Bioelectromagnetics 27:674 ^ 676 (2006) Comment Comment Concerning ‘‘Childhood Leukemia and Residential Magnetic Fields: Are Pooled Analyses More Valid Than the Original Studies?’’ (Bioelectromagnetics 27:1^7 [2006]) Leeka Kheifets, 1 * Gabor Mezei, 2 and Sander Greenland 1,3 1 Department of Epidemiology, University of California, Los Angeles, USA 2 Environment Department, Electric Power Research Institute, Palo Alto, California, USA 3 Department of Statistics, University of California, Los Angeles, USA Elwood [2006] contrasted the conclusions of three selected studies [Linet et al., 1997; McBride et al., 1999; UKCC, 1999] that there is no evidence for an association, to the positive findings of two pooled analyses by Ahlbom et al. [2000] and Greenland et al. [2000]. Elwood argued that these discrepancies may result from shortcomings of the pooled analyses, and suggested that the conclusions of the original studies may be more valid. While we agree that the issue of whether an effect is present remains far from settled, we believe that his analysis involves several conceptual and methodologic oversights which undermine his conclusion. First, the three studies were misinterpreted as ‘‘negative’’ because they found no statistically signifi- cant association, when in fact they did supply evidence for an association. It is unfortunately common to confuse lack of significance with lack of evidence or lack of positive association, but as well documented in textbooks [Royall, 1997; Rothman and Greenland, 1998, Ch. 12] it is a fallacy nonetheless, for the significance reflects numbers in categories as much as strength of the association. In reality, 11 of 12 studies used by Greenland et al. exhibit positive associations across a wide variety of analyses, as have three subsequent studies [Greenland and Kheifets, 2006]. Almost any association can be made to appear unstable and violate dose-response by dividing the extreme categories too finely, as Elwood does by using a 0.5 microtesla upper category. When analyzed using stable categories, however, results from the 15 studies are remarkably consistent, and remain positive when various bias adjustments are made [Greenland and Kheifets, 2006]. And as Greenland et al. showed, taking account of matching factors only increased the association. (Incidentally, Ahlbom et al. and Greenland et al. analyses included 9 and 12 studies with magnetic field exposure measures, not 10 and 13 as stated by Elwood; also, one milligauss (mG) equals 0.1 micro- tesla, not 10 microtesla as Elwood states.) Second, the three studies seemed to attract credi- bility in part because they had a large case numbers. But basic formulas [e.g., Rothman and Greenland, 1998, Ch. 14] show that statistical information (inverse variance) is mainly determined by the number of sub- jects in the smallest cell, which here is the number of cases in the highest exposure category. For example, although the UKCC study provided 33% of all cases (1,073/3,247) in the Ahlbom et al. pooled analysis, it contributed only 9% of the cases (4/44) exposed to 4 mG, and has fewer cases 4 mG than 4 of the 12 studies in Greenland et al. [2000]. Third, personal measurements only provide the best estimate of personal exposure during the actual measurement day. The etiologically relevant exposure, however, covers the years prior to disease development, and personal measurements may be far noisier than other surrogates for this exposure. This greater noisi- ness would likely lead to greater bias toward the null when using personal measurements. Furthermore, due to behavioral changes in children diagnosed with a debilitating disease, personal measurements taken after ß 2006 Wiley-Liss, Inc. —————— *Correspondence to: Leeka Kheifets, Professor, UCLA School of Public Health, Department of Epidemiology, 73-284 CHS, 650 Charles E. Young Drive South, Los Angeles, CA 90095-1772. E-mail: kheifets@ucla.edu Received for review 2 February 2006 DOI 10.1002/bem.20257 Published online 17 August 2006 in Wiley InterScience (www.interscience.wiley.com).