Methods

Rationale and Hypotheses

Drinking water at Camp Lejeune was contaminated with VOCs until the mid-1980s. Because a sizable population of young, married women were supplied with this water in their homes, concern has been raised about the potential adverse effects of VOCs on pregnancy outcomes. An association between VOC-contaminated drinking water and adverse late pregnancy outcomes is plausible, but further investigation is needed. The health significance of VOC contamination in drinking water is of particular interest at Camp Lejeune because some of the site's VOC plumes are still unremediated and might contaminate other supply wells in the future. Although Camp Lejeune supply wells are monitored on an annual or semi-annual basis as a protective public health measure, this action may not be protect all segments of the population. Fetal development can be disrupted by toxic chemical exposures in less time than the 6-month intervals at which the wells are being monitored. Hence, additional knowledge of the risks presented by these plumes might be useful in managing the potential risks from the unremediated plumes.

This cohort study examined the relationship between VOC exposure and fetal growth retardation (measured as SGA and decreased MBW) and preterm delivery in three groups with different exposures to contaminated drinking water and in an unexposed comparison population. The primary objectives of this investigation were to evaluate the following three hypotheses:

1. Residence in each of the exposed housing areas was associated with fetal growth retardation, measured as SGA and decreased MBW.
2. Residence in each of the exposed housing areas was associated with preterm birth.
3. Residence in each of the exposed housing areas was associated with late fetal death. This third hypothesis could not be tested because of poor data quality (see Data Quality).

The primary sources of data were birth and fetal death certificates at the North Carolina Vital Statistics Office. For each of the three categories of exposed births defined later, MBW, the prevalence of SGA and preterm births, and the ratio of fetal deaths per singleton live births were compared with these outcomes in unexposed births. In addition, the effects of timing and duration of exposure were examined by linking data from family base housing with birth and fetal death certificate data. The rationale and methods employed to complete three secondary objectives previously described (see Objectives) are included in Appendices A and B of this document.

Study Population

The study population consisted of all singleton live-born and stillborn infants delivered at 20 weeks of gestation during 1968-1985 to families residing in base family housing units at Camp Lejeune. Residents of Camp Geiger and Knox Trailer Parks were excluded because of incomplete housing records and ambiguity regarding their drinking water source. Approximately one-third of the women who sought prenatal care at the Navy Regional Medical Center at Camp Lejeune moved or were transferred before they delivered (CDR J. McGinnis, Camp Lejeune, personal communication). These women could not be identified; however, their exclusion from this analysis probably did not introduce selection bias because rates of mobility were not expected to be associated with being exposed to VOCs.

North Carolina Vital Statistics Data Files

Since 1968, the state of North Carolina has maintained computerized databases of live births and fetal deaths occurring at >20 weeks. During 1968-1985, the state used two versions each of the birth and fetal death certificates and three versions of the database file format.

The smallest recognizable unit for which birth and fetal death records could be selected from the data files was county of residence; information about the mother's ZIP code and residence on base was not available for the study period. Eligible births and fetal deaths, therefore, were identified by searching all records for Onslow County residents. For live births during 1975-1985, computerized records were searched for eligible street addresses. For the years 1968-1974, the mother's street address and city of residence were not included in computerized birth certificate files. In addition, other important information (e.g., exact birth weight in pounds and ounces) was included on the hard-copy certificate but not in the computerized file. Therefore, hard copies of records for Onslow County residents were searched by hand for addresses with eligible street names. Relevant information from records containing eligible street names was then entered into a computer file.

Some housing units on eligible streets were not eligible for inclusion in the study. For example, units 2000 through 2999 Onslow Drive were base family housing units, but housing units with numbers of 3000 or higher were privately owned. Therefore, after addresses for all eligible street names were computerized, a second electronic search was completed to remove addresses with an eligible street name but an ineligible housing unit number.

The mother's street address was not available in databases containing fetal death records. For fetal deaths occurring from 1968 through 1977, the mother's address was computerized from hard-copy records, and eligible records were identified in a process similar to that described for live births. Fetal death certificates for 1980-1985 were destroyed in accordance with North Carolina state law. Therefore, the housing record database was searched for matches with the father's name from the fetal death certificate, and eligibility was based on the housing records.

Base Family Housing Records

For each base family housing unit, Camp Lejeune maintains the following records on an index card: (1) the first and last names and middle initial of the active duty person to whom the housing unit is assigned; (2) the rank (e.g., seaman first class or captain) of the person to whom the housing unit is assigned; (3) the first and last dates of occupancy by the active duty person to whom the unit is assigned. For purposes of this study, approximately 88,000 names and addresses were identified. Birth certificate data were then matched to housing record data on the basis of the address listed on the birth certificate and the name of the father. For a match to be considered acceptable, the pregnancy interval had to have occurred during the period between the first and last dates of occupancy. Because it was possible that the father's name was spelled slightly differently in the birth and housing records, when no match was found for a particular birth certificate, a manual search was conducted comparing the father's last name from the birth certificate to alphabetized lists of names from the housing records. When the father's name did not match either by a computerized or a manual search, then a match was attempted on the mother's name. For the few parents who were dependents, matches were made by address and the last name of the father or the maiden name of the mother, and a notation was made that the parents were dependents in another household.

If a birth or fetal death record contained no address information, the housing record database was searched for matches with the father's name from the birth or fetal death record. A match was considered acceptable if the father's name on the housing record and the birth or fetal death record were the same, if the dates of occupancy for the housing record coincided with the time of the live birth or fetal death, and if no more than two other persons with the same first, last, and middle initial were identified in the housing record. Different persons with the same name were distinguished by comparing multiple housing records for dates of occupancy. In general, the dates of occupancy were contiguous for individuals who had more than one housing record. When the dates of occupancy were overlapping or separated by long time periods, it was assumed that different individuals with the same name had been identified. Although occurrences of several people with exactly the same first and last names and middle initials in the housing records were rare, excluding such person from the analysis minimized the possibility that fetal death records were randomly matched with a housing record coincidentally containing the same name as the death record.

Definition of Study Outcomes

The outcomes that were studied are MBW, SGA, and preterm birth. All data regarding these outcomes were obtained from North Carolina birth records.

Live births occurring at less than 37 weeks of gestation were defined as premature. Gestational age was based on the date of the last menstrual period. For observations with a valid month and year of last menstrual period, but a missing day, the day was interpolated to the value of 15. Last menstrual periods with a valid month and day but no year were assigned to the year that would yield the most biologically plausible gestational age. Last menstrual periods in which the month of the last period was not reported were excluded. The effects of missing gestational age were evaluated by comparing birth weight distributions and demographic information between births with missing and nonmissing gestational age data.

Birth weight in pounds and ounces was obtained from the birth certificate and converted to weight in grams. An SGA birth was defined as a singleton live-born infant weighing less than the 10th percentile based on published sex-specific growth curves. The standard published by Williams et al.1 for whites in the state of California was selected because it (1) was derived from birth certificate, (2) was based on a large group of live births occurring during the midyears of the study period, (3) was published in a reputable journal and in an easily read format, and (4) categorized approximately 10% of unexposed births as SGA. No other published standard met these criteria.

Although Williams published a standard for whites only, this standard was applied to births among women of all races. This decision was made because published race-specific standards were not readily available during the time period under investigation. Analyses were stratified on this variable to ensure that race did not confound the association. To provide further reassurance that use of the Williams data set for births to black women did not bias the results, some of the data were reanalyzed using an unpublished race-specific standard developed by the state of New Jersey for white and black births that occurred during 1985-1988. Because there was essentially no difference in the associations observed using the New Jersey race-specific standard and the associations observed using the Williams standard, only results based on the Williams standard are presented.

An attempt was also made to study late fetal deaths. A late fetal death was defined as any fetal death occurring at 20 or more weeks of gestation for which a North Carolina fetal death certificate was filed. Fetal deaths were not studied because of incomplete data (see Data Quality).

Definition of Exposure

Infants identified from birth and fetal death certificates were divided into three distinct exposed groups, which are summarized in Table 2. These groups are referred to as (1) PCE exposed; (2) long-term TCE exposed; and (3) short-term TCE exposed. The mothers of PCE-exposed children resided at Tarawa Terrace for at least 1 week before birth occurred. Mothers of long-term TCE-exposed infants resided at Hospital Point during 1968-1985 for at least 1 week before the children were born. The housing units that were supplied with TCE on a short-term basis were Berkeley Manor, Midway Park, Paradise Point, and Watkins Village. Requirements for inclusion of births in the short-term TCE-exposed group were (1) the mother resided in Berkeley Manor, Midway Park, Paradise Point, or Watkins Village at the time of birth and for a minimum of 1 week during January 27 through February 7, 1985; (2) the mother's last menstrual period was on or before January 31, 1985; and (3) the birth occurred after February 2, 1985. These dates were selected to ensure a minimum of 1 week of exposure to TCE during pregnancy. Births to the remaining residents of base family housing were considered unexposed. The unexposed group consisted of all residents of the Marine Corps Air Station, Rifle Range, and Courthouse Bay housing areas, as well as residents of Berkeley Manor, Midway Park, Paradise Point, and Watkins Village who were not in the short-term TCE-exposed group.

Several groups of infants whose exposure to PCE or to TCE on a short-term basis was unknown were excluded from the study. These included infants whose fathers had short-term exposure during spermatogenesis (i.e., based on the mother's last menstrual period); infants whose mothers resided in any of the exposed housing areas for less than 1 week while pregnant; and infants whose parents first moved into Berkeley Manor, Midway Park, Paradise Point, or Watkins Village during February 8-21, 1985. Although all of the contaminated wells were closed on or before February 7, 1985, the first day that water samples were entirely free of contamination was February 21, 1985. Therefore, residents who were not exposed during a known period of contamination, but who lived in the affected housing units during this ambiguous exposure period, were excluded.

Membership in each of these exposure groups was based on residence in a housing area known to have received contaminated water. Within each housing area receiving contaminated water, every housing unit probably received similar concentrations of contaminants; however, information regarding tap water concentrations in each housing unit was unavailable. Information about behavioral risk factors that would have affected exposure levels, such as showering and drinking water patterns, also was unavailable.

Analyses

Separate analyses were conducted for each of the exposure-outcome combinations defined previously. For dichotomous outcomes (SGA and preterm delivery), odds ratios (ORs) were computed relating exposure to outcome using the SAS software package2. MBW was also computed using the SAS software package. The two main criteria for identifying elevations in the dichotomous outcomes and decreases in MBW were the size and the plausibility of the association. The degree of variability in the data was examined by computing the 90% confidence interval (CI). For most analyses, 90% CIs were computed using the logit estimators produced by the SAS statistical package 34. However, when the number of either exposed or unexposed cases was fewer than 10, ORs and CIs were recomputed using the exact method. Exact computations were conducted using the StatXact software package. The mid-P approximation for CIs was reported5.

The following characteristics from the birth certificates were evaluated for their potential as confounders or effect modifiers: gestational age, maternal race, sex of infant, year of birth, mother's age, mother's educational level, father's age, father's educational level, parity, adequacy of prenatal care, maternal history of fetal deaths, and mother's marital status. In addition, military ranks obtained from housing records were standardized into nine enlisted pay grades (E1-E9), four warrant officer pay grades (WO1-WO4), and six officer pay grades (O1-O6). Pay grade is an accurate measure of income for the active duty member of the household, although information was unavailable about the occupation and income of the parent not listed on the housing card.

Each possible confounder or effect modifier was evaluated separately by using stratified analysis. Potential confounders were those variables that met all of the following conditions: (1) they were distributed differently in the exposed and unexposed groups, (2) they were risk factors for or protective factors against the study outcome, and (3) the association between exposure and outcome variables that was stratified on the potential confounder differed from the association between exposure and outcome variables that was not stratified on the potential confounder. The variables, maternal age, maternal and paternal educational levels, parity, year of birth, gestational age, and military pay grade, were collapsed to a minimal number of categories for the purposes of stratified analyses and, in some cases, for multiple regression (linear and logistic) analyses. Cut-off points for these variables were selected on the basis of both their distribution and their relevant social or biologic meaning. Histograms, scatter plots, and plots of means at different levels of each factor were employed to examine the distributions of these variables.

The potential for effect modification consisted of a simple inspection of ORs, or means in different strata, and a Breslow-Day test for homogeneity6. Ideally, effect modifiers would be identified a priori7, but current information is insufficient to determine which risk factors might act as effect modifiers. In the absence of a well-developed literature, the potential biologic and sociological relevance of each potential effect modifier were considered.

Each variable that was either an effect modifier or a confounder (i.e., based on the results of stratified analysis) was retained in analyses for multiple logistic regression or linear regression modeling, and was eliminated in a backwards fashion. For SGA and preterm births, potential confounders were eliminated from a model if their removal did not change the ORs relating exposure and outcome (or the ORs for different exposure-covariate combinations if interaction terms were used) by more than 10%. For MBW, potential confounders were eliminated from a linear regression model if their removal did not change the effect estimate by either a minimum of 20 g or 10% of the effect estimate, whichever was greater. For analyses of dichotomous outcomes, variables that indicated effect modification (also known as interaction terms), were included in logistic regression models if they were biologically plausible, described heterogeneous groups in which the ORs differed by more than 25%, and had P values less than 0.20. A less stringent P value for effect modifiers was used because of the low statistical power available to detect them8. The choice of a 25% change was arbitrary, but provided an effective decision rule for screening potential effect modifiers. For MBW, covariates for which at least one stratum-specific estimate showed a mean difference between PCE-exposed and PCE-unexposed births of -50 g or less were examined more closely for effect modification. Statistical significance was not helpful in identifying effect modifiers for MBW because the statistical power was so great overall that almost every parameter estimate was statistically significant.

Years of Exposure

Although exposure to VOCs probably occurred throughout the study period in the PCE- and long-term TCE-exposed groups, exposures before 1982 could not be documented. Because the PCE-exposed group was large enough, separate analyses were conducted for births that occurred in this group during 1982-1985. The number of births that occurred during 1982-1985 among persons in the long-term TCE-exposed group was too small to complete separate analyses in this group.

Timing and Duration of Exposure

The third trimester of pregnancy is usually regarded as the most important for fetal growth and toxicity resulting in delayed fetal growth9. However, Dejmek et al. recently observed an association between SGA and air pollutants that was greatest when exposure occurred during the second and third months of pregnancy10. A cumulative effect of exposure might also be possible11, and the influence of the exposure of timing on pregnancy outcome has not been fully determined12. In addition, the population at Camp Lejeune has always been unusually mobile. Approximately one-third of women receiving prenatal care at the Navy Regional Medical Center move to another base or into civilian life between the first prenatal care visit and delivery. If exposure is necessary in the early part of the pregnancy or during spermatogenesis to observe an outcome, then use of maternal residence at time of birth can, by including persons who were not exposed at these critical times, reduce or obscure an important association.

To address these concerns, the dates of occupancy for each household were examined to determine whether and when each family moved during the pregnancy. This information was used to explore the influence of the timing and duration of exposure on each study outcome. Within the PCE-exposed and long-term TCE-exposed groups, length of residence in the housing unit listed on the birth certificate was used as a surrogate for length of exposure. Based on discussions with Camp Lejeune personnel, it was determined that most women who had given birth while living at Camp Lejeune had remained in the same housing unit until after the delivery. Therefore, it was assumed that each family resided in only one base housing unit during the pregnancy, an assumption that was evaluated and is presented in a later section. Therefore, except for births occurring after the exposure ceased, length of exposure indicates the number of consecutive weeks before delivery that the mother lived in exposed housing. For example, a woman residing in exposed housing for 10 weeks during pregnancy was exposed during the last 10 weeks of the pregnancy and not the first 10 weeks. Among births in 1985, the year that the contamination ceased, timing of exposure was more heterogeneous. To maintain consistency regarding the meaning of the duration of exposure variable, births that occurred after the contamination ended were excluded from analyses of duration of exposure.

Births were categorized in the following groups depending on length of exposure: births to mothers exposed for 1-3 weeks, for 4-10 weeks, for 11-20 weeks, for 21-45 weeks, before conception and throughout pregnancy, and for 1 or more years before conception and throughout pregnancy. A number of duration-response relationships were considered to be biologically plausible. Because weight gain occurs most rapidly at the end of pregnancy, it was considered plausible that the last weeks of pregnancy would be the most important; in this case, birth outcome would not differ by exposure category. It was also considered to be biologically plausible that the effect measures would increase with duration during the last 20 weeks of pregnancy when most weight gain occurs, or that effects would increase with duration throughout pregnancy. Other biologically plausible scenarios were that an association would be observed only in infants born to mothers who were exposed during the entire first trimester or in infants of mothers who were exposed before conception and throughout pregnancy. Finally, because of the possibility of selective survival, it was considered plausible that no effect, or a very limited effect, would be observed among infants born to mothers who were exposed during the first 12 weeks of pregnancy when spontaneous abortion rates are highest, and then a duration-response relationship would be observed thereafter. There were also scenarios that were considered to be implausible. For example, it would be implausible to observe less of an effect among the women who were exposed for the longest time periods relative to women exposed for shorter periods, except where the pattern was consistent with selective survival. Because length of residence during pregnancy was a function of length of gestation, the percentage of the pregnancy during which exposure occurred and the gestational age at first entry into the exposed area were also examined. To examine the influence of timing of exposure in the short-term TCE-exposed group, both week of gestation at time of exposure and weeks elapsed between exposure and birth were examined.

Data Quality

A total of 12,493 live births and 83 fetal deaths that met the study selection criteria were identified. Figure 1 shows fetal death ratios for Camp Lejeune by race and year of birth compared with overall fetal death ratios for the United States13. Because of the small numbers of fetal deaths within some racial groups each year, 5-year averages were computed for each 5-year period starting with 1968 through 1972 (midpoint was 1970) and ending with 1981 through 1985. As illustrated in Figure 1, the fetal death ratio for whites in Camp Lejeune was about half of that expected for most of the study period; the fetal death ratio for nonwhites was close to that expected until 1972, but was approximately five times less than expected by 1982. In addition, less than half of the fetal death certificates listed a cause of fetal death. Given the low likelihood that these rates were accurate, and the expectation that all causes of fetal death would not be uniformly associated with VOC-exposure, no further analyses were conducted for the fetal death outcome.

Table 7 contains the distribution of live births in each exposure group. Forty-four births were deleted from analyses because the mothers were exposed for less than 1 week during pregnancy or only during spermatogenesis. Of the remaining 12,449 births, 479 (3.9%) were eliminated from all analyses because of poor data quality. The total numbers of observations included in or excluded from each analysis are presented by exposure group in Table 7. Some observations were eliminated for more than one reason, such as an infant who reportedly weighed less than 350 g and who had no data on the mother's last menstrual period. To provide a better sense of overall data quality for each critical field, observations that were eliminated are listed in each category into which they fell. Therefore, the numbers in Table 7 do not add up to 100%.

Live births at less than 22 weeks of gestation were eliminated only from the SGA analyses. These births were not eliminated because of poor data quality, but because the population standard selected to compute SGA began at 22 weeks. The number of births eliminated for this reason was less than 0.1%.

In addition to observations that were entirely excluded from MBW and SGA analyses, there were a number of observations with questionable values for gestational age. It is well recognized that, in most populations, there are a disproportionate number of infants who are classified as very preterm that are heavier than would be expected for their gestational ages; most of these heavy, very preterm infants are actually infants born at later gestational ages than indicated in the birth certificates14. To determine if the gestational ages for infants classified as very preterm were more commonly misclassified, the distribution of birth weights among these infants was compared with the standard birth weight distribution at each gestational age (75). Of live births at less than 28 weeks of gestation, 17% were above the 90th percentile reported for the standard population, although only 10% would have been expected. These data values were marked as unlikely but were not excluded from the MBW and SGA analyses because (1) although at a population level it was possible to determine that most of these values were misclassified, it was not possible to distinguish which individual observations were misclassified and which were correct but outlying, and (2) these observations represented a large proportion of early preterm births, but only 1% of all live births. Another set of questionable gestational ages were those that were estimated because the day of the last menstrual period was missing from the birth certificate (i.e., approximately 2.7% of the data). The final models for SGA and MBW were analyzed by both including observations with unlikely or interpolated gestational ages and excluding them. Unless explicitly discussed, reanalyzes without these data points had negligible impact on study results.

Preterm live-born infants that weighed more than the 90th percentile for birth weight at 36 weeks of gestation were excluded from the preterm birth analyses. Inclusion of these heavy, preterm infants substantially affected the total number of preterm infants in each exposure category.