Long-term impacts of an early childhood shock on human capital: Evidence from the 1999 economic crisis in Ecuador

This paper provides evidence on the lasting effects of the 1999 economic crisis in Ecuador on human capital formation. We show for children born during the crisis that the negative repercussions are still observable more than 10 years after macro-economic recovery. Taking advantage of micro-level data collected in 2012 and 2014, we assess long-term impacts on health and education. After controlling for age-in-months and survey effects as well as a linear birth year-cohort trend that varies by region, we find that after 12–16 years, the cohorts born during the recession report height-for-age Z -scores that are 0.003 standard deviations (SDs) lower for each month of exposure and have 0.002 fewer years of schooling per month exposed compared to the cohorts born outside the recession years. Children exposed to the entire crisis are 0.063 SDs smaller and have 0.042 years less of schooling. Girls have lower health outcomes than boys. Concomitantly, we show that selective childbearing or excess infant mortality are unlikely to drive our results. The persistence of the negative effects points to the existence of a poverty trap suggesting that policy interventions in response to (economic) crizes should be extended beyond macroeconomic recovery to counteract long-term, micro-level consequences.

protection, recessions also deteriorate the quality of healthcare services (Paxson & Schady, 2005). All these channels result in children being especially hurt by shocks like economic crizes, possibly resulting in a decrease in the returns to future investments in their human capital (Cunha & Heckman, 2007;Gluckman et al., 2008).
A question that remains is whether economic crizes impact children's human capital formation only during the measurable macroeconomic downturn, or whether the effects persist over time, leaving a mark on the affected populations for their lifetime and possibly perpetuating poverty. To answer this question, the study at hand presents evidence of the effects of the Ecuadorian economic crisis from 1999 on the long-term human capital formation of children who were born during the crisis. The long-term implications are measured 12-16 years after the crisis by combing data from two surveys -the National Survey of Health and Nutrition of 2012 (ENSANUT) and the National Survey of Life Conditions of 2013-2014 (LCS). After controlling for age-in-months and survey effects as well as a linear birth year-cohort trend that varies at the geographical level (city, region or province) we find that the cohorts born during the crisis report height-for-age Z-scores that are 0.003 SDs lower per month of crisis exposure. Being exposed to the entire crisis led to a loss of 0.063 SDs in height-for-age during early adulthood, representing a loss of 4.9-5.9 mm for a fully exposed child. Moreover, per extra month of exposure, children have 0.002 fewer years of schooling. Thus, exposure to the entire crisis reduced schooling by roughly 0.042 years. In the heterogeneity analysis we further find evidence that girls are most adversely affected; the negative health effects are largely carried by them.
The crisis was the result of a sequence of faulty economic decisions that partially derived from difficulties in the external sector stemming from the contraction of the oil price. The resulting macroeconomic instability led to a large fiscal deficit, a contraction in real economic activity, a severe banking crisis, and a rise in poverty (Parandekar et al., 2002). 1 year before the crisis, Ecuador's GDP per capita grew at a rate of 2% (Appendix A, Table A1); as a result of the crisis, GDP per capita decreased by 6.5% in 1999 and 40% of the population was affected by poverty (Vos, 2000). In addition, the crisis further aggravated unemployment, which increased from 9.2% in 1997 (before the crisis) to 14.4% in 1999. The increase in unemployment was exacerbated by an increase in the size of the informal economy and underemployment (Appendix A, Table A1).
Simultaneously, the monetary instability produced hyperinflation with rates of 10% per month in late 1999, leading to almost a doubling of prices from the beginning of 1999-2000. With the decrease in incomes and the rise in unemployment and prices, the crisis produced a humanitarian emergency with a serious deterioration of livelihoods, most notably for the most vulnerable populations. Household consumption which had increased 5% on average in the 5 years before the recession dropped by 12.2% in 1999. Because of the accompanying reduction of household expenditures on nutritious food, malnutrition was found in 12.5% of all children under the age of 5. Public social expenditure decreased from 881.1 million US$ in 1998 to 842.5 million US$ in 1999 reaching a minimum of 542.6 million US$ in 2000 (CEPAL-ILPES, 2001). 1 We assess the 1999 Ecuadorian crisis because of these huge macroeconomic impacts. The analysis allows us to contribute to the literature on persistent effects of recessions on human capital formation. The resulting contributions are five-fold: First, economic crizes directly weaken a household's economic status and concomitantly deteriorate the children's "environment" . Besides, the study of an economic crisis at national scale allows us to identify average effects estimated from nationally representative data, making the findings particularly relevant for policymaking. Yet, since the crisis is national it is also harder to identify a control group, leaving us to rely on cohort variation in exposure at birth. Second, the long-term −that is, more than 10 years− consequences of economic shocks on human capital formation are not well understood and existing evidence is inconclusive (Almond & Currie, 2010). To assess the long-term effects of the crisis we employ two national surveys: the 2012 ENSANUT and the 2013-2014 LCS conducted 14 and 16 years after the onset of the crisis, respectively. The identification of long-term, micro-level impacts complements studies of macroeconomic recovery that assess the dollarization of the economy, inflation dynamics, economic growth and poverty. Once the longevity of shocks is accounted for, the costs of an economic crisis are likely to be considerably larger. Third, to the best of our knowledge, the long-term consequences of economic crizes have not been explored using population representative data, but only for sub-samples. Existing research rests on relatively small samples of individuals who live in a specific area or belong to a specific socio-economic group, and tends to exploit geographic variation in exposure or variation in the subpopulation exposed for identification purposes (Alderman et al., 2006;Cutler et al., 2007;Hidrobo, 2014;Rosales-Rueda, 2018;Van Den Berg et al., 2006). Our estimates reflect the average effect based on nationally representative surveys exploiting cohort variation in exposure to the crisis at birth. Fourth, the literature tends to separately study health outcomes such as mortality (Cutler et al., 2007;De Cao et al., 2022;Van Den Berg et al., 2006) and anthropometric measures (Banerjee et al., 2010), and education outcomes such as schooling (Gutierrez, 2017). Human capital formation theories suggest health losses lead to future losses in education and eventually decrease productivity (Grossman, 1972). Therefore, we assess both outcomes.
Fifth, the effects of economic downturns on the short-term health of newborns have been examined for developed countries. While some studies find countercyclical effects on the health of those born during a crisis (Dehejia & Lleras-Muney, 2004;Van den Berg et al., 2020), others find that health deteriorates with economic downturns (Alessie et al., 2018;De Cao et al., 2022). Most of the positive effects seem to be driven by selective fertility (Dehejia & Lleras-Muney, 2004), and in contexts of strong social protection policies (De Cao et al., 2022). Our paper analyzes the impact of a recession in the context of a developing country with weak social protection policies.
The remainder of the paper is structured as follows: A literature review is presented in Section 2; the data sources and dataset construction are introduced in Section 3 and Section 4 outlines the empirical strategy. Descriptive statistics can be found in Section 5; Section 6 discusses the results and Section 7 concludes.

| LITERATURE REVIEW
The study of an economic crisis allows us to investigate the impact of a direct shock on income and consumption .
In terms of short-run effects, the literature finds a deterioration of nutrition indicators after economic crizes in developing countries. Lazzaroni and Wagner (2016) present evidence for Senegal showing that natural and economic shocks negatively affect child health; the shocks explain between 24.5% and 31.0% of the variation in weight-for-age. Similarly, for The Cameroon (Pongou et al., 2006), Latin America (Ferreira & Schady, 2009), and East Asia (Bhutta et al., 2008) economic crizes cause underweight, wasting, anemia, and a depletion of micronutrients.
Yet, the question whether crisis-induced, short-term, negative health outcomes persist is less studied. Banerjee et al. (2010) show that the phylloxera crisis in 19th century France generated a large income shock resulting in reduced long-term adult height by 1.8 mm for children born in regions affected by the insect. Similarly, Van Den Berg et al. (2006) demonstrate that in The Netherlands, adult survival was reduced by several years for those exposed to an economic recession at birth. Yet, both studies are conducted in developed countries. For the case of a developing country, Alderman et al. (2006) establish for a sample of roughly 600 Zimbabweans that they would be 3.4 cm taller in adolescence had they been of the median stature of preschoolers from a developed country. 2 However, these estimates result from the comparison with the international reference population and not children of the same age in the same country but not exposed to malnutrition. Cutler et al. (2007) do not detect any change in adult mortality for the birth cohorts affected by the Dustbowl era of the Great Depression of 1930 and Angelini and Mierau (2014) show that in Europe childhood health of individuals born during recessions is procyclical. Yet, the latter rely on self-reported, retrospective data that are likely to suffer from recall and selection bias (survivors only can be interviewed).
While overall, evidence about the long-term effects of economic crizes is limited, it is even more scares for developing countries (Gutierrez, 2017) and tends to derive from small samples (Alderman et al., 2006). Therefore, this paper complements existing evidence by providing average long-run effects of being born during an economic crisis using nationally representative data from a developing country.
The paper analyses the recession that Ecuador suffered in 1999. There are studies that examine the impact of this crisis, but none considers long-term effects. Vos (2000) and Salgado (1999) found immediate adverse effects on poverty, which escalated to 40%, and on children's undernutrition, which increased from 17% in 1998 to 20% in 2000.
Two studies assess the impacts of the Ecuadorian crisis using the data created to assess the cash transfer program Bono de Desarrollo Humano. Hidrobo (2014) presents evidence that 5 years after the crisis the impact was still observed in a sample of the poorest population quintile. Being exposed to the crisis for 1 year, significantly decreased HAZ, by 0.08 SDs. In addition, Hidrobo (2014) assesses the effect of the crisis on the receptive language skills of children of the same age but with different durations of crisis exposure. Compared to the baseline average she finds a 6% reduction in vocabulary test scores after 1 year of crisis exposure. Rosales-Rueda (2018) uses the same dataset to study the effects of the 1998 floods in Ecuador caused by the "El Niño" phenomenon; five to 7 years after the event these extreme floods are associated with lower stature of 0.03 SDs and lower scores on cognitive tests of 0.05 SDs among the children who were exposed in utero for 1 month.
However, both studies (Hidrobo, 2014;Rosales-Rueda, 2018) rely on the Bono de Desarrollo Humano data that was originally built to measure the impact of the program on eligible, low-income households. The children who were surveyed were poorer and the mothers in the sample younger than those of the general population. In contrast, the paper at hand examines the wellbeing of the entire cohort-generation. Moreover, we assess the persistence of impacts 12-16 years later and thus expand the study horizon of the earlier work.

| DATA SOURCES AND DATASET CONSTRUCTION
We rely on two surveys: the National Survey of Health and Nutrition of 2012 (ENSANUT), and the National Survey of Life Conditions of 2013-2014 (LCS). They provide anthropometric measurements for surveyed individuals. 3 The samples for both surveys were drawn based on the National Census of Population and Households of 2010 and are nationally representative, cross-sectional datasets. The 2012 ENSANUT is representative at the national and provincial level, as well as with respect to gender, ethnicity and age categories. Between 2011 and 2013 the survey collected information on 19,949 households and 92,502 individuals. Anthropometric information is available for individuals from 0 to 59 years of age. 4 The second dataset, the LCS 2013-2014, is a multipurpose survey which gathers information about different dimensions of household and individual welfare and is representative at the national and provincial level, for different age groups and areas. LCS includes multiple health variables among which are child weight and height. The LCS survey contains data for 28,970 households and 109,694 individuals from 0 to 99 years of age. 5 Anthropometric information is available for all individuals in the survey.
We assess the long-term impact of the 1999 economic crisis with the two surveys exploiting cohort variation in exposure to the crisis at birth. To this end we rely on the definition of a birth cohort as a group of individuals born in the same year. We are interested in the cohorts that were born during the 1999 economic crisis, that is, children born as of July 1998 andup to March 2000. In the 2012 ENSANUT the treated birth cohorts are between 12 and 14 years old. In the 2013-2014 LCS the treated cohorts are 14-16 years of age. By pooling these two surveys we can find different birth cohorts that are 12-14 years old in the 2013-2014 LCS that serve as comparison for the affected cohorts from the 2012 ENSANUT. Similarly, we find different birth cohorts of the same age for the 14-to 16-year-olds from the 2013-2014 LCS in the 2012 ENSANUT. This gives us an age range of 12-16 years for whom we can always identify a group of children born during the crisis and a group born outside the crisis years (Block et al., 2004). In addition, we add a control group of younger (10-11 years old) and older (17-18 years old) children. The children in these later two age groups are not affected by the 1999 economic crisis at birth no matter which survey they are from. Taken together, we have 11 birth year and 106 birth year-birth month cohorts from 1994 to 2004 (compare Appendix B for details).
We combine the data from the two surveys as they are both randomly selected from the same population (Bernal & Peña, 2011). The two surveys have already been combined to assess the nutritional status of the country, across provinces and over time (World Bank, 2007). Importantly, in addition to the birth-year cohort effects that we interact with city/region/province specific effects to allow for local birth cohort trends, we also account for age-in-months specific effects and survey specific effects.
As outcome variables, we focus on height-for-age and years of education. We employ height-for-age since this anthropometric measure is commonly considered as measuring the stock of nutritional inputs (Rieger and Wagner, 2015). Height-for-age reflects cumulative growth as a result of nutritional conditions, that is, a deficit in height-for-age reflects past suboptimal health or nutritional intake (WHO, 1997).
The anthropometric information in the surveys contains the height of each individual in centimeters. As children are in the growth process, the comparison of anthropometric measures between children of different ages and across gender is problematic. By transforming the raw anthropometric information from centimeters to Z-scores, we assure the comparability of the data. The Z-score measures how many SDs below or above the population mean the anthropometric measurement of an individual child is. The population distribution is calculated according to an international reference population (De Onis et al., 2007;WHO, 2007).
The value of the Z-score is linear, which means that any fixed interval of Z-scores is associated with a fixed height difference in centimeters at a given age. 6 The WHO (2007) establishes that children with less than minus two SDs are stunted. Stunting refers to growth retardation and is the result of long-term nutritional deprivation caused by poor diets or recurrent infections (WHO, 2010). We drop Z-scores bigger than 5 in absolute terms since they are empirically not feasible and likely represent measurement or data entry errors. 7 This leaves us with a sample of roughly 17,000 observations of children between 10 and 18 years.
According to the theory of human capital accumulation, a loss in the health stock will impact the education stock and in combination will deteriorate life-time productivity (; Ferreira & Schady, 2009). We evaluate both stages of human capital production; the education outcome allows us to detect whether the second stage of human capital deterioration is present. Yet, we do not have data to measure the long-term impact of the crisis on labor market productivity. Moreover, the availability of educational information is restricted because the 2012 ENSANUT survey focusses on health outcomes. The only educational variable that is present in the two surveys is completed years of schooling. Since this variable has been used in earlier studies to measure the impact of crizes on education (Gutierrez, 2017;Maccini & Yang, 2009), we consider it a valid educational measure as long as the highest grade completed and the reported years of attended schooling correspond to the grade-specific age of the child.

| EMPIRICAL STRATEGY
Key for our empirical strategy is defining the crisis period. We apply the following definition: A recession is defined as two consecutive quarters of negative economic growth. It ends with two consecutive periods of positive economic growth (Appendix A, Figure A1). Consequently, and in line with previous studies, the crisis is defined as taking place between July 1998 and March 2000 (Hidrobo, 2014;Parandekar et al., 2002). The end date coincides with the dollarization of the economy and a bundle of economic and social policies (Appendix A, Table A1).
Treated individuals are those who were exposed to the crisis at birth -that is, those born between July 1998 and March 2000. Our main explanatory variable is length of exposure to the crisis since birth, measured in months. The length of exposure measure allows us to account for the fact that cohorts that were born later during the crisis were less exposed (Hidrobo, 2014). The length of exposure measure ranges from 0 to 21 in line with the months an individual was exposed to the crisis since birth. For example, a child born in August 1998 will have been exposed to the crisis for 20 months. With this specification we measure the effect of one extra month of exposure. We compare the children born during the crisis with those born before or after the crisis. For the pre-and post-cohort control group the time of exposure takes a value of 0.
We include the cohorts born before the crisis to avoid an overestimation of the impact of the crisis due to the secular growth trend in population height (NCD, 2016 and2020). Because of the secular trend children born in 1998 should be taller than those born in the years before and should be shorter than children born in the years after. If we only consider the comparison between children born in 1999 and those born after the crisis, we might overestimate the impact of the crisis. Children born after the crisis are taller in part because younger cohorts are taller. If we include the cohorts born before the crisis, that should be shorter, and we nevertheless find that the cohorts born during the crisis are shorter, we can infer that it is a result of the exposure to the crisis.
However, this strategy generates a dilemma. A child who was born in the last month before the crisis has an exposure of 0. But even if she was not born during the crisis, she was exposed during her first months of life which are essential periods for human capital accumulation. As a robustness check, we compared children who were born during the crisis with only those born after the crisis.
Since a child born in 2000 will be 12 years old in 2012 and 14 years old in 2014 and anthropometric measures differ by age group, variation in age can confound the results. To avoid this problem, we account for age specific effects. Age is measured in months using information on the month of birth and the month in which the anthropometric data was collected.
In addition, we include birth-year cohort trends to capture trends in infant and child health, for example, those due to changes in technology. We allow this trend to vary across geographical regions (city/region/province).
Thus, the empirical specification compares the anthropometric outcomes of the children exposed to the crisis at birth with those not exposed to the crisis at birth. We identify the impact of the crisis by comparing children of the same age who were differentially exposed to the crisis, accounting for geographic-specific birth year cohort trends and age-in-months fixed effects. The empirical specification is estimated via linear regression and given by: where Y iltas is the human capital outcome of individual i born in location l in year t, with age in months a and observed in survey s. C tas is the number of months individual i was exposed to the crisis since birth. Child specific characteristics are collected in X iltas and include the gender of the child and whether she is an only child along with household characteristics such as the household's poverty level, 8 whether the household lives in a rural area, whether the household has access to clean water, the level of education of the household head and the number of family members and children in the household. These covariates are measured when the surveys were conducted; we have no information about the household situation at childbirth. We expect that in relative terms, that is, across the distribution, there was little change and that the contemporaneous variables are a good proxy for the ones at birth. Next, we include a linear birth year cohort trend that is interacted with a geographical indicator (i.e., city, province, region or rural/urban), ∑ =1 * Geo for x = 1, …, n. In addition, we include age-in-months fixed effects, .
This combination of a linear birth-year cohort trend that varies at the geographic level with age-in-months specific effects allows for the comparison of the height-for-age of children exposed to the crisis relative to those of the same age who were not born during the crisis. At the same time, it allows us to compare children of different ages who were born during the crisis. For instance, a child who was born in December 1999 and who was measured in the ENSANUT survey in February 2012 will be 12 years and 3 months old implying that the child belongs to the age-month group of 147 months and to the birth cohort 6. Likewise, a child who was born in September 2002 and who was measured in the LCS survey in December 2014 will be in the same age-month group of 147 months but in the birth cohort 9. In turn, a child who was born in December 1999 and who was measured in the LCS survey in December 2014 will be in the age-month group of 180 months but in the birth cohort 6 because of the year of birth.
In other words, we compare the height-for-age Z-score (HAZ) of the first child in our example who is 147 months old and was born during the crisis with the HAZ of the second child who is the same age, but who was born in a non-crisis period. In addition, we follow the birth year cohorts exposed during the crisis by measuring the HAZ of the first child who belongs to cohort 6 and was measured in the first survey with the HAZ of the third child who belongs to the same cohort but was measured in the second survey. Put differently, the effect of the crisis will be purged of age effects and cohort trends.
Furthermore, we include a survey effect, γ s , that is 1 if the observation stems from the 2012 LCS survey and 0 otherwise. The survey dummy captures survey specific differences that could drive the results, such as different training and/or questionnaires, etc. The error term is denoted by ε iltas . Similar to Hidrobo (2014), the error term contains unobserved individual and family/mother characteristics as well as unobserved time-variant community characteristics not captured by the trend. The error term is cluster wild-bootstrapped to account for the fact that we only have a small number of regional clusters, our level of clustering. We employ population weights in the analysis to derive nationally representative results.
We also implement a specification with family fixed effects. We can do so because we observe families with multiple children of whom at least one is born during the crisis. By employing family fixed effects, we capture time-invariant characteristics of the families, preferences and the home environment, that could impact the decision to have a child during the crisis as well as child outcomes. To control for selection into giving birth during the crisis, we compare siblings born into the same families, that is, with the same mother. Yet, the family fixed effects specification only accounts for fixed family/ mother characteristics but does not account for selection into childbearing based on unobserved time-variant family/mother characteristics.

| DESCRIPTIVE STATISTICS
For the main analysis, we consider 5236 children who were born during the crisis and 11,851 children who were either born before or after the crisis.
The individual level characteristics are presented in Table 1. Balance is obtained on 6 of the 9 characteristics/categories of characteristics and even for those variables where no statistical balance can be observed, the practical differences are small. The joint orthogonality test for all covariates has a p-value of 0.214, supporting the assumption of overall balance across the two groups.
The crisis-affected and unaffected children are balanced in terms of location with 54% of them living in rural areas. Access to clean water is 1.5% points higher among treated households and this difference is statistically significant at the 5% level. On average treated households have the same size as control households of 5 members and do not differ in their number of children; they tend to have 3 children. Similarly, across situations we observe that 12% of the sampled individuals are only children. Yet, we find more girls among the treated children. This difference amounts to 2% points, and is statistically significant at the 1% level. The sampled children are around 13 years old, with the treated children being some months older than the control children at the time of the surveys. Note that we have taken the definition of wealth quintiles used in the original surveys but since we consider a sub-sample of 10 to 18-year-old children, the share of individuals in a given quintile does not equal 20%. We observe an overrepresentation of the lowest wealth quintiles in the surveys with balance across the treatment and control group. Turning to the level of education of the household head more than half the sample has primary education and another 31% secondary education. The categories of educational attainment are balanced across the treatment and control group.
At the bottom of Table 1, we present the outcome variables: the children born during the crisis present a lower HAZ compared to the children not born during the crisis. The difference is 0.081 SDs and statistically significant at the 0.1% level. Years of education are higher for those born during the crisis. The difference is about 0.23 years (p-value ≤ 0.001). However, when comparing the level of education (none, primary, secondary) attained by the children, those born during the crisis have a lower level on average compared to those not affected (p-value ≤ 0.000). T A B L E 1 Descriptive statistics non affected and affected groups.

| RESULTS
The main results for the impact of the 1999 economic crisis in Ecuador on child health are presented in Table 2. Tables 3-5 and Appendices D to F present robustness checks and alternative specifications. Table 6 presents the results for the educational outcomes.

| Impact on health-Height-for-age
Even at a glance we can see that the crisis is negatively associated with child health (Table 2). Across specifications, for each additional month of exposure to the crisis, children's Z-scores are 0.003 lower. This implies that children exposed for the full duration of the crisis of 21 months are 0.063 (0.003*21) SDs smaller. 9 This effect is similar no matter whether we allow time trends to vary across cities, provinces, regions or between rural and urban areas.
(1) (2) Note: The individual level child and household characteristics that are included are as follows: (a) being an only child, (b) being female, (c) dummies for household income quintiles, (d) dummy for having access to clean water, (e) the number of family members in the household, (f) the number of children in the household, (g) dummies for the level of education of the household head and, (h) a dummy for living in a rural area. The coefficient estimates associated with the child and household level characteristics can be found in Appendix C, Table C2. Wild-bootstrapped standard errors clustered at the regional level are presented in parentheses. + p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001.

T A B L E 2
Main results: The effect of crisis exposure on height-for-age Z-score.
(1) (2) Note: The individual level child and household characteristics that are included are akin to the earlier presented specifications. For details see the note to Table 2. The coefficient estimates associated with the child and household level characteristics are available upon request. Wild-bootstrapped standard errors clustered at the regional level are presented in parentheses. + p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001.

T A B L E 3
The effect of crisis exposure on height-for-age Z-score excluding those born before the crisis from the control group.
These results are consistent with the ones obtained by Hidrobo (2014) who showed that three to 5 years after the crisis the poorer strata of the Ecuadorian population who were exposed for an additional month experienced a decrease in HAZ by 0.008 SDs (using a comparable specification with a province-specific birth year-birth month trend and age-fixed effects). Employing a specification that, similar to Hidrobo (2014), includes a birth year-birth month trend interacted at the geographical level and age-fixed effects as well as birth year-birth month clustering of the standard errors we equally obtain a crisis exposure effect of 0.003 SDs (Appendix C, Table C1). Differences in the coefficient estimates between our main results (Table 2) and the results using birth year-birth month trends (Appendix C, Table C1) tend to be only visible at the level of the fourth decimal; the statistical significance of the findings is reinforced. The impact is also similar to the one found by Rosales-Rueda (2018) 5 years after the crisis who shows that 1 month of in utero exposure decreased HAZ by 0.03 SDs. Since one Z-score is associated with a length of about 7.8 cm for children between 14 and 16 years, the results imply that the loss in stature caused by the crisis would be around half a centimeter.
The results complement earlier findings (Hidrobo, 2014;Rosales-Rueda, 2018) as we obtained them from population-representative samples and we show that the negative impacts are still observable 12-16 years after the crisis. Moreover, as growth is almost completed in adolescence, this implies that the Ecuadorian crisis of 1999 has impacts likely to last into adulthood.
The specifications discussed so far ( Table 2, Columns 1-5) do not account for family fixed effects. Yet, the childbearing decision could be influenced by the economic downturn leading to a potential bias in the estimated effect of the crisis on child outcomes. Therefore, we also implemented an empirical specification with family fixed effects to compare siblings, that is, children born by the same mother. We included only the child observations that have siblings along with the siblings' observations in the family fixed specification. This results in a sample of 16,125 children. It implies a reduction of 962 child observations or 5.6% of the original sample of 17,087 observations. In line with Hidrobo (2014) we argue that the family fixed effects capture time-invariant aspects such as preferences which are related to child outcomes as well as the length of crisis exposure. Further attempts to address possible selection into childbearing are discussed in Section 6.4. After including family fixed effects, the effect of an additional month of exposure to the crisis is a reduction of 0.003 SDs but its significance is reduced to the 10% level.
The coefficients on the individual-and household-level covariates suggest that children from richer households, from households that have access to clean water, and households where the head has a higher level of education are taller. In turn, girls, children from rural areas, and those coming from households with more children have a lower HAZ (Details: Appendix C, Table C2). All these coefficient estimates are statistically significant at conventional levels across specifications. At least some of the control variables such as the number of children, poverty, and location, could have been impacted by the crisis, and hence be post determined. Excluding these control variables and only keeping the gender and education of the household head, our main results are similar (compare Table 2, Column 1). The impact is −0.003 SDs and significant at the 10% level, implying that our results are not sensitive to the inclusion of potentially post-determined variables. 10 So far, we compared children born during the crisis with those born before or after the crisis. We assumed that birth is the critical period where bad nutrition can generate long-term impacts. However, those children born before the crisis might have been negatively affected by the crisis. In Table 3 we show results comparing only children born during the crisis with later born children. The estimated impacts tend to be slightly bigger in absolute terms (reduction of up to 0.005 SDs in HAZ per additional month of exposure) and less precisely estimated, suggesting that due to the secular growth trend the comparison of children born during the crisis with only children born after the crisis might result in an overestimation of the negative impact of the crisis. 11 At the same time, we fail to reject equality of the coefficient estimates associated with crisis exposure presented in Tables 2 and 3. 12

| Heterogeneous effects across boys and girls and in-womb exposure
Next we examine heterogeneous effects of crisis exposure by gender. In addition, we also analyze effects of in-womb exposure to the crisis.
The existing literature finds gender bias in resource allocation (Behrman, 1988;Chen et al., 1981;Gupta, 1987;Miller, 1997), making the impacts of economic crizes likely to be gender specific. For example, Chen et al. (1981) show that mortality and severe malnutrition among Bangladeshi girls up to 4 years of age is 1.45 and 2.82 higher, respectively, compared to boys. This pattern worsens during crisis. Behrman (1988) demonstrates that Indian households prioritize the nutrition of boys during the lean season.
We also find evidence that is consistent with gender bias in resource allocation (Table 4). Yet, it could be that females are more sensitive to stress in the household. Across specifications, we consistently find a reduction of 0.005 SDs in HAZ per additional month of exposure for girls (Table 4,. In turn, the estimates of the effects are not statistically significant for boys. While the coefficient associated with crisis exposure is negative for boys, it is smaller in absolute magnitude and not statistically significant (Table 4, . The results indicate that compared to the full sample and the sample of boys, girls are more adversely affected by the crisis. 13 This finding might be driven by gender selection bias at birth. Therefore, we calculated the impact of the crisis on the number of children, the number of girls and the ratio of girls over the total number of children. Implementing a time series analysis that consists of the 108 months of observations in our sample we calculated the monthly number of children, girls, and the ratio of both and regress this monthly data on an indicator variable that is equal to 1 for crisis months and 0 otherwise. We show that there is no significant relationship between the crisis and any of the three outcomes, indicating that the crisis has likely not affected the decision of parents to give birth to girls (Appendix C, Table C4).
Next, we turn to in-womb exposure. So far, we have assumed that the critical period of exposure is birth and the months after. In this section we explore the impact of the crisis for children exposed in utero. We constructed the following exposure indicators: We assigned the indicator of 1 to those children who were in-womb during the crisis, and 0 for those who were in-womb after or before the crisis. We show in Table 5, Column 1 that the effect is positive but statistically insignificant.
Next, we implement a specification where the control group consists only of those born 1 year after the end of the crisis, so we exclude the possibility of any crisis exposure while in womb. Yet, the effect is similarly small and imprecisely estimated suggesting that in-womb exposure is not what mattered for long-term lower stature (Table 5, Column 2).
The third and fourth specification are for the intensity of exposure in-womb, making the comparison relative to those individuals who were in-womb before and after the crisis (Column 3) or only after the crisis (Column 4). Again, we cannot detect any statistically significant effect. Finally, we estimated three specifications that account for the trimester of pregnancy during which the children were exposed to the crisis. The effects associated with crisis exposure at any trimester are again statistically insignificant (Columns 5-7). Thus, across specifications that account for in-womb exposure we do not identify any significant effect.

| Impact on education-Years of schooling
Since health is an important input for human capital formation, brain development, educational success, future labor supply and productivity (Ferreira & Schady, 2009;Gutierrez, 2017;Hoddinott et al., 2008;Maluccio et al., 2009), we now turn to educational outcomes. Strauss and Thomas (2007) and Barker (1998) show that children who suffer from malnutrition during early-life adapt their metabolism permanently to a lower level resulting in slower growth but assuring survival. This increases the probability of stunting since it produces lower future growth (Alderman et al., 2009). Related, WHO (2010) describes that children who suffer from long-term growth retardation often present delayed mental development, with subsequent lower school performance resulting in lower economic productivity. This is reinforced by Case and Paxson (2008) who followed a birth cohort until middle age, showing that child health is the most important mechanism for the intergenerational transmission of economic status. Other studies suggest that the length and height of newborns are strongly linked with future cognitive ability (Case & Paxson, 2008;Richards et al., 2002).
If the Ecuadorian crisis permanently lowered children's metabolism and growth rates (Strauss & Thomas, 2007) and delayed their mental development (WHO, 2010), we may observe less educational achievement. Results are presented in Table 6. Indeed, there is a significant negative impact of being exposed to the crisis on completed years of schooling. An extra month of exposure to the crisis is associated with 0.002 fewer years of schooling (Column 1). This means that a child exposed to the entire 21 months of crisis would have 0.042 fewer years of schooling. This is a small yet precisely identifiable impact. This impact is smaller in absolute magnitude compared to a comparable study by Gutierrez (2017) who finds 0.12 years of schooling lost in response to the Peruvian crisis in the 1980's. We further show that the effect of the crisis on schooling does not seem to be gender neutral. While we obtain the same coefficient estimates for girls and boys, only the estimate for girls is precisely identified which might aggravate the gender wage gap ( Table 6, Columns 5 and 6). Yet, we fail to reject equality of the coefficients on exposure for boys and girls (p-value = 0.520). This finding is weaker than earlier ones in the literature. Cameron et al. (2001) showed that in Indonesia, as a response to crop failure, households with girls reduced educational expenditures more than households with boys.

| Further robustness tests, the role of migration and selective childbirth and mortality
In Appendix D we provide further robustness tests implementing an alternative indicator for crisis exposure. One of the main challenges when assessing the impact of an economic crisis is precisely determining its start and end date. In the main specification we used the same periods as established in previous works (Hidrobo, 2014;Parandekar et al., 2002). However, according to the macroeconomic data, the contraction of GDP started one quarter later, that is, not in July 1998 but in October 1998. Moving the start of the crisis one quarter later and including only those periods as crisis when there was an observed  Table 2. The coefficient estimates associated with the child and household level characteristics are available upon request. Wild-bootstrapped standard errors clustered at the regional level are presented in parentheses. + p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001.

T A B L E 4
Heterogeneous effects of crisis exposure on height-for-age Z-score by gender.

T A B L E 5
The effect of in-utero exposure to the crisis on height-for-age Z-score.
contraction in GDP, we find that the identified effects on HAZ and years of schooling are similar to the results obtained with the main specification (Table D1, Columns 1 and 2). Yet, the effect on years of schooling is halved suggesting that it might partially pre-date the economic crisis (Column 2). We also used a second alternative crisis indicator that includes the period of the rainfalls caused by the "El Niño" phenomenon that lasted from February 1997 to August 1998. Thus, we account for the combined effect of "El Niño" and the economic crisis. We observe that the effect on HAZ is slightly smaller in absolute magnitude (−0.002, Table D1, Column 3) and less precisely estimated (p-value ≤ 0.1). The effect on years of schooling is larger in absolute magnitude and very precisely identified (−0.003, p-value ≤ 0.001, Table D1, Column 4). These results imply that the impact of the economic crisis should not fully be separated from the "El Niño" rainfalls.
We also performed a placebo analysis where we defined a placebo crisis in 1996, 1997, 2002, 2003 and 2004. We exclude observations that were exposed to the actual crisis from the placebo specifications and assume the placebo crisis began in the same calendar quarter and lasted the same length as the actual crisis. We do not find any statistically significant impact of the placebo crisis on either of the human capital measures. In fact, three of the five coefficients are positive and all are statistically insignificant (Table D2). The main effect of the actual 1999 crisis is an outlier relative to the distribution of the placebo effects.
Another concern involves policies that the government might have implemented to cope with the crisis. Not accounting for such policies could bias our results downward. A possibly confounding factor is the social assistance program Bono de Desarrollo Humano that was rolled out in response to the crisis. Yet, we show that controlling for the program does not change our main results (Table D3). The results of these estimations are consistent with the findings of previous studies that have evaluated the cash transfer program directly. While in the short-term no impact of Bono de Desarrollo Humano could be established on HAZ (Fernald & Hidrobo, 2011;León & Younger, 2007), a short-term impact on school enrollment could be found because people thought it was a condition to get the transfer. This positive education effect disappeared in the long-run after people learned that there is no enforcement mechanism to control for the compliance with this requirement (Oosterbeek et al., 2008). Similarly, previous studies have shown that the cash transfer program increased education related outcomes for the poorest quintile, that is, school enrollment, memory and number of words a child can say and combine, but such effects could not be found for higher income quintiles or urban areas (Fernald & Hidrobo, 2011;Paxson & Schady, 2010;Schady et al., 2008). Previous studies already question whether the limited effects of Bono de Desarrollo Humano would persist over time. Although the exact coverage of this program targeting stunted children is not available, Parandekar et al. (2002) indicate that the program only reached 15,000 children while the number of children at risk of malnutrition was as high as 209,000. Thus, as previous studies showed, the program implemented in Ecuador after the crisis was not strong enough to compensate the early childhood loss. Besides, after the crisis Ecuador adopted strictly neoliberal policies that restricted public spending. Public health expenditure represented 2% of GDP in 1997 but did not reach this share again until 2007. Moreover, we are not aware of any other, large-scale social policy that might have been put in place for coping.  Table 2. The coefficient estimates associated with the child and household level characteristics can be found in Appendix C, Table C5. Wild-bootstrapped standard errors clustered at the regional level are presented in parentheses. + p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001.

T A B L E 6
The effect of crisis exposure on years of schooling. Another concern relates to possible biases introduced by migration. We present a detailed discussion of the role of migration in Appendix E. In short, we can infer that the magnitude of any possible bias (if it exists) is not important. The Census records a total of 6000 migrant children who were born during the crisis. This number only represents 1.1% of the total number of children born during the crisis. 14 At the very least our results shed light on the long-term effect of the 1999 economic crisis for those who stayed in Ecuador.
Lastly, we are concerned about selection into childbearing and selective infant mortality. A detailed discussion is presented in Appendix F. Over the 10-year period, January 1993 to December 2003, we observe an increasing trend in child births with no reduction during the crisis period ( Figure F1). Moreover, comparing the contemporaneous characteristics of those mothers giving birth during the crisis and those before and after does not show any statistically significant differences (Table F1). Similarly, including the number of children born at the same year, that is, the cohort size, as a control variable in the main specification, we do not identify any relationship between that control variable and the Z-score (Table F2). More importantly, the main crisis exposure effect of −0.003 remains.
Concerning infant mortality, we discuss in detail in Appendix F that Vos et al. (2006) do not find any change in infant mortality trends in Ecuador in 1999 compared to the five previous years. Moreover, we do not find any statistically significant impact of the crisis on the total number of children (Appendix C, Table C4).

| CONCLUSION
This paper contributes new evidence on the long-term effects of crizes. We show that the economic crisis in Ecuador in 1999 had a long aftermath because the loss in human capital for those who were born during the recession appears to be permanent. If shocks and crizes experienced in early childhood are not counteracted, they might have long-term negative repercussions for exposed individuals despite macroeconomic recovery.
We assessed the impact of the crisis using two population-representative datasets, that is, our estimates reflect the average effect estimated from nationally representative data 12-16 years after the original shock. The results indicate that being exposed to the entire period of the crisis led to a loss of 0.063 SDs in height-for-age during early adulthood, representing a loss of 4.9-5.9 mm for a child exposed for the full duration of the crisis. This effect is larger than the one obtained by Banerjee et al. (2010) who calculated a loss of 1.8 mm for the cohorts exposed to phylloxera in France.
Furthermore, we assess the impact of the crisis on the "next stage" of human capital formation, namely education. Being exposed to the crisis for the entire period diminished the years of schooling by roughly 0.042 years.
For long-term health impacts of the crisis, we find gender bias. The impact on the HAZ of girls is five times bigger compared to boys. Lower female height is linked to higher probabilities of giving birth to children with lower birth weight, which increases the risk of intergenerational malnutrition (Victora et al., 2008).
Our analysis is not without limitations. First, we would have liked to analyze a panel dataset that followed the cohorts born during the crisis over time. However, there is no such data available. Second, we do not have information about the child, maternal and household characteristics at the time of childbirth. We rely on information collected years after the crisis. Third, as the crisis affected the whole country, we cannot exploit geographic variation for identification purposes. Fourth, we do not have information about health centers and can only account for access to healthcare indirectly via geographic-specific cohort trends and mother fixed effects. Similarly, regions/provinces/cities might have reacted differently to the crisis depending on their local capacity. However, health policies and budgets are concentrated in the hands of the central government.
Our micro-level results imply that despite the efforts of the Ecuadorian government to counteract the crisis with the Bono de Desarrollo Humano program, long-term growth retardation and loss in education among those children being born during the crisis could not be fully counteracted. One reason for the persistent negative impacts on human capital might be the limited coverage of Bono de Desarrollo Humano (Parandekar et al., 2002) or the lack of effectiveness of the cash transfer component of Bono de Desarrollo Humano with respect to health-related outcomes (Fernald & Hidrobo, 2011;Ponce & Bedi, 2010) and the modest effects for education-related outcomes (Paxson & Schady, 2010).
Future research should assess the persistence of the negative effects of the crisis further, follow the exposed cohorts and measure their performance in the labor market.
The findings from this research might inform sustained policy responses to COVID-19 to minimize lasting effects. The COVID-19 pandemic started as a health crisis and rapidly turned into an economic and social crisis, in particular in developing countries. According to the latest State of Food Security and Nutrition Report global hunger numbers rose to between 702 and 828 million in 2021, making it 46 million people more compared to 2020 when taking the middle of the projected range as base (FAO et al., 2022). The open question is whether the negative, (indirect) repercussions that the crisis has on children will persist over time. This paper suggests that the possibility of long-term poverty traps should not be underestimated.