by the model. The probability concerns the random interval produced by sampling; the parameter is treated as fixed.
Thus, a 95% confidence procedure covers the true parameter in at least 95% of hypothetical repetitions under the stated assumptions. Once a particular interval has been observed, the frequentist construction does not itself assign a 95% probability to the fixed parameter lying inside that particular interval.
This distinguishes confidence intervals from credible intervals in Bayesian inference, which express posterior probabilities under a specified prior and likelihood.
Neyman's framework also clarified the connection between interval estimation and testing: a confidence set can be constructed from the parameter values not rejected by a corresponding family of tests. Coverage alone does not determine a unique procedure, however; interval length and other performance properties also matter.
Survey sampling and experimental design
Probability-based sampling
Neyman's 1934 paper On the Two Different Aspects of the Representative Method: The Method of Stratified Sampling and the Method of Purposive Selection became a foundational contribution to survey methodology.
He explained how random sampling permits uncertainty to be evaluated through the sampling design, rather than relying solely on a judgment that a selected sample appears representative. His treatment of stratified sampling showed how dividing a population into groups and selecting samples within them could improve precision.
For a fixed total sample size and equal per-unit sampling costs, the allocation now known as Neyman allocation assigns stratum sample sizes according to
where is the population size of stratum and is its within-stratum standard deviation. Larger and more variable strata receive more observations. The optimal allocation changes when sampling costs differ or other constraints apply.
Randomized experiments and causal effects
Neyman's 1923 work on agricultural experiments provided an early formal treatment of treatment-specific potential yields, anticipating the modern potential-outcomes framework for causal inference.
For each experimental unit, the framework distinguishes the outcome under one treatment from the outcome under another. Only the outcome corresponding to the assigned treatment is observed. Randomization provides the basis for studying the behavior of estimators of average treatment effects.
This work helped connect experimental design to probability-based estimation. It also exposed a fundamental limitation: individual treatment effects are generally not directly observable because a unit cannot simultaneously reveal its outcomes under mutually exclusive treatments.
Statistical philosophy and disagreements
Neyman was a principal architect of frequentist inference, particularly the view that statistical methods should be assessed by operating characteristics such as coverage, error probabilities, and power.
His approach differed from that of Ronald Fisher, despite their shared reliance on probability models and randomization. Fisher emphasized significance testing as a means of assessing evidence against a null hypothesis. Neyman and Pearson developed rules for choosing between hypotheses with specified long-run error properties. Their disagreements also concerned the interpretation of probability and the purposes of statistical inference.
The distinction has practical consequences. Failure to reject a hypothesis need not establish its truth, particularly when a test has low power. Likewise, controlling a false-rejection rate does not establish the substantive importance of a detected effect. The validity of the calculations remains conditional on the probability model and the sampling or experimental design.
Research breadth and institutional influence
Beyond foundational inference, Neyman worked on applied problems in agriculture, astronomy, and weather modification. His research frequently involved constructing probability models for scientific mechanisms rather than treating statistical analysis as a purely mechanical calculation.
At Berkeley, he promoted close interaction between mathematical theory and substantive scientific problems. The Statistical Laboratory, the statistics department, and the Berkeley Symposia helped establish mathematical statistics as an independent academic discipline. His lasting technical contributions include likelihood-ratio optimality for simple-hypothesis testing, repeated-sampling coverage for interval estimation, and design-based reasoning in surveys and experiments.