Rethinking the U-Shaped hypothesis: evidence of a convex education–participation relationship in Sri Lanka

Over the past three decades, South Asia has experienced sustained economic growth alongside substantial public investment in women’s health and education (World Bank, 2025). Gender gaps in educational attainment have narrowed, fertility rates have declined and women’s human capital indicators have improved across the region. Conventional development theory predicts that such gains should translate into higher female labour force participation (FLFP), thereby contributing to inclusive growth and improving returns to public investment in education. However, this expected translation remains limited. As of 2024, FLFP in South Asia stood at approximately 29.1%, compared with 77.08% for men (World Bank, 2025), highlighting a persistent and economically significant gender gap.

The relationship between educational attainment and FLFP has traditionally been conceptualised through the U-shaped hypothesis Goldin, (1995). The hypothesis predicts relatively high participation among women with very low education, a decline at intermediate levels and renewed participation at higher levels as access to skilled employment expands. While this framework has been widely applied in development economics, emerging evidence across South Asia suggests considerable cross-country variation. This indicates that the education–employment relationship is shaped not only by educational attainment, but also by differences in structural transformation, labour demand conditions and institutional contexts.

Comparative evidence illustrates this heterogeneity. Bangladesh provides a clear example. Export-oriented industrialisation, particularly the expansion of the ready-made garment sector, generated large-scale employment opportunities for women with low to intermediate levels of education (Heath and Mobarak, 2015). By contrast, India has experienced a prolonged stagnation and decline in FLFP despite rising female educational attainment and economic growth (Deshpande and Singh, 2024; Klasen and Pieters, 2015), while Pakistan continues to exhibit persistently low participation rates, reflecting structural rigidities and normative constraints (World Bank, 2024). These divergent trajectories suggest that educational expansion alone is insufficient to increase FLFP. Outcomes are fundamentally mediated by labour demand conditions and institutional context.

Within this broader regional context, Sri Lanka represents a distinctive, comparatively under-examined case. The country has achieved high levels of female educational attainment and has undergone a gradual transition towards service-oriented sectors, creating conditions that may alter the expected education–FLFP relationship. However, existing research has predominantly focused on larger South Asian economies. Consequently, limited empirical attention has been given to whether Sri Lanka’s experience conforms to or diverges from the conventional U-shaped hypothesis, particularly regarding the functional form of the education–participation relationship.

This study, therefore, examines Sri Lanka as a country-specific case to assess whether the relationship between educational attainment and FLFP conforms to the U-shaped hypothesis or exhibits an alternative non-linear pattern. Specifically, it investigates whether increases in educational attainment translate uniformly into higher participation, or whether structural constraints produce uneven outcomes across education levels. By focusing on Sri Lanka, the study contributes to the development literature by demonstrating how country-level labour market structures shape the returns to education for women’s economic participation.

This study makes three contributions. Firstly, it provides new empirical evidence on the education–FLFP relationship in Sri Lanka using recent Labour Force Survey (LFS) data (2017–2024), addressing the country’s relative underrepresentation in the South Asian literature. Secondly, it advances methodological approaches in development research by explicitly modelling the functional form of the education–participation relationship using beta regression techniques appropriate for bounded outcomes (Ferrari and Cribari-Neto, 2004). Thirdly, by identifying a convex (J-shaped) pattern in a service-oriented developing economy, the study contributes to development theory by showing that the U-shaped hypothesis is context-dependent rather than universal (Goldin, 1995) and that the effectiveness of educational expansion depends critically on alignment with labour market demand.

The analysis adopts a country-specific approach based on Sri Lankan data. While the South Asian literature provides an important comparative backdrop, the findings are not intended to be generalised to the region. Instead, Sri Lanka is treated as an analytically informative case that highlights the importance of national labour market structures, institutional conditions and sectoral dynamics in shaping the relationship between education and FLFP.

The relationship between educational attainment and FLFP is commonly framed through the U-shaped hypothesis, first articulated by Goldin (1995). The hypothesis predicts relatively high participation among women with very low levels of education, a decline at intermediate levels and a resurgence at higher levels as access to skilled and formal employment expands.

In developing economies, particularly in South Asia, this framework has been widely used to explain persistently low FLFP despite improvements in female education. Empirical studies suggest that women with secondary education often withdraw from the labour force due to social norms, household responsibilities, safety concerns and limited access to suitable employment. In contrast, highly educated women are more likely to enter professional and formal-sector work (Afridi et al., 2024; Goldin, 1995; Jayachandran, 2021; Klasen and Pieters, 2015; Najeeb et al., 2020). As a result, the U-shaped hypothesis has become a dominant reference point in regional labour and development research.

However, much of the empirical support for the U-shaped relationship is derived from country-specific or micro-level analyses, often focusing on large economies such as India. These studies typically examine participation decisions at the individual or household level and may not explicitly test the functional form of the education–FLFP relationship at the aggregate level. Moreover, the persistence of the U-shaped pattern is frequently interpreted as a stylised empirical regularity. It is less often systematically evaluated across diverse institutional and labour market contexts. This raises questions about the extent to which the U-shaped hypothesis can be treated as a generalisable framework, particularly in smaller or structurally distinct economies within South Asia.

Several theoretical perspectives inform the relationship between educational attainment and FLFP, though their implications are not always consistent when considered in isolation. Human Capital Theory (Becker, 1964) posits that higher levels of education enhance productivity and expected returns to market work, thereby strengthening incentives for labour force participation. In contrast, the Dual Labour Market Theory (Doeringer and Piore, 1971) emphasises structural constraints, highlighting segmentation between informal, low-quality employment and more stable primary-sector jobs. From this perspective, the effect of education depends on whether it enables women to transition into higher-quality employment segments rather than merely increasing their qualifications. This is consistent with evidence highlighting spatial labour market constraints and gendered access to employment across rural–urban gradients, which shape women’s ability to translate education into labour market participation (Chatterjee et al., 2015).

Complementing these economic explanations, Role Congruity Theory (Eagly and Karau, 2002; Wood and Eagly, 2012) introduces a normative dimension. It emphasises how gendered expectations shape labour market behaviour. Even where education increases potential returns, prevailing social norms may discourage labour force participation, particularly at intermediate levels of education, where women face tensions between traditional roles and emerging economic opportunities. This suggests that the relationship between education and FLFP is determined not solely by economic incentives but also by the interaction between labour market structures and socially constructed gender roles.

These perspectives suggest a non-linear and context-dependent relationship between education and FLFP. At lower levels of education, participation may be driven by necessity in low-productivity sectors; at intermediate levels, rising household income and restrictive norms may reduce participation; and at higher levels, increased access to skilled employment and shifting norms may encourage re-entry into the labour force. However, existing theoretical applications often treat these mechanisms independently, with limited integration of economic and normative explanations into a unified framework, which may partly explain the persistence of mixed empirical findings across contexts.

Empirical evidence from South Asia shows patterns broadly consistent with the U-shaped hypothesis, particularly in India, where economic growth has coincided with persistent gender norms and labour market barriers at intermediate levels of education (Desai and Joshi, 2019; Klasen and Pieters, 2015). However, these findings are largely derived from large economies and may not fully capture variation across smaller or structurally distinct contexts. In Sri Lanka, earlier studies suggest that structural transformation initially favoured male-dominated sectors, while sectors employing women expanded more gradually (Gunatilaka, 2013; Jayaweera, 1997), indicating that both labour demand conditions and institutional factors shape women’s participation outcomes.

Recent research increasingly calls into question the uniform applicability of the U-shaped hypothesis across South Asia. Lahoti and Swaminathan (2016) argue that demographic change, educational expansion and shifting labour demand have altered the traditional U-shaped relationship in certain contexts. Similarly, Klasen et al. (2019) and Jayachandran (2015) suggest that in economies undergoing structural transformation, the relationship between education and FLFP may follow a linear or convex pattern rather than a pronounced U-shape. As labour markets modernise and urbanisation expands, educated women may be better positioned to access employment aligned with their qualifications (Afridi et al., 2024).

However, these emerging findings also highlight an important limitation in the existing literature: the tendency to treat the U-shaped hypothesis as a broadly applicable framework despite substantial variation in underlying labour market conditions. Differences in sectoral composition, the pace and nature of structural transformation, and the availability of suitable employment opportunities for women can fundamentally alter how education translates into labour force participation. As a result, the expected U-shaped pattern may not hold in contexts where labour demand is concentrated in specific sectors or where employment opportunities for moderately educated women remain constrained.

Cross-country evidence further underscores this heterogeneity (see Figure 1). Bangladesh, for example, expanded female employment through export-oriented industrialisation, particularly in the garment sector, generating relatively high participation among women with low to intermediate levels of education (Heath and Mobarak, 2015). In contrast, India and Pakistan have experienced stagnating or declining FLFP despite rising female education and economic growth (Desai and Joshi, 2019; Klasen and Pieters, 2015). These contrasting trajectories suggest that the education–FLFP relationship is mediated less by educational attainment alone and more by the interaction between labour demand structures and institutional contexts.

Taken together, this evidence indicates that the U-shaped hypothesis should be interpreted as a context-dependent empirical pattern rather than a universal relationship. This has important implications for empirical analysis, as it suggests the need for country-specific investigations that explicitly examine the functional form of the education–participation relationship rather than assuming its shape a priori.

Pakistan is included in Table 1 to illustrate regional patterns in women’s educational composition among the employed, rather than for direct comparison of education-specific labour force participation rates, which are not available in the Pakistan LFS.

A parallel strand of the literature highlights important measurement challenges in assessing FLFP, particularly in developing-country contexts. Women’s economic activity is frequently undercounted in official labour statistics because a substantial proportion of women engage in informal, unpaid or family-based work that is not fully captured by standard LFS classifications (Jayachandran, 2021; Klasen et al., 2019). In addition, conventional measurement frameworks have historically struggled to reflect the full scope and diversity of women’s economic contributions, particularly in rural and household-based production systems.

These limitations introduce potential biases in the interpretation of FLFP patterns, particularly at lower levels of education, where women’s participation may be underreported rather than absent. As a result, observed low participation among less-educated women may partly reflect measurement constraints rather than genuine economic inactivity. This has important implications for the interpretation of non-linear relationships, including the U-shaped hypothesis, as the apparent decline in participation at certain education levels may be influenced by how economic activity is defined and recorded.

Furthermore, differences in measurement practices across countries and over time can complicate cross-country comparisons, potentially contributing to the observed heterogeneity in the education–FLFP relationship. Variations in survey design, definitions of work and the treatment of informal and unpaid labour may produce divergent estimates of participation, even in structurally similar contexts.

Taken together, these considerations caution against strong causal interpretations of aggregate FLFP statistics and highlight the need for careful empirical modelling. They may also reinforce the importance of examining the functional form of the education–participation relationship using appropriate statistical techniques, rather than relying solely on descriptive patterns that may be sensitive to measurement error.

Against this regional backdrop, Sri Lanka represents a distinctive yet comparatively under-analysed case. The country has achieved near-universal female literacy and high levels of secondary and tertiary attainment, supported by its long-standing free education policy introduced in 1945 [Alawattegam, 2020; UNESCO and United Nations Girls’ Education Initiative (UNGEI), 2015]. Women now account for a greater share of tertiary enrolments, reflecting sustained progress towards gender parity in education.

Sri Lanka has also undergone a gradual structural shift from agriculture towards service-oriented sectors, including education, healthcare, tourism, apparel and ICT-enabled services. These sectors place greater emphasis on formal qualifications and literacy, potentially aligning labour demand more closely with women’s educational attainment. The services sector accounts for approximately 59% of GDP (Central Bank of Sri Lanka, 2024), and expanding industries such as apparel, tourism and business process outsourcing have created employment pathways across multiple educational levels.

Cultural and institutional features further shape this context. Educated women are overrepresented in sectors such as education, public administration and healthcare (Seneviratne, 2020), reinforcing a strong participation premium at higher levels of education. However, employment opportunities remain geographically concentrated in urban and export-oriented zones, contributing to underemployment among moderately educated women in rural areas (Department of Census and Statistics Sri Lanka, 2024).

Recent studies have examined related aspects of education and women’s economic participation in Sri Lanka, though often from sector-specific or micro-level perspectives. Research on educational policy interventions has analysed how institutional reforms influence students’ subject choices and enrolment patterns (Priyadarshana, 2026). Other studies have explored the socio-cultural and structural barriers affecting women’s participation in emerging labour market segments such as digital and freelance work (Liyanage et al., 2025).

These studies provide important insights into education pathways and labour market constraints. However, they tend to examine education and employment outcomes separately. Consequently, they do not explicitly analyse how educational attainment translates into aggregate FLFP across the full distribution of education levels.

Taken together, these characteristics suggest that Sri Lanka’s education–FLFP relationship may differ from the conventional U-shaped pattern observed elsewhere in South Asia. However, the functional form of this relationship remains insufficiently examined at the macro level, particularly using recent data and methods capable of capturing non-linear patterns. This study addresses this gap by providing a country-specific analysis of the education–FLFP relationship in Sri Lanka.

While the relationship between educational attainment and FLFP has been widely examined in South Asia, the existing literature has predominantly focused on large economies such as India and Bangladesh, where the U-shaped hypothesis has been extensively tested and debated (e.g. Klasen and Pieters, 2015; Desai and Joshi, 2019; Klasen et al., 2019). A growing body of research has also explored the role of social norms, labour demand conditions and structural transformation in shaping women’s participation outcomes across the region (Jayachandran, 2015; Afridi et al., 2024).

In the Sri Lankan context, prior studies have provided important insights into female labour market participation, including the role of sectoral change, public-sector employment and gender norms (e.g. Gunatilaka, 2013; Seneviratne, 2020). However, much of this work has focused on descriptive trends, micro-level determinants or sector-specific dynamics. Consequently, relatively little attention has been given to the functional form of the relationship between educational attainment and FLFP at the aggregate level. This limitation partly reflects data constraints, reliance on micro-level survey approaches and a tendency in the literature to treat the U-shaped hypothesis as a stylised empirical regularity rather than to subject it to context-specific testing.

More specifically, existing studies have not systematically examined whether Sri Lanka’s education–FLFP relationship conforms to the conventional U-shaped hypothesis or exhibits an alternative non-linear pattern. This remains unexplored using recent data and econometric techniques appropriate for bounded outcomes. This represents an important gap, particularly given Sri Lanka’s high levels of female educational attainment and its distinct labour market structure relative to other South Asian economies.

This study addresses this gap by providing a macro-level analysis of the education–FLFP relationship in Sri Lanka using aggregated LFS data (2017–2024) and beta regression techniques that allow for flexible modelling of non-linear patterns. In doing so, it contributes methodologically by explicitly modelling the functional form of the relationship, and theoretically by demonstrating how country-specific labour market structures can produce deviations from the conventional U-shaped pattern. Rather than extending regional generalisation, the study refines existing theoretical expectations by highlighting the context-dependent nature of the education–FLFP relationship.

This study develops an integrated macro-level framework to examine the relationship between educational attainment and FLFP in Sri Lanka. Rather than treating existing theories independently, the framework conceptualises the education–FLFP relationship as the outcome of three interacting mechanisms:

1. Human capital accumulation.

2. Labour market segmentation.

3. Normative constraints.

This integrated approach allows for a more systematic interpretation of non-linear participation patterns in developing-country contexts.

From a human capital perspective (Becker, 1964), higher levels of education increase productivity and expected returns to labour market participation. In isolation, this framework predicts a positive relationship between education and FLFP. However, this effect is contingent on the structure of labour demand.

Dual Labour Market Theory (Doeringer and Piore, 1971) introduces a structural dimension, highlighting segmentation between informal, low-quality employment and more secure, higher-wage primary-sector jobs. Education may facilitate women’s transition into the primary segment, but where such opportunities are limited, increases in education may not translate into higher participation, particularly at intermediate levels.

Gender-normative perspectives (Jayachandran, 2015; Wood and Eagly, 2012) further condition this relationship by shaping the social acceptability of women’s employment. Normative constraints may discourage labour force participation, especially among moderately educated women, where rising household income and social expectations interact to reduce labour supply.

These mechanisms operate simultaneously and interactively. They shape how educational attainment translates into labour market participation across different levels. At lower levels of education, participation may be driven by necessity in the informal sector. At intermediate levels, limited access to suitable employment combined with normative constraints may suppress participation. At higher levels, access to professional employment and shifting norms may lead to increased participation.

This integrated framework contributes theoretically by reconceptualising the education–FLFP relationship as a context-dependent outcome of interacting economic, structural and normative forces, rather than as a fixed U-shaped pattern. It therefore provides a basis for empirically testing whether the relationship follows a conventional U-shape or an alternative convex (J-shaped) pattern.

In this study, educational attainment is operationalised across five categories, and FLFP is measured using aggregated LFS data from the Department of Census and Statistics, Sri Lanka (2024). The framework also incorporates contextual factors common to developing economies, including sectoral composition and macroeconomic shocks (Heath and Jayachandran, 2018), unpaid care responsibilities and gender norms (Chatterjee et al., 2015; Jayachandran, 2015), institutional supports such as childcare provision (UN Women, 2022), and digitalisation-related labour market changes (World Bank, 2021).

The framework, therefore, conceptualises the education–FLFP relationship as an emergent outcome of interacting economic, structural and normative forces, rather than a predetermined functional form.

Figure 2 presents the integrated conceptual framework guiding the analysis of the relationship between educational attainment and FLFP in Sri Lanka. Rather than depicting a strictly sequential process, the framework conceptualises the education–FLFP relationship as the outcome of interacting mechanisms. Educational attainment influences labour market participation through three interrelated dimensions: human capital accumulation (skills and productivity), labour market structure (access to primary-sector employment) and normative constraints (gender norms and role expectations).

These mechanisms operate simultaneously and interact with one another, shaping how education translates into labour market participation across different levels of attainment. In the Sri Lankan context, sustained investment in female education and the expansion of service-oriented sectors such as education, healthcare and ICT create conditions under which higher levels of education are more closely aligned with labour market opportunities. At the same time, normative and structural constraints may limit participation at intermediate levels, contributing to a non-linear relationship.

Accordingly, the framework supports the possibility of a convex (J-shaped) pattern in aggregate FLFP, while recognising that this pattern is an empirical outcome rather than a predetermined structure. The relationship is further shaped by contextual moderators, including institutional supports (e.g. childcare and maternity protection), cultural context (evolving gender norms) and emerging opportunities associated with digitalisation and remote work. These factors influence the strength and direction of the relationship across different institutional and labour market settings.

Guided by the conceptual framework, this study examines whether FLFP in Sri Lanka (2017–2024) is associated with educational attainment in a manner consistent with a convex J-shaped pattern at the aggregate level. Given the use of grouped data, the hypotheses are framed in terms of associations and functional form, rather than individual-level causal effects:

The analysis uses aggregated data from the annual LFS conducted by the Department of Census and Statistics (DCS) of Sri Lanka for the period 2017–2024. The data set comprises 40 education–year observations. These observations represent the proportion of working-age women (aged 15–64) who are economically active (employed or actively seeking work) within five educational attainment categories: no schooling, primary education, lower secondary education (Grades 6–10), upper secondary education (Advanced Level) and tertiary education.

Educational attainment is the primary explanatory variable, and FLFP is the outcome of interest, measured as the share of women participating in the labour force within each education group. FLFP rates are survey-weighted estimates from the LFS.

FLFP rates are drawn from published annual LFS reports of the Department of Census and Statistics (DCS), Sri Lanka. These estimates are nationally representative and incorporate complex survey weights applied by the DCS. As the analysis uses published aggregate rates, no additional weighting adjustments are required at the regression stage.

The association between educational attainment and FLFP is estimated using a beta regression framework. This approach is appropriate for modelling proportions bounded between 0 and 1(Ferrari and Cribari-Neto, 2004). To allow for non-linear patterns in the education–FLFP relationship, orthogonal polynomial terms are included, enabling a formal assessment of curvature while mitigating multicollinearity (Greene, 2018).

Year fixed effects control for common macroeconomic or policy shocks affecting all education categories. Education–year interaction terms are estimated to assess whether the education gradient varies over the 2017–2024 period. Four nested specifications are compared: a linear model, a quadratic model, a temporal interaction model and a higher-order polynomial model, with model selection guided by the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC).

While the beta regression model estimates the conditional mean of FLFP, marginal effects and pairwise comparisons are presented to aid the interpretation of differences across education categories. These are derived from the estimated model and should be interpreted as descriptive contrasts rather than as separate estimates.

Given the small number of years of education observations, conventional cluster-robust standard errors may yield unreliable inference. To address this concern, the study uses wild-cluster bootstrap inference, which has been shown to yield more reliable p-values in settings with few clusters (Cameron et al., 2008). This approach improves the stability of statistical inference without imposing strong distributional assumptions.

A range of diagnostic checks, including tests for multicollinearity, functional form adequacy and residual behaviour, are conducted to assess model performance. Full model specifications, polynomial constructions, weighting procedures and diagnostic results are reported in  Appendix.

This study adopts an explicitly macro-level and exploratory analytical approach. The use of aggregated data precludes the inclusion of individual or household-level covariates such as marital status, fertility, income or regional characteristics. It also does not allow for causal identification. In particular, the analysis cannot distinguish between education as a determinant of labour force participation and selection into education based on anticipated employment outcomes. Nor can it address potential reverse causality or omitted-variable bias.

Accordingly, the results are interpreted as descriptive evidence of population-level patterns in the relationship between educational attainment and FLFP in Sri Lanka. The analysis is intended to characterise the functional form of this relationship over time rather than to estimate individual-level causal effects.

All statistical analyses were conducted using RStudio (version 2025.05.1 + 513). References were managed using Mendeley. Language-editing tools, including Grammarly and Microsoft Copilot, were used solely to assist with grammar and clarity; they were not used to generate, analyse or interpret any part of the research content. The authors retain full responsibility for the content, analysis and interpretation presented in this study.

Table 2 reports descriptive statistics for FLFP across five educational attainment levels in Sri Lanka over 2017–2024, based on 40 education–year observations (eight annual observations per education category). Mean FLFP differs substantially by educational attainment, ranging from 25.0% (SD = 3.5) among women with below Grade 5 education to 81.1% (SD = 1.8) among women with degree-level qualifications. Intermediate categories exhibit lower mean participation than tertiary education. There is a modest dip at the Ordinary Level (26.6%) relative to Grades 6–10 (29.0%), consistent with a non-linear education–FLFP pattern. The Advanced Level category shows the greatest year-to-year variability (SD = 4.1), suggesting relatively greater temporal sensitivity in FLFP for this group.

These summary statistics are descriptive and based on aggregated group-level rates; they do not adjust for changes in group composition or for individual- and household-level factors (e.g. marriage, children, household income or urban–rural location). As such, they provide an initial indication of the education–FLFP pattern rather than evidence of micro-level behavioural mechanisms.

The association between educational attainment and FLFP in Sri Lanka follows a positive quadratic functional form with a positive linear component, implying a convex (J-shaped) pattern rather than the U-shaped relationship commonly reported in the South Asian literature. To assess whether the education–FLFP association conforms to a convex or U-shaped pattern, beta regression models with orthogonal polynomial terms were estimated (Table 3). This modelling strategy is appropriate for proportion outcomes bounded between zero and one and allows the functional form of the relationship to be evaluated flexibly while limiting multicollinearity among polynomial terms. Given the small number of clusters (five education categories), conventional clustered standard errors are unreliable. Accordingly, p-values are based on wild-cluster bootstrap inference, clustered at the education-category level (five clusters; B = 999), which yields more reliable inference in few-cluster settings (Cameron et al., 2008). The standard errors reported in Table 3 are clustered at the education level.

While the beta regression model estimates the conditional mean of FLFP, marginal effects and adjacent-category comparisons are reported in subsequent sections to facilitate substantive interpretation of differences across education levels. These are derived from the estimated model and should be interpreted as descriptive contrasts rather than independent estimations.

Model selection is guided by the AIC and BIC, where lower values indicate a better fit relative to model complexity. Pseudo-R² is reported descriptively but is not used for specification comparison.

The baseline linear model (Model 1) shows a positive, statistically significant association between education and FLFP (β = 4.964, p < 0.001; AIC = −52.704), indicating higher participation at higher levels of education. However, the linear specification imposes a constant gradient across categories. Introducing a quadratic term (Model 2) substantially improves model fit (AIC = −128.963; BIC = −110.385), with both linear (β = 4.961, p < 0.001) and quadratic (β = 2.826, p < 0.001) components significant. The positive quadratic coefficient indicates a convex pattern, consistent with a J-shaped relationship.

Model 3 incorporates education–year interactions to test temporal variation. It performs worse than both Models 1 and 2 (AIC = −38.908), suggesting that the additional parameters do not meaningfully improve the model’s fit. This result indicates that the convex education–FLFP gradient is broadly stable over the 2017–2024 period.

The full polynomial specification (Model 4), which adds a cubic term, yields the lowest AIC and BIC values (AIC = −172.874). Although the cubic term is statistically significant (β = 1.003, p < 0.001), the small sample of 40 grouped observations raises concerns about overfitting. Model 4 is therefore treated as a robustness check. As it preserves the convex structure identified in Model 2, the quadratic specification is retained as the preferred parsimonious model.

The marginal effects reported in Table 4 are derived from the estimated beta regression model and provide an interpretable representation of the non-linear relationship between educational attainment and FLFP. These effects reflect model-based adjacent-category differences averaged over the 2017–2024 period and should be interpreted as descriptive contrasts rather than independent estimates.

The estimated marginal effects indicate a clear convex (J-shaped) trajectory in the education–FLFP relationship. Based on the model, the transition from Below Grade 5 to Grades 6–10 is associated with a modest increase in participation (+4.0 percentage points, p < 0.001), suggesting that lower secondary education yields limited labour market returns for women in Sri Lanka. The transition from Grades 6–10 to Ordinary Level is associated with a small but statistically significant decline (−2.4 percentage points, p < 0.001), indicating that completing lower secondary education does not translate into a participation premium and may reflect a period during which women remain in education rather than entering the labour force.

The pattern changes markedly at higher education thresholds. Model-based estimates show that the transition from Ordinary Level to Advanced Level is associated with a substantial increase in participation (+18.1 percentage points, p < 0.001), while the largest marginal effect occurs at the transition from Advanced Level to Degree and above (+36.3 percentage points, p < 0.001). These results indicate that the returns to education, in terms of FLFP, are heavily concentrated at the upper end of the educational distribution.

Taken together, these model-derived marginal differences are consistent with the estimated quadratic specification, providing evidence of a convex (J-shaped) relationship between education and FLFP. A U-shaped trajectory would imply a more pronounced rebound following mid-level decline and a relative flattening at higher education levels, a pattern not supported by the estimated marginal effects. Instead, the evidence points to a threshold effect, whereby substantial participation gains materialise primarily beyond Advanced Level completion, with tertiary education generating the strongest returns.

Diagnostic checks in Table 5 are based on the education-only quadratic specification, which excludes year fixed effects and therefore differs from Model 2 in Table 3; the small difference in AIC values (−128.37 versus −128.963) reflects this specification difference rather than any inconsistency in estimation. This education-only specification is used for diagnostics because it isolates the education–FLFP relationship without the additional parameters introduced by year fixed effects, providing a cleaner basis for assessing model fit and residual behaviour. The quadratic specification indicates a strong overall fit, with a log-likelihood of 68.186, AIC of −128.37, BIC of −121.62 and a pseudo-R² (Ferrari and Cribari-Neto, 2004) of 0.946. Residual analysis shows no concerning outliers: the Pearson residual standard deviation is 1.03; only one observation exceeds the ± 2 threshold; and none exceed ± 3. The Shapiro–Wilk test yields p = 0.835, consistent with normality of the residuals. The likelihood-ratio test comparing the linear and quadratic specifications yields LR = 67.17 (df = 1, p < 0.001), indicating that the quadratic term yields a statistically significant improvement in fit and supports the presence of curvature in the education–FLFP relationship. The comparison of quadratic and cubic specifications also yields a significant LR statistic (LR = 31.85, df = 1, p < 0.001), though the cubic model is treated as a robustness check given concerns about overfitting in a sample of 40 observations. Wild cluster bootstrap resampling (B = 999, Webb weights, G = 5 clusters) yields a quadratic coefficient estimate of 2.885 with a 95% CI of [−0.28, 6.05], which is marginally significant at p = 0.079. The wider interval reflects the limited number of education clusters and the known conservatism of cluster-robust inference with few clusters. The curvature finding is therefore interpreted in conjunction with the LR test evidence rather than solely on the bootstrap. Taken together, the results across specifications and diagnostics are consistent with a convex (J-shaped) association between education and FLFP.

After accounting for non-linearities, the transition from lower secondary education (Grades 6–10) to Ordinary Level is associated with a small negative difference in FLFP.

Table 6 reports adjacent-category pairwise comparisons of FLFP across educational transitions for the period 2017–2024. The transition from Below Grade 5 to Grades 6–10 is associated with a modest positive difference of + 4.0 pp (SE = 0.55, t = 7.30, p < 0.001, 95% CI [+2.7, + 5.3]), indicating a small but statistically significant participation premium at the lower secondary level. The transition from Grades 6–10 to Ordinary Level is associated with a small negative difference of −2.4 pp (SE = 0.24, t = −9.80, p < 0.001, 95% CI [−3.0, −1.8]), indicating a modest decline in participation at the Ordinary Level relative to lower secondary education. While conventional t-statistic–based inference suggests statistical significance, wild cluster bootstrap procedures with five education clusters yield a p -value close to conventional significance thresholds (p ≈ 0.052), indicating that evidence of a mid-level decline is statistically sensitive to small-cluster adjustments and not robust at conventional 5% significance levels.

The pattern changes substantially at higher educational thresholds. The transition from Ordinary Level to Advanced Level is associated with a large positive difference of + 18.1 pp (SE = 1.20, t = 15.15, p < 0.001, 95% CI [+15.3, + 21.0]), and the largest premium occurs at the Advanced Level-to-degree transition (+36.3 pp, SE = 1.17, t = 31.08, p < 0.001, 95% CI [+33.6, + 39.1]). Together, these findings provide evidence that participation gains are heavily concentrated at the upper end of the educational distribution, consistent with the convex trajectory identified in the beta-regression and marginal-effects analyses.

Consistent across Table 4 (marginal effects) and Table 6 (pairwise differences), the sharpest rise in FLFP is observed between Advanced Level and Degree and above (+36.3 percentage points; t = 31.08, p < 0.001). The next-largest gain occurs between Ordinary Level and Advanced Level (+18.1 pp), while the increase from Below Grade 5 to Grades 6–10 is relatively small (+4.0 pp). Taken together, these differences suggest that the education–FLFP relationship in Sri Lanka is characterised by gains that are heavily concentrated at higher educational transitions.

One plausible interpretation, consistent with the development and labour market literature, is that Sri Lanka’s labour demand structure and institutional pathways may better absorb highly educated women than moderately educated women, particularly through professional and public-sector employment channels. However, given the aggregated nature of the data, this study does not directly test mechanisms such as occupational sorting, household constraints or employer-side selection.

The empirical results reveal a clear non-linear (J-shaped) relationship between educational attainment and FLFP in Sri Lanka. While the U-shaped hypothesis predicts a pronounced mid-level withdrawal from the labour market followed by re-entry at higher levels of education (Goldin, 1995; Najeeb et al., 2020), the Sri Lankan case exhibits a more moderate decline at intermediate levels and a substantially stronger increase at higher levels of education. FLFP rises modestly from low education to lower secondary and declines slightly at the Ordinary Level (−2.4 percentage points). It then increases sharply at the Advanced Level (+18.1 percentage points) and tertiary education (+36.3 percentage points), resulting in a 56-percentage-point gap between the lowest and highest categories. Importantly, the mid-level decline is statistically fragile under few-cluster inference, reinforcing a convex (J-shaped) rather than a pronounced U-shaped pattern.

These findings are partially consistent with existing literature that emphasises the role of structural transformation and labour demand conditions in shaping women’s labour market participation (Klasen et al., 2019; Jayachandran, 2015). In line with evidence from developing economies, the results indicate that access to skilled employment is a key determinant of participation among educated women.

At the same time, the findings diverge from the traditional U-shaped framework commonly observed in larger South Asian economies such as India (Klasen and Pieters, 2015). Unlike the pronounced mid-education withdrawal documented in these contexts, Sri Lanka displays a more gradual transition, with participation gains heavily concentrated at higher levels of education. This pattern indicates that the education–FLFP relationship is not universal but is structurally contingent on the alignment between educational attainment and labour market demand. Where labour markets generate limited opportunities for moderately educated women, but strong demand for highly educated workers, participation becomes concentrated at the upper end of the educational distribution. This suggests that the U-shaped hypothesis may not fully capture the dynamics of smaller, service-oriented economies.

More broadly, the findings may reinforce the argument that the education–FLFP relationship is inherently context-dependent, shaped by the interaction between educational expansion and labour market structure. Where structural transformation generates demand for skilled female labour, particularly in service-oriented sectors, education may function as a progressive enabler of participation rather than as an inducement to withdrawal (Heath and Jayachandran, 2018).

The findings of this study have important implications for employers and labour market practitioners in Sri Lanka. The high level of FLFP among those with higher levels of education suggests that organisations in service-oriented sectors such as healthcare, education and ICT are well positioned to benefit from the growing pool of highly educated female talent. Expanding flexible work arrangements, including remote and hybrid models, may further enhance participation among educated women, particularly amid increasing digitalisations.

At the same time, the presence of a “missing middle” highlights the need for organisations to develop employment pathways for moderately educated women. This may include expanding mid-skill roles, vocational training partnerships and structured career progression frameworks. Firms can also support women’s participation through workplace policies such as childcare provision, safe transport arrangements and inclusive work environments.

Overall, these findings suggest that organisational practices play a complementary role alongside policy interventions in translating educational attainment into sustained labour market engagement.

Sri Lanka’s strong tertiary premium reflects a close alignment between education and service-sector labour demand, particularly in healthcare, education, administration and ICT (Department of Census and Statistics; Sri Lanka, 2024). By contrast, the modest gains at lower and intermediate levels of education indicate that schooling alone does not guarantee participation when suitable employment opportunities are limited.

The evidence also reveals a “missing middle”, where participation among women with Ordinary Level education remains below that of lower secondary education. This pattern aligns with existing evidence on structural labour market constraints in Sri Lanka (Gunatilaka, 2013), but the present findings provide new aggregate-level evidence that this gap reflects a mismatch between educational attainment and labour market opportunities. Women who complete Ordinary Level but do not progress further may find themselves overqualified for informal work and underqualified for formal-sector employment. This creates a participation trap that educational expansion alone cannot resolve.

These findings are also consistent with broader regional evidence emphasising the role of labour demand constraints (Afridi et al., 2024) and extend this literature by demonstrating how such constraints operate across education levels in a service-oriented economy. Complementary investments in childcare, safe transport and the formalisation of mid-skill occupations may therefore be necessary to convert educational attainment into sustained labour-market engagement.

More broadly, development strategies should align education policy with sectoral labour demand. The stability of the convex education–FLFP gradient across the 2017–2024 period, including during Sri Lanka’s economic crisis, suggests that the observed pattern reflects a structural feature of the labour market rather than a cyclical outcome. This may reinforce the importance of long-term policy interventions that prioritise female-intensive sectors as drivers of inclusive growth.

Sri Lanka’s J-shaped pattern provides important insights into the theoretical frameworks underpinning the education–FLFP relationship. Firstly, the findings support Human Capital Theory (Becker, 1964), demonstrating that higher levels of education increase women’s participation where labour markets generate demand for skilled labour.

Secondly, the results refine Dual Labour Market Theory (Doeringer and Piore, 1971) by showing that segmentation is not purely binary. Moderately educated women may remain excluded from both informal and formal sectors, creating a “missing middle” that reflects a gradient of access shaped by labour demand and qualification thresholds.

Thirdly, the findings extend Role Congruity Theory (Eagly and Karau, 2002; Wood and Eagly, 2012) by suggesting that increased participation in professional roles may contribute to gradual normative change, although this mechanism operates over longer time horizons than captured in the present analysis.

Taken together, these findings highlight the importance of integrating economic, structural and normative perspectives. The results support a context-dependent interpretation of the education–FLFP relationship, in which human capital accumulation, labour market structure and gender norms interact to produce non-linear participation outcomes.

Using aggregated LFS data (2017–2024) and beta regression methods, this study identifies a convex (J-shaped) relationship between educational attainment and FLFP in Sri Lanka. The modest mid-level decline contrasts with the pronounced withdrawal predicted by the U-shaped hypothesis, while sharp increases at higher education levels reflect strong alignment between education and labour demand.

The findings demonstrate that the education–FLFP relationship is fundamentally shaped by structural and institutional context. While higher education enhances participation, the persistence of a “missing middle” highlights the importance of complementary labour market opportunities and support systems. Sri Lanka’s experience illustrates that educational expansion alone is insufficient to generate inclusive labour market outcomes without corresponding demand-side and institutional adjustments.

This analysis relies on aggregated education–year data and therefore cannot establish causal mechanisms. The small number of education clusters limits the precision of inference despite appropriate bootstrap procedures, and compositional changes within categories may influence observed patterns. Individual-level microdata is needed to assess mechanisms such as occupational sorting, public-sector absorption and care-related constraints.

Future research should examine micro-level determinants and employment pathways to better understand the drivers of participation across education levels, and to distinguish long-term structural dynamics from short-term fluctuations. The analysis does not distinguish cohort effects or generational shifts in FLFP, which may influence observed patterns over time. Future research using cohort-disaggregated or panel data could provide further insight.

A.1 model specification

Let $FLFPit∈(0,1)$denote the female labour force participation rate for the education level $i$ in year $t$⁠. Because FLFP is a proportion bounded between zero and one, models are estimated using beta regression with a logit link function:

where $μit$denotes the conditional mean of FLFP and $ϕ$is the precision parameter.

Four nested specifications are estimated:

Model 1 (Linear): Includes first-order education term (O1).

Model 2 (Quadratic): Adds second-order term (O2) to test for curvature.

Model 3 (Temporal Interaction – Linear): Allows the linear education gradient to vary over time.

Model 4 (Full Model): Includes second- and third-order terms and temporal interactions.

Year fixed effects control for unobserved macroeconomic or institutional shocks common across education groups.

Convexity is assessed through the quadratic coefficient $β2$⁠. A positive and statistically significant $β2$⁠, combined with a positive linear term, implies convexity in the education–FLFP relationship. Given low baseline participation at intermediate levels, this convexity empirically corresponds to a J-shaped pattern rather than the symmetric U-shaped relationship predicted by standard theory.

A.2 Construction of orthogonal polynomials

Educational attainment is coded as an ordinal variable:

0 = Below grade 5

1 = Grade 6–10

2 = Ordinary level

3 = Advanced level

4 = Degree and above

To model non-linearities while avoiding multicollinearity, first-, second-, and third-order orthogonal polynomial terms were generated using standard orthogonalisation procedures implemented in statistical software. By construction, these terms are mutually orthogonal (uncorrelated in the sample), which stabilises coefficient estimates relative to raw polynomial specifications.

The “no schooling” (illiterate) category is excluded from the analysis. In Sri Lanka, this group constitutes less than 1% of the female population, resulting in very small education–year cell sizes and correspondingly high sampling variance in FLFP estimates. Excluding this category improves statistical precision and avoids undue influence from imprecisely estimated observations without materially affecting the overall education–FLFP relationship.

A.3 inference and robustness

The regression is based on 40 education–year observations (5 education groups × eight years). However, with only five education clusters, conventional cluster-robust standard errors may be downward-biased.

Inference is therefore based on wild-cluster bootstrap p-values clustered by education level (B = 999 replications).

Model robustness is assessed through:

• Likelihood ratio tests comparing polynomial orders.

• Information criteria (AIC/BIC).

• Residual-versus-fitted plots.

• Checks for influential observations.

All diagnostics indicate stable model specification and no evidence of severe misspecification. Given only five education clusters (G = 5), Webb weights are used in the wild-cluster bootstrap to improve reliability at small cluster counts (MacKinnon et al., 2023).

A.4 interpretation

Coefficients represent effects on the log-odds scale. For interpretability, results are converted to marginal effects. Statistical significance is evaluated at the 5% level using wild cluster bootstrap p-values.