Patterns of Attributes In Experienced Unfair Treatment In Healthcare Settings: A Machine Learning Approach

Background

Unfair treatment, encompassing perceived unjust or inequitable actions, such as discrimination and microaggressions, has been widely documented as detrimental to well-being. In healthcare settings, where individuals are particularly vulnerable, such experiences are especially consequential. Prior research has shown that perceived unfair treatment undermines trust between patients and healthcare providers and contributes to delayed or avoided care. However, these experiences are inherently complex and shaped by multiple, intersecting social and health-related factors. Traditional regression-based approaches may be limited in capturing such multidimensional interactions. Grounded in the intersectionality framework, this study aims to examine patterns of perceived unfair treatment in healthcare settings using a data-driven approach.

Methods

Using the 2017 California Health Interview Survey, we examined lifetime unfair treatment in healthcare (binary) (N = 21,150 after exclusions). Predictors include: age (18-54/55-64/65 or over years), sex (male/female), race/ethnicity (White/African American/Hispanic/Asian/Others), poverty status (<300% vs. > 300% FPL), education (< college or >college), region (rural/urban), self-rated physical health (good or above vs. fair or poor), mental health (measured by the Kessler Psychological Distress Scale), insurance (yes/no), citizenship (yes/no), English proficiency (high/low), height categories, and obesity status (underweight/normal vs. overweight). Data were split into training and holdout test sets using an 80/20 stratified design. To address class imbalance, the minority class was oversampled within sex strata in the training set. Three complementary approaches were employed: (1) survey-weighted logistic regression, (2) average marginal effects (AMEs) for probability interpretation, and (3) machine learning using a tuned HistGradientBoostingClassifier with 5-fold cross-validation (Optuna, log-loss optimization). Logistic regression and random forest models were used as benchmarks. Model interpretability was assessed using Shapley Additive Explanations (SHAP). Analyses were conducted in Python 3.11 using Google Colab.

Results

Overall, (27.2%) of respondents reported experiencing unfair treatment in healthcare. Survey-weighted logistic regression identified significant intersectional predictors, including African American race (OR = 2.24, p < .001), female sex (OR = 1.67, p < .001), poor self-rated health (OR = 1.36, p < .001), and poverty (OR = 1.29, p < .001). AMEs showed the strongest marginal effects for race (+0.16), sex (+0.09), and mental health (−0.11). Machine learning models demonstrated modest predictive performance (accuracy = 0.633; ROC-AUC = 0.579; F1 = 0.379), with substantial false negatives (n = 677; 16.0%). SHAP analyses identified age, race/ethnicity, English proficiency, poverty, physical health, sex, and insurance as key contributors to model predictions. Across analytic approaches, several core predictors were consistent, although their relative importance differed by method, suggesting that perceived unfair treatment is shaped by multiple intersecting social and health-related vulnerabilities.

Discussion

Consistent with intersectionality theory, findings suggest that perceived unfair treatment in healthcare reflects the combined influence of multiple social and health-related factors rather than a single dominant determinant. At the same time, the modest predictive performance of machine learning models highlights the limits of relying solely on observable attributes to capture complex, context-dependent experiences. For social work practice, these findings suggest incorporating assessment of unfair treatment into routine psychosocial evaluations and adopting broad-based screening approaches rather than relying solely on demographic risk indicators.