Regarding "meta-analysis and causal inference: a case study of benzene and non-Hodgkin lymphoma": an incomplete analysis.

This Journal recently published an article by Douglas Weed (1) dedicated to criticizing our meta-analysis, which identified evidence of a causal association between non-Hodgkin lymphoma (NHL) and benzene and between NHL and petroleum refinery work (2). The whole of Weed’s article is devoted to criticizing our paper, although under the pretext of using our work as a ‘‘case study’’. Although we appreciate Weed’s interest in our paper, his major points are either incomplete or incorrect. Weed suggests that meta-analysis can only play three roles in evaluating causal inference: increased precision, evaluations of heterogeneity, and publication bias. But meta-analysis involves so much more. For example, our meta-analysis involved an extensive literature search, evaluation of hundreds of studies, and quantitative or qualitative evaluations of precision, magnitude of effect, heterogeneity, biologic plausibility, healthy worker effect, confounding, multiple sensitivity analyses, publication bias, review of comparable work.and so on. Many other meta-analyses have quantitatively assessed dose-response (3), chance (e.g., p values), latency (4), effect modification (5), biologic plausibility (6), temporality (7), outcome misclassification (8), exposure misclassification (9), diagnostic bias (10), recall bias (11), selection bias (12, 13), confounding (14), funding source (15), and many of the other elements of causal inference. We don’t claim that every meta-analysis does all things. However, the suggestion that the process of metaanalysis is limited to the three things described by Weed is incorrect and even inconsistent with some of his previous statements (16). Weed states that because of a ‘‘lack of consistency (by which he means statistically significant heterogeneity), weak associations, no evidence of dose-response, no effort to provide an assessment of biological plausibility, and no new epidemiological evidence’’ we should not have concluded that the elevated relative risks we identified provide ‘‘further evidence that benzene exposure causes NHL’’. However, his review of each of these factors is wrong or incomplete. For example, it is implied that because our meta-analyses of NHL and refinery work involved statistically significant tests for heterogeneity, the data are inconsistent and ‘‘combining them was not warranted’’. However, numerous authors have warned against over-interpreting these tests (14, 17). In Weed’s own words from a previous article, ‘‘statistical heterogeneity.need not preclude