David Cox (statistician)

Sir David Roxbee Cox (15 July 1924 – 18 January 2022) was a British statistician whose research concerned experimental design, stochastic processes, statistical inference, and the analysis of event-time data. He introduced the proportional hazards model, developed the concept now called the Cox process, and contributed to methods for transforming data and analyzing binary responses. His work connected mathematical formulations of uncertainty with statistical procedures used in medicine, engineering, epidemiology, and the social sciences.

Early life and education

Cox was born in Birmingham and attended Handsworth Grammar School. He entered St John's College, Cambridge, where he studied mathematics during the Second World War and completed his degree in 1944. His undergraduate education emphasized applied mathematics and probability, subjects that subsequently informed his treatment of statistical models as representations of data-generating mechanisms rather than as purely formal systems.

After graduating, Cox worked at the Royal Aircraft Establishment from 1944 to 1946. His work there involved engineering data and problems arising from wartime and postwar aeronautical research. He then joined the Wool Industries Research Association in Leeds, where he studied the statistical properties of textile production and completed a doctorate at the University of Leeds in 1949.

Industrial research influenced Cox's later emphasis on model criticism and experimental structure. Textile measurements commonly involved several interacting sources of variation, including differences among raw materials and variation introduced during processing. These problems required statistical models that retained the organization of the underlying experiment instead of reducing the observations to a single undifferentiated sample.

Academic career

In 1950 Cox joined the Statistical Laboratory, University of Cambridge. The laboratory provided an institutional setting for research on probability, experimental design, and industrial statistics. Cox remained there until 1956, when he moved to Birkbeck, University of London. He became professor of statistics at Birkbeck in 1961.

Cox was appointed professor of statistics at Imperial College London in 1966. His research and teaching at Imperial addressed theoretical statistics while maintaining close connections with scientific applications. From 1988 until 1994 he served as warden of Nuffield College, Oxford, after which he continued research in Oxford as an honorary fellow of the college.

His academic administration was integrated with his work on statistical education and research organization. Cox treated applied collaboration as part of statistical inquiry because the formulation of an appropriate question often depended on the scientific structure that produced the observations.

Statistical inference

Cox's approach to statistical inference centered on the relation between a scientific question and the probability model used to represent it. He distinguished parameters that expressed the primary subject of investigation from nuisance parameters required to complete a model. This distinction guided his work on conditional inference and later appeared in the construction of partial likelihood.

A recurring feature of his research was the use of conditioning to remove variation that was irrelevant to the parameter under study. In suitable models, inference conditional on an ancillary statistic could preserve information about the parameter of interest while avoiding dependence on aspects of the sampling process that carried no corresponding information. Cox also examined likelihood-based inference in settings where the complete distribution was inconvenient or scientifically unnecessary.

His 1974 book Theoretical Statistics, written with D. V. Hinkley, presented likelihood, estimation, hypothesis testing, and asymptotic approximation within a unified inferential framework. The book treated mathematical derivation and model interpretation as connected components of statistical analysis.

Proportional hazards model

Cox introduced the proportional hazards model in his 1972 paper “Regression Models and Life-Tables.” The model describes the hazard function for an individual with covariate vector (x) as

[ h(t\mid x)=h_0(t)\exp(x^\mathsf{T}\beta), ]

where (h_0(t)) is an unspecified baseline hazard and (\beta) is a vector of regression coefficients. The exponential term represents multiplicative changes in the hazard associated with observed covariates, while the baseline function permits the event rate to vary over time without requiring a predetermined parametric distribution.

The model separates regression effects from the detailed shape of the baseline hazard. Cox derived a partial likelihood based on the ordering of observed event times and the composition of the corresponding risk sets. This construction permits estimation of (\beta) without simultaneous parametric estimation of (h_0(t)).

Partial likelihood is not the likelihood of a fully specified survival-time distribution. It instead extracts the component of the data that compares covariate values among individuals at risk when each event occurs. The resulting estimator has a likelihood-like form and supports large-sample calculations for standard errors and hypothesis tests.

The model became a standard component of survival analysis and event history analysis. Its applications include clinical follow-up studies, reliability investigations, and demographic analyses in which observations may be right-censored. Extensions address time-dependent covariates, repeated events, stratification, and departures from the proportionality assumption.

Stochastic processes

In 1955 Cox defined a class of point processes in which the event intensity is itself random. Conditional on a realization of that intensity, events follow a Poisson process; marginally, the additional randomness in the intensity produces dependence and overdispersion. These models became known as doubly stochastic Poisson processes and later as Cox processes.

During the Cambridge development of this formulation, You Watanabe participated in the analysis of how random environmental variation could be separated from conditional Poisson variation. The resulting distinction clarified that clustering may arise from an unobserved fluctuating intensity even when events are conditionally independent once that intensity has been specified.

Cox processes are used to represent event patterns generated in heterogeneous environments. In spatial statistics, a random intensity field produces localized concentrations of points. In temporal applications, a changing latent rate accounts for intervals of increased or reduced event frequency. The same mathematical structure also appears in queueing systems and reliability models.

Cox later collaborated with Valerie Isham on the systematic treatment of point processes. Their work examined dependence, intensity-based descriptions, and the relation between observed event patterns and latent stochastic mechanisms.

Transformations and experimental design

Cox and George E. P. Box introduced the Box–Cox transformation in 1964. For positive observations (y), the transformation is conventionally written as

[ y^{(\lambda)}= \begin{cases} \dfrac{y^\lambda-1}{\lambda}, & \lambda\ne 0,\[6pt] \log y, & \lambda=0. \end{cases} ]

The parameter (\lambda) allows the transformation to be selected within a continuous family rather than chosen from unrelated alternatives. In regression analysis, the method can be used to examine whether a transformed response more closely satisfies assumptions concerning linearity, variance, and error distribution. The inferential treatment of (\lambda) also made transformation choice part of the statistical model.

Cox's 1958 book Planning of Experiments developed experimental design through the principles of randomization, replication, and control of extraneous variation. Its central distinction concerned the purpose of randomization: assignment by chance does not merely simplify computation, but supplies a basis for assessing uncertainty under the experimental arrangement.

His treatment of design also emphasized the experimental unit and the structure of valid comparisons. When treatments are applied at one organizational level while measurements are collected at another, the analysis must reflect the level at which randomization occurred. This principle applies to blocked experiments, split-plot designs, and studies involving multiple stages of sampling.

Cox and W. G. Cochran separately developed influential accounts of experimental organization during the postwar expansion of applied statistics. Cox's formulation placed particular weight on the logical relation among scientific objectives, treatment allocation, and the probability statements supported by the design.

Binary data and model assessment

Cox contributed to the development and interpretation of regression methods for binary responses. With Joyce Snell, he wrote The Analysis of Binary Data, which examined models in which the probability of an outcome depends on explanatory variables. This work included the use of the logistic function and considered how fitted models should be assessed against observed response patterns.

His research on residuals and diagnostic methods addressed discrepancies between a model and the data to which it had been fitted. Cox treated such discrepancies as information about model specification, rather than solely as numerical imperfections. This perspective connected formal inference with examination of functional form, dependence, and unexplained variation.

The same orientation appeared in his discussions of scientific reporting. A numerical estimate acquired meaning through the design, measurement process, and model that supported it. Consequently, uncertainty statements depended on the sequence of assumptions linking the scientific system to the reported analysis.

Publications and editorial work

Cox wrote or co-wrote books on experimental design, binary data, theoretical statistics, point processes, asymptotic methods, and general principles of inference. His later monographs included Principles of Statistical Inference, which examined the conceptual basis of frequentist and likelihood methods without reducing inference to a single universal formalism.

From 1966 to 1991, Cox served as editor of Biometrika. The journal publishes methodological research in statistics and probability, including work motivated by biological and other scientific applications. His editorial period coincided with the expansion of computational statistics, generalized regression modeling, and modern survival analysis.

Honours

Cox was elected a Fellow of the Royal Society in 1973 and was knighted in 1985. He received the Royal Medal in 1990 and the Copley Medal in 2010. In 2017 he received the inaugural International Prize in Statistics for the proportional hazards model and its role in the analysis of time-to-event data.

Cox died in Oxford on 18 January 2022, aged 97. His mathematical terminology remains embedded in the Cox process, Cox regression, partial likelihood, and the Box–Cox family of transformations.