the etverse
An R ecosystem for triangulating causal evidence in epidemiology.
No single study design or analytic method is free of bias. The etverse gives each approach its own focused package, with a shared grammar, so that you can estimate the same effect several different ways — and trust it more when the answers converge.
Why triangulate?
Triangulation integrates results from approaches with different and unrelated sources of potential bias. When estimates from methods that would fail in different ways still point to the same answer, confidence in that answer grows. The etverse builds this idea into tooling across three levels:
Methodological
The same estimand, estimated several ways — g-computation, IPW, AIPW, g-estimation, matching — each leaning on different modelling assumptions.
Design
The same question, asked through different study designs — cohort, case-control, target-trial emulation, negative controls — each vulnerable to different biases.
Evidence
Findings integrated across data sources, populations, and assumptions, rather than resting on a single point estimate from a single dataset.
Packages
Every package ends in -tr — for triangulation. Each does one job well and composes with the others through a shared engine.
causatr available
Unified causal effect estimation: parametric g-formula and ICE g-computation, inverse-probability weighting, doubly-robust AIPW, structural nested mean models, and matching. The engine the rest of the etverse builds on.
survatr available
Causal survival analysis on person-period data: pooled-logistic hazard g-computation, IPW and ICE for time-to-event outcomes, risk-difference / RMST contrasts, competing risks, and IPCW.
matchatr in development
Classical and causal estimation for (matched) case-control, nested case-control, and case-cohort designs — conditional logistic and weighted Cox models, plus case-control–weighted marginal effects that compose with causatr and survatr.
negatr in development
Negative-control methods for detecting and correcting unmeasured confounding, selection bias, and measurement error — from simple null tests to proximal causal inference and empirical calibration.
biasetr planned
Quantitative bias analysis — modelling how much a finding could shift under plausible biases, for the general case not covered by the negative-control machinery in negatr.
separatr planned
Separable effects — decomposing a treatment effect into components acting through distinct mechanisms, for sharper, more mechanistic causal questions.
dagtr planned
Directed acyclic graphs for encoding causal assumptions and deriving adjustment sets to drive the estimators above.
emulatr planned
Target-trial emulation — turning an observational question into an explicit protocol and analysis.
triangulatr planned
The meta-package — collect estimates from across the etverse and assess whether they triangulate.
The etverse is early and evolving. Package scopes marked planned describe intended direction and may change.