Strategies for pooling data in occupational epidemiological studies.

In occupational epidemiology, most of the populations at risk are of limited size, and therefore pooling of the experience of different groups with similar exposure is desirable. We present and compare five different strategies for pooling of such data, including the "default" strategy of present reporting practices, which have the disadvantage that the experience of many small units is not recorded because of the low statistical power of small populations. A simple summation strategy including cases and controls from diverse groups, while attractively simple, may be fallacious when cases and controls have different proportions which are confounded with the risk ratios. Separate summation of observed and expected numbers of cases avoids the risk of fallacy, but may be unduly influenced by the results in one large plant. A "Clinical Trial" type of strategy in which all of the data are collected by a common protocol and therefore considered as a single data set is administratively complex and does not seem well suited for occupational epidemiology. Probability pooling seems to have attractive features, avoiding most of the disadvantages of the other strategies, and having the added feature that weighting for exposure and duration of follow-up can be included. Probability pooling requires exact probability statements for each study. Examples are given for pooling of data from exposures of operating room personnel to anaesthetic agents, cohorts of workers exposed to asbestos, and cohorts of workers exposed to man-made mineral fibres. A working group is proposed to recommend procedures and to assist in interpreting pooling activities.