Abstract
This study looks at the impact of exposure to natural disasters during pregnancy on the educational outcomes of North Carolina children at the third grade level. A broad literature relates negative birth outcomes to poor educational performance, and a number of recent studies have examined the effect of prenatal exposure to natural disasters on birth outcomes. This study takes the next step by considering how prenatal exposure affects later outcomes. Combining North Carolina administrative data on births and school performance with disaster declarations from the U.S. Federal Emergency Management Agency (FEMA) allows for the identification of children who were exposed to disasters during prenatal development. These children are compared with other children born in the same county who were not exposed to disasters while in utero. Regression results suggest that children exposed to hurricanes prenatally have lower scores on third grade standardized tests in math and reading. Those exposed to flooding or tornadoes also have somewhat lower math scores. Additionally, results suggest that these negative effects are more concentrated among children in disadvantaged subgroups, especially children born to black mothers. However, no evidence exists that these effects are mediated by common measures of birth outcomes, including birth weight and gestational age.
Introduction
In 2011, the Federal Emergency Management Agency (FEMA 2014) reported 99 major disasters across the United States. The immediate costs of natural disasters both for individuals and communities are substantial. FEMA distributed an average of $3.3 billion in Public Assistance (PA) grants annually from 1999 to 2010 (FEMA 2013). The costs borne by individuals, families, and insurance companies were even higher. The short-term effects of disasters—including injuries, evacuations, prolonged power outages, damage to buildings, and lost days of school and work—are well documented. However, much less is known about the long-term costs of disasters in terms of their effects on families and children.
Natural disasters may be particularly damaging to vulnerable populations, such as developing fetuses. A number of recent studies have found negative effects of prenatal exposure to natural disasters on birth outcomes and complications (Auger et al. 2011; Currie and Rossin-Slater 2013; Dancause et al. 2011; Glynn et al. 2001; Oyarzo et al. 2012; Simeonova 2009; Tan et al. 2009; Torche 2011; Xiong et al. 2008). A natural addition to this literature is to consider how effects of prenatal exposures on birth outcomes translate into later-life outcomes. Researchers, in a study called “Project Ice Storm” that looked at a severe ice storm that occurred in Canada in 1998, did just that by looking at the effect of prenatal exposure on cognitive development at ages 2 and 5 (Laplante et al. 2004, 2008). These researchers found negative effects on cognitive development. However, the generalizability of their results is limited by small sample sizes and the focus on a single disaster at a single point in time.
The current study adds to the literature in several ways. This study looks at the effects of 15 different disasters on the third grade educational outcomes of a large sample of children born in North Carolina over a period of 13 years. Three different categories of disasters are represented, which allows for some exploration of the differences in effects by disaster type. Additionally, the large sample size allows for the consideration of differences in effects across demographic subgroups, especially disadvantaged groups. The focus on educational outcomes at the third grade level also gives a longer-run perspective on the effects of prenatal shocks caused by disasters. This longer-run perspective provides a better estimate of the enduring impact of prenatal exposure to disasters and its cost to society.
Background
The conceptual framework for this study assumes that natural disasters are unpredictable exogenous shocks that may affect the prenatal development of babies who are in utero at the time of the disaster. Through effects on prenatal development, natural disasters may have long-term effects on the educational outcomes of children exposed prenatally. However, these effects may be moderated by the type and timing of disasters as well as by family characteristics.
Natural Disasters and Prenatal Development
Recent interest in the effects of prenatal shocks on birth outcomes has led to a number of studies looking at natural disasters. Most of these studies have examined pregnant women exposed to a particular disaster and compared their birth outcomes with other women who were not exposed because of geography or timing. An early study (Glynn et al. 2001) found that maternal stress resulting from an earthquake in California was associated with lower gestational ages. Later studies of earthquakes in other countries confirmed findings of reduced gestational ages (Torche 2011) as well as finding higher rates of preterm births (Oyarzo et al. 2012; Tan et al. 2009), higher rates of birth defects (Tan et al. 2009), and lower birth weights (Tan et al. 2009; Torche 2011). Similar studies of a severe ice storm in Canada and Hurricane Katrina also found reductions in birth weight and increases in preterm births (Auger et al. 2011; Dancause et al. 2011; Xiong et al. 2008).
In addition to these studies of single events, two studies have looked at average effects across a broader group of disasters. The first (Simeonova 2009) pooled the effects of all weather-related disasters in the United States from 1968 to 1988 and found small negative effects on preterm birth and gestational age. The second (Currie and Rossin-Slater 2013) used sibling fixed effects to examine the effects of several hurricanes in Texas and found an increase in birth complications. These studies, along with the studies of single disasters, raise concern about the effect of prenatal exposure to natural disasters on longer-term outcomes, including education.
Prenatal Development and Educational Outcomes
The relationship between poor health at birth and negative educational outcomes is well documented. Low birth weight and preterm birth are associated with higher risks of significant physical and mental impairments (Goosby and Cheadle 2006, 2009; Reichman 2005; Saigal et al. 1991), which translate into higher levels of special education placements. Newborn health has also been used to predict cognitive skills and achievement test scores (Aarnoudse-Moens et al. 2009; Andreias et al. 2010; Black et al. 2007; Boardman et al. 2002; Goosby and Cheadle 2009; Hack et al. 2002; Reichman 2005). Other school-related outcomes are also associated with health at birth, including retention, behavior, attention, and executive function (Aarnoudse-Moens et al. 2009; Conley et al. 2003; Saigal et al. 1991; Temple et al. 2010). Ultimately, early health may influence important life outcomes, including educational attainment and labor market success (Black et al. 2007; Conley et al. 2003; Hack et al. 2002). This literature suggests that health at birth may be an important factor to consider in attempts to improve educational performance.
However, a strong association also exists between birth outcomes and disadvantage, leading some researchers to question whether the relationship between newborn health and educational outcomes is causal or whether poor health at birth is simply a marker for other types of disadvantage that are difficult to measure (Almond and Currie 2010; Almond et al. 2005; Conley et al. 2003; Currie 2009). Yet, there are theoretical reasons to believe that health at birth may have a causal effect on educational outcomes. The “fetal origins” hypothesis asserts that poor health at birth may actually be a sign of negative developmental adaptations that have significant consequences for later life (Barker, 1999; Barker et al. 1993; Godfrey and Barker 2001; Rasmussen 2001; Shonkoff et al. 2009). According to this hypothesis, negative impacts that occur during critical periods of fetal development cause the fetus to adapt by making permanent changes that may ensure immediate survival at the expense of long-term welfare (Currie 2009; Ellis et al. 2011; Godfrey and Barker 2001; Shonkoff et al. 2009). The exogenous nature of shocks, such as those caused by natural disasters, provides an opportunity to examine support for the fetal origins hypothesis.
The stress hormone cortisol, in particular, has been cited as a possible mechanism connecting prenatal shocks, birth outcomes, and later-life outcomes. The release of cortisol resulting from maternal stress may be a contributing factor in preterm delivery (Wadhwa et al. 2001). Exposure to cortisol in utero has also been associated with lasting changes in hypothalamo-pituitary-adrenal (HPA) function and later levels of cortisol in the child (Kapoor et al. 2006). Maternal prenatal stress levels, in general, have been associated with negative birth outcomes (Dole et al. 2003; Hedegaard et al. 1996; Kinsella and Monk 2009; Lobel et al. 2000) and longer-term mental health (Beydoun and Saftlas 2008). In addition to stress, natural disasters may impact prenatal care usage, nutrition, substance use, and other important health factors during pregnancy.
Several studies in the economics literature have used various types of community disruptions to provide support for the fetal origins hypothesis in reference to educational and economic outcomes (Almond and Currie 2011). These studies have shown that prenatal exposure to the 1918 flu pandemic (Almond and Mazumder 2005) and the Chernobyl nuclear disaster (Almond et al. 2009), for example, have negatively affected educational or economic performance later in life. It is reasonable to hypothesize that prenatal exposures to natural disasters might result in similar long-run effects. Indeed, the researchers in Project Ice Storm studying women exposed to the 1998 ice storm in Canada found that those children had lower levels of intellectual ability and language skill at ages 2 and 5 (Laplante et al. 2004, 2008). However, the sample sizes were quite small—58 and 89, respectively—and the researchers did not look at actual school performance outcomes.
This study will extend the literature by linking individual prenatal exposure to a natural disaster to educational outcomes across a large group of children and a number of different disasters. In doing so, this study will provide evidence of longer-term effects accruing from prenatal exposure, beyond health effects measured at birth.
The size of the study will allow for a nuanced look at how effects vary by disaster type, subgroup, and trimester. The literature on social vulnerability has found that some population subgroups are more likely to experience large negative effects from natural disasters (Cutter et al. 2003; Zahran et al. 2008). Among the most vulnerable groups are racial and ethnic minorities, the less-educated, and those with lower socioeconomic status (SES) (Cutter et al. 2003; Cutter et al. 2000; Zahran et al. 2008; Fothergill and Peek 2004). This increased vulnerability is due to limited access to information, resources, and political representation (Cutter et al. 2003; Zahran et al. 2008). In addition, some studies of prenatal health effects have found that families of higher SES can compensate for poor birth outcomes (Almond and Currie 2010; Conley et al. 2003; Goosby and Cheadle 2006).
Research also indicates that an infant’s health may be more negatively affected by a shock during early pregnancy or late pregnancy (Almond and Mazumder 2011; Catalano and Hartig 2001; Glynn et al. 2001; Hedegaard et al. 1996; Lederman et al. 2004; Roseboom et al. 2001; Torche 2011) rather than having a uniform effect across trimesters. Potential differences in outcomes of these different exposures are critical for understanding mechanisms as well as possibilities for intervention.
Methods
Data
This study combined three administrative data sets to create a longitudinal data set containing prenatal disaster exposure, birth characteristics, and school outcomes. The sample included all singleton births from 1988 to 2000 in North Carolina that could be matched to third grade test scores in North Carolina public schools from 1997 to 2011. Although the sample initially included all children born alive in North Carolina during the time period, some individuals were dropped from the sample because they could not be matched to public school records. Some of the missing school records were a result of children who left the state or did not attend public schools. However, some who did attend public schools may not have been matched because of errors and discrepancies in the data recording. The full data set included 880,967 children for whom third grade education data were available, of a total of 1,323,489 births in North Carolina between 1988 and 2000.
Information on birth date and county of residence at birth came from detailed birth certificate information obtained from the North Carolina Department of Vital Statistics. These records included all children born in the state of North Carolina. Birth records also served as a source of demographic information about the parents, including the mother’s age, ethnicity, and education; marital status of the mother; and information on the father, if available.
School outcome variables came from administrative records for all school districts in North Carolina, provided by the North Carolina Education Research Data Center. The outcome variables were scores on third grade End of Grade reading and math tests and identification as special education and gifted in third grade. School data also provided information on English-language learners.
A matching procedure was performed, using student names and birth dates to link each individual student’s school records to the corresponding birth record for all students for whom both records are available. Across all years, an average of 67 % of individuals born in the state from 1988 to 2000 were matched with third grade school records. The match rate for births in each year was at least 64 %. Most unmatched records are probably children who moved out of the state before reaching third grade or those who were enrolled in private school. Indeed, among 8- to 10-year-olds in the American Community Survey Public Use Microdata Sample (ACS PUMS) from the years 2000 to 2011 who were born in the state of North Carolina, 22 % had moved to other states, and an additional 9 % were enrolled in private school. This leaves just 69 % of the ACS PUMS children who were born in North Carolina enrolled in public school in the state at the age that most children attend third grade. These results compare favorably with match rates for children in this study.
Table 1 provides descriptive statistics and compares demographics of the matched birth records with those that could not be matched. The matched sample had mothers who were somewhat less-educated, less likely to be married, and less likely to be immigrants than the unmatched sample. These differences were statistically significant but are probably accounted for by general trends in those who were more likely to move out of state or to attend private school.
The source of data on natural disasters was FEMA records of major disaster declarations, which are made (if made) at the request of a state’s governor. To be eligible for federal disaster assistance, the needs for recovery must exceed the combined resources of the state and local governments. Declarations designate eligibility for federal assistance at the county level. Because the degree of damage caused by a particular disaster is idiosyncratic to the specifics of the disaster and the landscape of the area, disaster declarations may more accurately reflect exposure to a disaster than weather-based measurements.
Using the date of birth and county of residence at birth for each child in the data set, a determination was made as to whether the child was exposed to a disaster declaration during the prenatal period. For the purposes of determining disaster exposure, gestation was assigned to have begun 40 weeks before birth for all children. The child was considered to be exposed to the disaster if the date of the initial disaster declaration fell between the beginning of gestation and the birth date. Reported gestational age was not used to avoid an artificial correlation between disaster exposure and gestational age. Children who were born early are gestating for fewer weeks. If there is a chance of a disaster in any given week of the year, more weeks of gestation provide more opportunities for a disaster; therefore, children who are born at full-term are more likely to have been exposed to a disaster than those born early. This does not indicate that the disaster caused the children to be born at full-term; rather, they just had more time to be exposed. In addition, gestational age is difficult to measure accurately and may be less accurate following a disaster if prenatal care usage was affected. All analyses were also conducted using an alternate measure of disaster exposure calculated by counting forward 40 weeks from the presumed date of conception based on birth date and gestational age. The results were similar (see Online Resource 1).
Between 1988 and 2000, North Carolina experienced 15 major disaster declarations. All 100 counties in North Carolina experienced at least one disaster declaration, with individual counties experiencing between one and seven disasters over the 13-year period. The types of disasters included hurricanes, winter storms, and severe storms associated with flooding and tornadoes. Table 2 provides details on the number of counties and the fraction of births affected by each disaster type.
Empirical Analysis
This study used regression analysis with county fixed effects to look at the relationship between prenatal disaster exposure and educational outcomes. Test score outcomes were examined using linear regressions, and special education and gifted placement were examined using logistic regressions. The assumption underlying the analytical strategy was that given residence in a particular county, the exposure to natural disasters in a particular year was random and unpredictable.
The frequency of disaster exposure varied considerably by geographic region, but disasters were not concentrated in any one part of the state. The percentage of births affected by disasters also varied considerably across years. The total percentage of affected births ranged from 0 % in several years that did not have any disasters to 80 % of births in 1996, when four separate disasters occurred.
Because individuals were unlikely to be able to anticipate the occurrence of a natural disaster, their reproductive decisions would not be influenced by disasters. Therefore, families with a pregnancy at the time of the disaster should be similar to other families with children of similar ages. Any other long-term effects of disasters on families should not have varied between families who had children born just before the disaster and families with a pregnancy during the disaster. The regression analyses should, therefore, provide unbiased estimates of the effects of prenatal natural disaster exposure on educational outcomes.
The independent variables in the main analysis were disaster declarations in each trimester of pregnancy. Exposure was divided by trimester to allow for differences in the consequences of exposure by the period of development during which it occurred. Disasters were also divided according to type: hurricanes, winter storms, and severe storms with tornadoes or flooding. The individual and community effects of different disasters may be different and may, therefore, result in different influences on prenatal development and school outcomes. Aid provided by FEMA for winter storms tends to be PA aid for use by local and state governments for purposes such as debris removal and repair of roads, bridges, and utilities. This is in contrast to hurricanes and severe storms, for which a large percentage of the federal aid consists of individual assistance grants for families. This difference suggests that hurricanes and severe storms are associated with significantly more individual property damage. The per capita aid funding for hurricanes in the sample is also much higher than for the other types of storms, suggesting that hurricanes may be more severe for a larger number of people.
The dependent variables were third grade reading and math scores and also identification as special education or gifted. The test scores for each subject were standardized to a statewide mean of 0 and standard deviation of 1 for each year of the test. Special education and gifted status were indicator variables for whether the child was identified as special education or gifted in third grade testing data.
As control variables, the regressions included demographic information, including an indicator for first birth to the mother; maternal age, education, race, immigrant status, and marital status; presence of a father on the birth certificate; and English-language learner status. County of birth, week of birth, and year of birth fixed effects were also included in the analysis. Week of birth fixed effects controlled for systematic differences in outcomes by season of birth that have been documented in previous literature (Doblhammer 2004). Year fixed effects controlled for changes over time. County fixed effects allowed for comparisons across children who were born in the same county at different times. The county fixed effects were important because they accounted for any unmeasured differences in county populations that are correlated with the frequency of disaster exposure. Standard errors in all regressions were calculated based on individuals clustered within county-year of birth to acknowledge that treatment status is assigned based on county of residence at a particular time rather than individual exposure.
The initial analyses included all individuals with matched records. The analysis was then extended to consider differences in the effects for different subgroups. Analyses were rerun for children born to white mothers or black mothers, as well as children born to mothers with no education beyond high school or mothers with at least some college. A mediation analysis was then conducted to determine the role played by common measures of health at birth in mediating the effects on educational outcomes. Finally, sensitivity tests were performed to eliminate the possibility of confounding influences.
Selection Bias
One serious concern for this study was the possibility that disaster occurrences altered the composition of births in counties exposed to the disaster. Disasters may have affected the composition of births in a county in three ways. First, pregnant women exposed to the disaster may have made the decision to move out of the county before the birth of their child. These women may have been more likely to be more-advantaged women with more resources, which would lead to the appearance of worse birth outcomes in the county. Because county of residence was determined at the time of the birth, there was no way to observe whether pregnant women were leaving the county as a result of the disaster.
Second, women exposed to a disaster may have been more likely to have a miscarriage than other women, resulting in fewer births. If this happened, it was likely to have been occurring among pregnancies that were at the highest risk and would have biased the results toward the appearance of better birth outcomes. Because many miscarriages occur before women even know that they are pregnant, it was impossible to test for this directly. Third, women who experienced significant disruption to their lives because of a disaster may have been more likely to decide to terminate a pregnancy.
Although testing directly for any of these mechanisms was not possible, testing for changes in the total number of births per month in counties exposed to a disaster was. An increase in miscarriages would have most likely led to fewer births among those exposed to disasters during the first trimester, and an increase in abortions would have most likely led to fewer births among those exposed during the first or second trimester. Finally, an increase in movement out of the county could have led to fewer births across all three trimesters.
Table 3 shows regressions of natural log of total births per month in each county on the percentage of births in the month exposed to a disaster during each trimester. This is a better outcome measure than the total number of births because the significance of an increase or decrease of a certain number of births depends on the overall birth rate. Column 1 indicates no statistically significant effect of winter storm or tornado or flood exposure on the total number of births. However, hurricane exposure in the first two trimesters appears to reduce the number of births by 3.6 % and 2.6 %, respectively. The remaining four columns show the effect of exposure on the number of births to different subgroups of mothers. The decrease in births following hurricane exposure is statistically significant only among mothers with more education. This could create a negative effect of hurricanes on test scores based purely on selection. However, the difference in the coefficients on hurricane exposure between the maternal education subgroups is not statistically significant, and the subgroup analysis will allow me to explore whether effects are persistent among different subgroups. I discuss the implications of the selection in more detail in the Results section.
A second serious concern is the possibility that exposure to a disaster affects the likelihood that a child born in the state appears in the sample of students matched to their school records. If children move out of the state or attend private school at higher or lower rates following a disaster, it could affect the composition of children included in the matched sample. Table 4 shows the results of logistic regressions of an indicator for being in the matched sample on disaster exposure variables for children of all mothers and children born to mothers in different demographic subgroups. The results indicate that hurricanes and winter storms have no effect on matching. However, flooding and tornadoes may increase the match rates, particularly among disadvantaged groups, perhaps because flooding and tornadoes are much less likely to have a community-wide effect but bring significant financial aid to the community. There is no consistent pattern of differences in matching between subgroups.
Results
Table 5 displays the basic results for the influence of prenatal disaster exposure on test scores and special education and gifted placement by disaster type and trimester of exposure. The negative effects of disasters are mostly concentrated among those exposed to hurricanes. Hurricane exposure in any trimester is associated with reduced math scores of about 2.3 % of a standard deviation and reduced reading scores ranging from 1.6 % to 2.2 % of a standard deviation. Being exposed to flooding or tornadoes in the first trimester reduces math scores by 3.2 % of a standard deviation. Winter storms in the later trimesters reduce the relative odds of placement into special education. None of the disaster types show significant effects on gifted placement. Overall, the effects on test scores are quite small but not trivial. The effects are similar in size to the effects of having a teacher with National Board Certification (Clotfelter et al. 2007), a qualification for which the state of North Carolina pays $3,000 extra in salary per year. For the “missing births” shown in Table 3 to have effects of this size, it would be necessary for the “missing” children to have had average scores of 0.75 to 1 standard deviation larger than the average. A selection effect this size is possible, but unlikely even if all of the “missing births” were among the most advantaged mothers. The following section will provide more detail on the differences between subgroups.
Subgroup Analysis
Rather than being distributed evenly across all groups, effects of prenatal disaster exposure are likely to be concentrated among disadvantaged groups, which are more vulnerable to negative consequences of disaster exposure. In the analysis described in this section, I tested this theory by conducting regressions separately for children of white mothers, black mothers, mothers with a high school diploma or less, and mothers with at least some college. A post-estimation test then compared the size of the coefficients on disaster exposures for the two racial groups and the two education levels.
Table 6 shows the subgroup results for math, reading, and special education. Each pair of subgroup columns is followed by a column that indicates the statistical significance of the difference in coefficient size. As shown in the first two columns, children born to black mothers suffer significantly greater negative effects on math test scores from exposure to hurricanes in the second or third trimester compared with children of white mothers. Children of black mothers also suffer greater negative effects on reading test scores from exposure to hurricanes during the third trimester, as shown in the two middle columns of Table 6. The size of the test score effects for children of black mothers was quite a bit larger than the average effects across all children, ranging from 2.6 % to 5.4 % of a standard deviation. There were no significant differences in test scores between children of mothers with different education levels. However, there were significant effects of hurricanes for both education groups. (See Online Resource 1 for a full subgroup analysis.) The fact that disaster exposure effects occur even when highly educated mothers are excluded suggests that the effect is not driven merely by selection. However, the lack of significant differences between the education subgroups may be due in part to negative selection among the more-highly educated mothers.
The last two columns of Table 6 show significant differences in the effects of disaster exposure on the probability of special education placement for children born to mothers with higher or lower education. The differences between these two groups were statistically significant for at least one trimester for all disaster types. However, children of highly educated mothers were more likely to be placed in special education if they were exposed to a hurricane, tornado, or flooding, but they were less likely to be placed in special education if they were exposed to a winter storm. This pattern of results could suggest that highly educated mothers are more likely to respond to small changes in their child’s cognitive abilities by seeking special education placement, or could merely be a result of selection among this group. It also raises the possibility that certain types of disaster exposure may be good (rather than bad) for educational outcomes. There was no difference in effects of exposure on special education placement between mothers of different races (not shown).
Mediation Analysis
The next section of the analysis considers the role played by common measures of health at birth. As a first step, Table 7 displays the results of regressions of birth outcomes on indicators of disaster exposure. The outcomes in these regressions are birth weight, gestational age, low birth weight (less than 5.5 pounds), preterm birth (earlier than 37 weeks), and small for gestational age (below the 10th percentile). There is little evidence that disaster exposure affects birth outcomes. The only significant effects are a 4.6 % increase in the relative odds of low birth weight and a 4.7 % increase in the relative odds of being small for gestational age related to exposure to a hurricane in the second trimester.
Given the limited evidence that disaster exposure affects birth outcomes, one would not expect these variables to mediate the effect on school outcomes. Indeed, regressions including birth outcomes as controls in the basic model do not show any changes in the effect of disasters on school outcomes (Online Resource 1).
This study’s finding of no impact of disaster exposure on common measures of birth outcomes is at odds with much of the literature. However, a recent study of the effects of hurricane exposure similarly found no effects on birth weight and gestational age but did find effects on birth complications (Currie and Rossin-Slater 2013). Additionally, a study using sibling fixed effects to examine the influence of prenatal cortisol exposure on educational outcomes found negative effects of cortisol exposure on adult educational attainment as well as IQ at age 7 but no effects on birth outcomes (Aizer et al. 2009). These types of findings, in addition to my own, suggest a need to understand the more subtle influences on prenatal cognitive development that may not be captured in broad summary measures, such as birth weight.
Robustness Checks
The next section addresses some potential concerns about the primary estimation strategy used in this study and attempts to assess to what extent those concerns may bias the results in the previous sections.
One concern is whether the effect on educational outcomes previously demonstrated is in fact due to effects on prenatal development or due to other influences of the disaster, such as family income loss. Although I cannot directly test for other disaster influences, I can compare the children who were exposed prenatally with a group of similarly aged children who may also have been exposed to these influences.
My first approach is to consider the effects on children born just before the disaster occurred. To do this, the basic regression equations are altered to include indicators for children up to three months younger or up to three months older than the children who were exposed prenatally. These groups include all children born in the three months before a disaster or conceived in the three months after in the counties where the disaster occurred. Because any effects of disaster exposure on children who were born before the disaster could not operate through prenatal development, any significant effects for older children would indicate disaster effects operating through channels other than prenatal development. The younger children may capture confounding effects as well but are also very likely to be affected by selection into having a baby immediately following a disaster.
The results of this analysis are shown in Table 8. Hurricane exposure among the younger children is associated with significant changes in test scores and special education placement that suggests significant selection into having a baby following a hurricane. Math scores also decreased significantly—by 1.7 % of a standard deviation—for the slightly older children exposed to a hurricane, suggesting that some of the effects of hurricane exposure on math test scores may be operating through a mechanism other than prenatal development, such as employment or mobility effects on the family. However, the size of the decrease in math test scores is smaller among the older children than those exposed prenatally, so some of the effect may indeed operate through prenatal development. The older children also show a decrease in special education and gifted placement associated with winter storms. It is difficult to interpret this finding given that there is no significant effect of winter storms in the three trimesters of prenatal development. Math and reading test scores both seem to show a negative effect of a tornado or flooding exposure among the older children, but there are also other reasons to believe caution is necessary in interpreting these results.
One important consideration when using slightly older children as a falsification test in this study is that many of the students born just before a disaster will attend school with students who were exposed prenatally, so they may also be experiencing spillover effects if they are, as a result, sharing a classroom with students with weaker cognitive skills. As a general conclusion, the results from this test suggest that although the effects of hurricanes on reading scores and special education are robust, caution should be used in interpreting the size of the effects on math scores.
In a second test for more generalized effects of natural disasters on families with young children, the group of children in the analysis is limited to those born in a two-year window immediately around the occurrence of a disaster in their county of birth. This two-year window includes those born in the year just before the disaster and those born in the year following the disaster, so that all families would have had a very young child during or soon after the disaster. Table 9 shows the results of this analysis. The effect sizes for math and reading scores among those exposed to hurricanes and tornadoes or flooding are quite a bit smaller, and some coefficients are no longer significant. However, the effects of hurricane exposure in the third trimester on math and reading scores are still significant although somewhat reduced in size. In this test, hurricane or tornado or flooding exposure in the third trimester are also associated with increased odds of being placed in special education. Additionally, with this more-limited sample, winter storms seem to be associated with large decreases in special education placement and large increases in gifted placement.
Again, students included in this more-limited control group are more likely than other students to experience spillover effects from attending school with children affected by disaster prenatally. Additionally, county, week, and month-of-birth fixed effects are likely to be less accurately estimated on this reduced sample.
In summary, the preceding evidence seems to suggest that mechanisms other than prenatal development may account for some of the effects of natural disasters on children exposed prenatally. However, there is still reason to believe that natural disasters, especially hurricanes, have a negative impact on prenatal development leading to decreases in school outcomes. Exact effect sizes, particularly those for math scores, should be interpreted with caution.
There is also a potential concern regarding the designation of disaster exposure to a particular trimester of pregnancy. The primary method used in this study is to assign the trimester of exposure based on the date of birth and a normal length of gestation of 40 weeks. This method is common in the literature but does lead to some measurement error given that not all children are born at exactly 40 weeks of gestation. It is possible to assign disaster exposure based on reported gestational age and birth date. However, this is not desirable for two reasons. First, gestational age is often an unreliable measure and may be systematically biased if disaster exposure leads to changes in healthcare usage. Second, using reported gestational age results in children with longer gestation having more weeks in which a disaster could occur and be considered prenatal. Therefore, actual gestational age tends to be positively correlated with disaster exposure, but the relationship is not causal.
This study takes two alternate approaches to assess the impact of the method of designating trimester of exposure on the results. First, the original analyses (reported in Table 5) are repeated with the sample restricted to children with a reported gestational age of 37 to 42 weeks. With this limited range of gestational age, the original method of designating trimester of exposure will have minimal measurement error. The results of this analysis (Online Resource 1) are nearly identical to the initial estimates.
The second approach is based on an approach used by Currie and Rossin-Slater (2013) to look at prenatal hurricane exposure in Texas. In this strategy, actual birth date and gestational age are used to calculate the beginning of gestation and actual exposure during each trimester. The date of the beginning of gestation is also used to calculate the expected exposure during each trimester if the pregnancy had lasted a normal 40 weeks. Expected exposure is then used to instrument for actual exposure, using an instrumental variable (IV) regression. The results of this analysis are displayed in Table 10. In general, the effects are quite similar or somewhat larger than those seen in the basic analysis. However, the results of hurricane exposure in trimester 3 are somewhat smaller than those seen in the basic analysis. In addition, the effect of hurricane exposure on special education is significant in all trimesters. Overall, this analysis suggests that measurement error in the assignment of exposure to particular trimesters may be biasing the basic results somewhat downward. However, this method is vulnerable to bias if exposure to disasters affects the accuracy of the measure of gestational age.
In sum, the tests performed in this section indicate that although some caution should be exercised in interpreting exact effect sizes, the overall result that prenatal exposure to some natural disasters, especially hurricanes, has a negative effect on school outcomes is relatively robust. The first set of robustness checks indicates that some of the effects of hurricanes may not be operating through prenatal exposure and that the initial results may represent an upper bound. However, the analysis of the measurement error surrounding placement into particular trimester indicates that this measurement error may be biasing initial estimates downward. Altogether, there is reason to believe that the initial estimates presented in this article represent moderate estimates of the impact of prenatal exposure to natural disasters across the whole population of births exposed. The effects on individual births probably vary a great deal.
Discussion
Natural disaster exposure is fairly common throughout the United States. In North Carolina from 1988 to 2000, more than 20 % of children were exposed to at least one disaster during their mother’s pregnancy. This study provides evidence that shocks to prenatal development caused by exposure to natural disasters can have impacts on educational performance in elementary school. Test scores in math and reading decrease between 1 % and 5 % of a standard deviation, and the probability of special education placement increases between 10 % and 20 %.
The test score effects are relatively small, and effect sizes should be interpreted with some caution, but they are of a magnitude comparable with other factors of concern to policymakers. As mentioned earlier, the increase in test scores associated with a teacher being National Board Certified is similar to the size of the effect associated with hurricane exposure. Controlling for other factors, the effect of birth weight on test scores is 3 % to 3.5 % of a standard deviation—again, an effect size similar to that of hurricane exposure. Finally, the largest effect sizes—those for hurricane exposure among children born to black mothers—are as large as 5 % of the total test score gap between black and white students.
These effects warrant policy concern. This is especially true if we consider that the average effects across entire exposed counties probably include many children who experienced only minimal adverse outcomes related to their disaster exposure as well as some children who experienced substantial impacts. Additionally, these outcomes are measured at a quite distant point in time, approximately nine years after exposure. Effects may have been larger in early grades and may have influenced student trajectories in school through mechanisms, such as track placement and grade retention.
The lack of significant effects on birth outcomes and the failure of these variables to mediate the effects on educational outcomes are important to note. An important potential mechanism that may relate prenatal disaster exposure to educational outcomes is cortisol exposure related to prenatal maternal stress. Housing damage, power failures, and evacuations may also play an important role in the effect of prenatal natural disaster exposure, but one biological mechanism mediating the effect of these disruptions of prenatal development is likely to be maternal stress as well. Other potential mechanisms are prenatal care usage, nutrition, and substance use, all of which may be affected by the community disruption following a disaster.
Previous studies have found effects of increased maternal cortisol on both birth outcomes and educational outcomes (Beydoun and Saftlas 2008). However, the literature does not suggest that birth weight or gestational age mediate the effects of maternal cortisol on educational outcomes. Rather, birth weight and gestational age simply serve as signals of potentially negative developmental changes that may also have other effects. In this study, maternal stress and resulting cortisol exposure are likely an important mechanism affecting cognitive development; but because of timing, intensity, or other biological factors, cortisol exposure in this study does not affect birth outcomes. Other studies have also documented effects of stressors that seem to bypass measures of health at birth and yet affect later outcomes (Almond et al. 2009; Aizer et al. 2009).
The effects of prenatal exposure to disasters on educational outcomes in this study also provide evidence for the theory that early development can have long-lasting impacts. Studies in the medical literature have suggested similar relationships between fetal development and later-life outcomes (Barker 1999; Barker et al. 1993; Godfrey and Barker, 2001; Rasmussen 2001; Shonkoff et al. 2009), and many studies have established associations between birth outcomes and educational performance (Aarnoudse-Moens et al. 2009; Andreias et al. 2010; Black et al. 2007; Boardman et al. 2002; Goosby and Cheadle 2009; Hack et al. 2002; Reichman 2005). However, the close relationship between health at birth and family SES has made it difficult to make causal connections between early development and school outcomes. The random and unpredictable nature of the shocks caused by natural disasters in this study provides strong support for the causal role of prenatal health impacts on school performance.
The variation in size of effects across disaster type, trimester of exposure, and subgroup provide some insight. Of disaster types, hurricanes show the most consistent negative effects. Severe storms associated with tornadoes and flooding also show signs of negative effects on test scores, but winter storms, if anything, may have some small positive effects. These differences illustrate the importance of understanding the mechanisms by which natural disasters influence prenatal development. Winter storms may have far less devastating effects in terms of housing displacement, property damage, and overall stress. As discussed earlier, federal aid figures indicate that hurricanes are associated with more overall damage and more individual property damage than winter storms. This difference hints that family disruptions associated with housing loss and displacement may be more important than short-term damage to roads and utilities in influencing longer-term educational outcomes of prenatal exposure.
As hypothesized, some disadvantaged subgroups seem to show larger negative effects on test scores from disaster exposure. In particular, children born to black mothers experienced negative effects from hurricane exposure that were twice as large as the average effects experienced by all children in the sample. On the other hand, children of white mothers appeared to experience smaller than average effects. This supports the hypothesis that some advantaged groups are less vulnerable to shocks such as those created by natural disasters. This relative invulnerability may be related to greater resources, which can be directed to replace or repair resources damaged in the disaster or a higher likelihood of avoiding damage in the first place (Cutter et al. 2003; Zahran et al. 2008). Contrary to expectations, test score results did not differ across mothers’ educational levels. The vulnerability of children born to black mothers in this study may be particularly great because children in this group are particularly likely to have adverse birth outcomes. In addition, black mothers may be more likely than white mothers at similar educational levels to live in vulnerable communities. Subgroup results suggest the importance of targeting additional resources to assist those most at risk in the wake of natural disasters. Future studies should work to better understand the mechanisms that mediate the effect of disasters on prenatal development.
Policy Implications
Millions of dollars and decades of effort have been invested in attempting to improve the educational outcomes of American school children. However, this study suggests that influences on outcomes may begin much earlier and far outside the classroom. No one doubts that natural disasters are costly, but this study suggests that the costs may be much larger and longer-term than typically assumed. Policymakers may wish to respond to these additional costs.
The costs to pregnant women and their children in terms of decreased educational performance should be included in cost-benefit analyses when deciding how much to invest in efforts to mitigate the impact of disasters. The long-term nature of these costs also suggests potential benefits that could be derived by investing resources to reduce the negative impact of disasters on pregnant women when they do occur, particularly targeting disadvantaged groups that may be at the highest risk.
Future Directions
Future work on this topic should focus on discerning and understanding the mechanisms that mediate the effects of natural disasters on prenatal development. The short-term consequences of natural disasters are many and varied, and understanding which mechanisms are most important for long-term outcomes may facilitate better responses to disaster occurrences. It would also be useful to extend this research to consider other long-term consequences from disaster exposure, including long-term health and economic outcomes that may be influenced by disaster exposure either prenatally or at other vulnerable periods of life.
