Inequalities in early childhood care and development in low/middle-income countries: 2010–2018

Background Inequalities in early childhood development (ECD) tend to persist into adulthood and amplify across the life course. To date, little research on inequalities in early childhood care and development in low/middle-income countries has been available to guide governments, donors and civil society in identifying which young children and families should be targeted by policies and programmes to improve nurturing care that could prevent them from being left behind. Methods Using data from 135 Demographic and Health Surveys and Multiple Indicator Cluster Surveys between 2010 and 2018, we assessed levels and trends of inequalities in exposure to risks of stunting or extreme poverty (under age 5; levels in 85 and trends in 40 countries), early attendance of early care and education programmes (36–59 months; 65 and 17 countries), home stimulation (36–59 months; 62 and 14 countries) and child development according to the Early Childhood Development Index (36–59 months; 60 and 13 countries). Inequalities within countries were measured as the absolute gap in three domains—child gender, household wealth and residential area—and compared across regions and country income groups. Results 63% of children were not exposed to stunting or extreme poverty; 39% of 3–4-year olds attended early care and education; and 69% received a level of reported home stimulation defined as adequate. Sub-Saharan Africa had the lowest proportion of children not exposed to stunting or extreme poverty (45%), attending early care and education (24%) and receiving adequate home stimulation (47%). Substantial gaps in all indicators were found across country income groups, residential areas and household wealth categories. There were no significant reductions in gaps over time for a subset of countries with available data in two survey rounds. Conclusions Available data indicate large inequalities in early experiences and outcomes. Efforts of reducing these inequalities must focus on the poorest families and those living in rural areas in the poorest countries. Improving and applying population-level measurements on ECD in more countries over time are important for ensuring equal opportunities for young children globally.


Chapter 2 Constructing micro-level variable on extreme poverty status in DHS and MICS
We followed the method used in a previous study on children exposed to stunting or extreme poverty. 1 All the Demographic & Health Surveys 2 and Multiple Indicator Cluster Surveys 3 data included variables indicating a household wealth index scores. We estimated the % of under-five children living in extreme poverty ($1.9 per day) with the following steps.
Firstly, we obtained a country's poverty rate at $1.9 per day (2011 PPP) from the World Bank. 4 There were 121 countries with estimates available in some years between 2000 and 2018. Among the 85 countries under this study, the World Bank has poverty estimates available for 34 countries.
Secondly, we followed previous studies 1 and imputed poverty rates for countries and years without available data using the following methods: average poverty rates by income group, regression with generalized linear model (GLM) for prediction, linear extrapolation, or replacing the missing with the value in the closest year. We compared these methods in terms of their rootmean-square errors (RMSE) using existing estimates between 2000 and 2018. The RMSE measures the difference between predicted values and the observed values and is calculated as the square root of the mean of the squares of the deviations between predicted values and observed values. The smaller the RMSE is, the closer is the imputed value to the observed value, and the more accurate of the imputed value is. 5 To calculate the RMSE, we first randomly selected 20% of the existing poverty estimates in 121 countries between 2000 and 2018 and assigned them with missing values. We then predicted the values for those 20% using the 80% of existing estimates and the four imputation methods. With the average income method, we obtained the average poverty ratio of the 80% data by income group and then assigned them to the 20% data according to a country's corresponding income group. With the closest-year method, we assigned the estimate in the year that was closest to the year with missing data. With the GLM method, we used a generalized linear model with a logit link and binomial distribution. This model is a commonly used with dependent variable as proportion data (values between zero and one). 6 Following previous studies, 1 we generated predicted values for the 20% missing values by regressing poverty estimates on GDP per capita. With the linear extrapolation or interpolation method, we first checked the assumption of approximately linear trends of poverty estimates by graphing the trends of poverty ratios for countries with 3 data points or more between 2000 and 2018. We found that most of the countries in most years followed approximately linear trends (Figure 1). This method has been used to impute missing values for poverty ratios in previous studies 7,8 and the limitation of this assumption is that we had to assume no change in distribution of poverty.
We obtained the RMSE for each method as shown in Table 1. Among the four methods, linear interpolation/extrapolation has the smallest RMSE, followed by the closest-year, GLM regression, and average income. We calculated the correlation between actual values and the predicted values derived from the four methods: the correlation between the actual values and the predicted values from the linear extrapolation is the highest (0.98), followed by predicted values derived from the closest-year method (0.97), average income (0.64), and GLM regression (0.64). Based on these findings, we decided that (1) for countries with two or more data points, we replaced the missing values with the predicted values derived from the linear extrapolation method. For those years with predicted values equal to or greater than 1 or less than zero, we replaced them with the existing poverty estimates in the closest year; (2) for countries with only one data point between 2000 and 2018, we replaced the missing values with the available poverty estimates; (3) for countries without poverty estimates between 2000 and 2018, we replaced the missing data with the predicted values derived from the GLM.
Thirdly, we derived % of population living in extreme poverty by ranking the wealth index score of the sampled household members with household weights. We applied a country's poverty rate to the ranked wealth scores and located the cutoff point of extreme poverty in the ranked wealth index. For example, according to the World Bank, the 2016 extreme poverty headcount ratio in Uganda is 42%. We used the Uganda DHS 2016 and ranked its household members' wealth index score with household weights. We then located the 42 th percentile of the ranked scores as the cutoff point for the extreme poverty. All under-five children with wealth score below the 42 th percentile were considered to be living in extreme poverty.

References
1. Lu C, Black MM, Richter LM. Risk of poor development in young children in lowincome and middle-income countries: an estimation and analysis at the global, regional, and country level.

Chapter 3 Measuring the level and trends of inequalities in four indicators (1) Calculating aggregate-level mean values and their 95% confidence intervals (95% CI) of the inequalities in the most recent years since 2010
Following previous studies (1,2), the mean values as well as 95% CI at the aggregate level were computed by averaging country-level estimates. Taking inequality of prevalence of children exposure to the two risk factors by gender in 85 countries as an example. In a first step, we estimated mean difference (absolute inequality) by gender (and their 95% confidence intervals) and adjusted for the sample design, including sample weights, primary sampling units (psu), and strata. In the second step, we obtained aggregate-level inequality by averaging over country-level mean difference across the countries and computed 95% CI of the aggregate-level inequality using ̅ √ (1) Where ̅ represents the aggregate-level mean inequality, is the z-value (Z=1.96 for 95% CI) and is the standard deviation of the means of the 85 countries. N is the number of countries (85).
Among the 85 countries, four countries (Uganda 2011, Lao 2017, Palestine 2010, and Sierra Leone 2017 have no information on psu or strata. We were not able to generate national-level estimates adjusted for these variables and therefore used the sample means and 95% CI in these countries by assuming that the sample was randomly drawn from the populations directly. (

2) Trend analysis for four indicators at the country level
We conducted trend analysis on inequalities at the country level for countries with data available in the two rounds.
Taking inequality of prevalence of risk exposure by gender as an example. For each country with data in two rounds, we first conducted logit regression (adjusted for sample design) with a dichotomous variable indicating a child's status of being not exposed to stunting and poverty as the outcome variable, and gender, round indicator, and interaction between gender and round indicator as the independent variable. Then we used linear combination to generate the absolute difference of inequalities (and their 95% confidence intervals) by gender between the two rounds.
BMJ Publishing Group Limited (BMJ) disclaims all liability and responsibility arising from any reliance

Chapter 4 Country-level estimates for four indicators in the most recent years since 2010
(1) Not exposed to stunting or extreme poverty Table 1. Prevalence (%) of children NOT stunted or in extreme poverty by gender with 95% confidence intervals, the most recent years in 85 countries

Country
Year  (2) Attendance of early care and early education *: A negative sign implies an increase in the % of children attending early care and education, a positive sign a reduction in the % of children attending early care and education  -6.1, 9.5) *: Gender gaps are defined as the difference between boy and girl averages. A negative sign implies a reduction of inequality, a positive sign implies an increase of inequality. Table 11. Level and trends of area inequalities in the prevalence (%) of attending early care and education with 95% confidence intervals in 16 countries with data available in two rounds  (3) Home simulation Table 13. Percentage of children exposed to high level of home stimulation by gender with 95% confidence intervals, the most recent years in 62 countries   Table 14. Percentage of children exposed to high level of home stimulation by place of residence with 95% confidence intervals, the most recent years in 62 countries *: A negative sign implies an increase in the % of children exposed to high stimulation, a positive sign a reduction in the % of children exposed to high stimulation.   Table 23. Level and trends of national prevalence (%) of ECDI with 95% confidence intervals in 13 countries with data available in two rounds Table 25. Level and trends of area inequalities in the prevalence (%) of ECDI with 95% confidence intervals in 13 countries with data available in two rounds

Country
Round 1 Round 2 Round1 -Round2 Rural Urban Gap Rural Urban Gap