The Impacts of Repeated Disasters on School Attendance and Learning

REPEATED DISASTER EXPOSURE AND HETEROGENEOUS SHOCKS

Both globally and within the US, people increasingly live in places prone to repeated flooding. A nascent body of research explores how families often move to a neighborhood because of institutions and social networks and are reticent to move away, even when it means placing themselves at risk for future floods. Residents often have strong attachments to schools and wish to see their children graduate from neighborhood schools, compelling them to renovate flooded homes and move back in (Kimbro 2021; Rhodes and Besbris 2022). The National Flood Insurance Program—the only flood insurance option available to US residents—has largely been unsuccessful at moving people away from floodplains due to a lack of political appetite and uneven local implementation (Elliott 2021). As a result, the problem worsens: in North Carolina, from 2006 to 2017, new construction in floodplains has outpaced floodplain buyouts by a ten to one margin (Hino et al. 2024). Moreover, storms sometimes offer opportunities for new development, increasing population density in devastated regions (Ellen et al. 2026, this issue; Pais and Elliott 2008).

As disasters become more commonplace and larger populations live in risky areas, a literature has emerged classifying how disasters overlap and interact with each other. For instance, scholars have defined the ways in which multiple disasters co-occur: consecutive disasters refer to disasters that occur in the same spatial region within a period where full recovery from the first event has not been achieved. These include compound disasters (two or more extreme events that occur simultaneously or successively) (Leppold et al. 2022); cascading disasters (events that occur in a chain or in reaction to each other, such as the 2011 earthquake in Japan, which triggered a tsunami and then a nuclear disaster); and recurring disasters (events that occur in the same region within a one-year period). These classifications lay the groundwork for research on heterogeneity in the effects of each disaster within a set of consecutive disasters.

A related literature has emerged on heterogeneity in the impacts of disruptive events on children. Studies on disruptive events show variation in the effects of shocks such as job loss, divorce, and disasters across socioeconomic status, within-group normativity, and other dimensions (Aquino et al. 2022). For instance, in many cases socioeconomic status is protective against negative impacts. In other cases, however, unexpected events create greater negative consequences, sometimes meaning that more advantaged children, who face low risks of experiencing these events, experience larger effects (Torche et al. 2024). The time frame of recoveries varies widely, with some occurring over days or months while others occur over years or decades. Features such as age and institutional support help to explain some of these heterogeneous effects. While heterogeneity has been important to understanding shocks in various contexts, there is little research examining how disaster repetition may contribute to heterogeneous impacts among children. In other words, we are missing evidence on differences in disaster impacts for children across levels of prior exposure.

CUMULATIVE OR SATURATED DISADVANTAGES?

Two theoretical approaches—cumulative disadvantage and disadvantage saturation—yield contrasting predictions for how children may be affected by repeated climate disasters. Cumulative disadvantage emerged out of Merton’s theory on academic success but has grown to cover impacts of multiple types of social (dis)advantages, including socioeconomic, racial, and spatial (dis)advantages (Diprete and Eirich 2006; Merton 1968). Thomas A. Diprete and Gregory M. Eirich (2006) note that cumulative (dis)advantage broadly refers to temporal processes in which relative benefits in one period are associated with further relative benefits in following periods. This has been called the “Matthew Effect” (Merton 1968). Here, I apply cumulative disadvantage to the impacts of consecutive disasters. Hurricane flooding is associated with material losses, socio-emotional challenges, disruptions to social networks and community, and other negative impacts which may compound. The cumulative disadvantage approach suggests that the material, socio-emotional, and social burdens created by the first disaster would place children in situations where they are less prepared for the second disaster than they would be otherwise.

Studies have found compounding disadvantages when it comes to disasters’ effects on health and finances. For example, a growing body of literature examines cumulative disasters in health contexts, and while the results are not uniform, the studies largely show that “multiple exposures were associated with increased risks to mental health (a cumulative effect)” (Leppold et al. 2022, e276). Micah B. Hahn and colleagues (2022) find that multiple disasters “compound” for outcomes such as self-reported poor mental health, high blood pressure, and asthma, with each additional disaster leading to risk ratio increases of 1–2 percentage points in these categories. Analyzing interviews of mother-daughter pairs in Louisiana who experienced multiple disasters, Lubna Mohammad and Lori Peek (2019) note that a lack of socioeconomic resources can lead to problems that “pile up” with each successive event. As a result of these piling burdens, the second disaster may bring about larger negative impacts than the first disaster.

Disadvantage saturation theory, on the other hand, predicts that existing burdens will dull the impacts of other disadvantages (Hannon 2003; Pinchak and Swisher 2022). Two mechanisms could plausibly explain saturation. First, the impacts of the first disaster may diminish societal advantages for affected groups, thus reducing the possible negative effects of a second event. In this vein, Christina J. Cross (2020) finds that preexisting socioeconomic disadvantages mitigate the stress imposed on children by parental absence, and Lance Hannon (2003) finds that delinquency has a smaller association with educational attainment for disadvantaged children. Put another way, the relationship between certain shocks and academic success may be stronger among advantaged children, for whom the shock leads to a larger relative difference in resources. Florencia Torche and colleagues (2024, 4) note that “as a negatively assessed event becomes more prevalent and normative in society, the stigma associated with it becomes less severe because the event represents a smaller deviation from the social norm.” Such findings suggest that students who are affected by the first disaster would feel smaller burdens from the second disaster.

Saturation could also occur because one disaster prompts institutions or individuals to prepare for similar situations. Regions hit by two disasters may have increased opportunities for preparation, as the first event provides a “window of opportunity for agencies wishing to enhance preparedness” (McClure et al. 2016, 192). At the individual level, Rachel Tolbert Kimbro (2021) calls this “flood capital,” in which people who initially did not imagine that their homes would flood are spurred to develop contingency plans and knowledge about saving personal belongings, evacuating, and navigating insurance or disaster relief agencies. Cheongil Kim and colleagues (2024) find that direct experience with an earthquake is associated with substantial gains in emergency preparedness and insurance uptake. These results highlight the plausibility of one disaster spurring precautionary actions that then mitigate the impacts of a second event.

PATHWAYS CONNECTING STORMS TO DISADVANTAGES

To evaluate the impacts of repetitive hazards, I focus on attendance and learning, which are important educational outcomes and consequential for school operations, governance, and finance. In the US, states set mandatory requirements for instructional days and attach funding to these requirements; schools that experience drops in attendance may experience diminished capacity to hire teachers or fund other school operations (Schlemmer 2020). Attendance is vital to school operation and academic activities and makes other school services possible: when students are in attendance, schools can provide meal services and extracurricular activities. In the wake of disasters, schools can appeal for exceptions to instructional day requirements when they are unlikely to reach the number of required days. In this way, schools may opt to offer fewer instructional days because of the disaster. Between 2011 and 2019, as many as 75 million student days may have been lost due to climate disasters and weather (Jahan et al. 2022).

I operationalize learning as statewide tests. Unlike indicators such as graduation, which might be subject to looser standards during periods of crisis (Harris et al. 2024), statewide tests capture rigid learning benchmarks that cannot be adjusted at the school or district levels. Past research in the US has focused on the impacts of Hurricane Katrina on learning and achievement and has estimated small negative impacts or even some benefits (Kousky 2016). For instance, Michael E. Ward and colleagues (2008) find displaced students received more disciplinary actions following the storm but higher levels of promotion. Bruce Sacerdote (2012) finds academic achievement increases among students displaced to higher-performing schools, and Douglas N. Harris and Matthew F. Larsen (2023) find academic benefits for students who returned to New Orleans. In the sections that follow, I explore pathways through which repeated floods may affect student learning and attendance.

DATA

To address the research questions, I analyze longitudinal student records from the North Carolina Education Research Data Center, which maintains a comprehensive collection of K–12 student information including details on residential blocks, enrollments, demographics, transfers, attendance, and test scores from 2014 to 2022. North Carolina contains several of the highest-ranking counties both in terms of population growth and hurricane flooding risks (Bhatia 2023; Pew Charitable Trusts 2017), placing the region at the forefront of climate challenges in the US. Two major storms, Hurricanes Matthew and Florence, hit regions of Eastern North Carolina in 2016 and 2018, respectively, with many people affected by both, but also each individually (Fuller and Davis 2021).

In the school sample, I include all public schools in North Carolina that were operational during the entire study period (2,929 schools). In the student sample, I focus on a cohort of students in grades K–9 during the 2015–2016 school year. These students were typically in grades 1–10 by the 2016–2017 school year (when Hurricane Matthew occurred) and in grades 3 or higher by the 2018–2019 school year (when Hurricane Florence occurred). I limit the student sample to students with sufficient information on the outcomes of interest (enrollment days, absences, reading and mathematics test scores), within districts that have sufficient residential information.² All together, these restrictions yield a sample of 544,118 unique students, or roughly 34 percent of all K–12 students in North Carolina’s public school system during a given year. Students who move schools or residences are included in the sample, but if students exit the sample districts during the study period, they are excluded. While uneven attrition is a plausible concern, exits are relatively stable across time, representing roughly 11 percent of students and 1 percent of schools each year, which are similar to rates reported by Morrill and Westall (2023). Online appendix A provides further information on attrition for schools and students.³

Measuring School Closures

I gather information on school closures from the Prolonged Unplanned School Closures (PUSC) dataset (Jahan et al. 2022), which contains closure and reopening dates for closures lasting at least five school days in the US. Figure 1 shows the spatial distribution of prolonged closure days for Hurricanes Matthew and Florence. The data mostly report district-level closures but also include school-level closures in the case that individual schools within districts operated differently.⁵ For Hurricanes Matthew and Florence, all recorded closures were district-wide. In online appendix B, I compare the prolonged closures information with data from the North Carolina Department of Public Instruction, which collected information on school-level days missed from Hurricane Florence (the number of total missed days minus those that were made up). The PUSC data capture 97 percent of prolonged closures reported to the state, and I make one correction to the data.⁶

• Download figure
• Open in new tab
• Download PowerPoint

Figure 1.

Prolonged School Closures, Hurricanes Matthew and Florence

Source: Jahan et al. 2022.

Measuring Flood Impacts

During Hurricanes Matthew and Florence, most storm damage was caused by flooding rather than wind. The National Oceanic and Atmospheric Administration notes that “most of the structural damage caused by Matthew’s wind was described as minor, which is a stark contrast to the moderate to severe structural damage that was associated with the storm surge” (National Weather Service 2017). Similarly, the flood damage from Hurricane Florence “caused more than 10 times as much loss as the damage from wind” (Grzadkowska 2019). These factors mitigate concerns about multiple treatments (flooding and wind damages) potentially violating the stable unit treatment value assumption that causal inferences require (Rubin 1974).

I identify impacts of storms on residential blocks by spatially joining block-level student residential information with detailed flood inundation information from First Street Foundation on the two storms of interest (Wing et al. 2018). For each block, I use flood inundation information and a structure inventory from the North Carolina Building Footprint Database (State of North Carolina Emergency Management 2021) to calculate the proportion of residential structures where inundation levels exceeded 20 cm, following Kevin T. Smiley and colleagues (2022, 8), who note that flood levels below this threshold “are unlikely to cause much damage from surface water or pluvial flooding.” The proportion of residential buildings flooded is positively correlated with average flood depths (ρ = 0.72 for both Matthew and Florence, separately) but is less likely to be skewed by extreme values.

Figure 2 shows residential block flooding across the state. To capture different levels of flood exposure, I consider two treatments: whether students experienced flooding for at least 20 percent and 50 percent of residential structures in their block, respectively. Alternative flood thresholds (10 percent, 80 percent) and an alternative measure of inundation (Schaffer-Smith et al. 2020), which uses a binary measure of “flooded” or “not flooded,” are presented in the online appendix.

• Download figure
• Open in new tab
• Download PowerPoint

Figure 2.

Block-Level Flooding, Hurricanes Matthew and Florence

Source: First Street Foundation.

METHODS

Several methodological challenges make it difficult to causally differentiate the effects of repeated hazards. Disasters can have effects that last several years (see Gibbs et al. 2019 for an example in the context of children and disasters), possibly interacting with the influences of subsequent events. Brian Y. An and colleagues (2026, this issue) describe problems related to how researchers define each disaster’s influence, which include indeterminate length, temporal overlap, and units switching in and out of treatments. If researchers define the influence narrowly, they may neglect lagged effects; if researchers define influences as permanent (a “staggered” approach), they may neglect consecutive events. The omission of lagged effects from events that occurred before the study period could also be problematic. These challenges lead Ram Krishna Mazumder and colleagues (2023, 90) to lament researchers’ “inability to analyze the impacts of multiple disaster events similar to the sequential events of Hurricanes Matthew and Florence and followed by the COVID-19 pandemic.”

However, when the number of events is small, a dynamic treatment design offers a viable approach for estimating effects of consecutive treatments. Under the dynamic treatment design, the sample is split into subsamples, each of which experience a different treatment sequence (such as unaffected, affected by Matthew only, affected by Florence only, and affected by both events). Table 1 shows the number of observations for each treatment path under different treatment specifications. I use such a design to measure repeated flooding impacts.

Before analyzing the impacts of closures on students, I note that schools and districts may respond differently, in terms of closure days, to singular and repeated flooding events. If they do, this could confound the estimation of repeated closures on student outcomes. Therefore, in online appendix C, I investigate the relationship between prior flooding and closure days. Because closures occur at the district level, the number of observations in each treatment path is small; however, I find no significant effect of prior flooding on closure length during Hurricane Florence. In other words, I find no differences in closure days by whether the district was previously flooded. I therefore proceed with the estimation of closures’ impacts on student outcomes.

To measure the relationship between prolonged closures and student outcomes, I focus on the school level, where a path-specific event study is appealing given the completeness of each school’s data over long time periods. Clément de Chaisemartin and Xavier D’Haultfœuille (2024) note that when the total number of treatment paths (in this case, four) is low relative to the number of groups (students), there may be sufficient information to estimate path-specific effects, rather than overall effects. For each of the four treatment groups (affected by Hurricane Matthew, affected by Hurricane Florence, affected by both, and affected by neither), this approach allows for estimation of treatment leads and lags. This event study approach is also less restrictive in terms of identifying possible confounders (because fixed effects net out unobservable characteristics), and relies on the parallel trends and no anticipation assumptions, which can be examined using placebos. The general event study model is:

Where α_j is a school fixed effect, γ_t is a time fixed effect, and D_jt is an indicator for school j’s treatment status at time t. K refers to a set of pre- and post-event periods such that β_k is either a placebo estimate or a treatment effect estimate. I use the modified event study model detailed by de Chaisemartin and D’Haultfœuille (2024), which is robust to bias when treatment effects are heterogeneous across groups or time periods.

I next estimate the impacts of residential flooding on students. At the student level, a straightforward estimation of dynamic treatment effects is possible under the assumption that treatment selection in each period is unconfounded, conditional on time-varying covariates (which may include past outcomes) and treatment histories. This assumption may appear problematic, as flooding is not random: proximity to flood-prone regions is strongly associated with a milieu of socio-demographic characteristics. However, the path of any storm is highly variable until shortly before landfall, and conditional on flood risks, it is difficult to anticipate storm flooding.

I use a combination of propensity score matching and Covariate Balancing Propensity Score weighting (Imai and Ratkovic 2014, 2015) to adjust samples for pretreatment differences and create comparable control and treatment groups. I include estimations of residential flood risks,⁷ official floodplain boundaries, as well as social and demographic information and past outcomes (including absences, enrollment days, and reading and mathematics test scores) in the propensity score matching. Within each school, I match treated students with up to five non-treated students. I use a limited set of variables, including floodplain boundaries, demographic information, and past outcomes, to weight the sample so that treatment groups are balanced across covariates. I then estimate effects with OLS using the following linear model:

Where Y_id is the outcome of interest for student i in district d, α_d is a district level fixed effect to control for differences in school closures and other district characteristics, Matthew_id is an indicator for whether the student experienced flooding during Hurricane Matthew and Florence_id is an indicator for whether the student experienced flooding during Hurricane Florence. Covariates, X_id, are the same as those used in weighting. I also report an unweighted version of Equation 2 using only covariates from 2016. Such a comparison is useful in assessing potential posttreatment bias (Blackwell 2013). Observations are grouped according to sequences of treatments at two time periods, 2016 and 2018, and outcomes are measured in the final period.⁸

RESULTS

Descriptive Findings

Table 2 presents geographic and demographic information on students and schools impacted by each sequence of treatments. Unsurprisingly, students affected by the storms lived in more flood-prone places than unaffected students. This is especially clear comparing students affected by both storms (for whom 2.0 percent live in low flood risk blocks) to unaffected students (84.8 percent of whom live in low flood risk blocks). Still, on average, students affected by both storms lived in blocks where only 36.9 percent of residences were within the 100-year floodplain, aligning with reports of substantial flooding occurring outside of official flood zones (National Weather Service 2017). Schools were mostly in low-flood risk areas. While there is often little official legislation preventing schools from being sited in flood zones (Pew Charitable Trusts 2017), few schools were located within the floodplains (ranging from 1.5 percent to 7.7 percent across treated groups), suggesting that local authorities often successfully located the school buildings in less risky locations.

View this table:

• View inline
• View popup
• Download powerpoint

Table 2.

Student and School Characteristics by Flooding Exposure

Several relationships between demographics and flooding exposure are noteworthy. Compared to students affected by Hurricane Florence alone, students affected by Hurricane Matthew alone tended to be disproportionately Native (29.6 percent versus 3.8 percent), Black (40.4 percent versus 29.2 percent), and economically disadvantaged (79.5 percent versus 59.7 percent). Students impacted by Florence alone are ethnoracially similar, but more disadvantaged than unaffected students (who are 51.8 percent economically disadvantaged), while students affected by both hurricanes fall between those affected by Matthew and Florence alone. Schools follow a similar pattern, with Matthew-impacted schools consisting of 56.9 percent Black students, compared to 22.8 percent for unaffected schools. There are also differences in achievement outcomes across treatment groups. Students affected by Matthew and by both storms scored roughly half a standard deviation below average on reading assessments (0.42 for both), whereas students affected by Florence alone scored only 0.09 standard deviations below average. Similar differences are present for mathematics assessments. Schools affected by prolonged closures from Florence alone also scored better than schools affected by Matthew or both storms in reading. Together, these differences highlight how various demographic and geographic factors are associated with flooding exposure across space. While researchers have noted the disproportionate risks that Black residents face in the US South, less attention has been given to flood risks for Native residents. As Max Besbris and colleagues (2026, this issue) note, residential segregation creates uneven exposures to climate hazards, but economic and political mechanisms can reinforce inequalities in disaster recoveries.

Figures 3 and 4 show school attendance and test scores, among the restricted sample, for the school years 2013–2014 through 2021–2022. For students in schools that experienced prolonged closures, there are large drops in enrollment days (ranging between five and nine days) and small rises in absences (less than one day) in the years immediately following the events. Schools affected by Matthew and by both storms experienced large increases in absenteeism during 2021 (rising from ten to twenty-one and ten to eighteen, respectively). This surge in absenteeism is partially—but not fully—explained by a broader trend of rising absences during the COVID-19 period, which has been noticed by education researchers (Dee 2024), and appears for the unaffected group (rising from ten to fifteen absences between the 2019 and 2022 school years). Figure 4 also shows that schools affected by Matthew experienced small declines in reading and mathematics test scores immediately following the hurricanes (0.03 and 0.05 standard deviations), while schools affected by Florence and by both storms saw dips in 2019 mathematics scores (both 0.04 standard deviations).

• Download figure
• Open in new tab
• Download PowerPoint

Figure 3.

Average Enrollments and Absences over Time, by School Closure Status

Source: Restricted-access NCERDC data, 2014–2022.

Note: Data are missing for the 2020 school year. “Affected” refers to students whose schools experienced a prolonged closure.

• Download figure
• Open in new tab
• Download PowerPoint

Figure 4.

Average Test Scores over Time, by School Closure Status

Source: Restricted-access NCERDC data, 2014–2022.

Note: Data are missing for the 2020 school year. “Affected” refers to students whose schools experienced a prolonged closure.

Pathway 1: Impacts of Closure Days

Turning to the event study analysis, figure 5 reports attendance and test score estimates for schools across time. I first note the placebo estimates, which assess differences in trends prior to the year of each group’s first hurricane (for example, years 2015, 2016, and 2017 for the Florence affected group). The goal of these placebo estimates is to test the plausibility of the parallel trends assumption. The placebo estimates (shown in online appendix D) are mostly null or very small, indicating a lack of meaningful pre-event differences in trends between unaffected and affected groups of schools. I therefore proceed to the post-event estimates.

• Download figure
• Open in new tab
• Download PowerPoint

Figure 5.

Effect Estimates of Prolonged Closures on Attendance and Learning over Time

Source: Restricted-access NCERDC data, 2014–2022.

The effects of prolonged school closures on enrollment days range between four and ten lost days. Longer-term impacts of the storms are also apparent: for example, schools affected by Hurricane Matthew experienced losses of five school days in the 2016–2017 school year, followed by additional losses of three days in the 2017–2018 school year, and four days in the 2018–2019 school year. Matthew-affected schools also experienced later rises in absences (5.2 days) during the 2020–2021 school year. Effects on test scores are smaller, but schools affected by Matthew and both storms experience losses in mathematics achievement (on the scale of 0–0.09 standard deviations) and schools affected by both storms experience losses in reading achievement (0.06 standard deviations) in the 2020–2021 school year (although in the same year, schools affected by Florence alone see gains of 0.05 standard deviations). Overall, there are slightly larger negative impacts for students who are affected by both storms than for singularly affected students, although Matthew-affected students and schools experience sizable attendance declines during Hurricane Florence. The impacts of school flooding, shown in online appendix C, are smaller and limited to enrollment days but generally mirror the prolonged closure results.

Figure 6 reports school effects across baseline (2016) achievement levels. Schools with pre-storm low and mid achievement levels (less than −0.5 and between −0.5 and 0.5, respectively, on a student standardized scale) show similar effects to the overall trends, with losses in enrollment days, gains in absences during COVID-19, and some small losses in mathematics test scores. Mid-achieving schools impacted by Matthew also show increases in absences in the 2021–2022 school year. However, high-achieving Matthew-impacted schools show little or no negative impacts in the years following the storm, but rather, experience long-run gains in enrollment days, losses in absences, and gains in reading and mathematics test scores. These surprising gains, which culminate with six additional enrollment days, losses of three absence days, and 0.2 and 0.3 standard deviation increases in reading and mathematics scores in the 2021–2022 school year, occur while the rest of the Matthew-impacted schools experienced negative or null impacts on the same outcomes.

• Download figure
• Open in new tab
• Download PowerPoint

Figure 6.

Effect Estimates of Prolonged Closures on Attendance and Learning over Time, by Test Scores

Source: Restricted-access NCERDC data, 2014–2022.

Note: “Low,” “Mid,” and “High” refer to schools with average 2016 test scores below −0.5, between −0.5 and 0.5, and above 0.5 (measured in standard deviations).

Pathway 2: Impacts of Residential Block Flooding, Net of School District Closures

Figure 7 presents estimates from weighted linear models, which account for selectivity in treatment and school features, to test for differences between students in flooded and non-flooded residential blocks. Beginning with students in blocks with at least 20 percent flooding, one observes that net of school districts, each of the treatment sequences is mostly unassociated with attendance days and absences, although students impacted by Matthew gained 0.5 school enrollment days and 0.4 absence days, during the 2018–2019 school year. For other treatment groups, and for students affected by flooding at the 50 percent residential block level, there are no significant effects on reading or mathematics test scores.

• Download figure
• Open in new tab
• Download PowerPoint

Figure 7.

Effect Estimates of Residential Flooding

Source: Restricted-access NCERDC data, 2014–2022.

The results also show losses in mathematics scores—but not reading scores—for some students in flooded blocks. Students affected by Hurricane Florence at the 20 percent residential block flooding level see declines of 0.07 standard deviations in mathematics scores. Unweighted models, reported in online appendix E, also point to small declines in mathematics for students affected by Matthew and by both storms. I also report results for student models in 2019, 2021, and 2022 separately. Except for increases in absences for Matthew-affected students, impacts were concentrated in 2019.

In short, these results suggest that residential block flooding had little impact on student outcomes, beyond the district-level effects of the storms, although some affected students experienced further declines in mathematics scores. There are few differences, in terms of attendance or test score impacts, between students who experienced singular flooding as opposed to consecutive residential flooding. In online appendix E, I investigate heterogeneity in student-level models across student characteristics (poverty status, ethnoracial status, and achievement). With a couple of exceptions, the effects are not statistically different across subgroups.⁹ It is possible that differences in residential information completeness led to confounding in the sample of students affected by residential flooding. To investigate this possibility, I also run the student analyses for districts with at least 85 percent residential completeness, and report results that are very similar to those in figure 7.

DISCUSSION AND CONCLUSION

As children inside and outside the US face increasing flooding risks, the successive hurricane flooding events in North Carolina between 2016 and 2018 provide a useful case for what the climate-changed future may look like. This article investigates the effects of these consecutive hazards on educational outcomes. Leveraging a dynamic treatment design, I estimate the effects of singular and repeated hurricane flooding through two pathways: school closures and residential flooding. I find that closures resulted in large declines in attendance, which were similar each time that schools were affected. On average, schools that experienced prolonged closures during Hurricane Florence singularly lost about ten enrollment days in the period from 2017–2022, whereas schools impacted by Hurricane Matthew lost about fifteen enrollment days and schools impacted by both storms lost about twenty-one enrollment days during the same period. The prolonged closures—whether singular or repeated—were related to long-run increases in absences and, for some schools, declines in test scores. Net of school closures, students who lived in residential blocks where flooding occurred experienced little or no negative impacts on test scores, which were similar each time the students were affected.

Overall, the results align with a middle of the road scenario where prior exposure to flooding or school closure neither heightens nor dulls the impacts of a second event. In other words, the storms caused disadvantages, mostly through lost school days, but I find no evidence that disadvantages from a second exposure are substantially different than those from initial exposure. Similarly, An and colleagues (2026) find that the effects of multiple disasters on rent hikes and vacancy are similar in scale for each successive storm. Even if the effects of a second storm are close in magnitude to the first, these can still lead to cumulative disadvantages if some groups are disproportionately affected by the hazards.

Some schools affected by Hurricane Matthew alone also experienced losses in enrollment days, higher absences, and lower mathematics scores during Hurricane Florence. Transportation—often schools’ “biggest challenge” in these circumstances (Schlemmer 2018)—could play a role in these effects if challenges in staffing and rerouting buses affected otherwise unaffected districts. In online appendix F, I investigate average school bus route times for Matthew-affected districts. Most of these districts had long bus routes relative to the state average, but the data are unable to provide decisive evidence on whether Florence-related transit disruptions resulted in spillover effects in Matthew-affected districts.

The COVID-19 pandemic introduces a third disruption which, to some extent, all students and all schools experienced.¹⁰ Figure 6 illustrates that during the pandemic, cumulative disadvantages fell on students and schools that were still recovering from prior events—for example, schools impacted by both flooding events appear to experience worse long-run trends than unaffected schools, including increases in absences and decreases in mathematics and reading test scores. The negative outcomes during COVID-19 could be the result of sustained disaster impacts: research has found long-lasting effects of disasters on students’ education (Deuchert and Felfe 2015; Gibbs et al. 2019) and increases in mental health support services needed by both teachers and students (Davis et al. 2021; Kousky 2016).

A small group of Matthew-affected schools that were already high performing appear to benefit from the storm sequence. These schools are all located in a single district (Pitt County Schools), and are less poor (30.7 percent), and Whiter (62.0 percent), on average, than the rest of the sample. Sometimes, due to insurance processes and relief programs, more affluent households benefit during the disaster recovery period (Rhodes and Besbris 2022; Howell and Elliott 2019). Here, one observes that some schools that were already advantaged were able to increase academic advantages, with better performance in terms of days enrolled, absences, and reading and math test scores. Relative deprivation may facilitate some of these gains: perhaps because disadvantaged students were displaced during the storms or because their peers at other schools fell behind, other students at these schools were able to excel. However, these students represent a small portion of the overall sample, and these academic gains could be unrelated to the storms.

Several limitations are noteworthy. The student-level model relies on the assumption of unconfounded treatment, given time-varying covariates. While I include relevant variables at multiple levels (a full list of covariates is available in online appendix G), unobserved variables could have created confounding here. Similarly, the event study models show some placebo effects, violating the no anticipation restriction. While these effects are small relative to the main effects, endogeneity in treatment groups may not be fully accounted for. Missingness in district residential reporting also limits the measurement of residential flood impacts. I provide a supplemental analysis of districts with at least 85 percent completeness in online appendix E, but beyond this threshold, the number of treated units becomes too small for covariate balancing. While attrition rates for schools and students were similar between affected and unaffected groups, storm-induced school closures or student exits may have changed the composition of the sample. In both cases, however, one would expect effect estimates to be biased toward zero due to these exits. Flood exposure estimates also may contain errors. In online appendix H, I use a secondary source of flooding estimates (Schaffer-Smith et al. 2020) to reproduce the main student effect estimates, which are similar but smaller. Lastly, there are several mechanisms by which storms could plausibly affect students that I do not test for in this article. For instance, storms have immediate and long-term consequences for various labor sectors (Francis and Horn 2026, this issue), which could influence teachers as well as students’ home lives. Storm-induced preparations for future events and cumulative burdens may also offset each other. Such mechanisms are outside the scope of this analysis but could be explored in future studies.

Limitations notwithstanding, the article extends previous research on consecutive disasters in several ways. First, this article offers a longitudinal analysis on consecutive disaster impacts for a population of children with minimal selectivity. Existing work has mostly studied the impacts of singular disasters in the school context, and little is known about how children experience cumulative shocks. Research has shown that disasters can have long-lasting impacts on educational outcomes (Gibbs et al. 2019; Deuchert and Felfe 2015); by analyzing students across a period of several years, I am able to disentangle the effects of singular and cumulative disasters. Second, I measure flood impacts at both the residential and school levels, allowing for more detailed exploration of mechanisms than previous work, which has mostly identified hurricane impacts in schools and districts (Sacerdote 2012; Fuller and Davis 2021).

One policy-relevant finding from this study is that school closures, rather than residential impacts, appear to be the main source of hurricane-related losses in attendance and learning. Like prior literature on community-level trauma (Trejo et al. 2024; Goldmann and Galea 2014), I find that disaster impacts are not limited to those who experience direct residential exposure. While more than 9,000 students experienced residential flooding (at the 20 percent residential block level), over 395,000 students experienced a prolonged school closure during the study period. These results beg the question: how many of these prolonged closures could have been avoided or shortened? Because the district closures affected so many students, and closure days appear to be responsible for much of the storms’ academic impacts, policies that prepare districts to safely reopen quickly following hurricanes would likely benefit students and families. For schools and governments to increase their preparedness strategies, research on municipalities suggests that low-cost incentives may be useful (Brody et al. 2009).

As children in the twenty-first century experience disasters with greater frequency, questions around how consecutive hazards impact students become increasingly relevant. Future generations affected by multiple storms may no longer be “exposure outliers” (Mohammad and Peek 2019), and further research is needed to fully explore the consequences of these repeated exposures. Disasters may occur in unexpected places, and a lack of expectation may contribute to the disaster itself (Quarantelli 1998). A community member who experienced flooding during Hurricane Matthew described anxiety during subsequent storms, saying, “I believe that anyplace can flood with enough water” (PBS North Carolina 2017). Given updates in flooding projections (Wing et al. 2018), large swaths of the US landscape could indeed be flooded.

While schools are positioned to be sites of resilience and protection during extreme weather events, this article shows that many are not prepared for the repeated disasters that a climate-changed future will bring. Schools are community assets and often a reason that families return and rebuild after being displaced by a natural disaster: Kimbro (2021, 208) argues that “the school must be viewed as a core component of family life, a key neighborhood institution that threaded itself through the families in ways that increased social cohesion as well as neighborhood attachment.” In Hurricanes Matthew and Florence, as well as in the recent Hurricane Helene, schools were some of the first organizations to coordinate information about relief services, water and material supplies, and charging stations. Because schools are sites of refuge and information, and often the first places that people go in times of crisis, it is especially important that we make them more resilient to future singular and repeated flooding events.