How Endpoint Definitions Shape Epidemiologic Inference on Tuberculosis Treatment Outcomes: A Secondary Analysis from Lima, Peru

Background/Objectives: Tuberculosis treatment outcomes are often dichotomized into heterogeneous composites that may obscure outcome-specific associations. We compared baseline characteristics across mutually exclusive out-comes—loss to follow-up (LTFU), death, and not evaluated/transferred versus documented cure—and examined the composite endpoints secondarily in a prospective cohort of adults with smear-positive pulmonary tuberculosis treated in Lima, Peru, in 2010–2011. Methods: The complete-case cohort included 1233 participants. The primary analysis used ridge-penalized multinomial logistic regression (lambda = 0.16); a mean bias-reduced multinomial model assessed sparse-data sensitivity. Results: Overall, 1016 participants were cured, 127 experienced LTFU, 30 died, and 60 were not evaluated/transferred. Drug use was associated primarily with LTFU (aOR 4.15; 95% CI 2.55–6.75), whereas HIV-positive status showed the largest, although imprecisely estimated, association with death (aOR 32.12; 95% CI 8.95–115.33; six deaths among 22 HIV-positive participants). The bias-reduced estimate was similar (aOR 35.56; 95% CI 9.80–129.08). Multidrug-resistant tuberculosis was associated with several non-cure categories; its association with the combined not evaluated/transferred category was largely driven by participants still in treatment at database closure. Composite endpoints changed the magnitude and interpretation of several associations and did not eliminate sparse outcome–exposure cells. Conclusions: Endpoint definitions therefore materially shaped epidemiologic inference. Dis-aggregated outcomes should be reported alongside composite indicators, and findings from this historical cohort should not be interpreted as contemporary risk estimates for Lima.