philosophy.qmd

---
title: "Philosophy"
subtitle: "Triangulating causal evidence in epidemiology"
toc: true
toc-depth: 2
---

## 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](https://doi.org/10.1093/ije/dyw314)).
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.

## Further reading

- Lawlor DA, Tilling K, Davey Smith G. Triangulation in aetiological epidemiology.
  *International Journal of Epidemiology*. 2016;45(6):1866–1886.
- Hernán MA, Robins JM. *Causal Inference: What If*. Boca Raton: Chapman & Hall/CRC, 2025.