Andrew S. C. Ehrenberg - Research Methods

Research Methods

An early interest in social science applications of statistics had already begun to show through in Cambridge (such as extensive experiments into the reliability of trained taste-testers for quality assessments and into price subsidies in the food industry). Also developed were two early aversions, the first to multivariate techniques imposed on simple data, and the second to mathematics for its own sake in applied statistics.

Ehrenberg's belief that the methods of physical science are applicable to social science was expressed in an article in the hard science journal Nature. In it he asserted that even in a field dominated by people's impulses to buy, that of marketing, there are striking regularities.

The discovery and development of such lawlike relationships, was described in series of papers. The definitive statement of this position came later in another paper read to the Royal Statistical Society. The summary stated baldly:

Ehrenberg always pointed out that very many models of very different functional form may be generated with almost identical goodness of fit. The selection of an appropriate model form to express a quantitative relationship is governed in his view by the need for:

1. Previous knowledge. This seems obvious but standard statistical modelling techniques seldom if ever start from any reference to any earlier model.
2. Simplicity. Models with many parameters cannot be tested in their variations by new data.
3. Adequate but not unnecessarily wide scope. Scope means the range of conditions in which the model has been found to apply. Empirical models are not of universal application. Decision on whether the scope is adequate depends on technical knowledge of the topic being modelled.
4. The absence of bias in predictions. Closeness of fit is secondary to this. Exceptions to the model may be noted.
5. A structural model, which remains meaningful under algebraic transformation. This rules out regression techniques - even simple regression produces two different equations.

The failure to use any technique of optimization either in model form selection or parameter determination has however been widely criticised. Ehrenberg has answered these criticisms by pointing to the absence of published and widely used models generated by conventional techniques. Another criticism is that Ehrenberg's published models deal with relationships between the averages of groups, and thus ignore variability between individuals.

The ideas in this section were used in one of Ehrenberg's principal fields of work, the study of buyer behaviour.