Human capital in a fragile state: an econometric and machine learning analysis of development aid, humanitarian assistance, conflict, and economic conditions in Somalia

1 Introduction

Understanding how human capital evolves in fragile and aid-dependent economies has become a central concern in development research (Andaish and Assadi, 2024; Seyoum, 2020). Early development thinking focused largely on capital accumulation and structural change, but over time attention shifted toward the quality of human resources as a key driver of economic performance (Li and Gao, 2025; Wang et al., 2022). Improvements in health, skills, and participation in productive activities are now widely recognized as

both outcomes of development and inputs into growth (Ding et al., 2021; Fufa, 2021). Yet these improvements rarely occur in isolation; they are shaped by the interaction of domestic conditions, external support, and structural constraints, particularly in contexts characterized by instability.

Within this broader setting, foreign aid has played a major role in shaping development trajectories worldwide (Ullah et al., 2025). Development aid emerged after the Second World War as a long-term instrument aimed at rebuilding economies, strengthening institutions, and supporting structural transformation (Grant, 1979). Over time, it expanded into sector-focused programs, especially in health and education, with the intention of improving human capital and productivity (Groß and Nowak-Lehmann Danzinger, 2022; Niño-Zarazúa, 2016). Humanitarian aid, by contrast, developed as a response to emergencies, focusing on immediate relief during conflicts, disasters, and economic shocks (Arcos González and Gan, 2024). While development aid is generally associated with long-term capacity building, humanitarian assistance is designed to preserve wellbeing in the short term. Across regions, both forms of aid have contributed to improvements in health and living standards, though their effects have varied depending on institutional quality and the broader policy environment (Andaish and Assadi, 2024).

In Africa, the interaction between aid, human capital, and economic performance has been particularly pronounced. Many countries have relied heavily on external assistance to support health systems, education, and social protection (Nonvignon et al., 2024). While this support has contributed to gains in life expectancy and access to services, concerns remain regarding sustainability, institutional dependence, and uneven outcomes across countries (Warsame et al., 2025). These issues are most evident in fragile and conflict-affected states, where governance constraints and instability shape how resources are allocated and utilized. As a result, the relationship between aid and human capital reflects a combination of supportive and limiting influences rather than a consistent pattern. From a dependency and fragile-state governance perspective, aid effectiveness depends not only on the amount of assistance received but also on whether external resources strengthen domestic institutions, improve state capacity, and support locally owned systems (Addison and Tarp, 2015). When aid is fragmented, externally managed, or weakly integrated into national structures, it may deliver short-term services while limiting longer-term institutional consolidation (Addison and Tarp, 2015; Gehring et al., 2017). Consequently, development aid, humanitarian aid, conflict, and economic conditions should not be viewed as isolated influences but as interconnected factors that jointly shape human capital outcomes in fragile environments.

Somalia offers a compelling context for examining these dynamics. Decades of conflict, institutional disruption, and recurrent shocks have constrained human capital development, while at the same time the country has received substantial development and humanitarian assistance (Hassan et al., 2026; Warsame et al., 2025). In such an environment, external support plays a dual role. It helps sustain basic services and livelihoods, yet its interaction with weak institutions and ongoing instability can influence longer-term outcomes. Conflict further complicates this relationship by disrupting service delivery, displacing populations, and limiting participation in economic activity (Nor and Raheem, 2025). These overlapping conditions make Somalia a relevant setting for exploring how external assistance, instability, and economic conditions jointly shape human capital over time. This makes Somalia a particularly important case for examining how aid, fragility, conflict, and economic conditions interact in shaping human capital. The country's prolonged dependence on external assistance, combined with persistent instability and institutional constraints, provides a unique setting for evaluating both the benefits and limitations of external support over time.

A growing body of research has explored these relationships from different angles. Studies such as Asaleye et al. (2023) and Wang et al. (2024) highlight the potential of development aid to improve sectoral outcomes, particularly in health and employment, while others, including Bethencourt and Perera-Tallo (2025) and Atitianti and Asiamah (2023), point to institutional challenges associated with sustained aid dependence. Evidence from Somalia, including Warsame et al. (2025) and Mohamed et al. (2025). provides insights into how aid and macroeconomic factors interact within a fragile setting, while Hussein et al. (2026) emphasizes governance and sustainability concerns. At the same time, the literature on humanitarian assistance, such as Morrison and Malik (2024) and Akpoghelie et al. (2024), underscores its role in maintaining human wellbeing during crises. Research on conflict, including Nor and Yusof (2025) and Hidalgo-Aréstegui et al. (2025), consistently documents its adverse effects on human development. Despite these contributions, much of the existing work considers these factors separately or focuses on specific sectors, making it difficult to capture their combined influence over time. Despite these contributions, three important limitations remain in the literature. First, many studies examine development aid, humanitarian aid, conflict, or economic conditions separately, providing limited evidence on their combined influence on human capital. Second, relatively little attention has been given to differences between short-run and long-run effects, despite growing evidence that the impact of aid may change over time. Third, most existing studies rely on a single analytical framework, offering limited insight into whether findings remain consistent across econometric, machine-learning, and Bayesian approaches. These gaps are particularly relevant in fragile settings such as Somalia, where institutional constraints and data limitations complicate empirical analysis.

Building on these insights, this study examines how development aid, humanitarian aid, conflict, and economic conditions interact in shaping human capital in Somalia over the period 1990–2023. The analysis brings these elements together within a unified framework and considers both short-term adjustments and longer-term patterns. The study is informed by Human Capital Theory, Dependency Theory, and fragile-state governance perspectives, which together explain why aid, conflict, and economic conditions may influence human capital differently across time horizons. In doing so, it reflects the mixed nature of development aid across different time horizons, the generally supportive role of humanitarian assistance in crisis contexts, the persistent adverse influence of conflict, and the less consistent relationship between economic growth and human capital.

By combining econometric, machine learning, and Bayesian approaches, the study provides a broader perspective on these relationships than is typically offered by single-method analyses. The ARDL model is used to examine short-run and long-run relationships, Random Forest is used to identify the relative importance of predictors, and Bayesian regression is used to account for parameter uncertainty. This integrated approach allows for the examination of dynamic interactions, the identification of influential factors, and the consideration of uncertainty in a data-constrained environment. The evidence generated contributes to a clearer understanding of how external support and internal conditions interact in fragile settings, offering insights that are relevant for policy discussions aimed at improving health outcomes, strengthening labor participation, and supporting human capital development over time.

1.1 Literature review

Several decades ago, development scholars began to question whether external assistance could reliably support long-term improvements in human welfare, particularly in fragile and low-income countries (Burnside and Dollar, 2000; Easterly, 2003, 2007). Early discussions often treated foreign aid as a direct instrument for enhancing human capital through investments in health, education, and infrastructure. However, as empirical evidence expanded across different contexts, it became increasingly clear that the relationship between aid and human capital is not uniform. Instead, it reflects a dynamic process shaped by time horizons, institutional conditions, and the structure through which resources are delivered and utilized.

In the case of development aid, the literature presents a notably mixed and evolving perspective. Some studies emphasize its potential to strengthen human capital, particularly when aid is directed toward specific sectors such as health and education. Evidence suggests that targeted interventions can expand healthcare capacity, improve service delivery, and contribute to long-term improvements in human development indicators (Asaleye et al., 2023; Wang et al., 2024). These findings support the view that development aid can play a constructive role when it is effectively aligned with national priorities and embedded within domestic systems.

At the same time, a substantial body of research highlights the limitations of development aid, particularly over longer periods. Several studies argue that sustained reliance on external assistance may alter domestic incentives, weaken institutional development, and reduce the effectiveness of public systems (Atitianti and Asiamah, 2023; Bethencourt and Perera-Tallo, 2025). In fragile settings such as Somalia, these concerns are further amplified. Evidence points to issues such as aid fragmentation, short-term project orientation, and parallel implementation structures, which can undermine state capacity and limit sustainability (Hussein et al., 2026; Mourad, 2023). This dual perspective suggests that development aid may produce different outcomes depending on the time frame considered, with short-term support potentially differing from longer-term structural effects.

These contrasting findings suggest that the effects of development aid may evolve over time. In the short run, aid can expand access to health services, finance essential programs, and support basic infrastructure, thereby contributing to improvements in human wellbeing (Leunig et al., 2024; Mohamed et al., 2025). Over longer periods, however, the effectiveness of aid may depend on institutional capacity and the extent to which external resources are integrated into domestic systems (Ehigiamusoe et al., 2025; Hussein et al., 2026). Where aid operates through parallel structures, weak coordination mechanisms, or prolonged external dependence, short-term gains may not translate into sustainable improvements. This perspective helps explain why development aid may generate positive immediate outcomes while producing less favorable long-term effects in fragile environments.

The literature on humanitarian aid presents a somewhat more consistent perspective, particularly in crisis-affected environments. Humanitarian assistance is designed to address immediate needs, and as such, it plays a critical role in preserving human capital under conditions of instability. Empirical evidence shows that humanitarian interventions can improve access to healthcare, reduce malnutrition, and support vulnerable populations during periods of crisis (Akpoghelie et al., 2024; Morrison and Malik, 2024). In contexts such as Somalia, where shocks are frequent and institutional capacity is limited, humanitarian aid often serves as a key mechanism for sustaining basic human development outcomes.

Nevertheless, the longer-term implications of humanitarian aid remain subject to debate. While it provides essential support in the short term, some studies note that its effectiveness can be influenced by broader governance challenges and coordination issues (Hussein et al., 2026). Despite these concerns, the overall body of evidence suggests that humanitarian aid tends to play a supportive role in maintaining human capital, particularly in fragile and conflict-affected settings.

Although both humanitarian aid and development aid involve external resource transfers, they differ in purpose and time horizon. Humanitarian aid is primarily intended to protect lives and maintain basic wellbeing during crises, whereas development aid seeks to support long-term structural improvements in economic and social conditions (Elayah, 2025; Kuijt, 2015; McGillivray et al., 2006). Consequently, humanitarian assistance is generally evaluated in terms of its ability to preserve existing human capital, while development aid is assessed according to its contribution to sustainable human capital accumulation (Githu et al., 2026; Rustad et al., 2020). Distinguishing between these objectives is particularly important in fragile settings where emergency needs and long-term development priorities often coexist.

Beyond external assistance, the literature consistently identifies conflict as one of the most significant constraints on human capital development. A wide range of studies demonstrate that conflict disrupts access to education, healthcare, and economic opportunities, leading to persistent declines in human capital. Evidence from Somalia and other conflict-affected regions shows that instability contributes to displacement, weakens service delivery systems, and increases vulnerability (Ahmed et al., 2024; Nor and Yusof, 2025). Moreover, the effects of conflict often extend across generations, influencing not only current outcomes but also future human capital formation through its impact on education and social development (Hidalgo-Aréstegui et al., 2025). These findings reinforce the view that conflict exerts a consistently adverse influence on human capital.

The relationship between economic growth and human capital has also received considerable attention, though with less consensus. Some studies argue that higher income levels improve access to essential services and create conditions that support human capital development (Chang et al., 2024; Tahir and Burki, 2023). Others, however, suggest that the impact of growth is conditional and may depend on how economic gains are distributed and whether they translate into investments in health and social infrastructure (Büttner et al., 2023; Patterson, 2023). In certain contexts, economic growth alone may not be sufficient to generate meaningful improvements in human capital without complementary policy interventions.

Taken together, the literature suggests that the relationship between external assistance, conflict, economic conditions, and human capital is complex and context-dependent. Development aid presents a dual perspective, with outcomes that may vary across time horizons and institutional settings. Humanitarian aid generally plays a supportive role in maintaining human capital during crises, while conflict consistently undermines human development outcomes. The role of economic growth appears less definitive, reflecting variations in structural conditions and policy environments. These insights highlight the importance of examining these factors jointly within a unified empirical framework, particularly in fragile contexts such as Somalia, where their interactions are likely to be more pronounced.

2 Methodology

This study is grounded in human capital theory, which explains that investments in health and productive capacity enhance individual productivity and support long-term economic development (Barro, 1991; Lucas, 1988; Schultz, 1961). Human capital is not only accumulated through improvements in health and skills but is also reflected in labor market participation. This distinction is particularly relevant in fragile and conflict-affected contexts such as Somalia, where economic disruptions often limit the effective use of human capital. To capture both the capacity and utilization dimensions, the study constructs a Human Capital Index (HCI) by combining life expectancy and labor force participation (Becker, 1962). Life expectancy reflects the health component of human capital, which is closely linked to productivity and economic performance (Cui et al., 2025; Fumagalli et al., 2024; Lyeonov et al., 2025), while labor force participation captures the extent to which human capital is actively engaged in economic activity (Ogbuabor et al., 2025; Umair et al., 2024).

The construction of the HCI is guided by both theoretical considerations and data availability. While global measures such as the World Bank Human Capital Index include education components, consistent long-term education data for Somalia remain limited. Although education is traditionally regarded as a core component of human capital (Becker, 1962; Schultz, 1961), the literature also recognizes health as a central dimension of human capital and commonly uses life expectancy as a proxy for health status (Barro, 2001; Bleakley, 2010). More recent human-capital measurement frameworks also emphasize employment as part of the realization of human capital over the working life (World Bank, 2024). Therefore, in the absence of consistent long-term education data for Somalia, this study focuses on two measurable dimensions: health, represented by life expectancy, and utilization, represented by labor force participation. . Therefore, combining life expectancy and labor participation provides a practical and context-appropriate proxy that reflects both the level and utilization of human capital. This approach aligns with the literature, which supports the use of alternative indicators when data constraints exist, particularly in fragile and data-scarce environments.

The study employs annual time-series data for Somalia covering the period 1990–2023, allowing for the analysis of both short-term fluctuations and long-term structural relationships. This period captures Somalia's experience of prolonged conflict, institutional fragility, and gradual recovery, providing a relevant context for examining the evolution of human capital under conditions of instability and external support. The Human Capital Index is constructed using normalized indicators to ensure comparability across variables measured in different units. The normalization procedure is expressed as: where X_*it* represents the observed value of a variable at time t, while X_min and X_max denote the minimum and maximum values within the dataset. The normalized values are then combined to form the Human Capital Index as: Where, LEI_*t* and LFPI_*t* represent the normalized life expectancy and labor force participation indices. Development aid and humanitarian aid are measured in constant US dollars, conflict is proxied by the number of battle-related deaths, and GDP per capita represents overall economic capacity.

Zit=Xit-XminXmax-Xmin

HCIt=LEIt+LFPIt2

To examine the relationship between human capital and its determinants, the study adopts a dynamic modeling framework that integrates econometric, machine learning, and Bayesian approaches. The purpose of combining these methods is methodological triangulation rather than treating them as competing substitutes. The ARDL model serves as the primary framework for examining short-run and long-run relationships. Random Forest is used as a complementary predictive tool to assess variable importance and possible nonlinear patterns, while Bayesian regression is used to incorporate parameter uncertainty and provide a probabilistic robustness check. Given the small sample size, the Random Forest and Bayesian results are interpreted cautiously and used to support, rather than replace, the main ARDL evidence. The baseline empirical model is specified as: where HCI_*t* is the dependent variable, and DA_*t*, HA_*t*, CF_*t*, and Y_*t* represent development aid, humanitarian aid, conflict, and GDP per capita, respectively. Before estimation, stationarity of the variables is tested using the following specification: This ensures that the variables meet the requirements of the Autoregressive Distributed Lag (ARDL) model, which allows for a mixture of level and first-difference stationary variables. Variables integrated at second difference, I(2), would invalidate the ARDL bounds testing procedure and are therefore excluded from the modeling framework.

HCIt=β0+β1DAt+β2HAt+β3CFt+β4Yt+εt

ΔYt=α+δYt-1+∑i=1pγiΔYt-i+ut

The ARDL model is employed as the primary econometric framework due to its ability to capture both short-run dynamics and long-run equilibrium relationships. It is also appropriate for relatively small samples and mixed integration orders, which makes it suitable for annual time-series data in fragile and data-constrained contexts. The lag length is selected using standard information criteria, particularly the Akaike Information Criterion (AIC), with the maximum lag restricted to preserve degrees of freedom. In this study, lag length 2 is selected because it provides the best balance between model fit and parsimony. The general ARDL specification is given as: To distinguish between short-run and long-run effects, the model is expressed in an error correction form: The coefficient λ measures the speed of adjustment toward long-run equilibrium following short-term deviations. A negative and statistically significant error correction coefficient confirms convergence toward the long-run equilibrium path. Before interpreting the ARDL estimates, additional validation checks are conducted. Multicollinearity is examined using the variance inflation factor (VIF), since aid variables and GDP per capita may be correlated. The VIF is expressed as: where Rj2 is obtained by regressing each explanatory variable on the remaining predictors. Values below common thresholds such as five or 10 indicate that multicollinearity is not severe. The study also acknowledges possible endogeneity and reverse causality, particularly because aid inflows may respond to deteriorating human capital conditions. Therefore, the estimated coefficients are interpreted as dynamic associations rather than definitive causal effects.

HCIt=α0+∑i=1pαiHCIt-i+∑j=0q1βjDAt-j+∑k=0q2ϕkHAt-k+∑m=0q3ψmCFt-m+∑n=0q4ωnYt-n+εt

ΔHCIt=...+λECTt-1+εt

ECTt-1=HCIt-1-θ1DAt-1-θ2HAt-1-θ3CFt-1-θ4Yt-1

VIFj=11-Rj2

To complement the econometric analysis, Random Forest regression is employed to capture potential nonlinear relationships and assess variable importance: Where, B represents the number of decision trees and f_*b*(X_*t*) denotes the prediction from each tree based on the explanatory variables. Because the sample consists of 34 annual observations, Random Forest is not treated as a standalone predictive model. To reduce overfitting concerns, the model is evaluated using a simple train-test split and cross-validation where feasible, with hyperparameters such as the number of trees, maximum tree depth, and minimum leaf size selected to preserve parsimony. The results are interpreted primarily through variable importance rather than causal effects.

HCI^t=1B∑b=1Bfb(Xt)

In addition, Bayesian regression is applied to incorporate uncertainty into the estimation process and provide probabilistic interpretation: This framework combines prior information with observed data to generate posterior distributions of the parameters. Weakly informative priors are used to allow the data to drive the estimates while preventing unstable parameter values in a small-sample setting. For example, the regression coefficients are assigned the following prior: The Bayesian model is estimated using Markov Chain Monte Carlo simulation, with multiple chains and sufficient iterations to assess convergence. Posterior predictive checks are also used to assess whether the Bayesian model adequately reproduces the observed data patterns.

p(β∣Y)∝p(Y∣β)p(β)

βj~N(0,102)

Model diagnostics are conducted to ensure the reliability and validity of the estimated results. Serial correlation is tested using: Heteroskedasticity is examined using: Normality of residuals is assessed using: Model stability is evaluated using: Cointegration is tested using the bounds approach: In this specification, the joint significance of the lagged level variables determines whether a long-run equilibrium relationship exists among human capital and its determinants.

ε^t=α0+α1Xt+∑i=1pρiε^t-i+vt

ε^t2=α0+α1DAt+α2HAt+α3CFt+α4Yt+ut

JB=n6(S2+(K-3)2 4)

CUSUMt=∑i=k+1twiσ^w

CUSUMSQt=∑i=k+1twi2∑i=k+1Twi2

ΔHCIt=α0+∑βiΔXt-i+δ1HCIt-1+δ2DAt-1+δ3HAt-1+δ4CFt-1+δ5Yt-1+εt

Bayesian diagnostics focus on convergence and sampling efficiency: where R^ close to one indicates convergence, and higher ESS values indicate more reliable posterior sampling. The Bayesian estimation is implemented using statistical software capable of MCMC estimation, while the ARDL and Random Forest analyses are conducted using standard econometric and machine-learning packages.

R^=V^W

ESS=S1 + 2∑k=1∞ρk

Model evaluation is based on performance metrics such as RMSE, MAE, and R², allowing comparison between the ARDL, Random Forest, and Bayesian approaches. While the ARDL model provides the main basis for interpretation, the other methods serve as complementary tools that enhance robustness. As illustrated in Figure 1, this integrated methodological structure combines econometric, machine learning, and Bayesian techniques within a unified framework. This approach improves both analytical reliability and policy relevance, particularly in a fragile context such as Somalia where uncertainty and structural challenges are significant.

3 Empirical analysis and multi-model evidence

3.1 Summary of the data

The descriptive statistics in Table 1 summarize the key characteristics of the variables used in the analysis. The Human Capital Index (HCI) has a mean close to zero due to its standardized construction, with moderate variation over time. Development aid (DA) and humanitarian aid (HA) show high variability, reflecting uneven inflows, particularly for humanitarian assistance which responds to crisis conditions. Conflict (BRD) also exhibits substantial fluctuations, indicating periods of instability, while GDP per capita (Y) shows wide variation, including negative values during economic downturns. The skewness and kurtosis statistics suggest that most variables deviate from normal distribution, which is confirmed by the Jarque–Bera test results. Overall, the data reflect a volatile and unstable environment, supporting the use of dynamic modeling techniques in the analysis.

Table 1. Descriptives results.

3.2 Stationarity and lag selection test

The unit root test results in Table 2 indicate that the variables exhibit mixed orders of integration. The ADF and PP tests were conducted using an intercept specification. HCI, development aid (DA), and GDP per capita (Y) are stationary at first difference, suggesting they are integrated of order one, while humanitarian aid (HA) and conflict (CF) are stationary at level. The ADF and PP test results are consistent across variables, supporting the reliability of these findings. The presence of both I(0) and I(1) variables supports the application of the ARDL model, which is appropriate for handling such mixed integration properties. The lower part of Table 2 reports the lag selection results. Lag length 2 is selected by the LR, FPE, AIC, and HQ criteria, while the SC criterion selects lag 1. Since the majority of the criteria support lag 2, and because it provides a more appropriate dynamic specification for the ARDL model, lag length 2 is retained for the estimation.

Table 2. Unit root test and lag selection results.

The ADF and PP unit root tests were estimated using an intercept specification.

^*, ^**, and ^***indicate statistical significance at the 10%, 5%, and 1% levels, respectively. For lag selection, ^*indicates the lag selected by each criterion.

LR, sequential modified likelihood ratio test; FPE, final prediction error; AIC, Akaike information criterion; SC, Schwarz criterion; HQ, Hannan–Quinn criterion.

3.3 Diagnostic test

The diagnostic test results in Table 3 indicate that the model satisfies the main econometric assumptions. The heteroskedasticity and serial correlation tests are not statistically significant, suggesting that the residuals do not suffer from non-constant variance or serial correlation. The Jarque–Bera normality test also indicates that the residuals are approximately normally distributed. In addition, the multicollinearity results show that all centered VIF values remain below the conventional threshold of 10, indicating that severe multicollinearity is not present among the explanatory variables. However, some lagged variables have VIF values above five, suggesting moderate collinearity, which is expected in dynamic time-series models that include lagged regressors. These results support the reliability of the estimated model for inference.

Table 3. Diagnostic tests and multicollinearity results.

Heteroskedasticity and serial correlation are reported using their test statistics and p-values. Normality is reported using the Jarque–Bera statistic and p-value. VIF values below 10 indicate that severe multicollinearity is not present, while values above five suggest moderate collinearity.

3.4 Cointegration test

The ARDL bounds test results reported in Table 4 provide evidence on the existence of a long-run relationship among the variables. Using a lag length of two, selected based on the appropriate lag selection criteria, the computed F-statistic (5.613) exceeds the upper critical bounds at the conventional significance levels. This indicates that the null hypothesis of no cointegration can be rejected. The comparison with critical values across different sample sizes further confirms this result, as the test statistic remains above the upper bound even at the 1% level. This finding supports the presence of a stable long-run equilibrium relationship between human capital and its determinants, allowing for the estimation of both long-run coefficients and short-run dynamics within the ARDL framework.

Table 4. ARDL bounds test for cointegration.

Since the computed F-statistic of 5.613 exceeds the upper critical bound at the 1% level, the null hypothesis of no long-run relationship is rejected. This confirms the presence of cointegration among human capital, development aid, humanitarian aid, conflict, and GDP per capita.

4 ARDL results

The ARDL estimates in Table 5 reveal distinct patterns across the short run and long run. In the long run, development aid (DA) is negative and statistically significant, indicating a persistent inverse relationship with human capital. This result suggests that short-term benefits from development aid weaken over time when externally funded activities are not sufficiently absorbed into domestic systems or sustained through local institutional capacity. In contrast, humanitarian aid (HA) shows a positive and significant association, suggesting its supportive role over time. Conflict (CF) also exhibits a negative and significant effect, highlighting its adverse influence on human capital. GDP per capita (Y), while positive, is not statistically significant in the long run.

Table 5. ARDL long-run and short-run estimates.

LDA, LHA, and LCF refer to the logarithmic forms of development aid, humanitarian aid, and conflict, respectively. Y refers to GDP per capita. ECT(−1) is the lagged error-correction term. Standard errors are reported in addition to coefficients and probability values to support statistical inference.

In the short run, the effects differ for some variables. Development aid shows a positive and significant coefficient, suggesting that its immediate influence contrasts with its longer-term relationship. This pattern indicates that development aid is associated with immediate improvements through program financing and service support, while the long-run negative coefficient reflects the difficulty of converting these gains into durable institutional and human development capacity. Humanitarian aid remains positive and significant, reinforcing its role in supporting human capital. Conflict continues to have a negative and significant effect, indicating consistency across time horizons. GDP per capita is positive and statistically significant in the short run, implying short-term improvements linked to income changes. The error correction term ECT(−1) is negative and highly significant, with a coefficient of −0.7580, indicating a strong adjustment process toward equilibrium. This suggests that deviations from the long-run relationship are corrected relatively quickly, confirming the stability of the model. Specifically, the coefficient implies that approximately 75.8% of short-run disequilibrium is corrected each year. Given possible endogeneity, reverse causality, and omitted-variable concerns, the coefficients are interpreted as dynamic associations rather than definitive causal effects.

4.1 Reliability of the model

The stability diagnostics illustrated in Figure 2 confirm the consistency of the estimated model over time. The CUSUM and CUSUM of Squares plots remain within the 5% critical bounds throughout the sample period, indicating no evidence of structural instability. The recursive residuals do not cross the significance limits, suggesting that the estimated parameters are stable and the model is well specified. These results support the reliability of the ARDL estimates and indicate that the relationships identified in the model remain consistent over the study period.

4.2 Random forest and Bayesian results

Table 6 brings together the Random Forest importance measures and the Bayesian coefficient estimates to provide a broader perspective on the role of each variable. The Random Forest results show that humanitarian aid (HA) has the highest importance, followed by development aid (DA) and conflict (CF), while GDP per capita (Y) contributes the least to predictive performance. Because Random Forest importance values do not represent coefficients, signs, or statistical significance, they are interpreted as evidence of relative predictive contribution. The ranking therefore suggests that humanitarian aid contains the strongest predictive information for variations in human capital, while the lower score for GDP per capita indicates that income contributes less to the model's predictive structure. This interpretation is particularly useful because it complements the ARDL results by identifying which variables are most influential for prediction, even when the model does not estimate direct marginal effects. Given the small sample size, the Random Forest results are interpreted as complementary and exploratory rather than as standalone predictive evidence.

Table 6. Comparative results of random forest importance and Bayesian estimates.

The Bayesian estimates provide information on the direction and uncertainty of these relationships. Development aid and conflict both display negative associations with human capital, with their credible intervals lying entirely below zero, indicating stable negative posterior associations. In contrast, humanitarian aid shows a positive relationship with a credible interval above zero, suggesting a stable positive posterior association. GDP per capita remains positive, but its interval includes zero, indicating a less certain effect. Beyond confirming coefficient direction, the Bayesian results show the degree of uncertainty surrounding each relationship. The narrower and non-zero credible intervals for development aid, humanitarian aid, and conflict indicate more stable posterior evidence, while the interval for GDP per capita reflects greater uncertainty. This suggests that aid and conflict variables provide more reliable evidence within the Bayesian framework than economic growth alone. The combined Random Forest and Bayesian evidence therefore strengthens the interpretation of the main findings by showing both predictive relevance and probabilistic consistency across the key variables.

As reported in Table 7, Random Forest hyperparameters were selected using grid search with 5-fold cross-validation. The sample was divided into 80% training observations and 20% testing observations, corresponding to 27 and seven observations, respectively. Given the small annual sample size, model complexity was restricted by limiting tree depth and requiring minimum samples per leaf and split. Overfitting was assessed by comparing training and testing performance, and the Random Forest results are interpreted as exploratory evidence of relative variable importance rather than causal effects. Bayesian regression was estimated using Python and R. Four Markov Chain Monte Carlo chains were run with 4,000 iterations per chain, including 1,000 warm-up draws and 3,000 post-warm-up draws per chain, resulting in 12,000 posterior draws. Convergence diagnostics indicated satisfactory sampling performance, with a maximum R-hat value of 1.0004, minimum bulk effective sample size of 9,397, and minimum tail effective sample size of 9,154. Posterior predictive checks were conducted to assess whether the model could reproduce the observed Human Capital Index patterns. These diagnostics indicate no evidence of non-convergence and support the reliability of the Bayesian estimates.

Table 7. Machine learning and bayesian model validation settings.

4.3 Model performance and comparison

The model results show complementary evidence across approaches rather than direct competition among the models. The ARDL model provides the strongest in-sample explanatory power, with an R² of 0.846, indicating a high ability to explain variations in human capital within the estimated dynamic specification. The Random Forest model shows moderate predictive performance, reflecting its role in identifying patterns rather than structural relationships. The Bayesian model also demonstrates moderate performance, with reasonable prediction accuracy. As shown in Table 8, the full Bayesian model performs better than the null model, as indicated by higher (less negative) WAIC and LOO values. These criteria are Bayesian-specific measures and are used only to compare the full Bayesian model with the Bayesian null model, not to compare Bayesian regression directly with ARDL or Random Forest. This confirms that the inclusion of explanatory variables improves predictive ability within the Bayesian framework. The Random Forest metrics are based on the train-test validation described in Table 6, while the ARDL R² reflects in-sample explanatory fit. RMSE and MAE are not reported for ARDL because the ARDL model is used mainly for short-run and long-run dynamic inference rather than out-of-sample prediction. Given the small sample size, the Random Forest and Bayesian performance results are interpreted as supporting evidence rather than standalone predictive validation. These findings suggest that ARDL serves as the main explanatory model, while the other approaches provide supporting evidence through predictive importance and uncertainty assessment.

Table 8. Model performance and comparison.

The findings presented in this study align with the expectations derived from the theoretical and empirical background. The analysis shows that development aid exhibits differing effects across time horizons, with short-term support contrasting with longer-term patterns. Humanitarian assistance consistently contributes positively to human capital, particularly in periods of instability, while conflict remains a persistent constraint on human development. The role of economic conditions appears less pronounced, with limited evidence of a stable long-term relationship.

These results confirm that the interaction between external assistance, instability, and economic conditions is not uniform but evolves over time. By examining these relationships within a unified and dynamic framework, the study provides a clearer understanding of how human capital is shaped in fragile settings such as Somalia.

5 Discussion

The results of this study provide a coherent understanding of how human capital evolves in a fragile and aid-dependent context, while also reflecting patterns identified in the existing literature. By combining dynamic econometric analysis with complementary modeling approaches, the findings offer a more nuanced perspective on the interaction between external assistance, instability, and economic conditions in Somalia.

A central finding is the contrasting role of development aid across time horizons. The positive short-run effect suggests that development aid contributes to immediate improvements, which is consistent with studies showing that targeted interventions can enhance service delivery and support human development in the short term (Asaleye et al., 2023; Wang et al., 2024). However, the negative long-run relationship indicates that these gains may not be sustained over time. This pattern is consistent with aid-effectiveness and dependency perspectives, which argue that external assistance can produce immediate welfare gains but weaker long-term outcomes when aid delivery remains fragmented, donor-driven, or weakly embedded in domestic systems (Bethencourt and Perera-Tallo, 2025; Burnside and Dollar, 2000; Easterly, 2003).

In addition, development aid is often delivered through externally driven programs that may operate parallel to national systems. While such approaches can generate quick results, they may limit the development of local administrative capacity and reduce policy ownership over time. Fragmentation of aid flows, short project cycles, and weak coordination across donors can further reduce effectiveness, especially in environments with limited governance capacity (Hussein et al., 2026). These conditions make it difficult for short-term gains to translate into sustained improvements in human capital. In fragile-state settings, this reflects a broader absorptive-capacity problem: external resources support service delivery more effectively when they are coordinated with domestic priorities and institutional continuity, rather than remaining dependent on temporary project cycles (Mothupi et al., 2025; Shroff et al., 2024).

Another important consideration is the interaction between aid inflows and domestic resource allocation. External assistance may influence government priorities, potentially shifting attention away from long-term investment in health and human development if aid is perceived as a substitute. This can weaken the efficiency of public spending and reduce the long-term impact of aid. In a context such as Somalia, where institutional capacity and accountability mechanisms remain constrained, these effects may become more pronounced. Taken together, these findings suggest that development aid can provide short-term support but requires stronger alignment with domestic systems to generate lasting benefits. This interpretation is also consistent with fragile-state governance literature, which emphasizes that the effectiveness of external assistance depends on domestic absorptive capacity, coordination, and institutional continuity (Alem and Shirakawa, 2022).

In contrast, humanitarian aid shows a consistently positive relationship with human capital across both time horizons. This finding reflects its primary role in preserving wellbeing during periods of crisis. Similar evidence has been documented in studies highlighting the effectiveness of humanitarian interventions in improving health outcomes and supporting vulnerable populations in fragile settings (Akpoghelie et al., 2024; Morrison and Malik, 2024). The importance of humanitarian assistance is further reinforced by the Random Forest results, which identify it as the most influential variable in predicting human capital. While the literature acknowledges challenges related to coordination and sustainability, its contribution to maintaining human capital under unstable conditions remains evident (Hussein et al., 2026). The result also reinforces the conceptual distinction between humanitarian aid and development aid: humanitarian aid primarily preserves existing human capital during crises, while development aid is expected to build longer-term capacity (Paulmann, 2013).

The negative impact of conflict on human capital is strong and consistent across all models, reinforcing findings widely reported in the literature. Conflict disrupts access to healthcare, education, and economic opportunities, leading to long-term declines in human capital (Hidalgo-Aréstegui et al., 2025; Nor and Yusof, 2025). The persistence of this effect across both short-run and long-run estimates indicates that the consequences of instability extend beyond immediate disruptions and continue to shape development outcomes over time. This finding is also consistent with studies showing that conflict and weak human capital can reinforce each other, as poor health, limited labor participation, and weak livelihood opportunities may increase vulnerability to instability and fragility (Nor and Raheem, 2025; Vesco et al., 2025). In the Somali context, this highlights the central role of stability in supporting human development.

The relationship between economic conditions and human capital appears less pronounced. While GDP per capita shows a positive effect in the short run, its long-run impact is not statistically strong. This pattern is consistent with studies suggesting that economic growth alone may not translate into improvements in human capital without complementary investments in social sectors and institutional capacity (Büttner et al., 2023; Patterson, 2023; Tahir and Burki, 2023). In fragile economies, income gains may remain insufficient when service delivery systems, labor markets, and public investment capacity are weak, limiting the conversion of economic growth into sustained human capital improvements (Afolabi and Raifu, 2025).

The comparison across modeling approaches further strengthens the interpretation of these results. The ARDL model captures both short-run adjustments and long-run relationships, providing the main basis for inference. The Random Forest model complements this by identifying the relative importance of variables, highlighting the dominant role of humanitarian aid and the relevance of development aid and conflict. The Bayesian model adds a probabilistic dimension by confirming the direction of key relationships and accounting for uncertainty. The consistency of findings across these approaches enhances confidence in the results. At the same time, the relatively small annual sample means that Random Forest and Bayesian estimates are best understood as complementary evidence that supports the ARDL results rather than as independent bases for causal inference.

Taken together, the findings indicate that human capital development in fragile settings is shaped by a combination of short-term support, long-term structural conditions, and persistent instability. External assistance plays an important role, but its effects differ depending on its type and time horizon. Humanitarian aid contributes to maintaining human capital during crises, while development aid requires stronger institutional alignment to produce sustained outcomes. At the same time, conflict remains a fundamental constraint, limiting both domestic capacity and the effectiveness of external support, while economic conditions alone appear insufficient to drive long-term improvements. These findings are most directly applicable to Somalia, where prolonged fragility, aid dependence, and recurrent conflict shape the relationship between external assistance and human capital.

These insights highlight the importance of aligning short-term interventions with long-term development strategies. Strengthening institutional capacity, improving governance, and addressing instability are essential for enhancing the effectiveness of external assistance. In the case of Somalia, where policy decisions are often made under uncertainty, the use of multiple analytical approaches provides a more comprehensive understanding of these relationships and supports more informed efforts to improve human capital over time.

6 Conclusion, policy directions, and limitations

This study examines how external assistance, instability, and economic conditions shape human capital in Somalia over the period 1990–2023 using a unified and dynamic framework. The evidence points to differentiated effects across variables and time horizons. Development aid contributes to short-term improvements but does not translate into sustained gains over time, while humanitarian assistance plays a consistent supportive role in maintaining human capital, particularly under crisis conditions. Conflict remains a persistent and dominant constraint, exerting a negative influence across both short-run and long-run dimensions. Economic conditions show a more limited and less stable relationship, suggesting that income growth alone is insufficient to drive lasting improvements in human capital without complementary structural support.

These patterns underline an important distinction between immediate support and long-term development. Short-term interventions can stabilize outcomes, but durable progress depends on how external resources interact with domestic institutions, policy priorities, and the broader environment. In fragile settings, where shocks are frequent and governance capacity is constrained, this interaction becomes especially critical.

From a policy perspective, the findings point to the need for stronger alignment between external assistance and domestic systems. Development aid would be more effective when it is embedded within national frameworks, supports institutional capacity, and prioritizes continuity beyond project cycles. This includes improving coordination across donors, reducing fragmentation, and strengthening accountability mechanisms so that aid complements rather than substitutes domestic efforts. Emphasis should also be placed on long-term investments in health systems and labor market participation, ensuring that short-term gains are translated into sustained improvements.

Humanitarian assistance should continue to play a central role in protecting human capital during periods of instability, but with closer integration into longer-term development strategies. Linking emergency interventions with recovery and system-building efforts can help avoid discontinuities between relief and development. This requires coordination between humanitarian and development actors, as well as planning frameworks that extend beyond immediate crisis response.

Reducing the impact of conflict is fundamental for any meaningful progress in human capital. Policies that support stability, strengthen institutions, and rebuild service delivery systems are essential. Investment in basic infrastructure, education, and healthcare systems in post-conflict settings can help restore human capital and reduce long-term vulnerabilities. Without addressing instability, the effectiveness of both domestic policies and external assistance remains limited. Future policy design should also recognize that conflict and human capital can reinforce each other, since weak health, low labor participation, and limited opportunities may deepen vulnerability to fragility.

Economic conditions also require attention, not only in terms of growth but in how resources are distributed and utilized. Strengthening the link between income and human capital outcomes involves improving access to healthcare, expanding employment opportunities, and ensuring that economic gains translate into broader social benefits. This suggests a focus on inclusive policies that enhance both the capacity and utilization of human capital.

The study also has several limitations that should be considered when interpreting the results. First, the analysis relies on available time-series data, which may be subject to measurement constraints, particularly in fragile and data-scarce environments. The sample is limited to 34 annual observations, which constrains degrees of freedom and limits the predictive strength of machine-learning approaches. Second, the Human Capital Index is constructed using health and labor participation indicators due to limited availability of consistent education data, which may not fully capture all dimensions of human capital. Third, while the methodological approach combines multiple techniques to enhance robustness, each model has its own assumptions and limitations, and the results should be interpreted within this context. Possible endogeneity and reverse causality also remain important concerns, since aid inflows may respond to deteriorating human capital conditions and conflict may both affect and be affected by human capital. Finally, the study focuses on a single country, which may limit the generalizability of the findings to other settings with different institutional and economic conditions.

Despite these limitations, the study provides a comprehensive and integrated perspective on the drivers of human capital in a fragile context. By combining multiple analytical approaches and examining both short-term and long-term dynamics, it contributes to a clearer understanding of how external assistance and internal conditions interact. The findings highlight the importance of coordinated, context-specific strategies that balance immediate support with long-term development objectives, offering insights that are relevant for policy design and future research.

Correction note

This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Oleksandra Werbowa, University of Warmia and Mazury in Olsztyn, Poland

Reviewed by: Abdirahman Ibrahim Abdi, SIMAD University, Somalia

Mohamud Hussein, SIMAD University Faculty of Economics, Somalia