Philosophy

Triangulating causal evidence in epidemiology

The problem

Every attempt to estimate a causal effect from observational data rests on assumptions that cannot be fully verified — no unmeasured confounding, correct model specification, no selection bias, accurate measurement. A single estimate, however carefully produced, inherits whatever bias its assumptions let through.

Triangulation

Triangulation is the practice of integrating results from several approaches that have different and unrelated sources of potential bias (Lawlor, Tilling & Davey Smith, 2016). The logic is simple: if two methods would be wrong in different directions for different reasons, their agreement is hard to explain by bias alone. Convergence is evidence; divergence is a signal to investigate.

The etverse builds this idea into tooling at three levels.

Methodological triangulation

Estimate one estimand several ways. G-computation leans on a correctly specified outcome model; IPW leans on a correctly specified treatment model; AIPW is doubly robust; g-estimation targets effect modification; matching trades efficiency for transparency. These assumptions fail independently, so agreement across them is informative.

→ causatr, survatr, separatr

Design triangulation

Ask the same question through study designs whose biases are unrelated — a cohort analysis, a case-control study, a target-trial emulation, a negative-control analysis. A confounding structure that biases one design need not bias another.

→ matchatr, negatr, emulatr, dagtr

Evidence triangulation

Integrate across data sources, populations, and assumptions rather than resting on a single point estimate. Quantitative bias analysis makes the residual uncertainty explicit; a meta-layer assesses whether the body of estimates actually triangulates.

→ biasetr, triangulatr

Design principles

One job per package. Each -tr package does one thing well; the suffix marks its place in the ecosystem.
A shared engine. causatr provides the estimation core (sandwich and bootstrap variance, contrasts); other packages compose with it rather than reimplementing it.
A common grammar. Fit, then contrast. The same two-step API recurs across packages, so moving between methods costs little.
Honest uncertainty. The point of triangulating is to surface disagreement, not to hide it.