Parental mortality and the permanent consequences for health and education

1

PARENTAL MORTALITY AND THE PERMANENT CONSEQUENCES

FOR HEALTH AND EDUCATION

Kevin Kabatsi – 11390964

December 2017

Abstract

This paper presents evidence of the long-term impacts of parental death on children. Additionally, it has the ability to explore whether the age of the child when orphaned is contributory to long-term outcomes. This paper uses 1706 children from the North-West Tanzanian region of Kagera as a sample. All respondents had both parents living at the baseline

period (1991-1994), however by the final wave (2010) only 54.45 percent remained non-orphaned. Those orphaned served as the treatment group, while the non-orphans were the control. Respondents’ height and completed years of schooling served as proxies for health and

education respectively. Height was found to have no significant association with orphanhood. However, years of schooling are significantly fewer for orphaned children. This result is exacerbated if the mortality was maternal or if the bereaved child was a girl. This research finds

that the height and education of young children (0-6 years) is not affected in the long term by orphanhood. Whereas, orphanhood is significantly detrimental to the educational attainment of

older children.1

University of Amsterdam

MSc Economics: Development Economics MSc ECO - 15 ECTS

Supervisor – Pauline Rossi

1 _{I would like to personally thank my supervisor Pauline Rossi for the guidance, patience, and direction, which I}

2

Statement of Originality

This document was written by Kevin Kabatsi, 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 the contents.

3

Table of Contents

1. Introduction ... 4

2. Literature Review ... 5

3. Data and Methodology... 9

3.1 Data ... 9

3.2 Empirical strategy ... 14

4. Results ... 17

5. Conclusion ... 21

6.

References ... 23

7. Appendix ... 25

7.1 Basic Specifications ... 25

4

1. Introduction

Parents provide guidance, protection, and discipline for their children. Therefore, children undergo severe trauma when their parent dies. This trauma unfortunately is not limited to the child’s psychological environment, but also to their economic one. The death of a parent has been long associated with negative socio-economic consequences for the bereaved child; both in the immediate aftermath and later in adulthood. In light of the current state of knowledge, this paper aims to answer the following research question: “What are the long-term socio-economic impacts of losing a parent for a child, and do they vary depending on the age of the child at the time of the loss?”

The channels through which orphanhood can increase risk factors for adult poverty are vital to understand the problem. Development agencies and governmental institutions often have limited resources. Their policies may have greater impact if they are able to target the most vulnerable children. This logic is at the heart of why it is vital to understand the varying impact orphanhood has on children; depending on differing factors like age or gender.

The Kagera Health and Development survey (KHDS) was the principal data used in this analysis. KHDS is a 19 year panel data survey (1991-2010) from the North West Tanzanian region of Kagera. It was carried out during the HIV/AIDS epidemic that plagued the region and

contributed significantly to the deaths of young adults. However, it is vital to mention that the survey does not identify causes of death. This paper may also be used to understand some of the consequences of HIV/AIDS can have on a household.2 Of the UNICEF-estimated 140 million orphans globally in 2015; approximately 52 million were in Africa.3 In the case that this phenomenon persists; it would be imperative that the impact of orphanhood be studied in an African context. (UNICEF, 2015)

Very similar research has been done in this context and with the same data set of children by Beegle et al (2004). This paper will be an extension Beegle et al (2010). This paper contributes

2 _{This paper will focus on orphanhood in general and not specifically HIV/AIDS victims. However because this}

research was done of an area severely hit by the HIV/AIDS epidemic; it could give some insight into possible consequences of the disease.

3 _{This depicts an overrepresentation of African orphans as Africa comparatively only makes up approximately 15.7}

5 to the literature by examining the permanent effect orphanhood has on infant children versus older ones. Due to the shorter length of the panel at the time; it was not possible for Beegle et al (2010) to observe outcomes of children younger than 7 years old. In contrast, this paper has the ability to follow children from birth to adulthood therefore can evaluate outcomes of children younger than 7 years. This helps to answer this paper’s central question of; whether the long term consequences of childhood orphanhood vary between different ages.

The long term outcome variables of interest are years of schooling and height as proxies for education and health; which are generally considered vital human capital investments.

(Ainsworth et al, 2006) Using an Ordinary least squared regression (OLS); this research finds that the death of the parent of an older child corresponds to 0.680 fewer years of schooling on average whilst having no significant impact on height. In contrast for a younger child, the death of a parent has no observed significant impact on either height or years of schooling.4 Possible reasons for the discrepancy between ages will be suggested.

Section 2 will review vital research done previously in this field. Section 3 will detail the type of data used in this study. Section 4 will explain the empirical strategy used to conduct this analysis. Section 5 will evaluate the vital results and findings of this paper. Finally Section 6 will conclude and discuss key aspects from this analysis

2. Literature Review

The impact the loss of a parent has on the schooling- or health- of a bereaved child has been thoroughly researched with varying results, contexts, and methods used. The use of panel data for longer term evaluation is not commonly used especially in a sub-Saharan context. This is often due to the fact that panel data is relatively more expensive to acquire and seldom publically available. Case and Ardignton (2006) have argued that cross sectional data may often be limited in evaluating the schooling consequences of orphanhood as one cannot fully observe the relative wealth and other possible schooling covariates of a child’s household before the parents’ death.

6 In contrast, longitudinal data does a better job of addressing these concerns as one has the child’s socio economic record prior to their parent’s death.

Gertler et al (2004) utilized cross sectional data from Indonesia, and identified that; potential selection bias may exist in households with deceased parents. In an attempt to address this, he used semi-nonparametric matching. This links children who are apparently very similar based on defined pre-determined characteristics with the exception that one child has lost a parent and the other has not.5 _{A comparison of outcomes is then evaluated, with the child who has both parents}

living serving as the control and the one without being the treatment case. While this

econometric technique is effective in this case, it is often accused of leading to biased estimators as covariates may not be sufficiently controlled for. Gertler et al (2004) found in this case that a child who recently lost a parent; is 2.0 times more likely to drop out of school than one with both parents living. The highest rate of drop outs occur in the transitional period between primary and secondary school.

In comparison, Beegle et al (2010) made use of the same longitudinal KHDS data set used in this paper6. Panel data sets allow for both cross-sectional and time series dimensions, which results in more accurate inferences of model parameters and capture more complex human behavior than single cross sectional or time series data. In the Sub-Saharan African context, it was the first data set of its kind, making Beegle et al (2010) the reference paper in this field. They found that the loss of a mother is associated with significant negative impacts on years of schooling and weakly significant negative outcomes for height for the bereaved child. The loss of a mother is

associated with 1.240 less years of schooling and 1.2% shorter in height. In comparison, paternal mortality had no significant effect on a child’s height or years of schooling. The magnitude of impact that maternal orphanhood has on years of schooling is considered very high and sparks awareness of necessity for greater specificity in future policies.

A limitation of Beegle et al (2010) is in the data sampling. This is due to the fact that they only evaluate outcomes for children aged 7-15 years old. All children younger than that were deemed incapable of being evaluated because they would still be of school age by their final wave

5 _{The characteristics used to link children; are based on the data available and intuitions of the author.} 6 _{Beegle et al (2010) was the main paper upon which this paper was based. It provided valuable direction both}

7 (2004).7 This paper uses the same data set but with a timescale up to 2010. This is vital because when evaluating the long term impact of orphanhood, one must evaluate all children capable of being orphans. This shortcoming is very common in a lot of the research because one cannot easily and precisely evaluate permanent health or schooling outcomes for pre-schooling age orphans with cross sectional data. To effectively evaluate their long term outcomes, a

comprehensive panel data set of at least 18 years is required. This allows those born in the first interview year to be of scholastic maturity by the final re-interview.

A similar study was done in a sub-Saharan Africa context using longitudinal data by Case and Ardignton (2006) in South Africa. They found that the loss of a mother was strongly associated with lower levels of school enrollment. While paternal orphans were associated on average with belonging to households of lower income. They observed a slightly negative correlation of paternal orphanhood to schooling outcomes. However, they argued that this was mostly due to the fact that households had lower income following the death of the father. The data was

collected over four years (2001-2004) so the results are not permanent. Additionally there was no age dimension in their analysis because they only observed older children’s outcomes.

In comparison, Case and Paxson (2004) utilized the data of children aged 14 and younger from 10 African countries and 19 Demographic and Health Surveys (DHS). They found that orphans were significantly associated with lower school enrollments than non-orphans. The level of significance and magnitude differed greatly between countries and time periods. However, there is a strong indication of the general direction of influence that the death of a parent has on the schooling of a child. They also find evidence to support the Hamilton’s rule.8 _{This notion argues}

that the degree of relatedness between children and their adult caregivers; is highly predictive of the children's outcomes. Orphans are relatively more likely to live with caregivers who they are not related to. This may be a possible channel to explain why orphans on average have worse schooling outcomes than non-orphans. Case and Paxson (2004) provided a vital macro-economic perspective on this issue in the African context. However, they utilized cross sectional data. This

7 _{There are 6 waves that span 19 years 1991-2010. Beegle et al (2010) used the first 5 waves, with the first 4}

(1991-1994) as a baseline and the 5th _{wave to evaluate final outcomes.}

8 _{Hamilton’s rule originates from the work of the British evolutionary biologist W.D. Hamilton. This was a central}

theorem of kin selection theory. Using mathematics; he argued that certain social behavior can be determined by relatedness, benefit and cost. It has since been tested across many academic disciplines.

8 means that selection into orphanhood from unobserved covariates; may not have been adequately controlled for in their models.

Similarly Mishra et al (2007), using data from rural Kenya; found that orphans are relatively less likely to attend school. However, they also analyzed the nutritional status of children using an age dimension. They looked at the impact orphanhood had on the height and weight of children in two groups (0-4 years and 5-14 years). They observe no consistent relationship between orphanhood and these health proxies regardless of age.9 _{They did not examine schooling}

outcomes for orphans dependent on their age. This leaves room for this paper’s contribution to the literature.

There are several pathways through which orphanhood may affect child outcomes. Ainsworth et al (2000) identify the income effect as the most pertinent. This refers to when household income is reduced following the death of a parent, especially in credit-constrained settings. This may lead to decreases in the investment in both the health and education of children. This could be further exacerbated when households incur significant costs before and after the death of the parent.10 Looking past direct wealth effects; there could also be an increase in the child’s value for household production. In other words, children may serve as substitutes for the deceased adults’ labor. This dynamic could likely prove detrimental to time spent at school.

There is an array of possible household behavioral responses that may arise following the death of a parent; and subsequently affect outcomes. Gertler et al (2004) has argued that after a child loses their biological parent, the household’s preference for the child’s “quality” might change. In other words, the child’s foster-caregivers might prefer their own biological children over the orphan when disseminating human capital investments. This phenomenon would be likely difficult to research as it would rely on self-reporting.11 _{If this effect is found by researchers,}

then the relevant actors could have better direction for policy.

9 _{This mirrors the results later discussed in this paper. Height as a proxy for health appears to be unaffected by}

orphanhood.

10 _{This refers to the medical expenses households often undertake before a death. Additionally following a death;}

there are many costs not limited to but include the funeral.

11 _{Foster parents are unlikely to admit during interviews; that they are treating various children differently within the}

household. Similar issues are encountered; when researching intra-household gender preferences for human capital investments.

9 The socio-economic consequences of orphanhood cannot be viewed in a vacuum. The trauma or psychological toll of a parent’s death on a child may significantly influence long term outcomes. This may be exacerbated in certain societies, where the children of HIV/AIDS victims face severe stigma. (UNICEF, 2015) Traditionally an orphan is defined as an individual who doesn’t have either biological parent living. However in this thesis, an orphan will be defined as any respondent who has lost at least one biological parent. Beegle et al (2010) used the term

“orphan” in a similar fashion and justified if for usability reasons. They argued; it was useful for grouping bereaved children who had experienced variations of parental mortality.

The death of a parent usually corresponds to a decrease in the household’s amount of income and personal resources available to develop children. Therefore, there must be counterbalancing phenomena that explains how some households mitigate the shock of orphanhood. Gertler et al (2004) define three examples of these possible phenomena; “intergenerational altruism, mutual insurance, and intra-household allocation.” Intergenerational altruism refers to when other surviving family members substitute both time and income to raise the bereaved child.12 He argues that these mechanisms can ensure that long term economic outcomes are not affected by orphanhood. Mutual insurance refers to; two or more separate households committing to aid one another if they face severe shocks such as the death of parents. Intra-household allocation refers to the community at large assisting bereaved households where they can. Exploring the validity of these channels of mitigation, is outside the scope of this paper. However they are vital subjects for future research; to assist in effective policy implementation.

3. Data and Methodology

3.1 Data

The KHDS was the primary data set used for this research paper. The KHDS dataset was collected by Economic Development Initiatives (EDI) and the World Bank based on the World Bank’s Living Standards Measurement Study (LSMS). The KHDS carried out a total of six waves, with the first four taking place annually in 1991-94, the fifth in 2004 and the last in 2010. In line with the objectives of the World Bank’s LSMS, KHDS data is very comprehensive in its

12 _{This is commonly done by the child’s grandparents, aunts and uncles. That is why this form of altruism is referred}

10 collection of socioeconomic information from respondents. This information encapsulates all society’s relevant dimensions, namely individual, household and community. The major success behind this panel data set is that significant resources were spent trying to ensure the tracking and re-interview of baseline respondents. The dataset achieved a re-contact rate in the final wave of 89 percent, excluding respondents who died.

The Kagera region of north western Tanzania borders Rwanda and Burundi to its west, Uganda to the north and Lake Victoria to the east. The population of this region both at baseline and by the final wave was primarily rural and agrarian. Rain fed crops such as maize and cotton are the most common commercial output in the region. Kagera was arguably one of the first regions in Africa to be significantly hit by the HIV /AIDS epidemic. This played a significant role in the increase in adult mortality witnessed at the time. It is for this reason that it is a useful context to study the intergenerational impact of the disease through the channel of orphanhood. The KHDS surveys, along with the relatively low attrition rate, covers a relatively long period of time (19 years), making it exceptional compared to other publically available panel surveys in similar contexts.

The analysis done on this data set used the first four waves (1991-1994) as a baseline and the sixth wave in 2010 to evaluate outcomes for re-interviewed respondents. The KHDS baseline sample was fairly similar to a random sample of households from Kagera at the time. However, it was not a random sample as households which were expected to experience an adult death were oversampled. The sampling at baseline was done in two stages; the first dealing with the selection of communities and the second with sampling households from within these

communities.

The first stage of sampling was done randomly with 51 primary sampling units (PSUs) selected from 550 geographically demarcated communities within the Kagera district.13 _{Using data from}

an enumeration study it conducted in March to June 1991 in the region, KHDS classified households as “sick” or “well”. These categories were based on whether an adult death due to illness had occurred in the household in the previous 12 months. The second stage of sampling

11 was then conducted. Within each of the 51 PSUs, 16 households were selected, with 14

randomly drawn from the “sick” category and 2 drawn randomly from the “well” group. Households that had experienced an adult mortality in the past 12 months only comprised 3.7 percent of the 29000 households in the enumerated area. This overrepresentation of “Sick” households raises concerns about the representative nature of the data set. The results of this paper must regarded while taking this into consideration.14 This was done because the impact of the HIV/AIDS epidemic was of research interest. A completely random sampling would likely have resulted in insufficient households who recently suffered adult deaths, being part of the dataset. The final sample drawn for the first passage in 1991 consisted of 816 households. This figure eventually rose to 897 by 1994.

The baseline sampling done for the purposes of this research paper used the data of the 816 households sampled in Wave 1 (1991) as well as the additional 81 households which had joined by the Wave 4 (1994). At baseline, these households contained both orphaned and non-orphaned children. When looking at the 0-15 age range, 29.75 percent (914 children) had already lost either one or both biological parents while 70.25 percent (2,158 children) had both biological parents living. All children aged 0-15 with two living parents in Wave 1 (1991) qualified for analysis. However only data from children aged 3-15 years could be used from Wave 4 (1994). All respondents should have been at least 19 years by the time they were re-interviewed in 2010. Of the 2,158 children who had both living parents in Waves 1-4 and would be at least 19 years of age in 2010; 1,706 were successfully re-interviewed in 2010.15

Table 1: Total respondents eligible for evaluation by Final Wave (2010)

Final Orphan status (2010) No. of respondents Percentage

At least one parent dead (Orphan) 777 45.55

Both parents living (Non-Orphan) 929 54.45

Total 1706

As shown in Table 1 above, 45.55 percent of all respondents who were re-interviewed had become orphans. Orphanhood is defined by the loss of at least one biological parent. The

14 _{The severe majority of respondents of this data set were drawn from a minority category of society. Therefore}

how much the results can be applied to the entire society; comes into question.

12 remaining 929 respondents in the sample still have both of their biological parents living. These two groups represent the treatment and control groups in the analysis.16

Figure 1: The composition of orphans by the re-interview period (2010).

Note: The total amount of orphans is 777. This refers to respondents who have lost at least one biological parent.

The figure above depicts the distribution of the 777 orphaned respondents. Whereby 58.56 percent had lost only their father, 17.12 percent just their mother and 24.32 percent had lost both parents. The predominance of paternal orphanhood in this context gives evidence of a disparity in life expectancy between men and women in Tanzania.

16 _{The treatment group are the orphans and the control group; the non-orphans.}

59% 17%

24%

13

Note: The figures in bold are the calculated means from the sample of respondents used for analysis. The figures in italics are the means’ standard deviations. Two tailed T-tests were run on all variables to see if the variable means are significantly different between orphans and non-orphans. Level of significance is denoted by stars. * refers to significant at the 5% level while ** refers to significant at the 1% level. Finally “NS” indicates no significance. Variables denoted with “+” are binary variables whose values range from 0-1.

The above table depicts the differences in means of key baseline characteristics of the two types of respondents. Both groups at baseline had both biological parents living. Respondents could either have become orphaned or remained non-orphaned if; they still had both parents living by the re-interview. It is vital to mention that the key outcome variables (years of schooling and height) had low recordings at baseline because the respondents were on average quite young (6.84 years for non-orphans and 7.78 for orphans). It appears that at baseline; there was no significant difference in average years of schooling for the two groups of children. Contrarily there seemed to be a significant difference in their average height. However, one may look at the higher average age of children who became orphans as an explanation for why they were taller on average at baseline.

Another vital aspect that should be highlighted is that per capita household consumption and quality of the dwelling’s floor were not significantly different between the control and treatment groups. These variables are proxies for household income or wealth levels. (Beegle et al, 2010) Therefore, this would indicate that, in this context a difference in socio-economic background may not have played a substantial role in a child’s selection into orphanhood. Additionally,

Level of Significance Years of schooling 0.83 1.71 0.91 1.66 ns

Height (in cm) 110.55 26.67 115.82 26.22 ** Age (in years) 6.82 4.52 7.78 4.52 **

Male + 0.49 0.5 0.53 0.5 ns

Mother resides in Household + 0.85 0.36 0.75 0.43 ** Father resides in Household + 0.76 0.42 0.71 0.46 ** Household head years of schooling 5.04 3.56 4.27 3.2 ** Household head Male + 0.91 0.3 0.86 0.42 ** Household head Age 46.55 14.84 53.73 16.73 ** Ln per Capita Consumption (TZ shillings) 7.03 0.68 7.09 0.64 ns Dwelling has Good flooring + 0.15 0.35 0.12 0.33 ns

Number of observations 921 777 Table 2: Mean differences in Baseline Characteristics by Future Orphan Status

14 respondents who remained non-orphans by the re-interview were more likely to have to been living with both their mom and dad at baseline. While those who became orphans had household heads who were more likely to be older and female. Unsurprisingly the gender of the child was not significant in determining selection into orphanhood with no gender gap apparent at

baseline.17

Note: The figures in bold are the calculated means from the sample of respondents used for analysis. The figures in italics are the means’ standard deviations. Two tailed T-tests were run on all variables to see if the variable means are significantly different between orphans and non-orphans. Level of significance is denoted by stars. * refers to significant at the 5% level while ** refers to significant at the 1% level. Finally “NS” indicates no significance.

The above table depicts the differences between means of the key outcome variables (height and years of schooling) at the final wave of the survey. The respondents are the two groups of interest: those that had become orphans in adulthood and those that had not. Average height between the two groups is not significantly different in this context. In contrast, years of schooling is significantly higher for children who remained non-orphans. It is pertinent to mention that this alone does not indicate a causal link between future orphan status and years of schooling nor does it disprove one with height. These links can be better identified when

modeled with controls for relevant covariates.

3.2 Empirical strategy

The impact orphanhood has on the long run health and education for children aged 0-15 years; is what is being examined in this paper. The basic specification for years of schooling in this analysis can be written as:

S

= α + βX

+γD

+ δS

+ ∂

ε

17 _{The selection into gender as a child at birth; is generally considered random. There also exists no obvious reason}

for why the gender of the child could affect parental mortality.

Level of Significance

Years of schooling 8.29 3.56 7.58 3.63 **

Height (in cm) 162.37 8.67 162.74 8.56 ns

Number of observations 921 777

Remains Non-orphan Becomes an orphan Table 3: Mean differences in adult height and school attainment by adult orphan status

15 Whereby Si1 is the years of school in 2010. The vector Xi0 is a set of baseline individual,

household and community control variables which were identified in the first four waves of KHDS (1991-1994).

∂

(districts) demarcated in the Kagera region. Di is a dummy variable of whether a respondent is an

orphan or not in 2010 and finally Si0 is years of schooling at baseline.18 Future orphan status (Di)

implies that between baseline and the final wave in 2010, respondents either had at least one parent die to be classified as an orphan. Alternatively they would need to have both parents living to be classified as non-orphan. This specification is estimated using OLS and therefore is equivalent to using a difference in differences specification with controls. This can be easily conveyed by subtracting Si0 from both sides of the equation. The basic specification for height as

a proxy for health is identical to that of years of schooling, except that height is modeled in logarithmic form as opposed to levels, as in years of schooling. This specification is very similar to that used by Beegle et al (2010).

It is vital that a respondent’s baseline years of schooling and height are respectively controlled for. It is vital because one must control for initial heterogeneity correlated with schooling history, health endowments and the child’s general developmental background. The above specification uses γ to quantify the average treatment effect of orphanhood on years of schooling. Aside from a small number of cases, both height and years of schooling barring a few examples are

considered quite permanent once adulthood is reached making them ideal for long term analysis. The utility of height as a measure is seen when comparing it to weight which can change

reflecting temporary circumstances, disqualifying it for longer term analysis.

Years of schooling are a proxy for investments in a child’s human capital. It is measured by each year of schooling completed by a respondent. There has been substantial research done in

varying contexts about the returns both long and short term of schooling. These returns often come in the form of income and labor-force participation, with contradicting levels of

significance and magnitude. Conclusive research on the real returns of education is still

18 _{It is vital to reiterate that the baseline period refers to the first four waves of the survey (1991-1994). The 6}th _and

16 outstanding. However, there is a virtual consensus that education is beneficial for child

development as its returns are arguably more than just economic.19 (Mishra et al, 2007)

Height was chosen as a proxy for health status because of its permanence and apparent positive impact on wage earning capacity. (Thomas et al, 1997) In this analysis, height in centimeters was converted to logarithmic form, making it easier to interpret, as results can be virtually understood as percentages.20

Covariates are important when modeling the long-term impact of orphanhood on height and years of schooling which are influenced by several baseline factors that may likely be correlated with future orphan status. These include individual characteristics of the respondent (sex and age), the household socioeconomic atmosphere at baseline (whether the child lived with their father, similarly for their mother and the household head’s educational attainment, age, and sex) and finally baseline indicators of wealth (log per capita household consumption, whether the domicile flooring was good quality). In addition to these covariate controls, all the regressions in this paper control for community fixed effects at baseline. The capture of community fixed effects helps to control for community specific factors such as access to schools and health services.

Subsequent further analysis was conducted using more specific subsets of orphanhood as dependent variables. The basic model specification applies except for the definition of the dummy variable of orphanhood. Paternal vs maternal orphans for example are regressed together on height and schooling as dummy variables.21 In the case of paternal orphans, a respondent scores 1 if their father has died and 0 if he is alive. The same structure applies for a maternal orphan. This means if a respondent has lost both parents then they would score 1 in both the paternal and maternal orphan dummy variables. Likewise if a respondent had both parents living they would score 0 in both.22 The dimension of age was coded similarly, with the two dummies

19 _{Non-economic benefits refer to societal benefits. Educated people often are found to be relatively more law}

abiding for example.

20 _{Logarithmic transformations aid in the interpretation of results for variables with large values. In this case of}

height, it is easier to understand for example that a group of children on average is 3 percent shorter than say 9 cm shorter.

21 _{Regressed together refers to both dependent variables of interest (paternal and maternal orphanhood) being in the}

same regression.

22 _{The full regression outputs are available in the appendix. This relates to both the parental gender and age}

17 “young orphan” and “old orphan”. Taking “young orphan” as an example, respondents score 1 if they were young at baseline (0-6 years) and have lost a parent, scoring 0 otherwise. This amounts to a comparison between young orphans and all the other respondents (young non orphans, old orphans, and old non-orphans). The same variable construction applies for the “old orphan” dummy, except that the baseline age range is 7-15 years. These two orphan configurations (age and gender of deceased parent) allow for substantial analysis that is useful for policy

formulation.

As is often the case with regression analyses, there are endogeneity concerns. Baseline outcomes for height and years of schooling may be highly correlated with future orphan status and,

therefore with the respondents’ final height and education outcomes in 2010. Beegle et al (2010), who had a very similar model in the same context, attempted to address this issue by using instrumental variables (IVs). They used past rainfall data and share of crop harvest lost in the year preceding the baseline surveys as instruments for schooling and height respectively. These instruments had questionable levels of relevance and strength but were sufficient in their capacity to circumvent concerns of endogeneity. The results, however, did not vary greatly in magnitude or significance; between when she used the IVs and when she used a basic OLS. For this reason the same exercise was not repeated in this analysis as the necessity of the instruments was not substantial.

4. Results

Table 4: Determinants of Height and Schooling in 2010 (basic model)

Variables Dependent variables

Ln height (1) Years of Schooling (2)

Orphan 0.000647 -0.505***

(-0.00246) (-0.177)

Observations 1,324 1,465

R-squared 0.477 0.319

Note: Both these regressions are estimated with community fixed effects. Robust standard errors are in parentheses. As described earlier, baseline height and schooling are controlled for in this equation. Both these regressions have controls. These include individual characteristics of the respondent (sex and age), the household socioeconomic atmosphere at baseline (whether the child lived with their father, similarly for their mother and the household head’s

18

educational attainment, age, and sex) and finally baseline indicators of wealth (Log per capita household

consumption, whether the domicile flooring was good quality i.e. cement or tiles). There also includes controls for the various baseline waves 1991-1994; as baseline information from some respondents was recorded in different years. The variable orphan is a dummy variable where 1 corresponds to having at least one parent deceased and zero otherwise.

Significance is denoted: *** p<0.01, ** p<0.05, * p<0.1.

The above table shows the basic regression model to assess the impact of being an orphan. This model looks at orphanhood’s impact (regardless of age, gender or the deceased parent’s gender) on human capital outcomes. These regression results represent the average treatment effect of the loss of a parent. In this basic form of the model, the loss of a parent is significantly associated with a decrease in years of schooling of approximately half a year. This result is significant at the 1% level and are in line with previous research looking at orphanhood’s impact on schooling. There are several possible pathways, both socio-economic and socio-psychological, by which a loss of a parent may negatively impact a child’s schooling. However, they are not identified by this model.

When investigating these results further, there is an apparent gender dimension with regards to standard orphanhood. The magnitude and significance of the coefficients is evidently higher for girls than boys. Orphaned girls are associated with 0.6 less years of schooling. This compared to boys who are associated with 0.4 less years of schooling when orphaned, with this relationship weakly significant at the 10% level. The other outcome variable of interest in this model (height) appears to not have any significant associations with the loss of a parent.

Table 5: Determinants of Height and Schooling in 2010 (Parental gender model)

VARIABLES Dependent Variables

Ln Height (1) Years of schooling (2)

Father is dead -0.0000495 -0.182 (-0.00244) (-0.183) Mother is dead 0.0012 -0.626*** (-0.00327) (-0.212) Observations 1,320 1,461 R-squared 0.478 0.32

Note: Both these regressions are estimated with community fixed effects. Robust standard errors are in parentheses. As described earlier, baseline height and schooling are controlled for in this equation. Both these regressions have

19

controls. These include individual characteristics of the respondent (sex and age), the household socioeconomic atmosphere at baseline (whether the child lived with their father, similarly for their mother and the household head’s educational attainment, age, and sex) and finally baseline indicators of wealth (Log per capita household

consumption, whether the domicile flooring was good quality i.e. cement or tiles). There also includes controls for the various baseline waves 1991-1994; as baseline information from some respondents was recorded in different years. The variable “father is dead” is a dummy variable where 1 corresponds to a respondent’s father being

deceased and zero otherwise. In that case, it is a comparison between paternal orphans and any respondent who has a father living (including maternal orphans). The same structure applies to “Mother is dead.”

Significance is denoted: *** p<0.01, ** p<0.05, * p<0.1.

The above table evaluates the impact of the loss of a parent on various outcomes for a child depending on the parent’s gender. The death of a father does not appear to have any long-term associations with the height or schooling of the bereaved child, regardless of the child’s gender. This null finding regarding the death of a father replicates that of Beegle et al (2010). However these outcome variables do not necessarily account for the possible psychological impacts associated with the loss of a father.

The loss of a mother, on the other hand, is associated with an average of a 0.62 year decrease in a child’s schooling. This may be attributed to the roles within households that women and men play when raising children. This result may prove that women invest relatively more of both income and time resources in their children. This would explain the relatively more severe effect of their death on their children’s schooling. This is very much in line with previous research, however the magnitude is much lower than the one examined by Beegle et al (2010). This may likely be because Beegle et al (2010) excluded children aged 0-6 from analysis while they have been included in this paper. This result may aid policy as children’s schooling outcomes on average are worse when they lose a mother versus a father.

Another gender dimension is discovered when these results are further unpacked. For both girls and boys, schooling is not found to be negatively associated with the loss of a father. However, when compares to boys, girls’ schooling appears to be more negatively respondent to the loss of a mother. When maternally orphaned, girls are associated an average 0.82 year decrease in schooling. Compared to boys who are associated with a 0.5 year decrease in schooling when their mother dies. This result may point to the increased opportunity cost of schooling for children when their parents die. Girls, who are more substitutable within a household for the roles of their mothers, are associated with larger decreases in schooling when their mothers die.

20 Household activities often performed by mothers or adult women, such as raising children, are the type of roles that might be believed to be better suited for girls. Thus, the opportunity cost of being at school is much higher for girls than boys when their mothers die. This finding may suggest that the increase of the opportunity cost of being at school is a likely channel where orphanhood impacts schooling. However, it cannot be definitively isolated in this model.

Table 6: Determinants of Height and Schooling in 2010 (Age model)

VARIABLES Dependent variables

Ln Height (1) Years of Schooling (2)

Young & Orphaned 0.00372 -0.247

(-0.00334) (-0.264)

Old & Orphaned -0.00132 -0.680***

(-0.00313) (-0.208)

Observations 1,328 1,470

R-squared 0.478 0.32

Note: Both these regressions are estimated with community fixed effects. Robust standard errors are in parentheses. As described earlier, baseline height and schooling are controlled for in this equation. Both these regressions have controls. These include individual characteristics of the respondent (sex and age), the household socioeconomic atmosphere at baseline (whether the child lived with their father, similarly for their mother and the household head’s educational attainment, age, and sex) and finally baseline indicators of wealth (Log per capita household

consumption, whether the domicile flooring was good quality i.e. cement or tiles). There also includes controls for the various baseline waves 1991-1994; as baseline information from some respondents was recorded in different years. The variable “Young & Orphaned” is a dummy variable where 1 corresponds to a respondent who was aged 0-6 years at baseline and lost at least one parent. And zero otherwise. In that case, it is a comparison between younger orphans and any respondent who was older and/or non-orphaned. The same structure applies to “Old & Orphaned” except the age range is 7-15 years at baseline.

Significance is denoted: *** p<0.01, ** p<0.05, * p<0.1.

The table above relates to the analysis of the long term impact of the loss of a parent

differentiated by the child’s age at baseline. The first notable finding was that for all outcome variables, losing a parent as a young child has no significant long-term impact, even when the gender of both the parent and the child are examined. With limited social welfare available from the state to help bereaved children, support is most likely to have come from within extended families and the community at large. This, in combination with possible beliefs that suggest that younger children are considered easier to integrate into new families, may arguably be enough to mitigate the likely negative impacts of orphanhood in the long term. (Cas et al 2014) This could be an avenue for future research: to evaluate whether informal insurance exists in rural

21 communities in Tanzania and its effectiveness in combating permanent household income

shocks.

The effect of orphanhood when a child is older, however, seems to be dramatically different. The loss of a parent is associated with 0.68 years less of schooling for children aged 7-15 years at baseline. This result is significant at the 1% level. There is a slight gender dynamic with older girls’ years of schooling slightly more adversely impacted by orphanhood than their male

counterparts.23 _{Similarly the loss of a mother for an older child is associated with 0.71 years less}

of schooling while the death of a father is not associated with any significant effect on the years of schooling of an older child.

An attempt was made to evaluate the effect orphanhood had on children dependent concurrently on their gender, age and the gender of their deceased parent. For example, the impact on the height and education for the average young girl who had lost her mother. This would have been ideal for understanding key scenarios and possibly isolating channels of influence. However, the data was limited in its number of its observations to evaluate these interactions reliably.

The channel through which the loss of a parent would affect schooling for older children has not been isolated in this paper. However, it is plausible that the relatively greater income earning capacity of older children likely makes the opportunity cost of them going to school higher after a parental death. This opportunity cost channel could also be working in combination with the notion that older children are less preferred for integration into new families resulting in negative outcomes for schooling.

5. Conclusion

According to the findings above, orphanhood has no observed long-term impact on height. However, it is negatively associated in varying magnitudes with the years of schooling for the bereaved children. Magnitudes vary dependent on the respondent’s age, gender and the gender of their parents. The death of a father was found not to have an impact on both height and years of schooling. However, the death of a mother is significantly associated with lower years of schooling particularly for girls. Orphanhood for younger children (0-6 years) was found to have

22 no permanent consequences for the child’s height or education. While orphaned older children (7-15years) corresponded to significantly worse schooling outcomes. This holds particularly if the child was a girl or if the child’s mother was the parent who died. One can conclude that, to varying degrees, older children, girls and maternal orphans are particularly vulnerable to the effects of orphanhood in the long term.

States, NGOs or any institutions that aim to invest in the mitigation of the effects of orphanhood in practice, are quite limited in their budgets. The findings of this paper and others like it can be used to direct policies towards those who are, on average most vulnerable. This paper is limited in that it would have been ideal if the data was large enough to identify which specific

demographic of orphanhood was most vulnerable.24 Evaluating simultaneous interactions would have made it easier to give more specific policy recommendations. That said, much can still be recommended despite these limitations. When interpreting the results of this paper, it is

important to remember that there are still many psychological consequences associated with orphanhood that are not taken into consideration by this survey. The loss of a father or being orphaned at a young age, while not apparently affecting schooling or height outcomes, does not imply that the bereaved child is in the same socio-psychologic condition as prior to the loss of their parent.

Following the identification of which orphans are most vulnerable, begs the question of what the best policy to address this issue may be. This is where further research that identifies the

channels by which orphanhood affects schooling is most valuable for policy makers. Income loss and intra-household labor substitution, for example, may have similar negative impacts for the child’s schooling but the policies required to address either scenario probably differ greatly. Policies that are tailored to which people are affected by orphanhood and how, will likely achieve the best results. In the end, orphanhood, as with most socioeconomic shocks, does not affect everyone equally. Therefore, the mitigation of its consequences needs to be appropriately focused if there is any hope of sufficiently addressing it.

24 _{For example examining who is more vulnerable between old girls who have lost a mother vs young boys who are}

23

6. References

Ainsworth, M., K. Beegle, and G. Koda. (2005). “The Impact of Adult Mortality and Parental Deaths on Schooling in Northwestern Tanzania.” Journal of Development Studies 41:412–39. Ainsworth, M. and D. Filmer. 2006. “Inequalities in Children’s Schooling: AIDS, Orphanhood, Poverty, and Gender.” World Development 34:1099–128.

Beegle, K. De Weerdt, J. Dercon, S. (2010) Orphanhood and human capital destruction: is there persistence into adulthood ? Demography, 47(1), pp. 163-180.

Beegle, K. Krutikova, S. (2008) Adult mortality and children’s transition into marriage. Demographic Research, 12(42), pp. 1551-1574.

Cas, AG. Frankenberg, E. Suriastini, W. Thomas, D. (2014). The impact of parental death on child wellbeing: Evidence from the Indian Ocean Tsunami. Demography 51(2), pp. 437-457 Case, A., C. Paxson, and J.H. Ableidinger. (2004). “Orphans in Africa: Parental Death, Poverty and School Enrollment.” Demography 41:483–508.

Case, A. and C. Ardington. (2006). “The Impact of Parental Death on School Outcomes: Longitudinal Evidence From South Africa.” Demography 43:402–20.

Centers for Disease Control (CDC). (2000). “Cause-Specific Adult Mortality: Evidence From Community-Based Surveillance—Selected Sites, Tanzania 1992–1998.” Morbidity and Morality Weekly Report 49:416–19.

Crampin, A.C., S. Floyd, J.R. Glynn, N. Madise, A. Nyondo, M.M. Khondowe, C.L. Njoka, H. Kanyongoloka, B. Ngwira, B. Zaba, and P.E. Fine. (2003). “The Long-term Impact of HIV and Orphanhood on the Mortality and Physical Well-being of Children in Rural Malawi.” AIDS 17:389–97.

Deininger, K., M. Garcia, and K. Subbarao. (2003). “AIDS-Induced Orphanhood as a Systemic Shock: Magnitude, Impact, and Program Interventions in Africa.” World Development 31:1201– 20.

Evans, D. and E. Miguel. (2007). “Orphans and Schooling in Africa: A Longitudinal Analysis.” Demography 44:35–57.

Fitzgerald, J., P. Gottschalk, and R. Moffitt. (1998). “An Analysis of Sample Attrition in Panel Data.” Journal of Human Resources 33:251–99

Gertler, P., D. Levine, and M. Ames. (2004). “Schooling and Parental Death.” Review of Economics and Statistics 86(1):211–25

Hargreaves, J.R. and J.R. Glynn. (2002). “Educational Attainment and HIV-1 Infection in Developing Countries: A Systematic Review.” Tropical Medicine and International Health 7:489–98.

24 Ksoll, C. (2007). “Family Networks and Orphan Caretaking in Tanzania.” University of Oxford Economics Working Paper Series 361. Department of Economics, University of Oxford, United Kingdom.

Kwesigabo, G., J. Killewo, W. Urassa, J. Lugalla, M. Emmelin, A. Mutembei, F. Mhalu, G. Biberfeld, S. Wall, and A. Sandstrom. (2005). “HIV-1 Infection Prevalence and Incidence

Trends in Areas of Contrasting Levels of Infection in the Kagera Region, Tanzania, 1987–2000.” Journal of Acquired Immune De¿ ciency Syndrome 40:585–91.

Mishra, V. Arnold, F. Otieno, F. Cross, A. Hong, R. (2007) Education and nutritional status of orphans and children of HIV-infected parents in Kenya. AIDS Education and Prevention. 19(5), pp. 383-395

Thomas, D. and J. Strauss. (1997). “Health and Wages: Evidence on Men and Women in Urban Brazil.” Journal of Econometrics 77:159–85

World Bank. (2010). “User’s Guide to the Kagera Health and Development Survey Datasets.” Document. World Bank, Washington, DC.

UNICEF. (2015) Press Centre: Orphans. [online]. Available at:

https://www.unicef.org/media/media_45279.html [Accessed 08 June 2017].

Yamano, T., and T.S. Jayne. (2005). “Working-age Adult Mortality and Primary School Attendance in Rural Kenya.” Economic Development and Cultural Change, vol.53, in press. Yamano, T., Y. Shimamura, and D. Sserunkuuma. (2006). “Living Arrangements and Schooling of Orphaned Children and Adolescents in Uganda.” Economic Development and Cultural Change 54:833–56.

25

7. Appendix

7.1 Basic Specifications

a) Below is the basic model specification for height. The outcome variable height is in logarithmic form for ease of interpretation. All the logarithmic variable results correspond to their results in level form. This refers to the death of either parent.

lnH

= α + βXi0

+γD

+ σH

+ ∂

ε

b) Below is the basic model specification for years of schooling and parental gender mortalities. FDi and MDi stand for the father and mother deaths in the final wave.

S

= α + βXi0

+

θFD

+ ζMD

+ δS

+ ∂

ε

c) Below is the basic model specification for log height and parental gender mortalities.

lnH

= α + βXi0

+

θFD

+ ζMD

+ σH

+ ∂

ε

d) Below is the basic model specification for years of schooling and the two age groups of orphaned children. YOi refers to the young orphan variable and OOi is old orphan.

S

= α + βXi0

+

λYO

+ ξOO

+ δS

+ ∂

ε

e) Below is the basic model specification for log height and the age groups for orphans.

26

7.2 Complete Regression Tables

Note: Both these regressions are estimated with community fixed effects. Robust standard errors are in italics. As described earlier, baseline height and schooling are controlled for in this equation. Both these regressions have controls. These include individual characteristics of the respondent (sex and age), the household socioeconomic atmosphere at baseline (whether the child lived with their father, similarly for their mother and the household head’s educational attainment, age, and sex) and finally baseline indicators of wealth (Log per capita household

consumption, whether the domicile flooring was good quality i.e. cement or tiles). There also includes controls for the various baseline waves 1991-1994; as baseline information from some respondents was recorded in different years. The variable orphan is a dummy variable where 1 corresponds to having at least one parent deceased and zero otherwise. All the dummy control variables are denoted with “+” Significance is denoted: *** p<0.01, ** p<0.05, * p<0.1.

VARIABLES

Ln Height Years of Schooling

Orphan 0.000647 -0.505*** -0.00246 -0.177 Baseline height 0.00181*** 0.0128 -0.000171 -0.0103 Baseline years of schooling 0.00225*** 0.559*** -0.000804 -0.0669 Age in years -0.00967*** -0.310*** -0.00106 -0.0655 Male+ 0.0613*** 0.741*** -0.0023 -0.16

Household Head Years

of Schooling -0.000693* 0.183***

-0.000417 -0.0285

Household Head male+ -0.00487 -0.603**

-0.0034 -0.259

Household Head age -4.22E-05 0.0138**

-7.85E-05 -0.00569 Ln per capita consumption 0.00688*** 0.375** -0.00209 -0.155 Household Size 0.00132*** 0.0206 -0.000338 -0.0219

Floor building material

good+ 0.00118 2.039*** -0.00392 -0.268 Wave4 control+ -0.0115** -0.381 -0.00469 -0.317 Constant 4.873*** 4.029*** -0.0193 -1.364 Observations 1,324 1,465 R-squared 0.477 0.319 Dependent Variables

27

Note: Both these regressions are estimated with community fixed effects. Robust standard errors are in italics. As described earlier, baseline height and schooling are controlled for in this equation. Both these regressions have controls. These include individual characteristics of the respondent (sex and age), the household socioeconomic atmosphere at baseline (whether the child lived with their father, similarly for their mother and the household head’s educational attainment, age, and sex) and finally baseline indicators of wealth (Log per capita household

consumption, whether the domicile flooring was good quality i.e. cement or tiles). There also includes controls for the various baseline waves 1991-1994; as baseline information from some respondents was recorded in different years. The variable “father is dead” is a dummy variable where 1 corresponds to a respondent’s father being

deceased and zero otherwise. In that case, it is a comparison between paternal orphans and any respondent who has a father living (including maternal orphans). The same structure applies to “Mother is dead. All the dummy control variables are denoted with “+””Significance is denoted: *** p<0.01, ** p<0.05, * p<0.1.

VARIABLES

Ln Height Years of Schooling

Fatherdead -4.95E-05 -0.182 -0.00244 -0.183 Motherdead 0.0012 -0.626*** -0.00327 -0.212 Baseline height 0.00181*** 0.0128 -0.000171 -0.0103 Baseline years of schooling 0.00225*** 0.559*** -0.000804 -0.0669 Age in years -0.00967*** -0.310*** -0.00106 -0.0655 Male+ 0.0613*** 0.741*** -0.0023 -0.16

Household Head Years

of Schooling -0.000693* 0.183***

-0.000417 -0.0285

Household Head male+ -0.00487 -0.603**

-0.0034 -0.259

Household Head age -4.22E-05 0.0138**

-7.85E-05 -0.00569 Ln per capita consumption 0.00688*** 0.375** -0.00209 -0.155 Household Size 0.00132*** 0.0206 -0.000338 -0.0219

Floor building material

good+ 0.00118 2.039*** -0.00392 -0.268 Wave4 control+ -0.0115** -0.381 -0.00469 -0.317 Constant 4.873*** 4.029*** -0.0193 -1.364 Observations 1,324 1,465 R-squared 0.477 0.319 Dependent Variables

28

Note: Both these regressions are estimated with community fixed effects. Robust standard errors are in italics. As described earlier, baseline height and schooling are controlled for in this equation. Both these regressions have controls. These include individual characteristics of the respondent (sex and age), the household socioeconomic atmosphere at baseline (whether the child lived with their father, similarly for their mother and the household head’s educational attainment and sex) and finally baseline indicators of wealth (Log per capita household consumption, whether the domicile flooring was good quality i.e. cement or tiles). There also includes controls for the various baseline waves 1991-1994; as baseline information from some respondents was recorded in different years. The variable “Young & Orphaned” is a dummy variable where 1 corresponds to a respondent who was aged 0-6 years at baseline and lost at least one parent. And zero otherwise. In that case, it is a comparison between younger orphans and any respondent who was older and/or non-orphaned. The same structure applies to “Old & Orphaned” except the age range is 7-15 years at baseline. All the dummy control variables are denoted with “+”. Significance is denoted: *** p<0.01, ** p<0.05, * p<0.1.

VARIABLES

Ln Height Years of Schooling Young & Orphan 0.00372 -0.247

-0.00334 -0.264

Old & Orphan -0.00132 -0.680***

-0.00313 -0.208 Baseline height 0.00181*** 0.0128 -0.000171 -0.0103 Baseline years of schooling 0.00225*** 0.559*** -0.000804 -0.0669 Male+ 0.0613*** 0.741*** -0.0023 -0.16

Household Head Years

of Schooling -0.000693* 0.183***

-0.000417 -0.0285

Household Head male+ -0.00487 -0.603**

-0.0034 -0.259

Household Head age -4.22E-05 0.0138**

-7.85E-05 -0.00569 Ln per capita consumption 0.00688*** 0.375** -0.00209 -0.155 Household Size 0.00132*** 0.0206 -0.000338 -0.0219

Floor building material

good+ 0.00118 2.039*** -0.00392 -0.268 Wave4 control+ -0.0115** -0.381 -0.00469 -0.317 Constant 4.873*** 4.029*** -0.0193 -1.364 Observations 1,324 1,465 R-squared 0.477 0.319 Dependent Variables