Longo, Cristina; Bartlett, Gillian; Barnett, Tracie A et Schuster, Tibor . Using Causal Diagrams to Visualize Solutions for the Built-In Selection Bias of the HR in the Context of Time-Varying Exposures: An Example from Asthma Research In: 33rd International Conference on Pharmacoepidemiology & Therapeutic Risk Management, August 26‐30 2017, Montréal, Canada.
Ce document n'est pas hébergé sur EspaceINRS.Résumé
Background: The Hazard Ratio (HR) is one of the most commonly reported effect measures in drug safety and effectiveness studies. Although it is well known that the HR’s built-in selection bias limits its causal interpretation, little work has been done to for- mally explore potential corrections for this bias in the time-varying exposure context.
Objectives: i) To investigate the built-in selection bias of the HR using an example in asthma research, and ii) to propose a weighting-based correction approach for this bias motivated by causal diagrams within the con- text of time-varying exposures.
Methods: We describe two possible causal structures underlying the hypothesized biological relationship between an exposure that may or may not vary over time, and an outcome that could potentially recur (i.e. excluding death). We use a real-life example from asthma research to work through the causal structures, where the exposure is obesity status, measured at base- line and subsequently updated throughout follow-up, and the event of interest is a composite endpoint for asthma exacerbations. Using causal diagrams, we ex- plore the consequences of the built-in selection bias of the HR with respect to the depletion of susceptibles phenomenon and other potential problems, such as non-collapsibility, that result from restricting our analysis to the first event. The resulting bias leads to a spurious time-varying HR and an underestimation of the true average causal effect.
Results: We propose a solution using a weighting ap- proach to correct for the selection bias inherent to the HR. Applying this approach, we emulate what the 'survival' experiences would have been had the deple- tion of susceptibles not occurred. In the asthma exam- ple, the marginal average HR for obesity without the correction was 1.35 (95% CI: 0.95–1.91), with a clinically relevant and statistically significant variation with the squared function of time. In contrast, the marginal average HR for obesity with the correction was 1.67 (95%CI 1.41–1.98), which did not show relevant variation with time.
Conclusions: We propose a bias correction motivated by causal diagrams for the built-in selection bias of the HR. Further research is warranted to determine the conditions in which the validity of this correction method holds.
| Type de document: | Document issu d'une conférence ou d'un atelier |
|---|---|
| Mots-clés libres: | - |
| Centre: | Centre INRS-Institut Armand Frappier |
| Date de dépôt: | 16 août 2018 03:00 |
| Dernière modification: | 16 août 2018 03:00 |
| URI: | https://espace.inrs.ca/id/eprint/7493 |
Gestion Actions (Identification requise)
 |
Modifier la notice |