The Migration Household Index (MHI) measures household mobility by capturing migration frequency, duration, and type, along with its drivers. It uses survey data on whether households have migrated, how often and how long, and the reasons–whether environmental (flooding, erosion), economic, or social.

The Migration Household Index is defined in Equation 1 and normalized in Equation 2. MHI = ∑ i = 1 N Weight t , i × Frequency i × Prop m , i Total Household Members , (1) where: N represents the total number of migration events reported by the household, Weight t , i represents the weight assigned to each type of migration based on its socio-economic impact (e.g., permanent migration may carry a higher weight than seasonal migration), Frequency i represents the number of times a specific type of migration occurred Prop m , i : The proportion of household members involved in each migration event. Total Household Members : Total number of individuals in the household.

The raw MHI is then normalized to ensure comparability across households: Normalized MHI = Raw MHI − Minimum MHI Maximum MHI − Minimum MHI × 100 , (2) where Minimum MHI and Maximum MHI represent the observed minimum and maximum values of the raw index in the dataset.

The Disease Burden Index (DBI) measures the prevalence and severity of health conditions in households from Northeast Bangladesh’s chars, standardizing assessment by considering both the share of affected members and illness intensity.

Following [27], we define the Disease Burden Index (DBI) for household i . The Disease Burden Index is computed using Equations 3, 4 and normalized in Equation 5. D B I i = T i N i × 100 , (3) where T i is the total illness types reported and N i is the household size. The index is positive if any member is ill, capturing both the proportion of ill members and their illness severity. Defining n i as the number of reported illness types, we decompose: D B I i = n i N i × T i n i × 100 = d b i i p × d b i i s (4)

With d b i i p measuring the ill proportion and d b i i s indicating severity (see Supplementary Appendix C.2).

The Wealth Household Index (WHI) combines three wealth dimensions: land ownership (value of farmland, homestead, and rented/mortgaged land; see Supplementary Table C.3), household goods (furniture, appliances, transportation), and livestock (e.g., cows, goats, poultry). Values are based on market prices summarized in Supplementary Table C.3 (in Bangladeshi Taka).

We illustrate wealth distribution using WHI histograms by treatment (Supplementary Figure C.7) and displacement status (Supplementary Figure C.8), comparing treatment chars (where Friendship NGO operates) to control chars and displaced to non-displaced households. Scatter plots (Supplementary Figures C.9, C.10) further explore the relationship between age and wealth.

The Hygiene Household Index (HHI) measures household hygiene by combining data on sanitation behaviors, infrastructure, cultural practices, water usage, and toilet facilities. It uses binary variables (e.g., handwashing, waste disposal) alongside categorical assessments of water source, accessibility, and toilet quality (see Supplementary Table C.4).

Histograms of the HHI–disaggregated by treatment (Supplementary Figure C.11) and displacement (Supplementary Figure C.12)– compare hygiene conditions between treatment chars (Friendship NGO beneficiaries) and controls, and between displaced and non-displaced households. Scatter plots (Supplementary Figures C.13, C.14) further examine how age affects hygiene practices.

The Educational Attainment Index (EAI) measures household education by weighting each member’s highest education level adjusted for age (see Supplementary Table C.5), allowing comparisons across age groups. Histograms by treatment (Supplementary Figure C.15) and displacement (Supplementary Figure C.16) compare households in treatment chars (Friendship NGO) versus controls and displaced versus non-displaced households, while scatter plots (Supplementary Figures C.17, C.18) highlight generational differences in educational access.

The survey results are summarized through the construction of five key indices–Disease Burden Index (DBI), Wealth Household Index (WHI), Migration Household Index (MHI), Hygiene Household Index (HHI), and Educational Attainment Index (EAI) – to evaluate health and socio-economic conditions. Table 2 provides a compact overview of each index, including its main dimension, scale and the interpretation of higher values. Table 3 then reports descriptive statistics for these indices in the full sample, while Supplementary Tables C.7, C.8 present separate descriptive statistics by treatment status and displacement, providing a comparative overview of the treatment and control chars, as well as displaced and non-displaced households.

Having constructed the study indices, we next estimate the causal effects of (i) household displacement and (ii) residence in chars with Friendship NGO interventions on health and socio-economic outcomes. Because neither displacement nor NGO program placement is randomized, simple comparisons may be confounded by systematic differences between treated and control groups. We therefore adopt a potential-outcomes framework and estimate Average Treatment Effects (ATEs) under a selection-on-observables assumption, using multiple complementary estimators (regression adjustment, inverse-probability weighting, doubly robust approaches, and matching). We also assess covariate balance and robustness to support the credibility of the causal interpretation.

To analyze the impact of displacement and Friendship NGO interventions, we employ the causal inference framework of [16] and define the Average Treatment Effect (ATE). For each household i , we define two potential outcomes, where y 0 ( i ) is the outcome if the household is not treated (e.g., non-displaced or residing in a control char), while y 1 ( i ) represents the outcome if the household is treated (e.g., displaced or residing in a treatment char).

In Model 1 (Equation 9), y 0 ( i ) represents the Disease Burden Index (DBI) for a non-displaced household, while y 1 ( i ) represents the DBI for a displaced household. In Model 2 (Equation 10), y 0 ( i ) and y 1 ( i ) correspond to the Migration Household Index (MHI) for households in control and treatment chars, respectively, allowing us to evaluate the impact of NGO interventions on migration dynamics.

The individual treatment effect and the Average Treatment Effect are defined in Equations 6, 7, with the sample estimator given in Equation 8. τ i = y 1 i − y 0 i (6) Since τ ( i ) varies across households, the ATE is defined as: ATE = E y 1 − y 0 (7) which represents the average impact of displacement or NGO interventions on the outcome variable across all households.

Assuming ignorability of treatment assignment and a sufficiently large sample size, the ATE is estimated as: ATE ̂ = 1 N ∑ i = 1 N y 1 i − y 0 i (8) where N is the total number of households in the sample.

This framework enables robust estimation of causal effects, providing reliable insights into the socio-economic and health challenges faced by vulnerable populations in the chars of Bangladesh.

To ensure robustness in estimating treatment effects, we employ several causal inference techniques. In this subsection we briefly describe, in intuitive terms, what each method does; more technical details and formal notation are provided in Supplementary Appendix E for interested readers.

Regression Adjustment (RA): Controls for confounding variables by adjusting for covariates in regression models [22].

These methodologies collectively form a robust framework for causal inference, allowing us to draw reliable conclusions about the effects of displacement and NGO interventions in the chars of Northeast Bangladesh. Further methodological details are provided in Supplementary Appendix E.

The causal interpretation of the estimated treatment effects relies on the standard identification assumptions used in the program-evaluation literature [16, 20, 21]. For both Model 1 and Model 2 we assume conditional ignorability, i.e., y 0 i , y 1 i ⊥ D i | X i , where D i denotes the treatment indicator (displacement status in Model 1, treatment char in Model 2) and X i is the vector of observed covariates used in the propensity-score and outcome models. In our application, X i includes the Wealth Household Index (WHI), Hygiene Household Index (HHI), Educational Attainment Index (EAI), age of the respondent, agricultural and non-agricultural income quintiles, and fixed effects for chars. Intuitively, this assumption requires that, conditional on these covariates, treatment assignment is as good as random.

We also assume positivity (or overlap): for all values of X i in the support of the data, the probability of treatment is strictly between zero and one, so that each type of household has a non-zero chance of being both treated and untreated. This assumption is crucial for weighting and matching estimators, which rely on comparisons between treated and control observations with similar propensity scores. To assess overlap, Supplementary Appendix E reports the distributions of estimated propensity scores for treated and control groups in both models; these figures show substantial common support and no extreme concentrations near 0 or 1.

To evaluate whether conditioning on X i achieves adequate balance, we compute standardized mean differences (SMDs) for all covariates before and after applying the IPW, AIPW and IPWRA weights and after nearest-neighbour matching. Supplementary Table E.1 summarises these diagnostics. Prior to adjustment, some covariates exhibit absolute SMDs above commonly used thresholds (e.g., 0.2), indicating important differences between treated and control households. After weighting or matching, absolute SMDs for all main covariates are reduced to well below 0.1 in both models, which is typically viewed as evidence of good balance. These diagnostics, together with the propensity-score overlap plots, support the plausibility of our identification strategy, while we acknowledge that residual bias from unmeasured confounders cannot be ruled out in a cross-sectional design.

We estimate treatment effects using two models, each addressing a distinct research question. Model 1 (Equation 9) estimates the relationship between displacement and health outcomes (DBI). Model 2 (Equation 10) evaluates the impact of residing in treatment chars on migration dynamics (MHI). Both models account for potential biases and confounding factors [20, 21].

Model 1, Equation 9, estimates the Average Treatment Effect (ATE) of displacement–measured by the binary variable Displ-NonDispl–on the Disease Burden Index (DBI), while controlling for a series of covariates, including observable char-specific characteristics to capture inter-Char variation.

The estimated ATE presented in Table 4 is significantly negative, ranging from − 4.85 to − 6.19 , indicating that, when controlling for covariates, displaced households report a substantial reduction in disease burden compared to non-displaced households. This pattern is counterintuitive in light of the broader literature, yet it is consistent with a combination of mechanisms. One possibility is that displaced households, particularly in treatment chars, have better access to Friendship NGO services or other providers (for example, through mobile clinics or health outreach), which may improve prevention, diagnosis and treatment and thus reduce their disease burden. Another possibility is that displacement changes health-seeking behaviour and the way symptoms are perceived and reported, leading to systematic differences in how illness episodes are recorded in the survey. Finally, selection mechanisms, including forms of “survivor bias”, in which healthier, more resilient households are more likely to remain observable in the displaced population, may also shape which households we observe in this cross-sectional sample.

The Potential Outcome Mean (POmean) for non-displaced individuals–ranging from 23.34 to 24.06 – indicates a consistently high disease burden in more stable habitats, likely exacerbated by systemic health and sanitation challenges and, in some cases, more limited contact with NGO services. Taken together, these results suggest that the lower DBI observed among displaced households should not be interpreted as evidence that displacement is benign, but rather that the combination of service access, reporting differences and selection may partially mask underlying health risks among the most vulnerable.

Because we observe households at a single point in time, we cannot empirically disentangle these mechanisms. In Section Discussion we therefore treat survivor bias and other forms of selection as plausible, but not definitive, explanations for the lower DBI among displaced households.

Table 4 summarizes Model 1 (Equation 9) treatment effect estimations:

The estimates in Table 4 indicate that, after adjusting for covariates, displaced households have on average between 4 and 6 index points lower DBI than comparable non-displaced households. Given that the DBI ranges from 0 to 100 and the expected DBI for the non-displaced group (POmean) is around 23–24 points, this corresponds to roughly a 20%–25% reduction in the average disease burden among displaced households relative to non-displaced ones in our sample. This magnitude helps quantify the practical importance of the negative ATE, beyond its statistical significance.

While displacement typically heightens vulnerability, these findings suggest that targeted interventions can mitigate health impacts, even though non-displaced groups still face significant risks from limited healthcare and sanitation.

Model 2, Equation 10, estimates the Migration Household Index (MHI) as the outcome variable, representing the migration index, with the treatment variable identifying chars where Friendship NGO implements health and educational activities. This model incorporates the same covariates as Model 1.

Across all methodologies, the ATE estimate reported in Table 5 is significantly positive, ranging from 9.93 to 10.45. Given that higher values of MHI indicate more intense or frequent migration, these results imply that households in treatment chars exhibit, on average, higher migration intensity than comparable households in control chars, after conditioning on the covariates in the model. One plausible explanation is that Friendship tends to operate in chars already experiencing stronger migration pressures, so that treatment status partly reflects programme targeting rather than the causal impact of the intervention itself. We therefore interpret these estimates as evidence of systematic differences in migration patterns between treatment and control chars, rather than as demonstrating that NGO activities increase or decrease migration.

The Potential Outcome Mean (POmean), also shown in Table 5, represents the expected migration index for the control group. It is consistently high (ranging from 8.59 to 8.76), which points to a non-trivial baseline level of migration activity even among households in chars without Friendship interventions and reflects pre-existing migration trends and pressures in the study area.

Table 5 summarizes Model 2 (Equation 10) treatment effect estimations:

From Table 5, the estimated ATEs of roughly 10 index points imply that households living in chars where Friendship operates have, on average, substantially higher migration intensity than similar households in control chars, conditional on the covariates in the model. Since the MHI also ranges from 0 to 100 and the expected value for the control group (POmean) is about 9 points, the positive ATE indicates that treatment-char households experience migration levels that are roughly double those of households in non-program chars. This interpretation makes the scale and direction of the MHI differences easier to relate to concrete migration pressures.

To ensure the robustness of our estimations, we performed a Variance Inflation Factor (VIF) analysis to examine potential multicollinearity among the constructed indexes; the results are presented in Supplementary Appendix D. Additionally, we computed the ROC curve and AUC for each model to evaluate their discriminatory performance. The following section presents the Receiver Operating Characteristic (ROC) curve.

Displaced households show significantly lower DBI scores than non-displaced ones, even after adjusting for wealth, hygiene, educational attainment, age, income quintiles and char fixed effects. At face value, this pattern is consistent with beneficial effects of targeted health interventions and improved access to services in some displaced settings, as well as with effective coping strategies that reduce the burden of illness among households that are able to relocate.

At the same time, this finding is counterintuitive given the broader literature documenting that displacement typically heightens health risks. Several, non-mutually-exclusive mechanisms may help explain the result. First, treatment chars are precisely those where Friendship NGO operates, so displaced households in these areas may enjoy better access to preventive and curative care than non-displaced households in less-served locations. Second, displacement may influence how symptoms are perceived, normalised or reported, generating potential reporting and recall differences between displaced and non-displaced groups. Third, selection mechanisms–including forms of “survivor bias” in which more resilient households are more likely to remain observable among the displaced–could lead the most vulnerable households to exit the sampling frame through severe morbidity, mortality or onward migration.

Because our data are cross-sectional, we cannot empirically separate these mechanisms, and we therefore present survivor bias and related selection processes as plausible rather than definitive explanations. The robustness checks and sensitivity analysis reported in Supplementary Appendix E help bound the extent to which unmeasured selection would need to operate to overturn our main findings and highlight the need for future longitudinal work that follows households over time.

This study provides valuable insights into the health, migration, and socio-economic conditions of char households in North-East Bangladesh, but it has several important limitations. First, our analysis is based on a single cross-sectional survey conducted between March and June 2022. As a result, the data provide a snapshot rather than a dynamic picture, and we cannot observe pre-displacement baselines, post-displacement adaptation, or long-run trajectories of health and migration. The causal interpretation of the estimated treatment effects is therefore confined to this cross-sectional setting, under the stated identification assumptions, and does not extend to changes over time.

Second, all outcome variables are self-reported, which may introduce recall error and reporting bias. Differences in how displaced and non-displaced households perceive and report symptoms, income, or migration events could affect the constructed indices, including the Disease Burden Index (DBI) and Migration Household Index (MHI). Although we attempt to mitigate these issues through careful questionnaire design and index construction, we cannot fully rule out systematic reporting differences.

Third, while the quasi-experimental estimators adjust for a rich set of observed covariates, they cannot eliminate bias from unmeasured historical, environmental, or socio-economic factors. We cannot directly observe households that may have dissolved, experienced mortality, or migrated out of the study area, so selection mechanisms—including possible forms of “survivor bias”— remain a concern. As a consequence, our estimates should be interpreted as associations that are consistent with causal effects under the conditional ignorability assumption, rather than as definitive causal effects that hold under all forms of unobserved heterogeneity.

Fourth, the implications for external validity are limited by the specific context of the chars and by our sampling frame. The studied chars are riverine islands with highly dynamic landforms, recurrent erosion, and particular settlement and livelihood patterns, and the treatment chars host long-standing Friendship NGO health and education programmes. These physical and institutional conditions differ from other displacement settings in South Asia, such as coastal deltas, urban informal settlements, or conflict-affected areas, where both hazards and service provision may look very different. In addition, the survey interviews one primary respondent per household, restricted to women aged 15–55, which reflects Friendship NGO’s programmatic focus and the fact that women in this age group are typically the main caregivers and best informed about health, hygiene and education decisions. Households without an eligible woman (for example, male-only or elderly-only households) are therefore not represented. Our quantitative estimates are thus most directly generalisable to similar char households in North-East Bangladesh, rather than to all displacement-prone contexts in South Asia.

Despite these drawbacks, the study lays a foundation for understanding the links between temporary displacement, health and socio-economic vulnerability in Bangladesh’s chars. It also highlights the need for longitudinal and multi-site research that follows households over time and compares different types of displacement-prone environments across South Asia.

Our study uses a cross-sectional quasi-experimental framework and a set of household indices (DBI, MHI, WHI, HHI, EAI) to examine how temporary displacement and NGO presence are associated with health and socio-economic outcomes in chars in North-East Bangladesh. Within this setting and under the stated identification assumptions, the estimated treatment effects provide evidence on how displacement status and residence in treatment chars relate to disease burden and migration dynamics at the time of the survey.

For health outcomes, we find that displaced households in areas with Friendship NGO interventions exhibit lower DBI scores than comparable non-displaced households. This pattern is consistent with targeted health measures and better access to services helping to reduce disease burden among displaced households, but may also reflect differences in reporting behaviour and selection in which households remain observable. For migration dynamics, chars where Friendship operates show systematically different—and, on average, higher—MHI scores than control chars, which likely reflects both underlying migration pressures and the targeted placement of programmes. Because our data are cross-sectional and programme placement is not random, these associations should not be interpreted as definitive evidence that NGO activities increase or decrease migration over time.

More broadly, our results reveal distinct patterns of migration, economic resources, hygiene and education across displaced and non-displaced households and between chars with and without NGO interventions. Conceptually, the integrated use of health, migration and socio-economic indices illustrates how different dimensions of vulnerability and resilience interact in a climate-vulnerable riverine environment. However, the quantitative estimates are directly applicable to the char context studied here and to similar settings with comparable hazards and programmatic conditions; they should be extrapolated to other displacement-prone areas in South Asia only with caution and appropriate empirical validation.

Future research should build on this work by collecting longitudinal data that track the same households over time and by comparing multiple types of displacement-prone environments across South Asia. Such studies would allow a more precise estimation of dynamic causal effects of displacement and interventions on health and migration, and would help to design adaptive strategies that are tailored to diverse local contexts. Within these limitations, our findings provide useful guidance for policymakers and practitioners seeking to reduce disease burden and support household resilience in Bangladesh’s chars.