Reporting of attributable risk in epidemiologic studies.

Abstract

To the Editor: “Attributable risk” and “population attributable risk” (hereafter referred to as AR) are the terms used most often in the literature to describe the epidemiologic measure quantifying the impact of a risk factor on the disease burden in a population, ie, the hypothetical proportion of cases avoidable after total removal of the risk factor. This concept was originally introduced by Levin in 1953, 1 but was rarely used in early days. 2 Nakayama 3 now presents results of a literature search in 20 epidemiologic journals concerning the reporting of AR and compares these frequencies with those of reporting relative risks (RR). Similar to an earlier analysis not restricted to epidemiologic journals, 4 he gives data on the number of occurrences of AR and RR in the journals studied subdivided in two time periods and then concludes that a substantial under-reporting of AR compared with RR can be observed, a difference that has increased in magnitude over time. Furthermore, he argues that the true under-reporting of AR is even more extreme as he did not include synonyms like “risk ratio,” “rate ratio,” or “hazard ratio” for RR in his search. In this note, we address methodological concerns with automatic literature analyses like the one presented. We suggest that such analyses are not suited for a semantically ambiguous and complex area like AR. In an exhaustive literature search, we identified all publications in the period 1966–1996 accessible by MEDLINE that indicated in their abstract that results of AR calculations in epidemiologic studies were presented. 5 We used this collection as a starting point for evaluating a variety of aspects concerning the (mal-)practice of AR computations in epidemiologic studies. 6 To compile this collection we employed an automatic abstract scanning of AR-related phrases as a first step. As a second, essential step, we read all identified papers and verified their eligibility, because sustained confusion with respect to using the term AR and its many synonyms was apparent, a confusion that has consequences for all automatic search analyses. On the one hand, Nakayama has missed some studies in his search, as more than the terms used by him have been attached to Levin’s AR in the epidemiologic literature (a list of 16 terms is given by Gefeller 7). This low terminologic sensitivity is not severe, however, and could be remedied by using a more complete list of keywords for an automatic search. On the other hand, a substantial number of authors use the term AR but attach a different meaning to this term. Instead of referring to Levin’s definition they deal with the difference of disease risks among the exposed and unexposed, a quantity that has also been given the name “attributable risk,” even in some influential epidemiologic textbooks. 8-10 In our literature search we found more than 40 publications of that kind. Most of them appeared in the late 1970s and during the 1980s. We excluded them and all publications discussing only methodological aspects of AR manually from further consideration after having read the paper, which cannot be done in an automatic procedure. This step might explain why in our analysis on the usage of Levin’s AR, which was eventually based on 334 publications, a steep increase over time was apparent, whereas in Nakayama’s automatic analysis, a different picture emerged. We conclude that automatic literature searches followed by mere counting of phrases in abstracts have only limited areas of applicability. Especially in more complex areas like the concept of AR, such approaches can be seriously misleading and should thus be abandoned. Annette Pfahlberg Olaf Gefeller Wolfgang Uter