Flood exposure and pregnancy loss in 33 developing countries

Floods have affected billions worldwide. Yet, the indirect health impacts of floods on vulnerable groups, particularly women in the developing world, remain underexplored. Here, we evaluated the risk of pregnancy loss for women exposed to floods. We analyzed 90,465 individual pregnancy loss records from 33 developing countries, cross-referencing each with spatial-temporal flood databases. We found that gestational flood exposure is associated with increased pregnancy loss with an odds ratio of 1.08 (95% confidence interval: 1.04 - 1.11). This risk is pronounced for women outside the peak reproductive age range (<21 or >35) or during the mid and late-stage of pregnancy. The risk escalated for women dependent on surface water, with lower income or education levels. We estimated that, over the 2010s, gestational flood events might be responsible for approximately 107,888 (CIs: 53,944 - 148,345) excess pregnancy losses annually across 33 developing countries. Notably, there is a consistent upward trend in annual excess pregnancy losses from 2010 to 2020, and was more prominent over Central America, the Caribbean, South America, and South Asia. Our findings underscore the disparities in maternal and child health aggravated by flood events in an evolving climate.


Robustness to the alternative source of flood exposure data
The Global Flood Database (GFD) 1 is an alternative source of data to the Dartmouth Flood Observatory (DFO) database and also can be used to identify flood events.However, the GFD has a limited time span and total flood events, as it adopted satellite imageries to identify the flood events.Furthermore, GFD dataset does not provide comprehensive coverage of all areas affected by flood events.Instead, it focuses on specific regions with flood inundation.In addition, as the GFD dataset only captured about 30% of the global flood events, some participants may be misclassified as unexposed.To address the potential misclassification issue, we implemented additional screening measures in the matching results in this part of sensitivity analysis.Since the DFO covered nearly all flood events compared to the Global Flood Database GFD, any records that overlapped with a flood event in the DFO but not in the GFD were excluded from this part of analysis.By doing so, we ensured that the records included in this part of the analysis fell into one of the following categories: records with no exposure to floods at all, records exposed only to floods identified by the GFD, or records demonstrating the simultaneous recognition of flood events from both the GFD and DFO.Our aim in implementing this procedure was to minimize misclassification errors that could arise from solely considering unexposed data based on the GFD dataset.
In all, we use the GFD data to evolute the impact of exposure of flood inundation during pregnancy period, to test for the robustness of our main results.Table S1 listed

Robustness to different controls
In the main model, we controlled maternal age in the delivery year and a categorical term for the year and month of conception, and gestational mean temperature and precipitation.Although this helps to improve the identification of the causal effect of flood events, such fixed effects restrict potentially important variation that occurs across time and space.Below, we test the sensitivity of our results to the inclusion of less restrictive control effects (Fig S2).All specifications imply that having flood exposure during the gestational period increases pregnancy loss risk, although the magnitude of the effect varies.

Robustness to potential migration
Our primary findings are derived from the analysis of DHS surveys, which provide information on the location of clusters representing villages or neighborhoods where the mothers were interviewed.It is important to note that the house or location where the interview took place may not necessarily be the same as the place where the women resided during their previous gestational period, particularly in cases of forced migration due to floods.To address this potential bias, we decided to focus our analysis on a subset of women who reported having lived in their current house for at least 10 years.By applying this restriction, we excluded 72.73% of the total cases from our sample.As depicted in Figure S4, the main results obtained from this restricted sample remain consistent, albeit with wider error bars due to the reduced sample size.

Effects of different types of floods on pregnancy loss
Fig S5 depicts the excess pregnancy losses due to three main types of flood events.As mentioned in the main text, in addition to the total flood events, we estimated the excess pregnancy losses due to different sources of flood events, including heavy rains or monsoon rains, tropical cyclones or storms, and levee/dem break or release.

1. 1 1 . 2 to different controls 1 . 3 1 . 4 3 . 2 3 . 4
Robustness to the alternative source of flood exposure data Robustness Robustness to leaving each region out individually Robustness to potential migration 2. Effects of different types of floods on pregnancy loss 3. Extra data 3.1 Example illustrating our case-control design Map of data selection 3.3 DHS Surveys used in the main analysis Statistics of matched flood events over different regions 1 Robustness of main results the differences between identified cases and controls based on these two databases.As shown in Fig S1, although the matched cases were reduced based on GFD, the estimated risks of pregnancy loss are very similar.

Fig S1 :
Fig S1: Change in the risk of pregnancy loss for women due to gestational flood exposure identified by two flood databases.The dots are point estimates of the odds ratio of pregnancy loss with gestational flood exposure.Error bars indicate 95% CIs.

Fig
Fig S2 Estimation with different sets of covariates in the main model.The dots are point estimates and the error bars are corresponding 95% confidence intervals.

Fig
Fig S3 Estimation with cases from each of the main regions.The dashed line represents the mean odds ratio over all study regions.The dots are point estimates and the error bars are corresponding 95% confidence intervals.

Fig S4 :
Fig S4: The main results and subsequently re-estimated the same parameters after excluding potential migrants from the sample.The dots are point estimates and the error bars are corresponding 95% confidence intervals.

Fig S5 :
Fig S5: Average annual excess pregnancy losses (per 10,000 deliveries) due to three types of flood exposure (2010-2020).Spatial distribution of annual excess pregnancy losses at a spatial resolution of 10 km × 10 km due to floods sourced from heavy rains or monsoon rains (A), tropical cyclones or storms (B), and levee/dem break or release (C).

3. Extra data 4 . 1 4 . 1
Fig S6: Example illustrating our case-control design.Bars represent the months of gestation for the cases.