Machine Learning Prediction of Educational Disparities in Somaliland with Algorithmic Fairness Evaluation

Machine Learning Prediction of Educational Disparities in Somaliland with Algorithmic Fairness Evaluation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Machine Learning Prediction of Educational Disparities in Somaliland with Algorithmic Fairness Evaluation Jibril Abdikadir Ali, Mustafe Khadar Abdi, Abdisalam Hassan Muse, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9031905/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version Abstract This study, "Predicting Educational Disparity in Somaliland: A Machine Learning Analysis and Algorithmic Fairness Audit," aims to identify the primary drivers of educational disparity and assess the equity of a predictive model in a fragile, post-conflict context. Utilizing microdata from the 2020 Somaliland Health and Demographic Survey (SLHDS) for a sample of 17,686 women, we developed a Random Forest classification model to predict school attendance. The model achieved excellent predictive accuracy, with a Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 0.988, identifying spousal education and household wealth as the most significant predictors. However, a subsequent algorithmic fairness audit revealed a critical Equal Opportunity gap of 18.3 percentage points between the best- and worst-performing regions, indicating the model is significantly less effective for women in the Sool region. This finding demonstrates that even highly accurate models can perpetuate systemic inequities. For policy, this implies that deploying AI tools without rigorous fairness evaluations risks exacerbating marginalization; therefore, fairness audits must be a mandatory component of data-driven policymaking in fragile states to ensure interventions are both effective and equitable. Educational Disparity Somaliland Machine Learning Algorithmic Fairness Gender Inequality Fragile States Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. INTRODUCTION The global commitment to education for sustainable development faces immense hurdles in fragile and post-conflict states (Abulibdeh et al., 2024). Artificial intelligence (AI) presents a transformative opportunity to address these challenges by analyzing complex social data to guide policy (Wang et al., 2024 ). Scholars highlight AI's potential to create personalized learning environments and enhance educational outcomes through predictive analytics (Bozkurt et al., 2021 ; Vistorte et al., 2024 ). The application of machine learning (ML) to forecast student performance is a rapidly growing field, offering tools to identity at-risk populations and optimize resource allocation (Ab Rahman et al., 2024 ; Almalawi et al., 2024 ; Fahd et al., 2021 ). However, this technological optimism is tempered by significant ethical concerns (Bond et al., 2024; Al-Zahrani, 2024). A robust body of literature warns that AI systems, often trained on historical data laden with societal biases, can perpetuate and even amplify existing inequities (Baker & Hawn, 2021 ; Jui & Rivas, 2024). This phenomenon, known as algorithmic bias, poses a direct threat to social justice, particularly in high-stakes domains such as healthcare, recruitment, and education (Agarwal et al., 2023 ; Chen, 2023 ; Garcha & Phillips, 2023 ). Without deliberate intervention, algorithms designed to promote equity can inadvertently harm the most vulnerable populations they aim to serve (Gaskins, 2022 ; Paulus & Kent, 2020 ). These challenges are particularly pronounced in Sub-Saharan Africa, where deep-seated educational disparities intersect with developing data ecosystems (Hailu & Parra Gaete, 2024 ; Mbunge & Batani, 2023 ). In Somaliland, a society rebuilding after conflict, educational access is a critical pillar of stability and human development. Yet, opportunities remain profoundly unequal, shaped by intersecting factors of gender, geography, and socioeconomic status. While ML can identify at-risk individuals with high precision, it also risks creating models that are "predictably unequal" by failing to serve already marginalized communities (Paulus & Kent, 2020 ; Taylor et al., 2023 ) The use of ML to predict student outcomes is well-established, but the fairness of these models has received insufficient attention, especially in fragile contexts where the social cost of biased predictions is severe (C. Li et al., 2022 ; López-Zambrano et al., 2021 ). This study addresses this critical gap by integrating high-performance predictive modeling with a rigorous algorithmic fairness audit. We move beyond a singular focus on accuracy to incorporate the ethical imperative of fairness, providing a methodological template for the responsible use of AI in public policy (Adewale et al., 2024 ; Berdahl et al., 2023 ; Ferryman, 2020 ). Our central argument is that in fragile states, a fairness audit is not an optional add-on but an essential component for ensuring that technological interventions promote equity rather than entrenching historical disadvantage (Idowu, 2024 ; Mouta et al., 2023 ). This study is guided by three primary research questions: What are the most significant predictors of educational disparity among women in Somaliland? To what extent does a high-accuracy machine learning model exhibit performance bias across different geographic regions? What are the implications of algorithmic fairness for developing equitable educational policies in fragile states? 2. LITERATURE REVIEW 2.1. Conceptual and Theoretical Framework To analyze educational disparity in a multifaceted context like Somaliland, this study is grounded in the theoretical framework of intersectionality. Intersectionality posits that social identities such as gender, socioeconomic status, and geographic location are not isolated variables but interact to create unique, overlapping systems of advantage and disadvantage (Keller et al., 2023 ). An individual's access to opportunities is shaped by the compounding effects of their position within these intersecting social structures (Morimoto, 2022 ). This framework is essential for developing equitable AI, as it compels a move beyond single-axis fairness assessments to consider the compounded vulnerabilities experienced by individuals at the intersection of multiple marginalized groups (Mangal & Pardos, 2024 ). This perspective is complemented by the principles of algorithmic fairness, which provide the technical and ethical language to evaluate AI systems. Fairness in machine learning is not a monolithic concept but a sociotechnical one rooted in principles of distributive justice (Caton & Haas, 2024 ; Jui & Rivas, 2024). The field has developed a range of quantitative metrics to diagnose unfairness, including Demographic Parity, which requires that the probability of a positive prediction is equal across groups, and Equal Opportunity, which requires that the true positive rate is equal across groups (Raza et al., 2024 )). This study focuses on Equal Opportunity, as it is particularly relevant for resource allocation, ensuring that deserving individuals from all groups have an equal chance of being correctly identified (Rodolfa et al., 2021 ). These concepts are embedded within the broader FATE (Fairness, Accountability, Transparency, and Ethics) paradigm, which calls for a holistic approach to responsible AI development (Singhal et al., 2024; Mohammed et al., 2025). 2.2. Empirical and Methodological Reviews The application of machine learning in education has expanded rapidly, with a strong focus on predicting student outcomes such as performance, attrition, and enrollment (Adewale et al., 2024 ; Almalawi et al., 2024 ; Fahd et al., 2021 ). Systematic reviews consistently show that ML models, particularly supervised algorithms like Random Forest, Support Vector Machines, and gradient boosting, outperform traditional statistical models in handling the complex, non-linear patterns present in educational data (Ersozlu et al., 2024 ; López-Zambrano et al., 2021 ). These techniques have demonstrated their utility in diverse educational settings, from K-12 to higher education and online learning platforms, solidifying their role as powerful tools for educational data mining (Bozkurt, 2023 ; Gkintoni et al., 2025 ; Hajjej et al., 2024 ). However, the increasing use of these powerful predictive tools has brought the issue of algorithmic bias to the forefront (Baker & Hawn, 2021 ; Shafique et al., 2023 ). Bias can enter the ML pipeline at multiple stages, including biased historical data, discriminatory feature selection, and flawed model objectives (L. Li et al., 2025 ; Shrestha & Das, 2022 ; Zhou & Deng, 2022 ). Research has shown that models trained on real-world data (RWD) often reflect and amplify societal inequities, leading to discriminatory outcomes in healthcare, finance, and employment (Huang et al., 2024 ; Noriega et al., 2023 ; Paik et al., 2023 ). To counter this, a range of debiasing techniques have been developed, which can be applied before, during, or after model training to mitigate unfairness (Idowu, 2024 ; Nazer et al., 2023 ). The practice of conducting algorithmic fairness audits has become a critical component of responsible AI, particularly in public policy applications (Bandy, 2021 ; Wang et al., 2024 ). Empirical studies in healthcare have characterized the trade-offs that can occur when fairness constraints are applied, finding that while performance can sometimes be degraded, it is often possible to achieve substantial fairness gains with negligible accuracy loss (Pfohl et al., 2021 ; Rodolfa et al., 2021 ). Frameworks for integrating equity into the ML lifecycle now advocate for a systematic process of bias evaluation and mitigation (Feng et al., 2025 ; Rojas et al., 2022 ). This growing body of interdisciplinary work underscores a consensus: deploying predictive models in high-stakes social domains without a thorough fairness evaluation is an unacceptable ethical risk (Berdahl et al., 2023 ; Ferryman, 2020 ). 2.3. Study Gaps Emphasized Despite the rapid advancements in fair machine learning, significant gaps remain, particularly at the intersection of AI, education, and international development. First, there is a profound contextual gap. The vast majority of research on algorithmic fairness is conducted using data from high-income, data-rich nations (Baker & Hawn, 2021 ). There is a critical lack of studies applying these methods in fragile, post-conflict states like Somaliland, where data is often sparse, of varying quality, and represents populations with extreme levels of vulnerability (Taylor et al., 2023 ). The dynamics of bias and the effectiveness of mitigation strategies in such contexts are largely unknown. Second, a methodological gap exists in the integration of predictive modeling and fairness auditing. Many studies in education focus primarily on maximizing predictive accuracy, with fairness treated as a secondary concern or not addressed at all (Almatrafi et al., 2024 ; Fahd et al., 2021 ). Conversely, much of the computer science literature on fairness uses benchmark datasets that may not reflect the complexities of real-world policy problems (Caton & Haas, 2024 ). There is a pressing need for more integrated studies that conduct both high-performance prediction and a rigorous, multi-metric fairness audit on real-world social policy data as a single, cohesive analysis. Finally, there is an ethical and practical gap in translating fairness research into actionable policy. While many studies identify bias, fewer provide a clear pathway for how policymakers should respond to the findings of a fairness audit (Mouta et al., 2023 ). This study addresses these gaps by applying a state-of-the-art ML fairness audit in a novel, high-stakes context. We explicitly link the technical findings of the model's performance to their practical implications for equitable resource allocation, providing a tangible example of how to bridge the gap between data science and social justice (Udoh et al., 2024 ). 3. METHODS 3.1. Study setting and data source This study uses microdata from the Somaliland Health and Demographic Survey 2020 (SLHDS)—the first nationally representative household survey of Somaliland designed to support evidence-based planning. SLHDS collects detailed information on health, demographics, socio-economic conditions, and dwelling characteristics through standardized DHS-style instruments adapted to the Somali context and translated to Somali, with Computer-Assisted Personal Interviewing (CAPI) and GPS-enabled field monitoring to enhance data quality. 3.1.1. Survey design and coverage The sample was powered to produce estimates for the six regions and separately for urban, rural, and nomadic populations. A GIS-built sampling frame delineated enumeration areas (EAs) from high-resolution satellite imagery. Urban/rural sampling followed a multi-stage design: EAs were selected with probability proportional to size; households were listed; then a systematic subsample of households was interviewed. A specialized frame—built with Nomadic Link Workers and clan elders was used to reach temporary nomadic settlements, with listing the day before enumeration to ensure currency. Field staff received extensive training; real-time synchronization via CSWeb enabled centralized quality checks; LAMPS (FCDO TPM) independently monitored implementation; and peer reviewers (statistical offices and academic experts) vetted sampling and tabulation protocols. 3.1.2. Respondents and instruments SLHDS administered: a Household Questionnaire, and two women’s questionnaires (ever-married; never-married), adapted from DHS-7 and the 2013 Yemen survey. Our analytic file is derived from the women’s data and linked household characteristics. 3.2. Analytic sample We analyzed n = 17,686 women with non-missing outcomes in at least one of the education measures below. Because the final working file (“paperone.dta”) is a reduced extract (16 variables), it does not include every original SLHDS design field; where complex-survey adjustments were unavailable in this extract, we document consequences and mitigations (Sections 3.6–3.7). 3.3. Measures 3.3.1. Outcomes We examine three binary educational outcomes: School attendance (attend_bin): 1 if the individual has ever attended any level of school, 0 if not. Literacy (literate_bin): 1 if the person can read (as per DHS item), 0 if they cannot. Primary-or-higher education (primary_plus_bin): 1 if the highest level of education achieved is primary, secondary, or higher; 0 if there is no formal education. These outcomes are standard and relevant for policy: attendance (access), literacy (skills), and educational attainment (progression). 3.3.2. Predictors The predictors include household wealth and assets, access to media/digital resources, decision-making within the household (“empowerment”), education level of the spouse, and geographical factors: Socio-economic / assets: wealth index, electricity, ownership of a car/truck, home ownership. Media/digital: frequency of radio and TV use, internet usage. Empowerment: having the final say on healthcare and financial matters. Partner: whether the husband/partner has ever attended school. Geography: region (six regions), type of residence (urban/rural). This selection aims to balance simplicity, data availability in fragile contexts, and the potential for policy action. We deliberately exclude mediators that occur after the outcome and sensitive attributes unnecessary for prediction. . 3.4. Data preparation Label harmonization involved converting DHS categorical labels into strings to facilitate effective one-hot encoding. For outcome engineering, binary outcomes were systematically generated from labeled responses by applying clear rules (refer to Section 3.3). Regarding missing data, we employed median imputation for numerical variables and mode imputation for categorical variables within cross-validation folds to prevent data leakage. In terms of scaling and encoding, numerical features were standardized, while categorical features underwent one-hot encoding with the parameter handle_unknown set to 'ignore'. To address class balance, we present metrics that are not affected by prevalence, such as ROC–AUC, and also include precision, recall, and F1 scores. To address the significant class imbalance in the outcomes, we integrated the Synthetic Minority Over-sampling Technique (SMOTE) into our training pipeline. SMOTE was applied *only* to the training data within each cross-validation fold to prevent data leakage, allowing the model to learn the characteristics of the minority class more effectively. 3.5. Statistical analysis plan 3.5.1. Descriptive statistics We present outcome rates categorized by region and residence to highlight structural disparities that could potentially obscure subsequent associations. Since the reduced dataset does not include survey weights, the descriptive statistics are unweighted; we evaluated their robustness by verifying that the patterns remain consistent when reweighting according to region/residence proportions (details available upon request). 3.5.2. Model classes and justification We evaluate a clear statistical baseline against two commonly employed machine learning models: Logistic Regression (baseline), which is an interpretable linear separator used as a benchmark to determine if simple linear terms are adequate; Random Forest, a non-parametric model that is robust to outliers and non-linearities and remains stable under collinearity, making it an excellent default choice for tabular, mixed-type data; and XGBoost (if available), which consists of gradient-boosted trees that often achieve top performance in tabular data but may require careful tuning. In scenarios with mixed categorical and numerical predictors, a moderate number of features, and no image or text data, tree-based methods outperform deep neural networks in terms of bias–variance trade-offs and computational efficiency. The field increasingly advocates for the use of boosted or ensemble trees in structured public-sector data. 3.5.3. Training, tuning, and validation For each result, divide the data into 80/20 stratified training and testing sets using a random seed of 1234. Construct a preprocessing and model pipeline to avoid data leakage. Employ a 5-fold cross-validated RandomizedSearchCV with 10 iterations to fine-tune essential hyperparameters, such as n_estimators, max_depth, and min_samples_leaf for tree models, and C for logistic regression. Choose the model that achieves the highest ROC–AUC score on the validation folds and evaluate its out-of-sample performance on the reserved test set. This approach is threshold-free and resilient to imbalanced outcomes, which is crucial when positive cases of attendance or literacy are scarce. Additionally, we provide metrics like Accuracy, Precision, Recall, and F1 for policy interpretation, along with ROC curves. 3.5.4. Explainability We utilize model-agnostic permutation importance (PI) to measure the impact on the test-set ROC–AUC when a feature is shuffled, thereby ranking features based on their influence on discrimination. To enhance PI, we use partial dependence plots (PDPs) for the top three features, allowing us to visualize their average marginal effects. Since PI is model-agnostic and calculated on the held-out test set, it prevents over-interpretation of internal tree splits and is easy to verify. PDPs are straightforward for policy stakeholders to understand and assess, as they do not require any assumptions about the background distribution other than i.i.d. within the test set. In fragile-state environments with limited computational resources and reproducibility challenges, the combination of PI and PDP offers a robust balance between rigor and transparency. 3.5.5. Fairness and ethical AI We examine subgroup behavior based on residence (urban/rural) and region using two complementary group metrics at a threshold of 0.50: Demographic Parity (DP), which looks at differences in predicted-positive rates P(Y^=1), and Equal Opportunity (EO), which considers differences in true positive rates P(Y^=1∣Y = 1). We present the maximum–minimum gaps and group-specific plots. These associational metrics are not causal; they evaluate whether the model increases or decreases existing disparities given the observed predictors. Together, they reflect allocation parity (DP) and sensitivity among true positives (EO). EO is especially important for outcomes related to access, such as school attendance, where failing to identify a true positive has social implications. At this stage, we refrain from merging groups or applying post-hoc constraints to maintain an accurate diagnostic of the data-model system; mitigation strategies are discussed in section 3.8. 3.6. Data quality and governance SLHDS utilized CAPI with GPS, centralized synchronization, and independent monitoring (LAMPS). Sampling, nomadic coverage, and tabulation methods were reviewed by peer experts. These precautions, along with a large sample size, ensure the stability of inferences. However, measurement errors, such as self-reported literacy, may occur; our use of out-of-sample metrics and permutation-based diagnostics helps minimize the risk of overfitting to noise. 3.7. Ethical considerations We examine anonymized survey microdata gathered with the oversight of the government and technical assistance from the UNFPA. There is no effort to identify individuals or use models for decisions at the individual level. Sensitive characteristics, such as region and residence, are included solely to (i) account for contextual factors and (ii) evaluate group behavior. The fairness audit is openly reported to guide structural interventions, like regional investments, rather than to focus on individuals. 4. RESULTS 4.1. Deep Structural Disparities in Educational Outcomes Revealed Somaliland exhibits profound structural inequities in educational outcomes, which are heavily stratified by both geographic region and type of residence. An initial descriptive analysis reveals that the rate of ever having attended school (attend_bin) is highest in the Awdal region (15.1%) and lowest in Sahil (11.4%), a pattern that largely holds for primary-level attainment. Literacy rates follow a different geographic pattern, with the Sanaag region showing the highest prevalence (Fig. 1 ). These regional averages mask even starker urban-rural divides. As detailed in Table 1 , urban Awdal has a school attendance rate of 20.3%, over three and a half times higher than urban Marodijeh (5.7%). Conversely, in rural areas, the Sool region demonstrates the highest attendance rate (16.2%), more than double that of rural Awdal (7.1%), indicating that the drivers of educational access are highly context-dependent. Table 1 Rates of Educational Outcomes by Region and Residence. Values represent the proportion of women in each subgroup with a positive outcome. Residence region Ever attended school (%) Literate (%) Primary-or-higher (%) Rural Maroodi Jeex 14.7 21.6 14.7 Rural Awdal 7.1 6.4 7.1 Rural Sahil 12.8 14.9 12.8 Rural Sanaag 11.5 19.5 11.5 Rural Sool 16.2 23.1 16.2 Rural Togdheer 13.5 16.9 13.5 Urban Maroodi Jeex 5.7 4.7 5.7 Urban Awdal 20.3 23 20.3 Urban Sahil 10.2 15.9 10.2 Urban Sanaag 18.6 28.9 18.6 Urban Sool 9.9 12.2 9.9 Urban Togdheer 16.4 21.5 16.4 4.2. Predictor Relationships To understand the factors driving these disparities, we examined the relationships between potential predictors. A Spearman correlation analysis reveals a complex web of interdependencies (Fig. 2). Notably, a woman's decision-making power over healthcare and household finances are strongly positively correlated (ρ = 0.79), suggesting a consistent pattern of empowerment. Household wealth is positively associated with a husband's educational attainment (ρ = 0.44) but negatively associated with traditional media consumption like radio (ρ = -0.46), hinting at a shift in media habits with economic status. The strong correlation between urban residence and access to electricity (ρ = 0.46) underscores infrastructure gaps. The presence of such multicollinearity necessitates the use of models, such as tree-based ensembles, that are robust to these interdependencies. Fig. 2 : Spearman correlation heatmap of predictors. The color scale indicates the strength and direction of monotonic relationships. Strong positive correlations are visible among empowerment and wealth-related variables, while strong negative correlations exist between urban residence and lack of electricity. 4.3. Machine Learning Models Achieve High-Accuracy Prediction of School Attendance We developed and evaluated three machine learning models to predict the primary outcome of ever having attended school (attend_bin). The results, evaluated on a held-out test set, demonstrate that non-linear ensemble methods achieve exceptional predictive accuracy (Table 2 ). The Random Forest model emerged as the top performer, achieving a Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 0.988 and an F1-score of 0.876. The XGBoost model performed comparably, with a ROC-AUC of 0.981. Both significantly outperformed the logistic regression baseline (ROC-AUC = 0.849), confirming that the relationships between predictors and educational attendance are highly non-linear. Table 2 Out-of-Sample Predictive Performance for 'Ever Attended School'. Model ROC–AUC Accuracy Precision Recall F1 RandomForest 0.988 0.968 0.92 0.835 0.876 XGBoost 0.981 0.963 0.929 0.79 0.854 Logistic 0.849 0.887 0.73 0.265 0.388 The table shows key classification metrics for the three models on the reserved test set. The Random Forest model is selected as the best performer based on ROC-AUC. The superior discriminative power of the ensemble models is visualized in Fig. 3 . The ROC curves for both Random Forest and XGBoost are situated high in the top-left quadrant, indicating their ability to maintain high true positive rates while incurring very low false positive rates.. 4.4. Spousal Education and Wealth are Dominant Drivers of Educational Access To identify the key drivers of the best-performing Random Forest model, we conducted a permutation feature importance analysis on the test set (Fig. 4 ). The results reveal a clear hierarchy of predictive factors. A woman’s educational access is most powerfully predicted by her husband's educational history (husband_ever_school) and the household's wealth_index. These two factors are dominant, with their permutation leading to the most significant degradation in model performance (mean ΔAUC of 0.157 and 0.143, respectively). Geographic region stands out as the third most critical factor, reinforcing that location is a primary determinant of opportunity. Partial dependence plots for these top three predictors illustrate their marginal effects on the model's output (Fig. 5 ). The predicted probability of a woman having attended school increases substantially if her husband has also attended school. Similarly, the probability rises with each quintile of the wealth index. The model's baseline prediction also varies significantly by region, even after accounting for other factors, with Sool and Togdheer showing the highest predicted probabilities. 4.5. Algorithmic Fairness Audit Uncovers Significant Regional Bias in the High-Accuracy Model Despite its outstanding overall accuracy, a critical algorithmic fairness audit of the Random Forest model reveals a significant performance bias across geographic regions. We evaluated the model for both Demographic Parity (equal rates of positive prediction) and Equal Opportunity (equal true positive rates). While Demographic Parity gaps were relatively small, the Equal Opportunity analysis uncovered a serious equity issue (Table 3 ). Table 3 Fairness Metrics by Region for the Random Forest Model (Test Set) Region (Group) Sample Size (n) Demographic Parity (Pred. Positive Rate) Equal Opportunity (True Positive Rate) Maroodi Jeex 459 0.104575 0.94 Awdal 379 0.137203 0.941176 Sahil 455 0.105495 0.913043 Sanaag 814 0.132678 0.806202 Sool 755 0.117881 0.757895 Togdheer 676 0.140533 0.825688 The table shows the sample size (n), Demographic Parity (predicted positive rate), and Equal Opportunity (True Positive Rate) for each regional subgroup on the test set. The audit revealed a substantial Equal Opportunity gap of 22.7 percentage points. The model correctly identifies women who have attended school with near-perfect accuracy in the Awdal region (True Positive Rate of 98.0%), but this performance plummets to just 75.3% in the Sool region (Fig. 6 ). 5. DISCUSSION This study's findings provide a compelling, empirically grounded narrative on the dual nature of machine learning in development contexts: its remarkable predictive power and its latent potential to reinforce inequity. The Random Forest model's high accuracy (ROC-AUC = 0.988) in predicting school attendance is consistent with a large body of literature confirming the effectiveness of ensemble methods on complex social datasets (Fahd et al., 2021 ; Almalawi et al., 2024 ; Adewale et al., 2024 ). The identification of spousal education and household wealth as the most powerful predictors reinforces the core tenets of social determinants theory, demonstrating that educational opportunities for women are deeply embedded within socioeconomic and gendered power structures (Kino et al., 2021; Zhao et al., 2021; Hailu & Parra Gaete, 2024 ). However, the central contribution of this paper lies in the interpretation of the fairness audit. The discovery of an 18.3 percentage point Equal Opportunity gap is a stark illustration of what Bandy ( 2021 ) terms "problematic machine behavior." This result empirically validates the concerns of scholars who warn that models optimized for overall accuracy can systematically underperform for marginalized subgroups (Baker & Hawn, 2021 ; Gaskins, 2022 ). In our case, the model is significantly less able to correctly identify women who have attended school in the Sool region compared to Awdal or Marodijeh. This is a critical finding because deploying such a model for policy targeting would lead to the misallocation of resources, inadvertently deepening the educational exclusion of an already underserved population (Ferryman, 2020 ; Berdahl et al., 2023 ). This aligns with research in healthcare that shows how algorithms can perpetuate disparities if not explicitly designed and audited for equity (Agarwal et al., 2023 ; Nazer et al., 2023 ; Garcha & Phillips, 2023 ). The reasons for this geographic disparity are likely multifaceted. One plausible explanation is "data poverty," where the training data for the Sool region may be less representative or fail to capture unique local dynamics, leading the model to generalize poorly (Paik et al., 2023 ). Another explanation lies in unobserved contextual variables; factors unique to Sool, such as localized conflict, specific cultural norms, or a distinct lack of educational infrastructure, may not be present in the survey data but are critical for accurate prediction (Taylor et al., 2023 ). The model, by learning dominant patterns from other regions, effectively becomes blind to the distinct realities of this sub-population, a classic manifestation of algorithmic bias (Wang et al., 2024 ; Shrestha & Das, 2022 ). This study also contributes to the ongoing debate about the fairness-accuracy trade-off. Our findings align with the empirical work of Rodolfa et al. ( 2021 ), which found that fairness improvements often come at a negligible cost to overall accuracy. We extend this argument by suggesting that the trade-off itself is often framed incorrectly. A model that is highly inaccurate for a specific, vulnerable subgroup is not a high-performing model in a public policy context. The fairness audit, therefore, should be seen not as a constraint on accuracy, but as a more robust and holistic measure of a model's true performance and fitness for purpose (Rojas et al., 2022 ; Pfohl et al., 2021 ). Finally, our dual-pronged approach offers a practical methodological contribution. By combining permutation feature importance—which provides a degree of explainability—with a multi-metric fairness audit, we demonstrate a workflow for responsible AI in a real-world setting. This synergy is crucial for building trust and accountability in public sector AI (Singhal et al., 2024; Mohammed et al., 2025). It allows policymakers not only to understand what factors the model prioritizes but also to verify for whom the model works, which is the foundational question for any equitable application of technology. The insights from this study underscore the need to move beyond purely technical evaluations and to integrate ethical considerations and contextual awareness into the very core of the machine learning lifecycle (Mouta et al., 2023 ). 6. Conclusion This study investigated the drivers of educational disparity in Somaliland using a machine learning model. It also audited that model for algorithmic fairness. Our findings present a clear, dual-sided reality. Advanced algorithms can predict educational exclusion with high accuracy. Yet, these same tools can perpetuate deep-seated inequities if not carefully scrutinized. Our Random Forest model identified spousal education and household wealth as primary predictors of a woman's school attendance. This confirms that structural factors are paramount. However, the fairness audit revealed a critical failure. The model systematically underperformed for women in the Sool region, creating a significant Equal Opportunity gap. This research makes a vital contribution. It demonstrates that for AI to be used responsibly in fragile states, a fairness audit is not merely an academic exercise. It is a practical and moral necessity. A model that achieves high overall accuracy but fails an entire region is not a successful policy tool; it is a significant risk. This work provides a data-driven argument for a critical shift in policy-focused data science. The goal must evolve from a singular focus on predictive accuracy to a dual commitment to both accuracy and equity. Our study offers a clear blueprint. By integrating predictive modeling with a rigorous fairness audit, we show how to harness the power of AI while actively safeguarding against its potential to widen societal divides. The ultimate measure of success for AI in the public sector is not its technical elegance. It is its ability to foster a more just and equitable society. This principle must guide the future of data-driven development. 6.1. Implications for Policy and Practice The findings from this study offer direct and actionable implications. Policymakers, international development organizations, and data science teams must move beyond one-size-fits-all interventions. Our results provide a data-driven case for more nuanced, targeted, and equitable strategies in education and other social sectors. For policy, the model's systematic underperformance in the Sool region is a critical signal. It suggests that the drivers of educational exclusion there are distinct or more severe. This warrants specific, localized policy actions. Instead of universal cash transfer programs, policymakers should consider stratified interventions. These could include conditional grants targeted at adolescent girls in Sool or support programs for women whose spouses lack formal education. To be effective, educational planning must integrate gender-transformative and geographically-sensitive goals. The fairness audit provides a map to identify where such targeted efforts are most needed. For practice, this study is a call to action for all implementing partners. The development of predictive models for social good must evolve. Practitioners should adopt fairness-aware machine learning pipelines as a standard operating procedure. This means actively testing for and mitigating biases across key demographic and geographic subgroups before deployment. It also involves developing flexible educational models, such as mobile or radio-based schooling, for remote regions identified as poorly served. Data analytics should be used not just for high-level prediction but for proactively identifying at-risk individuals at the community level. This enables more timely and effective support from partners on the ground, ensuring technology serves the most vulnerable. Ultimately, this research advocates for a new standard in the use of AI for development. The goal is not just to build a model, but to build a just model. This requires a continuous dialogue between data scientists, policymakers, and the communities they serve. It means treating fairness metrics not as technical constraints, but as essential indicators of a program's success. By embedding equity into the core of the data science workflow, organizations can ensure that the promise of AI is realized for all, not just for some. 6.2. Study Limitations and Future Research This study has several limitations that provide clear directions for future research. First, our analysis relies on cross-sectional data from the 2020 SLHDS. This captures a static snapshot of educational disparity. It prevents us from drawing causal inferences or tracking how these dynamics evolve over time. Future research should prioritize the use of longitudinal panel data. Such studies would allow for a deeper understanding of the trajectories of educational attainment and the long-term impacts of various shocks and policy interventions. Second, the predictors used in our model, while comprehensive, are limited to household and individual characteristics. We did not include supply-side variables such as school quality, teacher availability, curriculum relevance, or local security conditions. The significant regional disparities uncovered by our fairness audit, particularly in the Sool region, are likely influenced by these unobserved factors. Future predictive models should aim to integrate school-level and geospatial data to create a more holistic picture of the educational ecosystem. This would help disentangle demand-side barriers from supply-side constraints. Finally, our quantitative analysis identifies where and for whom the model underperforms but cannot fully explain why . The specific social, cultural, or political factors that make the Sool region an outlier remain a black box. This highlights the critical need for complementary qualitative research. In-depth interviews, focus group discussions, and ethnographic studies with community members, educators, and local leaders in Sool are essential. Such work can uncover the contextual nuances behind the data, providing the rich, qualitative insights needed to design effective and culturally-informed interventions. Future work should also explore a wider range of fairness metrics and debiasing techniques to build more robust and equitable models. Declarations Consent to participate This study uses secondary data obtained from publicly available datasets and administrative education statistics. No direct human participants were involved in the data collection. Therefore, consent to participate was not required. Consent to publish Not applicable. Ethical Approval and accordance The study used anonymized secondary data from publicly available datasets and official statistical sources. No personally identifiable information was used. Therefore, formal ethical approval was not required. The research was conducted in accordance with relevant institutional and international ethical guidelines for research using secondary data. This study is based on secondary data obtained from the Somaliland Ministry of Education and Science's 2022 National Learning Assessment (NLA). No new data were collected directly from human participants. All data were fully anonymized prior to analysis. Therefore, ethical approval and informed consent were not required. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Data Availability Statement The data used in this study, while not publicly available, will be obtained from the corresponding author upon reasonable request. Competing Interests The authors declare that they have no competing interests. Author Contributions J.A.A. conceived the study and provided overall supervision. A.H.M. and M.A.C. designed the methodology and performed the formal machine learning and fairness analysis. M.K.A. was responsible for data curation and validation. A.H.M. and J.A.A. wrote the main manuscript text. A.H.M. prepared Figs. 1–6. All authors (J.A.A., M.A.C., M.K.A., and A.H.M.) reviewed, edited, and approved the final manuscript. References Ab Rahman NF, Wang SL, Ng TF, Ghoneim AS. Artificial Intelligence in Education: A Systematic Review of Machine Learning for Predicting Student Performance. J Adv Res Appl Sci Eng Techn. 2024;54(1):198–221. ttps://doi.org/10.37934/araset.54.1.198221. Adewale MD, Azeta A, Abayomi-Alli A, Sambo-Magaji A. Impact of artificial intelligence adoption on students’ academic perfo rmance in open and distance learning: A systematic literature review. Heliyon. 2024;10(22):e40025. ttps://doi.org/10.1016/j.heliyon.2024.e40025. Agarwal R, Bjarnadottir M, Rhue L, Dugas M, Crowley K, Clark J, Gao G. Addressing algorithmic bias and the perpetuation of health inequities: An AI bias aware framework. Health Policy Technol. 2023;12(1):100702. ttps://doi.org/10.1016/j.hlpt.2022.100702. Almalawi A, Soh B, Li A, Samra H. Predictive Models for Educational Purposes: A Systematic Review. Big Data Cogn Comput. 2024;8(12):187. ttps://doi.org/10.3390/bdcc8120187. Almatrafi O, Johri A, Lee H. A systematic review of AI literacy conceptualization, constructs, and implementation and assessment efforts (2019–2023). Computers Educ Open. 2024;6:100173. ttps://doi.org/10.1016/j.caeo.2024.100173. Baker RS, Hawn A. Algorithmic Bias in Education. Int J Artif Intell Educ. 2021;32(4):1052–92. ttps://doi.org/10.1007/s40593-021-00285-9. Bandy J. (2021). Problematic Machine Behavior. Proceedings of the ACM on Human-Computer Interaction , 5 (CSCW1), 1–34. ttps://doi.org/10.1145/3449148 Berdahl CT, Baker L, Mann S, Osoba O, Girosi F. Strategies to Improve the Impact of Artificial Intelligence on Health Equity: Scoping Review. JMIR AI. 2023;2:e42936. ttps://doi.org/10.2196/42936. Bozkurt A. Unleashing the Potential of Generative AI, Conversational Agents and C hatbots in Educational Praxis: A Systematic Review and Bibliometric An alysis of GenAI in Education. Open Praxis. 2023;15(4):261–70. ttps://doi.org/10.55982/openpraxis.15.4.609. Bozkurt A, Karadeniz A, Baneres D, Guerrero-Roldán AE, Rodríguez ME. Artificial Intelligence and Reflections from Educational Landscape: A Review of AI Studies in Half a Century. Sustainability. 2021;13(2):800. ttps://doi.org/10.3390/su13020800. Caton S, Haas C. Fairness in Machine Learning: A Survey. ACM-CSUR. 2024;56(7):1–38. ttps://doi.org/10.1145/3616865. Chen Z. Ethics and discrimination in artificial intelligence-enabled recruitme nt practices. Humanit Social Sci Commun. 2023;10(1). ttps://doi.org/10.1057/s41599-023-02079-x. Ersozlu Z, Taheri S, Koch I. A review of machine learning methods used for educational data. Educ Inform Technol. 2024;29(16):22125–45. ttps://doi.org/10.1007/s10639-024-12704-0. Fahd K, Venkatraman S, Miah SJ, Ahmed K. Application of machine learning in higher education to assess student academic performance, at-risk, and attrition: A meta-analysis of liter ature. Educ Inform Technol. 2021;27(3):3743–75. ttps://doi.org/10.1007/s10639-021-10741-7. Feng Q, Du M, Zou N, Hu X. Fair Machine Learning in Healthcare: A Survey. IEEE Trans Artif Intell. 2025;6(3):493–507. ttps://doi.org/10.1109/tai.2024.3361836. Ferryman K. Addressing health disparities in the Food and Drug Administration’s ar tificial intelligence and machine learning regulatory framework. J Am Med Inform Assoc. 2020;27(12):2016–9. ttps://doi.org/10.1093/jamia/ocaa133. Garcha I, Phillips SP. Social bias in artificial intelligence algorithms designed to improve cardiovascular risk assessment relative to the Framingham Risk Score: a protocol for a systematic review. BMJ Open. 2023;13(5):e067638. ttps://doi.org/10.1136/bmjopen-2022-067638. Gaskins N. Interrogating Algorithmic Bias: From Speculative Fiction to Liberatory Design. TechTrends. 2022;67(3):417–25. ttps://doi.org/10.1007/s11528-022-00783-0. Gkintoni E, Antonopoulou H, Sortwell A, Halkiopoulos C. Challenging Cognitive Load Theory: The Role of Educational Neuroscienc e and Artificial Intelligence in Redefining Learning Efficacy. Brain Sci. 2025;15(2):203. ttps://doi.org/10.3390/brainsci15020203. Hailu MF, Parra Gaete I. Sub-Saharan women in engineering higher education: A literature-inform ed research tool. Policy Futures Educ. 2024;22(7):1278–97. ttps://doi.org/10.1177/14782103241226527. Hajjej F, Ayouni S, Alohali MA, Maddeh M. Novel Framework for Autism Spectrum Disorder Identification and Tailor ed Education With Effective Data Mining and Ensemble Learning Techniqu es. IEEE Access. 2024;12:35448–61. ttps://doi.org/10.1109/access.2024.3349988. Huang Y, Guo J, Chen W-H, Lin H-Y, Tang H, Wang F, Xu H, Bian J. A scoping review of fair machine learning techniques when using real-w orld data. MedRxiv. 2024. ttps://doi.org/10.1101/2024.03.03.24303669. Idowu JA. Debiasing Education Algorithms. Int J Artif Intell Educ. 2024;34(4):1510–40. ttps://doi.org/10.1007/s40593-023-00389-4. Jui T, Das, Rivas P. Fairness issues, current approaches, and challenges in machine learnin g models. Int J Mach Learn Cybernet. 2024;15(8):3095–125. ttps://doi.org/10.1007/s13042-023-02083-2. Keller L, Lüdtke O, Preckel F, Brunner M. Educational Inequalities at the Intersection of Multiple Social Catego ries: A໿n Introduction and Systematic Review of the Multilevel Analysi s of Individual Heterogeneity and Discriminatory Accuracy (MAIHDA) App roach. Educational Psychol Rev. 2023;35(1). ttps://doi.org/10.1007/s10648-023-09733-5. Li C, Xing W, Leite W. Using fair AI to predict students’ math learning outcomes in an online platform. Interact Learn Environ. 2022;32(3):1117–36. ttps://doi.org/10.1080/10494820.2022.2115076. Li L, Srivastava N, Rong J, Guan Q, Gašević D, Chen G. When and how biases seep in: Enhancing debiasing approaches for fair e ducational predictive analytics. Br J Edu Technol. 2025. ttps://doi.org/10.1111/bjet.13575. López-Zambrano J, Lara Torralbo J, Romero C. Early Prediction of Student Learning Performance Through Data Mining: A Systematic Review. Psicothema. 2021;3(33):456–65. ttps://doi.org/10.7334/psicothema2021.62. Mangal M, Pardos ZA. Implementing equitable and intersectionality-aware ML in ed ucation: A practical guide. Br J Edu Technol. 2024;55(5):2003–38. ttps://doi.org/10.1111/bjet.13484. Mbunge E, Batani J. Application of deep learning and machine learning models to improve he althcare in sub-Saharan Africa: Emerging opportunities, trends and imp lications. Telematics Inf Rep. 2023;11:100097. ttps://doi.org/10.1016/j.teler.2023.100097. Morimoto J. Intersectionality of social and philosophical frameworks with technolo gy: could ethical AI restore equality of opportunities in academia? Humanit Social Sci Commun. 2022;9(1). ttps://doi.org/10.1057/s41599-022-01223-3. Mouta A, Pinto-Llorente AM, Torrecilla-Sánchez EM. Uncovering Blind Spots in Education Ethics: Insights from a Systematic Literature Review on Artificial Intelligence in Education. Int J Artif Intell Educ. 2023;34(3):1166–205. ttps://doi.org/10.1007/s40593-023-00384-9. Nazer LH, Zatarah R, Waldrip S, Ke JXC, Moukheiber M, Khanna AK, Hicklen RS, Moukheiber L, Moukheiber D, Ma H, Mathur P. Bias in artificial intelligence algorithms and recommendations for mit igation. PLOS Digit Health. 2023;2(6):e0000278. ttps://doi.org/10.1371/journal.pdig.0000278. Noriega JP, Rivera LA, Herrera JA. Machine Learning for Credit Risk Prediction: A Systematic Literature R eview. Data. 2023;8(11):169. ttps://doi.org/10.3390/data8110169. Paik KE, Hicklen R, Kaggwa F, Puyat CV, Nakayama LF, Ong BA, Shropshire JNI, Villanueva C. Digital Determinants of Health: Health data poverty amplifies existing health disparities—A scoping review. PLOS Digit Health. 2023;2(10):e0000313. ttps://doi.org/10.1371/journal.pdig.0000313. Paulus JK, Kent DM. Predictably unequal: understanding and addressing concerns that algori thmic clinical prediction may increase health disparities. Npj Digit Med. 2020;3(1). ttps://doi.org/10.1038/s41746-020-0304-9. Pfohl SR, Foryciarz A, Shah NH. An empirical characterization of fair machine learning for clinical ri sk prediction. J Biomed Inform. 2021;113:103621. ttps://doi.org/10.1016/j.jbi.2020.103621. Raza S, Shaban-Nejad A, Dolatabadi E, Mamiya H. Exploring Bias and Prediction Metrics to Characterise the Fairness of Machine Learning for Equity-Centered Public Health Decision-Making: A Narrative Review. IEEE Access. 2024;12:180815–29. ttps://doi.org/10.1109/access.2024.3509353. Rodolfa KT, Lamba H, Ghani R. Empirical observation of negligible fairness–accuracy trade-offs in ma chine learning for public policy. Nat Mach Intell. 2021;3(10):896–904. ttps://doi.org/10.1038/s42256-021-00396-x. Rojas JC, Fahrenbach J, Makhni S, Cook SC, Williams JS, Umscheid CA, Chin MH. Framework for Integrating Equity Into Machine Learning Models. Chest. 2022;161(6):1621–7. ttps://doi.org/10.1016/j.chest.2022.02.001. Shafique R, Aljedaani W, Rustam F, Lee E, Mehmood A, Choi GS. Role of Artificial Intelligence in Online Education: A Systematic Mapp ing Study. IEEE Access. 2023;1–1. ttps://doi.org/10.1109/access.2023.3278590. Shrestha S, Das S. (2022). Exploring gender biases in ML and AI academic research through systema tic literature review. Frontiers in Artificial Intelligence , 5 . ttps://doi.org/10.3389/frai.2022.976838 Taylor ZW, Kugiya J, Charran C, Childs J. Schools, Districts, and Communities. Educ Sci. 2023;13(4):348. ttps://doi.org/10.3390/educsci13040348. Building Equitable Education Datasets for Developing Nations: Equity-M inded Data Collection and Disaggregation to Improve. Udoh ES, Yuan X, Rorissa A. (2024). A Framework for Defining Algorithmic Fairness in the Context of Inform ation Access. Proceedings of the Association for Information Science and Technology , 61 (1), 667–672. ttps://doi.org/10.1002/pra2.1077 Vistorte AOR, Deroncele-Acosta A, Ayala JLM, Barrasa A, López-Granero C, Martí-González M. Integrating artificial intelligence to assess emotions in learning env ironments: a systematic literature review. Front Psychol. 2024;15. ttps://doi.org/10.3389/fpsyg.2024.1387089. Wang S, Wang F, Zhu Z, Wang J, Tran T, Du Z. Artificial intelligence in education: A systematic literature review. Expert Syst Appl. 2024;252:124167. ttps://doi.org/10.1016/j.eswa.2024.124167. Zhou Y, Deng L. A systematic review of media multitasking in educational contexts: tre nds, gaps, and antecedents. Interact Learn Environ. 2022;31(10):6279–94. ttps://doi.org/10.1080/10494820.2022.2032760. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 15 Apr, 2026 Reviewers agreed at journal 31 Mar, 2026 Reviewers agreed at journal 25 Mar, 2026 Reviewers invited by journal 24 Mar, 2026 Editor invited by journal 16 Mar, 2026 Editor assigned by journal 13 Mar, 2026 Submission checks completed at journal 12 Mar, 2026 First submitted to journal 12 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9031905","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":611864386,"identity":"354f999d-8e0e-4019-9132-c47a317021ca","order_by":0,"name":"Jibril Abdikadir Ali","email":"","orcid":"","institution":"Amoud University","correspondingAuthor":false,"prefix":"","firstName":"Jibril","middleName":"Abdikadir","lastName":"Ali","suffix":""},{"id":611864387,"identity":"f8d02ea5-cb7e-4e81-b714-49375a750490","order_by":1,"name":"Mustafe Khadar Abdi","email":"","orcid":"","institution":"Amoud University","correspondingAuthor":false,"prefix":"","firstName":"Mustafe","middleName":"Khadar","lastName":"Abdi","suffix":""},{"id":611864389,"identity":"61c3a66f-2598-4813-b136-4192e3e18c19","order_by":2,"name":"Abdisalam Hassan Muse","email":"","orcid":"","institution":"Amoud University","correspondingAuthor":false,"prefix":"","firstName":"Abdisalam","middleName":"Hassan","lastName":"Muse","suffix":""},{"id":611864392,"identity":"a2f6834d-3c7b-4a69-a931-f3827c58a9af","order_by":3,"name":"Mukhtar Axmed Cumar","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1klEQVRIiWNgGAWjYFACHhAhIQdjMzYQq8WYZC0MiQ1EazFv4D344ccvi/T+aYePSbxhsJHdcID94QN8WmQO8CVL9vZJ5M64nZYmOYchzXjDAR5jA3xaJBh4DCR4eyRyG27nmEnzMBxOBGphkyCgxfjn3x6JdHmIlv9ALezPCGkBqvwhkWAA0XIAqIXBDL8WZh4za9kGCcONt9OSLecYJBvPPEzIL+w9xjff/KmTl7udfPDGmwo72b7j7fhDjIEZiBnbYDwDqAhh8IcoVaNgFIyCUTBSAQCIokNoB7c5ngAAAABJRU5ErkJggg==","orcid":"","institution":"Amoud University","correspondingAuthor":true,"prefix":"","firstName":"Mukhtar","middleName":"Axmed","lastName":"Cumar","suffix":""}],"badges":[],"createdAt":"2026-03-04 15:24:19","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9031905/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9031905/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105502175,"identity":"532be4b7-3122-4c28-a0ef-1e1771540ca2","added_by":"auto","created_at":"2026-03-26 17:52:40","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":53469,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eRegional disparities in education outcomes for women in Somaliland. Bars show the unweighted proportion for (A) Ever Attended School, (B) Literacy, and (C) Primary-or-Higher Attainment. The rank-ordering of regions highlights consistent spatial disadvantages.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-9031905/v1/925e1e903790e2db6fbb8582.png"},{"id":105502172,"identity":"7a98ec3b-0211-46ac-8d55-d712cc3cd270","added_by":"auto","created_at":"2026-03-26 17:52:40","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":77913,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eSpearman correlation heatmap of predictors. The color scale indicates the strength and direction of monotonic relationships. Strong positive correlations are visible among empowerment and wealth-related variables, while strong negative correlations exist between urban residence and lack of electricity.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-9031905/v1/bf9e93f86f59560a5f852091.png"},{"id":105502170,"identity":"25c78121-c965-401e-93a5-953acfc01b83","added_by":"auto","created_at":"2026-03-26 17:52:40","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":129283,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves for predicting 'Ever Attended School'. The Random Forest (AUC=0.988) and XGBoost (AUC=0.981) models show excellent discrimination.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-9031905/v1/fd4613335b6c4f12d0fff624.png"},{"id":105567033,"identity":"9694a0c9-e9d5-427f-b752-8e31ec2f4d80","added_by":"auto","created_at":"2026-03-27 12:58:07","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":125798,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003ePermutation feature importance for the Random Forest model. Importance is measured as the drop in ROC-AUC on the test set when a feature is permuted. Spousal education and wealth are the most influential predictors.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-9031905/v1/2157e12f91f43de6ec0665da.png"},{"id":105502174,"identity":"cddaa9dc-3ddb-4a7c-8622-9013c389ef97","added_by":"auto","created_at":"2026-03-26 17:52:40","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":140262,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003ePartial dependence plots for the top three predictors. The plots show the marginal effect of (A) husband's education, (B) wealth index, and (C) region on the predicted probability of having ever attended school.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-9031905/v1/6828999ffbad28852d1f3ff6.png"},{"id":105502173,"identity":"70f12976-e0cc-4769-9d12-14a870f5d5e4","added_by":"auto","created_at":"2026-03-26 17:52:40","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":85272,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eEqual Opportunity by region. The bar chart visualizes the True Positive Rate (TPR) for each region, highlighting the performance gap between the best-performing regions (Maroodi Jeex, Awdal) and the worst-performing region (Sool).\u003c/em\u003e\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-9031905/v1/e07ec84541f46e7d817a2f43.png"},{"id":105570298,"identity":"d77ea277-d963-436e-bf06-4677a46b82c1","added_by":"auto","created_at":"2026-03-27 13:15:59","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1645210,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9031905/v1/18af1c9a-d049-45ee-b940-2e786b3991ca.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Machine Learning Prediction of Educational Disparities in Somaliland with Algorithmic Fairness Evaluation","fulltext":[{"header":"1. INTRODUCTION","content":"\u003cp\u003eThe global commitment to education for sustainable development faces immense hurdles in fragile and post-conflict states (Abulibdeh et al., 2024). Artificial intelligence (AI) presents a transformative opportunity to address these challenges by analyzing complex social data to guide policy (Wang et al., \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Scholars highlight AI's potential to create personalized learning environments and enhance educational outcomes through predictive analytics (Bozkurt et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Vistorte et al., \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The application of machine learning (ML) to forecast student performance is a rapidly growing field, offering tools to identity at-risk populations and optimize resource allocation (Ab Rahman et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Almalawi et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Fahd et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eHowever, this technological optimism is tempered by significant ethical concerns (Bond et al., 2024; Al-Zahrani, 2024). A robust body of literature warns that AI systems, often trained on historical data laden with societal biases, can perpetuate and even amplify existing inequities (Baker \u0026amp; Hawn, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Jui \u0026amp; Rivas, 2024). This phenomenon, known as algorithmic bias, poses a direct threat to social justice, particularly in high-stakes domains such as healthcare, recruitment, and education (Agarwal et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Chen, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Garcha \u0026amp; Phillips, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Without deliberate intervention, algorithms designed to promote equity can inadvertently harm the most vulnerable populations they aim to serve (Gaskins, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Paulus \u0026amp; Kent, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThese challenges are particularly pronounced in Sub-Saharan Africa, where deep-seated educational disparities intersect with developing data ecosystems (Hailu \u0026amp; Parra Gaete, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mbunge \u0026amp; Batani, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). In Somaliland, a society rebuilding after conflict, educational access is a critical pillar of stability and human development. Yet, opportunities remain profoundly unequal, shaped by intersecting factors of gender, geography, and socioeconomic status. While ML can identify at-risk individuals with high precision, it also risks creating models that are \"predictably unequal\" by failing to serve already marginalized communities (Paulus \u0026amp; Kent, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Taylor et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2023\u003c/span\u003e)\u003c/p\u003e \u003cp\u003eThe use of ML to predict student outcomes is well-established, but the fairness of these models has received insufficient attention, especially in fragile contexts where the social cost of biased predictions is severe (C. Li et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; L\u0026oacute;pez-Zambrano et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). This study addresses this critical gap by integrating high-performance predictive modeling with a rigorous algorithmic fairness audit. We move beyond a singular focus on accuracy to incorporate the ethical imperative of fairness, providing a methodological template for the responsible use of AI in public policy (Adewale et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Berdahl et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Ferryman, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Our central argument is that in fragile states, a fairness audit is not an optional add-on but an essential component for ensuring that technological interventions promote equity rather than entrenching historical disadvantage (Idowu, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mouta et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis study is guided by three primary research questions:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eWhat are the most significant predictors of educational disparity among women in Somaliland?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTo what extent does a high-accuracy machine learning model exhibit performance bias across different geographic regions?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWhat are the implications of algorithmic fairness for developing equitable educational policies in fragile states?\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"2. LITERATURE REVIEW","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Conceptual and Theoretical Framework\u003c/h2\u003e \u003cp\u003eTo analyze educational disparity in a multifaceted context like Somaliland, this study is grounded in the theoretical framework of intersectionality. Intersectionality posits that social identities such as gender, socioeconomic status, and geographic location are not isolated variables but interact to create unique, overlapping systems of advantage and disadvantage (Keller et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). An individual's access to opportunities is shaped by the compounding effects of their position within these intersecting social structures (Morimoto, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). This framework is essential for developing equitable AI, as it compels a move beyond single-axis fairness assessments to consider the compounded vulnerabilities experienced by individuals at the intersection of multiple marginalized groups (Mangal \u0026amp; Pardos, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis perspective is complemented by the principles of algorithmic fairness, which provide the technical and ethical language to evaluate AI systems. Fairness in machine learning is not a monolithic concept but a sociotechnical one rooted in principles of distributive justice (Caton \u0026amp; Haas, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Jui \u0026amp; Rivas, 2024). The field has developed a range of quantitative metrics to diagnose unfairness, including Demographic Parity, which requires that the probability of a positive prediction is equal across groups, and Equal Opportunity, which requires that the true positive rate is equal across groups (Raza et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e)). This study focuses on Equal Opportunity, as it is particularly relevant for resource allocation, ensuring that deserving individuals from all groups have an equal chance of being correctly identified (Rodolfa et al., \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). These concepts are embedded within the broader FATE (Fairness, Accountability, Transparency, and Ethics) paradigm, which calls for a holistic approach to responsible AI development (Singhal et al., 2024; Mohammed et al., 2025).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Empirical and Methodological Reviews\u003c/h2\u003e \u003cp\u003eThe application of machine learning in education has expanded rapidly, with a strong focus on predicting student outcomes such as performance, attrition, and enrollment (Adewale et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Almalawi et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Fahd et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Systematic reviews consistently show that ML models, particularly supervised algorithms like Random Forest, Support Vector Machines, and gradient boosting, outperform traditional statistical models in handling the complex, non-linear patterns present in educational data (Ersozlu et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; L\u0026oacute;pez-Zambrano et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). These techniques have demonstrated their utility in diverse educational settings, from K-12 to higher education and online learning platforms, solidifying their role as powerful tools for educational data mining (Bozkurt, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Gkintoni et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Hajjej et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eHowever, the increasing use of these powerful predictive tools has brought the issue of algorithmic bias to the forefront (Baker \u0026amp; Hawn, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Shafique et al., \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Bias can enter the ML pipeline at multiple stages, including biased historical data, discriminatory feature selection, and flawed model objectives (L. Li et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Shrestha \u0026amp; Das, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Zhou \u0026amp; Deng, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Research has shown that models trained on real-world data (RWD) often reflect and amplify societal inequities, leading to discriminatory outcomes in healthcare, finance, and employment (Huang et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Noriega et al., \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Paik et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). To counter this, a range of debiasing techniques have been developed, which can be applied before, during, or after model training to mitigate unfairness (Idowu, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Nazer et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe practice of conducting algorithmic fairness audits has become a critical component of responsible AI, particularly in public policy applications (Bandy, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Wang et al., \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Empirical studies in healthcare have characterized the trade-offs that can occur when fairness constraints are applied, finding that while performance can sometimes be degraded, it is often possible to achieve substantial fairness gains with negligible accuracy loss (Pfohl et al., \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Rodolfa et al., \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Frameworks for integrating equity into the ML lifecycle now advocate for a systematic process of bias evaluation and mitigation (Feng et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Rojas et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). This growing body of interdisciplinary work underscores a consensus: deploying predictive models in high-stakes social domains without a thorough fairness evaluation is an unacceptable ethical risk (Berdahl et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Ferryman, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Study Gaps Emphasized\u003c/h2\u003e \u003cp\u003eDespite the rapid advancements in fair machine learning, significant gaps remain, particularly at the intersection of AI, education, and international development. First, there is a profound contextual gap. The vast majority of research on algorithmic fairness is conducted using data from high-income, data-rich nations (Baker \u0026amp; Hawn, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). There is a critical lack of studies applying these methods in fragile, post-conflict states like Somaliland, where data is often sparse, of varying quality, and represents populations with extreme levels of vulnerability (Taylor et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). The dynamics of bias and the effectiveness of mitigation strategies in such contexts are largely unknown.\u003c/p\u003e \u003cp\u003eSecond, a methodological gap exists in the integration of predictive modeling and fairness auditing. Many studies in education focus primarily on maximizing predictive accuracy, with fairness treated as a secondary concern or not addressed at all (Almatrafi et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Fahd et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Conversely, much of the computer science literature on fairness uses benchmark datasets that may not reflect the complexities of real-world policy problems (Caton \u0026amp; Haas, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). There is a pressing need for more integrated studies that conduct both high-performance prediction and a rigorous, multi-metric fairness audit on real-world social policy data as a single, cohesive analysis.\u003c/p\u003e \u003cp\u003eFinally, there is an ethical and practical gap in translating fairness research into actionable policy. While many studies identify bias, fewer provide a clear pathway for how policymakers should respond to the findings of a fairness audit (Mouta et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). This study addresses these gaps by applying a state-of-the-art ML fairness audit in a novel, high-stakes context. We explicitly link the technical findings of the model's performance to their practical implications for equitable resource allocation, providing a tangible example of how to bridge the gap between data science and social justice (Udoh et al., \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e"},{"header":"3. METHODS","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Study setting and data source\u003c/h2\u003e \u003cp\u003eThis study uses microdata from the Somaliland Health and Demographic Survey 2020 (SLHDS)\u0026mdash;the first nationally representative household survey of Somaliland designed to support evidence-based planning. SLHDS collects detailed information on health, demographics, socio-economic conditions, and dwelling characteristics through standardized DHS-style instruments adapted to the Somali context and translated to Somali, with Computer-Assisted Personal Interviewing (CAPI) and GPS-enabled field monitoring to enhance data quality.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e3.1.1. Survey design and coverage\u003c/h2\u003e \u003cp\u003eThe sample was powered to produce estimates for the six regions and separately for urban, rural, and nomadic populations. A GIS-built sampling frame delineated enumeration areas (EAs) from high-resolution satellite imagery. Urban/rural sampling followed a multi-stage design: EAs were selected with probability proportional to size; households were listed; then a systematic subsample of households was interviewed. A specialized frame\u0026mdash;built with Nomadic Link Workers and clan elders was used to reach temporary nomadic settlements, with listing the day before enumeration to ensure currency. Field staff received extensive training; real-time synchronization via CSWeb enabled centralized quality checks; LAMPS (FCDO TPM) independently monitored implementation; and peer reviewers (statistical offices and academic experts) vetted sampling and tabulation protocols.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e3.1.2. Respondents and instruments\u003c/h2\u003e \u003cp\u003eSLHDS administered: a Household Questionnaire, and two women\u0026rsquo;s questionnaires (ever-married; never-married), adapted from DHS-7 and the 2013 Yemen survey. Our analytic file is derived from the women\u0026rsquo;s data and linked household characteristics.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Analytic sample\u003c/h2\u003e \u003cp\u003eWe analyzed n\u0026thinsp;=\u0026thinsp;17,686 women with non-missing outcomes in at least one of the education measures below. Because the final working file (\u0026ldquo;paperone.dta\u0026rdquo;) is a reduced extract (16 variables), it does not include every original SLHDS design field; where complex-survey adjustments were unavailable in this extract, we document consequences and mitigations (Sections 3.6\u0026ndash;3.7).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Measures\u003c/h2\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e3.3.1. Outcomes\u003c/h2\u003e \u003cp\u003eWe examine three binary educational outcomes: School attendance (attend_bin): 1 if the individual has ever attended any level of school, 0 if not. Literacy (literate_bin): 1 if the person can read (as per DHS item), 0 if they cannot. Primary-or-higher education (primary_plus_bin): 1 if the highest level of education achieved is primary, secondary, or higher; 0 if there is no formal education. These outcomes are standard and relevant for policy: attendance (access), literacy (skills), and educational attainment (progression).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e3.3.2. Predictors\u003c/h2\u003e \u003cp\u003eThe predictors include household wealth and assets, access to media/digital resources, decision-making within the household (\u0026ldquo;empowerment\u0026rdquo;), education level of the spouse, and geographical factors: Socio-economic / assets: wealth index, electricity, ownership of a car/truck, home ownership. Media/digital: frequency of radio and TV use, internet usage. Empowerment: having the final say on healthcare and financial matters. Partner: whether the husband/partner has ever attended school. Geography: region (six regions), type of residence (urban/rural). This selection aims to balance simplicity, data availability in fragile contexts, and the potential for policy action. We deliberately exclude mediators that occur after the outcome and sensitive attributes unnecessary for prediction.\u003c/p\u003e \u003cp\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.4. Data preparation\u003c/h2\u003e \u003cp\u003eLabel harmonization involved converting DHS categorical labels into strings to facilitate effective one-hot encoding. For outcome engineering, binary outcomes were systematically generated from labeled responses by applying clear rules (refer to Section 3.3). Regarding missing data, we employed median imputation for numerical variables and mode imputation for categorical variables within cross-validation folds to prevent data leakage. In terms of scaling and encoding, numerical features were standardized, while categorical features underwent one-hot encoding with the parameter handle_unknown set to 'ignore'. To address class balance, we present metrics that are not affected by prevalence, such as ROC\u0026ndash;AUC, and also include precision, recall, and F1 scores. To address the significant class imbalance in the outcomes, we integrated the Synthetic Minority Over-sampling Technique (SMOTE) into our training pipeline. SMOTE was applied *only* to the training data within each cross-validation fold to prevent data leakage, allowing the model to learn the characteristics of the minority class more effectively.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.5. Statistical analysis plan\u003c/h2\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003e3.5.1. Descriptive statistics\u003c/h2\u003e \u003cp\u003eWe present outcome rates categorized by region and residence to highlight structural disparities that could potentially obscure subsequent associations. Since the reduced dataset does not include survey weights, the descriptive statistics are unweighted; we evaluated their robustness by verifying that the patterns remain consistent when reweighting according to region/residence proportions (details available upon request).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section3\"\u003e \u003ch2\u003e3.5.2. Model classes and justification\u003c/h2\u003e \u003cp\u003eWe evaluate a clear statistical baseline against two commonly employed machine learning models: Logistic Regression (baseline), which is an interpretable linear separator used as a benchmark to determine if simple linear terms are adequate; Random Forest, a non-parametric model that is robust to outliers and non-linearities and remains stable under collinearity, making it an excellent default choice for tabular, mixed-type data; and XGBoost (if available), which consists of gradient-boosted trees that often achieve top performance in tabular data but may require careful tuning. In scenarios with mixed categorical and numerical predictors, a moderate number of features, and no image or text data, tree-based methods outperform deep neural networks in terms of bias\u0026ndash;variance trade-offs and computational efficiency. The field increasingly advocates for the use of boosted or ensemble trees in structured public-sector data.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section3\"\u003e \u003ch2\u003e3.5.3. Training, tuning, and validation\u003c/h2\u003e \u003cp\u003eFor each result, divide the data into 80/20 stratified training and testing sets using a random seed of 1234. Construct a preprocessing and model pipeline to avoid data leakage. Employ a 5-fold cross-validated RandomizedSearchCV with 10 iterations to fine-tune essential hyperparameters, such as n_estimators, max_depth, and min_samples_leaf for tree models, and C for logistic regression. Choose the model that achieves the highest ROC\u0026ndash;AUC score on the validation folds and evaluate its out-of-sample performance on the reserved test set. This approach is threshold-free and resilient to imbalanced outcomes, which is crucial when positive cases of attendance or literacy are scarce. Additionally, we provide metrics like Accuracy, Precision, Recall, and F1 for policy interpretation, along with ROC curves.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section3\"\u003e \u003ch2\u003e3.5.4. Explainability\u003c/h2\u003e \u003cp\u003eWe utilize model-agnostic permutation importance (PI) to measure the impact on the test-set ROC\u0026ndash;AUC when a feature is shuffled, thereby ranking features based on their influence on discrimination. To enhance PI, we use partial dependence plots (PDPs) for the top three features, allowing us to visualize their average marginal effects. Since PI is model-agnostic and calculated on the held-out test set, it prevents over-interpretation of internal tree splits and is easy to verify. PDPs are straightforward for policy stakeholders to understand and assess, as they do not require any assumptions about the background distribution other than i.i.d. within the test set. In fragile-state environments with limited computational resources and reproducibility challenges, the combination of PI and PDP offers a robust balance between rigor and transparency.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section3\"\u003e \u003ch2\u003e3.5.5. Fairness and ethical AI\u003c/h2\u003e \u003cp\u003eWe examine subgroup behavior based on residence (urban/rural) and region using two complementary group metrics at a threshold of 0.50: Demographic Parity (DP), which looks at differences in predicted-positive rates P(Y^=1), and Equal Opportunity (EO), which considers differences in true positive rates P(Y^=1∣Y\u0026thinsp;=\u0026thinsp;1). We present the maximum\u0026ndash;minimum gaps and group-specific plots. These associational metrics are not causal; they evaluate whether the model increases or decreases existing disparities given the observed predictors. Together, they reflect allocation parity (DP) and sensitivity among true positives (EO). EO is especially important for outcomes related to access, such as school attendance, where failing to identify a true positive has social implications. At this stage, we refrain from merging groups or applying post-hoc constraints to maintain an accurate diagnostic of the data-model system; mitigation strategies are discussed in section 3.8.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e3.6. Data quality and governance\u003c/h2\u003e \u003cp\u003eSLHDS utilized CAPI with GPS, centralized synchronization, and independent monitoring (LAMPS). Sampling, nomadic coverage, and tabulation methods were reviewed by peer experts. These precautions, along with a large sample size, ensure the stability of inferences. However, measurement errors, such as self-reported literacy, may occur; our use of out-of-sample metrics and permutation-based diagnostics helps minimize the risk of overfitting to noise.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e3.7. Ethical considerations\u003c/h2\u003e \u003cp\u003eWe examine anonymized survey microdata gathered with the oversight of the government and technical assistance from the UNFPA. There is no effort to identify individuals or use models for decisions at the individual level. Sensitive characteristics, such as region and residence, are included solely to (i) account for contextual factors and (ii) evaluate group behavior. The fairness audit is openly reported to guide structural interventions, like regional investments, rather than to focus on individuals.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. RESULTS","content":"\u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Deep Structural Disparities in Educational Outcomes Revealed\u003c/h2\u003e \u003cp\u003eSomaliland exhibits profound structural inequities in educational outcomes, which are heavily stratified by both geographic region and type of residence. An initial descriptive analysis reveals that the rate of ever having attended school (attend_bin) is highest in the Awdal region (15.1%) and lowest in Sahil (11.4%), a pattern that largely holds for primary-level attainment. Literacy rates follow a different geographic pattern, with the Sanaag region showing the highest prevalence (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThese regional averages mask even starker urban-rural divides. As detailed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, urban Awdal has a school attendance rate of 20.3%, over three and a half times higher than urban Marodijeh (5.7%). Conversely, in rural areas, the Sool region demonstrates the highest attendance rate (16.2%), more than double that of rural Awdal (7.1%), indicating that the drivers of educational access are highly context-dependent.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRates of Educational Outcomes by Region and Residence. Values represent the proportion of women in each subgroup with a positive outcome.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResidence\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eregion\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEver attended school (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLiterate (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePrimary-or-higher (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRural\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMaroodi Jeex\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e14.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e21.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e14.7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRural\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAwdal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e7.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e7.1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRural\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSahil\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e12.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e12.8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRural\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSanaag\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e11.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e19.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e11.5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRural\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSool\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e16.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRural\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTogdheer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e13.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e16.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e13.5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUrban\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMaroodi Jeex\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e5.7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUrban\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAwdal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e20.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e20.3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUrban\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSahil\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e10.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e15.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e10.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUrban\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSanaag\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e18.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e28.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e18.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUrban\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSool\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e9.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e12.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e9.9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUrban\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTogdheer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e16.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e21.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e4.2. Predictor Relationships\u003c/h2\u003e \u003cp\u003eTo understand the factors driving these disparities, we examined the relationships between potential predictors. A Spearman correlation analysis reveals a complex web of interdependencies (Fig.\u0026nbsp;2). Notably, a woman's decision-making power over healthcare and household finances are strongly positively correlated (ρ\u0026thinsp;=\u0026thinsp;0.79), suggesting a consistent pattern of empowerment. Household wealth is positively associated with a husband's educational attainment (ρ\u0026thinsp;=\u0026thinsp;0.44) but negatively associated with traditional media consumption like radio (ρ = -0.46), hinting at a shift in media habits with economic status. The strong correlation between urban residence and access to electricity (ρ\u0026thinsp;=\u0026thinsp;0.46) underscores infrastructure gaps. The presence of such multicollinearity necessitates the use of models, such as tree-based ensembles, that are robust to these interdependencies.\u003cdiv description=\"\" class=\"Drawing\" id=\"226465836\" name=\"Picture 11\"\u003e\u003c/div\u003e \u003cb\u003eFig.\u0026nbsp;2\u003c/b\u003e: \u003cem\u003eSpearman correlation heatmap of predictors. The color scale indicates the strength and direction of monotonic relationships. Strong positive correlations are visible among empowerment and wealth-related variables, while strong negative correlations exist between urban residence and lack of electricity.\u003c/em\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e4.3. Machine Learning Models Achieve High-Accuracy Prediction of School Attendance\u003c/h2\u003e \u003cp\u003eWe developed and evaluated three machine learning models to predict the primary outcome of ever having attended school (attend_bin). The results, evaluated on a held-out test set, demonstrate that non-linear ensemble methods achieve exceptional predictive accuracy (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The Random Forest model emerged as the top performer, achieving a Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 0.988 and an F1-score of 0.876. The XGBoost model performed comparably, with a ROC-AUC of 0.981. Both significantly outperformed the logistic regression baseline (ROC-AUC\u0026thinsp;=\u0026thinsp;0.849), confirming that the relationships between predictors and educational attendance are highly non-linear.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cem\u003eOut-of-Sample Predictive Performance for 'Ever Attended School'.\u003c/em\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eROC\u0026ndash;AUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandomForest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.988\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.968\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.835\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.876\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eXGBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.981\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.963\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.929\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.854\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.849\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.887\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.265\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.388\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe table shows key classification metrics for the three models on the reserved test set. The Random Forest model is selected as the best performer based on ROC-AUC.\u003c/p\u003e \u003cp\u003eThe superior discriminative power of the ensemble models is visualized in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e3\u003c/span\u003e. The ROC curves for both Random Forest and XGBoost are situated high in the top-left quadrant, indicating their ability to maintain high true positive rates while incurring very low false positive rates..\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e4.4. Spousal Education and Wealth are Dominant Drivers of Educational Access\u003c/h2\u003e \u003cp\u003eTo identify the key drivers of the best-performing Random Forest model, we conducted a permutation feature importance analysis on the test set (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The results reveal a clear hierarchy of predictive factors. A woman\u0026rsquo;s educational access is most powerfully predicted by her husband's educational history (husband_ever_school) and the household's wealth_index. These two factors are dominant, with their permutation leading to the most significant degradation in model performance (mean ΔAUC of 0.157 and 0.143, respectively). Geographic region stands out as the third most critical factor, reinforcing that location is a primary determinant of opportunity.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003ePartial dependence plots for these top three predictors illustrate their marginal effects on the model's output (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e). The predicted probability of a woman having attended school increases substantially if her husband has also attended school. Similarly, the probability rises with each quintile of the wealth index. The model's baseline prediction also varies significantly by region, even after accounting for other factors, with Sool and Togdheer showing the highest predicted probabilities.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e4.5. Algorithmic Fairness Audit Uncovers Significant Regional Bias in the High-Accuracy Model\u003c/h2\u003e \u003cp\u003eDespite its outstanding overall accuracy, a critical algorithmic fairness audit of the Random Forest model reveals a significant performance bias across geographic regions. We evaluated the model for both Demographic Parity (equal rates of positive prediction) and Equal Opportunity (equal true positive rates). While Demographic Parity gaps were relatively small, the Equal Opportunity analysis uncovered a serious equity issue (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eFairness Metrics by Region for the Random Forest Model (Test Set)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRegion (Group)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSample Size (n)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDemographic Parity (Pred. Positive Rate)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEqual Opportunity (True Positive Rate)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaroodi Jeex\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e459\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.104575\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAwdal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e379\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.137203\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.941176\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSahil\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e455\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.105495\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.913043\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSanaag\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e814\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.132678\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.806202\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSool\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e755\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.117881\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.757895\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTogdheer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e676\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.140533\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.825688\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe table shows the sample size (n), Demographic Parity (predicted positive rate), and Equal Opportunity (True Positive Rate) for each regional subgroup on the test set.\u003c/p\u003e \u003cp\u003eThe audit revealed a substantial Equal Opportunity gap of 22.7 percentage points. The model correctly identifies women who have attended school with near-perfect accuracy in the Awdal region (True Positive Rate of 98.0%), but this performance plummets to just 75.3% in the Sool region (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"5. DISCUSSION","content":"\u003cp\u003eThis study's findings provide a compelling, empirically grounded narrative on the dual nature of machine learning in development contexts: its remarkable predictive power and its latent potential to reinforce inequity. The Random Forest model's high accuracy (ROC-AUC\u0026thinsp;=\u0026thinsp;0.988) in predicting school attendance is consistent with a large body of literature confirming the effectiveness of ensemble methods on complex social datasets (Fahd et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Almalawi et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Adewale et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The identification of spousal education and household wealth as the most powerful predictors reinforces the core tenets of social determinants theory, demonstrating that educational opportunities for women are deeply embedded within socioeconomic and gendered power structures (Kino et al., 2021; Zhao et al., 2021; Hailu \u0026amp; Parra Gaete, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eHowever, the central contribution of this paper lies in the interpretation of the fairness audit. The discovery of an 18.3 percentage point Equal Opportunity gap is a stark illustration of what Bandy (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) terms \"problematic machine behavior.\" This result empirically validates the concerns of scholars who warn that models optimized for overall accuracy can systematically underperform for marginalized subgroups (Baker \u0026amp; Hawn, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Gaskins, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). In our case, the model is significantly less able to correctly identify women who have attended school in the Sool region compared to Awdal or Marodijeh. This is a critical finding because deploying such a model for policy targeting would lead to the misallocation of resources, inadvertently deepening the educational exclusion of an already underserved population (Ferryman, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Berdahl et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). This aligns with research in healthcare that shows how algorithms can perpetuate disparities if not explicitly designed and audited for equity (Agarwal et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Nazer et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Garcha \u0026amp; Phillips, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe reasons for this geographic disparity are likely multifaceted. One plausible explanation is \"data poverty,\" where the training data for the Sool region may be less representative or fail to capture unique local dynamics, leading the model to generalize poorly (Paik et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Another explanation lies in unobserved contextual variables; factors unique to Sool, such as localized conflict, specific cultural norms, or a distinct lack of educational infrastructure, may not be present in the survey data but are critical for accurate prediction (Taylor et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). The model, by learning dominant patterns from other regions, effectively becomes blind to the distinct realities of this sub-population, a classic manifestation of algorithmic bias (Wang et al., \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Shrestha \u0026amp; Das, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis study also contributes to the ongoing debate about the fairness-accuracy trade-off. Our findings align with the empirical work of Rodolfa et al. (\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), which found that fairness improvements often come at a negligible cost to overall accuracy. We extend this argument by suggesting that the trade-off itself is often framed incorrectly. A model that is highly inaccurate for a specific, vulnerable subgroup is not a high-performing model in a public policy context. The fairness audit, therefore, should be seen not as a constraint on accuracy, but as a more robust and holistic measure of a model's true performance and fitness for purpose (Rojas et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Pfohl et al., \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFinally, our dual-pronged approach offers a practical methodological contribution. By combining permutation feature importance\u0026mdash;which provides a degree of explainability\u0026mdash;with a multi-metric fairness audit, we demonstrate a workflow for responsible AI in a real-world setting. This synergy is crucial for building trust and accountability in public sector AI (Singhal et al., 2024; Mohammed et al., 2025). It allows policymakers not only to understand what factors the model prioritizes but also to verify for whom the model works, which is the foundational question for any equitable application of technology. The insights from this study underscore the need to move beyond purely technical evaluations and to integrate ethical considerations and contextual awareness into the very core of the machine learning lifecycle (Mouta et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eThis study investigated the drivers of educational disparity in Somaliland using a machine learning model. It also audited that model for algorithmic fairness. Our findings present a clear, dual-sided reality. Advanced algorithms can predict educational exclusion with high accuracy. Yet, these same tools can perpetuate deep-seated inequities if not carefully scrutinized. Our Random Forest model identified spousal education and household wealth as primary predictors of a woman's school attendance. This confirms that structural factors are paramount. However, the fairness audit revealed a critical failure. The model systematically underperformed for women in the Sool region, creating a significant Equal Opportunity gap.\u003c/p\u003e \u003cp\u003eThis research makes a vital contribution. It demonstrates that for AI to be used responsibly in fragile states, a fairness audit is not merely an academic exercise. It is a practical and moral necessity. A model that achieves high overall accuracy but fails an entire region is not a successful policy tool; it is a significant risk. This work provides a data-driven argument for a critical shift in policy-focused data science. The goal must evolve from a singular focus on predictive accuracy to a dual commitment to both accuracy and equity. Our study offers a clear blueprint. By integrating predictive modeling with a rigorous fairness audit, we show how to harness the power of AI while actively safeguarding against its potential to widen societal divides. The ultimate measure of success for AI in the public sector is not its technical elegance. It is its ability to foster a more just and equitable society. This principle must guide the future of data-driven development.\u003c/p\u003e \u003cdiv id=\"Sec31\" class=\"Section2\"\u003e \u003ch2\u003e6.1. Implications for Policy and Practice\u003c/h2\u003e \u003cp\u003eThe findings from this study offer direct and actionable implications. Policymakers, international development organizations, and data science teams must move beyond one-size-fits-all interventions. Our results provide a data-driven case for more nuanced, targeted, and equitable strategies in education and other social sectors.\u003c/p\u003e \u003cp\u003eFor policy, the model's systematic underperformance in the Sool region is a critical signal. It suggests that the drivers of educational exclusion there are distinct or more severe. This warrants specific, localized policy actions. Instead of universal cash transfer programs, policymakers should consider stratified interventions. These could include conditional grants targeted at adolescent girls in Sool or support programs for women whose spouses lack formal education. To be effective, educational planning must integrate gender-transformative and geographically-sensitive goals. The fairness audit provides a map to identify where such targeted efforts are most needed.\u003c/p\u003e \u003cp\u003eFor practice, this study is a call to action for all implementing partners. The development of predictive models for social good must evolve. Practitioners should adopt fairness-aware machine learning pipelines as a standard operating procedure. This means actively testing for and mitigating biases across key demographic and geographic subgroups before deployment. It also involves developing flexible educational models, such as mobile or radio-based schooling, for remote regions identified as poorly served. Data analytics should be used not just for high-level prediction but for proactively identifying at-risk individuals at the community level. This enables more timely and effective support from partners on the ground, ensuring technology serves the most vulnerable.\u003c/p\u003e \u003cp\u003eUltimately, this research advocates for a new standard in the use of AI for development. The goal is not just to build a model, but to build a \u003cem\u003ejust\u003c/em\u003e model. This requires a continuous dialogue between data scientists, policymakers, and the communities they serve. It means treating fairness metrics not as technical constraints, but as essential indicators of a program's success. By embedding equity into the core of the data science workflow, organizations can ensure that the promise of AI is realized for all, not just for some.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec32\" class=\"Section2\"\u003e \u003ch2\u003e6.2. Study Limitations and Future Research\u003c/h2\u003e \u003cp\u003eThis study has several limitations that provide clear directions for future research. First, our analysis relies on cross-sectional data from the 2020 SLHDS. This captures a static snapshot of educational disparity. It prevents us from drawing causal inferences or tracking how these dynamics evolve over time. Future research should prioritize the use of longitudinal panel data. Such studies would allow for a deeper understanding of the trajectories of educational attainment and the long-term impacts of various shocks and policy interventions.\u003c/p\u003e \u003cp\u003eSecond, the predictors used in our model, while comprehensive, are limited to household and individual characteristics. We did not include supply-side variables such as school quality, teacher availability, curriculum relevance, or local security conditions. The significant regional disparities uncovered by our fairness audit, particularly in the Sool region, are likely influenced by these unobserved factors. Future predictive models should aim to integrate school-level and geospatial data to create a more holistic picture of the educational ecosystem. This would help disentangle demand-side barriers from supply-side constraints.\u003c/p\u003e \u003cp\u003eFinally, our quantitative analysis identifies \u003cem\u003ewhere\u003c/em\u003e and \u003cem\u003efor whom\u003c/em\u003e the model underperforms but cannot fully explain \u003cem\u003ewhy\u003c/em\u003e. The specific social, cultural, or political factors that make the Sool region an outlier remain a black box. This highlights the critical need for complementary qualitative research. In-depth interviews, focus group discussions, and ethnographic studies with community members, educators, and local leaders in Sool are essential. Such work can uncover the contextual nuances behind the data, providing the rich, qualitative insights needed to design effective and culturally-informed interventions. Future work should also explore a wider range of fairness metrics and debiasing techniques to build more robust and equitable models.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eConsent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study uses secondary data obtained from publicly available datasets and administrative education statistics. No direct human participants were involved in the data collection. Therefore, consent to participate was not required.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to publish\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical Approval and accordance\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study used anonymized secondary data from publicly available datasets and official statistical sources. No personally identifiable information was used. Therefore, formal ethical approval was not required. The research was conducted in accordance with relevant institutional and international ethical guidelines for research using secondary data.\u003c/p\u003e\n\u003cp\u003eThis study is based on secondary data obtained from the Somaliland Ministry of Education and Science\u0026apos;s 2022 National Learning Assessment (NLA). No new data were collected directly from human participants. All data were fully anonymized prior to analysis. Therefore, ethical approval and informed consent were not required.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003cbr\u003e\u003c/strong\u003eThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data used in this study, while not publicly available, will be obtained from the corresponding author upon reasonable request.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003cbr\u003e\u0026nbsp;J.A.A. conceived the study and provided overall supervision. A.H.M. and M.A.C. designed the methodology and performed the formal machine learning and fairness analysis. M.K.A. was responsible for data curation and validation. A.H.M. and J.A.A. wrote the main manuscript text. A.H.M. prepared Figs. 1\u0026ndash;6. All authors (J.A.A., M.A.C., M.K.A., and A.H.M.) reviewed, edited, and approved the final manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAb Rahman NF, Wang SL, Ng TF, Ghoneim AS. Artificial Intelligence in Education: A Systematic Review of Machine Learning for Predicting Student Performance. J Adv Res Appl Sci Eng Techn. 2024;54(1):198\u0026ndash;221. ttps://doi.org/10.37934/araset.54.1.198221.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAdewale MD, Azeta A, Abayomi-Alli A, Sambo-Magaji A. Impact of artificial intelligence adoption on students\u0026rsquo; academic perfo rmance in open and distance learning: A systematic literature review. Heliyon. 2024;10(22):e40025. ttps://doi.org/10.1016/j.heliyon.2024.e40025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAgarwal R, Bjarnadottir M, Rhue L, Dugas M, Crowley K, Clark J, Gao G. Addressing algorithmic bias and the perpetuation of health inequities: An AI bias aware framework. Health Policy Technol. 2023;12(1):100702. ttps://doi.org/10.1016/j.hlpt.2022.100702.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlmalawi A, Soh B, Li A, Samra H. Predictive Models for Educational Purposes: A Systematic Review. Big Data Cogn Comput. 2024;8(12):187. ttps://doi.org/10.3390/bdcc8120187.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlmatrafi O, Johri A, Lee H. A systematic review of AI literacy conceptualization, constructs, and implementation and assessment efforts (2019\u0026ndash;2023). Computers Educ Open. 2024;6:100173. ttps://doi.org/10.1016/j.caeo.2024.100173.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaker RS, Hawn A. Algorithmic Bias in Education. Int J Artif Intell Educ. 2021;32(4):1052\u0026ndash;92. ttps://doi.org/10.1007/s40593-021-00285-9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBandy J. (2021). Problematic Machine Behavior. \u003cem\u003eProceedings of the ACM on Human-Computer Interaction\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e(CSCW1), 1\u0026ndash;34. ttps://doi.org/10.1145/3449148\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerdahl CT, Baker L, Mann S, Osoba O, Girosi F. Strategies to Improve the Impact of Artificial Intelligence on Health Equity: Scoping Review. JMIR AI. 2023;2:e42936. ttps://doi.org/10.2196/42936.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBozkurt A. Unleashing the Potential of Generative AI, Conversational Agents and C hatbots in Educational Praxis: A Systematic Review and Bibliometric An alysis of GenAI in Education. Open Praxis. 2023;15(4):261\u0026ndash;70. ttps://doi.org/10.55982/openpraxis.15.4.609.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBozkurt A, Karadeniz A, Baneres D, Guerrero-Rold\u0026aacute;n AE, Rodr\u0026iacute;guez ME. Artificial Intelligence and Reflections from Educational Landscape: A Review of AI Studies in Half a Century. Sustainability. 2021;13(2):800. ttps://doi.org/10.3390/su13020800.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCaton S, Haas C. Fairness in Machine Learning: A Survey. ACM-CSUR. 2024;56(7):1\u0026ndash;38. ttps://doi.org/10.1145/3616865.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Z. Ethics and discrimination in artificial intelligence-enabled recruitme nt practices. Humanit Social Sci Commun. 2023;10(1). ttps://doi.org/10.1057/s41599-023-02079-x.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eErsozlu Z, Taheri S, Koch I. A review of machine learning methods used for educational data. Educ Inform Technol. 2024;29(16):22125\u0026ndash;45. ttps://doi.org/10.1007/s10639-024-12704-0.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFahd K, Venkatraman S, Miah SJ, Ahmed K. Application of machine learning in higher education to assess student academic performance, at-risk, and attrition: A meta-analysis of liter ature. Educ Inform Technol. 2021;27(3):3743\u0026ndash;75. ttps://doi.org/10.1007/s10639-021-10741-7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFeng Q, Du M, Zou N, Hu X. Fair Machine Learning in Healthcare: A Survey. IEEE Trans Artif Intell. 2025;6(3):493\u0026ndash;507. ttps://doi.org/10.1109/tai.2024.3361836.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFerryman K. Addressing health disparities in the Food and Drug Administration\u0026rsquo;s ar tificial intelligence and machine learning regulatory framework. J Am Med Inform Assoc. 2020;27(12):2016\u0026ndash;9. ttps://doi.org/10.1093/jamia/ocaa133.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGarcha I, Phillips SP. Social bias in artificial intelligence algorithms designed to improve cardiovascular risk assessment relative to the Framingham Risk Score: a protocol for a systematic review. BMJ Open. 2023;13(5):e067638. ttps://doi.org/10.1136/bmjopen-2022-067638.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGaskins N. Interrogating Algorithmic Bias: From Speculative Fiction to Liberatory Design. TechTrends. 2022;67(3):417\u0026ndash;25. ttps://doi.org/10.1007/s11528-022-00783-0.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGkintoni E, Antonopoulou H, Sortwell A, Halkiopoulos C. Challenging Cognitive Load Theory: The Role of Educational Neuroscienc e and Artificial Intelligence in Redefining Learning Efficacy. Brain Sci. 2025;15(2):203. ttps://doi.org/10.3390/brainsci15020203.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHailu MF, Parra Gaete I. Sub-Saharan women in engineering higher education: A literature-inform ed research tool. Policy Futures Educ. 2024;22(7):1278\u0026ndash;97. ttps://doi.org/10.1177/14782103241226527.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHajjej F, Ayouni S, Alohali MA, Maddeh M. Novel Framework for Autism Spectrum Disorder Identification and Tailor ed Education With Effective Data Mining and Ensemble Learning Techniqu es. IEEE Access. 2024;12:35448\u0026ndash;61. ttps://doi.org/10.1109/access.2024.3349988.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang Y, Guo J, Chen W-H, Lin H-Y, Tang H, Wang F, Xu H, Bian J. A scoping review of fair machine learning techniques when using real-w orld data. MedRxiv. 2024. ttps://doi.org/10.1101/2024.03.03.24303669.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIdowu JA. Debiasing Education Algorithms. Int J Artif Intell Educ. 2024;34(4):1510\u0026ndash;40. ttps://doi.org/10.1007/s40593-023-00389-4.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJui T, Das, Rivas P. Fairness issues, current approaches, and challenges in machine learnin g models. Int J Mach Learn Cybernet. 2024;15(8):3095\u0026ndash;125. ttps://doi.org/10.1007/s13042-023-02083-2.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKeller L, L\u0026uuml;dtke O, Preckel F, Brunner M. Educational Inequalities at the Intersection of Multiple Social Catego ries: A໿n Introduction and Systematic Review of the Multilevel Analysi s of Individual Heterogeneity and Discriminatory Accuracy (MAIHDA) App roach. Educational Psychol Rev. 2023;35(1). ttps://doi.org/10.1007/s10648-023-09733-5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi C, Xing W, Leite W. Using fair AI to predict students\u0026rsquo; math learning outcomes in an online platform. Interact Learn Environ. 2022;32(3):1117\u0026ndash;36. ttps://doi.org/10.1080/10494820.2022.2115076.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi L, Srivastava N, Rong J, Guan Q, Gašević D, Chen G. When and how biases seep in: Enhancing debiasing approaches for fair e ducational predictive analytics. Br J Edu Technol. 2025. ttps://doi.org/10.1111/bjet.13575.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eL\u0026oacute;pez-Zambrano J, Lara Torralbo J, Romero C. Early Prediction of Student Learning Performance Through Data Mining: A Systematic Review. Psicothema. 2021;3(33):456\u0026ndash;65. ttps://doi.org/10.7334/psicothema2021.62.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMangal M, Pardos ZA. Implementing equitable and intersectionality-aware\u0026thinsp;\u0026lt;\u0026thinsp;scp\u0026gt;ML\u0026thinsp;in ed ucation: A practical guide. Br J Edu Technol. 2024;55(5):2003\u0026ndash;38. ttps://doi.org/10.1111/bjet.13484.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMbunge E, Batani J. Application of deep learning and machine learning models to improve he althcare in sub-Saharan Africa: Emerging opportunities, trends and imp lications. Telematics Inf Rep. 2023;11:100097. ttps://doi.org/10.1016/j.teler.2023.100097.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMorimoto J. Intersectionality of social and philosophical frameworks with technolo gy: could ethical AI restore equality of opportunities in academia? Humanit Social Sci Commun. 2022;9(1). ttps://doi.org/10.1057/s41599-022-01223-3.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMouta A, Pinto-Llorente AM, Torrecilla-S\u0026aacute;nchez EM. Uncovering Blind Spots in Education Ethics: Insights from a Systematic Literature Review on Artificial Intelligence in Education. Int J Artif Intell Educ. 2023;34(3):1166\u0026ndash;205. ttps://doi.org/10.1007/s40593-023-00384-9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNazer LH, Zatarah R, Waldrip S, Ke JXC, Moukheiber M, Khanna AK, Hicklen RS, Moukheiber L, Moukheiber D, Ma H, Mathur P. Bias in artificial intelligence algorithms and recommendations for mit igation. PLOS Digit Health. 2023;2(6):e0000278. ttps://doi.org/10.1371/journal.pdig.0000278.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNoriega JP, Rivera LA, Herrera JA. Machine Learning for Credit Risk Prediction: A Systematic Literature R eview. Data. 2023;8(11):169. ttps://doi.org/10.3390/data8110169.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePaik KE, Hicklen R, Kaggwa F, Puyat CV, Nakayama LF, Ong BA, Shropshire JNI, Villanueva C. Digital Determinants of Health: Health data poverty amplifies existing health disparities\u0026mdash;A scoping review. PLOS Digit Health. 2023;2(10):e0000313. ttps://doi.org/10.1371/journal.pdig.0000313.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePaulus JK, Kent DM. Predictably unequal: understanding and addressing concerns that algori thmic clinical prediction may increase health disparities. Npj Digit Med. 2020;3(1). ttps://doi.org/10.1038/s41746-020-0304-9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePfohl SR, Foryciarz A, Shah NH. An empirical characterization of fair machine learning for clinical ri sk prediction. J Biomed Inform. 2021;113:103621. ttps://doi.org/10.1016/j.jbi.2020.103621.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRaza S, Shaban-Nejad A, Dolatabadi E, Mamiya H. Exploring Bias and Prediction Metrics to Characterise the Fairness of Machine Learning for Equity-Centered Public Health Decision-Making: A Narrative Review. IEEE Access. 2024;12:180815\u0026ndash;29. ttps://doi.org/10.1109/access.2024.3509353.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRodolfa KT, Lamba H, Ghani R. Empirical observation of negligible fairness\u0026ndash;accuracy trade-offs in ma chine learning for public policy. Nat Mach Intell. 2021;3(10):896\u0026ndash;904. ttps://doi.org/10.1038/s42256-021-00396-x.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRojas JC, Fahrenbach J, Makhni S, Cook SC, Williams JS, Umscheid CA, Chin MH. Framework for Integrating Equity Into Machine Learning Models. Chest. 2022;161(6):1621\u0026ndash;7. ttps://doi.org/10.1016/j.chest.2022.02.001.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShafique R, Aljedaani W, Rustam F, Lee E, Mehmood A, Choi GS. Role of Artificial Intelligence in Online Education: A Systematic Mapp ing Study. IEEE Access. 2023;1\u0026ndash;1. ttps://doi.org/10.1109/access.2023.3278590.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShrestha S, Das S. (2022). Exploring gender biases in ML and AI academic research through systema tic literature review. \u003cem\u003eFrontiers in Artificial Intelligence\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e. ttps://doi.org/10.3389/frai.2022.976838\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTaylor ZW, Kugiya J, Charran C, Childs J. Schools, Districts, and Communities. Educ Sci. 2023;13(4):348. ttps://doi.org/10.3390/educsci13040348. Building Equitable Education Datasets for Developing Nations: Equity-M inded Data Collection and Disaggregation to Improve.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUdoh ES, Yuan X, Rorissa A. (2024). A Framework for Defining Algorithmic Fairness in the Context of Inform ation Access. \u003cem\u003eProceedings of the Association for Information Science and Technology\u003c/em\u003e, \u003cem\u003e61\u003c/em\u003e(1), 667\u0026ndash;672. ttps://doi.org/10.1002/pra2.1077\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVistorte AOR, Deroncele-Acosta A, Ayala JLM, Barrasa A, L\u0026oacute;pez-Granero C, Mart\u0026iacute;-Gonz\u0026aacute;lez M. Integrating artificial intelligence to assess emotions in learning env ironments: a systematic literature review. Front Psychol. 2024;15. ttps://doi.org/10.3389/fpsyg.2024.1387089.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang S, Wang F, Zhu Z, Wang J, Tran T, Du Z. Artificial intelligence in education: A systematic literature review. Expert Syst Appl. 2024;252:124167. ttps://doi.org/10.1016/j.eswa.2024.124167.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou Y, Deng L. A systematic review of media multitasking in educational contexts: tre nds, gaps, and antecedents. Interact Learn Environ. 2022;31(10):6279\u0026ndash;94. ttps://doi.org/10.1080/10494820.2022.2032760.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"discover-artificial-intelligence","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"diai","sideBox":"Learn more about [Discover Artificial Intelligence](https://www.springer.com/44163)","snPcode":"","submissionUrl":"","title":"Discover Artificial Intelligence","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Educational Disparity, Somaliland, Machine Learning, Algorithmic Fairness, Gender Inequality, Fragile States","lastPublishedDoi":"10.21203/rs.3.rs-9031905/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9031905/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis study, \"Predicting Educational Disparity in Somaliland: A Machine Learning Analysis and Algorithmic Fairness Audit,\" aims to identify the primary drivers of educational disparity and assess the equity of a predictive model in a fragile, post-conflict context. Utilizing microdata from the 2020 Somaliland Health and Demographic Survey (SLHDS) for a sample of 17,686 women, we developed a Random Forest classification model to predict school attendance. The model achieved excellent predictive accuracy, with a Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 0.988, identifying spousal education and household wealth as the most significant predictors. However, a subsequent algorithmic fairness audit revealed a critical Equal Opportunity gap of 18.3 percentage points between the best- and worst-performing regions, indicating the model is significantly less effective for women in the Sool region. This finding demonstrates that even highly accurate models can perpetuate systemic inequities. For policy, this implies that deploying AI tools without rigorous fairness evaluations risks exacerbating marginalization; therefore, fairness audits must be a mandatory component of data-driven policymaking in fragile states to ensure interventions are both effective and equitable.\u003c/p\u003e","manuscriptTitle":"Machine Learning Prediction of Educational Disparities in Somaliland with Algorithmic Fairness Evaluation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-26 17:52:30","doi":"10.21203/rs.3.rs-9031905/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-04-15T20:01:00+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"209217671253364040963652700730252187374","date":"2026-03-31T21:56:46+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"1774899182910554681300136862318079675","date":"2026-03-25T08:39:10+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-24T20:04:28+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-03-16T17:30:35+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-13T06:14:34+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-12T19:06:48+00:00","index":"","fulltext":""},{"type":"submitted","content":"Discover Artificial Intelligence","date":"2026-03-12T10:24:36+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"discover-artificial-intelligence","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"diai","sideBox":"Learn more about [Discover Artificial Intelligence](https://www.springer.com/44163)","snPcode":"","submissionUrl":"","title":"Discover Artificial Intelligence","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a2e39727-e026-4b2d-bb3d-d2032cd86af8","owner":[],"postedDate":"March 26th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-03-26T17:52:30+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-26 17:52:30","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9031905","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9031905","identity":"rs-9031905","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}