A half-day meeting, organised by the History of Statistics Study Group, was held at Rothamsted Research in Harpenden on Thursday 18 June 2015 to discuss the question: How did the needs of agricultural researchers come to play such a key role in the development of modern statistical methods? The main speaker was Dr Giuditta Parolini, Postdoctoral Fellow at Technische Universität Berlin, who had completed a PhD on RA Fisher and subsequently used archive material at Rothamsted and interviews with former staff to assemble material for a book proposal.
The appointment of Fisher to Rothamsted in 1919 and his impact on statistical theory and practice is well known. Less well known is the subsequent history of the statistics department in the last century led by Frank Yates, John Nelder, John Gower and Vic Barnett. The meeting placed the Fisher period in perspective, with talks on the earlier history of Rothamsted before Fisher’s appointment to the current work and prospects under the present head of the Applied Statistics Group, Andrew Mead. An exhibition of early calculators, including Fisher’s Millionaire, was displayed, along with a collection of books and papers written by Rothamsted staff. The meeting ended with a field trip to the famous Broadbalk Wheat experiment, the subject of Fisher’s first analysis.
Andrew Mead introduced the meeting and described research projects involving statistical modelling and biomathematics relating to current work in the Rothamsted departments.
Dr John Jenkyn, co-archivist for the Lawes Agricultural Trust, described the early history of Rothamsted up to the appointment of Fisher in 1919. The founder of Rothamsted Experimental Station was Sir John Lawes, owner of Rothamsted Manor, who saw the opportunity to improve agricultural production by the use of mineral fertilisers, including phosphates for which he obtained a lucrative patent, allowing him to invest the subsequent wealth in agricultural research. The Broadbalk Wheat Experiment was begun in 1843 to demonstrate that the same crop could be grown on a piece of land provided its fertility was replenished each year. Other experiments on barley, grassland and root crops were begun and are known as the 'classical experiments'. Samples of crops and soils have been kept from the earliest years of the experiments and provide today a valuable archive. With his co-director Henry Gilbert a disciplined set of procedures for conducting experiments, keeping accurate records of yields and meteorological variables, was maintained. Lawes showed an interest in the wider use of statistics and served on the council of the Royal Statistical Society. In 1901, following the deaths of Lawes and Gilbert, Rothamsted entered a new era, needing new sources of funding to expand the work to include a range of disciplines from soil chemistry and agricultural botany to the study of pests and diseases. By 1919, the current director, Sir John Russell, saw the need for a statistician to summarise and interpret the long series of records and to advise on the wide range of statistical problems arising in current research activity.
Dr Parolini gave a wide-ranging account of the people who established statistics as a modern discipline, the computing tools required, and the influence of statistics on agricultural science and practice. She described the influence of Fisher’s predecessors, especially WS Gosset (‘student’) whose empirical work on sampling distributions had been given theoretical proof by Fisher before he came to Rothamsted. Fisher’s first task was to summarise the results of 70 years of long-term experiments, for which he developed the application of orthogonal polynomials, multiple regression and the analysis of variance. But Rothamsted scientists provide a wide range of statistical problems, from the design of annual experiments on fertiliser treatments, the analysis of insecticidal assays, and the estimation of bacterial numbers from dilution series trials. All these challenges helped to crystallise his ideas on the fundamentals of statistical inference (including the method of maximum likelihood) and the principles of experimental design.
Fisher’s work attracted worldwide attention, and many assistant staff and visitors such as John Wishart, Harold Hotelling and L Tippett helped to spread his ideas to other disciplines. Fisher appointed Frank Yates in 1931 and when he left to succeed Karl Pearson at the Galton Laboratory in London in 1933, Yates became head of the department. In the years up to 1939 Yates built up the department, appointing William Cochran, Frank Anscombe, Oscar Kempthorne and David Finney among others. Yates played an important role in the war effort, advising the government on sampling and survey methods, and in the 1940s created a statistical research service for British agricultural institutes and Ministry experimental farms. The acquisition of the first electronic computer, the Elliott 401 in 1954 saw the appointment of staff to programme the machine and develop new statistical methods for the computer age. Michael Healy, Howard Simpson and John Gower developed new methodology, and the department collaborated with statisticians at other institutes, including John Nelder at the National Vegetable Research Station at Wellesbourne, and Clifford Pearce at East Malling Research Station. John Nelder succeeded Frank Yates in 1968 and appointed Graham Wilkinson from Adelaide to develop his general algorithm for the analysis of variance, leading to the Genstat computer package, widely used today. With Robert Wedderburn, who died tragically early in his career, Nelder developed the methodology for Generalised Linear Models and the GLIM system. John Gower developed many innovations in multivariate analysis and cluster analysis, and succeeded Nelder as head of the department in 1980.
Dr Parolini paid special tribute to the assistant staff, mainly female, who worked with desk calculators and tape and card punches to ensure high quality data input, checked by two independent operators to reduce the possibility of mistakes in processing. The scientists who brought their problems to the statisticians also had an important role, both in stimulating research and in disseminating good statistical practice in the design of experiments and sampling and in the reporting and publication of scientifically acceptable work.
The tools available to Fisher in 1919 were little more than an old slide rule, Fuller’s Spiral Rule - useless for addition. Following Gosset’s recommendation, he asked for a modern calculator, the Millionaire, which he and Yates used up to the 1940s, and on which the calculations for Fisher and Yates’s Tables were made. For survey work, Yates acquired a Hollerith tabulator, card sorter and card punch, and in 1954 negotiated the delivery of the prototype Elliot 401 machine, lying unused in Cambridge. The transition from single task programmes to packages and networks took place as larger and faster machines were acquired over the years, while other institutes acquired their own machines and shared the software and ideas developed at Rothamsted.
In the final section of her talk, Dr Parolini discussed how agricultural problems stimulated statistical theory and methodology, not only in the advances in experimental design but in likelihood theory, sampling theory and multivariate analysis. The basic principles of experimental design, randomisation, replication, blocking and factorial arrangements, led to more efficient research and were copied in other disciplines such as medicine, industry and psychology. Fisher and Yates were involved in many arguments with other agricultural statisticians, such as Gosset and Jerzy Neyman, and in the compromise between the ideal and affordable practices. Rothamsted continued to include a statistician on each of its practical committees in the planning and approval of experimental proposals. There were strong links with agricultural work in developing countries, and many overseas visitors came to learn their skills at Rothamsted.
We look forward to more published results from Dr Parolini’s research.