A multilevel analysis of educational transition rates at secondary level in sub-Saharan Africa

Abstract

Estimates show that, in 2019, only 41 per cent of students completed lower secondary and 28 per cent upper secondary education in the sub-Saharan Africa (SSA) region (UNESCO, 2021). One of the reasons for the low completion rates is the poor transition across secondary education due to the significant impact of factors at individual, household, and community levels shaping demand and supply constraints. This article employs a three-level logit to investigate the key determinants for transitions and their variability across countries and communities, and explores whether less wealth inequality is at odds with increasing countries’ and communities’ performances. It finds that variation on transition rates is 40–50 per cent larger between communities within countries than between countries themselves, and that heterogeneity is larger for upper secondary transition. Leading sources of inequality are wealth, gender, and communities’ prevalence of early marriage. Further, the article finds that the equity–performance trade-off does not hold across countries, but it does at the community level where communities with stronger rates of transitions are more unequal. The analysis suggests policies to improve SSA youth chances to move up to the next level of secondary education, starting for narrowing heterogeneity across communities, boosting chances for the poorest groups and female youth living in communities with weak social norms, and measures to diminish the impact of community wealth on their transition performance.

Education coverage in sub-Saharan Africa (SSA) has grown significantly in the last 50 years. The gross intake rate for enrolment increased from 54 per cent to 99 per cent between 1970 and 2011, and the percentage of children completing primary school across the countries increased from 27 per cent to 67 per cent between 1970 and 2015 (Evans & Mendez Acosta, 2021; see also World Bank, 2022). Alongside this expansion of cohorts finishing primary education and then attending secondary education, there has been an increase in drop-out rates, partly explained by new cohorts entering the secondary system with wider challenges and bottlenecks from the supply side of SSA education systems. For instance, in 2020, the out-of-school rate was 33 per cent at lower secondary and 48 per cent at upper secondary and, more worryingly, these estimates have either stagnated during the last decade or have grown instead of declining (UNESCO, 2022).

The reasons for the poor uptake at secondary level in the SSA region include the additional challenges that young people face when they attempt to move through secondary levels, whether from lower to upper secondary or from upper secondary to higher education. In 2010, only 68 per cent of students who completed lower secondary were enrolled in upper secondary schools (Majgaard & Mingat, 2012) and projections are equally bleak. For example, UNESCO (2016) has projected that by the 2030 Sustainable Development Goals (SDGs) deadline, only 6 out of 10 young people will complete lower secondary in the SSA region, and fewer will finish upper secondary (4 out of 10), let alone move across these two education levels, on the back of both cumulative disadvantages building throughout the lifecourse and high-stakes exams (Bashir et al., 2018; Kellaghan & Greaney, 2019; Sayed & Kanjee, 2013). The latest (2019) available regional estimates for completion at lower and upper secondary levels are just 41 per cent and 28 per cent, respectively (UNESCO, 2021). Access at post-secondary is a key problem as well, with just 5 per cent of the population having the possibility to access tertiary education in SSA (Arias et al., 2019; Majgaard & Mingat, 2012).

Yet, the societal benefits of improving secondary education indicators can be far-reaching in terms of health and economic outcomes. In SSA, the under-five mortality rate is nearly double for children of mothers with no education compared to those who have completed secondary education (UNESCO, 2014), with increasing mother education also having significant impacts on vaccination status and visits for prenatal care (Guliani et al., 2014). On the economic impacts linked to education, Montenegro and Patrinos (2014), when estimating the educational returns for SSA, find that one additional year of education could increase earnings by 13 per cent, and Psacharopoulos and Patrinos (2018) estimate that the private returns to education (i.e., the increase in the earnings from an additional year of education) are very large in SSA (at around 10.5 per cent). In a recent study for Burkina Faso, Werner et al. (2022a, 2022b) find that, in comparison to individuals with primary education, individuals with secondary or higher education held 26 per cent more assets in the long term.

However, if the same weak pattern of progress (or lack of) for secondary education persists, most SSA countries would lag behind, and in the next two decades gaps will increase, compared to other countries with similar economic levels from either Latin America or Asia, making the achievement of SDG 4 for SSA countries unattainable (Arias et al., 2019). Furthermore, pressures on SSA secondary systems have been significant—and not matched in scale by the resources needed—due to large increments in the rate of primary completers over the last two decades of around 20 per cent (from 46.5 per cent in 2000 to 63.2 per cent in 2020) (UNESCO, 2023).

The SSA is a heterogeneous region, with education performance varying not only between countries by their income, but also within countries. For example, the mean lower secondary completion rate in upper middle income SSA countries, with an estimate of 83.5 per cent, is quite high, but it goes down to 56.3 per cent and 29.7 per cent in lower middle income and low-income countries from the region, respectively. Thus, country heterogeneity needs to be considered when modelling regional estimates and the determinants of secondary schooling indicators.

In this study, we advance the empirical literature on secondary school outcomes for SSA by looking at the key determinants and sources of heterogeneity (across countries and across communities within countries) for two transition rates or dependent variables, i.e., transition from lower to upper secondary and transition from upper secondary to higher education, with the latter being constructed as a new indicator proposed for the SDG 4 monitoring framework. Early studies (e.g., Kuepie et al., 2015; Momo et al., 2019) focus on access and completion rates while this paper, by examining secondary transition, fills this gap.

The analysis relies on data from the Demographic and Health Surveys (DHS) Program in 30 countries, and we follow a multilevel framework built up over three levels: countries, communities, and individuals. Specifically, the analysis reported in this paper attempts to answer the following three research questions (RQs):

• RQ1. What are the key determinants for transition rates from lower and upper secondary, and are these determinants different across SSA regions?

• RQ2. Is performance on transition rates heterogenous across countries, and what is its association with the variability of wealth effects (performance–inequality trade-off)? Are higher-performing communities more unequal, and, if so, does this inequality vary by community location and development?

• RQ3. What is the variability of transition likelihood across countries’ regions and their communities? And what are the leading determinants of the chances of transition by country?

The remainder of the paper is organized as follows. The Literature review section contains a review of leading determinants and bottlenecks when moving up through secondary school levels in resource-constrained education systems and in the SSA region, in particular. The Data section includes a description of the data used for the analysis and definitions of transition rates and key explanatory variables. The Empirical framework section outlines the multilevel methods employed. Results are presented in the Results section and, lastly, conclusions are offered in the Conclusions section.

Literature review

Low completion rates of secondary education have become one of the main public policy concerns in SSA. Free secondary education is aimed not only at preparing young people for entering the labour market but also at reducing income and wealth inequalities; the reason compulsory secondary education was included in SDG 4 (Bashir et al., 2018). Literature exploring the factors associated with the low transition rate between lower and upper secondary education in SSA is scarce. Most of the research focuses on retention and drop-out rates for basic and secondary education, with information about the transition from basic to secondary school or from lower-secondary to upper-secondary merely descriptive, and the variations between countries not fully explained (Bashir et al., 2018; Majgaard & Mingat, 2012). Different factors affect school drop-outs, which can also be determinants of low transition rates.

One of the main reasons for children dropping out of school in SSA is the cost of schooling and poverty (Evans & Mendez Acosta, 2021; Murtin et al., 2015; Pezzulo et al., 2022). In countries such as Kenya and Uganda, households dedicate 75 per cent of their expenditures on education to just one child (Bashir et al., 2018). Evaluations of different programmes aimed at increasing access and attendance in low- and middle-income countries have found that factors such as reducing prices and distance to schools and assigning merit scholarships are associated with reductions in drop-out rates (Duflo et al., 2021; Glewwe & Muralidharan, 2016; Kremer et al., 2013). A recent study estimates that, if tuition fees were eliminated in SSA, there would be an average increase of 6 per cent in upper secondary school enrolment (Asante, 2022).

In most SSA countries, governments cover the costs of education until grades 8 and 9 (lower secondary school). To continue in upper secondary education, students need to meet high academic standards and pay high tuition fees, as well as accessing school materials, uniforms, and transport. That is why countries such as Malawi have low rates of student transition to upper-secondary education (Hunsaker et al., 2022). Other countries, such as Uganda, have reduced tuition fees for upper-secondary education, but without an impact on enrolment, with the most advantaged students receiving this benefit instead of poorest students (Omoeva & Gale, 2016). In contrast, the elimination of secondary school fees in Gambia increased girls’ enrolment by 50 per cent as well as test scores (Blimpo et al., 2019). Similarly, in Kenya, the removal of fees has increased students’ educational access, but without an impact on students’ scores (Brudevold-Newman, 2021). Uganda eliminated secondary school fees in private schools via public funds and, as a result, boosted the number of students taking the exit exam by 16 per cent (Masuda and Yamauchi, 2018).

Direct and indirect costs are usually higher in rural areas, where, in a context of relatively higher poverty than in urban areas, children and young people spend more time and resources going to school (Abuya et al., 2013; Alegana et al., 2021; Bashir et al., 2018). A scholarship programme in Ghanian rural schools exhibited effects on educational attainment, knowledge, skills, and preventative health behaviours, while reducing female fertility (Duflo et al., 2021). Hunsaker et al. (2022) found that a scholarship programme in rural Malawi had impacts in reducing the costs of household participation in schooling and travelling costs to school, with scholarship recipients achieving between 1 and 1.5 years additional years schooling and higher rates of graduation. In Benin, a programme for building rural schools significantly increased enrolment in rural areas (Deschênes & Hotte, 2019), while in Zambia a programme to increase the salary of rural teachers had significant effects on student outcomes as well (Chelwa et al., 2019).

Contextual factors are a determinant for educational outcomes and the likelihood of students dropping out. In the poorest neighbourhoods of Nairobi (Kenya), drop-out rates are particularly high, with school fees a likely explanation for these higher out-of-school rates (Abuya et al., 2013). In a study of 30 developing countries, Huisman and Smits (2015) found associations between student drop-out and contextual and household-level characteristics, such as socio-economic resources, parental education, and father’s occupation. Fathers with lower levels of education are not able to help children with their assignments and this becomes a barrier for students whose fathers have only achieved low educational levels (Biddlecom et al., 2008; Huisman & Smits, 2015). Other associated factors found in the literature include employment status, living in a single-parent household, age, region of residence, and school average performance, a proxy for school quality (Momo et al., 2019).

Educational gender gaps are still large in SSA (for recent regional estimates using twin-samples of opposite sex, see: Delprato, 2022). Gender inequality is associated with greater school drop-out, usually linked to early marriage and pregnancy (Bashir et al., 2018; Biddlecom et al., 2008; Parsons et al., 2015; Wodon et al., 2018). Girls with reduced education due to early marriage will have lower earnings, productivity, and useful work skills during their entire life, transmitting inequality cross-generationally by having less healthy and less educated children (Baxter et al., 2022; Delprato et al., 2015, 2017; Parsons et al., 2015). In upper-secondary school, the effect of early marriage and pregnancy on dropping out is even higher than in lower secondary (Bashir et al., 2018). There is evidence that the permanence of girls in schools is usually related to their academic performance (Kuepie et al., 2015).

Girls who have married early largely do not have control over their earnings, are isolated from support networks, and do not have any social help to sustain their emotional well-being (Duflo, 2012; Raru et al., 2022). In contrast, women and girls from wealthier and more educated families tend to marry later, have fewer children, and get pregnant at an older age, but these differences vary significantly by SSA country (Channon & Harper, 2019; Corker et al., 2022; Kebede et al., 2022; Ndagurwa & Chemhaka, 2020). Children of couples who married early and became parents at a young age are also prone to get married and have children at younger ages (Odimegwu et al., 2020).

Data

This study is based on DHS data from the SSA region (Measure DHS, 2022). We include all countries from this region with survey years from 2010 onwards, yielding a total sample of 30 countries. The DHS data are ideal for carrying out cross-national analysis on the determinants of transition at secondary level because they have the double advantage of providing nationally representative samples that are also comparable across countries. At the same time, the surveys provide rich information on socio-economic, demographic, and health factors and social norms, all determinants linked to transition. Furthermore, the primary sampling unit (PSU) in DHS is the community, which is particularly suitable for our purpose as communities can be seen as proxies, capturing educational supply and labour market forces that explain the heterogeneity of transition rates within countries and inside regions.

DHS data have been widely used in literature to examine educational outcomes in SSA (see, for instance, Alegana et al., 2021; Anyamele et al., 2017), as well as providing community data as a root of heterogeneity through multilevel modelling (Delprato et al., 2016; Pezzulo et al., 2022; Raru et al., 2022; Schrijner & Smits, 2018). All studies, however, tend to focus on access/drop-out indicators and completion rates (Huisman & Smits, 2015; Kuepie et al., 2015; Momo et al., 2019) as key indicators, whereas empirical studies relying on secondary transition rates (i.e., the likelihood of moving from lower to upper secondary and from upper secondary to higher education) are absent. This paper attempts to fill this empirical gap.

Dependent variables

The primary sample is defined by the two outcomes (or dependent variables) with sample sizes for the whole SSA of 54,367 (transition from lower secondary to upper secondary) and 21,367 (transition from upper secondary to higher education). It should be noted that the transition rate from upper secondary is a novel indicator we introduce in this paper. Transition rates are defined as ratios between cohorts attending two different secondary education levels (lower secondary versus upper secondary, and upper secondary versus higher education). Specifically, employing the definitions given by UNESCO-GERM (with a mandate to monitor SDG 4), each secondary transition rate is calculated as follows,

$${Y}_{\text{tr, from lowsec}}=\frac{\sum_{i}^{N}1\left\{\left({\text{grade}}_{i}\ge \text{first upsec grade}\right) \bigwedge \left({\text{age}}\in {\text{age}}_{upsec}\right)\right\}}{\sum_{i}^{N}1\left\{\left({\text{grade}}_{i}\ge \text{last lowsec grade}\right) \bigwedge \left({\text{age}}\in {\text{age}}_{upsec}\right)\right\}}$$

(1)

$${Y}_{\text{tr, from upsec}}=\frac{\sum_{i}^{N}1\left\{\left({\text{grade}}_{i}\ge \text{last upsec grade +1}\right) \bigwedge \left({\text{age}}\in {\text{age}}_{HE}\right)\right\}}{\sum_{i}^{N}1\left\{\left({\text{grade}}_{i}\ge \text{last upsec grade}\right) \bigwedge \left({\text{age}}\in {\text{age}}_{HE}\right)\right\}}$$

(2)

where i denotes the ith individual in the DHS data, 1(cond) is the indicator function which is equal to 1 if the condition is true and 0 otherwise, and ${\text{age}}_{upsec}$ = [upsec entry age + 1, upsec finishing age + 4], and ${\text{age}}_{HE}$ = [upsec entry age + 1 + 4, upsec finishing age + 4 + 4]. Plainly, the transition from lower secondary is the number of young people attending the first grade of upper secondary school as a percentage of those attending the final grade of lower secondary school, using the corresponding age group (those in upper secondary age); and, equivalently, for the second transition rate. Note that age groups for each indicator are defined in such a way so as to consider populations after the official/technical finishing age for each educational level, to account for the large prevalence of overage completion rates in SSA.

Transition rates for SSA countries are shown in Table 1. First, as expected, the chances of moving from upper secondary to higher education are much lower than towards the last stage of secondary (i.e., to upper secondary). The overall rates for SSA are 76 per cent for transition from lower secondary and 50 per cent for transition from upper secondary school, clearly highlighting the barriers young African people face.

Table 1 List of countries/surveys used in the analysis and mean value for transition rates

Table 1 also shows that the gap between the two transition rates is acute in the case of the Western Africa (WA) region where transition from lower secondary is 82 per cent, but from upper secondary it is just 47 per cent; while the gap between the two transition rates in Middle Africa (MA) is 24 per cent, in Southern Africa (SA), 19 per cent and in Eastern Africa (EA) 18 per cent. Countries with the largest gap between the two rates (≥ 0.45) are Sierra Leone, Nigeria, Liberia, and Rwanda.

Furthermore, there is a substantial degree of heterogeneity as to the likelihood of moving on from these two secondary stages, not only in the whole SSA region, but also within each sub-region. In particular, SSA sub-regions’ range (that is, the difference between top and bottom performing countries in each region) for transition rates from upper secondary are large: 38.5 per cent (MA), 61.6 per cent (EA) and 56.2 per cent (WA); and, still for transition rates from lower secondary, there are regional gaps of 15.7 per cent and of 21.3 per cent for the MA and WA regions, respectively. This degree of variation in the sample justifies the multilevel approach we follow, which accounts for a country’s varying performance by using specific intercepts or country fixed effects.

Dependent variables

We include an array of covariates (which we also use in our multilevel modelling) to account for factors, operating at individual/household and community levels, that influence the transition rates of SSA’s young people (Table 2). The chosen covariates have been shown by studies, such as Evans and Mendez Acosta (2021), to be important drivers of educational outcomes at secondary level. Gender and household wealth (Kuepie et al., 2015), coupled with the degree development of a community (Huisman & Smits, 2009, 2015)—which can be proxied by poverty and the extent of bankarization (Steinert et al., 2018) as well as the location of a community (Kazeem et al., 2010)—are prime determinants of schooling outcomes for the region. Equally, community factors such as increasing fertility rates will diminish schooling transition (Webbink et al., 2012) because of the additional burden of resources placed on households, whilst (negative) social norms, such as high levels of early marriage, will lead to higher secondary drop-outs (Delprato et al., 2015) and weaker secondary transition.

Table 2. Background characteristics by low–high values of transition rates

Table 2 displays the mean values for covariates by low and high values of the two transition rates (below or above the whole sample mean). The table shows how the context of families and communities are related to youth transition rates. In comparison to the group of youth with low transition rates, members of the youth group achieving higher transition rates come from families with higher wealth (index differences of 0.4 from low secondary and 0.59 from upper secondary education), and there is a reverse gender gap between the two transition rates (of 3.1 per cent against girls for lower secondary and of 1.5 per cent in favour for girls for upper secondary ) (Table 2, Panel A).

At the community level (Panel B), higher transition rates for young people are observed for those communities with higher development, with development index gaps among low–high transition groups being larger for transition from upper secondary (=0.209) in comparison to lower secondary (=0.116). Also, transition rates are linked to where young people live, underscoring supply constraints at upper secondary level and above with 50.7 per cent (39.3 per cent) and 40.2 per cent (25.1 per cent) of the low-performing transition groups living in rural (urban) communities, and the community degree of bankarization being larger for youth with transition rates above average. Gaps on community fertility rates between the low and high transition rate groups are of 0.427 (at lower secondary) and 0.661 (at upper secondary). Using data from Panel C, the gender gap within a country’s regions are all shown to be weighted against young females (≤ 0.97), and they are increasing from ≈ 0.90 (transition from lower secondary) to ≈ 0.77 (transition from upper secondary). The wealth gaps (poor/rich) are also increasing from the latter transition rate; in other words, systemic regional inequalities become wider as young people attempt to move up through educational levels.

Empirical framework

To answer the three research questions (RQs), we rely on two kinds of multilevel model: one for the whole SSA region, and the other at the country level. In both models, the approach places the community at the core of variations in transition rates and educational opportunities, given its linkage with education demand (for skills gained after moving through the educational levels) and supply factors, which are all broadly captured by the unit of sampling in DHS data, i.e., the community. For each multilevel model, we also allow for random impacts of household wealth. Details of the methods employed are included below.

Pooled sample: Multilevel modelling

The analysis for the pooled sample (or SSA region) is based on a three-level logit model to account for data clustering, with level 1 defined by individuals (youth), level 2 by communities, and level 3 by countries. We rely on a logit model given the binary nature of the dependent variable (i.e., ${Y}_{\text{tr}}$ = 1 if youth or individuals move from lower to upper secondary or from upper secondary to higher education, and 0 otherwise).

First, to estimate the impact of individual/household, community, and regional factors on transition rates, we specify a three-level random intercept logistic regression model (RIM) for ${Y}_{\text{tr,ijk}}$ where individuals i (i=1,..., N) are nested in community j (j = 1,..., J) which are nested in country k (k = 1,..., K):

$$ \begin{aligned} {\text{logit}}\left( {{\text{Pr}}(y_{ijk} = 1|{\text{x}}_{ijk} ,\mu_{jk}^{\left( 2 \right)} ,\mu_{k}^{\left( 3 \right)} )} \right) & = IF{\text{x}}_{ijk} \beta + CR{\text{x}}_{jk} \theta \\ & \quad + \mu_{jk}^{\left( 2 \right)} + \mu_{k}^{\left( 3 \right)} + \epsilon_{ijk} \\ \end{aligned} $$

(3)

where covariates are split into individuals and families $(IF{\text{x}}_{ijk})$ and community/region $(CR{\text{x}}_{jk})$, ${\mu }_{jk}^{\left(2\right)}$ and ${\mu }_{k}^{\left(3\right)}$ are the community’s and country’s random intercepts, and ${\epsilon }_{ijk}$ is the level 1 error term following a logistic distribution with variance ${\pi }^{2}/3$ (≈ 3.29). These two random terms, which are assumed to be independently distributed both across level 2 and 3 units, provide a control for community and country variation (unobserved factors) on the probability of transition rates. The estimation of equation (Eq.) (3) is employed to answer RQ1.

Also, note that we run Eq. (3) by adding controls sequentially: first, we run an empty model (without) covariates (model 0: M0), and then we include level 1 covariates (M1) and further we add level 2 covariates (M2). We also estimate models by SSA sub-regions (i.e., MA, EA, SA, and WA regions). When adding explanatory variables across levels, we rely on the intraclass correlation (ICC) coefficient to gauge an estimation of the decomposition of the variance across the available levels and how much variance is explained at each level by different covariates (Hox et al., 2017). The two ICC, at community and country level respectively, are calculated as:

$${\rho }_{\text{community}}=\frac{{\mu }_{jk}^{\left(2\right)}}{{\mu }_{jk}^{\left(2\right)}+{\mu }_{k}^{\left(3\right)}+{\pi }^{2}/3}$$

(4)

$${\rho }_{{\text{coun}}{\text{try}}}=\frac{{\mu }_{k}^{\left(3\right)}}{{\mu }_{jk}^{\left(2\right)}+{\mu }_{k}^{\left(3\right)}+{\pi }^{2}/3}$$

(5)

Furthermore, to assess the potential heterogeneous impact of household wealth on transition, we assume that the relationship between the chances of transition and wealth (i.e., ${\text{wealth}}_{ijk}$) randomly varies across units, which defines a random slope model (RSM). We estimate two (simpler) alternative three-level RSM, with the household wealth explanatory variable as random (either across communities or countries).

That is, we estimate the following two models:

$$ \begin{aligned} {\text{logit}}\left( {{\text{Pr}}(y_{ijk} = 1|{\text{wealth}}_{ijk} ,\mu_{jk}^{\left( 2 \right)} ,\mu_{k}^{\left( 3 \right)} )} \right) & = \beta_{1} {\text{wealth}}_{ijk} + \mu_{jk}^{\left( 2 \right)} + \mu_{k}^{\left( 3 \right)} + \epsilon_{ijk} \\ & \quad + \mu_{k}^{\left( 3 \right)} \times {\text{wealth}}_{ijk} \\ \end{aligned} $$

(6)

$$ \begin{aligned} {\text{logit}}\left( {\left. {{\text{Pr}}(y_{ijk} = 1} \right|{\text{wealth}}_{ijk} ,\mu_{jk}^{\left( 2 \right)} ,\mu_{k}^{\left( 3 \right)} )} \right) & = \beta_{1} {\text{wealth}}_{ijk} + \mu_{jk}^{\left( 2 \right)} + \mu_{k}^{\left( 3 \right)} + \epsilon_{ijk} \\ & + \mu_{jk}^{\left( 2 \right)} \times {\text{wealth}}_{ijk} \\ \end{aligned} $$

(7)

Our interest behind the estimation of Eq. (6) and (7) is to analyse whether there is a trade-off (i.e., RQ2) between performance (given by fixed effects at the community and country levels) and inequality (wealth random effects above average). That is, whether more unequal systems are also high-performing systems or, alternatively, if there is a double disadvantage—i.e., lower than average community (country) effects on the probability of transition are coupled with strong household wealth effects on transition in the community (country).

Whether this trade-off holds can be analysed by plotting both kinds of random effects. And, more directly, it can be measured by the covariance term of the two random effects intercept–wealth slope correlations: ${\mu }_{\mathrm{1,2}}^{\left(2\right)}$ and ${\mu }_{\mathrm{1,2}}^{\left(3\right)}$ of Eq. (8) in the (symmetric) covariance matrices of the RSM’s random effects:

$$ {\varvec{\mu}}^{\left( 2 \right)} = \left[ {\begin{array}{*{20}c} {\mu_{1,1}^{\left( 2 \right)} } & {\mu_{1,2}^{\left( 2 \right)} } \\ {\mu_{2,1}^{\left( 2 \right)} } & {\mu_{2,2}^{\left( 2 \right)} } \\ \end{array} } \right]\quad {\text{and}}\quad {\varvec{\mu}}^{\left( 3 \right)} = \left[ {\begin{array}{*{20}c} {\mu_{1,1}^{\left( 3 \right)} } & {\mu_{1,2}^{\left( 3 \right)} } \\ {\mu_{2,1}^{\left( 3 \right)} } & {\mu_{2,2}^{\left( 3 \right)} } \\ \end{array} } \right] $$

(8)

Country analysis: Multilevel modelling

We estimate multilevel models for each SSA country. Here, we also employ a three-level hierarchical structure, though with a country’s regions replacing level 3 (k’), level 2 (as before) the community (j), and level 1 (i) individuals. Under this setting, we attempt to answer RQ3, exploring the degree of relative variation of secondary transition given by either a country’s regions or a country’s communities, by fitting a RIM. We then examine the impact of selected covariates (i.e., wealth, gender, community location, and development) on the probability of the two transition rates. For the dependent variable ${Y}_{\text{tr, lowsec}}$ and for the country analysis, the average number of units across the three levels is: individuals (= 1,875), communities (= 414) and regions (= 11). For the dependent variable ${Y}_{\text{tr, upsec}}$, these values are: individuals (= 737), communities (= 277) and regions (= 11).

Formally, for each country, we fit a null model (without covariates) and then a full RIM given by:

$$ \begin{aligned} {\text{logit}}\left( {{\text{Pr}}(y_{ijk^{\prime}} = 1|{\text{x}}_{ijk} ,\mu_{jk}^{\left( 2 \right)} ,\mu_{k}^{\left( 3 \right)} )} \right) & = \beta_{1} {\text{wealth}}{\text{.rich}}_{{ijk^{\prime}}} + \beta_{2} {\text{gender}}{\text{.male}}_{{ijk^{\prime}}} \\ & \quad + \beta_{3} {\text{comm}}{\text{.rural}}_{{jk^{\prime}}} + \beta_{4} {\text{comm}}{\text{.devhigh}}_{{jk^{\prime}}} \\ & \quad + \mu_{jk^{\prime}}^{\left( 2 \right)} + \mu_{k^{\prime}}^{\left( 3 \right)} + \epsilon_{{ijk^{\prime}}} \\ \end{aligned} $$

(9)

where all covariates are dichotomous variables. ‘Wealth.rich’ denotes households with wealth falling into the top two quintiles of its distribution, and ‘comm.devhigh’ measures communities with development above the SSA average.

Results

Research question 1: Key drivers of transition rates

Table 3 contains the RIM estimates for stepwise models (M0, M1, and M2) for the whole SSA region, sequentially adding controls. (LR tests for a standard logit model without using a hierarchical formulation against the multilevel approach, led to a rejection of the multilevel model being nested within a standard logit. Results for these tests, as well as RIM versus RSM tests, are available from the authors upon request.)

Table 3 Multilevel (logit) estimates for lower and upper secondary transition: Random Intercept Model (RIM), odds ratio

Panel B includes random effects. Estimates for lower secondary transitions are displayed in columns 1 to 3, and for upper secondary transition in columns 4 to 6. To begin with, ICC1 (country) and ICC2 (community) do not change much when new controls are added, except some minor changes for the upper secondary transition dependent variable where the ICC2 moves from 0.41 (null model–M0, column 4) to 0.38 (full model–M2, column 6). All variances (random effects, Panel B) at levels 2 and 3 are statistically significant regardless of the model fit.

Thus, variation on transition rates for all countries is simultaneously attributed to unobserved heterogeneity at the community and country levels, after accounting for individual and community characteristics. In particular, in the full model (columns 3 and 6), the ICC2, or proportion of the variance attributed to community level factors, is 25 per cent and 38 per cent for lower and upper secondary transition, respectively, and this variation is larger than the rate of variation between countries with the ICC1, 18 per cent and 25 per cent. Moreover, it should be noted that, the extent of community and country heterogeneity for upper secondary transition, is higher than for lower secondary transition, showing a higher impact of unobserved heterogeneity at the community and country levels for the former indicator.

We discuss estimates for the fixed part (Panel A of Table 3) based on the full model (M2) using odds ratio (OR). On the one hand, for the outcome lower secondary transition (column 3), household wealth is a key determinant of transition: an individual from a wealthy family in the top quintile (wealth-Q5) has nearly four times more chances ($\widehat{OR}$ = 3.95) of moving from lower to upper secondary in comparison to an individual from the bottom quintile; even an individual from the second wealth quintile (Q2) has 30 per cent more chance of moving to the next education level compared to an individual coming from the poorest household. Gender, too, is another source of inequality: male youth are 27 per cent more likely (i.e., 1.268 – 1) to move up to upper secondary than female youth. When it comes to the community factors, with $\widehat{OR}$ = 0.67, early marriage is a strong barrier against transition to upper secondary, and to a lesser degree increasing fertility rates, whilst individuals living in communities with a higher prevalence of more qualified jobs (upper non-farm) have larger transition probabilities ($\widehat{OR}$ = 1.16).

For the transition from upper secondary (column 6, Table 3), results are similar to those obtained for lower secondary; though, overall, drivers of inequality are more intense. For instance, the OR for young people coming from the top wealth quintile family is 5.26, the OR for community rate of qualified jobs is 1.31, and community factors such as development (1.12) and bankarization (1.68) are now statistically significant, although the gender gap in favour of male youth associated with ${Y}_{\text{tr, upsec}}$ is smaller in comparison to lower secondary transition (1.08 versus 1.27). Importantly, an increasing wealth-driven inequality measured at the regional level of a country has negative impacts on transition; plainly: more inequality in countries’ regions is related to weak transition rates from upper secondary.

For completeness, Table 4 displays estimates by SSA regions. First, as far as the fixed part of the model is concerned (Panel A), household wealth is still an important determinant for all SSA regions, with sizeable and increasing impacts of wealth shown by higher OR across quintiles. These impacts are more pronounced for the EA and SA regions (for lower secondary, columns 1 to 4); and in the case of upper secondary (columns 5 to 8), higher effects are found for the MA and the SA regions with ${\widehat{OR}}_{\text{wealth.Q5}}$ over 10, with a lack of impact for bottom wealth quintiles. The gender gaps (in favour of male youth) linked to transition hold for most regions, though for upper secondary transition outcomes, we find significant effects in two regions (i.e., MA and WA). Hence, gendered transition behaves differently across regions for older cohorts. Some dissimilar impacts of covariates across regions at the community level can be outlined here. For example, youth living in rural areas are 26 per cent and 44 per cent less likely to move across education secondary cycles in the MA region, and the community rate of bankarization has positive effects on transition for WA and EA, but not in the other two SSA regions.

Table 4 Multilevel (logit) estimates for lower and upper secondary transition by SSA regions: Random Intercept Model (RIM), odds ratio

Estimates for the random part of Table 4 (Panel B) show that there is heterogeneity in the variation of transition rates for the four SSA regions, but only at the community level (all variance–level 2 are statistically significant), as country variance estimates show that country variation does not hold for the MA and SA regions. This is also reflected in the low (below 5.3 per cent) ICC1 (country) for these two regions and, conversely, the large ICC2 (community), with around 25 per cent and 32 per cent of the total variation attributed to community factors for lower secondary, and 27.5 per cent and 46.2 per cent for upper secondary.

Research question 2: Transition heterogeneity across countries and wealth-driven inequality by countries and communities

In this section, we offer more insights on each countries’ performance on transition rates, and their associations with wealth effects stemming from countries and communities. Recall that in Table 3, for the whole SSA sample, we found statistically significant variation of transition rates for the two random components at the community $({\mu }_{jk}^{\left(2\right)})$ and country $({\mu }_{k}^{\left(3\right)})$ levels. Hence, RIM intercepts vary both across communities and countries.

This can also be seen in Figure 1, which contains random intercept estimates for each country, ranked by their values, for the null (M0) and full (M2) models (Eq. [3]). In the first plot for lower secondary (M2), out the 29 countries’ estimates, 20 countries’ intercepts are statistically significant, which can be further divided between 11 (9) positive (negative) random effects. Some of the countries with probabilities of transition from lower secondary that are much higher than the SSA average are: Nigeria, Zimbabwe, and Kenya, and the bottom four performing countries are: Mali, Tanzania, Uganda, and Mozambique. For upper secondary transition (Figure 1, Plot B), the top performing countries are not the same as for lower secondary (Ethiopia, Uganda, Senegal), but countries with a below average performance on lower secondary transition also have a below average performance on upper secondary transition (e.g., Mozambique, Chad, Burundi).

Figure 1

Whole sample random intercept model (RIM) estimates based on a 3–level model. Countries’ random intercept (transition rates performance) displayed

Notes: (1) Whole sample modelling: level 3 (countries); level 2 (communities); level 1 (individuals). (2) M0 denotes a null model (without covariates). (3) M2 comprises a full model (with covariates shown in Table 3, columns 3 and 6). (4) White bars denote non-statistically significant random intercepts at 10 per cent level

Figure 2 (Plot A and B) visually shows whether increasing transition at the country level is achieved at the expense of higher inequality. We do this by simultaneously plotting country random intercepts and random slopes for household wealth, based on the estimation of a RSM given by Eq. (6). In this way, disadvantaged countries with both low transition rates and large effects of wealth on transition can be identified.¹

Figure 2

Random slope model (RSM) estimates. (i) First panel (Plot A and B): Intercepts and wealth random effects at the country level. (ii) Second panel (Plot C and D): Communities intercepts and wealth random effect across communities

Notes: (1) Whole sample modelling: level 3 (countries); level 2 (communities); level 1 (individuals). (2) Wealth indicator equals to 1 if household wealth falls into quintiles 4 and 5, and 0 otherwise

Table 5 Random effects: RSM for whole SSA

For lower secondary transition (Figure 2, Plot A), some groups can be identified in terms of disadvantage. For the first group located in the top left quadrant and measuring double disadvantage (or low performance alongside positive impacts of wealth), there are nine countries, largely from the WA region. In a second group, countries such as Nigeria and Zimbabwe (in the top right quadrant) have large lower secondary transition probabilities but rooted in more unequal educational systems given by larger random effects of wealth. The bottom right quadrant, the ideal group exemplified by countries such as Ghana and Zambia, contains more equal systems (random effect below zero) with good average performances.

Further, for the upper secondary dependent variable, estimates in Figure 2 (Plot B) are less cluttered and more dispersed, although the direction of point estimates for the intercept and random slope goes from the top left quadrant to right bottom quadrant direction. This broadly implies that it is possible to boost transition from upper secondary level towards post-secondary education by lowering the degree of existing country inequality, measured by (positive) wealth random effects. In other words, more unequal countries are also weak performers for this transition indicator.

Plots C and D in Figure 2 examine community performance versus the impacts of wealth random effects by fitting Eq. (7). Here, as for the country analysis, dispersion is larger for the outcome lower secondary transition. Importantly, for both educational outcomes, results and the direction of scatter observations point towards a positive correlation of ${\mu 0}_{j}$ with${\mu 1}_{j}$. This suggests, in turn, that communities with stronger rates of transitions are more unequal and those communities with performance below the mean are less unequal; that is: higher transition rates are achieved at the cost of widening community inequality. Thus, a comparison of the two set of results may also suggest that the trade-off examined operates in a different fashion when proxied at the two levels of analysis.

Additionally, when we analyse in more detail in which type of community the trade-off is more obvious (see Figure 3), we find that it holds in relatively better-off communities. Indeed, the gradient (or correlation) between random intercept and wealth slopes is more pronounced in urban and high-developed communities in comparison to rural and less-developed communities.

Figure 3

Random slope models (RSM): Estimates by community location and development

(1) Whole sample modelling: level 3 (countries); level 2 (communities); level 1 (individuals). (2) Wealth indicator equals to 1 if household wealth falls into quintiles 4 and 5, and 0 otherwise

Research question 3: Variability of transition chances by countries’ regions and communities and their leading inequality drivers

So far, we have presented findings for the SSA region and its sub-regions. This section examines leading sources of variation and the key determinants for secondary transition rates at a country level, to provide guidelines for policies that address the different sources of inequality behind youth secondary transition chances.² The hierarchical structure followed for this country analysis is still of a three-level type, but with level 3 units being replaced by the administrative regions of a country where communities (level 2 units) are nested. Results are based on RIM formulations (for details, see Eq. [9]).

Country-level estimates for intraclass correlation coefficients for regions and communities are shown in Figure 4. A first thing to note is that SSA countries’ regions are a fair and sizable source of variation for youth transition rates, particularly from upper secondary. For the null model (M0), the mean ICC3 (region) across all countries for transition from lower secondary education is 2.9 per cent and from upper secondary it equals 7.4 per cent. Still, as shown by Figure 4, the community is the main source of inequality and variation at a country level with ICC2 of 12.7 per cent (from lower secondary) and 24 per cent (from upper secondary).

Figure 4

Random intercept model (RIM): 3-level model estimates for the intraclass correlation coefficient (ICC) for level 3 (regions) and level 2 (communities)

Notes: (1) Country modelling: level 3 (regions); level 2 (communities); level 1 (individuals). (2) For M0 and M2 model details, see Table 2

Also, when controls are added (i.e., M2), they account—on average for the whole sample—for around 52 per cent of the total variation of level 3 and level 2 variances for lower secondary transition, and leading to a smaller reduction for upper secondary transition (25 per cent of level 2 variance).³ Background household and community differences are vital to explain the high heterogeneity observed between regions and communities happening within countries.

In particular, the Plot A of Figure 4 (for lower secondary transition) shows several countries with a significant reduction on community variance after accounting for individual and community characteristics, illustrating how important mechanisms operating at the individual and community levels are in determining the within-country distribution of lower secondary transition. For instance, reductions on ICC2 (or level 2 variance) between the two models reaching over 50 per cent can be found in more than half (= 18) of SSA countries (Niger, Angola, Zimbabwe, Chad, Sierra Leone, and Benin have a reduction variance of 70 per cent) for lower secondary transition. In the case of upper secondary transition ( Plot B), the same reduction is obtained for one quarter of total sample of countries (= 7), with countries such as Zimbabwe, Liberia, Côte d'Ivoire, Rwanda and D. R. Congo, all having an ICC2 reduction of at least 60 per cent.

In Figures 5 and 6, we present OR estimates—with their confidence intervals (CI)—for the four chosen covariates (wealth, gender, community location, and development) for the two transition rates. For the models fit for the dependent variable transition from lower secondary (Figure 5), the highest root of inequality across countries is given by the level of wealth of families where SSA youth come from, as the largest estimates are for ${\widehat{OR}}_{\text{wealth.rich}}$, which are well over one. Specifically, in 20 countries, the impact of the CI for wealth’s OR does not overlap one (and so the wealth covariate is statistically significant), with values of OR, on average, approaching the value of 2, and for some countries the OR are even higher. For example, the chances of youth moving from lower to upper secondary in countries such as Angola, Cameroon, Gabon, Uganda, and Zimbabwe are around double in comparison to youth coming from poorer households (measured by the group defined by the bottom three quintiles of the wealth’s distribution). In another set of countries, the impact of wealth is even larger, for example in Kenya, Ghana, Nigeria, Tanzania, and Zambia, the $\widehat{OR}$ fall into a higher interval (2.30–4.00).

Figure 5

Countries’ leading determinants of lower secondary transition rates (3-level model), based on random intercept models

Notes: (1) Odds ratio displayed with 90 per cent confidence intervals

Figure 6

Countries’ leading determinants of upper secondary transition rates (3-level model), based on random intercept models

Notes: (1) Odds ratio displayed with 90 per cent confidence intervals

Additionally, gender impacts are more nuanced, from the point of view of statistical power, with estimates of ${\widehat{OR}}_{\text{gender.male}}\ge 1$ favouring transition for male youth relative to female youth holding for 12 countries. Notably, with OR values between 1.60–2.00, gender inequality is a strong source of inequality in WA countries (e.g., in Benin, 90 per cent, in Burkina Faso and Sierra Leone, 60 per cent, and in Togo, 100 per cent). Living in rural communities are equally a barrier for young people’s lower secondary transition (as $\widehat{OR}<1$, and statistically significant, in 11 countries), most of them also being countries that have gender gaps. In certain countries (e.g., Chad, Democratic Republic of the Congo, and Mozambique) individuals living in rural communities have less than half the chance (i.e., $1- \widehat{OR}\ge 0.5$) of moving up to upper secondary relative to those living in urban communities, while in Tanzania and Benin OR values are 0.54 and 0.63. Youth living in comparatively high-developed communities also shows increasing chances of transition from lower secondary, with the values of estimates of OR being similar as those for wealth, and even, for some countries (Zimbabwe, Mozambique, and Guinea), having higher effects than wealth.

Estimates for the four covariates for the indicator upper secondary transition to higher education are displayed in Figure 6. Here it is apparent that many of countries’ point estimates for gender and for community location lack statistical power (90 per cent CI overlap one). Nonetheless, location and gender disparities persist in some countries for upper secondary transition. In Nigeria and Burkina Faso, lack of equality for girls is given by OR values of 3.1 and 1.3; and location is a powerful drawback for the likelihood of transition in Chad and Democratic Republic of the Congo ($\widehat{OR}$ = 0.12), Gabon ($\widehat{OR}$ = 0.31), and Uganda ($\widehat{OR}$ = 0.53). Wealth and community development are again—and more strongly—leading determinants of transition from upper secondary. For instance, most of the OR increase for transition at this educational level shift towards values well above 2.

Conclusions

In the SSA region, progress towards the SDG 4 goals is weak. Secondary school attainment indicators show low completion rates (≈ 46 per cent from lower secondary and ≈ 27 per cent from upper secondary in 2022 [UNESCO, 2023]) and deep inequalities (e.g., parity indices for wealth show poorest to richest quintile from lower secondary is 0.15 and for upper secondary 0.07 [UNESCO, 2021]). It is, therefore, vital to know what hurdles young people face as they transition through secondary education. These two stepping-stones across youth education incorporate specific and cumulative inequalities and barriers, set alongside rising opportunity costs of staying in school, and the added complexity of high-stakes exams.

Hence, this paper examines these factors—both from a regional and country perspective—using a sample of 30 SSA countries and carrying out multilevel modelling for two transition rates: transition from lower secondary to upper secondary, and from upper secondary to post-secondary. As far as we are aware, this study is the first regional analysis of this type investigating transition rates, and it also introduces a new definition/indicator (i.e., transition from secondary to post-secondary education). The paper attempts to answer various questions, from leading drivers of transitions to sources of variation and heterogeneity on transition chances across SSA regions and countries, as well as the relationship between inequality in education systems and their performance.

The paper’s results can be summarized in four broad findings. First, the degree of heterogeneity on transition rates is significant at both country and community levels, albeit variation between communities on transition rates is more substantial than variation between countries, and more heterogeneity is found for the indicator transition from upper secondary. Second, socio-economic factors (proxied by household wealth) are the leading barrier for secondary transition (with wealth having a more prominent impact when modelling youth’s probability of moving from upper secondary to higher education), with gender inequality (in detriment of female youth) operating negatively at the individual level and community level (proxied by community prevalence of early marriage). Operating community factors linked to development (rates of qualified jobs and rate of bankarization) are statistically significant when measuring their impacts at the point when youth attempt to move up to post-secondary education. Third, a substantial amount of SSA countries—mostly from WA—suffer from double disadvantage: low performance on lower secondary transition and above average impacts of wealth. However, in the case of transition from upper secondary, overall, it is possible (in quite a few countries) to boost this rate with lower country inequality (negative wealth random effects). Fourth, communities with high rates of transitions are more unequal and those communities with performance below the mean are less unequal, plainly: higher performance is achieved at the cost of widening community inequality.

After accounting for different factors influencing SSA transition rates, there is still a significant variance attributed to communities of 25 per cent (38 per cent) for lower (upper) secondary transition, much larger than variability across countries (of 18 per cent and 25 per cent). The large magnitude of community variation (ICC) suggests that transition chances are influenced more strongly by community level factors and by the composition of the youth and families living in those communities. Also, the net differential on ICC at the community level between SSA countries (ranging from 15 per cent in Togo, Namibia, and Nigeria, to less than 3 per cent in Gabon, Senegal, and Gambia, among other countries, for lower secondary transition) further indicates the relative importance of dispersion of transition inside specific countries, shaped by local communities characteristics and how specific countries’ context matters for the probability on moving up across secondary levels. With communities ICC for upper secondary transition over 20 per cent for 9 (out of 29) SSA countries, it points towards additional challenges for countries’ policy-makers, to improve transition opportunities for youth living in different communities. Still, reductions on ICC after controls (of around 50 per cent in level 2 community variance) indicates some fair progress if community context (shaping demand and transition supply) differences are minimized between communities. Also, in some countries, for the indicator transition from upper secondary (e.g., Burkina Faso, Chad, and Gabon) heterogeneity operates in equal terms across regions and communities.

When looking at individuals’ disadvantages at the SSA regional level, the entry point of policies should be on wealth and gender disparities. This is aligned with the literature review presented (see, for instance, Blimpo et al., 2019; Duflo et al., 2021; Hunsaker et al., 2022; Masuda & Yamauchi, 2018). In fact, young people originating from richer households have between 3.95 and 5.26 more chance to move up to the next education level during secondary than those coming from the poorest households; and young females have from 27 per cent to 8 per cent fewer chances to do so in comparison to young males. Of course, these gaps exist on the back of comprehensive country disparities and would require specific country policies. For instance, wealth impacts for lower secondary transition are over 3 in Tanzania, Zambia, and Ghana, but between 1 and 2 in Democratic Republic of the Congo, and Gambia; and, likewise, Chad and Benin have larger gender gaps on transition in comparison to other countries.

Lessening inequalities for transition would require policies also tackling further challenges on differential disadvantages of communities based on social norms and development (as shown, for instance, by the negative impacts of early marriage prevalence in the community [$\widehat{OR}$ = 0.67], and the positive impacts of bankarization [$\widehat{OR}$ = 1.68]). This, in turn, may resolve—or at least alleviate—the tendency we found for the community trade-off between inequality and performance on transition rates (i.e., communities with stronger rates of transitions being more unequal).

Future directions for research may entail an examination of transition likelihood linked to labour-market opportunities and their compositions in each country. Also, a conceptual analysis is needed on why performance on secondary transition rates and inequality operates slightly differently by indicators and by level of analysis, as well as in better communities (urban and more developed). A further theme to research is the role of intersectionality and cross-level effects (e.g., wealth impacts by a country’s region, gender gaps by community location) framing transitions likelihood.

This paper has some limitations. Because of the cross-sectional nature of the dataset employed, findings should be interpreted with caution, as correlations and not implying causality. There is also the possible measurement error of transition rates due to errors on reported age (intervals) used to derive these indicators. Another limitation is that, because of DHS design, when employed for modelling educational indicators for age groups later on the lifecourse, they are more prone not to have the same range of covariates as when the education indicators being modelled are defined earlier on the education lifecourse. For instance, in the analysis, we could not account for demographic factors and family composition as well as the literacy of mothers, all key drivers behind transition rates. That is why we found that the reduction explained community variance, when accounting for individual and contextual features, did not reduce significantly and quite a lot of unobserved community transition heterogeneity still remains.