The Impact of Flooding on Education of Children and Adolescents: Evidence from Pakistan

1. Introduction

In the 21st century, though the nations around the world are serious and taking initiative to provide quality education for their up-coming generations, the access to education is still a serious challenge to many developing countries, especially countries in Sub-Saharan Africa and South Asia. Millions of children around the world were out of school in 2017 including 61, 62, and 141 million of age groups 6–11, 12–14, and 15–17, respectively (UIS2017). Pakistan, after Nigeria, has the second highest number of primary aged out-of-school-children in the world. About half of the Pakistani children (5–16) that is equivalent to 23 million did not attend any schools; out of which, 18 million had never enrolled while 5.4 million had dropped out from schools (Mughal et al.2019).

Like other countries, Pakistan has also incorporated ‘free and compulsory education for all of its citizens’ in its constitution. Article 25-A of Constitution of Pakistan obligates “the state shall provide free and compulsory education to all children of the age five to sixteen years in such a manner as may be determined by law”. The out-of-school children in Pakistan is a grave threat for the economy. The common factors that restrain children to attend schools in Pakistan include: (a) School-related factors (e.g. parents’ disappointment in the quality of education, corporal punishment, mismatch between skills needed for job markets and education provision by schools especially in rural Pakistan, lack of government control on school management, lack of parent–teacher interaction, location of schools, etc.). (b) Economicfactors (e.g. poverty, i.e. children in rural areas help their parents during the harvest season in which every hand matters — a short-run benefit for the family at the cost of their long run benefits associated to education, female education is to be considered a waste of resources, etc.). (c) Cultural, social and religiousfactors (e.g. lack of integration of religious education in formal education system, background of parents, etc., an inclination of parents to send their children to Madrasahs because of their free education, boarding and lodging facilities), (Kamran and Ul-Deen2017; Buergi et al.2018; Manzoor et al.2018; Mughal2018; Mughal et al.2019; Mughal2020).

On top of that, natural disasters such as floods usually disrupt education system of children and adolescents due to the destruction of schools’ buildings, displacement of families and most immediate needs on time compelling them to discontinue their education and do child labor by helping their families meet basic necessities in such distressed periods (Kousky2016). Due to climate change and its geographical location, Pakistan is one of the most vulnerable countries in the world that faces extreme natural disasters, namely floods, drought, and earthquakes, etc. (Ahmad and Afzal2020). Pakistan is ranked 9th in flood-affected countries and since its inception it has faced 22 major floods, starting from 1950 to 2014: It has faced severe floods in 1950, 1956, 1957, 1973, 1976, 1978, 1988, 1992, 2010 and now 2022.

Recently Pakistan has been facing a heavy flood in 2022. Since mid-June 2022, floods caused by monsoon rains have affected almost all provinces of Pakistan. So far, 1,033 people have died and 1,527 people have been injured in the massive floods. The estimated number of people affected is around 30 million, and about one million houses have been completely or partially destroyed and millions are in urgent need of shelters (Turkish Red Crescent Society, 2022). OCHA (August 26, 2022) Pakistan’s Floods Situation Report No. 03 states that as of 25 August, Pakistan has experienced 375.4mm of rainfall — 2.87 times higher than the national 30-year average of 130.8mm. 116 districts have been affected, of which 66 have been officially declared “calamity hit”. Data from the provincial education departments show that 17,566 schools were damaged or destroyed by the flood 2022: 15,842 in Sindh, 544 in Balochistan and 1,180 in Punjab. Additionally, about 5,492 schools are reportedly being used to accommodate displaced people.

In 2010, Pakistan had faced a similar huge flood which affected every tenth individual of the country. The flood had spread over 78 out of 140 districts. According to Government of Pakistan, it affected 18 million people directly, destroyed 1.7 million houses, killed 1,984 and wounded 2,946 people (Government of Pakistan2011). The event demolished so many essential infrastructures and structures of the country that worsened the already impoverished economy of the country further and disrupted the process of human capital accumulation. The South and Asia Pacific regions are the most flood-prone areas of the World and according to one estimate these regions are twenty-five times more likely to be affected by the floods than any other continents of the world (Hirabayashi et al.2013; Ireland2016). Pakistan is one of those countries which has been badly affected by natural disasters particularly earthquakes, droughts, floods and the most recent pandemic COVID-19, etc. (Abbas et al.2015; Raza et al.2020). Therefore, the effect of such natural disasters on human capital accumulation needs to be studied.

In this study, we examine the effect of the flood on education outcomes of children and adolescents (ages 5–16) across flooded and non-flooded households of Pakistan by utilizing a quasi-natural experimental design and a difference-in-differences (DID) estimation technique, together with incorporating the inverse probability of treatment weights (IPTWs). This approach allows us to calculate the effect of the disaster on education outcomes of those households who were exposed to the flood comparing with the same outcomes of those children and adolescents in non-flooded households. Given the availability of household level data such as out-of-school or dropout children and adolescents — from large and nationally representative household surveys, namely Pakistan Social and Living Standards Measurement (PSLM) surveys, this study examines the effect of a natural disaster (flood) on human capital accumulation in Pakistan which was not possible in a small sample, otherwise.

The results of this analysis show that the rural communities in flood-prone districts were severely affected by the flood and the process of their human capital accumulation was disrupted during the flood. Our findings reveal that the flood further reduced admission in rural areas significantly; out of 1000 admissions of children and adolescents, 40 fewer admissions happened in rural areas of Pakistan. In other words, out-of-school children and adolescents increased in rural flooded households during the flood year.

The study contributes to the literature in the following ways: First, a huge body of literature is available on the effect natural disasters have on education outcomes (e.g. Maccini and Yang2009; Cuaresma2010; Kim2010; Chang et al.2013; Frankenberg et al.2013; Khan and Ali2014; Mudavanhu2014; Abbas et al.2015; Kousky2016; Ireland2016; Buergi et al.2018; Chuang et al.2018; Ilumin and Oreta2018; Paudel and Ryu2018; Gibbs et al.2019; Proulx and Aboud2019; Andrabi et al.2020; Drzewiecki et al.2020; Hoffmann and Blecha2020; González et al.2020; Mian and Chachar2020; Shah et al.2020); this study uniquely contributes to literature by using a very huge and nationally representative survey dataset from Pakistan to examine the impact of flood on educational outcomes (out-of-school or dropout) of children and adolescents (aged 5–16) in the flooded districts. Second, this study attempts to tease out the effect of the flood on educational outcomes of children and adolescents based on their residential location (rural vs. urban) or their gender (male vs. female). Third, the novelty of the study in the context of developing countries is that it utilizes a quasi-experimental research design and a difference-in-differences (DID) estimation technique to disentangle the effect of the disaster on education of children and adolescents in disaster-prone households not only for short-run or immediate, but also for the medium- to long-term (2–4 years) aftermath. Fourth, this study also tries to channel the causes of out-of-school children dropped out from schools in the flood-affected households during and after the flood. Finally, the findings of this study might have some policy implications in other similar settings (e.g. developing countries) or similar disasters (e.g. current pandemic COVID-19) which similarly cause the closure of schools and delay schooling of children and adolescents.

2. Literature Review and Theoretical background

The UN disaster risk reduction unit reported that the human cost of disasters in the last 20 years was huge that accounted for about US$ 2.97 trillion and affected over 4 billion people worldwide (CRED and UNDRR2020). The extant of literature studied the impact of natural disasters on different micro and macro level outcomes including education (Frankenberg et al.2013; Gibbs et al.2019; Proulx and Aboud2019), human capital productivity (Paudel and Ryu2018), infrastructure (Marto et al.2018), income and national growth (Panwar and Sen2019) and social settings (Ahmed2018). Among other major impact outcomes, natural disasters hamper human capital development as education systems of countries (i.e.; school infrastructure, school enrollment, student performance, teachers training etc.) are badly affected by disasters.

Theoretically, the impact of natural disasters on educational investment is yet ambiguous (Cuaresma2010). Because on the one hand, when natural hazards have a tendency to decrease the expected returns of physical capital, rational people tend to divert their resources to accumulate human capital (Okuyama2008). But on the other, “in a framework of models of agents with finite lives”, the possible risk on human death in fact lowers education investment in the affected areas of disasters (Cuaresma2010). We would rather follow Checchi and Garcıa-Peñalosa (2004), who postulated a very simple and refined theoretical model in which cumulative production risk determines the average level of education and its distribution. Greater hazard is linked with greater disparity and a less average attainment of education and potential future earning capability. As suggested by Checchi and Garcıa-Peñalosa, if we assume that the flood 2010 in Pakistan — a type of natural disaster is as a part of total production risk in the economy, then the flood affected households that were severely affected by the flood should reflect a reduction in human capital accumulation in terms of educational outcomes (i.e. on average, higher number of out-of-school and drop-outs from school children and adolescents) compared with the educational outcomes of individuals of same age groups in non-flooded households. A disaster not only destroys the landscape, structures and infrastructure of a household but it also influences the investment decision of the household in long-run human development (Kim2010).

The empirical evidence conforms to the validity of this theoretical model in different settings. For example, González et al. (2020) in the context of Argentina, we found that natural disasters affected individual personal experiences and human development outcomes, e.g. the first year of disaster reduced the schooling years and increased the chances of adults’ unemployment as well as reduced the living standards of people in the country. They further found that floods were the main cause for the reduction of schooling achievements. Similarly, in context of Zimbabwe, Mudavanhu (2014) found that schools were severely vulnerable to flood disasters which severely decrease children’s education performance because of losing learning hours and qualified human resources, out-breaking of diseases, higher absenteeism and incompletion of courses. By using data from the Demographic and Health Survey (DHS) and Living Standard Measurement Survey (LSMS) for Cameroon, Burkina Faso, and Mongolia, Kim (2010) found that extreme natural hazards have long-term adverse effects on education attainment and secondary school completion.

Moreover, Drzewiecki et al. (2020), while using disaster risk reduction (DRR), studied the relationship between educational attainment and resilience in West Indies. They found a significant association between professionally educated adults and resilience to natural hazard-induced disasters, but not for adults with secondary school education compared to adults with no more than primary school education. Further, researchers suggest that employing DRR education is an efficient way to increase resilience and mitigate the negative effects of natural disasters. Moreover, Shah et al. (2020), in the context of Pakistan, argued that Pakistan has severely been affected by natural disasters and disaster-related education in schools is needed to be incorporated in the educational system of the country. Further, they analyzed knowledge, perception, and preparedness of the school children in disaster-prone areas of Pakistan and found that school-level preparedness for disasters is very poor. In another study, Shah et al. (2020) assessed the resilience of schools during catastrophic flood disasters and found that schools’ resilience can be attained by having an integrated and improved mechanism for disaster management or flood resistance. Further, authors suggested that involving key stakeholders for planning and creating a resilient education sector has become vital.

Given the aforementioned background, both theoretically and empirically, this study tries to answer the following key questions:

	Q1.Did the flood 2010 affect education of children and adolescents (aged 5–16) in flood affected households in Pakistan?
	Q2.If yes, did the flood have a disproportional effect on education of children and adolescents depending upon their residential location (rural vs. urban) or gender (male vs. female)?
	Q3.How long had the impacts of the flood on children and adolescents lasted in the flooded households?
	Q4.Through which channels did the flood cause drop-out of children and adolescents from schools in flooded households during and after the flood?
	Q5.What can be done at the policy level to alleviate the adverse impact of floods on education of children and adolescents in flood-prone areas of developing countries?

3. Data sources and Descriptive Statistics

3.1. Household data

Data on educational outcomes of children and adolescents (age 5–18) — out-of-school children, dropout from school, and the reasons for not going to school at present and other control variables at household level that are — age of children and adolescents (in years), number of individuals in household (in numbers), and maximum education of head of family (in years/completed grade; 1$=$class 1, 2$=$class 2, 3$=$class 3 $\dots$, 20$=$PhD) — were extracted from the Pakistan Social and Living Standards Measurement (PSLM) surveys. The PSLM dataset at the district level consists of six waves representing the year 2004–05, 2006–07, 2008–09, 2010–11, 2012–13, and 2014–15. The PSLM is a household-level nationally representative survey conducted in each alternative year by the Pakistan Bureau of Statistic (PBS). There is a one-year gap between each survey. The out-of-school children and adolescents are individuals of age 5–16 years in a household who were never admitted in any school or educational institution. The dropout is a child or adolescent of age 5–16 years in a household who was not currently studying in any institution during the survey year. The reasons for dropping out from school are; the school was far away from home, individuals helped in household and domestic works, parents did not permit them to school, and the child was not willing to go to school. All of the outcome variables of interest that are associated with reasons are dummy variables which hold a value of 1 for an individual (age 5–16) who dropped out from school due to that specific reason and 0 if he/she was currently studying in any institution during the survey.

The variable “maximum parental education of children and adolescents” is defined as the level of education in years (1 $=$ class 1, 2 $=$ class 2 $\dots$, 20 $=$ PhD) of head of the family of the children and adolescents with the highest level of education. The variable “age” is defined as the age of children and adolescents in years and its upper and lower bounds are 5 year and 16 years, respectively.

3.2. Flooded and non-flooded districts

Due to unavailability of a precise variable classifying flooded and non-flooded districts, we combined three different sources of information to identify flooded and non-flooded districts: MapAction (2010), Critical Threats Project (2010), and National Disaster Management Authority (Government of Pakistan2011). Geospatial data for treated (flooded) and control (non-flooded) districts for this study were produced by utilizing the maps of the MapAction (2010) and the Critical Threats Project (2010) as well as data from National Disaster Management Authority. During the time of the disaster, the MapAction seriously coordinates with the humanitarian crisis teams of the United Nation Disaster Assessment and Coordination (UNDAC), the United Nations High Commissioner for Refugees (UNHCR), the Red Cross/Red Crescent Movement and international NGOs. Its emergency mapping services help in providing more instant support and response to more than 60 types of humanitarian crisis including floods. The MapAction’s map was produced on September 6, 2010, with coded colors to districts as per their level of severity. The districts were categorized as severely, moderately and non-affected based on the people affected in the district. The second map was produced by the Critical Threats Project (CTP). The CTP’s main objective for mapping the flooded districts was to track the flooded areas, their level of damages, possible consequences of the flood on the population and also to assist and guide international donors for handling issues in response to the flood by aids and other assistances for examples. This map also categorized the districts as severely, moderately and non-flooded districts. With very few exemptions,¹ the above-mentioned sources declared almost the same districts as flooded districts and also their level of severity. For instance, Mardan was declared as a severely affected district in the CTP while it was a moderately affected district in the MapAction. Only one district, Lasbela (excluded from the study), was declared as flood-affected district in the MapAction while not affected by the flood in the CTP. The intervention variable of interest for this study is the flooded district — a dummy variable which holds a value of 1 if it is a district which was moderately or severely affected by the flood 2010 and 0 otherwise as per the above sources.

The aim of both organizations was to clearly identify households in need of immediate assistance and to assist other humanitarian organizations in prioritizing and locating rehabilitation districts. The distribution of areas into flood-affected and flood-unaffected districts are almost same on both maps. By utilizing the geospatial data from the maps, the study formed a treatment group of 63 flood-affected districts and a comparison group of 52 flood-unaffected districts (see Figure A.1 in Appendix).

From the maps, it is clear that these affected districts were tightly clustered across the Indus River. For further robustness checks and sensitivity analysis, this study also utilized another source of geospatial data for the treatment assignment, the National Disaster Management Authority (NDMA, 2011), along with aforementioned two maps to create a new treated group of flood-exposed districts. These districts were severely affected by the 2010 floods and needed immediate aid response. For comparison group, we still used the maps’ unaffected districts. Here, the study excluded the moderately affected districts. Therefore, information of severely flooded districts (i.e. 29 districts) by the NDMA — as a treatment group — were utilized for robustness checks (see Government of Pakistan2011 for further details). Here, the intervention variable of interest was the severely flooded district which holds a value of 1 if a district was listed as severely flooded district by NDMA and 0 if it was declared as a non-flooded district by CTP and MapAction. The moderately flooded districts were excluded for the robustness checks.

A note of caution: This paper uses multiple data sources (maps from international organizations and data from the NDMA) to identify the impact of floods on education for children and adolescents in flood-prone and non-flood-prone districts. It has the limitation of using continuous variables such as number of household members affected by floods, number of deaths in each district or losses. Although the primary selection criterion for dividing districts on maps was population (and color-coded): severely affected, moderately affected, and no affected districts,no other data at the district level on this variable is available for use.

3.3. Descriptive statistics

The unit of observation for this analysis is children and adolescents of age 5–16 years in a household. After cleaning the data, the dataset consists of 1,026,186 individual observations, of which 52.66% are male and 47.34% are female. The average age of children and adolescents is 10.13 years. On an average, a household consists of 7.53 individuals and the average length of education of head of household is 4.3 completed grades. Table 1 shows descriptive statistics of the educational outcomes the variables of interest for this study. From Table 1, a significant difference between out-of-school children and adolescents in flood-affected and unaffected households before the flood 2010 was observed. For instance, among children and adolescents of age 5–16 years, an estimate of 27.60% of individuals in non-flooded and 36.66% in flooded districts were never admitted in any educational institution. Out of which, 9.48% of them in non-flooded and 10.19% of them in flooded districts were dropped out from schools before the flood. Overall, the percentage of out-of-school children and adolescents, on an average, in both flood-affected and non-affected households have reduced by 3.39% and 1.11%, respectively, after the flood. Descriptive statistics in Table 1 also reveal that the out-of-school and dropout children and adolescents in both flooded and non-flooded districts either before or after the flood 2010 vary significantly for the samples of individuals (age 5–16) based on their geographical location (rural vs. urban) or gender (male vs. female). For instance, the average differences of out-of-school children and adolescents between rural-non-flooded vs. urban-non-flooded households and rural-flooded vs. urban-flooded households were 18.17% (i.e. 34.59–16.42%) and 18.56% (i.e. 41.47–22.91%), respectively. Irrespective of the flood, children and adolescents in Pakistan, if they are female or living in rural areas are more prone to be out-of-schools or to drop out from schools.

4. Empirical Analysis Strategy

The study utilizes the Pakistani flood 2010 as a quasi-natural experimental design to examine the impact of flood on education of children and adolescents (age 5–16). The study tests a number of hypotheses by using a difference-in-differences (DID) approach with inverse probability of treatment weighting (IPTW) based on the following province-year fixed effects regression model :

where EO${}_{idpt}$ is the educational outcome variable (i.e. out-of-school children which represents whether a child or adolescent aged 5–16 years $i$ in district $d$ of province $p$ was out of school in the survey year $t$. The variable FD_d is whether a district was affected by the flood in 2010 in Pakistan and Y11, Y13 and Y15 are dummy variables showing during and after (2010–11), the second year (2012–2013), and the fourth year (2014–15) following the flood, respectively. The unit of observation is a child and an adolescent (aged 5–16 years) in a household who was whether never admitted in any educational institution or dropped out from the school during the surveys’ year. The $\beta s$ in above regression model are the DiD estimates of the coefficients of interest which show the impact of a flood on educational outcomes of children and adolescents in flooded households compared to educational outcomes of children and adolescents in non-flooded households during, two and four years following the flood 2010. The model also includes province-year fixed effects ${\lambda }_{pt}$ to capture time-varying or policy-related shocks at the provincial level. ${\epsilon }_{idpt}$ is the idiosyncratic term. The model is estimated by OLS estimation technique and robust errors are clustered at the district level.

Parallel Trends Assumptions

To justify the use of the DID method, we test the parallel trends assumption that flooded and non-flooded individuals have similar trends before the flood. Figure 1 shows the time trends of out-of-school children and adolescents (age, 5–16 years) in each of the flooded and non-flooded districts during 2004–2015. It shows that there exist almost 10% level difference of out-of-school children and adolescents between flooded and non-flooded districts before the flood 2010. This is because the flooded and non-flooded districts are systematically different due to their geographical locations and socioeconomic conditions of people. Poor or marginalized people are more vulnerable to natural disasters because they often live in more dangerous environments. The number of out-of-school children increased as poor families could no longer afford to enroll them, and they preferred to join the labor force and support the family (Hyder and Iqbal2016). Despite differences in level, the assumption of common trends in out-of-school children and adolescents between flooded and non-flooded districts before the flood 2010 prevailed (see Figure 1). However, surprisingly the out-of-school children and adolescents have decreased further in the flooded districts two years (2012–13) after the flood 2010 and then increased in the years 2014–15 while they have increased in the non-flooded districts in flooded year and then increased after that. This may be due to the fact that another flood occurred in 2013 which also affected districts that were previously unaffected by the 2010 flood and distorted the trend. By utilizing the EM-DAT datasets, once we took flood-prone districts which were affected by the 2010 flood and not by any other disasters (i.e. earthquakes and other floods) and non-flood-prone districts are those which were not affected by any type of disaster including the flood 2010 during 2004–2015, we observed the number of out-of-school children and adolescents increased only in the survey year 2010–11 (see Figure A.2). Due to the exclusion of districts that have been affected by other natural disasters, the sample size is roughly reduced to half; therefore, we use this sample for the robustness check only, not for the main analyses (see Column 8, Table 3).

The established assumption of parallel trends in difference is differences analysis which implies that if the event had not occurred, the average outcome of the exposed group would have been evolved in parallel with the average outcome of the control group. In the context of DiD analysis, relaxing or assessing the parallel trend assumptions is highly relevant and should be considered for the purpose of inferences (Roth et al.2022). By utilizing the event study specification, we tested the parallel trends assumption. Existing empirical literature uses lead terms of the treatment variable to test this assumption (Autor2003; Gao et al.20182019; Gao and Zheng2020). If we find no lead effects, then the parallel trends assumption is upheld to do our DiD analysis. In the model specification; age of children and adolescent, number of household members and maximum education of parent were also included as controls and errors are clustered around districts. The findings of event study are shown in Table A.1. The test results indicate that there is not a lead effect of the flood 2010 on education outcomes (out-of-school children and adolescents and dropouts from schools) in both flooded and non-flooded districts of Pakistan at the conventional level of significance, therefore, the above evidence support the argument of using DID estimation technique for this study.

Balancing of Individuals’ Observable Covariates in Flooded and Non-flooded

In a randomized controlled trial experiment, the effect of treatment on outcomes of individuals can be measured directly by comparing the average outcome of treated individuals with the average outcome of comparison individuals. This is because randomization by definition ensures similarities between the observed and non-observed covariates of the treated and control groups. In contrast, in an observational study, estimating the impact of treatment on outcome is not possible directly because of the treatment selection bias (Austin and Stuart2015). Therefore, though one can assume that a flood is a type of natural disaster which affects people exogenously, the 2010 flood affected people in districts of Pakistan systematically in different ways due to the geographical locations or pre-existing natural variations in districts. For example, people whose living sustenance comes from the agriculture sector usually dwells in and around regions where the main rivers (and which mostly cause flooding) flow. They also lack the basic facilities of lives including access to education, health facilities, foods, water and sanitation facilities, etc. The flood of 2010 in Pakistan severely affected these people (Deen2015). It is clear from the map (Figure 1) that the flooded districts in 2010 were mostly agglomerated across the Indus and other major rivers of Pakistan.

Therefore, in order to obtain unbiased or reduced-biased estimates, this study applied a strategy for balancing individuals by using their covariates in flooded and non-flooded districts. Based on individuals’ observable covariates in flooded and non-flooded districts, a reweighted-balanced-pseudo-sample is obtained by using the inverse probability of treatment weights (IPTWs). The IPTWs take into account measurable confounding and selection bias based on time-varying covariates when estimating the exposure effect on outcome (Cole and Hernán2008). The objective of weighting individuals based on IPTWs is to create an artificial sample in which the treatment assignment is independent of the measured baseline covariates (Austin and Stuart2015). Following Lunt’s (2014) guide for propensity analysis, propensity scores for the IPTWs are estimated by using a logistic model in which the treatment assignment (flooded district) is regressed on the 33 covariates along with survey-year fixed effects. The predicted values of outcome variable are the propensity scores of individuals. Austin and Stuart (2015) define propensity score as the probability of an individual being treated conditionally on observable baseline covariates. Next, we estimate the IPTWs by the following way; individuals in the flooded districts get a weight of 1/Propensity Score and individuals in the non-flooded districts get a weight of 1/(1 − Propensity Score). One of the benefits of using propensity score methods is that they almost mimic randomized controlled experiments in which research design is separated and independent from the main analysis of the treatment effect on outcome variable. The weights obtained by the inverse probability of treatment weighting method actually alter the distribution of covariates of the flooded and non-flooded districts in such a way that reduces their distributional gap (Lunt2014). According to the author “the IPT weighted analysis therefore compares “what we would expect to see if everyone received treatment” to “what we would expect to see if no-one received treatment”. Weighting, furthermore, allows us to use the entire sample without discarding any individual from the sample while reducing the bias by giving more weights to those individuals with closer propensity scores. In order to estimate the propensity scores for reweighting the sample, model misspecification may be one of the concerns but it is difficult to validate this assumption. So, Austin and Stuart (2015) suggest that rather than focusing on the model misspecification, we should focus on the balancing of the covariates between flooded districts and non-flooded districts. Baseline diagnostic checks — whether IPTWs have removed the systematic covariates differences between the flooded and non-flooded districts — are essential. Lunt’s user written program for checking the covariates’ balance between treated and control groups is used. Given the data constraints and repeated cross-sectional data from the PLSM surveys, we did limit here our analysis on five possibly “time invariant” covariates including number of household members, parental education of children and adolescents (in years), age of children and adolescents (in years), location (rural vs. urban) and family head as a male.

Table 2 shows that sample reweighting, on an average, reduces the observable covariates gap between the flooded and non-flooded individuals quite significantly. For instance, the standardized difference of age between flooded and non-flooded individuals was $-$0.064 before reweighting which reduced to 0.001 (a 102% reduction) after reweighting the sample. Reweighting also did well in balancing of other covariates — including age of children and adolescents (in years), number of household members, parental education of children and adolescents (level of education in years), location (rural vs. urban) and family head as a male — by reducing the covariates’ gap between flooded and non-flooded individuals up to approximately 100% (for more details see Table 2). This diagnostic assessment indicates that weighing the original sample created a pseudo-sample which means values of the covariates between flooded and non-flooded individuals remain almost similar. Thus, this study uses these IPTWs for obtaining biased-reduced estimates throughout the entire analysis, if not mentioned otherwise. Since the weights were estimated rather than known, a more conservative robust standard errors estimation method for inferences as suggested by (Joffe et al.2004) is used; that is the errors are clustered at the district level.

5. Results and Discussion

Education is one of the main pillars for the sustainable economic growth, poverty alleviation, and human development of nations. The classroom-based formal education is very crucial for human capital accumulation in developing countries, because schools deliver education to children with trained teachers and sophisticated learning environment (Ireland2016; Hoffmann and Blecha2020). Unfortunately, out-of-school children are a grave threat for the world’s economy in general and Pakistani economy in particular to flourish. The fundamental reasons of out-of-school children include lack of access to schools, poverty, lack of awareness, non-availability of schools, ghost schools, ghost teachers, etc. In addition to that, the dropout rate of children from schools in Pakistan is alarmingly high (UNICEF2013).

5.1. Impact of flooding on education: Out of school children and adolescents

This study — by utilizing a quasi-natural experiment research design with a difference-in-differences (DID) estimation approach — examines the impact of the flood 2010 on education outcomes of children and adolescents in Pakistan and the results are given in Table 3. In Column 1, the estimated coefficients from the baseline regression (1) indicates that the flood 2010 in Pakistan seemingly affected the education outcomes of children and adolescents negatively in flooded districts with reference to non-flooded districts in Pakistan during the flood year 2010–2011. More precisely, 51 more children and adolescents (age 5–16 years) out of 1000 remained out of schools in the flooded districts in Pakistan compared to non-flooded districts during the flood year (2010–11). Interestingly, the adverse effect of the flood on the education outcomes vanished immediately after the flood year.

Columns (2) and (3) of Table 3 represent the estimates of the geographical heterogeneous effect of the flood on education outcomes of children and adolescents respectively in urban and rural regions of Pakistan. The findings show that the flood affected negatively the education of children and adolescents in rural regions of Pakistan only. In other words, in the years 2010–11, 75 rural children and adolescents out of 1000 remained out of schools in 2010–11 in the flooded districts compared with their counterparts in the non-flooded districts. These findings are also aligned with the findings of a recent study by Jamshed et al. (2021) who found that the rural communities of Punjab province, Pakistan had been severely affected by the flood in Pakistan.

Furthermore, the analysis was further extended to investigate the gender heterogonous effect of the flood on education of children and adolescents in flooded districts of Pakistan. The estimated coefficients in Columns (4) and (5) of Table 3 show that the flood affected the educational outcomes of male children and adolescents adversely in the flooded districts compared to their counterparts in non-flooded district during the flood year (2010–11). However, the adverse effect of flood on the education of female children and adolescents was statistically insignificant. More specifically, about 52 more male children and adolescents were out-of-school in flooded districts in 2010–11. Here, our findings are aligned with those of Kim (2010) but contradict with Maccini and Yang (2009). Kim found that adult women in India when exposed to climate shocks during their birth were 19.4% less likely to join primary school. In contrast, Maccini and Yang found that Indonesian women, if exposed to a 20% higher rainfall in their year and location of birth, have attained 0.22 more completed grades of schooling.

Our main regression results are based on education outcomes of those individuals (5–16 years) who should get constitutionally “free and compulsory education” by the state in Pakistan. In order to tease out the effect of the flood on education of children and adolescents separately, we did the analysis by dividing the observations into two samples; children (age 5–9 years) and adolescents (age 10–16 years). The regression estimates in Columns (6) and (7) of Table 3 show that the flood adversely affected the education of both children and adolescents almost equally (that is 46 children and 52 adolescents were out of school in the flood year). A huge number of reduction in the enrolments of adolescents and children in schools during the flood may be due to the fact that they might have helped their parents and families in their household works or became their earning hands during the distressed situation. This line of argument is validated by the results represented in Column (3) of Table 3; 29 out of 1000 children and adolescents including 46 females and 18 males discontinued their education during the flood year and that seemingly due to the fact that they helped their parents and families in the household and domestic works. In addition to that, we further validated our findings that about 40–45 adolescents out of 1000 were dropped out from schools in flooded districts because they most probably helped their parents and families in household and domestic works.

In addition, the 2010 floods severely affected the education sector, costing the sector US$311.3 million in immediate costs and US$504.8 million in long-term rehabilitation of damaged buildings (Khan and Ali2014). It damaged 10,348 educational institutions including 3,741 which were completely damaged and 6666 were partially damaged and more than 90% and 25 of them were primary and girls’ schools, respectively. While the total damaged schools accounted for only 6.2% of total institutions in the country and 12% of total institutions in flood-affected districts, partially destroyed institutions accounted for about 64% of total damaged schools.

Furthermore, a large number schools (about 4,935) were used as shelters for flood victims, including 2,169 in Punjab and 2,372 in Sindh, affecting 1.6 million children and 32,000 teachers (Khan and Ali2014). Though the immediate impact of the flood on education has disappeared after one year of the flood and this may be because many humanitarian organizations have provided educational materials such as teachers, textbooks and lessons to the victims for immediate relief, “The floods exposed what was lacking in terms of education” narrated by Ms Nafisa Shah, the co-chair of the United Nations Girls’ Education Initiative (UNGEI) in Pakistan and Chairperson of the National Commission for Human Development (UN Children’s Fund2011). According to her, the flood brought attention to some hidden problems in Pakistan’s education system such as unequal access to state services for children living in poor communities. She identified lack of school infrastructure, insufficient number of trained teachers and local traditions that keep children engaged in farming as some of the obstacles keeping children out of school.

Most of the affected districts were in the summer zone except the northern hilly region and the flood occurred during summer vacations. Hence, there was negligible loss of life among students, teachers and staff, though a large number of students and teachers were emotionally traumatized by the floods (Khan and Ali2014). This meant that academic contact hours were not affected as a result of the flood, with the only exception being that the summer holidays were extended by a few extra days. The flood caused the most damage to infrastructure, affecting Pakistan’s already fragile education sector.

5.2. Impact of flooding on education: Dropout of school children and adolescents

The estimates in Table 4 show that the flood 2010 seemingly affected the education of children and adolescents adversely; increased dropping out children and adolescents from the school. A number of potential reasons were identified which may cause their dropping out from the schools in the flooded districts. Column 1 of Table 4 shows the evidence that the flood seemingly caused discounting of education of children and adolescents in the flooded districts during the food year (2010–11). Comparing with their counterparts in non-flooded households, 16 out of 1000 children and adolescents were dropped out from the school enrolment seemingly due to the flood (see Panel A of Column 1 in Table 4). In Pakistan, about 73% of children and adolescents (aged 5–16) drop out from schools before completion of their secondary school (Mughal et al.2019). Some possible channels through which the dropouts of children and adolescents from the schools were happened due to the flood 2010 — for which we have data — are further elaborated in the next section.

5.3. Major possible reasons of dropout of school children and adolescents in flooded households

Columns (2)–(5) of Table 4 show several reasons that explain how the dropping out of children and adolescents from the schools might takes place in the flooded districts during and after the flood. Column 2 shows that 32 out of 1000 children and adolescents in overall (43 females and 16 males) could not continue their education during the flood year (2010–11) due to the fact that the flood affected families felt that the distance between schools and their home was far away for their children to get education. Though the matter of dropping out from schools is equally serious for both genders, our analysis reveals that the cause of school at a distance even pushed further down the situation of out-of-school female children and adolescents not only in the flood year (2010–11) but also subsequently in the second (2012–13) and fourth year (2014–15) of the flood. The distance to schools also mattered for boys but only during the flood year and that was seemingly resolved in the subsequent second and fourth year of the flood. One possible explanation may be due to the fact that either boys’ schools in the flooded areas were given priority in the process of repairing and rehabilitation or it became a new normal for those affected to travel for a long distance in order to access those functioning schools located in the adjacent non-flooded areas. Unfortunately, the rate of dropping out of female children and adolescents from schools due to distancing schools had not been limited to only in the flood year (2010–11) but extended and increased after the second (2012–13) and fourth (2014–15) year of the flood with an estimates of 44 and 61, respectively. A possible explanation for this huge impact of the flood may be due the fact that the flood destroyed the school infrastructure and buildings including toilets and other facilities which are generally more essential for female children and adolescents than male. On the other hand, religious and cultural constraints and barriers (e.g. for walking to schools at distance, etc.) are much more lenient for boys than girls in Pakistani society and that may be the reason that compelled parents to be reluctant sending their daughters to the adjacent non-flooded schools during and after the flood. The male children and adolescents of the flooded areas might either have opted to go to the nearest adjacent non-flooded or slightly flooded schools or they might also have dropped out from the school enrolment.

Another possible reason for children and adolescents dropping out of school in flood-prone districts during the flood year may be that drop-outs and adolescents can help their families with household chores. Column (3) of Table 4 shows that 29 out of 1000 children and adolescents in total, 46.4 female, and 17.5 male were dropped out from the schools during the flood year seemingly due to the reason that they might help their families in household and other domestic works. The coefficients in Column (4) of Table 4 indicate that parents in the flooded households compared to their counterparts in the non-flooded households strictly prohibited their daughters from continuing their education during and after the flood. The coefficients of parental limitations on their daughters’ education during and after the flood are not only statistically but also economically significant. In addition to that, children and adolescents of flooded households also lost their interests in getting education and eventually dropped out from the school enrolment during the flood year (see Column 5 in Table 4 for further detail).

5.4. Robustness checks and further extensions

5.4.1. Placebo experiments

A serious possible issue of time trends, due to several confounding factors in the data, may raise several issues, so the findings from the difference in differences estimation technique would be biased. In this regard, we did two placebo experiments to test this assumption and the robustness of our findings. The basic assumption for the placebo experiments is that the devastating flood did happen in 2006–07 (first placebo experiment) and 2008–09 (second placebo experiment) rather than the actual flood that happened in 2010–11. Having a sufficient number of observations, we could estimate the effect of the pseudo flood impact on education outcomes of children and adolescents. The results from this exercise further validates our findings that all the coefficients of interest in both placebo experiments are statistically insignificant (see Columns 1 and 2 of Table 5). We also performed a fictitiously spatial placebo experiment in which we divided the actual non-flooded districts randomly into two pseudo groups (pseudo-flooded and comparison district) by using a STATA command for randomization. Results are shown in Column 3 which indicate that all coefficients of interests are insignificant. Evidence derived from these placebo experiments further proposes that a reasonable portion of adverse effects of schooling children and adolescents was due to the flood 2010 in Pakistan.

5.4.2. Restricted analysis based on common support and trimming data

One of the requirements of propensity score approaches is that there is common support. Common support bounds the sample within a limit which “implies that the test of the balancing property is performed only on the observations whose propensity score belongs to the intersection of the supports of the propensity score of treated and controls” (Becker and Ichino2002). Estimates without incorporating the common support that is by inclusion of households beyond the range of common support would be biased. Followed by Lunt (2014), we limit this analysis in the range of propensity scores for both children and adolescents in the flooded and non-flooded districts. Based on the common support restriction limits, only a small number of observations are dropped out from this analysis. That is why the reduction of the individuals — having extreme propensity score values — did not change the magnitudes and directions of all of variable of interest at all (see Column 4, Table 5).

Also suggested by Lunt (2014), we also used a data trimming approach by including all children and adolescents in flooded and non-flooded districts under certain propensity scores restrictions. First, we restricted this analysis to the first centile (Column 5, Table 5) and then to fifth centile (Column 6, Table 5). Despite excluding about 8.93% and 24.92% of observations in the first and fifth centiles analysis, respectively, the magnitudes of the 1st and fifth centile cases increased by 7.97% and 16.71% points, respectively.

5.4.3. An alternative source for the selection of flooded districts

In our previous analysis, it was assumed that the flooded districts were equally damaged by the disaster and consequently it affected the educational outcomes (i.e. out-of-school children and adolescents) equally across all flooded districts of Pakistan that may not be a good assumption for relying on the main findings. In order to test the validity of our findings, we do another robustness check by utilizing another source of data from National Disaster Management Authority (Government of Pakistan2011) for the selection of the flood-affected districts. As per NDMA, 78 districts of Pakistan were affected; out of which 49 and 29 districts were affected by the flood moderately and severely, respectively. The distribution of districts was done for the sake of providing urgent relief and support for their rehabilitation (Government of Pakistan2011). This study utilizes the NDMA data of severely affected households in 20 flooded districts and 55 in moderately affected districts in Pakistan, as the treatment group and previously the unaffected districts as the control group. The columns 7 and 8 of Table 5 present the findings. As expected, the continuity of education of children and adolescents in districts which were severely affected by the flood and disrupted badly in the flood year: the number of out-of-school children and adolescent in the flooded districts in the flood year 2010–11 increased.

This might happen because the situation of schools in Pakistan was already disturbing before the flood and the heavy flood in 2010 made it even worse. Around 11,000 schools were destroyed, and thousands of schools were replaced with shelter homes for community (Chuang et al.2018). The climate change badly affected the school-going children with a higher rate of absenteeism and lowered academic performances; comparing with other provinces of Pakistan, schools in Sindh Province were badly effected (Chuang et al.2018). According to Khan and Ali (2014), the education sector was badly affected in the 2010 Pakistani flood; counting the loss on education sector of approximately Rs. 26,464.3 million. Though schools have potential to add value for functioning activities during and post disasters, these are usually used as evacuation centers during disasters (Ilumin and Oreta2018). Wherever such disasters happen, the educational sector shall be given a top priority among other recovery strategies because, schools are the backbone of humanitarian effectiveness, decreasing susceptibilities and minimizing risk for future deathtraps (Chang et al.2013).

5.4.4. Covariates as additional controls in regression analysis

We used the inverse probability of treatment weighting (IPTW) method to adjust for confounding in our observational study. This method uses the propensity score to balance baseline individual characteristics in the flood and non-flooded districts by weighting each individual under study by the inverse probability of being affected by the flood 2010. In addition to that we also included covariates including age, number of household members, parental education of children and adolescents, location (rural vs. urban), family head a male, as additional controls to test the sensitivity of our main results. Our main findings remain robust (see Column 9 Table 5).

5.4.5. Regression based on district-year fixed effect

Our main analysis is based on province-year fixed effect, and one can argue that there might be some time trends prevailing in districts already independent of the flood which may distort the findings. Since our explanatory variable of interest is across districts that may create the problem of multi-collinearity in the main analysis, here, we test whether our findings are still robust if the estimates come from regression while controlling the district-year fixed effect. The results are still positive and significant at the 5% significance level (see Column 10, Table 5).

5.4.6. Consideration of other floods and earthquakes: since 2004 to 2015

Since Pakistan’s geographical location makes its population the most vulnerable to floods and earthquakes, we have also tested the hypothesis that, the floods and earthquakes during the last decade (2004 to 2015) have affected the education of children and adolescents. The results of this analysis are based on switching “on”/“off” (1/0) principle depending upon whether or not a district had been affected by any event starting from 2004 and it remained “on” once affected by a disaster. For this analysis, we used the EM-DAT’s data,² and ran five regressions separately by using the DiD model specification (two groups and two periods). The coefficients of interest (i.e. interaction terms) indicate that in almost all samples, (except the last one for which the control group squeezed significantly), the disasters affected the education of children and adolescents in the disaster-affected districts compared to their counterparts in unaffected districts (see Table A.2). These findings further support our main argument of this study, that is, the disasters in Pakistan are one of the leading causes that restrain children and adolescents from being in schools and continuing their education.

Due to its location, Pakistan is highly prone to natural calamities like floods, earthquakes and cyclones and without considering them, our main results will be biased. To identify the true impact of the 2010 flood on out-of-school children and adolescents, we consider other events (both floods and earthquakes from 2004 to 2015) using EM-DAT datasets and exclude districts that have been affected by floods or earthquakes in Pakistan. Here, flooded districts are those that were affected by the flood 2010 and have never been affected by any other floods or earthquakes before or after the flood 2010, while non-flooded districts are those that have never been affected by any floods or earthquakes including the flood 2010 during 2004–2015. Although the sample size is reduced to about half, it shows that the 2010 disaster has even a greater impact once isolated from other disasters. Excluding the effects of other natural disasters, our main results show an improvement that is not only statistically significant but also more than fourfold in magnitude (see Column 8 of Table 3).

5.5. Implications associated with other similar natural disasters: Particularly Pakistani flood 2022 and the recent pandemic COVID-19

Our findings are relevant at a time when Pakistan is facing a relatively heavy flood in 2022. Since mid-June 2022, the flood caused by monsoon rains have affected almost all provinces of Pakistan. So far, 1,033 people have been killed and 1,527 injured in the floods. About 30 million people have been affected and 1 million houses have been completely or partially destroyed. Figures from the provincial education department show that 17,566 schools were damaged or destroyed by the 2022 floods: 15,842 in Sindh, 544 in Balochistan and 1,180 in Punjab (UN Office for the Coordination of Humanitarian Affairs, 2022). The closure of schools due to summer vacations are postponed. Because the recent 2022 flood is comparable to the 2010 flood, our results suggest that without the support of humanitarian organizations and other government interventions, the impact of the 2022 flood on children and adolescents’ education will be catastrophic.

Our findings are also relevant in a time in which the current pandemic COVID-19 has been devastatingly affecting the economies of developing countries like Pakistan and disrupting their processes and systems of human capital accumulation. Having an impoverished population, the closure of schools due to COVID-19 shutdowns may further widen the gap between the expected years of schooling and adjusted years of schooling, possibly upsurge dropout rates and out-of-school children (Mian and Chachar2020). The impact of schools’ closure would extend beyond schools affecting not only children’s current education and learning but also their future potential earning capabilities. Mian and Chachar state that the closure of schools in Pakistan may worsen the existing divides in terms of enlarging the prevailing rural–urban, gender and socioeconomic divide with possible escalation of child labor both in the short run and in the long run in the country. Interruption of schooling which mostly happened during the time of disasters has profound and long-lasting adverse effects on children education. For example, the closure of schools during the earthquake 2005 in Pakistan, the areas accounted for 10% of loss in test score and even if the children went to schools when the situation became normal, their tests scores were even poorer and this may be due to children falling behind in curriculum and facing difficulty in catching up with their counterparts in unaffected regions (Andrabi et al.2020). The author also estimated that if such losses continue to their adult lives, they will lose about 15% of their earning per annum throughout their lives. The findings of this paper may help policymakers for identifying the most vulnerable children and adolescents (especially female and people living in rural areas) and for devising interventions to minimize the impact of out-of-school children and dropout due to schools’ closure either during or after the COVID-19 outbreak.

6. Conclusion and Recommendations

Natural disasters such as floods, inter alia, are major causes of destruction of many public and private infrastructure and structures including road and electricity networks, irrigation systems, water supplies, schools, and hospitals. It takes several years for a country to come to a normal situation after being affected by a disaster because the impacts of disasters on society remain permanent or persistent. Recently, Pakistan has been experiencing severe floods since June 2022 that have affected the country’s nearly 30 million population, affecting not only economic and business activities, but also destroying educational infrastructure and summers. Holidays and school closures are being postponed. Thousands of schools have been destroyed by the floods and others are being used as flood shelters. In 2010, Pakistan experienced a similar devastating flood which killed thousands of people and destroyed millions of houses, crops, schools and hospitals. The aftermath impacts of the flood were too extreme that individuals or sectors in every corner of the flooded areas in Pakistan experienced the aftermath shocks. The existing dilapidated education system in the country, particularly girls’ schools and schools in rural regions are more susceptible to such disasters.

In this study, we investigated the impacts of the flood 2010 — beyond the physical damages and economic losses — on accumulation of human capital in the flooded communities of Pakistan. From our analysis, we found that the flood undermined the education outcomes (i.e. increased out-of-school children and adolescents and dropouts from schools) in flooded households further. We found a short-run adverse impact of the flood on admissions of children and adolescents in flooded households compared to their counterparts in non-flooded households during the flood year (2010–11). On an average, 70 fewer children and adolescents out of 1000 in the flooded households were not admitted during the flood year. The flood severely affected the educational outcomes of female and rural children and adolescents more than their respective male and urban children and adolescent counterparts.

We further expanded our analysis to check the potential causes of dropping out of children and adolescents from schools during and after the flood and we found that the distance between schools and home, household and domestic works, and parental approval for continuing education are the most pertinent factors that caused dropouts from schools in general and a huge number of female dropouts in particular in the flooded districts during the flood; in most cases, the effects remained persistent to 2–4 years after the flood. For a country like Pakistan which has the second highest number of out-of-school children after Nigeria, the impacts of flood on education outcomes seriously needed policy interventions. In addition, the matter of dropping out from schools due to the disasters make the situation even worse to achieve the education targets of the Sustainable Development Goals (SDGs) by 2030. In order to make communities in developing countries safer and more resilient to disasters under the Sendai Framework for Disaster Risk Reduction 2015–2030 agreement, the analysis of socioeconomic impacts of disasters at a community or household level provides a better understanding of assessments of possible policy interventions during and after the disasters.

As per the current situation of COVID-19 pandemic, which has also been disrupting schooling of children due to stringent measures of lockdown/shutdown in many developing countries, the findings of this paper have greater implications in how we respond to the issues associated to educational outcomes of children and adolescent in developing countries during and after the pandemic COVID-19. Educational policymakers should consider ways for identifying the most vulnerable children and adolescents in disaster-affected areas during and after the disasters. In developing countries, children and adolescents, if they are female (than male) or living in rural areas (or urban areas) are more vulnerable to access to basic education, particularly during or after a disaster, need policymakers’ immediate attention and assistance for intervening to delay the interruption of their education.

Due to limitation of data on the quality of education in Pakistan — in terms of students’ learning outcomes which may continue in the very long-run after a disaster — the findings of this study are only relevant to policy implications of access to and interruption of schooling of children and adolescents in developing countries. But still, access to education in developing countries especially countries like Pakistan — which has a huge number of out-of-school children is crucial to achieve the targets in UN SGDs.

Annexure

Notes

¹ Since the intervention variable of this study is a dummy which holds a value of 1 if a district was flooded (either severely or moderately) and 0 otherwise, it can only matter when the district was declared flooded in one map and not flooded in the other and there is only one such district, Lasbela which we excluded from the study.

² EM-DAT, CRED/UCLouvain, Brussels, Belgium (www.emdat.be)