The impact of experiencing an earthquake in the first 1000 days of life on child growth

(1)

The impact of experiencing an earthquake

in the first 1000 days of life on child growth

Abstract

This thesis analyzes the effect of experiencing an earthquake in the first 1000 days of life on the growth of young children. A difference-in-differences strategy with multi-ple time periods and treatments is used to investigate the effect of five earthquakes that occurred in Indonesia between 2000 and 2014. The main results show that there is no significant effect of experiencing an earthquake on height and stunting. When inspecting the impact of the different earthquakes separately, all effects are insignif-icant, except for one. Surprisingly, the earthquake on West Java appears to have a positive effect on height. This could be the result of an effective aid response that outweighed the negative consequences of the earthquake.

Lotte Westerbeek

University of Amsterdam

Student number: 10343032

Supervisor: Hessel Oosterbeek

(2)

Statement of Originality

This document is written by Student Lotte Westerbeek who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document are original and that no sources other than those mentioned

in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for

(3)

|

1. Introduction

This thesis evaluates the impact of experiencing an earthquake in the first 1000 days of life on the growth of young children. Environmental factors are extremely important for the development of a child (Black et al., 2013). If infants suffer from negative shocks, resulting in nutritional deprivation or frequent occurrences of infectious diseases (e.g., diarrhea), they may fail to reach their full growth potential. When a child’s growth falters and height-for-age drops at least two standard deviations below the World Health Organization standard, (s)he is considered ‘stunted’ (Dewey & Begum, 2011). Stunting can thus be viewed as an extreme form of growth faltering.

Extreme growth faltering is associated with various adverse long-run effects. As Grantham-McGregor et al. (2007) show, stunting is a key marker of poor development at various levels. Children who suffer from malnutrition in utero and infancy are at greater risk of delays in development. They lag behind in the motor system, language, cognition, personal and social behaviors (Alderman, Hoddinnot & Kinsey, 2006; Dewey & Begum, 2011; Kar et al., 2008; Nurani, Sitaresmi & Ismail, 2011). This has large negative effects on school achievement and consequential productivity and wages. Furthermore, according to Dewey and Begum (2011), the children of stunted women experience negative effects from their mother’s growth faltering. They have lower chances of survival and are at higher risk of experiencing health problems. This implies that growth faltering has intergenerational consequences. Moreover, high levels of stunting do not solely affect the specific stunted in-dividual, but the country as a whole. Due to the negative adverse consequences of stunting and growth faltering on a child’s development, a high prevalence in a country will prevent the economy from reaching its productivity potential, hindering economic growth.

Up until now, various studies investigated the effects of negative shocks early in life on growth faltering and stunting. Many of these papers focus on the effects of experiencing a drought. Results indicate that this type of shock has a significant negative impact on growth. Children who live in areas hit by a drought end up being shorter than their peers. Estimates range from a child losing 0.9 to 2 centimeters in height (Godoy et al., 2008; Hoddinnot & Kinsey, 2001; Yamano, Alderman & Christiaensen, 2005). Other scholars focus on the impact of earthquakes. Nurani et al. (2011) conduct a descriptive study where they perform a risk analysis in order to identify the main risk factors for stunting after the 2006 Yogyakarta earthquake. They find a correlation between unmonitored growth in the previous three months and malnutrition, and between acute respiratory infection and stunting. Additionally, Rydberg, Marrone, Strömdahl and von Schreeb (2015) perform a multilevel regression analysis where they compare different birth cohorts and different shaking intensity zones after the 2001 Peru earthquake. Their findings show that chances of being stunted increase after the earthquake in the high-intensity zone, while they decrease in the medium- and low-intensity zones. However, there is still limited knowledge on the effects of an earthquake on childhood development, with only a few studies focusing on this

(5)

relationship (Rydberg et al., 2015). This thesis adds to the literature on early childhood development by focusing on multiple earthquakes within one country.

Child growth can be affected by an earthquake in different ways. On the one hand, the earthquake and its aftermath may negatively influence the growth of a child by caus-ing food shortages (Brancati, 2007). Secondly, earthquakes may lead to the outbreak of infectious diseases, affecting children’s health and growth (Brancati, 2007). Thirdly, an earthquake may lead to the loss of livelihood for people, indirectly affecting their children (Kun, Han & Yao, 2009). On the other hand however, an earthquake often sparks in-fluxes of aid into the affected region. This aid has the potential of countering the negative consequences that a child experiences from the earthquake (Raschky & Schwindt, 2009).

It is important to understand the net effects of an earthquake. According to McGuire (2013) global warming will lead to an increase in the number of earthquake occurrences. The majority of these earthquakes are expected to happen in developing countries, since most of these countries are located in areas especially prone to natural disasters (Alcantara-Ayala, 2002). In the 20th century, 42 percent of the registered natural disasters occurred in Asia. This geographical concentration makes it crucial to understand the net effect of an earthquake on growth faltering, since an increase potentially aggravates the economic inequality between countries. If an earthquake indeed has a negative impact on child growth, this might be an indication that aid flows are not sufficient to combat the ad-verse consequences of an earthquake. Given the intergenerational consequences of growth faltering this could have broad implications for economic growth.

This thesis focuses on the case of Indonesia; one of the countries particularly prone to seismic activity. Data of the Indonesian Family Life Survey (IFLS) is exploited to assess the impact of five earthquakes that occurred between 2000 and 2014 in different provinces of Indonesia. A difference-in-differences model with fixed effects for the provinces, periods and age cohorts will be used to examine the effect of these earthquakes on both height and stunting of children that were in their first 1000 days at the time of the earthquake. The results found in this study are inconclusive. The mean effect of the earthquakes combined is insignificant regardless of the specification. Yet, an investigation of the separate impact of the five earthquakes reveals that effects are heterogeneous. All estimates indicate an insignificant, negative impact of an earthquake on stunting. However, the earthquake in West Java shows a significant and positive effect on height, whereas the other four earth-quakes show an insignificant, negative effect. Further research into the various mechanisms at work here needs to be conducted before any definite conclusions can be drawn from these results.

The rest of this thesis will proceed as follows. In section 2, both the causes and consequences of growth faltering will be discussed, as well as the channels through which an earthquake may influence growth. Next, in section 3 an overview of the data will be provided. Section 4 contains an explanation of the model and discusses the main assumptions of the model. In section 5 the results will be presented, followed by the conclusion of this thesis.

(6)

|

2. Growth faltering

and earthquakes

The focus of this thesis is the impact of earthquakes on growth faltering. Growth faltering is mainly caused by malnutrition and repeated suffering from infectious diseases. The associated consequences of growth faltering are delays in brain development and higher risk of heart-related health problems. This, in turn, results in lower cognitive skills, poor school performance, and lower productivity and adult wages. An earthquake is expected to affect growth faltering negatively through shortages in food, health threats and loss of livelihood of parents. Aid is expected to offset these negative effects to a certain degree.

2.1 Causes of growth faltering

Unquestionably, genetic inheritance has an important influence on attained height. Yet, as argued by Grantham-McGregor et al. (2007), average growth potential does not differ significantly across countries. In terms of genetics, Indonesian children from poor families have the same growth potential as European or American children with wealthy parents. In reality however, the odds of reaching this potential are not the same across or within countries due to differing environmental factors, such as supply of food and quality of sanitation.

Both in utero and as an infant, absorbing the right nutrients is crucial for healthy growth. When still in utero, a child’s growth is affected by the nutritional status of the mother (Walker et al., 2007). If food is scarce during pregnancy, or the nutritional value of the food is low, this can affect the growth of a child even before (s)he is born. After birth, this issue resumes: if sufficient food supplies are lacking during early childhood, as is often the case in developing countries, the child cannot reach his/her growth potential. Insufficient food supply is not the only issue that children in developing countries face. Sanitation, hygiene and access to healthcare are often not optimal either, leading to relatively high rates of infectious diseases (Prendergast & Humphrey, 2014). Infections prevent proper intestinal absorption, which decreases the amount of nutrients available for the growth process. Besides, the already scarce nutrients are diverted away from growth purposes towards the immune response (Stewart, Ianotti, Dewey, Michaelsen, & Onyango, 2013). Together this exacerbates growth faltering and stunting.

Moreover, a downward spiral between undernutrition and infectious diseases exists: a child that suffers from malnutrition is more prone to get infected by a virus or bacteria. The infection will in turn influence the child’s appetite and intestinal absorption, worsening the nutritional status even further.

(7)

2.2 Consequences of growth faltering

In order to distinguish growth faltering from genetically determined short stature, the literature on early childhood development focuses on stunting; an indicator of substantial growth impairment. A child is considered stunted when height-for-age falls below 2 stan-dard deviations of the threshold of the World Health Organization (Stewart et al., 2013). The first 1000 days after conception are a period where stunted growth starts and where height status later in life is largely determined, making it an important period for growth (Prendergast & Humphrey, 2014).

This period corresponds to a crucial period for brain development, affecting cognitive ability and school performance. Many scholars found negative effects of poor nutrition and infectious diseases on the development of the brain (Black et al., 2013). Brain development occurs in different phases, all building on each other (Grantham-McGregor et al., 2007). Disturbances in the early growth process can thus have a long-run impact on the structure and functional capacities of the child’s brain. Berkman, Lescano, Gilman, Lopez and Black (2002) found that children in Peru that were severely stunted in the second year of life scored significantly lower on a cognitive functionality test at age 9. Mendez and Adir (1999) found similar results for children in the Philippines. Chang, Walker, Grantham-McGregor and Powell (2002) did not merely find an effect of underdevelopment on cognitive skills, but on school attainment and achievement as well. This implies that stunted children not only go to school less frequently than their non-stunted peers, but they also learn less while there. Moreover, Gardner, Grantham-McGregor, Himes and Chang (1999) found that stunted children in Jamaica showed higher levels of apathy and lower levels of enthusiasm in exploring while playing. In addition, their attention span was significantly shorter. This potentially interacts with the lower cognitive skills, leading to stunted children learning even less in the classroom and being less productive later in life.

Another issue associated with stunted growth is health later in life. The process that potentially causes health problems for stunted children is often called the ‘metabolic syndrome’ (Prendergast & Humphrey, 2014). Malnourishment during fetal or infant life triggers a reaction from the body to protect vital organs by permanently changing the metabolism system. Nutrients are directed away from growth towards the preservation of organ functionality. This process provides survival benefits during the period of undernu-trition. However, it simultaneously increases the risk of hypertension, heart-related dis-eases and type 2 diabetes in a period of sufficient nourishment (Prendergast & Humphrey, 2014). These effects are exacerbated if weight-gain is very rapid later in life or in case of obesity. This means that children that are stunted early in life are at higher risk of suffering from the abovementioned diseases, even, and potentially even more so, when nourished more than sufficiently later in life.

Stunting additionally generates an intergenerational cycle of poverty and reduced hu-man capital, as stated by Prendergast and Humphrey (2014). Women that are themselves stunted have an increased risk of complications during pregnancy and giving birth to

(8)

underweight infants (Prendergast & Humphrey, 2014). Moreover, lower educational at-tainment is associated with higher fertility and poorer knowledge on nutritional practices, exacerbating the problem of growth faltering even further. This attributes to the inter-generational transmission of poverty (Grantham-McGregor et al., 2007).

These issues underline the far-reaching consequences of growth faltering. These conse-quences do not halt at the individual or household level. It affects the economic situation of a nation as a whole. As mentioned, stunted children have lower cognitive skills, edu-cational attainment and grades. The poorer learning abilities combined with lower levels of schooling leads to lower productivity levels. As a result, wages are lower among these individuals. Additionally, the ‘metabolic syndrome’ leads to higher rates of diabetes heart-related diseases. Again, this lowers productivity, assuming that healthy workers are more productive than sick ones. Besides, government expenditures increase in order to cover the medical costs, playing into the expansion of budget deficits.

2.3 Channels

To understand why an earthquake might affect a child’s growth process, one must recognize potential channels through which this happens. In this subsection four possible channels are identified; shortages of food and water, health threats and loss of livelihood all influence growth in a negative way, whereas the aid that flows into the affected areas has the potential of countering these adverse effects.

2.3.1 Shortages

One of the difficulties that an earthquake causes is the scarcity of necessities, particularly in developing countries (Brancati, 2007). The vibrations of an earthquake often trigger various landslides. These landslides subsequently destroy crops and kill cattle. This cuts off an important source of food, fertilizer and draught power (Brancati, 2007). In addition, inventory of both farmers and shop owners may fall prey to the destructive effects of landslides and fires following an earthquake, reducing the supplies even further. Moreover, damage to water pipes and sewage systems may prevent the population from access to clean drinking water (Brancati, 2007).

Oftentimes, emergency response teams cannot readily bring food and water supplies into the affected regions to combat the most urgent nutritional needs. First of all, the emergency response needs to be set up, which takes time. Secondly, earthquakes often damage roads, making it hard to access the more remote areas (UN Resident and Human-itarian Coordinator for Indonesia, 2009).

All these problems affect the availability of essential nutrition. As explained in sub-section 2.1, nutritional deprivation is one of the main drivers of stunted growth, which is why it is expected that these shortages are an important channel through which an earth-quake affects growth. However, the magnitude of the effect depends on the duration of the scarcity situation. It takes time to grow crops anew and to fix broken water pipes. There

(9)

might not be enough funds to import enough food from other regions. Also, more remote areas will most probably be hit harder due to the aggravated problem of accessibility.

2.3.2 Health threats

Another channel through which an earthquake influences child growth is the enhanced spread of infectious diseases. As mentioned before, earthquakes often damage the sewage and water system. Pipes break and manholes are damaged, leading to the contamination of drinking water and the water in sanitation systems (Brancati, 2007). Bacteria, parasites and viruses from the sewage end up in the clean water, considerably increasing the risk of incurring communicable diseases. The spread of these diseases will be exacerbated by overcrowding in public places. Operational hospitals will be flooded by injured people and public spaces that avoided heavy damages will be sought by people forced to leave their homes.

As explained in subsection 2.1, frequent instances of infectious diseases at a young age hamper healthy growth, which is why it is expected that this channel is important in explaining the effect of an earthquake on growth.

2.3.3 Loss of livelihood

As a result of a serious earthquake, many people lose their job or business (Kun et al., 2009). As explained before, the agricultural sector often experiences crop damage and loss of inventory and cattle. Consequently, farmers incur financial losses that cannot readily be restored. Generally, it becomes difficult to obtain seeds and the consequent scarcity pushes seed prices up (Government of Indonesia, 2009). Also, growing back the crops takes time. In addition, great damage to other industries can be expected. Oftentimes the tourism industry suffers great losses. Facilities are damaged or destroyed, and tourist arrivals and reservations plummet (Government of Indonesia, 2009). Furthermore, small-and medium-sized businesses are usually affected. Fires, lsmall-andslides small-and power losses all damage their commercial activities.

The consequences for economic livelihood are predominantly secondary and long-term effects that may lead to nutritional deprivation of children in the household for an extended period. These are mostly related to the loss of the main source of income for the family. Parents in a developing country that lose their livelihood possibly have trouble providing adequate nutrition for their children. Moreover, if many casualties arise as a result of the earthquake, a number of children becomes orphaned. These children potentially suffer from malnutrition too in the absence of a caring parent, leading to increased growth faltering and stunting.

2.3.4 Aid flows

In the last three subsections, the negative effects of an earthquake on child growth have been discussed. Yet, there will usually be a counterforce as well, consisting of aid flowing

(10)

into the affected regions. Depending on the severity of the earthquake and the publicity it receives, this aid flow may be of a considerable size (Raschky & Schwindt, 2009). This potentially offsets part of the negative impact of an earthquake. The aid usually takes the form of manpower, food and funds.

Generally, part of the aid flow is dedicated to providing food and water for the popu-lation. If the aid teams manage to get the supplies into the region, this can help combat the negative impact of the shortages discussed in section 2.3.1. Another part of the aid is usually dedicated to rebuilding damaged buildings, water and sewage systems, as well as infrastructure (Raschky & Schwindt, 2009). This may temper the negative effects of contaminated water and loss of livelihood discussed in sections 2.3.2 and 2.3.3. The mag-nitude of the counter-effect that aid generates depends on the size and duration of the received aid.

(11)

|

3. Data

3.1 Growth faltering and earthquakes in Indonesia

In 2012, Indonesia was ranked fifth in number of children under 5 years old that were moderately or severely stunted (UNICEF, 2013). However, absolute numbers do not tell the complete story, since Indonesia is a large country. Having said that, the World Health Organization ranked Indonesia 25th in terms of percentages in that same year: still a very alarming position (Index Mundi, 2013). In 2011, the Indonesian government joined the Scaling Up Nutrition Movement, committing itself to a world free from malnutrition by 2030 (www.scalingupnutrition.org/indonesia). Despite this commitment to fighting malnutrition, the World Bank still estimates the economic losses due to growth faltering to be 2-3% of Indonesia’s GDP (World Bank Group, 2015).

As stated before, Indonesia lays in an area particularly prone to seismic activity. This thesis investigates whether the resulting earthquakes exacerbate the challenge of combat-ting growth faltering and stuncombat-ting in the country. The earthquakes that will be studied all occurred in the period between 2000 and 2014, and hit provinces whose residents were recorded in the RAND Indonesian Family Life Survey (IFLS). The most severe earthquakes of this period were selected for the study. Severity is based on the magnitude of the seismic activity on the Richter’s scale, total damage recorded, or both, depending on the available information. This information is drawn from the ‘Significant Earthquake Database’ by the National Geophysical Data Center (NGDC). In this database the magnitude on the Richter’s scale is reported. In addition, the NGDC scaled estimated damages from 1 to 4, ranging from no damage to extreme damage. The earthquakes that are selected for this research are reported to have caused either severe or extreme damage in the NGDC database. Severe damage, indicated by the number 3, corresponds to an amount of 5 to 24 million US dollars. Extreme damage, indicated by the number 4, corresponds to an amount of 25 million US dollar or more. Table 1 provides an overview of the magnitude and the level of damage for the five earthquakes selected for this study. Unfortunately, there is no information available on the exact magnitude of the aid response. Further-more, note that the 2004 Indian Ocean earthquake is excluded from the analysis despite its severity. The reason for omitting this natural disaster is the fact that practically all damage was done in the province of Aceh; a province not sampled by the RAND.

3.2 Data source and sample

The data that is used for the analysis in this thesis is drawn from the RAND Indonesian Family Life Survey. The IFLS is an ongoing survey that started in 1993 and collects infor-mation on fertility, education, migration, employment and anthropometrics. The waves used in this thesis are 3, 4 and 5, fielded in 2000, 2007/2008 and 2014/2015 respectively. The main outcome variables of this study are height and stunting. Data on height is

(12)

Table 1: Information on selected earthquakes Year Magnitude Level of damage South Sumatra 2000 7.9 3

Bali 2004 5.8 3

Yogyakarta 2006 6.3 4

West Java 2006 7.7 Unreported* West Sumatra 2009 7.5 4

*Despite the fact that the level of damage is not reported in the Significant Earthquake Base, this earthquake is still included in the analysis. As argued by Reese et al., the earthquake caused severe damage to West Java (2007).

readily available in the data set and no modifications are needed. In order to investigate stunting, however, a variable is needed that captures the number of standard deviations that a child is above or below the World Health Organization’s reference for height-for-age. This variable was obtained by using the ‘zanthro’ package in STATA. This function is created by Vidmar, Cole and Pan (2013) and it transforms data on child anthropomet-rics to z-scores by using the LMS method and the reference data of the World Health Organization. By common practice, a child is considered stunted when it is 2 standard deviations or more below the reference height.

In order to maximize both sample size and representativeness of the population, all provinces recorded in the IFLS are included in the analysis. The survey covers the majority of provinces in Indonesia. Nevertheless, certain areas were excluded by the RAND in order to curb survey costs. Appendix A contains a list of all the provinces that were ultimately included in this research. Five of these were hit by an earthquake in the examined period and are assigned to the treatment area: South Sumatra, Bali, Yogyakarta, West Java and West Sumatra. The remaining 19 provinces serve as counterfactuals. The five smallest provinces are pooled together in order to ensure an acceptable number of observations per group.

Furthermore, the sample is restricted to children aged 0.72-9.88 at the time of mea-surement; either 2000, 2007/2008 or 2014/2015. This age range is obtained by selecting the children that lived in one of the five provinces and that were in utero or an infant at the time that an earthquake hit that particular province. These children will from now on be named ‘treatment children’. As explained in section 2.2, the first 1000 days after conception are a crucial window for development. Since the exact moment of conception is unobserved however, the first 1000 days of life are approximated by selecting the children that were between -9 months and +2 years old at the time of the earthquake. Table 2 provides an overview of these children, per province. The table reports the year in which the earthquake took place in the listed province and the survey wave in which measure-ment of the children that experienced the earthquake in their first 1000 days of life took place. The last row reports the corresponding age range of these in the listed province. As can be seen from the table, age is converted to a continuous variable. This ensures a more precise estimation of the treatment effect.

(13)

Table 2: Treatment children

South Sum Bali Yogy West Java West Sum

Year of earthquake 2000 2004 2006 2006 2009

Year of survey wave 2007/2008 2007/2008 2007/2008 2007/2008 2014/2015 Age 6.67-9.88 3.45-6.33 0.96-3.69 0.72-3.77 4.29-7.43

The control group consists of all children that comply with the following requirements: 1. The child is not part of the group of ‘treatment children’ as listed in table 2 2. The child did not experience any earthquake before the point of measurement 3. The child is within the age range listed in table 2 at any point of measurement For example; a child that lives in South Sumatra, was measured in 2014, and is between 6.67 and 9.88 years old at the time, is assigned to the control group.

All children that experienced an earthquake but were not in their first 1000 days of life at that time, are excluded. Namely, even though slightly older at the time of the disaster, these children are exposed to the negative effects of an earthquake as well. Leaving them in the sample would lead to a negative bias. Additionally, in order to account for the fact that growth processes can vary among age groups, age cohorts were constructed. These age cohorts correspond to the age range of the ‘treatment children’ at the time of the post-earthquake height measurement as listed in table 2.

A final restriction is placed upon children for whom data on height is missing. There is no particular reason to suspect that the lack of measurement of height is system-atic. Nonetheless, a test was performed to investigate selectivity. Using a difference-in-differences framework, the effect of an earthquake on non-measurement was estimated. No significant effect was found, which indicates that non-measurement was indeed random (see: Appendix B). After all abovementioned exclusions, 15,819 observations remain in the control group. There are 1,967 observations in the treatment area pre-earthquake and 844 observations in the treatment area post-earthquake. Hence, this means there are 844 ‘treatment children’ in the sample.

3.3 Descriptive statistics

Table 3 presents an overview of pre-earthquake observables. The mean values of both the treatment and control area are reported, as well as the differences between the two. The first three columns show these statistics for the first period. Note that ‘treated 1’ (i.e. South Sumatra, Bali, Yogyakarta and West Java) represents the group of children from the four provinces that were hit by an earthquake in the period 2000-2007. The last three columns show the same statistics for the second period. ‘Treated 2’ refers to West Sumatra; the province hit by an earthquake in the period 2007-2014.

As shown in the table, there are various significant differences between the treatment and control areas at baseline. For instance, the parental characteristics in 2000 differ

(14)

between ‘treated 1’ and the control group. Namely, both mother and father of the children in the treatment areas are younger and of larger stature than their counterparts in the control areas. Also, children from the group ‘treated 1’ are significantly younger and shorter than the children from the control areas.

In comparison, the differences regarding parental characteristics are smaller when com-paring ‘treated 2’ and the control areas in 2007, as can be seen from the columns 4-6. The only significant difference is in the average level of maternal education, where the share of mothers staying in school beyond middle school education is about 17 percentage points higher in the treatment area. In contrast, all child characteristics differ between the group ‘treated 2’ and the control areas. Children in the treatment area are approximately 6.3 centimeters taller than the ones in the control areas, they have a lower chance of being stunted, and they are older. Also, children from the treatment area live in slightly larger households.

The importance of using a difference-in-differences strategy with state and age cohort fixed effects is underlined by the abovementioned differences. As can derived from this section, whether a child experienced an earthquake or not depends on the province they live in, and the people in these provinces differ considerably. Both observed and unobserved differences will be accounted for in the difference-in-differences model as long as they are time-invariant. Note, however, that a further investigation of the suitability of the control areas will be conducted. The discussion on this can be found in section 4.2.

(15)

Table 3: Pre-earthquake descriptive statistics

2000 2007

Treated 1 Control Difference Treated 2 Control Difference Child characteristics height 91.3 101.5 -10.21 108.5 102.2 6.28 % stunted 0.35 0.38 -0.03 0.17 0.33 -0.16 age 3.5 5.2 -1.72 5.8 5.2 0.58 household size 5.4 5.4 0.03 5.4 4.9 0.57 Number of observations 736 4159 134 5344 Mother’s characteristics height 151.4 150.2 1.20 151.1 150.7 0.38 age 30.1 31.8 -1.75 29.5 29.3 0.25

% >middle school education 0.38 0.34 0.04 0.58 0.41 0.17

Number of observations 698 3846 132 5056

Father’s characteristics

height 162.5 161.1 1.35 160.7 161.5 0.88

age 34.5 37.0 -2.49 34.5 34.0 -0.56

% >middle school education 0.39 0.37 0.01 0.47 0.42 -0.05

Number of observations 606 3356 108 4411

Note: ’Treated 1’ is a group consisting of all children living in a province in which an earthquake occurred in the period 2000-2007; either South Sumatra, Bali, Yogyakarta or West Java. ’Treated 2’ consists of children living in the province in which an earthquake occurred in the period 2007-2014; West Sumatra. ’Control’ is a group consisting of all children from the control provinces. The first, second, fourth and fifth column report mean values. The third and sixth column report mean differences. The underlined differences are statistically significant at the 5% level.

(16)

|

4. Method

In this thesis, a difference-in-differences framework with multiple time periods and provinces will be used in order to identify the effect of experiencing an earthquake on child growth. The main concerns regarding bias of the results are a violation of the parallel trend as-sumption, potential spillover effects and a change in the composition of the treatment groups in the period after the earthquake.

4.1 The model

To study the impact of an earthquake on growth, a difference-in-differences analysis will be performed. This method allows for the estimation of a treatment effect with non-experimental, cross-sectional data (Angrist & Pischke, 2008). This is a useful feature for this particular study. First of all, the data on earthquake victims is (obviously) observa-tional. Secondly, the focus is on exposure of very young children to negative shocks and the time between survey waves is approximately seven years. This means that it is impos-sible to follow the same children at different time periods; they were simply not born yet in the survey wave before the earthquake. Hence, the data can be viewed as cross-sectional. The fact that the earthquake is an exogenous shock that only affects certain provinces in Indonesia, and not others, is exploited. All provinces that were affected by an earthquake in the period 2000-2014 are assigned to the treatment area, while the other provinces are assigned to the control area.

In a basic difference-in-differences model, the treatment effect is identified by compar-ing the changes in the outcome of interest before and after the treatment for the treatment relative to the control group (Stock and Watson, 2012). Essentially, the trend of the con-trol group serves as the counterfactual trend of the treatment group. In this way, potential pre-treatment, time-invariant differences between the control and treatment group are ac-counted for. The simple difference-in-differences estimator is given by:

β1 = (E[Ytreat,1] − E[Ytreat,0]) − (E[Ycontrol,1] − E[Ycontrol,0]) (4.1)

Where the subscript treat stands for the treatment group and control for the control group. The subscript 1 denotes the period after the treatment and 0 denotes the period before. Since in this thesis various earthquakes are evaluated over a longer period of time, the difference-in-differences estimator will capture the mean effect of the examined earthquakes combined. The corresponding regression equation to approximate this estimator is:

Yista= α + β1 T REATsta + 20 X k=2 βk P ROVks + 3 X j=2 βj P ERIODjt + 5 X l=1

βl AGEla + Xista + εista

(17)

Where subscript i denotes an individual child, subscript s denotes the province in which the child lives, subscript t indicates the period and subscript a denotes the age cohort the child is assigned to. PROV_sis a set of dummy variables that indicate the province in which the particular child lives.PERIODtdenotes a set of dummies for the different time periods.

The variable AGE_a is a set of dummies for various age cohorts. This variable is added to account for the fact that not all examined children are of the same age. Namely, since the earthquakes occur in different years, the age of the children that were infants at the time of the earthquake differs at the next date of measurement. The age cohorts correspond to the age range of the children that experienced an earthquake in a particular province when they were between -9 months and 2 years old. The age ranges of the different cohorts can be found in table 2. These age cohorts are selected in favor of smaller, more precise age cohorts in order to maintain a sufficient number of observations per cohort.TREATsta

denotes a dummy that is 1 for the first measurement after the earthquake for every child that lives in an affected province and was between -9 months and 2 years old at the time of the earthquake in that particular province. This means that β1 represents the average

treatment effect. Furthermore, Xista represents a set of background characteristics of a

child (age, gender, household size, height of parents, parents’ education). εista indicates

the individual-specific error term.

The outcome variables are denoted by Yist. The main outcome variables to measure the

impact of an earthquake on growth are height and stunting. Height is a straightforward, continuous variable of the height measurements conducted in the IFLS. The above-noted Ordinary Least Squares regression can readily be conducted. Stunting, in contrast, is a dummy variable equal to 1 if a child is stunted and 0 otherwise. When the dependent variable is a binary, the question arises which regression specification ought to be used; a linear probability, a probit or a logit specification.

On the one hand, many scholars argue that a linear probability specification cannot be used due to the violation of the assumption of normally distributed errors (Bartlett, 2015). When a linear regression is used to model binary outcomes, the probabilities pre-dicted by the model possibly lie outside the usual range of probabilities [0,1]. On the other hand, the estimated marginal effects in logit and probit models depend on the val-ues of the observation in other variables, making the estimates particularly difficult to interpret. Marginal effects in the linear probability model are constant however, leading to a preference for this regression specification in terms of clarity and avoidance of unnec-essary complexities. Moreover, since we are solely interested in marginal effects, and not predicted probabilities, we are not concerned with the violation of the probability range. Hence, the empirical strategy of this thesis will consist of linear probability specifications. Also, robust standard errors will be reported in order to allow for heteroscedasticity.

(18)

4.2 Potential biases

The internal validity of the difference-in-differences model explained above hinges on a number of assumptions. This section will evaluate the validity of a number of these assumptions. The first and most crucial one is the parallel trend assumption. Secondly, spillover effects ought to be absent and thirdly, the composition of treatment and control group should not change over time (Gertler, Martinez, Premand, Rawlings & Vermeersch, 2016). It is suspected that a potential violation of the third assumption may lead to an underestimation of the treatment effect.

The first assumption, the parallel trend, is difficult to investigate in a framework with various earthquakes and time periods. Since the earthquakes occur at different points in time, one cannot simply follow common practice by performing a visual inspection of the similarity of the trend of the treatment and control groups before the earthquakes. Having said that, a placebo test that mimics the visual inspection can in fact be carried out (Gertler et al., 2016). This test is a replicate of the main difference-in-differences regression, but with fake outcomes. Instead of regressing the treatment on height and stunting, the same regression is performed with outcomes unaffected by the treatment as the dependent variable. Would the regression output imply an effect of the treatment on these fake outcomes, the comparison group must be flawed (Gertler et al., 2016). In this thesis, the placebo test was performed on a list of outcomes that are unaffected by the earthquake. First, the test was performed on the variables age and sex, since these characteristics cannot change as a result of an earthquake. Also, the height of the parents of the child will not change after the earthquake, thus height mother and height father are also used as fake outcomes. For none of these variables a treatment effect was detected, meaning there is no reason to suspect a flaw in the choice of control group (see: Appendix C). Hence, no bias as a result of the violation of this assumption is expected. Note that this is at the same time a first indication that the treatment and control group do not change in terms of composition over time. Later in this section, this assumption will be investigated further.

The second assumption concerns spillovers. As stated, the absence of spillover effects is a necessary condition for an estimator to be unbiased (Gertler et al., 2016). However, in the case of an earthquake, a reallocation of government funds may lead to a downward bias on the results. Namely, the Indonesian government may choose to relocate funds from unaffected regions (i.e. the control provinces) to the affected areas in order to curb the negative impact of the earthquake. This leads to a reduction in government spending in the control areas, which may affect the growth process of children unaffected by an earthquake. However, there is no reason to suspect this issue to arise. Namely, to the best of my knowledge, there are no recordings of such reallocations. There is, however, no other way to check for spillover and thus it has to be assumed to be absent.

A third assumption of the model is that the composition of the treatment and control group does not change over time (Gertler et al., 2016). It cannot readily be inferred that

(19)

this assumption holds. First of all, since cross-sectional data is considered in this study, a change in the composition of the groups might arise as a result of failed randomization. However, as discussed above and shown in appendix C, there is no flaw in the choice of control group, ruling out this possibility. Secondly, an earthquake possibly affects the rate of successful pregnancies in the affected provinces. As argued by Liu, Liu and Tseng (2015), natural disasters increase the likelihood of fetal losses. In addition, their results are indicative of positive selection, meaning that the strongest fetuses have the highest chances of survival. This could lead to a so-called ‘survivor bias’, where the post-earthquake sample is on average healthier than the pre-post-earthquake sample. Since this bias puts downward pressure on the treatment effect, this would imply that any reported result is a lower bound estimate of the true impact. This bias can be tested for by performing a difference-in-differences analysis similar to the one described in section 4.1, but with the outcome variable being the number of stillbirths/miscarriages of an individual woman. The children’s age cohorts are omitted. No effect of the earthquakes on the number of stillbirths and miscarriages per woman was detected. The survivor bias is therefore assumed to be absent (see: Appendix D).

(20)

|

5. Results

This section presents the estimates of the effect of experiencing an earthquake early in life on a child’s height and stuntedness. The section is divided into three different subsections: the main results, heterogeneous effects of the different earthquakes and robustness. The main results suggest that there is no significant effect of an earthquake on either height or stuntedness. The estimates vary in magnitude and sign for the five earthquakes separately. After increasing our sample size by relaxing the stringent age restriction, the findings do not change considerably.

5.1 Main results

The main results of this thesis come from the model as described in section 4.1; a difference-in-differences model with province-, period- and age cohort fixed effects. The focus of this subsection is the effect of experiencing an earthquake in the first 1000 days of life on height and stunting of children from several provinces in Indonesia. As mentioned before, ‘height’ is a continuous variable, whereas ‘stunting’ is binary. Effects of the treatment on both of these variables are estimated by the use of an OLS regression. In columns 1 and 2 of table 4 the basic OLS regression estimates are reported. Column 3 and 4 show estimates for the same specifications, but controlling for background characteristics. These characteristics include age, gender, the height of both parents and the household size. The coefficient on ‘Earthquake’ indicates the mean treatment effect of the five examined earthquakes combined.

Table 4: Main results

Without controls With controls

(1) (2) (3) (4)

Height Stunting Height Stunting Earthquake 0.2460 0.0268 0.0183 0.0243

(0.3617) (0.0192) (0.2271) (0.0188) Province fixed effects YES YES YES YES Period fixed effects YES YES YES YES Age cohort fixed effects YES YES YES YES Observations 18629 18629 18629 18629

R2 0.68 0.04 0.91 0.08

Note: The regressions of column 3 and 4 include controls for age, gender, height of parents and household size. When possible, missing values on parents’ height are replaced by values recorded in the previous wave and otherwise replaced by the average of the complete sample. A dummy variable indicating the missing value is added to all specifications that include controls. Robust standard errors are reported in parentheses. *, ** and *** denote significance at the 10%, 5% and 1% level respectively.

(21)

Against expectations, the results from table 4 indicate that there is no causal relationship between experiencing an earthquake and child growth. As reported in column 1, the effect of an earthquake on height is estimated to be 0.25 centimeters, yet insignificant. When including controls, as in column 3, the estimated coefficient drops to 0.02 centimeters, and remains insignificant. As reported in column 2, the chance of being stunted increases by 2.68 percentage points for children that experienced an earthquake. Again, this result is highly insignificant. This effect decreases to 2.43 percentage points after the inclusion of control variables in column 4 and remains insignificant.

The straightforward explanation for the insignificance of the treatment effect is the actual absence of an impact of experiencing an earthquake on growth. As discussed in section 2.3, there are several channels through which child growth may be affected by an earthquake. On the one hand, scarcity, health threats and the loss of livelihood may limit the growth of a child. On the other hand, large influxes of aid have the potential of affecting growth positively. These opposing effects might offset each other in the particular setting analyzed here.

Another explanation for these results is low statistical power. Both the expected effect and the sample size of the ‘treatment children’ is fairly small, making it hard to detect an effect. However, the effect sizes of 0.9 to 2 centimeters that are found by scholars studying the impact of a drought on growth are ruled out by the regression output of this study. These effect sizes fall outside the confidence interval. Perhaps the effect of experiencing an earthquake is indeed significantly smaller than the effect of experiencing a drought. In case the size of the standard errors is the reason why no effect has been detected however, it is noteworthy that the estimated effects on height and stunting are working in opposite directions. The coefficient on treatment when considering height is positive, implying a positive effect of experiencing an earthquake on the growth process. Yet, when the independent variable is ‘stunting’, the coefficient on the treatment is also positive, implying a negative effect of experiencing an earthquake on the growth process. In the next section this anomaly will be further discussed.

5.2 Heterogeneous effects

Up until now, the treatment effects of the earthquakes have been assumed to be homoge-neous. However, the earthquakes considered in this study differ in both magnitude and reported damages. Therefore, heterogeneous effects are examined, in order to obtain fur-ther understanding of the variation in impact across the five earthquakes. Table 5 reports on these heterogenous results. Earthquake 1 refers to the earthquake on South Sumatra in 2000. Earthquake 2 refers to the earthquake on Bali in 2004. Earthquake 3 refers to the earthquake on Yogyakarta in 2006. Earthquake 4 refers to the earthquake on West Java in 2006. Earthquake 5 refers to the earthquake on West Sumatra in 2009.

Given the results presented in section 5.1, the reported estimated effects of the earth-quakes on stunting are not surprising. Column 2 and 4 of table 5 show that the impact is

(22)

Table 5: Heterogeneous results

(1) (2) (3) (4)

Height Stunting Height Stunting Earthquake 1 -0.5958 0.0192 -0.5543 0.0071 (0.8069) (0.0501) (0.6064) (0.0497) Earthquake 2 -0.2122 0.0327 -0.4660 0.0473 (0.8244) (0.0485) (0.5729) (0.0461) Earthquake 3 -0.5385 0.0572 -0.6713 0.0642 (1.1154) (0.0574) (0.6598) (0.0561) Earthquake 4 0.9324* 0.0140 0.6219* 0.0024 (0.5565) (0.0296) (0.3411) (0.0290) Earthquake 5 -0.0043 0.0426 -0.2764 0.0524 (0.8495) (0.0429) (0.5124) (0.0412) Province fixed effects YES YES YES YES Period fixed effects YES YES YES YES Age cohort fixed effects YES YES YES YES Observations 18629 18629 18629 18629

R2 0.68 0.04 0.90 0.08

Note: The regressions of column 3 and 4 include controls for age, gender, height of parents and household size. When possible, missing values on parents’ height are replaced by values recorded in the previous wave and otherwise replaced by the average of the complete sample. A dummy variable indicating the missing value is added to all specifications that include controls. Robust standard errors are reported in parentheses. *, ** and *** denote significance at the 10%, 5% and 1% level respectively.

(23)

insignificant for all earthquakes, regardless of the inclusion of control variables. There is some variation in magnitude, but all coefficients are positive and insignificant. In contrast, the effects of the different earthquakes on height are not as uniform. Column 1 shows that the effect of the treatment on height is negative, small and insignificant for all earthquakes, except the one in West Java (i.e. ‘Earthquake 4’). Children who experienced the 2006 earthquake in West Java when they were in their first 1000 days of life are estimated to be 0.93 centimeters taller than their peers in the control group. This estimate is significant at the 10% level. When accounting for background characteristics, as in the specification reported on in column 3, this effect drops to 0.62 centimeters, which is significant at the 10% level as well.

Almost half of the group of ‘treatment children’ is made up of children from West Java. Hence, it is not surprising that this group drives the results of the main analysis. A plausible suspicion would be that the impact on height as estimated in section 5.1 would have been negative if the earthquake in West Java is disregarded. In order to test this hypothesis, the analysis is repeated without the inclusion of West Java. The regression output is presented in appendix E. As expected, the estimate now implies a negative effect of an earthquake on growth. This results becomes significant at the 10% level after controlling for confounding factors. Moreover, the effect of an earthquake on stunting, the other dependent variable investigated in this study, becomes significant at the 10% level as well. Although the effects are small, this is still an indication that the 2006 earthquake in West Java is an outlier compared to the other selected earthquakes. The aid flows might have been more substantial or the aid response faster. Also, as pointed out in section 3.1, the NGDC did not report on the estimated damages of this particular earthquake. Damages may have been considerably less severe compared to the other earthquakes. If the aid response was indeed exceptionally effective and fast, or the damages relatively little, this provides an explanation for the divergence of the effect of the earthquake in West Java from the other four earthquake effects. Having said that, omitting West Java from the regression substantially reduces the sample size of the ‘treatment children’, which in turn reduces the reliability of the analysis.

(24)

Table 6: Robustness check

(1) (2) (3) (4)

Height Stunting Height Stunting Earthquake 0.1008 -0.0055 0.2390 -0.0059

R2 0.57 0.04 0.91 0.08

Note: The regressions of column 3 and 4 include controls for age, gender, height of parents and household size. When possible, missing values on parents’ height are replaced by values recorded in the previous wave and otherwise replaced by the average of the complete sample. A dummy variable indicating the missing value is added to all specifications that include controls. Robust standard errors are reported in parentheses.*, ** and *** denote significance at the 10%, 5% and 1% level respectively.

5.3 Robustness

The results discussed in the last two sections are based on a sample of children that were in their first 1000 days of life at the time of the earthquake. As explained before, environmental factors are extremely important in the beginning of life. Shocks are assumed to have the greatest effect in these first 1000 days. To claim that children’s growth is not affected by negative shocks beyond these 1000 days is rather strong, however. In order to relax this assumption and to obtain more observations in the treatment group, the main analysis will be repeated with a sample of children that were approximately in their first 2000 days of life at the time of the earthquake. When including these children in our sample as well, the sample size of the ‘treatment children’ increases from 844 to 1,718. On the one hand, more observations will increase the statistical power of our analysis. On the other hand, the expected effect of an earthquake on growth will be slightly smaller, making it harder to detect an effect.

The results are reported in table 6. All coefficients remain very small and insignificant, implying that an earthquake has no effect on the growth of children that were in their first 2000 days of life. Although insignificant, the sign of the estimates would indicate a positive effect on growth (if significant). An explanation for this result is that the humanitarian response is more effective for slightly older children. The positive impact associated with the aid response may not have only offset, but completely outweighed the negative consequences of the earthquakes. This explanation needs to be approached with care, however. More research needs to be conducted on the relationship between aid responses and early childhood development before dependable inferences can be drawn from the results presented in this section.

(25)

|

6. Conclusion

This thesis applies a difference-in-differences strategy with multiple time periods and treat-ment groups in order to investigate the impact of experiencing an earthquake in the first 1000 days of life on the growth of a child. Five major earthquakes that all occurred in a different province of Indonesia between 2000 and 2014 were selected for this analysis. Data on anthropometry from the Indonesian Family Life Survey was exploited in order to examine the effect of these earthquakes, both pooled and separate, on height and stunting. The results found in this study are inconclusive. The output of the main regressions indicate that there is no effect of an earthquake on child growth. The coefficients of treat-ment on height are positive, yet insignificant. The coefficients of treattreat-ment on stunting are positive (indicating a negative impact of the earthquake on growth), but insignificant as well. Most coefficients remain insignificant when zooming in on the five earthquakes separately. However, the estimated effect of the earthquake in West Java on height lies around 1 centimeter, whereas the estimates for the other earthquakes show an insignificant negative effect.

One explanation for the results is that the aid response has been more effective for the children in West Java compared to the other treatment provinces. Also, the damages caused by this earthquake were possibly less severe. In turn, the positive impact of the aid response potentially outweighed the negative consequences of the disaster in West Java, but not in other treatment provinces. Having said that, the size of the treatment sample is rather small, meaning the effect of potential outliers is large. Hence, another explanation for the estimated results could be low power of the test and a considerable influence of outliers.

In conclusion, no direct inferences can be drawn from the outcomes of this study. Given the absence of reliable data on aid responses and precise data on damages, it is difficult to explain the divergence in treatment effects across the different earthquakes. More research needs to be conducted on the underlying mechanisms through which an earthquake affects child growth before any definite conclusions can be drawn. This thesis treated the earthquakes as an event with homogeneous consequences, but there are various dimension to it that might be essential to uncover. Future research could focus on the ratio between aid and damages in order to assess the importance of the opposing mechanisms discussed in this thesis. Also, the variation in the time that passed since the earthquake can be exploited, in order to deduct the implied relation between this passed time and the treatment effect.

(26)

|

7. References

Alcantara-Ayala, I. (2002). Geomorphology, natural hazards, vulnerability and prevention of natural disasters in developing countries. Geomorphology, 47 (2-4), 107-124. Alderman, H., Hoddinott, J., & Kinsey, B. (2006). Long term consequences of early

childhood malnutrition. Oxford economic papers, 58 (3), 450-474.

Angrist, J. D. & Pischke, J. (2008). Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton University Press.

Bartlett, J. (2015, January 17). Why shouldn’t I use linear regression if my outcome is binary? The Stats Geek. Retrieved from www.thestatsgeek.com/2015/01/17/why-shouldnt-i-use-linear-regression-if-my-outcome-is-binary

Black, R. E., Victora, C. G., Walker, S. P., Bhutta, Z. A., Christian, P., De Onis, M., Ezzati, M., Grantham-McGregor, S., Katz, J., Martorell, R., & Uauy, R. (2013). Maternal and child undernutrition and overweight in low-income and middle-income countries. The Lancet, 382 (9890), 427-451.

Berkman, D. S., Lescano, A. G., Gilman, R. H., Lopez, S. L., & Black, M. M. (2002). Effects of stunting, diarrhoeal disease, and parasitic infection during infancy on cog-nition in late childhood: a follow-up study. The Lancet, 359 (9306), 564-571.

Brancati, D. (2007). Political aftershocks: The impact of earthquakes on intrastate con-flict. Journal of Conflict Resolution, 51 (5), 715-743.

Chang, S. M., Walker, S. P., Grantham-McGregor, S., & Powell, C. A. (2002). Early childhood stunting and later behaviour and school achievement. Journal of Child Psychology and Psychiatry, 43 (6), 775-783.

Dewey, K. G., & Begum, K. (2011). Long-term consequences of stunting in early life. Maternal child nutrition, 7 (s3), 5-18.

Gardner, J. M. M., Grantham-McGregor, S. M., Himes, J., & Chang, S. (1999). Behaviour and development of stunted and nonstunted Jamaican children. The Journal of Child Psychology and Psychiatry and Allied Disciplines, 40 (5), 819-827.

Gertler, P. J., Martinez, S., Premand, P., Rawlings, L. B., & Vermeersch, C. M. (2016). Impact evaluation in practice. The World Bank.

Godoy, R., Tanner, S., Reyes-García, V., Leonard, W. R., Mcdade, T. W., Vento, M., ... & Jha, N. (2008). The effect of rainfall during gestation and early childhood on adult height in a foraging and horticultural society of the Bolivian Amazon. American Journal of Human Biology, 20 (1), 23-34.

Government of Indonesia (2009, October 31). Indonesia: West Sumatra and Jambi nat-ural disasters – damage, loss and preliminary needs assessment. Retrieved from: https://reliefweb.int/report/indonesia/indonesia-west-sumatra-and-jambi-natural-disasters-damage-loss-and-preliminary

(27)

International Child Development Steering Group. (2007). Developmental potential in the first 5 years for children in developing countries. The Lancet, 369 (9555), 60-70. Hoddinott, J., & Kinsey, B. (2001). Child growth in the time of drought. Oxford Bulletin

of Economics and statistics, 63 (4), 409-436.

Index Mundi (2013). Global Database on Child Growth and Malnutrition. Retrieved from: https://www.indexmundi.com/facts/indicators/SH.STA.STNT.ZS/rankings Kar, B. R., Rao, S. L., & Chandramouli, B. A. (2008). Cognitive development in children

with chronic protein energy malnutrition. Behavioral and Brain Functions, 4 (1), 31. Kun, P., Han, S., Chen, X., & Yao, L. (2009). Prevalence and risk factors for posttrau-matic stress disorder: a cross-sectional study among survivors of the Wenchuan 2008 earthquake in China. Depression and Anxiety, 26 (12), 1134-1140.

Liu, E. M., Liu, J. T., & Tseng, T. Y. H. (2015). The Impact of a Natural Disaster on the Incidence of Fetal Losses and Pregnancy Outcomes. Draft, July.

McGuire, B. (2013). Waking the Giant: How a Changing Climate Triggers Earthquakes, Tsunamis, and Volcanoes. Oxford: Oxford University Press.

Mendez, M. A., & Adair, L. S. (1999). Severity and timing of stunting in the first two years of life affect performance on cognitive tests in late childhood. The Journal of nutrition, 129 (8), 1555-1562.

National Geophysical Data Center, World Data Service (NGDC/WDS): Significant Earth-quake Database. National Geophysical Data Center, NOAA. doi:10.7289/V5TD9V7K Nurani, N., Sitaresmi M. N., & Ismail, D. (2011). Risk factors for malnutrition in under-five children: one year after the Yogyakarta earthquake. Paediatrica Indonesiana, 51 (6), 327-31.

Prendergast, A. J., & Humphrey, J. H. (2014). The stunting syndrome in developing countries. Paediatrics and international child health, 34 (4), 250-265.

Raschky, P. A., & Schwindt, M. (2009) Aid, natural disasters and the Samaritan’s dilemma. (Working Paper No. 4952). The World Bank.

Reese, S., Cousins, W. J., Power, W. L., Palmer, N. G., Tejakusuma, I. G., & Nugra-hadi, S. (2007). Tsunami vulnerability of buildings and people in South Java? field observations after the July 2006 Java tsunami. Natural Hazards and Earth System Science, 7 (5), 573-589.

Rydberg, H., Marrone, G., Strömdahl, S., & von Schreeb, J. (2015). A promising tool to assess long term public health effects of natural disasters: combining routine health survey data and geographic information systems to assess stunting after the 2001 earthquake in Peru. PLOS ONE, 10 (6), 1-11.

Stewart, C. P., Iannotti, L., Dewey, K. G., Michaelsen, K. F., & Onyango, A. W. (2013). Contextualising complementary feeding in a broader framework for stunting preven-tion. Maternal child nutrition, 9 (S2), 27-45.

Stock, J. H. & Watson, M. M. (2012). Introduction to Econometrics. Harlow, Essex: Pearson Education Limited.

(28)

UN Resident and Humanitarian Coordinator for Indonesia, (2009). West Sumatra Earth-quake: Humanitarian Response Plan 2009 in coordination with the Government of In-donesia. Retrieved from: www.reliefweb.int/report/indonesia/west-sumatra-earthquake-humanitarian-response-plan-2009-coordination-government

UNICEF (2013). Improving child nutrition: the achievable imperative for global progress. New York: United Nations.

Vidmar, S. I., Cole, T. J., & Pan, H. (2013). Standardizing anthropometric measures in children and adolescents with functions for egen: Update. Stata Journal, 13 (2), 366-378.

Walker, S. P., Wachs, T. D., Gardner, J. M., Lozoff, B., Wasserman, G. A., Pollitt, E., Carter, J.A., & the International Child Development Steering Group. (2007). Child development: risk factors for adverse outcomes in developing countries. The Lancet, 369 (9556), 145-157.

World Bank Group (2015, April 23). The Double Burden of Malnutrition in Indone-sia. Retrieved from: http://www.worldbank.org/en/news/feature/2015/04/23/the-double-burden-of-malnutrition-in-indonesia

Yamano, T., Alderman, H., & Christiaensen, L. (2005). Child growth, shocks, and food aid in rural Ethiopia. American Journal of Agricultural Economics, 87 (2), 273-288.

(29)

Appendices

Appendix A. List of provinces

1. Bali 2. Bengkulu 3. DKI Jakarta 4. Jambi

5. Jawa Barat (West Java) 6. Jawa Tengah (Central Java) 7. Jawa Timur (East Java)

8. Kalimantan Barat (West Kalimantan) 9. Kalimantan Tengah (Central Kalimantan) 10. Kalimantan Selatan (South Kalimantan) 11. Kalimantan Timur (East Kalimantan) 12. Lampung

13. Nusa Tenggara Barat (West Nusa Tenggara) 14. Riau

15. Sulawesi Utara (North Sulawesi) 16. Sumatera Utara (North Sumatra) 17. Sumatera Barat (West Sumatra) 18. Sumatra Selatan (South Sumatra) 19. Yogyakarta

20. Banten - Kepulauan Bangka Belitung - Kepulauan Riau - Sulawesi Barat - Sulawesi Selatan (pooled)

(30)

Appendix B. Missing data check

Table 7: Missing data check (1) Missing height

measurement

Earthquake -0.0034

(0.0137) Province fixed effects YES Period fixed effects YES Age cohort fixed effects YES Observations 20089

R2 0.03

Robust standard errors are reported in parentheses. *, ** and *** denote significance at the 10%, 5% and 1% level respectively.

Appendix C. Placebo test

Table 8: Placebo test

(1) (2) (3) (4)

Age Sex Height mother

Height father Earthquake 0.0628 0.0120 -0.2036 -0.4081

R2 0.14 0.00 0.01 0.01

Robust standard errors are reported in parentheses. *, ** and *** denote significance at the 10%, 5% and 1% level respectively.

(31)

Appendix D. Survivor bias check

Table 9: Survivor bias check (1) Stillbirths and miscarriages Earthquake 0.0163 (0.0275) Province fixed effects YES Period fixed effects YES Observations 31383

R2 0.003

Robust standard errors are reported in parenthe-ses. *, ** and *** denote significance at the 10%, 5% and 1% level respectively.

Appendix E. Hypothesis test

Table 10: Hypothesis test (West Java excluded) Without controls With controls

(1) (2) (3) (4)

Height Stunting Height Stunting Earthquake -0.3417 0.0382 -0.5208* 0.0451*

R2 0.66 0.05 0.89 0.09

Note: The regressions of column 3 and 4 include controls for age, gender, height of par-ents and household size. When possible, missing values on parpar-ents’ height are replaced by values recorded in the previous wave and otherwise replaced by the average of the com-plete sample. A dummy variable indicating the missing value is added to all specifications that include controls. Robust standard errors are reported in parentheses. *, ** and *** denote significance at the 10%, 5% and 1% level respectively.

The impact of experiencing an earthquake in the first 1000 days of life on child growth