Of the original 269,746 singleton births, we were able to geocode 94.6%, with only negligible differences in this rate between case and control subgroups. A further 4.6% of these records were excluded because estimated gestational age was incompatible with birth weight. From the remaining births, we identified 465 ASD cases plus 6,975 matched controls. ASD cases were 85.2% male; for controls this proportion was 51.4% (further information on demographic characteristics is shown in Table 1 ). Eight cases and 100 controls had missing data for at least one covariate of interest; in nearly all instances this covariate was maternal education. For each regression model, only subjects with complete information for all necessary covariates were included.
Coefficients for ASD risk comparing the fourth nonzero quartile of exposure to no exposure among children born in selected California counties, 1996–1998. Only coefficients for which a minimum of 20 subjects had nonzero exposure are shown. Model controls for maternal education, maternal race/ethnicity, and RC of diagnosis (imputed for controls). Open circles represent coefficients for organochlorine pesticides applied during the CNS period; closed circles represent all others.
ORs (solid lines) and lower 95% confidence limits (dotted lines) for ASD comparing nonzero quartiles of organochlorine pesticide applications within 500 m to no applications for overlapping 8-week temporal windows. Models control for maternal education, maternal race/ethnicity, and RC of diagnosis (imputed for controls). x-Axis is the date in the center of each temporal window relative to fertilization date; shading indicates clinical first trimester; gaps indicate no ASD cases occurred for that category.
Dose–response curve for ASD risk with organochlorine pesticides applied within 500 m of residence during a posteriori temporal period (26–81 days postfertilization), controlling for maternal education, maternal race/ethnicity, RC of diagnosis (imputed for controls), and sex of child. Plus signs indicate data points from which curves have been calculated. Solid line, OR estimate; dotted lines, 95% confidence limits.
A total of 249 combinations of compounds, buffer radii, and temporal periods met the requirement of five exposed cases and controls per cell. The coefficients comparing the fourth nonzero quartile of exposure to the reference category are presented for the eight combinations where the p-value was < 0.05 (the unadjusted α) in Table 2 .
Regardless of buffer radius, all fourth nonzero quartile coefficients meeting our numeric criterion for significance adjusted for multiple testing were for the category of organochlorine pesticides with applications occurring during the CNS period; furthermore, only regressions with this compound–temporal period combination yielded p-values that met this criterion. Generally, these coefficients had p-values an order of magnitude smaller than those for the next most significant coefficients.
The p-values and fourth nonzero quartile coefficients are plotted in Figure 1. As expected because of multiple testing, most of these coefficients are randomly distributed around zero, with a few having p-values close to 0.05. The coefficients for organochlorine exposure during the CNS period, in contrast, have p-values substantially smaller than their nearest neighbors on the graph and are consistently positive.
Organochlorine pesticides were found to be associated with ASD regardless of the buffer radius used. The effect becomes monotonically smaller as the radius gets larger; when the buffer radius is extended to 1,750 m, the fourth nonzero quartile odds ratio (OR) finally becomes nonsignificant (p > 0.05; data not shown). For the a posteriori analysis, we selected the radius of 500 m, which was the smallest for which there was at least one case for each exposure category.
Only organochlorine compounds met the criteria for inclusion in a posteriori analyses. Using a 500-m radius around residential locations, we allowed the 8-week temporal window to be centered anywhere between 300 days before and 300 days following estimated date of conception. Although significant coefficients (α = 0.05) were found for alternative time periods and among nonzero quartiles besides the fourth, these are dwarfed in magnitude and significance by those occurring during the first trimester of gestation among the highest quartile of exposure (Figure 2). Shifting the temporal window so that it starts just following neural tube closure (day 26) yielded the largest fourth nonzero quartile OR = 7.6 [95% confidence interval (CI), 3.1–18.6]. ORs from regression modeling using both a priori and a posteriori time periods for the organochlorine category of pesticides are presented in Table 3 .
In the study area, dicofol and endosulfan accounted for > 98% (by poundage) of the organochlorines applied. During the temporal period identified through the a posteriori analysis (i.e., days 26–81), 88 subjects resided within 500 m of a dicofol application and 27 within 500 m of an endosulfan application. Because of these small numbers, a full set of ORs could not be calculated separately for each of the two compounds. Analysis using radii > 500 m suggested magnitudes of association slightly higher for endosulfan than for dicofol; the association of each compound with ASD appeared to be largely independent of the other (data not shown).
Our initial model controlled for maternal race and ethnicity, education, and RC of diagnosis (recorded or imputed). To assess model sensitivity, we employed the a posteriori time window and the 500-m buffer and investigated models using no covariates, our original covariates plus maternal age and child sex, and various combinations of these. ORs were not significantly altered under any model, although we did observe some attenuation of the association when sex was included in the model. Given the observed sex ratio among cases and our low exposure prevalence (1.5%), nearly all exposed cases were male, so this attenuation should not necessarily be construed as evidence for confounding or effect modification by sex. Inclusion of covariates besides sex nonsignificantly increased, rather than decreased, the observed association. Choice of the initial covariates plus sex in the model yielded a fourth nonzero quartile OR of 6.1 (95% CI, 2.4–15.3). Repetition of the simulated RC assignment for controls 100 times yielded a median estimate for this number of 6.1 (95% CI, 2.4–15.5), minimum 5.8 (95% CI, 2.3–14.6), and maximum 6.7 (95% CI, 2.6–17.2).
Characterization of the dose–response relationship between organochlorine pesticide applications and ASD risk is shown in Figure 3. Risk appears to increase monotonically up to the application amount of approximately 22 lb during the 8-week period with the highest OR determined a posteriori. This poundage is equivalent to the 87th percentile for the nonzero applications in the sample; beyond this magnitude of application, data are too sparse to allow for the calculation of risk, as evidenced by the widening of CIs and the attenuation of the OR back to the null.