Childhood Leukemia, Military Aviation Facilities, and Population Mixing

In the November 2003 issue of EHP, Liu et al. (2003) concluded that " relatively low concentrations of gaseous air pollutants are associated with adverse effects on birth outcomes. " Although this may be true from a purely statistical sense, there appear to be limitations of this research that suggest cautious interpretation of the findings. Liu et al. (2003) evaluated individual-level birth certificate data, which is an improvement over the ecologic designs of past time-series studies on pollution. However, birth records do not contain most of the variables that are important predictors of low weight and preterm births. These include smoking, alcohol and/or drug abuse, low socioeconomic status (SES), small maternal weight or height, complications of the current or previous pregnancy [e.g., pregnancy-induced hypertension, previous low birth weight (LBW), spontaneous abor-tion], insufficient weight gain during pregnancy , maternal illness (e.g., fever), and job-related exertion (Berkowitz and Moore 2003). Many of these are major factors that substantially affect risk. For example , maternal smoking during pregnancy, which has a prevalence of 10–20% in the United States (Ebrahim et al. 2000; O'Campo et al. 1995), is associated with a 2-to 4-fold increase in risk of LBW or growth restriction (Kramer 1987; Lang et al. 1996; Nordentoft et al. 1996). Therefore, there is considerable room for uncontrolled confounding that might account for the small odds ratio of 1.05–1.10 observed by Liu et al. (2003). Liu et al. (2003) argued that uncontrolled or residual confounding is an unlikely explanation for their results because a) there is no evidence that these factors are associated with air pollution; b) ecologic measures of SES did not modify the associations; and c) " there were only slight differences between crude and adjusted estimates, " and " individual characteristics ... did not attenuate the risk estimates. " However, these arguments have limitations. First, there may not be evidence that important risk factors co-vary with pollution , but it seems reasonable that many might correlate with residential location. Liu et al. (2003) linked pollution measurements in 13 census subdivisions to births within those subdivisions. If gaseous pollutant measurements and other factors (e.g., SES, smoking prevalence) co-vary by census subdivision, then confounding could occur. Second, ecologic measures are poor surro-gates for individual-level ones, which can result in confounder misspecification and Third, the individual-level covariates included in some of the models did appear to have substantive impacts. For example, the odds ratio …