Methods

Design and study population

The study presented here used data from the National Birth Defects Prevention Study (NBDPS) which is a large, collaborative case–control study of birth defects currently being conducted in the United States. All appropriate Institutional Review Board approvals were obtained. The goal of the NBDPS is to evaluate the association between specific birth defects and environmental and genetic factors. A total of 10 birth defects centres (Arkansas, California, Centers for Disease Control and Prevention (Atlanta), Iowa, Massachusetts, New Jersey, New York, North Carolina, Texas and Utah) have contributed data for the years 1997–2003, which are included here. The methods used by the NBDPS have been previously described.18

This spectrum analysis included all cases of isolated (a single major defect) and multiple (more than one unrelated major defect present in an infant) birth defects and all control births included in the NBDPS from the participating centres with estimated dates of delivery from 1 October 1997 through 31 December 2003. The NBDPS does not collect information on single gene conditions and chromosomal abnormalities. Defect cases considered complex (multiple embryologically related defects), with the exception of amniotic band sequence and heterotaxy, were excluded because they may have different aetiologies from isolated/multiple cases. All cases were identified using each centre’s birth defects surveillance system and were reviewed by clinical geneticists using case-specific criteria including standardised definitions of defects, and confirmatory diagnostic procedures.19 Controls consisted of non-malformed live-born infants, randomly selected from either birth certificates or birth hospitals. In the NBDPS, the overall ratio of cases to controls is approximately 3:1. As the same control group was used for analysis of each defect, the ratio of cases to controls will vary. However, for some centres, data on certain defects were collected for a limited number of years. For these defects, only similarly collected control data were used. Infants were excluded for the following reasons: they were adopted, they were in foster care, or their mother did not speak English or Spanish.

Data collection

A computer-assisted telephone interview with mothers was conducted in English or Spanish by trained female interviewers. Participants were enrolled using a standard procedure.18 During the maternal interviews information was collected on maternal health status, including chronic diseases, medications used, pregnancy history and complications, maternal diet, vitamin use, caffeine, tobacco, alcohol or illicit drug use, household exposures, work history, family demographics, and water use patterns. Timing of exposures during pregnancy was recorded in the way the mother recalled best, for example by date, month or trimester. A pregnancy calendar was used to aid recall. Interviews were completed between 6 weeks and 2 years after the estimated date of delivery, an interval chosen with regard to the situation of mothers of newborns, especially those with defects, and also to allow sufficient time to ascertain all defect cases, which may take up to 2 years.

Outcomes

A total of 45 specific defects were included in the study (table 1). These defects are the major defects for which the NBDPS collects data. They were summarised into nine systems (shown in table 1): amniotic band (non-system-specific), central nervous system, ears, eyes, gastrointestinal, genitourinary, cardiovascular, musculoskeletal and orofacial. Only amniotic bands and ear defects do not contain specific defect subgroups. In the case of neural tube defects (under “Central nervous system”), limbs (under “Musculoskeletal”) and clefts (under “Orofacial”), additional subgroups of specific birth defects were analysed in this study.

Exposure

A combination of occupation (job title) and industry was used to create groups. Information collected from the interview included company name, job title, what the company made or did, job duties, month/year of starting job, and month/year of ending job. A team of two occupational epidemiologists and two industrial hygienists used this information to complete occupational coding using the 2000 Standard Occupational Classification System for occupation and the 1997 North American Industry Classification System for industry. Up to six jobs held during pregnancy and the 3 months before pregnancy were recorded and coded for each woman. However, only jobs held during the critical period of fetal development, defined as 1 month prior to pregnancy through the end of the third pregnancy month, were included in these analyses. Similar occupation and industry codes were grouped into a manageable number of categories. Table 2 shows the distribution of the 24 final occupational groups among cases and controls. Women who held more than one job during the critical period were included in more than one occupational group, if appropriate.

Some of the occupational groups were broad and encompassed several subgroups, which are also listed in table 2. For example, the public servants group was made up of police, firefighters and military personnel. To restrict the analyses to a manageable number, only the major occupational groups (n = 24) were used. Any job held during the critical period of fetal development was considered for exposure. Mothers who did not work during the critical period were excluded. Only month and year were recorded for job start date and job end date. For calculation of the critical period, a day of the month was assigned. To be as inclusive as possible and be consistent with other NBDPS exposure assessments (ie, medications), start day was assigned as the first day of the month and end day was assigned as the last day of the month. Jobs considered to be in the critical period fall into one of four start/end date scenarios: (1) started before the critical period and ended after the critical period; (2) started before the critical period and ended during the critical period; (3) started during the critical period and ended during the critical period; and (4) started during the critical period and ended after the critical period. In the case of missing dates, as long as either the start or end date was within the critical period the job was included. If both start and end dates were missing, the job was excluded.

Statistical analysis and potential confounders

A standard set of a priori covariates was determined for use in all logistic regression models based on literature review. This set included: study centre, maternal age at delivery, maternal pre-pregnancy body mass index, maternal race/ethnicity, maternal education, parity, folic acid use, maternal smoking and maternal alcohol use. The latter three variables were specific to the critical period. Logistic regression models were run, adjusted for potential confounders, to determine the association between each of the nine birth defect systems and each of the 24 occupational groups as well as each of the 45 specific birth defects and each of the 24 occupational groups. In each analysis, all other employed women were used as the exposure reference category.

In addition to standard logistic regression, we used a Bayesian approach to logistic regression developed by Greenland20 21 for the analysis of the 45 specific defects. This method is useful in situations such as this where multiple comparisons are made and also when there are small numbers. The Bayesian approach requires that the researcher have a “prior” belief about the effect of the exposure and covariates on the outcome before the actual analysis is performed. Prior values were set to 1, assuming no effect on the outcome, for all exposures and covariates except for maternal age greater than 35, smoking and drinking status, which had priors of 2. The prior of 2 was set for covariates known in the literature to have positive associations with the outcome. Since a variance must be specified for all priors, a conservative approach was taken and a wide variance of 1.125 was chosen in order to slightly minimise the effect of the priors on the data. In the presence of small numbers the prior distribution is not always normal, so a rescale factor of 10 was incorporated.21 A dataset was created with the prior specifications and imported into the actual data to make an augmented dataset. In addition, an indicator variable was created in the augmented dataset to indicate whether a record was from the prior data or the actual data. The Bayesian analysis was conducted by running logistic regression with the augmented dataset.

Results

The overall response rate for the NBDPS was 65.6% (69% for cases and 65% for controls). Of those who responded, approximately 72% of case mothers and 72% of control mothers were employed in a wide variety of occupations and industries. Of those mothers employed during the critical period, most (88.9%) held only one job during the critical period, while 10.1% had two jobs, 0.9% had three jobs, 0.1% had four jobs, and less than 0.1% had five or six jobs. After condensing these various occupations into meaningful occupational groups, we found that the largest group, for both cases and controls, was administrative support workers (over 20%). Other large groups making up roughly 12–13% of both cases and controls were sales workers, food industry workers (mostly food service workers) and healthcare workers (mostly nurses). Ten of the 24 groups were extremely small, each containing approximately 1% or less of the total cases and controls and included artists, public servants, landscapers, construction workers, metal workers/welders, dry cleaners, chemical/semiconductor workers, textile/paper workers, electronic equipment operators and manufacturing/transportation workers.

Table 3 shows the demographic characteristics of cases (n = 8977) and controls (n = 3383). Case mothers were more likely to be over 35, obese, nulliparous, have pre-pregnancy diabetes, and smoke, while control mothers were more likely to be college educated.

Table 4 shows adjusted odds ratios from the standard logistic regression for the risk of birth defects (grouped into nine systems) among 24 occupational groups. Not all combinations could be analysed due to small sample size. Janitors and cleaners had elevated risks for four birth defect systems (amniotic band, central nervous system, ear and eye). Teachers showed reduced risks for two categories of birth defects (central nervous system and heart).

Table 5 is a summary of the standard logistic regression analysis and the corresponding Bayesian results of the 24 occupational groups and the 45 specific birth defects included in the study. Only results that were statistically significant (p<0.05) in at least one of the analyses are presented because of the large number of comparisons made.

Several occupational groups showed a risk for multiple defects, including electronic equipment operators, janitors and scientists, among others. For instance, janitors had significant increased risks not only for amniotic bands and ear defects (anotia/microtia) as previously mentioned, but also for five other birth defects including two eye defects, one musculoskeletal defect, one gastrointestinal defect and clefts. All but amniotic bands and bladder exstrophy (musculoskeletal system) remained significant in the Bayesian analysis. Scientists had an increased risk for five specific defects including two heart defects, two musculoskeletal defects and one gastrointestinal defect. All but sacral agenesis (musculoskeletal system) remained significant in the Bayesian analysis. Conversely, teachers had significantly reduced risks for four specific defects, all of which remained significant in the Bayesian analysis. Interestingly, although healthcare workers had reduced risks for three specific defects, they also had an increased risk for two specific defects, all of which remained significant in the Bayesian analysis. Other occupational groups with a mix of positive and negative results in the standard analysis were administrative support workers, food workers and other office workers. Only food workers and other office workers showed this mix in the Bayesian analysis.

A total of 17 (16 positive and one negative) results were no longer significant in the Bayesian analysis. This was not unexpected as the Bayesian analysis considers multiple comparisons and is stricter than the standard method of logistic regression that was used for the presented analysis. Most of the results that were significant in the standard analysis and not significant in the Bayesian analysis were those performed on groups with very small numbers of exposed cases (13 of 17 comparisons had three or fewer exposed cases); the remaining four comparisons were of borderline significance in the standard analysis.

Discussion

Using a large population-based case–control study, we were able to analyse a spectrum of occupations and 45 specific birth defects for hypothesis generating purposes. Using standard logistic regression, these analyses resulted in 42 (26 for Bayesian) significantly elevated risks of birth defects in the offspring of working mothers. The job positively associated with the largest number of birth defects was that of janitors/cleaners. Cleaners and housekeepers have previously been positively associated with neural tube defects (NTDs), spina bifida and oral clefts in three separate studies.3 8 23 Our study also found a positive association between cleaners and oral clefts. In addition, although the association was not significant, we also observed an elevated risk for NTDs and more specifically, spina bifida (data not shown). At the system level, central nervous system defects (which include NTDs and spina bifida) were positively associated with cleaners.

In addition to the positive associations discussed above, we also found several reduced risks. Using standard logistic regression, we found 12 (11 for Bayesian) significantly reduced risks of birth defects among the offspring of working women. The job negatively associated with the largest number of birth defects was teaching. Teachers have not been commonly studied in occupational studies of birth defects, likely due to their lower perceived risk. However, two studies of clefts looked at teachers as a group; one study found a reduced effect similar to our cleft results23 and one study found non-significant results very close to the null for various cleft subgroups (OR 1.1–1.3).4

Due to its large size and breadth of interview-based information, including data on specific potential confounders, the NBDPS is a good study to use to conduct a spectrum analysis. This study has included a large number of women in order to study 45 specific birth defects. Most of the women in the NBDPS were workers in a variety of occupations. With the help of industrial hygienists we were able to create meaningful occupational groupings with sufficient size to conduct analyses. These groupings also tend to group workers based on other attributes such as education and socio-economic status. Extensive information was collected on potential confounders, such as socio-demographic factors and maternal behaviours during pregnancy. This overall assessment of the association between various occupations and all birth defects is useful as a hypothesis generating tool. A future paper will look at teachers, janitors and healthcare workers in more detail by breaking these broad groups down into more specific occupations and considering hours worked. Also, this spectrum analysis provides guidance on future analyses of specific defects and specific maternal occupations that can be performed with NBDPS data. Although this analysis was adjusted for several potential confounders, future studies looking at specific outcomes and exposures should include other confounders applicable to those analyses, such as medication use, diet and home exposures.

This analysis used job title to create occupational groups. The analysis did not consider hours/days worked or possible chemical exposures. A separate project being conducted within the NBDPS is working on a more detailed exposure assessment for pesticides, solvents and metals using this information.

We carried out a total of 1080 different comparisons for this analysis (24 occupational groups×45 specific defects). With so many comparisons being made, one would expect a number (up to 5%) of significant results to occur simply by chance. In addition, some estimates provided in table 5 are unstable due to the small number of exposed cases, such as those for artists, farm workers, metal workers/welders, landscapers, public servants and textile/paper workers. These point estimates tended to be larger with standard logistic regression. The Bayesian method used is useful when dealing with both multiple comparisons and small numbers. The Bayesian method can pull or shrink coefficient estimates toward prior values or the null, thus limiting the occurrence of type 1 errors. Stable coefficients experience very little pull, while unstable estimates are more strongly pulled toward prior values. The degree of adjustment is controlled by specifying the prior variance to achieve more precise and reasonable confidence intervals than in traditional methods. By using conservative prior estimates in our Bayesian analysis, we found that the results of analyses with small numbers of exposed cases or of borderline significance were the results that became non-significant when applying the Bayesian analysis. Therefore, the significant Bayesian results seem to be stable and likely have less type 1 error, as the method suggests. Thus, these results seem less likely to be due to chance alone. Although the results using standard logistic regression may be useful for hypothesis generating, which is the main purpose of our analysis, the results from the Bayesian analysis are likely more precise. Given the novelty of the Bayesian approach, we felt that presentation of both standard logistic regression and Bayesian logistic regression results was warranted.

Because this was a spectrum paper, the analysis was less detailed than would have been ideal. For example, in order to have a more manageable number of analyses we chose to use 24 occupational groups, each analysed with our 45 specific defects. Creating just 24 occupational groups could have introduced some misclassification bias. However, we began our analysis with 59 occupational groups, and those results did not differ greatly from the results presented here. The major advantage of using the 24 groups for this spectrum paper is that the results presented are more easily described and understood. Similar occupational spectrum studies have used up to 73 different occupational groups but were only looking at one outcome.15 Other occupational studies have grouped all birth defects together.1 11 For our study, industrial hygienists were consulted to develop the 24 groups. Some groups are more homogenous than others and were unchanged from the more expanded analysis (eg, dry cleaners and sales workers). However, other groups may be more mixed (eg, manufacturing/transportation workers and construction workers). These groups tended to aggregate job titles which very few women held. Therefore, our only chance to examine them, even in such a large study, was to group them. Groups that were large and where strong or multiple effects were seen (such as healthcare workers, janitors and teachers) will be followed up in subsequent in-depth analyses. Finally, although we presented results for nine physiological groups that were created by clinicians expert in birth defects, there may be some heterogeneity in these groups. Thus, the results for the 45 specific birth defects may be more useful.

Some misclassification may result from our assignments of job start and end day when calculating the critical period. We felt that choosing the first day of the start month and the last day of the end month would be most inclusive. Furthermore, this strategy was consistent with the method used for medication use in the NBDPS study. Lastly, the same strategy was applied to both cases and controls. Therefore, any potential misclassification should be non-differential with respect to the outcomes.

Because the overall response rate for the study was 66%, selection bias is a potential concern. As information on non-responders is not available, it is not possible to compare responders and non-responders to assess this potential bias related to response rate. Also, in any retrospective study, recall bias is a concern. In this study, women who frequently changed jobs would be of particular concern. Women were asked between 6 and 23 months after the birth of their child to report on jobs held during their entire pregnancy and just before pregnancy. A pregnancy calendar was used to aid recall. Approximately 1% of our population (1.1% of cases and 0.9% of controls) reported three or more jobs during the critical period. Therefore, recall bias may be less likely. In addition, this analysis is based on job title and description rather than questions about specific chemical or physical exposures. A general question about job title is less likely to be affected by the mother’s knowledge of her child’s case/control status than specific questions about chemical or physical exposures. Finally, although this analysis adjusted for education, race and ethnicity among other covariates, this may not adequately control for the complex mixture of factors related to socio-economic status.

Conclusion

This paper presents results from analyses on a spectrum of occupations and birth defects for hypothesis generating purposes and presents information useful for guiding future investigations of occupational exposures and birth defects. Several occupations have been found to be positively associated with one or more specific birth defects in a sample of working women, including those of janitors/cleaners, scientists and electronic equipment operators. In addition, several other occupations were found to be negatively associated with one or more birth defects including those of teachers and healthcare workers.