A Prospective Cohort Study Examining the Associations of Maternal Arsenic Exposure With Fetal Loss and Neonatal Mortality

Abstract

Arsenic crosses the placenta, possibly increasing the risk of adverse reproductive outcomes. We aimed to examine the association between maternal arsenic exposure and fetal/neonatal survival using data from a prospective cohort study of 1,616 maternal-infant pairs recruited at a gestational age of ≤16 weeks in Bangladesh (2008–2011). Arsenic concentration in maternal drinking water was measured at enrollment. Extended Cox regression (both time-dependent coefficients and step functions) was used to estimate the time-varying association between maternal arsenic exposure and fetal/neonatal death (all mortality between enrollment and 1 month after birth). In a sensitivity analysis, we assessed gestational arsenic exposure using maternal urine samples taken at enrollment. We observed 203 fetal losses and 20 neonatal deaths. Higher arsenic exposure was associated with a slightly decreased mortality rate up to the middle of the second trimester, and then the mortality rate switched directions around 20 weeks’ gestation. In the step function model, the hazard ratios for combined mortality (fetal loss and neonatal death) per unit increase in the natural log of drinking water arsenic concentration (μg/L) ranged from 1.35 (95% CI: 1.08, 1.69) in weeks 25–28 to 0.81 (95% CI: 0.65, 1.02) in weeks 9–12. This nonlinear association suggests that arsenic may exert survival pressure on developing fetuses, potentially contributing to survival bias, and may also indicate that arsenic toxicity differs by fetal developmental stage.

Naturally occurring arsenic in drinking water is a worldwide health hazard. In Bangladesh, groundwater has become the primary source of inorganic arsenic exposure following the promotion and adoption of groundwater as a microbiologically safer alternative to surface water (1, 2). National surveys conducted in 2012–2013 indicated that one-eighth of drinking water sources had arsenic concentrations above the Bangladeshi national guidelines of 50 μg/L and one-quarter exceeded the World Health Organization guidelines (3) of 10 μg/L (4). These results suggest that approximately 57 million people in Bangladesh are exposed to arsenic concentrations above World Health Organization standards (5). While not all tube wells are affected, elevated arsenic exposure is well documented. Data from surveys in the arsenic-affected regions of Bangladesh show that 95% of people have elevated levels of arsenic metabolites in their urine, indicating a substantial exposure burden (6).

Fetal development is a time of intense cellular changes and proliferation. Arsenic crosses the placenta, exposing the fetus to arsenic from the mother (7–9). Given the sequential nature of fetal development, the possible impacts of gestational arsenic exposure on fetal health may compound over time. For example, in a recent meta-analysis, Quansah et al. (10) estimated a greater risk of miscarriage (odds ratio (OR) = 1.98) than of stillbirth (OR = 1.77), neonatal mortality (OR = 1.51), or infant mortality (OR = 1.35) when pregnant women were exposed to groundwater arsenic levels above the Bangladesh regulatory cutoff (≥50 μg/L). However, the studies included in this meta-analysis primarily included only women with high exposures and made comparisons for persons above and below the regulatory cutoff levels (e.g., (≥50 μg/L vs. <50 μg/L) (10), hindering causal conclusions because of a lack of exploration of the arsenic-mortality dose response. For example, if gestational arsenic exposure contributes to early-life mortality, it would induce a selection bias in pregnancies, greatly biasing studies that explore the possible childhood effects of arsenic exposure by enrolling children at birth. Furthermore, additional prospective research is needed across a wider range of arsenic exposures, especially at lower doses that extend into infancy (11). Therefore, we examined the association between arsenic exposure during pregnancy and fetal loss and neonatal death using survival analysis in a prospective pregnancy cohort study in Bangladesh. Specifically, we estimated the time-varying association of arsenic exposure with the probabilities of fetal loss and neonatal death in a population with primarily compliant (below regulatory cutoffs) arsenic exposure. We hypothesized that each unit increase in arsenic exposure during pregnancy would be associated with an increased hazard of death earlier in pregnancy after adjustment for confounders.

METHODS

Data and study population

Between January 2008 and June 2011, a total of 1,616 pregnant women were enrolled in a maternal-child cohort study in Sirajdikhan and Pabna Sadar upazilas, Bangladesh. This cohort study was intended to study the health impact of chronic arsenic exposure on pregnant women, and details have been published previously (12, 13). To be eligible for this study, women needed to be at least 18 years of age and have an ultrasound-confirmed singleton pregnancy with a gestational age of ≤16 weeks. Women also had to report having used the same drinking water source for at least 6 months prior to enrollment and to be planning to live in the same location throughout the pregnancy.

The Dhaka Community Hospital Trust (DCHT) provided arsenic awareness training to these communities through rural clinics. Study participants were told the arsenic concentration in their drinking water, provided with counseling on how they could reduce their exposure to arsenic, and given safe drinking water options. Prenatal vitamins and transportation to Dhaka Community Hospital for pregnancy-related emergencies were provided to participants free of charge. After the initial enrollment visit, participants were visited every month in their homes by clinic staff. At that time, prenatal vitamin supplies were restocked and participants underwent a brief examination. Additionally, participants came to the clinic for medical follow-up at approximately 28 weeks of gestation and within 4 weeks after delivery. DCHT-trained midwives and/or clinic medical staff who resided in the recruitment areas attended home births; otherwise, deliveries occurred in DCHT-affiliated clinics and hospitals. Biological specimens were collected and trained examiners administered questionnaires at each of these visits.

Consent was obtained from each participant before the beginning of study activities. Consent documents were printed in Bangla and were read aloud to the participants by study staff to account for different literacy levels in these communities. The study protocol was approved by the human research committees/institutional review boards at DCHT (Dhaka, Bangladesh), Oregon State University (Corvallis, Oregon), and Harvard T.H. Chan School of Public Health (Boston, Massachusetts).

Arsenic exposure assessment

Our main exposure assessment used the concentration of arsenic in the mother’s primary source of drinking water collected at the time of enrollment (≤16 weeks’ gestation). Water was analyzed as described previously, following US Environmental Protection Agency Method 200.8 (14), by Environmental Laboratory Services, North Syracuse, New York (15). If the concentration of arsenic in a water sample was below the limit of detection (1 μg/L), it was assigned a value of 0.5 μg/L for analysis (6% of samples). Quality control procedures demonstrated that the average percentage of recovery of arsenic from PlasmaCAL multi-element Quality Control Standard 1 solution (SCP Science, Baie D’Urfé, Quebec, Canada) was 101% (range, 92%–110%).

As an additional sensitivity analysis, we used arsenic concentrations in maternal urine samples that were also collected at enrollment. Briefly, urine was frozen to −20°C and shipped to Taipei Medical University (Taipei, Taiwan), where it was analyzed as previously described (16) using high-performance liquid chromatography (Waters 501; Waters Associates, Milford, Massachusetts) and hydride-generated atomic absorption spectrometry (Flow Infection Analysis System 400AA 100; Perkin-Elmer, Waltham, Massachusetts) for quantification of arsenite, arsenate, monomethylarsonic acid, and dimethylarsinic acid. Levels of these 4 arsenic species were summed to calculate total urinary arsenic concentration (in μg/L), which was statistically analyzed as a continuous variable. Creatinine was included in urinary models to account for variation in urine concentration (17).

Fetal loss and neonatal death outcome

The main outcome was the number of weeks from enrollment (≤16 weeks’ gestational age) to fetal loss (a fetus that never showed signs of life outside the womb) or neonatal death (mortality up to 30 days after having been alive outside the womb). Gestational age at enrollment was measured via ultrasonography at the first visit and was calculated using either mean gestational sac diameter at 4–6 weeks or crown-rump length at 7–16 weeks. If a pregnancy was found to be nonviable, the woman was not enrolled in the study. Two ultrasound technicians were used for all participants, and all measurements were verified by an obstetrician at Dhaka Community Hospital.

Covariates

On the basis of previous literature, the following variables were considered as possible confounders in this analysis: self-reported maternal age (dichotomized into 18–34 years or ≥35 years, since exact birthdays are often unknown to rural residents of Bangladesh), monthly family income (in 2,000-taka increments, as reported by the mother’s financial provider), and self-reported maternal education (categorized as no school attendance, primary or secondary education, and higher secondary education or beyond). We also included gravidity (the number of previous pregnancies reported by the mother, regardless of the outcome of those pregnancies) and clinic location. Birth weight and gestational age are known to be associated with early childhood survival. However, it is thought that both are also affected by arsenic exposure (12, 13), making them intermediaries rather than potential confounders. Therefore, these variables were not included in this analysis. Finally, survival time since enrollment was the outcome of interest in this analysis, so gestational age was already accounted for.

Statistical analysis

While fetal loss and neonatal death are biologically distinct outcomes, arsenic does contribute to preterm birth. It is unclear whether the role of arsenic exposure in mortality differs meaningfully for a stillbirth that occurs at 35 weeks’ gestation and a child born at 35 weeks who then dies shortly after birth. Therefore, we developed both combined and parallel models for fetal loss and neonatal death. General descriptive statistics were calculated, and arsenic exposure data were natural log-transformed. We also examined the collinearity between covariates by means of condition indices and variance decomposition proportions (18, 19), using the “perturb” (20) and “car” (21) packages in R (R Foundation for Statistical Computing, Vienna, Austria). We decided a priori that conditional indices greater than 30 indicated the presence of collinearity and that variance decomposition proportions over 0.5 indicated which covariates were involved in the collinearity. Where these conditions were met, the variable with the highest variance decomposition proportion would be removed and the model refitted until no collinearity was indicated. We used extended Cox proportional hazards regression models to estimate the time-varying association between arsenic exposure and any mortality (fetal loss or neonatal death), arsenic exposure and fetal loss, and arsenic exposure and all-cause mortality from birth to 30 days postnatal. We explored the time-varying arsenic exposure association using 2 different analytical approaches: 1) continuous time-dependent coefficients and 2) step functions. The proportional hazards assumption of the time-constant variables was assessed graphically and using Schoenfeld residuals (22) via the survfit() and coxph() functions in the R “survival” package (23). Data analysis was conducted in R using the “survival” package (23).

Continuous time-dependent coefficients

In the models that examined continuous time-dependent coefficients, arsenic exposure was included a priori as a time-varying variable by including the interaction between maternal drinking water arsenic and time since enrollment (in weeks). This can be thought of as estimating the cumulative effect of arsenic exposure. The use of time-varying coefficients allowed for nonproportional hazards; hence the use of extended Cox regression, since standard Cox regression assumes proportional hazards, or a constant hazard over time. Adjusted models included the known confounders.

The fully parameterized statistical model was as follows:

$$\begin{array}{c}h\left(t|\mathit{X}\right)=h_0\left(t\right)\exp [{\mathrm{\delta }}_1\left(\mathrm{As}\right)\left(t\right)+{\mathrm{\beta }}_2\left(\mathrm{age}\right)+{\mathrm{\beta }}_3\left(\mathrm{gravidity}\right)\\ +{\mathrm{\beta }}_4\left(\mathrm{clinic}\right)+{\mathrm{\beta }}_5\left(\mathrm{income}\right)(t)+{\mathrm{\beta }}_6\left(\mathrm{edu}\right),\hfill \end{array}$$

where As = arsenic exposure and t = weeks since conception. This is interpreted as the instantaneous hazard of fetal loss/neonatal mortality at week t given covariates X. The baseline hazard (when all covariates are equal to 0) is given by $h_0$. The overall association with arsenic, or the association with arsenic at all times in the study, is given by ${\mathrm{\delta }}_1$. By specifying a specific time and exponentiating ${\mathrm{\delta }}_1$, it is possible to obtain the hazard ratio for a 1-unit increase in arsenic concentration at that given time (24).

Step functions

When step functions were used to obtain an estimated hazard ratio for each month of gestation, we split the observation period into (up to) 11 months of 4 weeks each. Coefficients were then estimated for each time period (“time-varying coefficients”), and models included the same covariates. The time-period–specific hazard ratios for the exposure are then interpreted as the association of arsenic exposure only during that specific time period.

RESULTS

Of the 1,616 pregnant women enrolled in the study, 8 were excluded for being pregnant with twins, 21 were dropped from the analysis for incomplete information on visit date (required for estimated death date), 7 were lost to follow-up before the initial ultrasound examination, 4 were dropped for having unrealistically long pregnancies (≥44 weeks), and 2 had missing data on water arsenic level; this left 1,574 pregnant women in our analysis. Of these women, 203 experienced a fetal loss, 20 experienced a neonatal death, and 239 were lost to follow-up before censoring at 1 month after birth. Of the remaining woman-child dyads, 31 were dropped from the analysis because of missing covariate data, including 21 fetal losses. Table 1 shows the proportions of fetal loss and neonatal mortality among participants according to quartile of maternal drinking water arsenic exposure as measured at enrollment. Households where women had higher levels of arsenic in their drinking water tended to have lower monthly incomes (P < 0.001), less education (P < 0.001), and fewer cesarean deliveries (P < 0.001) when compared with households with the lowest arsenic levels.

Table 1.

Characteristics of a Pregnancy Cohort Monitored for Fetal Loss and Neonatal Death According to Maternal Exposure to Arsenic in Drinking Water, Bangladesh, 2008–2011

Covariate	Quartile of Drinking Water Arsenic Exposure at Study Enrollmenta								P Value
	First Quartile (n = 393) (0.69 μg/L (0.50–0.89 μg/L))b		Second Quartile (n = 410) (1.48 μg/L (0.90–2.00 μg/L))		Third Quartile (n = 386) (9.64 μg/L (2.10–34.00 μg/L))		Fourth Quartile (n = 385) (115.62 μg/L (35.00–1,400.00 μg/L))
	No.c	%d	No.	%	No.	%	No.	%
Outcome
Fetal loss	48	12.2	65	15.9	47	12.2	43	11.2	0.208
Neonatal death	2	0.5	4	1.0	8	2.1	6	1.6	0.228
Maternal age, years									0.371
18–34	387	98.5	401	97.8	376	97.4	381	99.0
≥35	6	1.5	9	2.2	10	2.6	4	1.0
Graviditye,f	0.98 (1.13)		0.84 (1.08)		1.18 (1.21)		1.18 (1.17)		0.001
Monthly household income, taka									0.001
≤3,000	40	10.3	23	5.7	103	27.3	85	22.7
3,001–4,000	85	21.8	76	18.9	115	30.5	113	30.2
4,001–5,000	137	35.1	155	38.6	79	21.0	113	30.2
5,001–6,000	79	20.3	83	20.7	47	12.5	42	11.2
≥6,001	49	12.6	65	16.2	33	8.8	21	5.6
Maternal education									0.001
No school attendance	48	12.2	39	9.5	85	22.0	64	16.6
Primary or secondary education	331	84.2	354	86.3	285	73.8	310	80.5
Higher secondary education or beyond	14	3.6	17	4.2	16	4.2	11	2.9
Type of delivery									0.001
Vaginal	162	55.7	122	43.6	227	78.3	244	81.6
Cesarean	129	44.3	158	56.4	63	21.7	55	18.4
Clinic									0.001
Pabna	91	23.2	24	5.9	282	73.1	319	82.9
Sirajdikhan	302	76.8	386	94.2	104	26.9	66	17.1

Abbreviation: IQR, interquartile range.

^a Quartile sample sizes vary slightly because of duplicate values for the continuous exposure variable. All observations with a given value are in the same quartile.

^b Mean (range) gestational arsenic exposure. Water arsenic values of 0.5 μg/L reflect samples that were below the initial limit of detection (1.0 μg/L) and were assigned a replacement value of 0.5 μg/L (6% of all samples), whereas values between 0.5 μg/L and 1.0 μg/L are actual measured arsenic concentrations.

^c Number of mother-infant pairs.

^d Percentages may not match column totals because of missing covariate data.

^e Values are expressed as mean (standard deviation).

^f Median gravidity by arsenic quartile was as follows: first quartile, 1 (IQR, 2); second quartile, 1 (IQR, 1); third quartile, 1 (IQR, 2); fourth quartile, 1 (IQR, 2).

Condition indices and variance decomposition proportions were examined. No predictor variables were found to be highly collinear with one another. Via survival probability plots and Schoenfeld residuals, income clearly violated the proportional hazards assumption; the remaining variables (maternal age, gravidity, maternal education, and clinic) did not violate the proportional hazards assumption (see Web Figure 1 and Web Table 1, available at https://academic.oup.com/aje). Clinic was explored and did not function as an effect-measure modifier of the arsenic-mortality relationship. While there was no conceptual reason to believe that clinic was associated with the outcome (and indeed adjustment did not meaningfully change the exposure estimate), clinic was included in the final models to help adjust for residual confounding by community.

The results of the analysis using continuous time-dependent coefficients are presented in Table 2. This model examined the associations of arsenic exposure with fetal and neonatal mortality while adjusting for age, gravidity, education, clinic, and time-varying income. Overall, maternal arsenic exposure was not significantly associated with the hazard of offspring death over the observation period, regardless of biomarker used (i.e., water or maternal urine) or timing of outcome (i.e., fetal loss, neonatal death, or both combined) (hazard ratio (HR) = 1.00, 95% confidence interval (CI): 1.00, 1.00). However, the coefficient estimate of time-varying arsenic exposure did vary across time (Figure 1). This time-varying association of maternal arsenic exposure was nonlinear. Namely, maternal arsenic exposure was associated with modestly decreased mortality in early pregnancy. The association between arsenic and mortality then changed direction (though not statistically significantly) after about 24 weeks of gestation, and maternal arsenic exposure was associated with a slightly increased risk of mortality. This elevated risk of mortality then attenuated and approached the null towards the end of pregnancy. We note that clinic was significantly associated with increased probability of death during the observation period (for combined fetal loss/neonatal mortality, HR = 1.70 (95% CI: 1.24, 2.34); for fetal loss only, HR = 1.74 (95% CI: 1.26, 2.41)). Similar results were observed when maternal urine was used for exposure measurement (Web Table 2, Web Figure 2).

Table 2.

Estimated Hazard of Combined Offspring Mortality (Fetal Loss and Neonatal Death) and Fetal Loss Only per Unit Increase in Maternal Drinking Water Arsenic Concentration During Pregnancy (Adjusted Extended Cox Regression^a), Bangladesh, 2008–2011

Variableb	Coefficient (Robust SE)	HRc	95% CI
Combined Mortality (Fetal Loss and Neonatal Death)
Cumulative arsenic exposure (time-varying)	0.000 (0.000)	1.000	1.000, 1.000
Age	−0.149 (0.610)	0.862	0.261, 2.846
Gravidity	−0.081 (0.071)	0.923	0.803, 1.059
Maternal education	−0.328 (0.184)	0.720	0.502, 1.034
Monthly household income (time-varying)	−0.005 (0.003)	0.995	0.990, 1.000
Clinic	0.532 (0.162)	1.702	1.238, 2.340
Fetal Loss Only
Cumulative arsenic exposure (time-varying)	0.000 (0.000)	1.000	1.000, 1.000
Age	−0.091 (0.623)	0.913	0.269, 3.095
Gravidity	−0.110 (0.073)	0.896	0.776, 1.034
Maternal education	−0.265 (0.190)	0.767	0.529, 1.113
Monthly household income (time-varying)	−0.005 (0.003)	0.995	0.990, 1.000
Clinic	0.555 (0.166)	1.742	1.258, 2.411

Abbreviations: CI, confidence interval; HR, hazard ratio; SE, standard error.

^a Continuous time-dependent coefficient model.

^b For variable units and categories, see Table 1.

^c Exponentiated coefficient.

Figure 1.

Combined offspring mortality (fetal loss and neonatal death) (A) and fetal loss (B) according to maternal exposure to arsenic in drinking water at ≤16 weeks’ gestation, Bangladesh, 2008–2011. The graph shows the covariate-adjusted estimate for the time-dependent arsenic exposure coefficient.

This pattern was also observed when the data were modeled as a step function. After adjustment for the same covariates, there was a negative association between arsenic exposure and mortality until approximately the 20th week of gestation, after which there was a positive association between arsenic exposure and mortality (Table 3). In the step function model for fetal loss and neonatal death combined, the hazard ratios for each unit increase in natural log drinking water arsenic ranged from 1.35 (95% CI: 1.08, 1.69) in weeks 25–28 to 0.81 (95% CI: 0.65, 1.02) in weeks 9–12. For the step function model of fetal loss only, the hazard ratios were still highest and lowest in weeks 25–28 and 9–12, respectively (HR = 1.39 (95% CI: 1.10, 1.75) and HR = 0.81 (95% CI: 0.64, 1.02), respectively). Similar results were observed when maternal urinary arsenic concentration was used in the model (Web Table 3).

Table 3.

Estimated Hazard of Combined Offspring Mortality (Fetal Loss and Neonatal Death) and Fetal Loss Only per Unit Increase in the Natural Log of Maternal Drinking Water Arsenic Concentration During Pregnancy (Adjusted Extended Cox Regression^a), by 4-Week Month of Exposure, Bangladesh, 2008–2011^b

Natural Log Arsenic Exposure		HRc	95% CI	SE of Coefficient Estimate
By Month	By Week
Combined Mortality (Fetal Loss and Neonatal Death)
1	1–4	N/Ad	N/A	N/A
2	5–8	1.21	0.95, 1.54	0.12
3	9–12	0.81	0.65, 1.02	0.12
4	13–16	0.96	0.78, 1.18	0.11
5	17–20	0.83	0.63, 1.10	0.14
6	21–24	1.12	0.81, 1.54	0.16
7	25–28	1.35	1.08, 1.69	0.12
8	29–32	1.10	0.83, 1.45	0.14
9	33–36	1.03	0.76, 1.40	0.16
10	37–40	0.93	0.78, 1.11	0.09
11	41–44	1.02	0.78, 1.34	0.14
Fetal Loss
1	1–4	N/Ad	N/A	N/A
2	5–8	1.21	0.95, 1.54	0.12
3	9–12	0.81	0.64, 1.02	0.12
4	13–16	0.95	0.77, 1.17	0.11
5	17–20	0.83	0.63, 1.09	0.14
6	21–24	1.11	0.81, 1.53	0.16
7	25–28	1.39	1.10, 1.75	0.12
8	29–32	1.17	0.88, 1.55	0.15
9	33–36	0.96	0.68, 1.35	0.18
10	37–40	0.92	0.77, 1.10	0.09
11	41–44	1.04	0.78, 1.39	0.15

Abbreviations: CI, confidence interval; HR, hazard ratio; N/A, not applicable; SE, standard error.

^a Step function model.

^b Only the results for the exposure variable are presented. Full models are available from the authors.

^c Exponentiated coefficient.

^d No events in first month.

DISCUSSION

The results of this prospective study suggest that maternal arsenic exposure at the time of enrollment and early pregnancy was associated with all-cause mortality in offspring, but the exposure-response relationship was nonlinear and varied across the course of pregnancy. Specifically, arsenic exposure was associated with a lower probability of fetal loss up to the middle of the second trimester after controlling for maternal age, gravidity, education, clinic, and family income. However, after about the 20th week of gestation, the association between maternal arsenic exposure and mortality switched directions, to the point where maternal arsenic exposure was associated with an increased risk of fetal mortality. Regardless of the type of extended Cox proportional hazards modeling approach or exposure metric used, we consistently observed a time-varying association between maternal arsenic exposure and all-cause mortality throughout our observation period, which extended from enrollment (≤16 weeks’ gestational age) to 30 days postnatal.

Our results suggest that both the dose and timing of arsenic exposure influence the risk of mortality for the developing fetus. These findings compliment previous epidemiologic research showing that arsenic exposure is associated with higher risk of pregnancy loss and/or infant mortality, including a possibly larger association for stillbirths compared with miscarriages (25–29). It is also plausible that arsenic exposure exerts survival pressure on fetuses, which would be akin to the “healthy worker” survival effect reported in environmental occupational studies (30, 31). This phenomenon illustrates how harmful exposures can lead to temporal variations in mortality rates. For example, if arsenic is fetotoxic, then only the most robust pregnancies from the highest exposure group would be enrolled in a birth cohort, which induces selection bias. Additionally, there may be less survival pressure placed on lower-arsenic-exposure groups that could lead to more variability in pregnancy robustness. This could explain why we observed the highest rates of fetal loss in the middle quartiles of arsenic exposure. Thus, arsenic exposure early in pregnancy could effectively create a bottleneck whereby more susceptible fetuses do not survive the first half of pregnancy and those that do survive are more resistant to the negative consequences of fetal arsenic exposure. This could explain the change in the direction of the estimated coefficient for arsenic over time that we observed in this cohort. If this is true, then this fetal survival bias phenomenon has implications for deriving an unbiased effect estimate for arsenic’s developmental toxicity.

In this study, we assessed the risk of all-cause mortality from conception to fetal/neonatal death, yet we enrolled women at ≤16 weeks’ gestation and had to estimate the date of conception from ultrasound measurements. This probably introduced some error in our estimates of gestational age, potentially increasing bias, since arsenic is associated with decreased fetal growth (32–34). Subsequently, it is possible that gestational age was underestimated, especially among women with the highest arsenic exposure (since gestational arsenic exposure has been linked to retarded fetal growth), in which case the pattern of arsenic-associated mortality observed in this cohort may occur later in pregnancy. Additionally, our enrollment strategy would not have captured the earliest pregnancy losses, and subsequently our analysis may have underestimated the association between arsenic exposure and mortality. We attempted to address this issue by adjusting for left-truncation, whereby women only contributed person-time at risk from the point at which they were enrolled. By combining all types of pregnancy losses in this analysis (e.g., miscarriage, stillbirth, and neonatal death), we were able to provide an estimate of the combined developmental toxicity of arsenic exposure.

Additionally, our primary analysis estimated arsenic exposure from maternal drinking water samples. Drinking water is the primary route of exposure to arsenic in this population, and drinking water arsenic levels are highly correlated with biomarkers of internal dose in this population (35). However, relying on water samples probably introduces exposure misclassification, as water samples do not capture individual variability in internal dose. Arsenic measurements from biomarkers such as urine samples provide a better cumulative estimate of internal dose from all sources. These models yielded results similar to those observed for drinking water and mortality, which lends consistency to our findings.

Additionally, women were informed of their drinking water exposures during the course of this study and were counseled on how to reduce their exposure. This resulted in some women reducing their arsenic exposure over the course of their pregnancies. We did not take this variability into consideration. Instead, we modeled the risk of fetal/neonatal mortality on the basis of maternal arsenic exposure measured near the time of conception. Our rationale for this choice was practical and reflected the fact that women cannot go back and change exposures that occurred at conception.

For comparison, we also explored survival from enrollment to neonatal death only, using exposure measurements from water and urine (Web Tables 1–4, Web Figure 2). However, given the limited number of events of interest (n = 20), those results must be interpreted with caution.

Despite these limitations, this study had several strengths. It was prospective and modeled the association of arsenic exposure with fetal/neonatal mortality across a range of exposure levels. Specifically, 63% of our sample had drinking water arsenic levels below the 10-μg/L World Health Organization standard. All participants lived in the same rural recruitment areas, where health care options are limited. All participants received the same prenatal and postnatal care and had the same access to health care. Additionally, this population generally does not smoke or drink alcohol, which eliminated unmeasured confounding from those factors. We also used 2 different extended Cox regression models that allowed for time-varying coefficients with both approaches, yielding similar results. Finally, we controlled for several important confounders, including household income and maternal education.

In conclusion, we observed a persistent nonlinear association between maternal arsenic exposure measured near the time of conception and the probability of offspring mortality from enrollment (≤16 weeks gestational age) to 1 month postnatal. These results suggest that arsenic exposure is fetotoxic. Researchers in epidemiologic studies examining the association of arsenic exposure with child health should recognize that developmental arsenic exposure may exert survival pressure, which would contribute to selection bias. Finally, the isk of mortality increased when arsenic exposure was above 1 μg/L in this population, suggesting that any detectable level of arsenic exposure poses a risk to the developing fetus.

Abbreviations

CI

confidence interval

DCHT

Dhaka Community Hospital Trust

HR

hazard ratio

OR

odds ratio