Data were drawn from the Health and Aging in Africa: A Longitudinal Study (HAALSI), a population-based longitudinal study conducted in the Agincourt Health and Socio-Demographic Surveillance System (HDSS) site in rural Mpumalanga, South Africa (2). This region is undergoing a rapid epidemiological transition, characterized by a high prevalence of HIV alongside an increasing burden of non-communicable diseases (NCDs).

The analysis utilizes data from three waves of data collection: Wave 1 (Baseline, 2015), Wave 2 (2018), and Wave 3 (2021–2022). The study received ethical approval from the University of the Witwatersrand and the Harvard T.H. Chan School of Public Health, and informed consent was obtained from all participants.

The baseline cohort consisted of adults aged 40 years and older. For this analysis, the analytic sample was constructed as an unbalanced panel, including all individuals with valid data on the primary outcome (household health expenditure) in at least one wave. The final sample comprised 4,371 unique individuals, contributing a total of 13,437 person-wave observations, thereby maximizing the use of available longitudinal data.

The HAALSI cohort was designed to include adults aged 40 years and older to capture aging-related transitions beginning in late midlife, including the onset of non-communicable diseases and early functional decline in a setting shaped by long-term socioeconomic disadvantage and historically lower life expectancy. Retaining the full ≥40 cohort preserves representativeness and statistical power while allowing age to be modeled continuously. Because “older adulthood” is often operationalized as ≥60 years in global and African policy contexts (3), we include a sensitivity analysis restricted to participants aged ≥60 years to assess whether key associations are robust to this alternative definition.

Variables were conceptualized using the Adapted Andersen Model of Health Services Utilization (4), which distinguishes between predisposing, enabling, and need factors.

The primary outcome was total monthly household out-of-pocket health expenditures. This was calculated by aggregating self-reported costs for doctor/nurse visits, hospital fees, traditional healers, and medication purchases. To address the “two-part” nature of health spending data (characterized by a large mass of zero-spenders and a right-skewed distribution of positive costs), the outcome was operationalised in two forms (5): (a) Market Entry (Binary): a variable indicating whether the household incurred any OOP cost (>0) vs. none (0). (b) Financial Toxicity (Continuous): for the sub-sample of individuals who incurred costs, the total amount was log-transformed [ln (cost + 1)] to normalize the distribution for linear modeling.

“Need” was operationalized across three distinct domains to test the hypothesis that functional decline drives costs differently than disease diagnosis: (a) Chronic Disease Diagnoses: Binary variables were generated for Hypertension (measured blood pressure BP ≥140/90 mmHg or current medication use) and Diabetes (fasting glucose ≥7.0 mmol/L or current medication use). HIV status and History of Stroke were determined via self-report. (b) Objective Physical Function: Grip Strength: measured using a digital dynamometer (kg). The maximum reading from two trials on each hand was utilized as an objective marker of physical robustness. Activities of Daily Living (ADLs): a cumulative count (0–6) of reported difficulties in six core activities: walking, dressing, bathing, eating, getting in/out of bed, and toileting. Cognitive and Mental Health: Cognitive Function: assessed using a standard Delayed Recall task, scored as the total number of words (0–10) correctly recalled from a list after a delay. Depression: assessed using the 20-item Center for Epidemiologic Studies Depression Scale (CES-D; 6). In Wave 1, participants completed an 8-item CES-D (range 0–8), whereas Waves 2–3 administered the 20-item CES-D (range 0–60). To enable pooled analyses across Waves 1–3, we harmonized depressive symptom burden by converting each wave-specific score to a per cent-of-maximum-possible (POMP) and rescaling it to a 0–60 metric (6). Accordingly, higher values indicate greater depressive symptom burden; however, because the harmonized score combines different CES-D versions, it should be interpreted as a measure of symptom burden rather than a diagnostic threshold. For context, a CES-D-20 score ≥16 is commonly used to flag elevated depressive symptoms.

Control variables included Age (continuous), Sex (Female/Male), Marital Status (Married/Partnered vs. Single/Widowed/Divorced), and Education Level (No formal education to Secondary+). Household Wealth was assessed using a standard asset-based wealth index quintile derived from principal component analysis.

The decision to use single (simple) imputation was pragmatic and aimed at minimizing listwise deletion in an unbalanced longitudinal panel, which can disproportionately exclude older and more vulnerable participants and reduce precision. We acknowledge that single imputation can underestimate uncertainty (i.e., standard errors) and may attenuate associations; therefore, results should be interpreted cautiously and we highlight this as a limitation.

To handle missing data in predictor variables across the three waves, we employed imputation (series mean for continuous variables and mode imputation for categorical variables) to minimize selection bias; observations with missing values for the primary outcome were excluded.

To account for the correlation of repeated measures within individuals over time, Generalised Estimating Equations (GEE) were employed (7). GEEs provide population-averaged estimates and are robust to misspecification of the correlation structure. We adopted a two-part model framework: (a) Model 1 (Probability of Spending): a GEE with a binomial distribution and logit link function was used to estimate Adjusted Odds Ratios (AOR) for the likelihood of incurring any OOP expenditure. (b) Model 2 (Magnitude of Spending): a GEE with a normal distribution and an identity link function was used to estimate coefficients for log-transformed expenditure in the subsample of spenders.

Importantly, given the observational design, all coefficients are interpreted as population-averaged associations rather than causal effects. GEE estimates average differences in outcomes associated with predictors across the population and does not, by itself, establish causality or individual-level trajectories.

Both models utilized an exchangeable correlation structure with robust standard errors (Huber-White sandwich estimator) to account for within-subject clustering. All models were adjusted for time effects (Wave). Data preparation and analysis were conducted using IBM SPSS Statistics, Version 30. Statistical significance was defined at p < 0.05.

To address concerns that including participants aged 40–59 may dilute aging-related effects, we specify a robustness check in which both GEE models are re-estimated in a restricted sample aged ≥60 years (using the same modeling strategy and covariate set).

The final analytical sample consisted of 4,371 unique individuals contributing a total of 13,437 person-wave observations across the three waves of data collection (2015–2022). At baseline, the cohort's mean age was 63.9 years (SD = 12.6), and 38% of respondents were female.

The prevalence of multimorbidity was high: 2,841 (65.0%) participants were classified as hypertensive, 1,399 (32.0%) as diabetic, and 612 (14.0%) were living with HIV.

In terms of functional health, the cohort showed relatively high levels of independence at baseline, with a mean Functional Score of 5.8 out of 6 (where 6 indicates total independence). However, mental health vulnerabilities were evident, with a mean harmonized CES-D symptom score of 6.6 (SD = 8.8), which is well below commonly used CES-D-20 screening thresholds (e.g., ≥16).

Across the three waves, 60.7% of person-time observations incurred out-of-pocket (OOP) health expenditures. Among those who incurred costs, spending was highly skewed, necessitating the use of log-transformed expenditure in subsequent linear models (see Table 1).

The first Generalised Estimating Equation (GEE) model examined factors associated with the likelihood of incurring any out-of-pocket health expenditure (Table 2).

Socioeconomic status and chronic disease diagnoses acted as the primary “gatekeepers” to the fee-bearing healthcare system. Wealth exhibited a steep gradient; wealthier individuals had significantly higher odds of incurring health costs (AOR 1.17, 95% CI 1.13–1.21, p < 0.001). Similarly, a diagnosis of Hypertension was a robust driver of spending entry (AOR 1.26, 95% CI 1.14–1.40, p < 0.001), while HIV-positive status showed a smaller but positive association (AOR 1.16, p = 0.046).

However, measures of functional and mental vulnerability showed an inverse relationship. Contrary to the hypothesis that disability increases the probability of spending, higher ADL limitations were significantly associated with a lower likelihood of incurring costs. For every additional ADL limitation, the odds of entering the spending market decreased by 16% (AOR 0.84, 95% CI 0.77–0.91, p < 0.001). Similarly, higher Depression scores (AOR 0.97 per point, p < 0.001) and poorer cognitive function (lower Recall scores) were associated with a reduced probability of spending. This pattern suggests that the most functionally and cognitively vulnerable older adults may face barriers to accessing fee-based care or rely heavily on free public services.

Among those who incurred costs (n = 4,565), the drivers of spending intensity differed notably from the drivers of entry (Table 3).

The “ability to pay” remained the strongest predictor of expenditure magnitude. Wealth (B = 0.131, p < 0.001) and Education (B = 0.110, p < 0.001) were positively associated with higher spending, indicating a regressive system in which those with greater resources bear higher costs.

Unlike the binary model, objective physical weakness emerged as a significant cost driver. Grip Strength was negatively associated with spending (B = −0.006, p = 0.012), indicating that physically weaker individuals incurred higher costs. This supports the hypothesis that physical frailty translates into tangible financial toxicity for those who remain within the care system.

Consistent with Model 1, markers of severe vulnerability continued to predict lower spending intensity. Higher ADL counts (B = −0.223, p < 0.001) and Depression scores (B = −0.078, p < 0.001) were associated with significantly lower expenditure levels. This reinforces the finding of an “exclusion zone,” where the most severely disabled individuals may be unable to afford or access the high-cost services utilized by their healthier or wealthier counterparts.

Notably, HIV-positive status was associated with significantly lower OOP spending (B = −0.155, p = 0.008) compared to HIV-negative counterparts. This likely reflects the protective financial effect of South Africa's free public antiretroviral therapy (ART) program compared to the cost burden associated with NCDs and frailty.

Consistent with the Andersen Model's enabling factors, wealth and education were robust predictors of both market entry and spending magnitude. This aligns with extensive evidence from South Africa and the wider region showing that, despite the removal of some user fees, healthcare access remains highly regressive (8, 31). Crucially, hypertension emerged as a primary driver of entry into the paying system. This likely reflects the “medicalisation” of aging in South Africa, where mass screening programs successfully identify NCDs and link patients to care (9). However, once diagnosed, these patients incur recurrent costs (transport, medication co-pays, and ancillary fees) that constitute a “chronic tax” on their household budgets, a phenomenon well documented in the syndemic literature (10).

Our most novel finding is the divergence between early physical frailty and severe disability.

Among those who accessed care, weaker grip strength was associated with significantly higher costs. This supports the hypothesis that physical frailty is financially toxic, consistent with global gerontological data suggesting that early functional decline drives service utilization for rehabilitation and management (11).

Paradoxically, as vulnerability deepened into overt disability (higher ADL counts), cognitive impairment, or depression, the likelihood and magnitude of spending decreased. This inverse relationship contradicts standard health economic theory in high-income countries, where disability is a primary driver of exponential cost growth (12). In this resource-constrained rural setting, this likely signals unmet need. Studies on disability in Africa have consistently identified physical barriers (e.g., terrain, transport) as primary determinants of exclusion (13). Furthermore, the “treatment gap” for mental health in low-resource settings is well-known; Chisholm et al. (14) note that despite the high burden of depression, expenditure remains negligible due to a lack of available services. Our data suggest that the “poorest and sickest” are not spending less because they are healthy, but because they have dropped out of the formal financial economy of health care.

The finding that HIV-positive status was associated with lower spending magnitude is a testament to the success of South Africa's public antiretroviral therapy (ART) program. Unlike NCDs or frailty, which often requires piecemeal OOP expenditure, HIV care is vertically integrated and largely free at the point of service (15). This creates a disparity where “ageing with HIV” may be more financially protected than “ageing with frailty” (28). This validates calls from the Lancet NCDI Poverty Commission to extend the financial risk protection models of HIV care to severe non-communicable conditions and geriatric care (16).