Examining associations between occupation and health by using routinely collected data

When examining a large number of associations simultaneously, as happens when routinely collected data are used to assess associations between occupation and health, it is not obvious how best to identify associations requiring further investigation since some risks may be high, or low, by chance alone. We have developed an approach to deal with this problem which is relatively easy to apply and appropriate to applications where data are not too sparse. Observed to expected ratios are estimated using an empirical Bayes procedure. Anomalous associations can be seen as outliers in a normal probability plot of the log-ratios. The method is illustrated in the analysis of 252000 cancers registered in men in England during 1981-87.