Abstract
Leukemia is the most common childhood cancer in industrialized countries, and the increasing incidence trends in the US suggest that environmental exposures play a role in its etiology. Neighborhood socioeconomic status (SES) has been found to be associated with many health outcomes, including childhood leukemia. In this paper, we used a Bayesian index model approach to estimate a neighborhood deprivation index (NDI) in the analysis of childhood leukemia in a population-based case-control study (diagnosed 1999 to 2006) in northern and central California, with direct indoor measurements of many chemicals for 277 cases and 306 controls <8 years of age. We considered spatial random effects in the Bayesian index model approach to identify any areas of significantly elevated risk not explained by neighborhood deprivation or individual covariates, and assessed if groups of indoor chemicals would explain any elevated spatial risk areas. Due to not all eligible cases and controls participating in the study, we conducted a simulation study to add non-participants to evaluate the impact of potential selection bias when estimating NDI effects and spatial risk. The results in the crude model showed an odds ratio (OR) of 1.06 and 95% credible interval (CI) of (0.98, 1.15) for a one unit increase in the NDI, but the association became slightly inverse when adjusting for individual level covariates in the observed data (OR = 0.97 and 95% CI: 0.87, 1.07), as well as when using simulated data (average OR = 0.98 and 95% CI: 0.91, 1.05). We found a significant spatial risk of childhood leukemia after adjusting for NDI and individual-level covariates in two counties, but the area of elevated risk was partly explained by selection bias in simulation studies that included more participating controls in areas of lower SES. The area of elevated risk was explained when including chemicals measured inside the home, and insecticides and herbicides had greater effects for the risk area than the overall study. In summary, the consideration of exposures and variables at different levels from multiple sources, as well as potential selection bias, are important for explaining the observed spatial areas of elevated risk and effect estimates.