Income Tied to Cervical Cancer Survival in SEER Causal Analysis

Women with cervical cancer who lived in lower-income neighborhoods died at a markedly higher rate than wealthier patients, according to a retrospective analysis of more than 3,800 cases from the Surveillance, Epidemiology, and End Results (SEER) database. The study, published in Discover Social Science and Health, applied causal inference methods borrowed from epidemiology and machine learning to ask whether area-level income itself influences survival, rather than merely tracking alongside it. The answer the authors report is yes, with important caveats about what observational data can prove.

The research team, led by Anjana Eledath Kolasseri and B Venkataramana, drew on 3,810 cervical cancer patients diagnosed between 2013 and 2015. Each patient was assigned to a household income category—low, medium, or high—based on census tract-level median income linked to their residential address at diagnosis. This is a common proxy for individual socioeconomic status in population-based cancer registries, though it introduces ecological measurement error, since any individual’s income may differ from the neighborhood average. The outcome was survival time in months from diagnosis to death, with patients still alive at last contact right censored.

Before any adjustment, the raw differences were stark. Low-income patients presented more often with advanced-stage disease and larger tumors, and their average survival was dramatically shorter. In the high-income group of 1,545 patients, mean survival was 75.2 months; the medium-income group of 1,210 patients averaged 110.4 months; and the low-income group of 1,055 patients averaged just 30.2 months, a difference the authors report as statistically significant at p < 0.001. Those unadjusted numbers, however, say little about causation, because income correlates with age, race, marital status, tumor characteristics, and treatment patterns.

To disentangle these threads, the researchers estimated each patient’s propensity score—the probability of belonging to a given income category given their baseline characteristics—using multinomial logistic regression. They then constructed stabilized inverse probability weights that rebalanced the income groups, mimicking a randomized comparison. Standardized mean differences, which exceeded 0.3 for many covariates before weighting, fell below 0.1 for nearly all variables afterward, and a Love plot confirmed the improvement. Propensity score distributions overlapped substantially across groups, supporting the positivity assumption required for causal estimation, though the authors note the density plot showed limited overlap between the lowest and highest income categories.

With the weighted dataset in hand, the team fitted a marginal structural model, a Cox proportional hazards model applied to the IPW-weighted population. Compared with high-income patients, low-income patients had a mortality hazard ratio of 1.77 (95% confidence interval 1.33–2.14), and medium-income patients had a hazard ratio of 1.48 (95% CI 1.29–1.93). These are relative hazards, not absolute risks, and they represent population-averaged causal estimates under the model’s assumptions. Schoenfeld residual tests flagged violations of the proportional hazards assumption for age and tumor size, but income itself satisfied the assumption, and the authors argue that violations among adjustment factors do not invalidate the income hazard ratio as an average causal effect over follow-up.

Weighted Kaplan-Meier curves told the same story visually: high-income patients had the best survival over time, with a graded gradient running down through the medium and low-income groups. The log-rank test across income categories yielded p < 0.001. The authors observed that the medium-income group had a slightly higher proportion of administratively censored observations, likely reflecting differences in year of diagnosis and available follow-up time; because SEER follow-up is administratively determined rather than exposure-dependent, they judged this unlikely to introduce systematic bias, though it may shape the curves’ appearance.

The second major analytical pillar was targeted maximum likelihood estimation, a doubly robust method that combines outcome modeling—here powered by a SuperLearner ensemble of random forests, gradient boosting machines, and generalized additive models—with propensity score adjustment. TMLE requires a fixed binary outcome, so the team defined survival at 36 months after diagnosis, a time point chosen because many patients were followed beyond it, providing adequate risk-set support while limiting instability from heavy censoring. Patients with insufficient follow-up were excluded from this analysis. The TMLE estimates indicated that low-income patients survived an average of 4.34 months less than high-income patients (95% CI 2.55–6.13), while medium-income patients showed a deficit of 4.33 months (95% CI 2.53–6.12), both with p < 0.001.

Sensitivity analyses reinforced the pattern. G-computation, the parametric G-formula, produced a similar average treatment effect of a 4.12-month survival deficit for low versus high income, converging with the TMLE result across a third causal framework with different modeling assumptions. Extreme stabilized weights below 0.1 or above 10 were trimmed, and the analyses were rerun. The E-value for the low-versus-high income hazard ratio was 3.5, meaning an unmeasured confounder would need to be associated with both income and survival by a risk ratio of at least that magnitude to explain away the observed effect entirely. Confidence intervals for the average treatment effects came from nonparametric bootstrapping with 1,000 resamples.

What might drive the association beyond stage and treatment delay? The authors point out that even after accounting for stage in the marginal structural model, income effects persisted, suggesting additional routes: differences in treatment quality, access to high-volume cancer centers or guideline-concordant care, cost-related nonadherence, transportation barriers, missed visits, inadequate symptom management, and limited access to palliative care or post-treatment monitoring. Psychosocial stress, low health literacy, and insufficient navigation assistance may compound financial constraints. SEER records none of comorbidities, treatment quality, or these social factors directly, which is precisely where the causal argument is most vulnerable.

The limitations deserve emphasis. As a retrospective cohort study, it cannot eliminate residual confounding; the authors themselves state the findings should be viewed as relationships conditional on measured factors rather than conclusive causal effects. Selection bias is plausible, since SEER captures only patients who reached diagnosis, potentially excluding the poorest women facing screening barriers and thereby underestimating the true impact of inequality. Differential censoring cannot be ruled out if loss to follow-up relates to both income and survival risk. The use of census tract income as an SES proxy may bias estimates toward the null, making the observed differences a conservative approximation. And because the analysis draws on a single US registry population, generalizability to other healthcare systems is uncertain. Still, the convergence of MSM, TMLE, and G-computation estimates, backed by an E-value of 3.5, gives the finding unusual weight for observational cancer research: where a woman lives, and how much money circulates around her address, appears to shape how long she survives cervical cancer, and the authors argue that targeted policies—financial support for treatment, transportation assistance, patient navigation, and expanded insurance coverage—are needed to close the gap.

Subject of Research: Causal analysis of area-level household income and survival outcomes among cervical cancer patients in the SEER database

Article Title: Assessing the causal influence of socioeconomic status on cervical cancer survival using the SEER database

Article References: Kolasseri, A. E., & Venkataramana, B. (2026). Assessing the causal influence of socioeconomic status on cervical cancer survival using the SEER database. Discover Social Science and Health, 6(1), Article 124. https://doi.org/10.1007/s44155-026-00403-9

Image Credits: AI Generated

DOI: 10.1007/s44155-026-00403-9

Keywords: cervical cancer, socioeconomic status, household income, SEER database, survival analysis, causal inference, marginal structural model, TMLE, health disparities, cancer outcomes, Assessing, causal