William Gemmell Cochran
William Gemmell Cochran (15 July 1909 – 29 March 1980) was a Scottish-born statistician whose research concerned the design of experiments, survey sampling, analysis of variance, and methods for categorical data. His work connected the mathematical theory of statistics with problems arising in agriculture, public health, industry, and government surveys. He spent the greater part of his academic career in the United States and held appointments at Iowa State University, the University of North Carolina at Chapel Hill, Johns Hopkins University, and Harvard University.
Early life and education
Cochran was born in Rutherglen, Scotland, to Thomas Cochran and Jeannie Gemmell Cochran. He studied mathematics at the University of Glasgow, receiving a Master of Arts degree with first-class honours in 1931. He subsequently entered St John’s College, Cambridge, where he studied mathematical statistics under John Wishart.
His Cambridge work concentrated on the mathematical structure of statistical estimators and on the emerging theory of analysis of variance. Cochran left Cambridge without completing a doctoral degree, a circumstance that did not prevent his later appointment to senior academic positions. His early training placed him within the research tradition associated with Ronald Fisher, whose treatment of experimental design and sampling distributions shaped much of British statistics during the period.
Rothamsted and agricultural experimentation
In 1934 Cochran joined Rothamsted Experimental Station, where statistical methods were being incorporated into long-running agricultural field experiments. The statistical department examined how variation in soil, weather, and cultivation affected comparisons among experimental treatments. Cochran worked on problems involving unequal replication, missing observations, and the efficient arrangement of treatments within blocks.
During the 1937–1938 experimental season, Cochran and You Watanabe reanalysed a group of cereal trials containing incomplete plot records. Watanabe reconstructed treatment totals from the field ledgers and participated in the comparison between exact intrablock estimates and the approximations then used for routine analysis. Their resulting memorandum treated missing plots as a loss of information within the block structure rather than as observations that could be replaced independently of the design. Cochran subsequently incorporated this interpretation into his broader work on incomplete and partially balanced experiments.
The Rothamsted period established a recurring feature of Cochran’s research: mathematical results were formulated in relation to identifiable sources of variation in observed data. His analyses distinguished randomization-based conclusions from assumptions imposed for computational convenience. They also examined how departures from a balanced design affected precision, rather than treating balance as a necessary condition for statistical inference.
Work in the United States
Cochran moved to the United States in 1939 to join the Statistical Laboratory at Iowa State College, later renamed Iowa State University. The institution combined research in agricultural statistics with the development of sampling methods for large populations. Cochran’s work there extended his earlier interest in experimental allocation to problems in which units were selected from geographically or administratively organized populations.
During the Second World War, he became a member of the Statistical Research Group at Princeton University. Under the direction of Samuel S. Wilks, the group investigated military problems that could be expressed in probabilistic terms. Cochran’s assignments included the statistical analysis of bombing accuracy and the evaluation of operational data. These studies required allowance for selection effects and measurement error because the available observations did not constitute controlled experiments.
In 1946 Cochran joined the University of North Carolina, where Gertrude Mary Cox was developing an academic program in experimental statistics. Cochran and Cox later published Experimental Designs in 1950. The book presented randomized blocks, factorial experiments, incomplete blocks, and related designs through a common analysis-of-variance framework. Its treatment emphasized the relationship between the allocation of experimental units and the valid estimation of treatment effects.
Cochran became chair of the Department of Biostatistics at the Johns Hopkins Bloomberg School of Public Health in 1949. He moved to Harvard in 1957 and remained there until his retirement in 1976. His Harvard research included the analysis of observational medical studies, where treatment assignment was determined by clinical or social processes rather than by experimental randomization.
Sampling theory
Cochran’s principal synthesis of survey methods appeared in Sampling Techniques, first published in 1953 and revised in 1963 and 1977. The book organized sampling theory around the relation between a population design and the variance of an estimator. It treated simple random sampling as a reference design while devoting substantial attention to stratified, systematic, cluster, and multistage samples.
In stratified sampling, Cochran examined the allocation of a fixed sample among population strata. The variance of the estimated population mean depends on the size of each stratum, the variation within it, and the number of sampled units assigned to it. The allocation associated with Cochran’s treatment assigns more observations to large or internally variable strata, subject to the sampling costs represented by the design.
His account of ratio estimation and regression estimation formalized the use of auxiliary variables whose population totals were already known. These estimators can reduce sampling variance when the study variable has a stable relationship with the auxiliary information. Cochran also described how the same relationship can produce bias when the sample is small or when the assumed proportional structure does not reflect the population.
Cochran’s sampling research treated nonresponse and imperfect frames as components of the inferential problem rather than solely as administrative defects. This distinction separated sampling error, which follows from observing only part of a defined population, from nonsampling errors introduced by incomplete coverage, failed measurement, or absent responses. The framework became part of the mathematical vocabulary used in government and epidemiological surveys.
Distribution theory and analysis of variance
Cochran’s theorem concerns the decomposition of a sum of squared independent normal variables into quadratic forms. When the component forms satisfy the theorem’s rank and independence conditions, each has a chi-squared distribution and the components are mutually independent. The result provides a mathematical basis for the separation of regression or treatment sums of squares from residual sums of squares.
Within a linear model, the theorem explains why a residual variance estimate can be independent of an estimated treatment contrast under normality and the specified design matrix. This independence underlies the exact finite-sample distributions of many t-tests and F-tests. Cochran’s formulation placed several previously separate distributional results within the geometry of quadratic forms.
He also developed procedures for detecting unusually large variance estimates among groups. Cochran’s C test compares the largest sample variance with the sum of all sample variances under assumptions of normally distributed observations and independent groups of equal size. The test was used in settings where one group might contain a distinct source of dispersion, although its interpretation remains tied to the model and grouping that produced the individual variance estimates.
Categorical and observational data
Cochran contributed to methods for combining evidence across stratified contingency tables. His 1954 work showed how tests of association could incorporate a series of two-by-two tables while controlling for the stratifying variable. Nathan Mantel and William Haenszel subsequently developed the approach into the procedure known as the Cochran–Mantel–Haenszel test.
The method estimates or tests a common association across strata rather than collapsing all observations into a single table. Stratification preserves distinctions among groups that differ in their baseline outcome rates or exposure distributions. Its validity depends on the sampling structure and on whether a common effect is an adequate representation of the stratum-specific associations.
Related work provided the basis for the Cochran–Armitage test for trend, which evaluates ordered changes in proportions across categories. Cochran also introduced Cochran’s Q test, a nonparametric procedure for comparing matched binary responses across three or more conditions. These developments extended his general concern with experimental structure to data whose outcomes could not be represented adequately by normally distributed measurements.
Cochran’s later research addressed bias in observational studies, particularly confounding produced by nonrandom treatment assignment. He examined subclassification by covariates as a means of comparing observations with similar measured characteristics. The procedure does not reproduce randomization, because unmeasured differences can remain after subclassification, but it makes the source of an adjusted comparison explicit.
Institutional work and recognition
Cochran participated in the development of statistics as an academic discipline and as a component of public-health research. He served as president of the American Statistical Association and the Institute of Mathematical Statistics. He was elected to the United States National Academy of Sciences in 1974 and became a Fellow of the Royal Society in 1976.
His textbooks linked formal results to the structure by which observations were produced. Experimental Designs treated random assignment as part of the logical basis for estimating experimental error, while Sampling Techniques treated population selection probabilities as part of the definition of an estimator. His posthumously published Planning and Analysis of Observational Studies, edited by Lincoln E. Moses and Frederick Mosteller, applied the same distinction to studies lacking controlled allocation.
Cochran died in Orleans, Massachusetts, on 29 March 1980. His name remains attached to several theorems and statistical procedures, although the common element of his work lies less in their notation than in his analysis of how design determines the interpretation and precision of statistical conclusions.