📊 statistics

Machine Learning Analysis of Socioeconomic Stratification and Health Vulnerability in the United States

Using a supervised machine learning pipeline on a synthetic dataset calibrated to US national surveys, this study demonstrates that structural socioeconomic factors, particularly education and occupational class, are the dominant predictors of health vulnerability—explaining 82% of feature importance and supporting fundamental cause theory over behavioral explanations.

Original authors: Tamim Anowar

Published 2026-07-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Tamim Anowar

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the United States as a massive, high-stakes video game where every player is trying to stay healthy. For years, the "experts" (and many of us) have argued about what makes the biggest difference in the game: Is it the character's gear and stats (like how much money they have, what school they went to, and what job they hold), or is it the player's choices (like whether they eat their veggies, don't smoke, and exercise)?

A researcher named Tamim Anowar decided to settle this debate using a super-smart computer brain called Machine Learning. But here's the twist: instead of using real people's private data (which is locked up tight), the researcher built a perfectly realistic simulation of 1,200 American players. This "fake" group was programmed to look exactly like real national surveys, acting as a test lab to see which factors actually control the game's outcome: a score called the Health Vulnerability Index (HVI). Think of the HVI as a "Danger Meter" from 0 to 100, where 100 means you are in deep trouble and 0 means you're safe.

The Big Reveal: It's the Gear, Not the Player

When the computer analyzed the data, it found a massive imbalance. The "structural" factors—your education, your job class, your income, and whether you have insurance—explained about 82% of why people had high danger scores. In contrast, all the "behavioral" choices combined (diet, smoking, exercise, sleep) only explained about 8%.

To put it in game terms: If you are playing a character with terrible stats (low education, low income, no insurance), it doesn't matter if you press the "eat healthy" button perfectly. The game is rigged against you. The computer found that structural factors are 10 times more important than individual choices in predicting who gets sick.

However, there is a crucial catch to this number. Because the "Danger Meter" (HVI) itself is built by adding up scores for low income, low education, and lack of insurance, it is almost guaranteed that the computer will say those factors are the most important. The authors are very clear: this result is a confirmation that their computer pipeline works correctly, not a standalone proof that lifestyle choices don't matter in the real world. It proves the method is sound, but the specific 82% vs. 8% split is a feature of the simulation's design, not necessarily a final fact about real human biology.

The Education and Job "Power-Ups"

The study looked closely at two specific stats: Education and Job Class.

The Education Gap: The difference in danger between someone with less than a high school diploma and someone with a professional or doctoral degree was huge. The low-education group had an average danger score of 68.96, while the high-education group was at 25.75. That's a gap of 43.2 points. In the world of statistics, this is a massive jump (a "Cohen's d" of 3.63), meaning education is a super-powerful shield.
The Job Gap: Similarly, moving from an unskilled manual job to an executive or elite job lowered the danger score by 34.5 points.

Even when the computer "turned off" the income variable to see if education and jobs mattered on their own, they still had a strong, independent effect. This supports an old idea called "Fundamental Cause Theory," which suggests that having resources (like a degree) lets you dodge any health risk that comes your way, whether it's a virus or a bad diet.

The "Intersectional" Twist: Race Changes the Rules

The study also checked if these rules applied equally to everyone. It found that while the average danger score was similar across different racial groups, the gap between rich and poor was different.

For Hispanic/Latino players, the gap between low-class and high-class players was the widest: 20.9 points.
For Black Non-Hispanic players, it was 18.7 points.
For White Non-Hispanic players, it was 18.5 points.
For Asian/Pacific Islander players, it was the narrowest at 16.2 points.

This suggests that being a racial minority makes the "health shield" of a high-class job slightly less effective. It's like having a powerful weapon that gets jammed more often if you belong to a certain group. This aligns with "intersectional theory," which says that race and class mix together to create unique, often harsher, challenges.

Can the Computer Predict a Crash?

The researchers also asked: "Can we spot someone about to have a rapid health crash?"
The answer was a cautious "yes, but." The computer models (specifically Gradient Boosting and Random Forest) were incredibly good at reading the map, with a success rate (AUC-ROC) of 0.969 for binary yes/no questions and 0.893 for predicting exact scores.

But why were they so good? Because the "Danger Meter" (HVI) was literally built using the same ingredients (income, education, insurance) that the computer was looking at. The high score doesn't mean the computer is a psychic; it means the computer is very good at math when the answer is already written in the question. The study explicitly warns that this high accuracy is structurally expected because of how the data was defined, not because the models discovered some hidden magic.

However, the models did find a stark reality: the risk of a rapid health decline was about 24 times higher for those with the lowest education (35.9%) compared to the highest (1.5%).

What the Paper Rules Out (and What It Doesn't)

It is crucial to understand what this study does not say.

It does NOT say that lifestyle choices don't matter at all. It just says that in this specific simulation of 1,200 people, they were far less important than the structural factors.
It does NOT say that these results are a final, proven fact about the real world. The authors are very clear: this was a simulation using synthetic data. They built this "fake" dataset to test their computer pipeline before they try it on real, restricted government data.
It does NOT say that the results are "solved." The authors explicitly state that the findings are a "proof-of-concept." They are confident in the method (the way they used the computer), but they warn that the specific numbers (like the 82% vs 8% split) might shift when they run the same test on real human data.

The Bottom Line

In this simulated game, the "Health Vulnerability Index" showed that the playing field is tilted heavily by your starting stats (education, job, money). While the computer models were excellent at reading the map, the study serves as a warning: fixing health problems by just telling people to "eat better" might not work if the game itself is rigged against them.

The authors suggest that to really win the game, we need to fix the "gear" (upstream structural issues like insurance and education) rather than just coaching the players on their moves. But remember, this is a simulation designed to prove the computer works. The real test is coming soon, when they apply this same logic to real, restricted national data. Until then, we know the method is solid, but the exact numbers for the real world are still waiting to be discovered.

Technical Summary: Machine Learning Analysis of Socioeconomic Stratification and Health Vulnerability in the United States

Problem Statement
The paper addresses the persistent and significant health disparities in the United States, where morbidity, mortality, and healthcare access remain tightly correlated with social class and education levels. A central theoretical tension exists between two explanatory frameworks: the structural-materialist view (Fundamental Cause Theory), which posits that socioeconomic status (SES) acts as a flexible resource determining exposure to health risks, and the individualist-behavioural view, which attributes inequalities primarily to lifestyle choices. Existing quantitative research often relies on linear regression models that fail to capture high-dimensional interactions and non-linear relationships hypothesized by intersectional sociology. Furthermore, there is a lack of empirical studies that directly compare the predictive power of structural socioeconomic predictors against individual behavioural factors within a single, unified model.

Methodology
The study employs a supervised machine learning pipeline on a synthetic analytic dataset ( $N=1,200$ ) calibrated to major US national surveys (NHANES, BRFSS, MEPS, ACS, CPS, PSID/SCF, EPA EJScreen, USDA) covering the period 2019–2023. The synthetic nature of the data allows for pipeline validation and pre-registration without disclosure risks, serving as a proof-of-concept before application to restricted-use microdata.

Outcome Variable: The primary outcome is the Health Vulnerability Index (HVI), a validated composite score (0–100) weighted by educational disadvantage, occupational class, income deprivation, poverty status, insurance gaps, and age-related risk.
Predictors: The model incorporates structural variables (education, occupational class via EGP schema, income, poverty, insurance), social determinants (food/housing insecurity, transportation barriers, environmental hazards), and individual behavioural/biomedical factors (BMI, smoking, diet, physical activity).
Models: The study compares Logistic Regression and LASSO (as interpretable parametric benchmarks) against ensemble methods: Random Forest (RF) and Gradient Boosting (GB).
Evaluation Strategy: Data is split into training (70%), validation (15%), and held-out test (15%) sets. Models are tuned using Bayesian optimization (Optuna) with 5-fold stratified cross-validation. Performance is measured via AUC-ROC, RMSE, $R^2$ , and F1 scores.
Interpretability: To address the "black box" nature of ensemble methods, the study utilizes SHAP (Shapley Additive exPlanations) values for global feature importance, dependence plots, and interaction analysis. Predictor importance is triangulated across RF permutation importance, LASSO coefficients, and OLS standardized betas.

Key Contributions

Empirical Triangulation: The study directly quantifies the relative significance of structural versus behavioural predictors in a single model, moving beyond theoretical debate to empirical estimation.
Methodological Innovation: It applies ensemble ML techniques to sociological health research, utilizing SHAP values to capture non-linear relationships and complex interactions (e.g., class-race dynamics) that standard additive regression models miss.
Intersectional Analysis: It operationalizes race/ethnicity as a measure of structural exposure rather than a biological factor, examining whether the protective returns of higher class positions vary across racial groups (intersectional amplification).
Pipeline Validation: It establishes a reproducible analytical pipeline for predicting sociological health vulnerability, designed for future deployment on restricted-use national survey microdata.

Results

Predictor Importance: Structural socioeconomic predictors explain approximately 82% of the total feature importance in Random Forest models, whereas all combined behavioural and biomedical variables account for only 8%. This 10:1 ratio is consistent across RF, LASSO, and OLS models.
Education and Class Gradients:
- The education gradient spans 43.2 HVI points (from 68.96 for <High School to 25.75 for Professional/Doctoral), with a large effect size ( $d=3.63$ ).
- The occupational class gradient spans 34.5 HVI points (from 66.22 for Unskilled/Manual to 31.74 for Executive/Elite), with $d=2.54$ .
- Both education and class remain significant independent predictors even after controlling for income (partial $r = -0.383$ and $-0.137$, respectively).
Intersectional Amplification: While marginal mean HVI scores were similar across racial groups, the within-group class-health gaps varied significantly. Hispanic/Latino respondents exhibited the largest gap (20.9 points), followed by Black Non-Hispanic (18.7) and White Non-Hispanic (18.5), supporting the theory that racial minority status attenuates the health returns of higher class positions.
Rapid Deterioration: The risk of rapid health deterioration is approximately 24 times higher for those with less than a high school education (35.9%) compared to those with professional/doctoral degrees (1.5%).
Model Performance:
- Binary Classification (High Vulnerability): Gradient Boosting achieved an AUC-ROC of 0.969, and Logistic Regression achieved 0.974.
- Continuous Regression (HVI): Gradient Boosting achieved an $R^2$ of 0.893, outperforming LASSO ( $R^2=0.885$ ) and capturing non-linear interactions.
- The near-equivalence in performance between parametric and non-parametric models suggests the prediction surface is predominantly linear, validating the interpretability of regression coefficients.

Significance and Claims
The paper claims that its findings provide quantitative support for Fundamental Cause Theory, demonstrating that structural socioeconomic position is the dominant driver of population-level health vulnerability, outweighing individual behavioural factors by a factor of ten. The study disputes purely behavioural explanations of health inequality.

The authors explicitly frame this work as a proof-of-concept. They caution that the results are derived from a synthetic dataset calibrated to national distributions and should not be interpreted as novel empirical claims about actual US population health without validation on restricted-use microdata. The primary contribution is the validation of the analytical pipeline and the convergence of results across different modeling frameworks (parametric and non-parametric). The paper concludes that policy strategies focusing solely on individual behavioral change are unlikely to significantly reduce health inequalities without addressing the upstream structural circumstances (education, occupational safety, universal coverage) that generate vulnerability. The authors intend to deploy this pipeline on real, restricted-use national survey microdata to produce potentially stronger effect sizes.