Competing-risk analysis of tuberculosis outcomes in Liangshan

A retrospective analysis of 340 closed tuberculosis treatment episodes among people living with HIV in Liangshan Yi Autonomous Prefecture, China, estimated that the cumulative incidence of an unfavorable treatment outcome reached 7.96% by 12 months, with a 95% confidence interval of 5.30% to 10.91%. The study treated cure or treatment completion as competing events because, once either occurred, a later unfavorable outcome could no longer occur within the same treatment episode. The result represents the observed probability of an unfavorable outcome in the presence of favorable outcomes, rather than treating completion as ordinary censoring. Only 33 participants remained at risk at the 12-month landmark, and the unfavorable-outcome composite was dominated by deaths not attributed to tuberculosis, limiting the precision and clinical meaning of the estimate.

The researchers analyzed patient-level records from routine tuberculosis case management in Liangshan, Sichuan Province. The source workbook contained 521 tuberculosis/HIV records, from which the investigators selected 340 episodes with a closed status. They excluded 153 episodes that were still under treatment and 28 re-registered episodes, analyzing the closed cohort over an administrative observation and censoring window from 20 October 2024 to 30 June 2026. Treatment initiation was defined as time zero, and elapsed days from the recorded initiation date formed the time scale. Follow-up ended at the earliest recorded treatment-stop date or 30 June 2026. The cohort was overwhelmingly composed of Yi patients, who represented 332 of 340 participants, or 97.6%, and 260 participants, or 76.5%, were male. Mean age was 41.04 years, with a standard deviation of 10.89 years, and the median age was 41 years with an interquartile range of 36–47 years.

Among the 340 episodes, 27, or 7.9%, ended in the composite unfavorable outcome; 303, or 89.1%, ended in cure or treatment completion; and 10, or 2.9%, were administratively censored. The unfavorable group comprised 22 records coded as non-tuberculosis death, three transfers to multidrug-resistant tuberculosis treatment, one treatment failure and one loss to follow-up. Thus, non-tuberculosis deaths accounted for 81.5% of unfavorable outcomes. The source field contained no separately coded category for tuberculosis-related death, so the analysis did not estimate all-cause mortality and could not distinguish the mortality burden attributable to tuberculosis. Transfer to multidrug-resistant treatment was classified as unfavorable because it ended the observed episode, not because the investigators interpreted it as biological treatment failure.

The Aalen–Johansen estimator showed that most of the unfavorable cumulative incidence accrued early. It was 4.41% at 3 months, with a 95% confidence interval of 2.35% to 6.77%; 6.47% at 6 months, with an interval of 3.82% to 9.12%; 7.65% at 9 months, with an interval of 5.00% to 10.60%; and 7.96% at 12 months, with an interval of 5.30% to 10.91%. By 3 months, 55.5% of the 12-month unfavorable cumulative incidence had already accumulated. In parallel, the cumulative incidence of cure or treatment completion rose from 0.29% at 3 months to 14.71% at 6 months, 55.73% at 9 months and 81.83% at 12 months. Pointwise 95% confidence intervals for these cumulative-incidence functions were generated from 2,000 individual-level bootstrap resamples. The analysis did not perform formal pairwise or Gray-type comparisons between the cumulative-incidence functions.

The statistical approach addressed a central feature of tuberculosis treatment data: favorable and unfavorable outcomes compete for the opportunity to occur. In an earlier analysis, cure and treatment completion were treated as censoring observations in Kaplan–Meier and Cox analyses. The authors explained that this practice can alter the estimand and overestimate absolute event probability because it effectively redistributes some probability among participants who remain under observation. Their reanalysis instead used Fine–Gray subdistribution hazard models to examine associations with the cumulative-incidence estimand. A prespecified three-variable sensitivity model included age, sex and patient source recoded as tracing versus all other sources, while a cause-specific hazards model tested whether the principal estimates were directionally consistent. The full model contained 14 parameters and was treated as exploratory because only 27 unfavorable outcomes occurred.

Older age produced the most consistent exploratory association. In the full 14-parameter model, each 10-year increase in age was associated with a 1.54 times higher subdistribution hazard, with a 95% confidence interval of 1.02 to 2.32 and a p value of 0.039. The corresponding association in the parsimonious Fine–Gray model was 1.70, with an interval of 1.18 to 2.43 and a p value of 0.004. A cause-specific hazards model yielded a similar estimate of 1.73, with an interval of 1.20 to 2.50 and a p value of 0.003. These results identify an association within the analyzed cohort, not an independent causal effect or a validated clinical prediction rule. The surveillance extract did not include measurements that could test proposed mechanisms such as physiological reserve, comorbidity, polypharmacy, delayed presentation or treatment toxicity, so the analysis cannot determine why older age was associated with the outcome.

Patients identified through tracing also had elevated estimates in selected analyses. Compared with direct presentation, tracing had an unadjusted subdistribution hazard ratio of 5.45, with a 95% confidence interval of 1.63 to 18.29 and a p value of 0.006. In the full exploratory model, the adjusted estimate was 3.12, with an interval of 0.81 to 11.93 and a p value of 0.097, making it inconclusive in that specification. The parsimonious model gave a subdistribution hazard ratio of 4.73, with an interval of 1.89 to 11.86 and a p value of 0.001, while the cause-specific model gave a hazard ratio of 4.62, with an interval of 1.84 to 11.59 and a p value of 0.001. The investigators did not calculate a crude cumulative incidence stratified by source, so these regression estimates should not be recast as a directly measured difference in absolute risk. They also did not measure possible explanations involving previous disengagement, delayed presentation, disease severity, mobility, stigma or structural barriers.

Associations involving incomplete documentation were much less stable. Five of seven patients whose fixed-dose-combination field was missing and all three patients whose bacteriological result was unavailable experienced unfavorable outcomes. The full model consequently produced an adjusted subdistribution hazard ratio of 5.41 for missing fixed-dose-combination status, with a 95% confidence interval of 1.48 to 19.77, and 79.77 for unavailable bacteriology, with an interval of 11.63 to 547.26. These extreme estimates arose from very small cells and should not be interpreted as effects of treatment formulation or microbiology. The source system did not document why the fields were missing. Early deterioration, death, diagnostic workflows, treatment interruption or service processes could all be related to incomplete records, creating information bias and possible reverse causation. No imputation was performed because the missingness mechanism was unknown. Missing and unavailable values were retained as explicit data-completeness categories rather than treated as biological exposures.

Several design and modeling limitations constrain the findings. The retrospective cohort included only 27 unfavorable events but 14 parameters in its full model, making overfitting, wide confidence intervals, separation and unstable p values likely. The composite endpoint was clinically heterogeneous and primarily reflected mortality rather than a balanced combination of failure, loss to follow-up and transfer. Routine cause-of-death coding may be inaccurate, and the analysis could not include a separate tuberculosis-related death category. The data also lacked CD4 count, HIV viral load, antiretroviral therapy timing and adherence, body mass index, anemia, opportunistic infections, tuberculosis site and severity, drug susceptibility, adverse reactions and socioeconomic measures. The nearly entirely Yi population from one prefecture further restricts generalizability, while selection effects, residual confounding and errors in event dates or classifications remain possible.

The proportional-hazards assessment did not settle whether a constant effect was appropriate over time. A joint interaction test that incorporated uncertainty in the estimated censoring weights suggested non-proportionality, principally involving sex, with a Wald statistic of 9.096, 3 degrees of freedom and a p value of 0.028. An inverse-information sensitivity calculation for the same model did not support that conclusion, producing a statistic of 2.536 and a p value of 0.469. Because only three women experienced unfavorable outcomes, these asymptotic tests were unstable, and proportional subdistribution hazards could not be regarded as established. Overall, the analysis demonstrates how a competing-risk framework can estimate absolute outcome probability while recognizing that favorable treatment completion permanently ends the relevant risk period. Its age and tracing associations warrant evaluation in larger linked cohorts with more events, fuller clinical data, prespecified models, outcome-specific analyses and external validation; the current results do not establish causal effects, clinical prediction performance or a generalizable treatment-failure rate.

Subject of Research: Competing-risk estimation of tuberculosis treatment outcomes among people living with HIV in Liangshan, China

Article Title: Cumulative incidence and factors associated with unfavorable tuberculosis treatment outcomes among people living with HIV in Liangshan, China: a competing-risk analysis

Article References: Yang, N., Li, T., Li, Y., & Xia, L. (2026). Cumulative incidence and factors associated with unfavorable tuberculosis treatment outcomes among people living with HIV in Liangshan, China: a competing-risk analysis. Frontiers in Public Health, 14, Article 1958022. https://doi.org/10.3389/fpubh.2026.1958022

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1958022

Keywords: tuberculosis, HIV, competing risks, cumulative incidence, Liangshan, treatment outcomes, Fine–Gray model, retrospective cohort, mortality, survival analysis, Cumulative, incidence