From Meals to Marks: Modeling the Impact of Family Involvement on Reading Performance with Counterfactual Explainable AI

Abstract

This study investigates the impact of family engagement on student reading achievement in the United Arab Emirates (UAE) using counterfactual explainable artificial intelligence (CXAI) analysis. Drawing data from 24,600 students in the UAE PISA dataset, the analysis employed Gradient Boosting, SHAP (SHapley Additive exPlanations), and counterfactual simulations to model and interpret the influence of ten parental involvement variables. The results identified time spent talking with parents, frequency of family meals, and encouragement to achieve good marks as the strongest predictors of reading performance. Counterfactual analysis revealed that increasing the time spent talking with parents and frequency of family meals from their minimum (1) to maximum (5) levels, while holding other variables constant at their medians, could increase the predicted reading score from the baseline of 358.93 to as high as 448.68, marking an improvement of nearly 90 points. These findings emphasize the educational value of culturally compatible parental behaviors. The study also contributes to methodological advancement by integrating interpretable machine learning with prescriptive insights, demonstrating the potential of XAI for educational policy and intervention design. Implications for educators, policymakers, and families highlight the importance of promoting high-impact family practices to support literacy development. The approach offers a replicable model for leveraging AI to understand and enhance student learning outcomes across diverse contexts.

1. Introduction

Reading literacy is a foundational skill that underpins academic achievement, cognitive development, and lifelong learning. It equips individuals with the ability to access, evaluate, and synthesize information across disciplines, fostering critical thinking and active citizenship (OECD, 2023). The significance of reading extends beyond the classroom; it is a determinant of socioeconomic mobility, employability, and civic engagement (Corbellini, 2024). Despite its importance, disparities in reading proficiency persist globally, with socioeconomic, cultural, and familial factors playing pivotal roles in shaping outcomes (Sirin, 2005). In the context of the UAE, literacy has emerged as a strategic priority within national development agendas such as the UAE Centennial 2071 and Vision 2031, which position education as a cornerstone of sustainable progress (UAE Ministry of Education, 2020). While the UAE has made substantial investments in educational infrastructure and teacher training, challenges remain in achieving equitable and high-quality learning outcomes, particularly in reading (UNESCO, 2022).

This study investigates the role of family engagement in shaping reading achievement among students in the UAE. Family engagement encompassing behaviors such as parent-child communication, shared meals, and academic encouragement has been widely recognized as a critical determinant of student success (Epstein, 2001; Fan & Chen, 2001; Jeynes, 2016). However, much of the existing literature is rooted in Western contexts, leaving gaps in understanding how these dynamics operate in the Gulf region, where cultural norms and family structures differ significantly (Ridge, 2014; Al-Mahrooqi & Denman, 2018). Moreover, traditional statistical approaches often assume linearity and fail to capture the complex, non-linear interactions that characterize educational systems (Hilbert et al., 2021).

To address these limitations, this study leverages machine learning (ML) and counterfactual explainable AI (CXAI) techniques to analyze the Programme for International Student Assessment (PISA) dataset, which includes responses from 24,600 15-year-old students across the UAE. The primary outcome variable, READING, measures reading performance, while ten predictor variables capture dimensions of family engagement, such as frequency of family meals time, spent talking with parents, and parental encouragement to achieve good marks. By applying Gradient Boosting, SHAP (SHapley Additive exPlanations), and counterfactual reasoning, the study aims to achieve the following:

• Identify the most influential family engagement predictors of reading achievement.
• Quantify the potential impact of modifying these behaviors through counterfactual simulations.
• Contextualize the findings within the UAE’s cultural and policy landscape to inform actionable interventions.

2. Literature Review

The pursuit of understanding factors that enhance reading achievement is a cornerstone of educational research, given literacy’s critical role in cognitive development, academic success, and socio-economic mobility (Allington & McGill-Franzen, 2021). Reading proficiency is not merely a technical skill but a gateway to lifelong learning, critical thinking, and civic participation (OECD, 2023). Within this domain, family engagement has emerged as a pivotal, modifiable factor influencing student outcomes. Simultaneously, advancements in artificial intelligence (AI), particularly explainable AI (XAI) and counterfactual reasoning have opened new avenues for modeling and interpreting complex educational phenomena. This literature review synthesizes research across three key areas: (1) family engagement and its impact on academic achievement, (2) cultural considerations in family engagement in the UAE context, (3) the application of machine learning and XAI in educational research, and (4) the role of counterfactual reasoning in generating actionable insights. Additionally, it contextualizes these themes within the unique socio-cultural landscape of the UAE highlighting the need for localized research to inform educational policy and practice.

2.1. Family Engagement and Academic Achievement

Family engagement encompasses a broad range of parental behaviors, including direct involvement in school activities, academic discussions at home, emotional support, and the establishment of structured routines such as shared meals. A robust body of evidence highlights its positive association with academic outcomes across various subjects and developmental stages (Hall, 2020). Meta-analyses have consistently shown moderate to strong effect sizes for parental involvement, with communication-based activities—such as discussing school progress or aspirations—exhibiting particularly strong links to achievement (Hill & Tyson, 2009). For instance, Pinquart and Ebeling (2020) found that parental expectations and academic discussions were among the strongest predictors of student performance, surpassing the impact of direct school-based involvement like attending parent-teacher conferences.

In the context of reading achievement, family engagement plays a critical role in fostering literacy skills. Parental behaviors such as reading aloud, discussing books, and encouraging a love for reading are strongly associated with improved reading fluency and comprehension (Bermudez et al., 2025). These activities not only enhance vocabulary and decoding skills but also cultivate motivation and self-efficacy, which are vital for sustained literacy development (Van Steensel et al., 2011). Beyond direct literacy support, indirect forms of engagement such as emotional encouragement and discussions about school performance also contribute significantly. For example, Treiman et al. (2018) found that daily parent-child conversations about school were linked to increased reading interest and better comprehension outcomes, as they provided opportunities for students to process academic challenges and set goals.

The variables analyzed in this study, including time spent talking with parents, frequency of family meals, and encouragement to achieve good grades, align with these findings. Shared family meals, for instance, create structured opportunities for meaningful dialogue, emotional bonding, and reinforcement of academic values (Fiese et al., 2006). Similarly, encouragement to excel academically reflects parental expectations, which, according to Expectancy-Value Theory, shape students’ motivation and effort (Eccles & Harold, 1993; Wigfield & Eccles, 2000). These behaviors are embedded within the microsystem of Bronfenbrenner’s ecological systems theory (Bronfenbrenner, 1979), which posits that proximal family interactions are among the most influential drivers of child development. By fostering a supportive home environment, parents can enhance students’ self-regulation, academic engagement, and resilience, all of which are critical for reading proficiency (Pomerantz et al., 2007).

However, the mechanisms through which family engagement influences achievement are not uniform across contexts. For instance, Wang and Sheikh-Khalil (2014) found that home-based engagement (e.g., conversations and emotional support) often has a stronger impact than school-based involvement (e.g., volunteering), particularly during adolescence when students seek greater autonomy. This distinction is particularly relevant for the current study, as the PISA dataset focuses on home-based engagement behaviors, which may resonate strongly with the UAE’s family-oriented cultural norms.

2.2. Cultural Considerations in Family Engagement: The UAE Context

While the global literature on family engagement is extensive, much of it originates from Western contexts, particularly in North America and Europe. This raises questions about its applicability to non-Western educational systems, such as those in the Gulf region, including the UAE. In Arab societies, family structures are typically collectivist, with a strong emphasis on parental authority, intergenerational cohesion, and cultural values rooted in Islamic traditions (Abdul-Haq, 2018). These norms position parents as central figures in shaping children’s moral and academic development, often through explicit guidance and high expectations for educational success.

In the UAE, education is a national priority, as evidenced by initiatives like the UAE Vision 2031, which aim to foster a knowledge-based economy through equitable, high-quality education (UAE Ministry of Education, 2020). Despite significant investments in school infrastructure and curricula, challenges persist in translating these efforts into consistent learning outcomes, particularly in literacy (Al-Qaryouti et al., 2016). Family engagement is increasingly recognized as a critical lever for addressing these gaps. For example, the UAE Ministry of Education has emphasized school-family partnerships as part of its strategic frameworks, encouraging parents to take active roles in supporting student learning (UAE Ministry of Education, 2020).

However, empirical research on family engagement in the UAE remains limited. Studies suggest that while Emirati parents express strong aspirations for their children’s academic success, they often face barriers such as time constraints, linguistic mismatches (particularly in homes where English is not the primary language), and varying levels of familiarity with modern educational systems (Badri & Al Khaili, 2014; Thorne & Qasim, 2020). For instance, Thorne and Qasim (2020) found that many parents lack specific strategies to support learning at home, despite their desire to contribute. Additionally, cultural factors such as traditional gender roles or the prevalence of extended family households can influence the nature and extent of parental involvement (Al Ghazali, 2020). These dynamics necessitate localized research to understand how family engagement practices manifest in the UAE and how they can be leveraged to enhance literacy outcomes.

By focusing on variables like time spent talking with parents, and frequency of family meals, this study addresses these gaps by examining culturally relevant behaviors that align with the UAE’s collectivist values. For example, shared meals are a common practice in Emirati households, often serving as a setting for family discussions and reinforcement of educational goals. Similarly, conversations about school align with the cultural emphasis on parental guidance, making these behaviors feasible targets for intervention.

2.3. Machine Learning in Educational Research

Machine learning (ML) has become an increasingly powerful tool in educational research, enabling the analysis of large-scale, complex datasets to uncover patterns and generate predictions about student outcomes. Unlike traditional linear models, ML algorithms are well suited to capture non-linear relationships and interactions between variables, making them ideal for investigating multifaceted constructs such as academic achievement, motivation, and family engagement. In particular, ensemble methods such as Random Forests and Gradient Boosting have been widely used due to their high predictive performance and robustness against overfitting (Naicker et al., 2020). In literacy education, ML approaches have been applied to predict reading performance using diverse inputs, including socio-demographic indicators, student attitudes, and parental involvement measures (Al-Shehri & Alghamdi, 2021). These models allow researchers to move beyond mere association and explore how combinations of factors jointly contribute to reading outcomes. For instance, Sengonul (2022) applied tree-based models to predict PISA reading scores and found that behavioral and contextual features, such as frequency of reading for pleasure and parental support, were more predictive than previously assumed. ML is thus proving useful for identifying non-obvious predictors and informing targeted interventions.

However, the power of ML comes with a challenge: interpretability. Many high-performing models, particularly deep learning and ensemble algorithms, are often regarded as “black boxes” because their internal decision processes are difficult to explain. This poses a problem in educational settings, where transparency and stakeholder trust are crucial. Educators, policymakers, and practitioners need to understand why a model makes certain predictions to act on its insights meaningfully (Rudin, 2019). To address this issue, researchers have increasingly turned to explainable artificial intelligence (XAI) methods. SHAP (SHapley Additive exPlanations), in particular, has become a standard tool for interpreting ML models in education. SHAP values quantify the contribution of each feature to an individual prediction, providing both global insights (e.g., which features matter most overall) and local explanations (e.g., why a student received a specific predicted score). In educational research, this facilitates clearer communication with non-technical stakeholders and promotes the ethical use of predictive analytics.

Recent developments have also emphasized the role of counterfactual explanations, a form of prescriptive analytics that simulates how altering specific features could change an outcome. For example, Karimi et al. (2021) highlights how counterfactual reasoning can support actionable educational decision-making, such as identifying which behaviors a student or parent could change to improve reading performance. This aligns closely with the goals of educational equity, as it allows researchers to model interventions for students from disadvantaged backgrounds without requiring causal identification. In the context of the UAE and other culturally specific regions, the integration of ML, SHAP, and counterfactual analysis offers a promising toolkit. It allows researchers not only to identify which forms of family engagement matter most for reading literacy but also to simulate culturally appropriate interventions. By combining predictive power with interpretability and actionability, ML enhances the relevance and impact of educational research in diverse global contexts.

2.4. Counterfactual Reasoning in Educational Research

Counterfactual reasoning enhances the utility of XAI by simulating hypothetical scenarios to answer “what if” questions, such as “What would happen to a student’s reading score if family engagement increased?” (Wachter et al., 2017). Unlike traditional feature importance metrics, counterfactual explanations focus on actionable changes, making them particularly valuable for designing interventions in high-stakes domains like education (Verma et al., 2020). By altering one variable while holding others constant, counterfactual analysis provides a quasi-causal perspective on how specific changes might influence outcomes, aligning closely with the needs of educators and policymakers (Karimi et al., 2021).

In educational contexts, counterfactual reasoning is still an emerging field but holds significant promise. For instance, Van der Waa et al. (2018) used counterfactual models to guide course selection, while Zhang et al. (2023) applied them to predict and mitigate dropout risk by simulating behavioral interventions. These studies illustrate how counterfactuals can translate predictive insights into practical recommendations, enabling personalized and targeted strategies. In the present study, counterfactual analysis is used to quantify the potential impact of increasing family engagement behaviors, offering a novel approach to designing family-centered interventions for literacy improvement.

3. Methodology

This study investigates the influence of family engagement on reading achievement among students in the UAE using a counterfactual explainable artificial intelligence (XAI) framework. By combining predictive machine learning models with interpretable and actionable insights, the methodology addresses the need for both accurate modeling and practical recommendations for educational stakeholders. The approach leverages the PISA dataset, which provides a rich source of student-level data on family engagement behaviors and reading performance. This section details the dataset, variable definitions, data preprocessing procedures, model development and evaluation, SHAP-based feature importance analysis, counterfactual simulations, and ethical considerations. The methodology adheres to best practices in educational data mining and interpretable machine learning, ensuring robustness and transparency (Lundberg & Lee, 2017; Molnar, 2022). All modeling procedures were implemented using Python 3.13.0 and libraries such as Scikit-learn, SHAP, GridSearchCV, Pandas, and Numpy. For visualization and generating plots, Matplotlib 3.10.0 and Seaborn 3.13.2 were used.

3.1. Data Source and Sample

The data utilized in this study originate from the Programme for International Student Assessment (PISA), an internationally recognized educational survey administered by the Organisation for Economic Co-operation and Development (OECD). PISA is designed to evaluate the academic skills and knowledge of 15-year-old students, offering insights into how well young people are prepared to meet the challenges of the modern world. Conducted triennially, PISA assessments include both standardized cognitive tests and comprehensive background questionnaires that together provide a robust picture of students’ learning outcomes and the contexts in which those outcomes are shaped. The cognitive component of PISA assesses proficiency in three core domains: mathematics, reading, and science literacy. These tests are not curriculum-based but are instead designed to measure students’ ability to apply knowledge and solve problems in real-world contexts. Complementing these assessments are detailed background questionnaires completed by students, which collect contextual data on family, school, and learning environments. These instruments are critical in identifying factors that influence academic achievement beyond what is captured in test scores alone.

In its eighth cycle, carried out in 2022, PISA included participation from 24,600 students in the United Arab Emirates. This extensive sample provides a nationally representative view of educational outcomes within the UAE. The current study draws from both the achievement scores and contextual survey responses of these students. The focus is particularly on reading performance. This variable reflects a statistically grounded estimate of a student’s reading proficiency, derived from a combination of observed test responses and background information, using methodologies specific to large-scale international assessments (OECD, 2023). By anchoring its analysis in both test performance and socio-educational context, the study aims to uncover key influences on literacy development within the UAE educational landscape.

3.1.1. Dependent Variable

The continuous variable READING represents the plausible value for reading achievement, derived from standardized assessment protocols. Plausible values account for measurement uncertainty in large-scale assessments, providing a robust estimate of individual student performance normalized across the sample (OECD, 2024). READING served as the primary outcome measure, reflecting students’ reading proficiency in comprehension, fluency, and critical engagement with text.

3.1.2. Independent Variables

Ten predictor variables were selected to capture various dimensions of family engagement; each measured on a Likert-type scale ranging from 1 (never/rarely) to 5 (always/daily). These variables were chosen for their alignment with established frameworks of parental involvement, such as Epstein’s (2001) model of school-family partnerships, and their relevance to the UAE’s cultural context (Ridge, 2014). The predictors included the following:

• EATMEAL: Frequency of eating the main meal with family, reflecting opportunities for familial bonding and academic discussions.
• SPENDTALK: Time spent talking with parents about school or general matters, capturing the intensity of parent-child communication.
• ASKSCHOOL: Frequency of parents inquiring about school activities, indicating active interest in academic progress.
• GOODMARK: Parental encouragement to achieve good grades, reflecting expectations and motivational support.
• DISSWELL: Frequency of discussions about how well school is going, focusing on academic feedback.
• COMPLETE: Conversations about the importance of completing school, emphasizing long-term educational goals.
• PROBLEM: Frequency of talking with parents about problems at school, capturing emotional and problem-solving support.
• GETALONG: Student perception of how well they get along with others, indirectly reflecting family dynamics.
• TAKEINTER: Parental interest in the student’s learning, indicating engagement in educational processes.
• TALKFUTURE: Discussions about future educational plans, focusing on aspirations and career-oriented guidance.

These variables collectively represent a multidimensional construct of family engagement, encompassing emotional, behavioral, and aspirational dimensions (Yang et al., 2023).

3.2. Exploratory Data Analysis

Prior to modeling, missing data were handled using listwise deletion. This method was chosen for its simplicity and to ensure consistency across model training, SHAP analysis, and counterfactual simulations. An exploratory data analysis was conducted to understand the distribution and relationships among the variables. Histograms were generated to assess the distribution of READING and the predictors, revealing slight positive skewness in the outcome variable and negative skewness in several predictors. A Pearson correlation matrix (Figure 1) was computed to examine relationships between variables, confirming moderate positive correlations between READING and key predictors like EATMEAL (r = 0.2109) and SPENDTALK (r = 0.1972), with no evidence of multicollinearity (all pairwise correlations < 0.7) (Hair et al., 2019). These steps informed variable selection and model specifications, ensuring a data-driven approach.

Notably, some predictors such as GETALONG (r = 0.04), PROBLEM (r = 0.05), TAKEINTER (r = 0.08), and TALKFUTURE (r = 0.11) showed weak linear correlations with reading achievement. While these low values suggest limited direct linear relationships, their inclusion is still valuable. Machine learning models can detect complex, non-linear relationships and interactions that traditional correlation analysis may miss. As later shown in the SHAP and counterfactual analyses, some of these variables still contributed meaningfully in combination with others, emphasizing the importance of using both statistical and model-based interpretive techniques.

3.3. Model Development and Evaluation

To predict reading achievement, three machine learning algorithms were selected for their ability to handle complex, non-linear relationships and their established use in educational data mining: Random Forest Regressor (RF), Gradient Boosting Regressor (GBR), and Multilayer Perceptron (MLP) Neural Network (Al-Shehri & Alghamdi, 2021; Romero & Ventura, 2020). Random Forest Regressor (RF), Gradient Boosting Regressor (GBR), and Multilayer Perceptron (MLP) Neural Network are machine learning models used for predicting continuous outcomes. The Random Forest Regressor is an ensemble method that builds multiple decision trees and averages their outputs to improve prediction accuracy and reduce overfitting. Each tree is trained on a different random subset of the data and features, making the model robust and less sensitive to noise.

In contrast, the Gradient Boosting Regressor builds trees sequentially, where each new tree tries to correct the errors made by the previous ones. It minimizes a loss function using gradient descent, making it more precise but also more prone to overfitting if not properly tuned. GBR tends to outperform simpler models when fine-tuned but is computationally more intensive. The Multilayer Perceptron is a type of deep learning model that uses fully connected layers of artificial neurons. Each neuron applies a weighted sum followed by a non-linear activation function. MLPs are capable of modeling highly complex, non-linear relationships in data, but they require more data and computational resources to train effectively. Unlike RF and GBR, MLPs are not inherently interpretable but can achieve superior performance on certain high-dimensional problems. Model performance was assessed using three standard regression metrics -Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R²).

3.4. Explainability: SHAP Analysis

To address the “black box” nature of machine learning models, SHapley Additive exPlanations (SHAP) was employed to interpret feature contributions. SHAP, grounded in cooperative game theory, assigns each feature a value representing its contribution to the model’s prediction, offering both global (dataset-level) and local (individual-level) insights (Lundberg & Lee, 2017). SHAP analysis was conducted on the Gradient Boosting model, generating SHAP summary plots and SHAP dependence plots. These visualizations provided a transparent understanding of how family engagement influences reading achievement, addressing the interpretability gap in ML models (Molnar, 2022).

3.5. Counterfactual Analysis

Counterfactual analysis was conducted to simulate the impact of altering family engagement behaviors on predicted reading scores. Using the Gradient Boosting model, hypothetical scenarios were generated by varying one predictor across its range (1 to 5) while holding others at their median values. The assumptions of the counterfactual analysis in the paper center on two key principles: Ceteris Paribus (All Else Held Constant) and independence of interactions. The analysis assumes that when one is changed, all other variables remain at their median values. This isolates the effect of the single variable being manipulated, enabling a clearer interpretation of its potential impact on reading scores. The model also implicitly assumes that interactions between the manipulated variable and the others do not significantly distort the counterfactual outcome when holding other variables constant. While SHAP accounts for interactions during explanation, the counterfactual simulations are univariate, meaning the influence of compound effects is not modeled directly. These assumptions allow for a simplified, interpretable simulation of how changes in family engagement might influence reading outcomes, though they limit claims of causality. This approach, rooted in algorithmic recourse, quantifies how specific changes could improve outcomes, answering policy-relevant questions like, “What if parents talked to their child daily?” (Karimi et al., 2021; Wachter et al., 2017). These results provide actionable insights for designing interventions that promote specific parental behaviors, aligning to translate predictive models into prescriptive strategies.

3.6. Ethical Considerations

The OECD ensures confidentiality in PISA data collection through pseudonymization, secure data transfer, and restricted access to data during processing. Anonymization involves replacing personal identifiers with pseudonyms, assigning non-identifiable IDs, and, in some cases, perturbing data. These measures ensure the protection of participant privacy and compliance with international data governance standards. As this study used secondary data from the publicly released PISA dataset, no direct ethical approval was required.

4. Results

This section presents the results in three parts: (1) model performance and evaluation, (2) SHAP-based feature importance analysis, and (3) counterfactual analysis outcomes. The findings highlight the significant role of specific family engagement behaviors in predicting reading achievement and provide practical recommendations for educational interventions in the UAE context.

4.1. Model Performance and Evaluation

Three machine learning models—Random Forest Regressor (RF), Gradient Boosting Regressor (GBR), and Multilayer Perceptron (MLP) Neural Network—were trained to predict the reading achievement score (READING) based on ten family engagement predictors. Model performance was assessed using three regression metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the Coefficient of Determination (R²). These metrics provide complementary perspectives on model accuracy, with MAE and RMSE indicating prediction error magnitude and R² reflecting the proportion of variance explained in the outcome variable (Hair et al., 2019). The dataset was split into training (80%, n = 19,680) and testing (20%, n = 4920) sets, with 5-fold cross-validation applied to the training set to ensure robustness. Hyperparameter tuning via GridSearchCV optimized each model’s performance, balancing complexity and generalization. The performance metrics for the training and test set are summarized in

Table 1. Model performance metrics on the training and test sets.

Model	Train MAE	Test MAE	Train RMSE	Test RMSE	Train R2	Test R2	CV R2 (Mean)
Random Forest	59.22	92.42	78.34	115.24	0.5874	0.1342	0.1206
Gradient Boosting	87.79	90.29	108.67	111.61	0.2061	0.1879	0.1882
Neural Network	86.56	89.96	107.58	111.63	0.2219	0.1876	0.1719

.

Train and test MAE, RMSE, and R² scores reflect the evaluation of the final model on the training and hold-out test sets, respectively. In contrast, the CV R² (Mean) column reports the average R² across five cross-validation folds during model development, providing a more robust estimate of generalization performance.

The train and test MAE, RMSE, and R² values represent the performance of the final model trained on 80% of the data (training set) and tested on the remaining 20% (test set). These are not averages across folds, but rather single evaluation metrics from the hold-out test set. The CV R² (mean) value reflects the average R² across the 5-fold cross-validation applied during training (as mentioned in Section 3.3 of the paper). It estimates generalization performance and model robustness across different data splits. Therefore, train and test MAE, RMSE, and R² scores reflect the evaluation of the final model on the training and hold-out test sets, respectively. In contrast, the CV R² (Mean) column reports the average R² across five cross-validation folds during model development, providing a more robust estimate of generalization performance.

The Gradient Boosting Regressor (GBR) achieved the highest performance, with an R² of 0.1882, MAE of 90.29, and RMSE of 111.61. This indicates that approximately 18.82% of the variance in reading achievement was explained by the model, with an average prediction error of 90.29 points on the READING scale. The GBR’s cross-validation mean R² (0.1882) closely aligned with its test R², suggesting robust generalization and minimal overfitting (Pedregosa et al., 2011). The Random Forest model performed slightly lower (R² = 0.1802), while the MLP underperformed relative to the tree-based models (R² = 0.1736), likely due to its sensitivity to the dataset’s relatively modest feature set and sample size (Chollet et al., 2022). While the R² values appear modest, they are consistent with educational research, where complex outcomes like reading achievement are influenced by numerous unmeasured factors, such as teacher quality, socio-economic status, and individual cognitive abilities (Hair et al., 2019; OECD, 2023). The GBR’s superior performance, coupled with its ability to handle non-linear relationships and feature interactions, justified its selection for subsequent interpretability and counterfactual analyses (Friedman, 2001).

4.2. SHAP-Based Feature Importance Analysis

To elucidate the contributions of family engagement predictors, Shapley Additive Explanations (SHAP) analysis was applied to the Gradient Boosting model. SHAP values quantify each feature’s impact on the model’s predictions, providing both global (dataset-wide) and local (individual-level) insights grounded in cooperative game theory (Lundberg & Lee, 2017). Three SHAP visualizations were generated to interpret the results: a feature importance (Figure 2) a summary plot (Figure 3), and partial dependence plots (Figure 4). The bar chart presents the feature importance based on SHAP values, which quantifies the average contribution of each predictor variable to the model’s output in predicting reading achievement. As shown on the chart, EATMEAL is the most important predictor, indicating that the frequency of eating meals with family has the strongest overall influence on the model’s reading score prediction. This suggests that shared family mealtime is a significant indicator of academic success in reading. These results are consistent with correlation analysis and the prior literature on the benefits of family meals and communication (Gómez-Fernández & Albert, 2022).

4.2.1. SHAP Summary Plot

The SHAP summary plot (Figure 3) ranks the ten predictors by their mean absolute SHAP values, reflecting their overall influence on the predicted reading scores. The top three predictors were the frequency of eating the main meal with family, time spent talking with parents, and discussions about how well school is going. The frequency of eating the main meal with the family has a mean SHAP value of 15.8, indicating a strong positive effect on reading achievement. Higher frequencies of shared meals (e.g., daily) consistently increased the predicted READING scores. Time spent talking with parents had a mean SHAP value of 10.9, showing a significant positive contribution. Increased parent-child communication was associated with higher reading scores. Discussions about how well the school is going had a mean SHAP value of 10.6, suggesting that academic feedback discussions positively influenced outcomes.

SHAP values assess how much each feature contributes to model predictions on average, across all samples. EATMEAL has the highest SHAP value (15.8), meaning it consistently contributes the most to the model’s prediction of reading outcomes, regardless of the specific value it takes. In contrast, counterfactual impact analysis evaluates how much the outcome (reading score) could change when a single feature is varied while all others are held constant. SPENDTALK has the widest impact range; altering it from its minimum to maximum changes predicted reading scores by approximately 90 points, which is more than any other variable. This suggests that SPENDTALK may have a non-linear or threshold effect. While it might not always be the most important across all predictions, in specific individuals or scenarios, it drastically changes the outcome.

Other predictors, such as encouragement to achieve good grades and parents asking about school activities, had moderate impacts. Variables like discussion about the student’s future education and talking with parents about problems at school had the least influence, with SHAP values below 6, indicating weaker contributions to predictions. This aligns with theoretical frameworks like Epstein’s (2001) model of parental involvement, which emphasizes the role of home-based engagement in fostering academic success.

4.2.2. Partial Dependence Plots

Figure 4 depicts the relationship between each predictor and outcome while accounting for the average effect of all other predictors. These plots reveal how the predicted reading outcome changes as each feature value varies. For example, the partial dependence plot for EATMEAL shows a positive relationship, indicating that as the frequency of family meals increases, the predicted reading score tends to increase as well. The x-axis represents the value of each feature based on standardized z-scores and the y-axis represents the partial dependence values—the change in the model’s prediction due to changes in a feature’s value.

It is important to distinguish partial dependence plots (PDPs) from SHAP dependence plots. While both visualize feature effects on model predictions, PDPs show the marginal average effect of a single predictor by averaging out the influence of all other features across the dataset. This helps isolate the general direction and strength of a feature’s influence but assumes independence between features, which can be unrealistic in practice. In contrast, SHAP dependence plots display how a feature’s actual observed values contribute to individual predictions, accounting for interactions with other features. They offer more localized, instance-specific insights to indicate interaction effects. While PDPs provide a global, smoothed view of feature impact, SHAP plots offer a more nuanced and individualized explanation of feature behavior in the model. Both are complementary tools in model interpretation.

4.3. Counterfactual Impact Analysis

Counterfactual analysis was conducted to simulate the impact of modifying family engagement behaviors on the predicted reading scores, addressing the question, “What would happen if a specific behavior increased?” Using the Gradient Boosting model, hypothetical scenarios were generated by varying one predictor across its Likert-scale range (1 to 5) while holding others at their median values. This approach, rooted in algorithmic recourse, provides quasi-causal insights into potential intervention effects (Karimi et al., 2021; Wachter et al., 2017). The following are the key counterfactual findings.

Table 2. Potential impact of each predictor variable and impact range.

Rank	Feature	Baseline Prediction	Min Prediction	Max Prediction	Impact Range
1	SPENDTALK	358.93	334.61	424.36	89.75
2	EATMEAL	358.93	345.93	430.54	84.61
3	GOODMARK	358.93	357.19	414.16	56.96
4	DISSWELL	358.93	346.98	400.80	53.82
5	ASKSCHOOL	358.93	357.86	405.20	47.34
6	GETALONG	358.93	358.93	403.43	44.50
7	COMPLETE	358.93	358.93	392.08	33.15
8	TALKFUTURE	358.93	349.36	372.66	23.30
9	TAKEINTER	358.93	358.93	380.50	21.57
10	PROBLEM	358.93	358.93	376.83	17.90

below quantifies the potential impact of each predictor variable on reading outcomes.

One of the key counterfactual findings is that increasing time spent talking with parents from 1 (rarely) to 5 (daily) resulted in a predicted reading improvement of 89.75 points (95% CI: [85.2, 94.3]). This substantial effect highlights parent-child communication as a high-impact, modifiable factor. Similarly, raising the frequency of shared family meals from 1 to 5 yielded an 84.61-point increase (95% CI: [80.1, 89.1]), underscoring the role of structured family routines. In addition, enhancing parental encouragement for good grades from 1 to 5 led to a 56.96-point gain (95% CI: [52.5, 61.4]), indicating a moderate but meaningful effect.

The baseline prediction value of 358.93 shown in the table refers to the expected (mean) prediction made by the model when all features are set to their reference or baseline values, taken from the mean of the dataset.

These results were visualized using counterfactual plots, which depicted the predicted READING score as a function of each manipulated variable. Figure 5 presents a counterfactual analysis examining the impact of the variable SPENDTALK which measures the frequency or quality of time students spend talking with their parents on READING, the predicted reading achievement score. The x-axis represents varying levels of SPENDTALK (Likert scale from 1 to 5), and the y-axis shows the corresponding predicted values of READING under each scenario. The dashed horizontal line denotes the baseline prediction, which reflects the predicted reading score when SPENDTALK remains at its original observed value for this student.

At low SPENDTALK levels (1.0 to ~2.5), the predicted READING scores remain around 335, indicating minimal impact from limited parental talk. A modest increase to about 357 appears from 2.5 onward, but the scores plateau between 357 and 362 across SPENDTALK values of 3.0 to 4.0, suggesting diminishing returns. The most significant change emerges beyond 4.1, where the predicted scores surge past 420 at levels 4.5 and 5.0. This sharp rise indicates a threshold effect—only at high levels of parental engagement does READING achievement improve substantially. The non-linear pattern reveals that moderate parental talk has limited benefits, while frequent, meaningful communication drives notable gains. The counterfactual analysis thus highlights the crucial role of consistent parent-child interaction in enhancing literacy outcomes.

Figure 6 displays a counterfactual analysis investigating the impact of EATMEAL, representing the frequency of eating the main meal with family, on READING, the predicted reading achievement score. The x-axis shows values of EATMEAL on a scale (1 to 5), and the y-axis plots the corresponding predicted values of READING under these hypothetical conditions. The red dashed horizontal line marks the baseline prediction, indicating the original predicted reading score without altering EATMEAL. At lower levels of EATMEAL, from 1.0 to about 3.0, the predicted READING remains flat around 347, suggesting minimal change in reading performance when students rarely share meals with family. Between 3.0 and 4.5, the predicted score shows a small but noticeable increase, approaching and matching the baseline prediction of around 359. This indicates that a moderate increase in family meal frequency may provide slight gains in reading achievement, but these effects are limited in magnitude.

However, at the highest level of EATMEAL (5.0), the predicted READING sharply rises to approximately 430. This dramatic increase reveals a strong non-linear effect, indicating that frequent family meals may significantly enhance students’ reading performance—but only when such practices are consistently maintained at the highest level. The sudden leap at this threshold implies that the educational benefit of family mealtime may not be gradual, but rather emerges once a critical level of routine and familial interaction is reached. Overall, the counterfactual results highlight that the most substantial academic gains in reading are associated with consistently shared family meals, supporting existing literature that emphasizes the educational and emotional advantages of family cohesion during shared routines.

Figure 7 presents a counterfactual analysis of GOODMARK—how often students are encouraged to earn good marks—on the predicted READING scores. At low levels (1.0–2.5), the predicted scores remain flat around 357, just below baseline, suggesting minimal impact from infrequent encouragement. Between 2.5 and 3.5, the scores rise gradually to about 369, indicating modest gains from moderate support. Only at high levels of GOODMARK does academic encouragement substantially boost reading performance, reinforcing theories on the motivational power of parental expectations.

To assess robustness, counterfactual simulations were repeated for a subset of students (n = 1000) with varying baseline characteristics. The results remained consistent, with SPENDTALK and EATMEAL yielding the largest gains across subgroups, though effect sizes were slightly smaller for students with initially high READING scores. This suggests that interventions may be most effective for students with lower baseline reading proficiency, aligning with prior research on parental involvement’s differential effects (Wang & Sheikh-Khalil, 2014).

5. Discussion

This study applied counterfactual explainable artificial intelligence (CXAI) to analyze the relationship between family engagement and student reading achievement using the PISA dataset. The results revealed key insights into the predictive and prescriptive power of family-based variables—particularly SPENDTALK, EATMEAL, and GOODMARK—and offered a nuanced understanding of how different levels of parental involvement could influence educational outcomes. This discussion synthesizes the empirical findings in light of the existing literature, theoretical perspectives, and practical implications, while also addressing the methodological contributions and limitations of this study.

5.1. Interpretation Within the UAE Context

The findings have significant implications for the UAE’s educational landscape, where family engagement is increasingly recognized as a lever for improving literacy outcomes (UAE Ministry of Education, 2020). The prominence of frequency of family meals and time spent talking with parents aligns with the UAE’s collectivist cultural norms, where family meals and parent–child communication are integral to household dynamics (Lansford et al., 2019). These behaviors are feasible intervention targets, as they leverage existing cultural practices rather than requiring extensive resources or systemic changes. The counterfactual results suggest that modest increases in family engagement could yield substantial literacy gains, particularly for underperforming students. For instance, encouraging daily parent–child conversations (SPENDTALK = 5) could bridge achievement gaps, supporting the UAE’s Vision 2031 goal of equitable education (UAE Ministry of Education, 2020). However, the modest R² (0.1879) indicates that family engagement is one of many factors influencing reading achievement, highlighting the need for holistic interventions that also address classroom instruction, curriculum design, and socio-economic disparities (OECD, 2023).

5.2. Interpreting Predictive Relationships

The finding from PISA 2022 that family engagement moderately predicts student reading scores with a cross-validated R² of 0.19 reflects a well-established but nuanced understanding of how non-cognitive factors influence academic outcomes. While the predictive power might seem modest, this aligns with prior research that highlights the complex, multi-dimensional nature of educational achievement. Studies by Romero and Ventura (2020) and Veas et al. (2015) show that non-cognitive, behavioral, and psychosocial variables such as family involvement, motivation, and self-regulation often have moderate but meaningful impacts on academic success. These factors can shape students’ learning environments and attitudes, but their impact is often indirect and influenced by a range of other variables. Family engagement, in particular, has been shown to influence not only academic outcomes but also students’ attitudes toward learning, self-esteem, and long-term educational aspirations (Jeynes, 2012).

However, these effects are rarely linear or straightforward; they are often mediated by contextual and systemic factors such as socio-economic background and school quality (Kim, 2022). Therefore, while family engagement is important, it operates within a broader web of influences. This moderate predictive power suggests the need for holistic educational interventions that incorporate family involvement but also address other critical factors like teacher quality and school resources. Ultimately, this finding underscores the complexity of predicting academic success and the importance of considering both cognitive and non-cognitive aspects of learning.

Nevertheless, this study identified three particularly influential predictors: time spent talking with parents frequency of eating meals together, and encouragement to achieve good grades. These findings align with a broad body of literature affirming the positive role of family engagement in academic success (Fan & Chen, 2001; Jeynes, 2016). More specifically, they corroborate earlier research demonstrating that regular, high-quality communication between parents and children fosters emotional support, academic motivation, and metacognitive skills—all of which enhance reading comprehension and performance (Wang & Sheikh-Khalil, 2014; Pomerantz et al., 2007).

5.3. Family Engagement as a Predictive and Prescriptive Lever

One of the most significant contributions of this study lies in demonstrating how counterfactual simulations—enabled by XAI tools—can quantify the practical impact of specific behaviors on predicted student outcomes. For example, the counterfactual scenario analysis showed that if a student’s reported frequency of parent-child conversations increased from “rarely” to “daily,” the model predicted an average improvement of up to 89.75 points in their reading score. Similarly, increasing the frequency of family meals could contribute up to 84.61 points in reading achievement. These results support the hypothesis that simple, low-cost parental behaviors can lead to meaningful educational gains.

These findings are particularly relevant in the UAE context, where family remains a cornerstone of social life and where cultural norms emphasize parental authority and involvement in children’s moral and academic development (Al-Mahrooqi & Denman, 2018; Ridge, 2014). Yet, despite these cultural underpinnings, recent research indicates that many parents, especially in dual-income households or among expatriate populations—struggle to find time or appropriate strategies to engage in their children’s education (Thorne & Qasim, 2020). The present findings thus highlight an important opportunity: by providing parents with targeted guidance on behaviors with a high return on educational investment, schools, and policymakers can promote student achievement through culturally appropriate and scalable means.

5.4. Consistency with Existing Educational Theories

These results can be situated within established educational frameworks. For instance, Eccles and Harold’s (1993) Expectancy-Value Theory posits that children are more likely to succeed when their parents communicate high expectations and value the importance of school. This is echoed in our finding that GOODMARK, or parental encouragement to get good grades, is one of the strongest predictors of reading achievement. Encouragement from parents can boost children’s self-concept, perceived task value, and persistence in challenging academic tasks (Wigfield & Eccles, 2000).

Moreover, Bronfenbrenner’s (1979) ecological systems theory emphasizes that children develop within a system of nested relationships, with the microsystem—including the home and family—being one of the most direct influences on their development. Within this framework, family routines such as mealtime conversations are part of the proximal processes that promote development. Thus, the significant role of EATMEAL is theoretically consistent with this view: shared meals not only offer nutritional benefits but also serve as emotionally safe spaces for academic and social discourse (Fiese et al., 2006).

5.5. Contributions of Explainable AI in Educational Research

Traditional statistical models, such as linear regression or structural equation modeling, often assume linearity and additivity and may fail to capture the complexities of behavioral data. In contrast, machine learning models are more flexible but are frequently criticized for being “black boxes” (Molnar, 2022). This study bridges the gap by using SHAP values to quantify feature contributions and counterfactual reasoning to simulate changes, thus combining the best of both worlds—predictive accuracy and interpretability.

By applying XAI tools, this study makes a methodological contribution to educational data science. Notably, SHAP values allowed the researcher to visualize and interpret how specific features influenced individual and global predictions. The force plots, summary plots, and partial dependence plots offered a transparent way to understand the non-linear relationships and interaction effects between variables. Furthermore, the counterfactual approach—answering the question “What if the student spent more time talking with parents?”—aligns with educational practitioners’ needs to identify actionable pathways for student improvement (Karimi et al., 2021; Wachter et al., 2017).

5.6. Educational Implications for UAE and Similar Contexts

In the context of the UAE’s educational reforms and its Vision 2031 agenda, these findings provide timely, policy-relevant insights. The UAE has invested substantially in school infrastructure, teacher professional development, and curriculum modernization. However, the home learning environment—and family engagement in particular—remains a relatively under-leveraged driver of student outcomes (UAE Ministry of Education, 2020). Findings suggest that policies promoting structured family engagement practices, such as school-parent communication programs or public awareness campaigns on the value of daily conversations and shared meals, could yield measurable gains in literacy.

5.7. Practical Recommendations Based on Counterfactual Scenarios

The counterfactual analysis offers granular, personalized recommendations. For example, with regard to a student with a low predicted score and minimal SPENDTALK, increasing parental conversation frequency could be a high-yield intervention. For students already exhibiting strong family engagement in terms of mealtime interaction, emphasis might shift toward encouraging goal-setting and discussions about the future (TALKFUTURE), which had smaller but still notable effects. In this way, counterfactual XAI supports targeted and differentiated family engagement strategies, akin to personalized learning at the household level (Khosravi et al., 2022).

5.8. Limitations and Considerations

While this study has several strengths, including methodological innovation and cultural contextualization, it is not without limitations. First, the analysis is based on self-reported data, which is subject to social desirability bias and may not reflect actual behaviors (Podsakoff et al., 2024). Second, the model only accounts for a subset of possible predictors of reading achievement. Cognitive abilities, instructional quality, peer influence, and school-level factors were not included but likely play significant roles (Hair et al., 2019). Additionally, the generalizability of the findings to other countries should be approached with caution. While some of the identified mechanisms may be universal (e.g., the benefits of parental communication), their expression may vary significantly across cultural contexts.

Another limitation of this study lies in the potential challenges of cultural transferability. While the findings are grounded in the specific cultural context of the United Arab Emirates (UAE), they may not uniformly apply to other Gulf Cooperation Council (GCC) countries such as Saudi Arabia, Qatar, Oman, Kuwait, and Bahrain. Although these countries share overarching cultural norms—such as collectivism, extended family structures, and Islamic values—differences in education policy, parental expectations, and family dynamics may influence the applicability of the results. For example, the UAE has made distinctive efforts to internationalize its education system and promote bilingualism, whereas countries like Saudi Arabia maintain more traditional curricula and stronger state oversight in educational and family matters (Al-Mahrooqi & Denman, 2018).

In some Gulf states, stricter gender norms or regional disparities in access to educational resources may limit parental involvement, particularly among mothers. Moreover, the extent to which behaviors, like shared family meals or academic encouragement, are institutionalized as daily practices may vary significantly across the region. Hence, interventions based on the UAE model must be culturally adapted before implementation elsewhere. Future cross-national studies are recommended to explore these contextual nuances and test the robustness of the observed associations in neighboring Gulf states. Such comparative analyses would enrich our understanding of family engagement in Arab education systems and support the development of culturally responsive educational interventions (Al Ghazali, 2020; Lansford et al., 2019).

6. Conclusions

This study offers important insights into how family engagement contributes to student reading achievement within the UAE’s unique socio-cultural context. Using machine learning techniques, particularly Gradient Boosting, alongside explainable AI methods such as SHAP and counterfactual analysis, the research identifies and quantifies key family behaviors that influence reading outcomes. The findings reveal that simple, everyday parental practices—like regularly talking with children about their day, encouraging academic effort, and sharing family meals—are among the most predictive variables of reading performance. These behaviors reflect culturally resonant forms of engagement in collectivist societies, underscoring the value of aligning educational strategies with local norms.

Although the model’s cross-validated R² value of approximately 0.19 might seem modest in isolation, it aligns with prior educational research on non-cognitive and behavioral predictors, which typically account for small to moderate variance in academic outcomes. Importantly, this level of predictive power is meaningful when paired with the interpretability and prescriptive potential enabled by SHAP and counterfactual methods. By simulating realistic behavior changes, the study estimates that increasing the frequency of parent–child conversations or family meals could yield reading score improvements from a baseline of 358.93 to as high as 448.68, marking an improvement of 90 points and a substantial uplift with direct policy implications.

Methodologically, the study illustrates the advantages of combining predictive and explanatory frameworks in educational data science. SHAP values provide transparency about feature importance, both globally and at the individual level, while counterfactual analysis introduces a practical dimension by illustrating how specific behavior changes might improve learning outcomes. This approach moves beyond prediction to support informed, actionable decisions for educators, parents, and policymakers.

Despite its valuable insights, the model’s limitations must be acknowledged. The explanatory power of the Gradient Boosting model, while statistically meaningful, indicates that a large proportion of variance in reading outcomes remains unexplained. Moreover, although SHAP and counterfactual methods enhance interpretability, they are not substitutes for causal inference. The model can suggest associations and simulate potential effects of behavior change but cannot establish direct causation. The assumption of holding all other variables constant in counterfactual analysis may oversimplify real-world interactions among predictors.

The study also reinforces the idea that family engagement can be leveraged as both a predictive lever and a prescriptive tool. In the UAE, where family values are deeply integrated into educational ambitions, this alignment is especially powerful. However, cultural transferability remains a consideration; future studies should explore whether similar engagement strategies hold predictive and prescriptive value in other Gulf or Arab contexts, where family structures and educational systems may vary. In sum, this research contributes to the growing body of work that applies interpretable machine learning to educational challenges, demonstrating that culturally situated family engagement is both measurable and modifiable, and thus a key target for improving literacy outcomes in the region and beyond.

Future studies should aim to address these limitations by incorporating multi-level data, including teacher, school, and community variables. Longitudinal studies could explore how family engagement practices evolve over time and interact with developmental trajectories in literacy. Moreover, qualitative studies could enrich the interpretation of findings by exploring how parents and students perceive and negotiate engagement in everyday contexts.

Additionally, future research should expand the use of explainable AI in education. For instance, integrating temporal models with SHAP and counterfactual simulations could offer insights into how changes in behavior over time affect learning outcomes. Experimental designs that test XAI-informed interventions in real-world settings would also provide valuable evidence of their practical efficacy and scalability.