Societal determinants of flood-induced displacement

Flood disasters exhibit distinct trends across space and time. Statistics by the International Disaster Database, EM-DAT (19), reveal that flood disasters are especially prominent in South and East Asia, whereas slightly populated polar and dry areas are largely exempt from severe flood risk (Fig. 1A).^* Over the past 50 y, the global frequency of floods has risen steeply, from an annual average of 25.8 flood disasters in the 1970s to 153.9 disaster events per year on average since 2011 (Fig. 1B, blue area). Anthropogenic climate change constitutes a key driver of this growth, with increasing human exposure, land use changes, and improved quality of reporting over time representing other major explanations (20, 21).

In marked contrast to the global surge in flood events, the severity of floods in terms of lives lost is declining. Over the past half century, global average mortality per flood event dropped by nearly 90% (Fig. 1B, orange line), from around 272 casualties per disaster in the 1970s (258 floods totaling 70,215 deaths) to 31.5 deaths per event in the last decade (1,539 floods; 48,477 deaths). In the same period, the global population nearly doubled, and growth has been especially high in hazard-prone regions (20).

The remarkable decline in the human cost of floods is evidence of major progress in flood management and disaster risk reduction in high-exposure areas (26–29). Fig. 1C documents a negative relationship between level of development and flood mortality. Consistent with earlier findings, it reveals that the deadliest floods concentrate in the lower income deciles, especially in the first half of the period. Average flood mortality for the 10% wealthiest countries remains comparatively low throughout this period.

Mortality rate is not the only relevant metric of flood impact. For every fatality, almost one thousand persons are displaced by flooding according to Dartmouth Flood Observatory (DFO) (25), the leading provider of flood-related statistics. A displaced person in this context is someone who is forced to leave their home or place of residence as a result of disaster (30). Flood displacement can occur both reactively and preemptively, for example as planned evacuation in anticipation of an advancing storm. However, definitions and measurement practices vary (31), implying that available statistic are prone to measurement uncertainty.

Fig. 1D displays global trends in floods and displacements since 1985 as recorded by the DFO. Consistent with EM-DAT, the DFO data reveal a peak in flood occurrence in the early 2000s followed by a decline, and the downward trend in displacements matches the falling flood mortality shown in Fig. 1B. The contour plot (Fig. 1E) reveals that large-scale displacement events concentrate in lower income groups, although the clustering pattern is slightly weaker than for mortality (Fig. 1C). However, global average patterns disguise considerable heterogeneity in flood impact across events, between regions, and over time. What explains this variation remains understudied.

In order to advance scientific understanding of the conditions under which floods generate widespread human displacement, we draw on the tenets of coupled human and natural systems (32–34). Consistent with this framework, flood-induced risk is conceptualized as a product of complex interactions between natural hazards, human exposure, and societal determinants of vulnerability and resilience that are experienced and play out differently across social scales, from the individual to the nation-state at large. Our analysis focuses on the magnitude of human displacement from distinct flood events as a function of hazard exposure while accounting for prevailing societal contexts at the time of disaster.

Fig. 2 provides a simple schematic illustration of how these features are related to each other. People directly exposed to flooding as well as those affected by compounding and cascading hazards jointly comprise the pool of latent movers. The subset of exposed people that are forced to flee to avoid or minimize flood-induced harm are classified as displaced persons. Some exposed people may be unable to relocate and are involuntarily immobile while others may decide to stay even when confronted with imminent danger, for example due to strong attachment to place (35), challenging simplistic notions that natural hazards increase individual-level likelihood of displacement in a mechanistic and universal fashion. Common barriers to mobility in the context of disaster include destruction of transportation networks, political restriction on movement, and adverse security conditions (16, 36, 37).

The ratio of exposed to displaced persons will vary greatly across cases, depending on the nature and severity of the flood event (inundation speed and depth, extent of physical destruction, duration of flooding); the prevalence of compounding and cascading hazards within and beyond the flooded areas (severe windstorm, outbreak of diseases, breakdown of services); the preparedness and resilience of the exposed population (disaster management plans, robustness of built infrastructure, early warning systems), as well as contextual factors that may cushion or amplify mobility outcomes (financial resources, government response, armed conflict).

The abrupt and potentially life-threatening nature of severe floods implies that human agency is less central here than for migration in response to gradual environmental change, although aspects of power and privilege may influence the viability and range of mobility options even in the context of rapid-onset disasters (38–40). On the basis of theoretical and empirical relevance, we consider three core contextual dimensions that are expected to affect aggregate displacement levels for a given level of flood exposure. This approach privileges conditions that are amenable to policy input over terrain- and weather-related moderators of the flood–displacement relationship:

(i) All else equal, socioeconomic development should lower the need for disaster-driven mobility. Higher income levels facilitate investing in more comprehensive adaptation schemes, more robust service provision and infrastructure, and better flood protection systems (41, 42). These features are likely to reduce exposure of people and assets to flooding, but they also enable greater opportunities for in situ coping. Although some studies point to low incomes being a barrier to mobility (43, 44), that applies more to high-agency migration than to rapid-onset disaster displacement, which usually is temporary and occurs over short distances (12).

(ii) Good governance likewise should moderate displacement, although this expectation is less unambiguous. Well-functioning and accountable institutions are associated with better protection of vulnerable groups, inclusive urban and rural planning, and more effective disaster preparedness and response, which jointly increase sustainability and resilience (45–47). A more developed civil society also offers greater opportunities for support from nonstate actors in times of crisis. At the same time, democratic governments are more sensitive to domestic audience costs (48) and therefore may be especially likely to initiate inclusive evacuation operations preemptively, potentially inflating overall displacement numbers in an effort to minimize loss of lives.

(iii) A third key contextual factor is security: Armed conflict is an important driver of displacement in its own right (49), but it erodes coping capacity also among those who do not flee the fighting by obstructing markets and commodity flows, destroying assets and livelihoods, and reducing the state’s ability (and willingness) to deliver services and relief aid. The result can be a vicious circle of environmental vulnerability, political violence, and disaster loss (50–52).

A systematic investigation into the empirical merit of these expectations requires, first of all, detailed data on the spatial extent of inundation for each flood event and associated displacement estimates at global scale. The best available data for this purpose are provided by the Global Flood Database (GFD). The GFD is based on the DFO flood event catalogue and provides high-resolution (250 m) raster data of inundated terrain for 913 severe flood events worldwide, 2000 to 2018, based on Moderate Resolution Imaging Spectroradiometer (MODIS) images from NASA’s Aqua and Terra satellites (53). The GFD data represent a subset of all DFO-recorded floods in this period, since adverse meteorological conditions and complex topography sometimes prevent remote sensing. Because many floods transcend national borders, we create unique event-country observations for all floods (N = 1,682) by superimposing the WorldPop population raster data, which contain country ID for each 100-m grid cell. The WorldPop data compare favorably with other population raster data in fine-scale accuracy applications (54), so we replace the GFD flood exposure variable with our own estimates of the number of persons residing in the flooded areas using these data.

Estimates of the number of displaced persons per flood event are provided by DFO, based mostly on news reports and official country statements. Around ¼ of the GFD events (N = 221) lack verified information on the magnitude of human displacement. These observations are dropped from the analysis. DFO does not offer estimates of involuntarily immobile populations. Lack of systematic information on compounding and cascading hazards prevents assessing indirect flood exposure and implications for disaster mobility here.

There are many ways to operationalize socioeconomic, political, and security contexts. We choose to measure these concepts broadly, relying on indicators that are widely used to assess societal progress at global scale. To increase analytical rigor, we combine country-level information on each contextual dimension with corresponding georeferenced variables tapping the situation in the flooded areas. Socioeconomic development is represented by gross domestic product (GDP) per capita as well as a satellite-based index of average nighttime luminosity in the flooded area as proxy for density of local economic activity and infrastructure. The political context is represented by a country-level index of electoral democracy, supplemented by a measure reflecting whether some or all of the flood-exposed population is formally excluded from participation in national politics. To capture the prevailing security situation, we measure the total number of battle-related deaths (BRD) in the country over the previous 10-y period, as well as the total number of BRDs from armed conflict events in the flooded areas during the previous 6 mo. The socioeconomic, political, and security variables jointly speak to the most salient drivers of societal resilience and comprise common sustainability metrics (55).

Lastly, we include a set of flood-specific variables: the number of people directly exposed to flooding; the duration of the flood in days; and a count of the total number of DFO flood events in the country over the previous 10 y to capture accumulated impacts of repeated disasters.

To assess the extent to which these societal conditions moderate flood-induced displacement, we rely on negative binomial (NB) regression analysis with Bayesian estimation techniques, due to their powerful postestimation diagnostics and flexibility in accommodating overdispersed count data. Variable importance and model performance are assessed via both in-sample (2000 to 2014) prediction and out-of-sample (2015 to 2018) cross-validation. Given the complex interplay of multiple hierarchical factors, data limitations, and likely measurement errors for the outcome variable, this is a more robust approach for identifying relevant factors than standard econometric analysis.

In building the statistical models, we first estimate the base model with flood-specific characteristics only (base), and then sequentially specify more complex models that additionally incorporate the economic (e), political (p), and conflict (c) components separately and in combination (e.g., c-e, c-p, full). A subset of the models that contain all societal dimensions are limited to country-level (clvl) or local-level (llvl) predictors. Some models also consider nonlinear tensor spline interactions (t2) between the two key flood-specific characteristics since the effect of population exposure on displacement may be nonlinearly related to flood duration and vice versa. Lastly, to better account for heterogeneity in flood characteristics and impacts across regions (cf. Fig. 1A), we specify continent-level random intercepts (ri) and random slopes for the flood-specific characteristics (rs) in most models. See Materials and Methods and SI Appendix for further details on measurements, modeling approach, and performance evaluation.

Fig. 3 shows the estimated in-sample conditional effects from the preferred full+ri+rs model on the posterior predictive distribution for each regressor, holding all other variables at their mean (see SI Appendix, Table S7 for regression output). The detected functional forms are mostly in line with expectations. Flood displacement is highly influenced by the number of people exposed; the upper bound of the prediction interval rises steeply with increasing population in the flooded areas, as well as with increasing flood duration and extent of previous flood exposure. The latter pattern indicates that flood displacement often is endemic in regions with recurring extreme weather events (e.g., densely populated coastal zones and flood plains in the tropics), and could reflect lack of improvements in flood protection in many disaster hotspots but also increasingly effective early warning systems and life-saving evacuation routines (56).

The socioeconomic indicators reveal that displacement is inversely related to the country’s level of economic development, although the effect is small in substantive terms. At the local level we find a positive association between nighttime light emission and displacement. We interpret this effect as driven partly by population exposure, given the positive correlation between luminosity and local population density, although it also could reflect superior mobility opportunities and evacuation capacity in more urbanized areas. The political variables suggest that democracies are associated with lower levels of displacement than other political systems; the upper bound of the estimated effect entails a reduction in displacement risk for severe floods by two-thirds by shifting from low to high democracy score, holding other factors constant. We also find that floods that affect excluded populations generate lower displacement on average than those that exclusively affect coethnics of the ruling regime. This result breaks with research that finds ethnopolitical discrimination to be an important driver of vulnerability (57, 58), although it may be reflective of lower de facto risk of inundation from flooding in rugged hinterlands where marginalized minority groups often reside (59). As expected, conflict-related casualties—both at the local level and in recent national history—are positively associated with levels of flood displacement.

Collectively, the upper bounds of the predictions reveal meaningful associations between societal characteristics and flood-induced displacement and underscore the relevance of economic development, inclusive governance, and peace for lowering vulnerability to climate-related hazards. That said, we note that the median predictions remain low across the full range of values for all covariates. This reflects the very large spread in observed displacements, even for seemingly comparable contexts, and is representative of the significant challenge of producing accurate predictions of human mobility with macrolevel models (60, 61).

Fig. 4 shows the calibration of our predictions by comparing the 10th, 50th, and 90th percentiles in-sample prediction based on the full+ri+rs model with the observed displacement numbers for the same percentiles for each continent. Despite large differences in the distributions of flood displacements across continents that our model needs to accommodate, the density of the predictions fall quite close to the observed values for most subsamples. Europe is a clear outlier in these data, with median displacement comprising less than 5% of the median observed in the other continents. The 90th percentile event in Europe (4,319 displaced persons) is dwarfed by that in Asia (300,000 displacements). The full+ri+rs model tends to overpredict outcomes at higher ranges, especially in the Americas, while underpredicting displacement in the least severe floods. Other model specifications documented in SI Appendix (SI Appendix, Figs. S5–S7) perform better for in-sample posterior prediction for some regions but are less successful in predicting displacement levels on new data.

The empirical analysis uncovered reasonable statistical effects of the contextual variables, broadly in line with expectations. Although the wide heterogeneity in observed displacement levels for a given level of exposure challenges accurate prediction, the analysis was able to establish meaningful upper bounds on the magnitude of flood-induced displacement as a function of shifting socioeconomic, political, and security conditions. These results can be used to indicate plausible worst-case outcomes of flooding in various societal contexts. Controlling for the number of people directly exposed as well as the duration of flooding, the upper limit of the displacement prediction is substantially higher when the flood affects conflict-affected areas in nondemocratic, low-income societies. Ongoing efforts embedded in the Sustainable Development Goals (SDG) agenda toward, inter alia, eradicating poverty and strengthening peaceful and inclusive governance thus also are likely to contribute to lessening future human cost of extreme weather events. Even so, out-of-sample tests demonstrated that the societal indicators struggle to improve our ability to accurately forecast flood-induced displacement on new data. We believe there are both conceptual and analytical reasons for the modest out-of-sample predictive performance.

On the conceptual side, we are confronted with challenges related to legal and operational ambiguity; there is no globally agreed-upon definition of what classifies as a displaced person (and what does not) in terms of primary motive for leaving, required distance traversed, minimum duration of relocation, etc. Accordingly, providers of displacement statistics often must navigate inconsistent, conflicting, patchy, or missing information. The displacement estimates in the GFD data that this analysis relies on originate mostly from media reports, leaving it to the interpretation of the journalist to determine whether and to what extent human mobility within contexts of flooding is newsworthy and portrayed as displacement. In addition, media coverage exhibits considerable geographical variation, with better reporting in urban sites and areas with wire services (66). In the absence of verified information, the GFD sometimes uses estimated numbers of people residing in the flooded areas as a surrogate for displacement, although that could both underestimate and overestimate true displacement.

A related source of uncertainty is the quality of available information. In countries without a free and independent press, information about human consequences of disaster may be deliberately withheld or grossly inaccurate, reshaped by political incentives to inflate or deflate the gravity of the situation. Without transparent reporting mechanisms, such biases are hard to ascertain. Improved penetration of digital information, including social media and remote sensing technologies, provides better coverage over time, but comparisons across societies remain challenging. Satellite imageries are no perfect substitute for high-quality reporting on the ground when studying human impacts of natural hazards.

On the analytical side, extreme precipitation events and surges are sometimes compounded by destructive windstorms that force people to flee across much wider areas than the flooded terrain alone would suggest (67, 68). Uncertainty about the exact location and severity of an approaching storm may motivate inclusive early warnings and preemptive evacuation based on worst-case assessments as a precautionary principle. For example, 375,000 citizens in low-lying parts of New York City were ordered mandatory evacuation ahead of the 2012 Hurricane Sandy, where the resulting flooding ended up displacing “tens of thousands” city residents (69). Depending on whether predisaster evacuation and impacts of compounding and cascading hazards are accounted for, the number of displaced persons registered in the DFO catalogue may be many times higher than the number of residents in the flood-affected area.

We can substantiate this discussion by looking deeper into the GFD data. As shown in Fig. 5A, there is a strong, positive correlation between numbers of fatalities and numbers of displacements, but we also detect a significant share of flood events that depart from this general trend. Stacked along the vertical axis are flood events without reported casualties, although some of these had very high displacement figures. Eight nonfatal flood events in the GFD database displaced more than 100,000 persons, three orders of magnitude above the median estimate for nonfatal floods of 161 displacements. Most high-impact events occurred in Asia, but otherwise there is no clear geographical pattern.

Fig. 5B compares numbers of displacements with the estimated number of people directly exposed to flooding. Consistent with the importance of population exposure as a predictor of displacement (Fig. 3), the plot reveals a strong positive association between these variables, although with notable spread in displacement for any given exposure level that partly reflects important regional differences. European floods tend to have comparably small shares of displaced to exposed persons, whereas Africa has the highest average ratio. This could indicate a measurement problem, for example that the georeferenced demographic data are less reliable for informal and rural settlements in developing countries (70), resulting in systematic underestimation of people exposed to flood events in many African regions. However, we also believe this pattern is driven by systematic variation in societal sensitivity to climate hazards that is only partly captured by our predictors. For example, tropical regions may be more exposed to compounding hazards since floods often are related to hurricane or cyclone activity. Although a simple comparison of tropical with nontropical floods in the GFD data failed to reveal a consistent difference in relative displacement outcomes (SI Appendix, Fig. S1), this issue deserves further scrutiny. In vulnerable contexts, a severe flood also is more likely to generate cascading impacts across sectors, such as breakdown of public services and destruction of critical infrastructure. Such knock-on effects might compel people far beyond the flooded areas to relocate in search of humanitarian assistance. Accounting for local variation in vulnerability at global scale requires better information than what is currently available.

Another analytical reason for the weak out-of-sample results is complex temporal patterns. Global average exposure per flood event rose during the first half of the analysis period but has since dropped, whereas displacement numbers have fallen sharply throughout the period. This implies that the displacement to exposure ratio shifts significantly over time and is considerably lower in the test period than in the training data for the out-of-sample cross-validation. Most of the contextual variables tested here represent relatively inert structural features and are as such much less suited to capture short-term changes in flood impacts than identifying configurations of societal factors that are capable of mitigating adverse effects of flood hazards across longer time scales. However, the poor out-of-sample predictive performance is also indicative of recent progress in flood risk reduction that has occurred much more rapidly than sustainable development more broadly defined. Here, technological innovation, knowledge sharing, and multiscale public–private collaboration have played important roles (71–73).

The analysis studies georeferenced floods worldwide, 2000 to 2018, as recorded in the GFD (25). The unit of analysis is each distinct flood-country event, represented by high-resolution raster images of the maximum spatial extent of the inundation area, derived from daily 250 m resolution MODIS satellite data. The dependent variable is the number of displaced persons per flood-country event, derived from the related DFO catalogue. Because the displacement estimates originally are aggregated to each flood event, we split the numbers for multicountry floods such that the displacements in each flood-country event reflect the share of the total population exposed to the flood residing in that country. We calculate those directly exposed to flood inundation based on UN-adjusted top-down constrained settlement data from WorldPop (www.worldpop.org), aggregated from 100-m resolution grid cells, calibrated, and overlaid with the GFD flood data. GFD flood events that are listed with 0 displacements are dropped from analysis since 0 in this case reflects lack of verified information.

The statistical models draw on four sets of thematic components. The base model contains two indicators specific to each flood-country event—the number of persons exposed to flood and the duration of flooding in days—and a third country-level flood experience variable, measuring the 10-y moving count of flood events in the country, based on the DFO data. The socioeconomic component contains a calibrated index of average nighttime luminosity in the flooded area based on satellite imageries (84) and country-level GDP per capita (85). The political component includes a dichotomous measure indicating whether any share of the flood-exposed population is formally excluded from national politics, using geo-referenced data from the Ethnic Power Relations Dataset (86), supplemented by level of democracy in the country, measured using the Varieties of Democracy electoral democracy index (87). The conflict component contains a count of the number of BRD from armed conflict events within 20 km from the flooded areas during the previous 6 mo, and a complementary conflict exposure indicator counting the total number of BRD in the country over the previous 10-y period. Both estimates are calculated from data provided by the Uppsala Conflict Data Program (88). The socioeconomic indicators are measured in the year prior to each flood event, and all variables except political exclusion and democracy score are log-transformed to improve the linear fit and reduce outlier influence.

We are grateful to the coorganizers of the PNAS Special Feature, Neil Adger, Sonja Fransen, and Ricardo Safra de Campos, for encouraging this contribution, and to the Editor and three anonymous reviewers for constructive comments on our manuscript. We also thank participants at the 2022 International Conference on Environmental Peacebuilding, the 2022 International Studies Association Annual Convention, the 2022 Peace Science Society (International) meeting, and the Nordic climate-conflict seminar series for valuable feedback on earlier drafts. This research is supported by the European Research Council through grant no. 101055133 awarded to H.B.