Using causal frameworks to reduce bias in observational TB research: a comparison of model-building approaches

Background: Observational studies investigating causal (aetiological) questions often address confounding bias using data-driven predictive models where variables are selected regardless of their causal role, challenging interpretation. We compared the estimated effect of HIV co-infection on end-of-treatment outcomes among people with multidrug/rifampicin-resistant tuberculosis using a causal framework model and a data-driven predictive model.Methods: The causal framework guided confounder adjustment. Results were compared with models that applied alternative variable selection strategies (based on p-values and change-in-estimate), which do not draw on domain knowledge to identify relevant variables.Findings: The model informed by a causal diagram indicated that people living with HIV had a 31% lower probability of achieving a successful outcome compared to those without HIV (adjusted risk ratio [aRR] 0.69, 95% CI: 0.41 – 0.98). In contrast, data-driven models produced attenuated associations (aRR 0.78, 95% CI 0.50 – 1.06 and aRR 0.80, 95% CI 0.51 – 1.09 for the p-value and change-in-estimate strategies, respectively), reflecting omission of key confounders, and inappropriate inclusion of mediators and colliders.Interpretation: When the research question is aetiological, using a causal approach to guide variable selection ensures proper adjustment, improves interpretability, and establishes a stronger foundation for future observational research.