Academic resilience in the Latin America region post COVID-19 pandemic -- an explainable machine learning analysis of its determinants and heterogeneity using alternative definitions

arXiv:2509.24830 · econ.GN, q-fin.EC · Submitted 2025-09-29 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Academic resilience in the Latin America region post COVID-19 pandemic -- an explainable machine learning analysis of its determinants and heterogeneity using alternative definitions".

Warren: Here's the gist of the paper. The paper identifies leading factors behind student academic resilience using explainable machine learning methods and alternative definitions to address educational inequality in Latin America post-COVID-19 Pruning redundant reasoning tokens The learning crisis in the Latin American region is worrying, particularly post-pandemic given the stronger role of inequality behind achievement The concept of student academic resilience (SAR), students who despite coming from disadvantaged backgrounds reach good performance levels, and an analysis of its determinants, are policy relevant Using as a benchmark the bottom two quintiles of the socio-economic in the distribution of the economic, social and cultural status (ESCS) index of households in the region to capture disadvantaged students and level 2 as the cutoff used for basic performance, this paper contributes to the body of literature on academic resilience by presenting new empirical evidence for the LAC region How the training signal is built The analysis relies on four alternative definitions of students academic resilience given the lack of agreement on how to measure it These definitions fall into a definition-driven approach, which contains two measurement sub-approaches: (i) a high-risk and high achieving group membership and (ii) resilience residuals Specifically, SAR1 is defined as students reaching level 2 or above (in the three subjects) among the LAC’s region bottom two SES quintiles SAR2 measures as academic resilient a student whose estimated probability falls into the top 40% of the estimated probability of the level 2 achievement distribution Where the gains come from The analysis utilizes Gradient Boosted Trees (GBT) as the chosen ML model for initial estimations and subsequent explainable results based on the Shapley Additive Explanations (SHAP) method The GBT model shows clear performance dominance across all four SAR indicators, with differential performance between 6% to 17% (AUROC score) and 25%-58% (AUPRC score) for SAR2 For SAR1, the leading predictive drivers are most from to the student’s group, such as the number of digital devices at home, gender, homework, satisfaction with life, primary and lower secondary repetition, and engagement with paid work Pruning redundant reasoning tokens The analysis also investigates COVID-19 background variables and personality traits through partial dependence plots For COVID-19 background variables, partial dependence plots show an uniformly decreasing tendency on the SAR’s likelihood for longer schools shutdowns, with an increment of 170 days leading to reductions on estimated odd-ratio (ORc) between 10% to 25% Personality traits like curiosity and perseverance are found to be protective factors for resilience against adversity, with students with positive values for curiosity having a 10% larger probability of being academic resilient compared to those with an index equal to zero Where the gains come from Sub-sample results reveal heterogeneity across education sub-systems; for instance, based on SAR1, student’s assertiveness and rate of certified teachers are more prominent for students attending public schools Furthermore, local estimations for students at the extremes of SHAP index distribution show that a student with the highest chance of being academic resilient attends a private school who has not repeated and is engaged with homework The paper concludes that knowing which are the pull up factors for students to be more likely to fall into the academic resilient group can be employed as a baseline to aid those poorer students at the bottom of the learning distribution Pruning redundant reasoning tokens The paper contributes to the literature by relying on four different versions of the student academic resilient (SAR) indicator and introducing two novel SAR indicators It is also noted that this is the first study investigating the determinants of students academic resilience for the LAC region The empirical analysis focuses on drivers by education sub-systems and relies on novel explainable artificial intelligence (XAI) methods, specifically Shapley values The findings suggest how the pandemic powerfully shaped SAR probabilities in the LAC region through factors like school closure length and remote learning barriers The paper seeks to assess equality of opportunity in education and its determinants for the LAC region as a policy tool to boost upward social mobility for the most deprived students Where the gains come from The analysis provides granular local explanations by focusing on students at extremes of SHAP index distribution, showing specific profiles for high and low chance students For example, a student with the highest probability of being academic resilient attends a private school with high rates of certified teachers and teachers attending PD equals to 70% This detailed analysis helps in identifying key features behind SAR that can be used to lift disadvantaged students from lower achievement groups towards being academic resilient The results offer insights for policy designs where key features behind SAR can be used to lift disadvantaged students from lower achievement groups towards being academic resilient Pruning redundant reasoning tokens The analysis of the paper has some limitations, as estimates denote associations but not causal claims because of the crosssection nature of the PISA data employed A first future line of investigation could be a more thorough analysis of the role of intersectionality on the likelihood of a student being academic resilient Another venue for future research is to look how different conceptualisations of academic resilience and their determinants are aligned with specific educational countries policies targeting the most disadvantaged students The paper's findings add to the scare regional literature contributing to future policy designs where key features behind SAR can be used to lift disadvantaged students from lower achievement groups towards being academic resilient The paper's findings adds to the scare regional literature contributing to future policy designs where key features behind SAR can be used to lift disadvantaged students from lower achievement groups towards being academic resilient Where the gains come from

Paper summary: Ruth: So we're looking at this paper titled Academic resilience in the Latin America region post COVID-nineteen pandemic -- an explainable machine learning analysis of its determinants and heterogeneity using alternative definitions <ref:2509.24830#pg1,Academic resilience in the Latin America region post COVID-19 pandemic -- an>. Basically, it’s addressing the learning crisis in Latin America, which is getting worse after the pandemic because inequality plays a bigger role than before >

Warren: Exactly. They introduce student academic resilience, or SAR, which is defined as students who come from disadvantaged backgrounds but still achieve good performance levels >

Ruth: And they tackle this by using different ways to measure resilience. They use four alternative definitions for SAR, which gives them a really robust look at the problem >

Iris: That’s interesting because having multiple definitions lets them see if the results are consistent across different ways you can frame what resilience actually is >

Warren: Right. So, they use explainable machine learning methods and specific data from PISA two thousand twenty-two to find the leading factors behind these four SAR indicators > <ref:2509.24830#pg1>

Ruth: They claim that household inputs like books and digital devices, gender, homework load, repetition in school or work intensity are the main drivers for one of these indicators >

Iris: And they look at how COVID-nineteen background variables and personality traits might affect this resilience using partial dependence plots > <ref:2509.24830#pg1>

Warren: The results show that longer school shutdowns actually reduce the likelihood of academic resilience, with a one hundred seventy day increase in shutdown leading to a reduction in the estimated odd ratio between ten percent and twenty-five percent >

Ruth: And personality traits like curiosity and perseverance are found to be protective factors, meaning students who score positively on those traits have a higher chance of being academic resilient >

Conclusion: Iris: Looking at the conclusion of this paper, they really emphasize that they’ve contributed by using four different versions of the SAR indicator and introducing two new ones >

Ruth: And they specifically highlight that this is the first study looking at the drivers of student academic resilience for the LAC region >

Warren: The implication here is that knowing what pulls students up into this resilient group can serve as a baseline to help those poorer students at the bottom of the learning distribution >

Iris: They also point out how their analysis focuses on drivers by education sub-systems and uses explainable artificial intelligence methods, specifically Shapley values >

Ruth: So, they’re trying to assess equality of opportunity in education and what determines it for Latin America as a policy tool to boost upward social mobility for the most deprived students >

Warren: It’s interesting how they tie the findings about school closure length and remote learning barriers back to how much the pandemic reshaped SAR probabilities in this region >

Department of Education, University of Bath · Instituto de Investigaciones Educativas, Universidad Nacional de Chilecito

econ.GN, q-fin.EC

Submitted: 2025-09-29

Updated: 2026-03-05

Comments: 33 pages, 11 figures

Journal ref: Cogent Education, 2026

DOI: 10.1080/2331186X.2026.2758566

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: The paper identifies leading factors behind student academic resilience using explainable machine learning methods and alternative definitions to address educational inequality in Latin America…

Key concepts

Student Academic Resilience (SAR)
This concept refers to students from disadvantaged backgrounds who still achieve good academic performance despite facing challenges. The study explored four different ways to measure this resilience, including group membership and probability distributions, to capture its complexity.
Explainable Machine Learning (XAI) / SHAP
The researchers used Gradient Boosted Trees combined with Shapley Additive Explanations (SHAP) to understand which factors most strongly predict academic resilience. This technique helps move beyond simple prediction by showing exactly how each variable contributes to the final outcome.
Alternative Definitions of SAR
Since there was no single agreed-upon way to measure academic resilience, the paper used four distinct definitions (SAR1 and SAR2). These definitions allowed researchers to test different lenses—such as high-risk groups versus probability distributions—to see which definition yielded the most meaningful results.
Heterogeneity Across Sub-systems
The analysis showed that the factors predicting resilience are not uniform across all educational settings. For example, the drivers for SAR1 differed significantly between students in public schools and those in private schools, highlighting that context matters for intervention.

Terminology

Summary

The paper identifies leading factors behind student academic resilience using explainable machine learning methods and alternative definitions to address educational inequality in Latin America post-COVID-19

Pruning redundant reasoning tokens

The learning crisis in the Latin American region is worrying, particularly post-pandemic given the stronger role of inequality behind achievement The concept of student academic resilience (SAR), students who despite coming from disadvantaged backgrounds reach good performance levels, and an analysis of its determinants, are policy relevant Using as a benchmark the bottom two quintiles of the socio-economic in the distribution of the economic, social and cultural status (ESCS) index of households in the region to capture disadvantaged students and level 2 as the cutoff used for basic performance, this paper contributes to the body of literature on academic resilience by presenting new empirical evidence for the LAC region

How the training signal is built

The analysis relies on four alternative definitions of students academic resilience given the lack of agreement on how to measure it These definitions fall into a definition-driven approach, which contains two measurement sub-approaches: (i) a high-risk and high achieving group membership and (ii) resilience residuals Specifically, SAR1 is defined as students reaching level 2 or above (in the three subjects) among the LAC’s region bottom two SES quintiles SAR2 measures as academic resilient a student whose estimated probability falls into the top 40% of the estimated probability of the level 2 achievement distribution

Where the gains come from

The analysis utilizes Gradient Boosted Trees (GBT) as the chosen ML model for initial estimations and subsequent explainable results based on the Shapley Additive Explanations (SHAP) method The GBT model shows clear performance dominance across all four SAR indicators, with differential performance between 6% to 17% (AUROC score) and 25%-58% (AUPRC score) for SAR2 For SAR1, the leading predictive drivers are most from to the student’s group, such as the number of digital devices at home, gender, homework, satisfaction with life, primary and lower secondary repetition, and engagement with paid work

The analysis also investigates COVID-19 background variables and personality traits through partial dependence plots For COVID-19 background variables, partial dependence plots show an uniformly decreasing tendency on the SAR’s likelihood for longer schools shutdowns, with an increment of 170 days leading to reductions on estimated odd-ratio (ORc) between 10% to 25% Personality traits like curiosity and perseverance are found to be protective factors for resilience against adversity, with students with positive values for curiosity having a 10% larger probability of being academic resilient compared to those with an index equal to zero

Sub-sample results reveal heterogeneity across education sub-systems; for instance, based on SAR1, student’s assertiveness and rate of certified teachers are more prominent for students attending public schools Furthermore, local estimations for students at the extremes of SHAP index distribution show that a student with the highest chance of being academic resilient attends a private school who has not repeated and is engaged with homework The paper concludes that knowing which are the pull up factors for students to be more likely to fall into the academic resilient group can be employed as a baseline to aid those poorer students at the bottom of the learning distribution

The paper contributes to the literature by relying on four different versions of the student academic resilient (SAR) indicator and introducing two novel SAR indicators It is also noted that this is the first study investigating the determinants of students academic resilience for the LAC region The empirical analysis focuses on drivers by education sub-systems and relies on novel explainable artificial intelligence (XAI) methods, specifically Shapley values The findings suggest how the pandemic powerfully shaped SAR probabilities in the LAC region through factors like school closure length and remote learning barriers The paper seeks to assess equality of opportunity in education and its determinants for the LAC region as a policy tool to boost upward social mobility for the most deprived students

The analysis provides granular local explanations by focusing on students at extremes of SHAP index distribution, showing specific profiles for high and low chance students For example, a student with the highest probability of being academic resilient attends a private school with high rates of certified teachers and teachers attending PD equals to 70% This detailed analysis helps in identifying key features behind SAR that can be used to lift disadvantaged students from lower achievement groups towards being academic resilient The results offer insights for policy designs where key features behind SAR can be used to lift disadvantaged students from lower achievement groups towards being academic resilient

The analysis of the paper has some limitations, as estimates denote associations but not causal claims because of the crosssection nature of the PISA data employed A first future line of investigation could be a more thorough analysis of the role of intersectionality on the likelihood of a student being academic resilient Another venue for future research is to look how different conceptualisations of academic resilience and their determinants are aligned with specific educational countries policies targeting the most disadvantaged students The paper's findings add to the scare regional literature contributing to future policy designs where key features behind SAR can be used to lift disadvantaged students from lower achievement groups towards being academic resilient The paper's findings adds to the scare regional literature contributing to future policy designs where key features behind SAR can be used to lift disadvantaged students from lower achievement groups towards being academic resilient

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Adaptive SAR modeling for policy design can be achieved by using key features behind SAR can be used to lift disadvantaged students from lower achievement groups towards being academic resilient as a baseline, allowing AI systems to identify specific predictors for targeted interventions.

  2. Explainable machine learning models like SHAP values allow AI systems to provide local explanations of predictions, enabling researchers and policymakers to understand the degree to which each feature (or variable) contributes to a given prediction value when conditioning on that feature.

  3. The system can perform comparative analysis across different definitions of academic resilience, specifically by utilizing the four alternative definitions of students academic resilience given the lack of agreement on how to measure it as inputs for determining leading factors.

  4. AI systems can predict the likelihood of a student being academic resilient using Gradient Boosted Trees (GBT) models, which show a clear performance dominance of GBT over logistic regression and neural networks in capturing non-linear and interaction effects.

  5. The system can generate partial dependence plots to show the marginal effect of a given covariate on the prediction of the ML model by calculating the relative probability (RP) of SAR outcomes taking the value 1, providing actionable insights into how specific factors influence resilience chances.

  6. The system can identify heterogeneity in determinants by comparing results across education sub-systems, such as finding that students attending either private or urban schools are more likely to be academic resilient based on the SAR formulation used.

Related papers