The orphan impact: HIV-AIDS and student test scores from sub-Saharan Africa

ABSTRACT

In sub-Saharan Africa over 52 million children are living with the death of one or both parents. Drivers of this parental mortality include afflictions at levels endemic to the region, including: HIV; malaria and other parasites; lower respiratory infections; diarrhoeal illnesses; and road accidents, among others. This paper examines the impact of orphanhood on learning outcomes among girls and boys in sub-Saharan Africa, conditional on school enrolment. By analysing test scores for approximately 60,000 pupils in 12 countries, we estimate the effect on student test scores by comparing paternal, maternal, and double orphans to non-orphans in the sample, specifically for the subjects of reading, mathematics, and HIV-AIDS knowledge. No previous study has analysed how orphanhood might influence learning by using student test score data, making this paper’s approach unique in the literature. This study employs two estimation techniques: Coarsened exact matching calculates the sample average treatment effect on the treated, while matching on students’ family structure, household wealth, school resources, and geographic location; and double lasso (DL) regression applies applying machine-learning for variable selection with high-dimensional controls for regional and school identifiers, school location, and student age. Our results show both CEM and DL consistently report a significant negative impact of orphanhood on test scores among specific countries, especially those which faltered in addressing the HIV-AIDS crisis.

KEYWORDS:

• Orphan
• Africa
• SACMEQ
• double-lasso regression
• coarsened exact matching
• education

Acknowledgments

We acknowledge the useful suggestions by Professor Yuichiro Yoshida in strengthening the methodology of this paper. The author Benjamin K. Blevins acknowledges the support of the Hiroshima University TAOYAKA Program for creating a flexible, enduring, peaceful society, funded by the Program for Leading Graduate Schools, Ministry of Education, Culture, Sports, Science and Technology, of the Government of Japan.

Author contributions

The first author analyzed the data and wrote the paper. The second author assisted in the selection of the methodology and provided comments on the text.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. This figure represents a subset of the 52 million orphans living in the whole of sub-Saharan Africa.

2. Examining the rates of paternal versus maternal mortality, the SACMEQ data contains on average 2.5 times the paternal mortality compared to maternal, with a low ratio of 1.8 (Namibia) and a high of 3 times (Kenya). This is consistent with other scholars’ estimations using Demographic and Health Survey (DHS) data (1995–2000) on a slightly younger cohort (0–15 years old), closely matching the ratio of roughly double the rate of paternal to maternal mortality in seven of the fourteen SACMEQ countries (Ainsworth & Filmer, 2006; Bicego, Rutstein, & Johnson, 2003). These high rates of paternal mortality are not reflected in national health statistics of the World Bank, which shows an average rate of male mortality at only 1.1 times that of female mortality (World Bank, 2012). However, the World Bank data address adult mortality (between 15 and 60 years of age) and therefore include adults who are no longer rearing their own children, and as such would not be reflected in the SACMEQ data. This suggests that paternal mortality is high when children are young enough to be enrolled in primary school, with maternal mortality “catching up” later in the child’s life.

3. The study reviewed nearly 20 cross-sectional demographic and health survey (DHS) data from ten countries in sub-Saharan Africa, examining the enrolment rates of orphans adjusting for proximity of relatedness among their primary guardians.

4. Malawi, Zambia, Namibia, Lesotho, and Uganda each displayed significant gaps in enrolment between the first and fifth wealth quintile.

5. Cross-sectional data were used from over 100 household surveys representing over 50 countries spanning from 1992 to 2003, primarily from sub-Saharan Africa.

6. The study accessed a panel dataset from rural Kenya covering approximately 20,000 children over a five-year period (1998–2002).

7. Panel data were used from Tanzania consisting of two surveys (the first where 718 non-orphans were interviewed between 1991–1994, and the second survey where the same children, as adults, were interviewed in 2004).

8. The geographic scope of the studies are as follows: Ardington & Leibbrandt, 2010 (South Africa); Beegle et al., 2009 (Tanzania); Evans & Miguel, 2007 (Kenya); Ha et al., 2015 (Zimbabwe); Lloyd & Blanc, 2006 (Cameroon, Kenya, Malawi, Namibia, Niger, Tanzania, and Zambia); Nyambedha & Aagaard‐Hansen, 2010 (Kenya).

9. The ministries of Botswana, Kenya, Lesotho, Malawi, Mauritius, Mozambique, Namibia, Seychelles, South Africa, Swaziland (Eswatini), Tanzania (Mainland), Tanzania (Zanzibar), Uganda, Zambia, and Zimbabwe.

10. SACMEQ I (1995), included seven Ministries of Education (MoEs). The assessment captured reading test scores from approximately 20,000 learners; 3,000 teachers and 1,000 school principals. SACMEQ II (2000), included fourteen MoEs; assessing reading and mathematics for 40,000 learners; 4,000 teachers; and 2,000 school principals.

11. The reading test includes 55 questions, testing pupils’ ability to read narrative prose, expository prose, and documents, ranging in difficulty from Level 1 (pre-reading) to Level 8 (critical reading). Mathematics includes 49 questions on numbers, measurements, and spatial data, ranging from Level 1 (pre-numeracy) to Level 8 (abstract problem solving). HIV-AIDS awareness includes 86 test questions including diagnosis/treatment, and myths and misconceptions.

12. The first stage identified schools in the target frame, with larger schools given a higher probability of inclusion. At this stage, pupil weights were applied to the inverse of the selection probability for each student. The second stage selected 25 students from the 6th grade classroom using computerised randomisation (Murimba, 2005).

13. In 2018 Swaziland was renamed Eswatini, however this study will continue with Swaziland for consistency with the SACMEQ III dataset.

14. For example, an OLS regression of reading scores compared to the county-level mortality rate indicates a ten-point reduction (0.9 standard error) in test scores for every 1,000 deaths attributed to HIV-AIDS.

15. The observed characteristics included are: biological sex (male/female); age (10- to 27-years-old, measured in three-year increments); homework given (none/1–2 month/1–2 week/most days); repeated grades (never/once/twice/three or more); pupil home quality (lower-/middle-/upper-tercile); extra tuition (no/yes); speak English at home (never/sometimes/most of the time/all the time); dropout of school (never/sometimes/often); parents’ mean education (2–12 years, below and above six years); living with family (no/yes); the school location (isolated or rural/small town/city); and the school building condition (poor/good).

16. In contrast to other matching methods, such as Mahalanobis distance matching (MDM) and propensity score matching (PSM). For both MDM and PSM, the balancing of covariates occurs after the pruning of unmatched sample units, discarding valuable sample information. Compared to MDM- and PSM-type estimators (Iacus et al., 2012; King & Nielsen, 2016; King, Nielsen, Coberley, & Pope, 2011).

17. Or LASSO: least absolute shrinkage and selection operator. The single lasso extends the standard linear regression by introducing a penalty term for variables that are uncorrelated with the outcome variable, minimising their influence. Using the single lasso alone can introduce bias by the under-estimation of non-zero coefficients.

18. The double lasso requires continuous predictors to be normalised as = (X–mean(X))/S.D.(X), with X equal to the student’s age in months.

19. Some of these coefficients are statistically significant only at the lower threshold of 90%, as seen by error bars that intersect the zero line.

20. For double orphans, 96.3%; single orphans, 93.9%; and vulnerable children, 90.6%.

21. Significant at 90%.