Statistical basis of public policyBMJ 1997; 314 doi: https://doi.org/10.1136/bmj.314.7073.72 (Published 04 January 1997) Cite this as: BMJ 1997;314:72
Epidemiology does not need Bayesian inference
- Paul Brennan, Lecturer in epidemiology and medical statisticsa
- a Arthritis and Rheumatism Epidemiology Research Unit, University of Manchester Medical School, Manchester M13 9PT
- b Nuffield College, Oxford OX1 1NP
- c Department of Statistical Science, University College London, London WC1E 6BT
- d Department of Epidemiology and Public Health and Department of Statistical Science, University College London, London WC1E 6BT
- e Centre for Social and Economic Research on the Global Environment, University of East Anglia, Norwich NR4 7TJ
- f 37 College Avenue, Melton Mowbray LE13 0AB
Editor–R J Lilford and D Braunholtz propose that a Bayesian perspective in the interpretation of epidemiological results may prove fruitful.1 Their central thesis is that public health decisions are often based on the results of significance tests. Instead they propose that other prior knowledge of the association be incorporated into the calculation of a subjective Bayesian probability. It is unlikely for several reasons that this will happen in epidemiology.
Firstly, they are wrong in suggesting that conventional significance tests are a basis for public health policy. Policy decisions made on the basis of epidemiological data are determined by a whole host of reasons, including the perceived validity of the results, the absolute risks, and the consequences of inaction. It is rare that the probability of the association, whether frequentist or Bayesian, plays more than a minor part.
Secondly, Bayesian statistics does not address the main challenges when interpreting epidemiological data. Any effect measure can be contaminated by three factors–confounding, bias, and chance. Con- founding presents the most serious challenge because its possibility can never be excluded. Lilford and Braunholtz do not mention the role of confounding when interpreting results, and it seems that Bayesian statistics has little to offer in this area. A Bayesian assessment of bias also seems to offer little. The authors present an example of how the potential for bias can be incorporated into a final analysis by saying that if epidemiological studies tend to overestimate results by 30% then this may be corrected by reducing the effect estimates by a similar amount. An alternative is to anticipate which selection and information biases are likely to be problematic and to design better studies that minimise them.
While Bayesian statistics has proved valuable in certain areas–for example, in diagnostic medicine, where prior beliefs are based on hard …