The CHDS was a prospective cohort study that enrolled about 20,000 pregnant women either attending prenatal clinics or giving birth at Kaiser Foundation Health Plan Medical Centers in the San Francisco Bay Area during the 1960's 202122. The Kaiser Health Plan, a prepaid integrated health care system, at that time had enrolled 90 percent of membership through employer or union groups (G Friedman, Kaiser-Permanente, personal communication,1996). The population served was thus largely persons having stable employment, with wealthy and very poor persons generally not well represented. Otherwise, members represented a broad cross-section of the San Francisco Bay Area population.

The 399 woman/child pairs in this study were selected from 3,412 CHDS children born April 1964 through April 1967 who participated in an extensive examination at five years of age, described previously 2324. To be eligible for our study (Table 1), women had to have completed the enrollment interview before delivery and donated a second or third trimester blood specimen, for which a sufficient volume of serum still remained. As the study was designed to examine a range of birth and developmental outcomes, exclusions were made for: 1) non-singleton deliveries, 2) children with severe anomalies or who did not complete two cognitive exams and a hearing screening; 3) mothers who were deaf, had rubella during pregnancy, were taking thyroid medication in the 60 days prior to blood draw, or took iodine-containing medications during pregnancy or in the six months prior to conception 25; and 4) infants born at gestational ages that were unknown, less than 35, or greater than 45 completed weeks. Based on regularly updated contact information provided by the CHDS investigators, we selected children residing in eight counties in or surrounding the greater San Francisco Bay Area. The study was approved by the Human Subjects Review Board at the University of California, Davis.

After these exclusions, there remained a group of 1,291 children. We sampled all children with low scores on at least one of two cognitive tests (below 10^th percentile on either the Peabody Picture Vocabulary Test or an abbreviated Raven's Progressive Matrices), or who failed a hearing screening, and a 17 percent random sample of all others, to achieve the targeted sample size. The motivation for this design was to enhance power for the hypotheses related to neurocognitive and neurosensory deficits (which were of relatively low prevalence) while not adversely affecting efficiency for other hypotheses. Several authors have published papers on "reusing" data from case- control studies – designs that are, in essence, a special case of stratified sampling 2627. At the analysis stage, we adjusted for this design by applying normalized weights in all regression model fitting (detailed below). Finally, if women had more than one eligible child, we sampled no more than one child for each mother. Serum PCB concentrations were determined for 399 women in this final sample.

Laboratory methods have been described previously 1728. Briefly, serum specimens collected during pregnancy were aliquoted and stored at -20° Celsius at National Institutes of Health (NIH) facilities. For determination of PCBs the protocol was as follows: Samples were thawed and then homogenized; surrogate standards (e.g, 2,3,4,4'-tetrachorobiphenyl (PCB International Union of Pure and Applied Chemistry (IUPAC) #66)), and 2,3,3',5,5',6-hexachlorobiphenyl (PCB #165) and glacial acetic acid were added, and the solution was mixed. The PCBs were extracted three times with 90 percent hexane/10 percent dichloromethane, and the extract was passed through a 0.5 percent deactivated florisil column. The PCBs were eluted with 60 milliliters of hexane and then 60 milliliters of 50 percent hexane/50 percent dichloromethane. After concentration by rotary evaporation, the internal standard 2,2',3,4,4',5,6,6'-octachlorobiphenyl (PCB #204) was added. Serum extracts were analyzed by gas chromatography with electron capture detection (Hewlett-Packard 6890 Series) in parallel using an RTX-5MS and an RTX-1701 column. All samples were analyzed between January 14, 1998 and September 22, 1999 in batches of 10 to 20 samples. Concentrations were adjusted for percent recovery for each sample. Specimens analyzed on batch dates with extremely high or low variability were re-analyzed, and the two values averaged.

Before analyzing any samples from our subjects, noncritical samples from the same cohort (i.e. of women in CHDS who were lost to follow-up) were analyzed to assess the integrity of the stored serum specimens. Organochlorine determinations demonstrated comparability with other historical samples, and lipids were in reference ranges.

We included PCBs that: (a) were measured in mixtures reported (as of the late 1990's) to have associations with reproductive, developmental or endocrine alterations in animals or humans; (b) did not co-elute with other compounds on either of the two columns; and (c) were present in sufficient concentration in these samples 28. These included: PCB #101, #105, #110, #118, #137, #138, #153, #156, #170, #180, and #187. Values below the limit of quantitation (LOQ) were imputed as described previously 17. The total of PCBs was calculated as the sum of the nine congeners that had fewer than 30% below the LOQ (PCB IUPAC #105, #110, #118, #137, #138, #153, #170, #180, and #187).

Sex of the infant was collected at the time of birth and noted in the CHDS database. Because of our design, only data on pregnancies that resulted in a live birth were included.

Potential covariates were selected from CHDS databases which included data on sociodemographics and maternal health and pregnancy. To adjust for the potential effects of maternal sex hormone or corticosteroid use during the periconception period, all prescriptions and associated dates were extracted for women in this study. A complete list of medications prescribed during the index period between thirty days prior to and four months after conception was reviewed using references pertinent for the time period of the study 293031. Active ingredients in medications from earlier years were checked against current references 3233. Mothers were coded (yes/no) according to whether they had been prescribed sex hormones or corticosteroids. Other potential adjustment variables for the multivariate analysis were specimen characteristics. These were: batch-date of laboratory analysis and an indicator for sample storage history (some specimens had been immediately shipped for storage at an NIH facility while others had been stored locally and later shipped).

Log-binomial and logistic regression analyses were conducted to estimate the relative risk (RR) and odds ratio (OR), respectively, of male birth as a function of maternal PCB concentrations. Since the odds ratio is a poor approximation of the relative risk when the outcome is common 34, the relative risk of male birth was estimated using log-binomial regression 35, and forms the basis of our primary analysis. As a secondary analysis and for purposes of comparison with previous research, we estimated the odds of male birth in relation to maternal PCB concentrations using logistic regression. All regression models were adjusted for specimen characteristics and prescription for sex hormones or corticosteroids. Some maternal factors have been implicated with altered sex ratio such as maternal age, parity, and race 36. However, since inclusion of these measures did not alter PCB coefficients by more than 10 percent, they were not included in our final multivariate models. Separate models were fit for total PCBs and for the 9 specific PCB congeners that had less than 30 percent of their values below the LOQ. In all regression models, PCB concentration was modeled as a continuous variable in micrograms per liter of serum (μg/L) (parts per billion (ppb)). To provide a more interpretable measure of the PCB "effect size," we calculated a risk ratio comparing sex ratio at the 90^th vs. 10^th percentiles of PCBs by multiplying the per-unit PCB coefficient by the concentration difference between the 90^th and 10^th percentiles. In addition, normalized weights were applied to adjust for our stratified sampling procedure; these weights were inversely proportional to the sampling fraction in each stratum. When applied to the regression analysis, the results are as if the sample of 399 children was drawn completely at random from the subset of 1291 children. Thus, any potential bias created by the stratified sampling procedure is removed. As in standard survey research, this procedure ensures that all parameter estimates are valid with respect to the larger sample or population from which the stratified sample has been selected, and eliminates any bias introduced by use of a stratified sample. To account for potentially greater homogeneity within sampling strata, robust variance estimates were calculated for both log-binomial and logistic regression models. All analyses were conducted using the GENMOD procedure in SAS (version 9.1; SAS Institute, Inc., Cary, North Carolina).

Table 2 compares the sociodemographic characteristics of the original CHDS cohort, the 3,412 children who appeared for the 5-year examination, the 1,291 eligible after exclusions, and the final study sample of 399. As compared with the original cohort, the mothers of those examined at five years were of similar age, more highly educated, and less likely to be white. Those eligible and selected for our study differed from the original cohort on the same factors, with a greater percentage of African-Americans meeting our eligibility criteria. Female children were slightly more prevalent among those participating in the 5-year examination than in the original cohort, and among those eligible after our exclusions; finally, more females were included in our study because of poorer average performance on the Raven's Progressive Matrices, one of the cognitive tests for which children with low scores were over-sampled. Our study sample included 53.4 percent females as compared with 48.8 percent of the original cohort. As described above, normalized weights were used to account for stratified sampling; since the sampling altered the sex ratio, these weights also resulted in correction for the differential selection by gender.

The mean, median, 10^th and 90^th percentile of total maternal PCBs were: 5.4, 4.7, 3.1, and 8.7 μg/L (ppb) of serum, respectively. PCB #153, the most abundant congener in this study, had a median of 1.1 μg/L (10^th percentile: 0.6 μg/L serum; 90^th percentile: 1.9 μg/L serum). These PCB concentrations are similar to levels reported by others for the United States from the same time period 37, and higher than most but not all populations under study today 18.

Log-binomial regression analyses, in which we adjusted for specimen characteristics and hormonally active medications, revealed that total maternal PCB concentration was associated with a reduction in male births (p = 0.02). Specifically, the relative risk of a male birth decreased by 33 percent comparing women at the 90^th percentile of total PCBs with those at the 10^th percentile of total PCBs (RR = 0.67; 95% CI: 0.48, 0.94), or by approximately 7 percent for each 1 μg/L (1 ppb) increase in total PCBs (Table 3). In log-binomial regression models for specific PCB congeners, the predicted relative risk of a male birth decreased with higher maternal PCB concentration in each and every congener measured, though most congeners were only marginally associated with male birth. The results from logistic regression models were similar in direction, showing, as expected, larger odds ratios than the relative risks derived from the log-binomial regression analyses.