Haves and have-nots: Socioeconomic position improves accuracy of machine learning algorithms for predicting high-impact chronic pain

Introduction

Disparities in chronic pain are driven, in part, by systems of social stratification that distribute resources, prestige, as well as adversity unequally across groups[19]. Socioeconomic position (SEP) is a multifaceted construct that captures an individual’s material and social resources (e.g., income, wealth, educational attainment) and social rank[28; 29]. Lower SEP is associated with increased risk for developing chronic pain conditions[27; 55], reporting elevated pain severity[38], and experiencing pain-related disability[14]. Despite strong theoretical grounding [19; 28] and broad empirical support for impacts of socioeconomic gradients on pain[25; 42], there remains little consensus on how best to operationalize SEP for purposes of describing chronic pain inequalities[19].

SEP indicators may be subjective or objective[16], describe individuals or neighborhoods[39], and impact chronic pain risk differentially across the life course[29]. Individual SEP indicators are not interchangeable: relations between SEP and pain are stronger for some SEP indicators and pain outcomes than others [18]. Relations between SEP and pain may also differ in direction across sociodemographic groups[23]. For example, among individuals with chronic low back pain, higher perceived social status was associated with lower pain interference among White individuals but was not associated with pain interference among Black individuals[3]. Despite facing multilevel social barriers, persons from marginalized groups who climb the social ladder may experience differences in how they are treated and rewarded relative to persons from dominant groups at the same level, thereby increasing risk for inequitable health outcomes[23]. Improved understanding of which SEP indicators are most strongly associated with chronic pain outcomes - and for whom - is critical for monitoring chronic pain inequalities and understanding how best to intervene, potentially through multiple contributing pathways.

The primary aim of this cross-sectional study was to determine aspects of SEP that help to identify individuals currently suffering from high-impact chronic pain in the United States. Traditional regression-based approaches hinder meaningful progress toward accurate pain prediction because they are limited by (1) the number of variables that can be compared within models and (2) a priori identification of variables by clinicians and/or theoretical models, thereby limiting discovery of new predictors. Machine learning approaches, in contrast, are data-driven, well-suited to testing biopsychosocial models of chronic pain, and have improved prediction of other complex phenomena such as suicidal thoughts/behaviors [6] and incident diabetes [8]. In addition, the relative importance of SEP indicators for identifying high-impact chronic pain was examined in models that included demographic, physical health, mental health, and geographic factors implicated in chronic pain risk[17]. The secondary aim of this study was to evaluate potential racial and sex differences in the importance of SEP indicators of high-impact chronic pain (i.e., pain limiting life and work activities). Together, these aims sought to further understanding of how SEP domains are associated with high-impact chronic pain, and whether their influence differs across demographic groups with well-documented disparities in pain severity and prevalence (i.e., Non-Hispanic Black and White adults; males and females) [11; 35; 36; 56].

Methods

Data.

The National Health Interview Survey (NHIS) is a cross-sectional survey of US civilian households that is conducted annually by the National Center for Health Statistics; NHIS is considered to be the best available national pain surveillance system in the United States[15]. The present study focused on the 2019 adult interviews (n=31,997; response rate = 61.1%).¹ The 2019 NHIS data release was selected because it is the most recent year including both the chronic pain and mental health assessments. After data cleaning, 716 participants were removed due to missing values for high-impact chronic pain (see variable definition below) and another 3 participants with missing sex, leaving a total of 31,278 participants included in analyses.

Participants included 14,422 males (46%) and 16,856 females (54%). The mean age was 52.0 (range: 18–85; note: NHIS top-coded age at 85+). There were 23,783 White only (76%), 3,466 Black or African-American only (11%), 4,035 Hispanic (13%), 1,617 Asian only (5%), 264 American Indian and Alaska Native (AIAN) only (1%), 284 AIAN and any other group (1%), and 362 other single and multiple races (1%). These proportions were unweighted and should not be considered representative of the United States population. This sample included 2,666 (9% of the total sample) reported ongoing high-impact chronic pain.

Variables.

Summary statistics and detailed descriptions of study variables - including coding and reference groups for categorical variables - are provided in Table 1.

Table 1.

Study variable descriptions, coding, and summary statistics.

	Description	Coding	NHIS variable	N (%)	Mean (SD)
High-impact chronic pain	Frequency of pain in past 3 months	1=never;2=some days; 3=most days; 4=every day	paifrq3m_a
		Chronic pain (3 or 4)		7,174 (23%)
	How often pain limits life/work in past 3 months	1=never;2=some days; 3=most days; 4=every day	paiwklm3m_a
		High-impact chronic pain (3 or 4)		2,666 (9%)
Demographic
Age	Age	Dichotomous variables (0=no; 1=yes) for each group	agep_a
		Less than 25		2,116 (7%)
		25–34		4,785 (15%)
		35–44		4,853 (16%)
		45–54		4,746 (15%)
		55–64		5,692 (18%)
		65–74		5,214 (17%)
		75 and over		3,872 (12%)
		(note: NHIS top-coded age at 85+ years)
Hispanic	Hispanic ethnicity		hisp_a
		Yes		4,035 (13%)
Orientation	Sexual orientation	Dichotomous variables (0=no; 1=yes) for each orientation:	orient_a
		Gay or Lesbian		452 (1%)
		Straight		29,771 (95%)
		Bisexual		412 (1%)
		Unknown		643 (2%)
Marital status	Marital status	Dichotomous variables (0=no; 1=yes) for each category:	marital_a
		Married		14,714 (47%)
		Living with a partner		2,130 (7%)
		Neither married nor living with a partner		14,287 (46%)
		Unknown marital status		146 (0.5%)
Race	Racial groups	Dichotomous variables (0=no; 1=yes) for each group:	raceallp_a
		White only		23,783 (76%)
		Black/African American only		3,466 (11%)
		Asian only		1,617 (5%)
		AIAN only		264 (1%)
		AIAN and any other group		284 (1%)
		Other single and multiple races		362 (1%)
		Unknown		1,502 (5%)
Sex	Sex of adult	Female (0)	sex_a	16,856 (54%)
		Male (1)		14,422 (46%)
Geographic
Rurality	Urban-Rural Classification Code (2013)	Nonmetropolitan (1)	urbrrl	4,997 (16%)
		Medium and small metro (2)		9,981 (32%)
		Large fringe metro (3)		7,207 (23%)
		Large central metro (4)		9,093 (29%)
U.S. region	Region of United States	Dichotomous variables (0=no; 1=yes) for each region:	region
		Northeast		5,278 (17%)
		Midwest		6,987 (22%)
		Southern		11,414 (36%)
		West		7,599 (24%)
Physical Health
Self-rated health status	Self-reported general health status (SRH)	1=poor; 2=good; 3=very good; 4=excellent	phstat_a		2.6 (1.1)
BMI	Body Mass Index (BMI) category	Underweight	bmicat a	498 (2%)
		Healthy weight		9,745 (31%)
		Overweight		11,219 (36%)
		Obese		9,816 (31%)
Smoking	Cigarette Smoking Status	Never smoker (0)	smkcigst_a	19,023 (61%)
		Former smoker (1)		7,964 (25%)
		Current some day smoker (2)		926 (3%)
		Current every day smoker (3)		3,365 (11%)
Opioid use	Prescription opioid use in past 12 months	Yes	opd12m_a	4,210 (13%)
Arthritis	Ever Told By Doctor Had Arthritis	Yes	arthev_a	8,050 (26%)
Cancer	Ever Told By Doctor Had Cancer	Yes	canev_a	3,776 (12%)
Diabetes	Ever Told By Doctor Had Diabetes	Yes	dibev_a	3,281 (10%)
Hypertension	Ever Told By Doctor Had Hypertension	Yes	hypev_a	11,229 (36%)
Stroke	Ever Told By Doctor Had a Stroke	Yes	strev_a	1,169 (4%)
Respiratory disease	Ever Told By Doctor Had Respiratory Disease	Yes	copdev_a	1,745 (6%)
Heart disease	Ever Told By Doctor Had Coronary Heart Disease	Yes	chdev_a	1,814 (6%)
High cholesterol	Ever Told By Doctor Had High Cholesterol	Yes	chlev_a	9,026 (29%)
Mental Health
Anxiety disorder	History of anxiety disorder	Yes	anxev_a	4,653 (15%)
Anxiety frequency	Frequency of anxiety	0=never; 1=a few times as year; 2=monthly; 3=weekly; 4=daily	anxfreq_a		1.4 (1.4)
Anxiety medication	Taking medication for anxiety	Yes	anxmed_a	3,985 (13%)
Anxiety severity	Anxiety severity over past 2 weeks (Generalized Anxiety Disorder 7-item)	None/minimal (1)	gadcat_a	26,477 (85%)
		Mild (2)		2,933 (9%)
		Moderate (3)		1,046 (3%)
		Severe (4)		822 (3%)
Depression	History of depression	Yes	depev_a	5,347 (17%)
Depression frequency	Frequency of depression	0=never; 1=a few times as year; 2=monthly; 3=weekly; 4=daily	depfreq_a		0.7 (1.1)
Depression medication	Taking medication for depression	Yes	depmed_a	3,427 (11%)
Depression severity	Depression severity over past 2 weeks (Patient Health Questionnaire)	None/minimal (1)	phqcat_a	25,397 (81%)
		Mild (2)		3,693 (12%)
		Moderate (3)		1,316 (4%)
		Severe (4)		872 (3%)
Socioeconomic position
Family income (continuous)	Family income (continuous)	Continuous up to $219,999 (NHIS top-coded at $220,000+)	faminctc_a		73,404 (56,707)
Family income (categorical)	Family income (categorical)	$0–34,999 (1)	incgrp_a	9,149 (29%)
		$35,000–49,999 (2)		4,082 (13%)
		$50,000–74,999 (3)		5,656 (18%)
		$75,000–99,999 (4)		3,908 (12%)
		$100,000+ (5)		8,483 (27%)
Income to poverty (continuous)	Ratio of family income to poverty threshold (continuous)	Continuous up to 9.99 (NHIS top-coded at 10.00)	povrattc_a		4.0 (2.8)
Income to poverty (categorical)	Ratio of family income to poverty threshold (categorical)	0.00–0.49 (1)	ratcat_a	913 (3%)
		0.50–0.74 (2)		929 (3%)
		0.75–0.99 (3)		1,596 (5%)
		1.00–1.24 (4)		1,253 (4%)
		1.25–1.49 (5)		1,489 (5%)
		1.50–1.74 (6)		1,298 (4%)
		1.75–1.99 (7)		1,597 (5%)
		2.00–2.49 (8)		2,693 (9%)
		2.50–2.99 (9)		2,429 (8%)
		3.00–3.49 (10)		2,068 (7%)
		3.50–3.99 (11)		2,214 (7%)
		4.00–4.49 (12)		1,598 (5%)
		4.50–4.99 (13)		1,747 (6%)
		5.00+ (14)		9,454 (30%)
Educational attainment	Highest level of education completed	Dichotomous variables (0=no; 1=yes) for each category:	educ_a
		Less than high school completion		2,865 (9%)
		GED (general educational development) or equivalent		756 (2%)
		High school diploma		7,227 (23%)
		Some college but no postsecondary degree		5,271 (17%)
		Associate degree (technical/vocational)		1,239 (4%)
		Associate degree (academic)		2,847 (9%)
		Bachelor’s degree		6,813 (22%)
		Master’s or higher degree		4,260 (14%)
Family educational attainment		Dichotomous variables (0=no; 1=yes) for each category:	maxeduc_a
	Highest level of education completed by any adult in family	Less than high school completion		1,845 (6%)
		GED (general educational development) or equivalent		544 (2%)
		High school diploma		5,669 (18%)
		Some college but no postsecondary degree		4,795 (15%)
		Associate degree (technical/vocational)		1,398 (5%)
		Associate degree (academic)		3,187 (10%)
		Bachelor’s degree		7,839 (25%)
		Master’s or higher degree		6,001 (19%)
Health insurance	Health insurance coverage	Not covered (1)	notcov_a	2,796 (9%)
Problem paying medical bills	Problems paying medical bills in the past 12 months	Yes	paybll12m_a	3,850 (12%)
Worry about medical bills	Worried about paying medical bills if get sick or have accident	0=not at all worried; 1=somewhat worried; 2=very worried	payworry_a		0.6 (0.7)
Delayed medical care	Delayed medical care due to cost in past 12 months	Yes	meddl12m_a	2,586 (8%)
Skipped medication	Skipped medication doses to save money in past 12 months	Yes	rxsk12m a	1,218 (4%)
Took less medication	Took less medication to save money in past 12 months	Yes	rxls12m_a	1,314 (4%)
Delayed filling prescriptions	Delayed filling prescription to save money in past 12 months	Yes	rxdl12m a	1,632 (5%)
Skipped filling needed prescriptions	Needed prescription medication but did not get it due to cost in past 12 months	Yes	rxdg12m_a	2,079 (7%)
Delayed counseling or therapy	Delayed counseling or therapy due to cost in past 12 months	Yes	mhthdly_a	1,395 (4%)
Skipped needed counseling or therapy	Needed counseling or therapy but did not get it due to cost in past 12 months	Yes	mhthnd_a	1,330 (4%)
Worked last week	Worked last week	Yes	empwrklswk_a	18,899 (60%)
Adults in family working	Number of adults in family who are working	0 adults	pcntadtwkp_a	8,970 (29%)
		1 adult		11,759 (38%)
		2 adults		8,806 (28%)
		3+ adults		1,743. (6%)
Food stamps in past 12 months	Received food stamps in past 12 months	Yes	fsnap12m_a	3,185 (10%)
Food security	Food security category	High food security (1)	fdscat4_a	27,187 (87%)
		Marginal food security (2)		1,595 (5%)
		Low food security (3)		1,424 (5%)
		Very low food security (4)		1,072 (3%)
Duration in residence	Duration lived in house or apartment	Less than 1 year (1)	houyrsliv a	4,021 (13%)
		1–3 years (2)		6,806 (22%)
		4–10 years (3)		7,807 (25%)
		11–20 years (4)		5,964 (19%)
		More than 20 years (5)		6,680 (21%)
Owner or renter	Own or rent residence	Rented (0)	houtenure_a	9,920 (32%)
		Owned or being bought (1)		21,358 (68%)

High-impact chronic pain.

High-impact chronic pain was defined as pain reported on most days or every day during the previous 3 months that limited life or work activities on most days or every day during the previous 3 months [11].

Socioeconomic position (SEP).

The SEP indicators reflected features in multiple domains. Education features included educational attainment and highest education level completed by any adult in the home. Categories for both education variables were defined based on prior work in the NHIS dataset [55]. Family income features included imputed family income (continuous variable top-coded at $220,000), imputed family income group (categorical variable), and imputed ratio of family income to poverty threshold (continuous variable top-coded at 10.0 and categorical variable).² Family income was determined as the sum of income (before taxes) for all family members living in a household, excluding capital gains or noncash benefits (e.g., food stamps). The poverty threshold (dollar amount) for 2019 was determined based on family size, number of children, and presence of a person 66 years or over. Health care utilization features included problems with medical bills, medications, and counseling/therapy. Employment features included working in the past week and number of adults in the family working. Housing features included length of time in current house or apartment and owning or renting current residence. Food security was determined by 10 questions about household food security in the past 30 days. Geographic characteristics based on respondent residence included 2013 urban-rural classification and region of the United States.

Mental health indicators.

Depression features included history of depression, frequency of depressed mood, medication for depression, and severity of depressive symptoms in the past two weeks according to the 8-item Patient Health Questionnaire (PHQ-8 total score). Anxiety features included history of anxiety disorder, frequency of anxious mood, medication for anxiety, and severity of anxiety symptoms in the past two weeks according to the 7-item Generalized Anxiety Disorder scale (GAD-7 total score).

Demographic and physical health indicators.

Demographic characteristics included age (a categorical variable was adopted based on evidence for non-linear associations between age and chronic pain [22; 54]; NHIS top-coded age at 85), sex (information about gender identity was unavailable in the 2019 NHIS survey), race and ethnicity, sexual orientation and marital status. Physical health characteristics included self-reported body mass index (BMI) category, cigarette smoking status, general self-rated health (SRH), opioid use in the past 12 months, and ever being told by a doctor they had arthritis, cancer, diabetes, hypertension, high cholesterol, coronary heart disease, a stroke, or respiratory disease.

Subgroups.

Secondary analyses evaluated feature importance for high-impact chronic pain separately for Non-Hispanic Black/African Americans and Non-Hispanic Whites as well as for males and females.

Data curation.

We used a light-touch approach to participant exclusion and feature imputation to avoid affecting subsequent model interpretations with any prior assumptions. Data manipulation was minimized in the process of training and testing models. Based on guidelines[31], variables of interest with >30% missing data were excluded. Missing features imputation was performed by replacing missing values with the median for numerical variables and the mode for categorical variables. Imputation strategies (multiple or machine learning-based) were eschewed due to concerns that learning from the existing correlational structure could affect the determination of feature importance. Preliminary analyses using Python’s Scikit-learn iterative random forest imputer [41] did not improve model accuracy. As a result, all missing values were imputed using the median or mode. Sex was not reported for three individuals; these were removed from analyses because it is not feasible to impute this information accurately.

Modeling algorithms and evaluation.

We evaluated three modeling algorithms for classifying high-impact chronic pain among NHIS participants. These models were compared using Area Under the Curve (AUC) and accuracy metrics. AUC was used to evaluate the discriminatory power of models across different classification thresholds and provided a comprehensive measure of overall model performance. Accuracy, on the other hand, was used to determine the overall correctness of model predictions and provided insight into the proportion of correctly classified cases. Together, these two metrics offer a balanced assessment of both discriminatory power and overall classification accuracy. In case of disagreement between accuracy and AUC, we prioritized AUC because it considers the true positive rate (sensitivity) and false positive rate (1-specificity) across different threshold settings, which makes it particularly valuable in scenarios where the dataset is imbalanced.

H2O Auto ML library [33] was used to automate the model training and hyperparameter tuning processes for Gradient Boosting Decision Trees (GBDT), Random Forests (RF), and Deep Neural Networks (DNN). H2O AutoML systematically explores and optimizes hyperparameters for various machine learning models using random grid search, which enhances prediction accuracy and performance. For GBDT, it tunes parameters such as the number of trees, learning rate, maximum depth, minimum number of observations for a leaf, and sampling ratio. For RF, it optimizes the number of trees, maximum depth, and sampling rate. For DNN, it adjusts the number of layers, neurons in each layer, dropout rate, optimization algorithm, learning rate, activation function, and batch size. This automated approach eliminates the need for manual hyperparameter tuning and ensures all models are optimally configured. Model evaluation was conducted in Python 3.8 using PyTorch 1.10 and H2O 3.42. The following three modeling algorithms were used in the present study: GBDT, RF, and DNN.³ Each of these models offer unique strengths to the predictive task in terms of accuracy and robustness.

Gradient Boosting Decision Trees.

GBDT are powerful ensemble models that sequentially build weak learners to correct errors made by preceding models by training on residuals in boosting rounds. In this study, GBDT was implemented to enhance prediction accuracy. The algorithm iteratively builds decision trees, optimizing for prediction performance.

Random Forest (RF).

RF is an ensemble learning method that constructs a multitude of decision trees and merges their predictions. Each tree is built on a bootstrap sample of the original dataset, and randomness is introduced by considering a subset of features at each split. This ensemble approach helps mitigate overfitting (i.e., performing well on training data but poorly on unseen data) by taking an average of predictions across many trees, which reduces variance and enhances generalization. In addition, this approach improves robustness by maintaining high performance on the test set, thereby ensuring stability and reliability.

Deep Neural Networks (DNN).

DNN were employed to harness the power of neural network-based models. DNNs learn complex patterns and dependencies directly from the data, without assuming any specific stochastic process. The hyperparameters, including the number of layers, neurons in each layer, dropout rate, optimization algorithm, learning rate, activation function, and batch size, were tuned for optimal model performance.

Model interpretability.

Machine learning models are often viewed as black-box procedures yielding little insight or interpretability except for predictions of outcomes. However, recent improvements have been made in the generation of robust and interpretable insights from complex machine learning models[1]. Shapley Additive Explanation (SHAP[32]) values have gained attention because they can facilitate interpretation of complex machine learning models with high accuracy and robustness. By comparing SHAP values generated for input features, it is possible to assess the extent to which changes in the inputs influence the final model’s prediction and, hence, to evaluate feature importance for complex models. Importantly, SHAP values must be interpreted within the context of the model’s scale (i.e., values are in the same units as the model’s output) and the specific dataset; SHAP values cannot, therefore, be compared between models.

The present study evaluated SEP, mental health, demographic, physical health, and geographic factors as input features in GBDP, RF, and DNN models for high-impact chronic pain risk prediction. After training models on a training set, model hyperparameters were optimized on an evaluation set. Model accuracy and AUC were then reported on a hold-out test set to ensure robust performance metrics. Before calculating SHAP values, the model was retrained on the full dataset using the same parameters, increasing the number of samples for model training. Retraining on the full dataset (versus relying on the limited training set) is necessary for accurate interpretation of feature contributions and interactions. The K-means clustering algorithm (K=100) was applied to the dataset; this provided a total of 100 cluster centroids to be used as a manageable background sample for SHAP value calculation.⁴ K-means also ensured that this background sample captured the dataset’s diversity while reducing computational complexity. SHAP values were then generated for input features based on the trained model and the background samples; they provide instance-specific contributions. Feature importance was then reported as the mean of the absolute SHAP values for each feature across all instances in the dataset. Figures displayed the top 50 features for each model, ranked by mean absolute SHAP values. Supplemental figures showing positive and negative SHAP values are included for high-impact chronic pain in the full sample (Supplemental Figure 1), Non-Hispanic Blacks (Supplemental Figure 2), Non-Hispanic Whites (Supplemental Figure 3), Females (Supplemental Figure 4), and Males (Supplemental Figure 5). Whereas negative SHAP values indicate that a feature decreases the probability of high-impact chronic pain, positive SHAP values indicate that a feature increases the probability of high-impact chronic pain. The magnitude of each SHAP value reflects its importance and influence on the model’s predictions.

Results

A comparison of model accuracy and discriminatory power results is presented in Table 2. Of the machine learning algorithms evaluated, GBDT achieved the highest overall accuracy and discriminatory power for high-impact chronic pain (accuracy=92.63%, AUC=0.88).

Table 2.

Model Evaluation: Accuracy and Discriminatory Power

Algorithm	High-impact chronic pain
	Accuracy	AUC
Gradient Boosted Decision Tree (GBDT)	0.9263	0.8882
Random forest (RF)	0.9238	0.8558
Feed Forward Deep Neural Network (DNN)	0.9180	0.8467

Note: AUC = Area Under the Curve.

Feature importance.

Relative importance of features, sorted by mean absolute SHAP values, is presented for high-impact chronic pain (Figure 1). SRH was the most important feature for classification of high-impact chronic pain. This variable captured how each respondent perceived their current health, ranging from poor to excellent. Other physical health measures, including respondent ever being told by a doctor they had arthritis and any prescription opioid use in the past year, were also highly important features. Mental health measures were also highly important features for high-impact chronic pain, including severity of depressive symptoms over the past two weeks (PHQ-8 total score), depression frequency, severity of anxiety symptoms in the past two weeks according to the 7-item Generalized Anxiety Disorder scale (GAD-7 total score), and anxiety frequency.

Figure 1.

Relative importance for the study features predicting high-impact chronic pain status by mean absolute SHAP value.

Six SEP variables were among the 15 most relevant features for high-impact chronic pain status. These included whether or not the respondent worked in the previous week, family income (continuous variable), ratio of family income to poverty threshold (both continuous and categorical variables), number of working adults in the family, and duration respondents had been living in their house or apartment.

Subgroup comparisons of high-impact chronic pain.

Race.

Relative importance of features, sorted by mean absolute SHAP values, are presented for high-impact chronic pain separately for Non-Hispanic Black/African Americans (Figure 2) and Non-Hispanic Whites (Figure 3). The top 5 most relevant features were the same for both groups. The relative importance of BMI, rurality/urbanicity, and owning or renting a residence for high-impact chronic pain was higher among Non-Hispanic Black/African American adults than among Non-Hispanic White adults. In contrast, the relative importance of number of working adults in the family, housing stability (i.e., duration living in their house or apartment), and food insecurity for high-impact chronic pain was higher among Non-Hispanic White adults than among Non-Hispanic Black/African American adults.

Figure 2.

Relative importance for the study features predicting high-impact chronic pain status among Non-Hispanic Black/African Americans by mean absolute SHAP value.

Figure 3.

Relative importance for the study features predicting high-impact chronic pain status among Non-Hispanic Whites by mean absolute SHAP value.

Sex.

Relative importance of features, sorted by mean absolute SHAP values, are presented for high-impact chronic pain separately for females (Figure 4) and males (Figure 5). The top 7 most relevant features were the same for both groups. The relative importance of anxiety symptom severity, BMI, and cigarette smoking status for high-impact chronic pain was higher among females than among males. In contrast, the relative importance of housing stability (i.e., duration living in their house or apartment) and frequency of anxiety and depression for high-impact chronic pain was higher among males than among females.

Figure 4.

Relative importance for the study features predicting high-impact chronic pain status among females by mean absolute SHAP value.

Figure 5.

Relative importance for the study features predicting high-impact chronic pain status among males by mean absolute SHAP value.

Discussion

The burden of chronic pain is not borne equally across socioeconomic strata. Socioeconomic position (SEP) is both a key social determinant and consequence of chronic pain [19]. Lower SEP is associated with worse pain outcomes [14; 27; 38; 55] and can limit access to quality health care [9; 26]. However, SEP is a multifaceted construct and there is no consensus on how best to capture SEP features most relevant to chronic pain [18; 42; 43]. Selecting SEP measures based on convenience and assuming they are interchangeable with respect to health outcomes [2] could hinder progress toward identifying unique, potentially modifiable targets for interventions to address socioeconomic disparities in chronic pain. The present cross-sectional study sought to identify the most relevant SEP indicators for classifying prevalent high-impact chronic pain in a large sample of U.S. adults from the 2019 National Health Interview Survey. Consistent with prior research [7; 13; 21; 24; 30; 55], SRH, arthritis, and prescription opioid use were the most important indicators of high-impact chronic pain. Notably, a variety of SEP facets (e.g., income, employment) were among the most important indicators of high-impact chronic pain – more so even than well-established chronic pain risk factors such as cigarette smoking status and body mass index (BMI) [57; 58]. The present study adds to this literature by showing that distinct SEP facets, including material resources and employment, were highly relevant indicators for classifying high-impact chronic pain.

Overall, income indices were more relevant for high-impact chronic pain classification than other financial risk factors for chronic pain previously examined in the NHIS, including financial worries [52] and food insecurity [50]. Prior work suggests that the benefit of additional income on health outcomes is greatest for those with the lowest SEP[2]. Together, these findings highlight the potential for interventions that help those with lower SEP to achieve greater balance between household income and expenses (i.e., financial stability) through cash (and cash equivalent) transfers, which include short-term consumption support (e.g., for health care, food, and housing) and long-term asset building (for stability and wellbeing) [49].

Challenges measuring and interpreting SEP are well-documented in health disparities research [42; 43]. For example, education is generally associated with greater access to resources and is less likely than other SEP factors (e.g., income, employment status) to change with onset of high-impact chronic pain in adulthood. However, chronic pain in adolescence is associated with lower subsequent educational attainment in young adulthood [40]. There is also evidence that the economic returns of education are not distributed equally, favoring Non-Hispanic White males over minoritized racial groups and females, and being less relevant for predicting health outcomes among those living below the poverty line [29]. Together, these characteristics make education less useful for understanding how short-term changes in SEP influence chronic pain among more vulnerable segments of the population [43]. Wealth, in contrast, better captures financial resilience following sudden economic shocks such as traumatic injuries and is more closely linked to social class than income [43]; however, available measures are vulnerable to individual differences in both willingness and ability (i.e., financial literacy) to accurately report. The present findings identified a selection of indicators that captured multiple SEP facets, including access to material resources, employment status, and exposure to financial strain.

Prior theoretical reviews have also recommended stratifying data analyses by race and sex to evaluate similarities and differences in associations between SEP features and pain [42; 43]. Key findings from subgroup analyses revealed features that were relevant for high-impact chronic pain among both Non-Hispanic Black and White adults and among both men and women: SRH, arthritis, prescription opioid use in the past 12 months, depression severity, family income, having worked in the past week, and ratio of family income to poverty threshold. However, several important differences emerged between Non-Hispanic Black and White adults and between men and women in the relative importance of SEP features. Whereas BMI and homeownership had higher relevance as an indicator of high-impact chronic pain among Non-Hispanic Black as compared to White adults, housing stability and food security had higher relevance as indicators among Non-Hispanic White compared to Black adults. These results conflict with research reporting no consistent racial or ethnic differences in associations between BMI and chronic pain [46]. However, results are somewhat consistent with work in African-American and White males showing that taking less medications due to their cost predicted pain among both groups, but food insecurity, financial strain, and difficulty paying bills were only associated with pain among White males [45]. Evidence suggests that racial inequities in SRH can be explained, in part, by racial differences in homeownership [4]. Whether racial differences in the association between homeownership and high-impact chronic pain can be attributed, in part, to residential segregation should be investigated in future studies.

Sex differences emerged in the relative importance of predictive features for high-impact chronic pain. Anxiety symptom severity, BMI, and smoking had higher relevance as indicators of high-impact chronic pain among females compared to males, whereas housing stability and frequency of anxiety and depression had higher relevance as indicators among males as compared to females. These findings are consistent with evidence for stronger associations between BMI and chronic pain among women than men [46]. Although prior work highlights the relation between housing insecurity and poor physical health outcomes (e.g., lower SRH), with stronger relations observed for Black as compared to White adults [5], to our knowledge potential sex differences in the association between housing insecurity and chronic pain have yet to be examined. The present study highlights the potential for machine learning algorithms to advance health equity research, which has typically relied on stratified multivariate logistic regression models [44], by leveraging SHAP values to compare the relative importance of SEP features within different groups. Machine learning is well-suited to enhancing prediction accuracy for complex health outcomes [6; 8] such as chronic pain because algorithms can simultaneously examine large numbers of variables and their interactions while allowing for non-linear relationships with outcomes [34].

Limitations of this study provide directions for future research. First, cross-sectional data precluded inferences about the strength or direction of causal relations between SEP features and high-impact chronic pain. It is likely that lower SEP increases risk for new onset or worsening of chronic pain, which, in turn, would have a negative impact on a variety of SEP features. The cross-sectional design also precludes causal inferences about relations between other important features and chronic pain. For example, both individual differences in prescription opioid use and SRH may be consequences of the development of chronic pain [10; 37]. Prospective studies are needed to clarify the mechanisms through which socioeconomic pain inequalities are produced and maintained. Second, subjective SEP features (e.g., perceptions of social status) were underrepresented compared to objective SEP features in the NHIS dataset. Although subjective and objective SEP measures tend to be moderately correlated, subjective SEP has a stronger association with perceived wellbeing [51]. Subjective SEP features were less important than objective SEP features for classifying chronic pain in the present study.

Third, several potentially important SEP domains were neglected, such as wealth [22; 47], assets (e.g., savings, stocks and bond), debt (e.g., credit card debt, payday loans), and financial capability (i.e., financial knowledge, access and behavior) [48]. Further prospective research is warranted to provide a thorough and accurate assessment of socioeconomic factors in order to clarify whether they precede and predict changes in high-impact chronic pain; ultimately, this understanding could inform the development of effective prevention and intervention programs that target multiple contributing pathways through which SEP likely affects pain. Fourth, health insurance type predictors were not included in models due to differences in eligibility between types that would make comparisons of feature importance challenging. Fifth, subgroup analyses focused on comparisons between Non-Hispanic Black and White adults, which were driven by prior work contrasting chronic pain risk in these groups [11; 56] but neglects other groups likely to be impacted by chronic pain [53]. Findings may differ for other racial and ethnic groups who were excluded from subgroup analyses but comprised over 20% of the 2019 NHIS dataset. Finally, results focused on age and educational attainment should be interpreted with some caution due to the inclusion of emerging adults still enrolled in degree programs [55].

Chronic pain conditions substantially reduce quality of life and exert an enormous economic burden on society[20]. Pharmacologic and nonpharmacologic therapies for chronic pain typically yield significant but small-to-medium effects[12], suggesting the importance of examining additional biopsychosocial factors as potential targets for intervention – particularly those that promote health equity by focusing on socially vulnerable groups. The present findings highlight key SEP features that contributed to highly accurate classification of prevalent high-impact chronic pain status. Importantly, some of the most important SEP indicators of prevalent high-impact chronic pain were potentially modifiable, including family income. If these SEP features are found to precede and predict changes in pain experiences and interference, they could help to inform individual- and policy-level interventions to mitigate disparities in chronic pain.

Data availability:

The National Health Interview Survey (NHIS) 2019 dataset is publicly available, as are the datasets generated from it and analyzed for the current study (upon request from the corresponding author).