The effects of cash transfers on adult and child mortality in low- and middle-income countries – Nature

We performed analyses of changes in adult and child mortality associated with implementation of cash transfer programmes between 2000 to 2019, a study period when many countries introduced cash transfer programmes.

Mortality data

To estimate mortality, we generated two individual-level longitudinal datasets—one for adults aged ≥18 years and one for children aged <18 years—using demographic and health surveys (DHS)^36,37,53. The DHS are conducted in many LMICs about every five years. They use a two-stage cluster sampling design to produce national and sub-national estimates for a variety of indicators that are representative of their target populations⁵⁴. The first stage involves systematic selection of enumeration areas drawn from census files with probability proportional to population size, and the second stage involves a random sampling of households from each enumeration area. Primary respondents were all female household members of reproductive ages (15–49 years). Procedures and questionnaires for DHS have been reviewed and approved by the ICF Institutional Review Board. All analysed data were anonymized. In accordance with standard procedures for secondary data analysis, the University of Pennsylvania Institutional Review Board waived ethical review.

We used surveys that included a maternal mortality module to create the adult dataset. This module collects information from all primary respondents about every sibling born to her biological mother—sex, current vital status, year of death if deceased, current age (or age at death), and for female siblings whether the death was pregnancy-related (death while pregnant or within two months following termination of pregnancy, irrespective of the cause). As there are limited, heterogeneous, and inconsistent data available about other causes of death, we focus on mortality from all causes. Using previously established methodology^36,37,53, we first restructured the dataset such that there was one observation per sibling, and then again such that each observation corresponded to one person-year from one sibling. Each observation included a binary variable indicating the sibling’s survival status during that person-year. We excluded observations from incomplete years (that is, the year of the survey). To minimize recall bias, we excluded observations earlier than ten years before the survey. We excluded person-years during which a sibling was aged <18 years for the purposes of this adult dataset. Of note, because primary respondents in the DHS were 15–49 years of age, older adults were underrepresented.

We created a child dataset from the same set of surveys using the birth history module, which asked female respondents for information about all births—sex, birth date, survival status, and death date. As above, we constructed a longitudinal dataset with observations at the level of the person-year, including an indicator variable for survival and excluding incomplete years and observations earlier than ten years prior to the survey. We excluded person-years during which a child was >17 years old.

We extracted additional data about the primary respondent (sibling in the adult dataset, mother in the child dataset)—age, rural or urban setting, wealth quintile, and schooling attainment (categorized as none, primary, secondary, or greater than secondary). Respondents were classified into wealth quintiles using the DHS Wealth Index, a composite measure of households’ cumulative living standard generated using a principal components analysis based on ownership of certain assets, materials used for housing construction, and types of water access and sanitation facilities⁵⁵.

Cash transfer programme data

We identified all major, government-led cash transfer programmes within included countries using previously established methods²⁶. We manually searched a variety of sources to identify the programmes as well as the years in which they were implemented, the population targeted by the programmes (for example, older adults, families with young children), whether the programmes had behavioural conditionalities, amounts of annual cash transfers, and most recently available number of beneficiaries^{56,57,58,59,60}. Data sources included social protection databases from the World Bank, United Nations, and non-governmental organizations, as well as primary documentation and reporting from individual programmes. We excluded countries with pre-existing cash transfer programmes at the start of the study period.

We calculated the impoverished population coverage for each programme as the most recent estimate of the number of programme beneficiaries divided by the number of individuals in a country with income less than the international poverty line of US$1.90 per day (2011 purchasing power parity). To do this, we divided the most recent estimate of total household beneficiaries by the impoverished population size. If estimates for total beneficiaries were not available, we multiplied direct beneficiaries by the average household size to estimate total beneficiaries⁶¹. In general, the number of beneficiaries was available during only a limited number of years. Impoverished population sizes were calculated by multiplying the percentages of the populations with income less than the international poverty line (that is, the poverty headcount) prior to programme implementation by the mid-year population from the year of the total beneficiaries estimate⁶². We used the poverty headcount prior to programme implementation because poverty headcount estimates after programme implementation may be decreased by the programmes themselves. For example, if a cash transfer programme began in 2012, we divided the most recent estimate of beneficiaries (numerator) by the poverty headcount in 2012 (denominator) to calculate the impoverished population coverage.

We also calculated the maximum transfer amounts as percentages of GDP per capita in the most recent year the maximum transfer amounts were reported.

Additional country-level data

We obtained additional time-varying covariates for each country and year that are known to be or are likely to be associated with changes in cash transfer programmes and mortality: GDP per capita⁶², total health expenditures per capita⁶², life expectancies at birth⁶², PEPFAR funding budgeted⁶³, and six Worldwide Governance Indicators from the World Bank that are composite indicators based on 30 data sources: voice and accountability, political stability and absence of violence, government effectiveness, regulatory quality, rule of law, and control of corruption⁶².

Statistical analysis

We used a difference-in-differences approach, a quasi-experimental technique that can be used to estimate causal effects from observational data by comparing the differences in outcomes between intervention and comparison groups during pre-intervention and post-intervention periods, under an assumption of parallel trends (that is, that in the absence of cash transfer programmes, trends in outcomes would be similar in intervention and comparison countries). To do this, we estimated multivariable modified Poisson regression models with the unit of observation being the person-year and a binary outcome variable indicating whether an individual died in a given year⁶⁴.

Our primary explanatory variable was a binary variable set to 1 if a cash transfer programme (or combination of programmes) with total impoverished population coverage greater than 5% was active in the respondent’s country during that year. We were prevented from considering coverage as a continuous, time-varying exposure because beneficiary data were available only during a limited number of years for most programmes. We chose 5% based on our prior analyses showing this threshold was associated with improvements in HIV-related outcomes²⁶, but conducted subgroup analyses (described below) to explore the association with different levels of coverage. We excluded intervention countries that lacked at least two years of mortality data prior to the cash transfer period.

To optimize our comparison country-years, we excluded from our analysis country-years during which cash transfer programmes (or combination of programmes) were implemented with coverage between 2% and 5%. Comparison country-years were therefore defined as those during which there were no active cash transfer programmes, or cash transfer programmes (or combination of programmes) had coverage <2%.

Our effect measure of interest was the risk ratio denoting the association between the cash transfer programme exposure and mortality. In addition to overall estimates, we also evaluated the temporal relationship between cash transfer programmes and mortality by creating a series of binary indicators for each year before and after the cash transfer period began.

We included in the models country- and individual-level covariates that were likely to confound relationships between cash transfer programmes and mortality. For country-level covariates, we included GDP per capita, PEPFAR funding budgeted, and three Worldwide Governance Indicators: control of corruption, political stability and absence of violence, and voice and accountability. The other three Worldwide Governance Indicators were left out of the models because they displayed substantial multicollinearity with the other covariates as evidenced by variance inflation factors >5. We also considered inclusion of health expenditures per capita, but this variable was not available for all years and adding it to the models minimally impacted the effect estimates.

For individual-level covariates, we included age and rural or urban setting in all analyses. In the child analyses, we also included sex, mother’s age, and birth order. We did not include other individual-level variables that were likely to be affected by receipt of cash transfers and/or potentially mediate relationships between cash transfer programmes and mortality (for example, wealth quintile, schooling attainment).

We included country fixed effects to control for time-invariant differences among countries, and year fixed effects to control for secular trends in mortality. We used robust standard errors clustered at the country level to relax the assumption of independently and identically distributed error terms^65,66.

We stratified the adult mortality analysis by sex because of previously identified sex-specific effects of cash transfers^26,43,44,56, and the child mortality analysis by age (<5 years, 5–9 years, 10–17 years) because of highly varying mortality rates by child age⁶⁷.

We explored heterogeneity of the effect of cash transfer programmes using subgroup analyses based on the beneficiary, cash transfer programme design, and country factors. For the beneficiary, we considered wealth quintile (of the sibling for the adult analysis and the mother for the child analysis), age (for the adult analysis, categorized as 18–40, 41–60, and >60 years), educational attainment (of the sibling for the adult analysis, and the mother for the child analysis), rural or urban setting, and cause of death among women (pregnancy-related versus not pregnancy-related). For cash transfer design features, we considered conditionality (unconditional, mixed, or conditional), and four subgroups characterized by most recent impoverished population coverage above or below the median (30%) and maximum annual transfer above or below the median (11% of GDP per capita). For country factors, we considered subgroups characterized by being above or below the median at the start of the cash transfer period for the following: each of the Worldwide Governance Indicators, current annual healthcare expenditures per capita (US$118 purchasing power parity), and life expectancies at birth (62 years). We also stratified by region (sub-Saharan Africa versus outside of sub-Saharan Africa). Finally, we generated country-specific estimates for adult women to allow for informal evaluations of heterogeneity across a range of dimensions.

We also conducted additional sensitivity analyses. First, we assessed the validity of the parallel trends assumption in two ways. We used the previously described temporal analysis to visualize pre-trends, and we estimated regression models using only data prior to the cash transfer period in each country and including an interaction term between an indicator of whether the country was in the intervention group and a linear time trend.

Second, while we used modified Poisson regression models based on conceptual justifications and to be consistent with prior literature assessing changes in mortality using DHS datasets^36,37,53, we assessed for robustness of the results when using logistic and linear models.

Third, recent advances in difference-in-differences analyses with variation in intervention timing have shown that estimates may be biased particularly if there is heterogeneity in intervention effects over time^38,39,63. When there is effect heterogeneity only in time since the intervention, this concern can be mitigated through use of temporal analysis with dynamic effect estimates (as described above), although there can still be bias present if there are heterogeneous treatment effects over overall calendar time⁶⁸. To address this, we assessed whether a proposed alternative linear estimator not vulnerable to this bias was consistent with our primary findings⁴¹. In addition, this bias tends to be influenced by later country-years during the intervention period, so to assess the possible magnitude of this bias we conducted a sensitivity analysis by repeating the primary analysis after excluding country-years after year 5 of the cash transfer programme⁶⁹.

Fourth, we assessed whether individual countries might be outliers for key outcomes by assessing whether estimates for women changed substantially after excluding each country individually.

Fifth, we repeated our primary analyses with inclusion of the respondent’s wealth quintile and educational attainment.

We did not use statistical methods to predetermine sample size. We performed statistical analyses using SAS V.9.4, R V.3.5.2 using the ggplot2 and forester packages, and STATA V.17 using the did2s package.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.