Levels and determinants of DDT and DDE exposure in the VHEMBE cohort

BACKGROUND: Although indoor residual spraying (IRS) is an effective tool for malaria control, its use contributes to high insecticide exposure in sprayed communities and raises concerns about possible unintended health effects. OBJECTIVE: The Venda Health Examination of Mothers, Babies and their Environment (VHEMBE) is a birth cohort study initiated in 2012 to characterize prenatal exposure to IRS insecticides and exposures’ impacts on child health and development in rural South Africa. METHODS: In this report, we describe the VHEMBE cohort and dichlorodiphenyltrichloroethane (DDT) and dichlorodiphenyldichloroethylene (DDE) serum concentrations measured in VHEMBE mothers when they presented for delivery. In addition, we applied a causal inference framework to estimate the potential reduction in population-level p,p0 -DDT and p,p0 -DDE serum concentrations under five hypothetical interventions. A total of 751 mothers were enrolled. RESULTS: Serum concentrations of p,p0 isomers of DDT and DDE were above the limit of detection (LOD) in ≥98% of the samples, whereas the o,p0 isomers were above the LOD in at least 80% of the samples. Median (interquartile range) p,p0 -DDT and p,p0 -DDE serum concentrations for VHEMBE cohort participants were 55.3 (19.0–259.3) and 242.2 (91.8–878.7) ng/g-lipid, respectively. Mothers reporting to have lived in a home sprayed with DDT for malaria control had ∼ 5–7 times higher p,p0 -DDT and p,p0 -DDE serum concentrations than those who never lived in a home sprayed with DDT. Of the five potential interventions tested, we found increasing access to water significantly reduced p,p0 -DDT exposure and increasing the frequency of household wet mopping significantly reduced p,p0 -DDT and p,p0 -DDE exposure. CONCLUSION: Our findings suggest that several intervention approaches may reduce DDT/DDE exposure in pregnant women living in IRS communities.

VHEMBE is a birth cohort study based in the rural Vhembe district of South Africa’s Limpopo Province. The study aims to investigate the potential effects of IRS insecticide exposure on child growth and development. Between August 2012 and December 2013, we enrolled mother–newborn dyads at the time of maternal presentation for delivery at Tshilidzini Hospital. Eligible women were  ≥ 18 years old, spoke Tshivenda at home, lived within 20 km of the hospital, planned to remain in the area, had not been diagnosed with malaria during pregnancy, had contractions  > 5 minutes apart, and gave birth to a viable singleton. We obtained informed consent by verbally explaining the study procedures prior to the collection of study data. All human subject protocols were approved by institutional review boards at the University of California, Berkeley; McGill University; the University of Pretoria; the Limpopo Department of Health and Social Development; and Tshilidzini Hospital. Out of the 1,649 women approached to participate in the VHEMBE study, 920 were eligible ( ∼ 57%). Of those eligible, 152 refused enrollment ( ∼ 16%), 14 did not complete a baseline questionnaire ( ∼ 1%), and three did not provide a sufficient blood sample for DDT analysis ( < 1%). In total, 751 mothers completed a baseline questionnaire and provided a blood sample, and 722 were visited at their homes by our staff one week after delivery ( ∼ 96%). On average, mothers enrolled in the VHEMBE study were 1.6 years younger and had given birth to 0.2 fewer children prior to the index child than had eligible mothers who refused enrollment (p - values < 0.05). Tshivenda-speaking study staff administered a baseline questionnaire before hospital discharge to collect data on demographic characteristics (e.g., maternal age, primary language, marital status, education, and household income), parity, length of cumulative breastfeeding, hygiene/cleaning habits, and housing and IRS-use history. Household income was compared with the food poverty line determined by Statistics South Africa (W. Ruch, written communication, May 2014; Statistics South Africa 2014). We also assessed nutrient intake by administering a quantitative food frequency questionnaire (FFQ) validated in the Limpopo population (MacIntyre et al. 2001a,b; MacIntyre et al. 2001c). FFQ parameters were generated using the Food Finder 3 program (Nutritional Intervention Research Unit and Biomedical Research Division). Maternal height was measured using a wall-mounted stadiometer (Charder HM210D; Taichung City, Taiwan), and weight was measured with a digital scale (Beurer PS06; Ulm, Germany). All measurements were performed in triplicate, with the mean values used to calculate body mass index (BMI). At the one-week visit, we performed home inspections to collect information on household water source, building type, and homes’ latitude and longitude coordinates. We used 2009 Spot 5 satellite imagery to create spatial variables to test for the association between location and p,p′-DDT/E serum concentrations. We calculated the minimum distance from each participant’s home to the nearest body of water using ArcGIS’s ‘Near’ tool. Water bodies were defined based on publicly accessible national datasets but were supplemented with manual additions, drawn using ArcScan based on the Spot 5 imagery. The distance-to-body-of-water variable was created based on the hypothesis that participant homes located near bodies of water (potential mosquito habitats) would be more likely to undergo IRS applications and proximity would result in higher exposure to participants. In addition, we used the kernel density ArcTool to calculate the number of structures per hectare within 250 and 1,000 m buffers from the participant’s home. Our structure density variable was created based on the hypothesis that density of IRS use in the area (spraying is done by structure) would influence the exposure of the participant within that area. This variable was generated using ArcScan to extract imagery pixels with a radiometric resolution of 220 or higher. The resulting extracted pixel layer was cleaned by hand (to minimize misclassification of other features such as roads and structures). The pixels were then converted to points, and the kernel density was completed. Maternal blood was collected into red-top vacutainer tubes by study nurses prior to delivery (n = 595) or immediately after delivery (n = 156). Samples were immediately processed and stored at −80°C. Serum aliquots were sent on dry ice to Emory University’s Rollins School of Public Health for the measurement of p,p′ and o,p′ DDT/E using gas chromatography-tandem mass spectrometry (GC-MS) with isotope dilution quantification (Barr et al. 2003). The limit of detection (LOD) and limit of quantification (LOQ) for p,p′-DDT, o,p′-DDT, and o,p′-DDE were 0.01 and 0.05 ng/mL, respectively. For p,p′-DDE, the LOD and LOQ were 0.03 and 0.15 ng/mL, respectively. Total lipid concentrations were estimated based on triglycerides and total cholesterol concentrations (Phillips et al. 1989), measured using standard enzymatic methods (Roche Chemicals, Indianapolis, IN). Quality-control procedures included field spikes, field blanks, matrix-matched calibrants, and laboratory-prepared serum and reagent blanks analyzed concurrently with participants’ samples. The Supplemental Information (SI) describes the laboratory method used to quantify the p,p′ and o,p′ isomers of DDT/E and provides detailed quality control information. We used Spearman’s correlation or Kruskal-Wallis tests to examine the bivariate relationships of participant characteristics and potential determinants of exposure with DDT/DDE serum concentrations. Only the p,p′ isomers of DDT and DDE were considered for these analyses due to lower detection frequencies in the o,p′ isomers of DDT and DDE. For the p,p′-DDT/E serum concentrations below the laboratory’s LOQ, but above the LOD, we assigned those values the GC-MS machine-read value (n = 27 for p,p′-DDT and n = 12 for p,p′-DDE). For p,p′-DDT serum concentrations below the laboratory’s LOD (n = 15), we imputed those values from maximum likelihood estimates of the lognormal distribution of the detected serum values (Lubin et al. 2004). Spearman’s correlation tests were also used to examine the correlation between p,p′-DDT and p,p′-DDE serum concentrations. Associations were considered statistically significant if p - values were  < 0.05. For the causal inference analysis, we aimed to estimate the marginal geometric mean difference in p,p′-DDT/E serum levels (Y) if, contrary to fact, all VHEMBE mothers were given an intervention (A = 1) relative to a scenario in which none of the mothers were given that intervention (A = 0): E[E(Y|A = 1, W) − E(Y|A = 0, W)], where W is a matrix of covariates. In addition, we tested the effect of potential interventions by whether the mother reported ever living in a home sprayed with DDT to explore exposure reduction effect modification by spray status. Covariates used in the TMLE analysis included the following: if the mother ever lived in a home sprayed with DDT for malaria control (W1, categorical); if the mother lived in a home sprayed with DDT for malaria control during pregnancy (W2, categorical); the frequency of IRS in the home where the mother lived during pregnancy (W3, ordinal); if the mother lived in a village sprayed for malaria control during pregnancy (W4, categorical); the frequency of IRS in the village where the mother lived during pregnancy (W5, ordinal); the time spent in an IRS home (W6, no. of years); the mother’s age at delivery (W7, years); the education level of mother at delivery (W8, ordinal); household income (W9, Rands per household member per month); whether the pregnancy home was a rondavel with earthen walls and thatched roof (W10, categorical); parity of mother at delivery (W11, no. of previous births); breastfeeding history (W12, no. of months); presence of a rondavel on homestead (W13, categorical); if the household owned livestock (W14, categorical); proximity of mother’s home to the nearest body of water (W15, meters); structure density within 250-m radius of the mother’s home (W16, no./hectare); and maternal BMI after delivery (W17,  kg/m2). The potential interventions that we evaluated included: 1) living in a home with piped water [A1=A1(A4,5,W1−17)]; 2) living in a home in which floors were mopped more than seven times weekly (median frequency reported by mothers) [A2 = A2(A1, A3−5, W1−17)], 3) washing bed sheets more than two times per month (median frequency reported by mothers) [A3=A3(A1,2,A4,5W1−17)]; 4) avoiding a high-fat diet (<the 75th percentile of fat intake among VHEMBE women) [A4 = A4(A1−3, A5, W1−16)]; and 5) avoiding local dairy/meat/poultry fish products during pregnancy [A5 = A5(A1−4, W1−17)]. The potential interventions for this analysis were selected because they may be modifiable characteristics that were hypothesized to reduce DDT/DDE exposure, while maintaining effective malaria control. The marginal geometric mean difference of p,p′-DDT/DDE serum concentrations for each intervention was evaluated in separate models using targeted maximum likelihood estimations (TMLE), a doubly robust substitution estimator that generates unbiased estimates if either models for the estimation of the exposure [E(Y|A, W)] or determinant mechanisms [E(A|W)] are correct (Rose and van der Laan 2011; van der Laan 2006; van der Laan and Rubin 2006). A directed acyclic graph (DAG) was generated to conceptualize the estimation of serum levels and interventions and to identify potential confounders (Figure S1) (Textor et al. 2011). Missing covariate values ( < 5%) were imputed at random based on their observed probability distributions. To estimate [E(Y|A, W)] and E(A|W), we used the Super Learner algorithm, an ensemble machine learning algorithm that uses a weighted combination of algorithms to return a prediction function that minimizes cross-validated mean squared error (van der Laan et al. 2007). We assessed positivity using the propensity score for each intervention and found that our positivity assumption holds for all interventions as the lowest propensity score was 0.07, and the median propensity scores across all interventions ranged from 0.53 to 0.75 (Table S1). To estimate E(Y|A, W) and E(A|W), we used the Super Learner algorithm with the following candidate algorithms: generalized linear models, generalized additive models, Bayesian linear model, support vector machine, recursive partitioning and regression trees, elastic net, neural network, local polynomial regression, and random forest. The associated weights used by Super Learner for estimating E(Y|A, W) and E(A|W) are presented in the Supplemental Material (Tables S2 and S3). We used bootstrapping (n=1,000) to estimate 95% confidence intervals (CI) based on the percentile method (Efron 1979). Data analyses were performed using the statistical programs R (version 3.1.3; R Development Core Team) and ArcGIS (version 10.3; ESRI Corporation). We compared VHEMBE lipid-adjusted p,p′-DDT and p,p′-DDE serum concentrations to serum/plasma levels previously reported in 1) adults living in IRS communities and 2) pregnant women from non-IRS communities in the United States. The median and inter-quartile ranges (IQR) were used to compare serum/plasma concentrations across studies, as those descriptive statistics were the most commonly reported. Because Aneck-Hahn et al. (2007) reported only the arithmetic mean and standard deviation (SD) of men living and not living in DDT-sprayed homes, the geometric mean (GM) and geometric standard deviation (GSD) were estimated according to equations presented in Jean and Helms (1983). We sampled 1,000 values from a log-normal distribution using the estimated GM and GSD to estimate the median and IQR of population from Aneck-Hahn et al. (2007). As only wet-weight concentrations (ng/mL) were presented by Whitworth et al. (2014), the Study of Women and Babies (SOWB) researchers graciously provided the lipid-adjusted distributions for comparison (K.W. Whitworth, written communication, October 2014). We compare only the Van Dyk et al. (2010) results for p,p′-DDE in adults living in home sprayed  ∼ 60 days prior to blood collection because the detection frequency for p,p′-DDT in sprayed homes (5%) and p,p′-DDT/E in unsprayed communities were low (0 and 33%, respectively). Only the lipid-adjusted values from the control group (n = 283) were used from the case-control study of Bhatia et al. (2004). The p,p′-DDT/E serum concentrations from pregnant women who participated in the 1999–2000, 2001–2002, and 2003–2004 National Health and Nutrition Examination Study (NHANES) (DDT n = 263, DDE n = 277) were combined (Center for Disease Control 2000, 2002, 2004). In the three NHANES surveys, p,p′-DDE was detected in 100% of the samples, and p,p′-DDT was detected in 37% of the samples (LOD ∼ 5.1 ng/g - lipid).

• Dima Health