Does the procedure really improve survival — or is it patient selection?
A 1.2-million-patient survival analysis that used instrumental-variable methods to separate a real treatment effect from confounding.
The published work
The challenge
Observational studies had suggested that multiarterial grafting during coronary artery bypass surgery improves survival compared with single arterial grafting. But observational comparisons are vulnerable to confounding: the patients who receive the more complex procedure, and the surgeons who choose it, differ in ways that also affect outcomes. Does the procedure cause the benefit, or does selection?
The approach
Across more than 1.2 million Medicare beneficiaries, the team used flexible parametric survival models with time-dependent effects and regression standardisation to compute standardised survival probabilities, and compared two analytical strategies side by side: a conventional risk-adjusted model, and an instrumental-variable model that used each surgeon's multiarterial-grafting rate as the instrument.
Alessandro Gasparini, then at Red Door Analytics, provided the statistical methodology and analysis. The work was published in the Journal of the American College of Cardiology.
What it found
The conventional model showed a modest survival advantage for multiarterial grafting, in line with previous observational research. The instrumental-variable model did not — suggesting the apparent benefit may reflect unmeasured confounding tied to patient selection and surgeon decision-making, rather than the procedure itself.
Why it matters
Two defensible methods, two different conclusions: it is a clean illustration of how much the analytical strategy matters when the question is causal and the data are observational — exactly the scrutiny that regulatory and HTA audiences apply to real-world evidence.
When the strategy decides the answer, it matters a great deal that the strategy was fixed before the answer was known. That is the argument for writing the analysis plan as a structured register — the estimand, the confounding structure and the adjustment set each recorded as a field, and checked, before any result exists — rather than as prose reconstructed afterwards.
It also runs on the same flexible-parametric survival and standardisation machinery that powers the merlin family.