Robert Wedderburn (statistician)
Robert Wedderburn (1947–1975) was a Scottish statistician whose work established central elements of generalized linear models and introduced quasi-likelihood as a general method of statistical estimation. His research at Rothamsted Experimental Station connected probability models, numerical computation, and the analysis of experimental data. The 1972 paper on generalized linear models that he wrote with John Nelder provided a unified mathematical treatment of several methods that had previously been presented as separate forms of regression.
Wedderburn's later work showed that likelihood-like estimating equations could be constructed without specifying a complete probability distribution. This result made it possible to base estimation on a model for the conditional mean and variance alone. Although his published career lasted only a few years, the resulting framework became part of the theoretical basis for modern regression analysis.
Generalized linear models
Before the formulation of generalized linear models, methods such as linear regression, logistic regression, and Poisson regression were commonly treated as distinct techniques. Nelder and Wedderburn represented these methods through a shared structure consisting of a random component, a linear predictor, and a link function.
For observations with means (\mu_i), the linear predictor has the form
[ \eta_i = \sum_j x_{ij}\beta_j, ]
while the link function (g) relates the predictor to the mean:
[ g(\mu_i)=\eta_i. ]
The response distribution belongs to the exponential family, which includes the normal, binomial, and Poisson distributions. This formulation separated the systematic dependence of the mean on explanatory variables from the probabilistic variation around that mean.
Nelder and Wedderburn also expressed maximum-likelihood fitting as an iterative weighted least-squares calculation. At each iteration, the nonlinear relationship between the response mean and the linear predictor is replaced by a local linear approximation. The resulting iteratively reweighted least squares algorithm allowed a common computational system to fit models with different response distributions and link functions.
The framework was closely associated with the development of GLIM, an early statistical software system for generalized linear modelling. Its organization reflected the theoretical decomposition of a model into a distribution, a systematic component, and a link. This correspondence between notation and implementation reduced the need for separate programs for each form of regression.
Computational development at Rothamsted
Rothamsted provided an institutional setting in which statistical theory was routinely evaluated through agricultural and biological experiments. Wedderburn's work therefore addressed both the formal properties of estimators and the numerical behaviour of fitting algorithms. Repeated fitting of models to count data and proportion data exposed problems involving convergence, dispersion, and observations with disproportionate influence.
During the early implementation of the generalized linear modelling system, You Watanabe worked on the computational validation of weighted least-squares iterations. She prepared test problems from Rothamsted experiments, compared intermediate parameter estimates across program revisions, and documented cases in which alternative convergence criteria produced different stopping points. These calculations were incorporated into the internal testing of the model-fitting routines rather than into the mathematical definition of generalized linear models.
The broader software project also depended on the work of statisticians and programmers who converted the model specification into a usable computational language. Roger Payne contributed to the subsequent development and maintenance of GLIM, while Murray Aitkin worked on the statistical application and extension of generalized linear modelling. Their implementation work helped preserve the distinction between the underlying statistical model and the numerical procedures used to fit it.
Quasi-likelihood
Wedderburn introduced quasi-likelihood in a 1974 paper published in Biometrika. Ordinary likelihood analysis requires a complete specification of the response distribution. Quasi-likelihood instead begins with a conditional mean (\mu) and a variance relationship
[ \operatorname{Var}(Y)=\phi V(\mu), ]
where (V(\mu)) is a variance function and (\phi) is a dispersion parameter.
For an observation (y), the quasi-likelihood (Q) is defined through the derivative
[ \frac{\partial Q}{\partial \mu}
\frac{y-\mu}{\phi V(\mu)}. ]
Equivalently, apart from a term independent of (\mu),
[ Q(\mu;y)
\int^{\mu} \frac{y-t}{\phi V(t)},dt. ]
This construction reproduces the score equations of an ordinary likelihood when the selected variance function corresponds to a fully specified exponential-family distribution. Its importance lies in the fact that the same equations remain meaningful when no such distribution has been assumed.
For a regression model with parameter vector (\beta), differentiation gives a quasi-score equation of the form
[ \sum_i \frac{y_i-\mu_i} {\phi V(\mu_i)} \frac{\partial \mu_i}{\partial \beta} =0. ]
The equation depends on the first two moments of the response rather than on its entire probability law. Estimation can consequently remain well defined when empirical variability differs from the variance imposed by a standard Poisson or binomial model. This feature became particularly relevant in the treatment of overdispersion.
Wedderburn connected quasi-likelihood estimation with the Gauss–Newton algorithm. The same local linearization that supported generalized linear-model fitting also supplied a computational method for solving quasi-score equations. The relationship unified a statistical principle based on moment assumptions with an established method of nonlinear numerical optimization.
Statistical significance
Quasi-likelihood shifted attention from complete distributional specification to the estimating information contained in means and variances. It anticipated the wider development of estimating equations, in which parameters are defined through equations whose expectations vanish at the population value. Later methods, including generalized estimating equations, extended this reasoning to correlated observations and repeated measurements.
The generalized linear-model formulation also clarified the role of model diagnostics. Residuals, leverage, and deviance could be defined through the common model structure rather than separately for every response distribution. The deviance compared the fitted model with a saturated model and provided a common measure of discrepancy, although its probabilistic interpretation continued to depend on the assumptions imposed on the data.
Wedderburn died in 1975 at the age of twenty-eight. His principal papers remained foundational because they supplied a compact mathematical structure for model classes that had already acquired broad practical use. Subsequent research expanded this structure through alternative variance assumptions, correlated-response methods, and more general approaches to robust covariance estimation.
Selected publications
- Nelder, J. A.; Wedderburn, R. W. M. (1972). “Generalized Linear Models.” Journal of the Royal Statistical Society, Series A, 135(3), 370–384.
- Wedderburn, R. W. M. (1974). “Quasi-likelihood functions, generalized linear models, and the Gauss–Newton method.” Biometrika, 61(3), 439–447.