Re: Body Mass Index and Risk of Lung Cancer Among Never, Former, and Current Smokers

Abstract

Using a standardized series of dose– response cancer-specific meta-analyses, we previously reported positive associations between body mass index (BMI) and the risks of several cancers (1). Two smokingrelated cancers, namely lung and esophageal squamous cell carcinoma, were notable exceptions: the risks of these cancers were inversely associated with BMI. Given the established observation that BMI values are generally lower in current smokers than in never smokers, we reasoned that the observed inverse associations reflected residual confounding by smoking. In a recent article in the Journal, Smith et al. (2) have comprehensively addressed this question in almost half a million individuals from the NIH AARP cohort. Among current smokers, they found an inverse association between BMI categories and the risk of incident lung cancer both in men and women when only adjusting for age. The authors extended their age-adjusted models to control for detailed smoking status and dose, cigar or pipe smoking. This adjustment hardly changed the five hazard ratios and the 95% confidence intervals when comparing different BMI categories with the 22.5–24.9 kg/m 2 reference category. Additional inclusion of race or ethnicity, education level, history of emphysema, physical activity, and alcohol intake in the analysis also hardly changed the strength of the estimated associations. They concluded that “our inability to attribute the inverse association between BMI and the risk of lung cancer to residual confounding by smoking or to bias suggests the need for considering other explanations.” We propose, as alternative explanation, a larger degree of measurement error for the assessment of smoking than for BMI. Under many conditions, nondifferential measurement error in the main exposure variable produces bias toward the null. However, this is not necessarily the case if the main exposure and another strong risk factor, such as a confounder, are correlated. Using methods similar to those described by others (3), we created a hypothetical 100 000-person population with a 2% risk of lung cancer to investigate, in logistic regression analyses, the size of the apparent exposure–outcome association that might occur, when in truth, the main exposure (BMI, in this example) has no causal effect, but the other risk factor (smoking) is strongly associated with the outcome (Table 1). When we assumed a modest inverse correlation (Pearson correlation coefficient of −0.10) between true BMI and the true number of cigarettes smoked per day, we obtained effect sizes for the inverse association of BMI with the risk of lung cancer that were similar to those for women current smokers, which were reported by Smith et al. (2) In the absence of measurement error, the association with BMI becomes null after adjustment for smoking. However, when allowing for plausible levels of measurement error in both variables (4), a substantial inverse association between BMI and the risk of lung cancer remains even after controlling for smoking. Smith et al. (2) advanced interesting biological hypotheses about the role of estrogens among smokers. However, these hypotheses would not explain the contrasting observation (in the same cohort) that greater waist circumference, another anthropometric measure of body adiposity, is associated with an increased risk of lung cancer mortality (5). The explanation is probably simpler: incomplete removal of confounding due to measurement error in a strong confounder.