Is the impact of paid maternity leave policy on the prevalence of childhood diarrhoea mediated by breastfeeding duration? A causal mediation analysis using quasi-experimental evidence from 38 low-income and middle-income countries



There is a growing interest in supporting evidence-based policy-making by estimating the impact of public policies on population health, with increasing experimental and quasi-experimental evidence. The difference-in-difference method has emerged as a powerful quasi-experimental method to address specific sources of confounding in observational studies and evaluate the impact of population-level interventions, as it relies on a less strict exchangeability assumption. By comparing the trend in an outcome in a treated group before and after a policy change to that in an unaffected control group, the difference-in-difference method assumes that the control group mimics the behaviour of the treated group had it not been treated (the ‘parallel trends’ assumption), thus serving as a counterfactual.

Few studies have examined the mediating pathways linking policies to population health. Understanding the causal mechanisms through which policy reforms impact population health outcomes not only supports the plausibility of the association but may also facilitate theory building and inform the choice of alternative or complementary interventions. Causal mechanisms can be identified by specifying mediators on the causal pathway between policy and outcomes.1 One approach involves the application of causal mediation analysis, which has been employed to test pathways within the evaluation of health interventions.2 3

The controlled direct effect (CDE), commonly the interest in policy settings, fixing the intermediate to level M=m, CDE(m), captures the effect of the exposure on the outcome if the intermediate was set, possibly contrary to fact, to level m.3 This quantity is measured based on the assumption of no two assumptions: no unmeasured confounding between exposure and outcome and no unmeasured confounding between mediator and outcome.3 The assumptions can be examined by assessing the tenability of the parallel trends of the mediator and the outcome. The violation of the parallel trends assumption in either case would suggest that there are other factors besides the intervention that are affecting the mediator or outcome.

Conceptual framework

We illustrate this approach using the example of maternity leave policy and the prevalence of severe diarrhoea, with breastfeeding duration as the mediator.

Globally, diarrhoea is responsible for the unnecessary loss of 0.5 million young lives each year and is a threat to sustainable development for the poorest nations.4 Diarrhoea ranks the fourth-leading cause of death and second-leading infectious cause of death in children under 5 years of age.5 6 Diarrhoea mortality is disproportionately concentrated in low-resource settings. Low-income and middle-income countries (LMICs) have 62% of the world’s under-5 population, but account for more than 90% of the global diarrhoea deaths.4

There are proven interventions that can go a long way toward reducing diarrhoea incidence and mortality.7–14 Childhood diarrhoea can be effectively prevented by exclusive breast feeding for the first 6 months of life,15 continued breast feeding,9 15 vitamin A supplementation,16 immunisation,17 18 safe drinking water, sanitation and hygiene.19–21 Nonetheless, there continue to be a large number of deaths every year. And, due to the lack of policy data on a sufficiently large number of nations to allow comparative analyses, the evidence for informing specific national policy strategies to lower childhood diarrhoea morbidity and mortality remains limited.7 22 23

There is a growing body of evidence evaluating the impact of paid maternity leave in LMICs. Paid maternity leave, defined as the ‘leave that the country guarantees employed women in connection with the birth of a child,’24 enables women to take time off from work following childbirth while maintaining a partial income. In 2015, 185 out of 193 United Nations (UN) member states had a national maternity leave policy in place, but only 57% (n=105 countries) met the 14-week minimum standard25 set by the International Labour Organization (ILO) in convention C183. The empirical evidence shows that increasing the generosity of paid maternity leave policies is associated with lower infant mortality,26 improved breastfeeding practices27 and increased vaccination uptake in LMICs.28 A recent study showed a 1-month increase in the legislated duration of paid maternity leave was associated with 61 fewer cases of bloody diarrhoea (95% CI –98.86 to –22.86) per 10 000 children under 5 years of age in LMICs.29 However, existing studies have not explored the pathways through which paid maternity leave might affect child health. In this study, we build on the existing literature by applying causal mediation methods to examine the pathways linking increases in the duration of paid maternity leave to decreases in childhood diarrhoea prevalence.

We hypothesise that longer paid maternity leave policies have the potential to lower the prevalence of childhood diarrhoea in LMICs through several possible mechanisms (figure 1). First, paid maternity leave may affect the caregiving environment for newborns and infants. Mothers with access to paid maternity leave may be able to dedicate more time to the care of their children in the first few weeks of their lives, which might protect children from early infection in group care settings, such as daycare,30–33 or at workplaces, in contexts where daycare facilities are lacking. Second, access to paid maternity leave may facilitate preventive care. Mothers who can take leave from work are more likely to initiate and continue breast feeding27 34 and to have their children vaccinated on time,28 both of which boost immunity and protect children from infection. Third, access to paid maternity leave may facilitate access to postnatal health services, which could prevent infection and progression to more severe disease. Fourth, access to paid maternity leave provides psychological benefits to women who used to work full time before childbirth.35–41 Longer maternity leave is associated with less depressive symptoms, lower likelihood of severe depression and better overall maternal health.40 41 Children are more likely to receive good care and stay healthy in a stress-free home environment. Additionally, healthy mothers are more likely to facilitate preventive care. Finally, paid maternity leave may increase income, and thus, support both prevention and early treatment.42

Figure 1
Figure 1

Possible mechanisms that paid maternity leave policy impact the prevalence of childhood diarrhoea.

The protective effect of breast feeding against mortality and morbidity from childhood diarrhoea has been widely studied.43 A systematic review on the benefits of breast feeding showed about half of all diarrhoea episodes would be avoided by breast feeding.43 The protective effect of breast feeding is attributable to the presence of substances unique to human milk that have anti-infective, anti-inflammatory and immunoregulatory functions. Human milk oligosaccharides, unconjugated complex carbohydrates that are highly abundant in human milk, have been suggested to prevent the development of gastrointestinal infections by blocking the attachment of pathogens to the infant’s mucosa.44–49 Lactoferrin, one of the major multifunctional agents in human milk, appears to have a major role in blocking bacterial virulence by disruption of the integrity of the bacterial outer membrane.50 Many different specific secretory antibodies, produced by mothers who have been exposed to such pathogens, are found in human milk and in varying quantities, protecting the infant from developing an infection.50 51 In addition, children who die from diarrhoea often suffer from underlying malnutrition, which makes them more vulnerable to diarrhoea infection. Optimal breastfeeding practices can prevent undernutrition associated with repeated infections in low-income settings.52 Furthermore, breast milk is a safe food source for infants. Infants who receive only breast milk for the first 6 months have less exposure to pathogens than non-exclusively breastfed infants, as foods offered to infants are often contaminated with microbial pathogens that could lead to gastrointestinal infection.

For the assessment of the short-term consequences of breast feeding, the comparison between ever-breastfed subjects and those never breastfed would tend to underestimate any association because there is usually a cumulative effect of breast feeding, rather than a critical-window effect.43 In this study, we estimated whether the protective effect of increasing the duration of paid maternity leave on childhood diarrhoea was mediated by the duration of breast feeding (part of red path in figure 1). First, we exploited legislated increases in the duration of paid maternity leave in LMICs to estimate impacts on the prevalence of childhood diarrhoea using a difference-in-difference design. Second, we applied causal mediation methods to estimate the extent to which breastfeeding duration explained this total effect.


Data sources

Longitudinal data on national maternity leave policies for each UN member state were made available by the University of California Los Angeles’ WORLD Policy Analysis Center and collected retroactively to 1995 by McGill University’s Policy-Relevant Observational Studies for Population health Equity and Responsible Development project.53 54 Further details on the collection and coding of global maternity leave policies are available elsewhere.24

Demographic and Health Surveys (DHSs)55 were used to obtain individual-level information on childhood diarrhoea, breastfeeding duration and other covariates for children under 5 years of age. DHS is conducted approximately every 5 years in many LMICs, using a nationally representative two-stage cluster sampling design, with the first stage selecting clusters and the second selecting households. In each household, trained interviewers and structured questionnaires were used to obtain socioeconomic, demographic and health information from women aged 15–49 and anthropometric information for children younger than 5. Standardised measurement techniques were used to ensure the comparability of surveys across countries and survey waves. Further details regarding the sampling and survey techniques are available elsewhere.56 57


For the analysis, we linked the national maternity leave policies between 1995 and 2013 (inclusive) to information on 692 452 children from 38 LMICs who were under 5 years of age at the time of interview. The five treated countries (1) experienced at least one change in the duration of paid maternity leave policy between 1995 and 2015 and (2) had at least one survey before and after the policy change. We also included 33 control countries with data available from at least two DHS between 1995 and 2015 (inclusive). These inclusion criteria allowed us to compare changes in the duration of breast feeding and the prevalence of bloody diarrhoea over time for treated and control countries that did or did not reform their paid maternity leave policies, respectively. Our sample included 3 countries in East Asia, 2 in South Asia, 2 in the Middle East, 7 in Latin America and 24 in sub-Saharan Africa.

A 1-year lag was used to respect the temporality between policy year and children’s birth year, leaving 639 153 children born between 1996 and 2014 (inclusive) from 126 DHS in 38 LMICs in the final sample (table 1). For the analyses including the mediators breast feeding for at least 6 months or at least 12 months after birth, we restricted the sample to children older than 6 months and 12 months, respectively.

Table 1

Sample description



The exposure of interest in our study was the legislated length of paid maternity leave for each sampled country between 1995 and 2013 (inclusive). We recorded the legislated length of paid leave available to mothers only. To ensure temporality between exposure and outcome, and to reduce exposure misclassification, each observation was assigned the legislated length of paid maternity leave 1 year prior to the birth year. We did not distinguish between leave that could be taken before or after birth.


Our primary outcome variable was whether the child had bloody stools in the 2 weeks prior to the interview. This measure was used as an indicator of severe diarrhoea because the frequency of loose stools in breastfed infants can be difficult to distinguish from pathological diarrhoea based on survey data. Information on bloody diarrhoea (clearly pathologic) was extracted from at least two DHS for each sampled country. Briefly, mothers surveyed in the DHS were asked to provide information on the prevalence of diarrhoea for all children in the household under the age of 5. If a child had diarrhoea in the previous 2 weeks, the presence of blood in the stools was also recorded.


Our three mediators were breastfeeding duration in months, and whether the child was breastfed for at least 6 months or at least 12 months after birth. At the time of interview, mothers provided breastfeeding information for all children in the household under the age of 5. If a child was ever breastfed after birth, the interviewer recorded the number of months of breast feeding. If the child was still being breastfed at the time of interview, the interval between the child’s date of birth and date of interview was used as the breastfeeding duration.


Based on several empirical studies on diarrhoea in children,6 7 58–61 we identified potential confounders and other household, maternal and child-level determinants of childhood diarrhoea and breastfeeding duration in LMICs. Covariates at the household level included the number of listed household members, number of children under 5 years of age living in the household, place of residence (eg, urban or rural) and an indicator for drinking water source (eg, unimproved or improved), based on the new scale for household drinking water used by the WHO and the UNICEF.62 We also included the DHS household wealth index, which was created and used as a standard by the DHS and UNICEF Multiple Indicator Cluster Surveys to capture the within-country relative wealth standing of each household. At the maternal and child levels, covariates included mother’s education in years, mother’s age at delivery, mother’s number of living children, mother’s working status at the time of the interview, number of antenatal visits during pregnancy, place of delivery, child’s sex, child’s age at interview and child’s birth order. Information on all variables was obtained from the DHS.

In addition, we included country-level indicators from the World Bank’s World Development Indicators and Global Development Finance databases that may be associated with paid maternity leave policy reforms and also affect breastfeeding duration or childhood diarrhoea prevalence.63 These variables included the gross domestic product (GDP) per capita (constant 2011 international dollar) based on purchasing power parity (PPP), female labour force participation rate (percentage of female population ages 15–64), percentage of unemployed female labour force, government health expenditures per capita based on PPP (constant 2011 international dollar) and total health expenditure (percentage of GDP).

Mediation analysis

Our inferential goal was to estimate the CDE of increasing the duration of paid maternity leave on the risk of childhood diarrhoea had we intervened to fix the duration of breast feeding in treated and control countries. Identification of the CDE assumes no unmeasured confounding of (1) the total effect of paid maternity leave on the risk of diarrhoea and (2) the relation between the mediator, breastfeeding duration and the risk of diarrhoea.2 3

We examined the mediating role of breastfeeding duration using a three-step process.

Step 3: calculate proportion eliminated

The ‘proportion eliminated’ measures what proportion of the total effect of the exposure on the outcome would be eliminated had we intervened to set the mediator to a fixed value, M=m, for the population.67 We calculated the proportion eliminated on the excess RR scale after taking the inverse of the reported RRs for the total and CDEs, which were protective (RR<1):

Embedded Image

Patient and public involvement

Our study does not involve the participation of patients or any members of the public. All data used in this study are aggregated and publicly available DHS anthropometric data and can be accessed through the DHS programme online repository.


Descriptive statistics

In the study sample, information on breastfeeding duration has the most missing values. Characteristics of observations with and without breastfeeding information are presented in online supplemental table 1. Observations without breastfeeding information were older, as breastfeeding information was based on maternal recall. The distributions of the other covariates among observations without breastfeeding information were similar to those among the observations with breastfeeding information.

Supplemental material

The weighted proportion of children under 5 with bloody diarrhoea in the 2 weeks prior to the interview was 1.81% (SD=0.65) in treated countries that had changed their policy, as compared with 1.40% (SD=0.17) in control countries that had not. On average, 98.65% (SD=0.32) and 95.61% (SD=0.97) of children under 5 were breastfed for at least 6 and at least 12 months, respectively, in treated countries, compared with 94.25% (SD=1.76) and 86.88% (SD=3.04) in control countries.

On average, each household had six listed members and two children under 5 years of age. More than 65% of households had improved drinking water sources, and 62% were located in rural areas. The mean age of mothers in the sample was 27 years, with three children on average. More than 52% of mother were currently working at the time of the interview. More than 47% of mothers had four times or more antenatal visits during pregnancy and about 53% of delivery happened in an institution (table 2).

Table 2

Demographics and characteristics of the independent variables in the study sample, 1996–2014, N=639 153

Between 1995 and 2013, the average weeks of paid maternity leave among the 33 countries that did not change the duration of leave was 12.8 (online supplemental figure 1). Among the five countries that changed the duration of leave available (ie, Bangladesh, Kenya, Uganda, Zambia, Zimbabwe), paid maternity leave increased on average from 10.0 weeks in 1995 to 14.4 weeks in 2013 (figure 2).

Figure 2
Figure 2

Legislated paid maternity leave policy in weeks in sampled countries that changes policies.

Mediation analysis

A 1-month increase in the legislated duration of paid maternity leave was associated with a 34% (RR 0.66, 95% CI 0.47 to 0.91) reduction in the risk of bloody diarrhoea (table 3). The CDEs are 0.68 (95% CI 0.64 to 0.73), 0.67 (95% CI 0.62 to 0.73) and 0.67 (95% CI 0.56 to 0.81) when setting breastfeeding duration to at least 6 months, at least 12 months and the same duration, respectively. Breast feeding for at least 6 months and 12 months mediated 10.56% and 7.43% of this effect, respectively, whereas breastfeeding duration mediated 6.74% of the total effect.

Table 3

Effect of a 1-month increase in legislated length of paid maternity leave and potential mediators on the risks of bloody diarrhoea for children under 5


Causal mediation analysis, considered as an approach to understand casual mechanisms, is becoming increasingly popular in many disciplines of the social and medical sciences, including epidemiology, psychology and political science.1 However, to date, there is limited empirical evidence of its application for the evaluation of health and social policy. Building on an existing study, we tested the causal pathways whereby paid maternity leave policy affected severe childhood diarrhoea in LMICs. Our results show that breast feeding for at least 6 months and at least 12 months mediated 10.56% and 7.43% of the effect of a 1-month increase in the legislated duration of paid maternity leave on childhood diarrhoea in LMICs. This is the first study, to our knowledge, to examine the pathways through which paid maternity leave policy affects child health.

In our study, despite evidence of strong exposure–mediator and mediator–outcome associations in the literature, only a small proportion of the impact of paid maternity leave policy on childhood diarrhoea is shown to be mediated through breastfeeding duration. This could be a result of later-life exposures mitigating the benefits of breast feeding during childhood.

A large proportion of the impact of paid maternity leave policy on childhood diarrhoea is not mediated through breastfeeding duration. In addition, longer breastfeeding duration was not strongly associated with lower incidence of diarrhoea in our sample, contrary to other reports. This implies other explanations. Lack of access to paid maternity leave may force mothers back to work earlier, which might result in infants being left home alone, with other family. In many low-resource settings, parents who have to work to provide for the family’s basic needs of food, clothing and shelter, have no choice but to leave their preschool children home alone or in the care of their older siblings.33 The older siblings, most of the times young children themselves, are not able to provide adequate care for children younger than 5. In these situations, young children left home are more likely to become malnourished due to irregular feeding and have a higher risk of being exposed to pathogens in raw or rotten food.33 In addition, lacking an adult at home lowers the chance of proper medical care when the child is sick. For example, oral rehydration therapy, a simply made rehydration solution consisting of water, sugar and salt, can keep children from dehydration due to diarrhoea. But the solution has to be administered by an adult who can give the child one sip at a time for hours. Not taking proper medication increases the chances that children suffer repeated episodes of diarrhoeal disease and consequent malnutrition.33

Several limitations should be noted. First, the difference-in-difference approach relies on the assumption of parallel trends between intervention and comparison groups, in this case in relation to outcomes as well as potential mediators. This assumption is difficult to check visually in the generalised fixed-effects difference-in-difference design with a continuous treatment, as countries experienced policy changes at different time points.68 We examined the parallel trends assumption but lacked longitudinal preintervention data on our outcome and mediators for all sampled countries, as some countries only had one DHS available during the preintervention time. However, the trends in the natural logarithm of the outcome and mediators for treated and control countries appeared parallel in the preintervention period before 2000, suggesting that the assumption was not violated. Second, although we specified a comprehensive set of covariates at the individual, household and country levels, residual confounding by time-varying exposure–outcome and mediator–outcome factors is still possible, which may have biased the total and CDEs. For example, exposure to the Baby-Friendly Hospital Initiative may be an unmeasured mediator–outcome confounder. Women who gave birth in a Baby-Friendly Hospital may have more guidance and support on practicing breast feeding and having access to prenatal and postnatal care might also have affected the risk of diarrhoea for newborns. Moreover, we did not find reliable data to account for country-specific time trend for all the countries included in the study. Third, we did not account for changes to other health policies that may have coincided with changes in paid maternity leave policy. For example, reforms making healthcare coverage more universal, which may have promoted access to preventive interventions, could also affect the risk of childhood diarrhoea. Fourth, information related to our outcome, mediators and a few covariates was collected based on maternal recall. Recall bias may be less of a concern for our outcome measure because mothers were asked to recall a significant event that had happened in the previous 2 weeks. Breastfeeding measurements in mothers with older children are more prone to recall bias, although this is likely non-differential by treatment status. Fifth, we did not account for other types of leave (eg, parental leave) that might be available to mothers, but misclassification of our exposure is unlikely because paid parental leave is relatively short (eg, less than 4 weeks) among the sampled countries. Sixth, the possibility of survivor bias may not be completely ruled out because information on outcome, mediators and several covariates was only collected on children who were alive at the time of the interview. Seventh, due to the lack of information on policy compliance and implementation, the intention-to-treat estimate obtained in our study may be downwardly biased. Furthermore, ILO reported that 92% of employed women in low-income countries and 85% of employed women in lower-income to middle-income countries were informally employed in 2016.69 An average population effect, as we estimated in this study, is likely to underestimate the true policy effect since women in an informal economy may not be protected by paid maternity leave policy. Finally, generalisation of our results may be limited to countries with sociodemographic profiles similar to those of the 40 sampled countries.

In conclusion, few quasi-experimental evaluations have applied causal mediation methods to examine the pathways linking public policies to population health. Demonstrating the use of causal mediation methods within difference-in-difference analysis, we illustrate that mediation analysis can be useful for examining how the policy might impact health outcomes by changing intervening behaviours. Additional research is needed to assess other pathways linking paid maternity leave policy to childhood diarrhoea and other child health outcomes. Further studies should also examine population-level mediators related to behaviour changes, rather than individual-level mediators. From a policy perspective, these studies could help develop effective early-life interventions to ensure positive health outcomes for mothers and infants in LMICs.

This post was originally published on