Causal Inference

---
name: causal-inference
description: Estimate causal effects from observational and experimental data — DAGs, confounding control, quasi-experimental designs, and sensitivity analysis. Use when you need "what causes what," not just correlation.
category: ai-research
---

# Causal Inference

Correlation is cheap; causation is the question that matters for decisions. Causal inference is the 
toolkit for answering "what would happen if we did X?" — from experiments when possible, and from 
observational data with explicit assumptions when not.

## Overview

Every causal claim rests on assumptions — the craft is making them explicit and testing them. 
Draw the causal graph (DAG) to see what confounds what. Choose the identification strategy that 
fits: randomization, or quasi-experiments (difference-in-differences, regression discontinuity, 
instrumental variables), or adjustment with careful confounder selection. Then probe robustness: 
sensitivity analysis asks how wrong the assumptions would need to be to overturn the conclusion.

## When to use

- Deciding whether an intervention works: policy, product change, treatment.
- Estimating effects from data you didn't experimentally control.
- Reviewing causal claims: checking whether the design supports the conclusion.
- Choosing between running an experiment and analyzing existing data.

## Core concepts

- **Potential outcomes**: the causal effect is the difference between what happened with treatment 
and what would have happened without it. The fundamental problem: you never observe both. All 
methods are ways around this.
- **DAGs**: directed acyclic graphs mapping assumed causal structure — confounders (common 
causes), mediators (mechanisms), colliders (conditioning on them biases). Draw it before choosing 
an analysis.
- **Confounding control**: adjust for confounders (backdoor paths), never for mediators or 
colliders. The DAG tells you which variables to include — "control for everything" is wrong.
- **Quasi-experiments**: difference-in-differences (parallel trends assumption), regression 
discontinuity (threshold assignment), instrumental variables (exclusion restriction). Each trades 
randomization for a different assumption — name it, defend it.
- **Propensity methods**: matching, weighting, or doubly-robust estimation to balance treated and 
control groups on observables. Only as good as the observed confounders.
- **Sensitivity analysis**: how strong would unmeasured confounding need to be to explain away the 
effect? Report it; it's the honest answer to "but what if you missed something?"

## Practical workflow

1. Define the treatment, outcome, and the decision the estimate informs.
2. Draw the DAG: hypothesized confounders, mediators, colliders. Get domain experts to critique it.
3. Choose identification: experiment if feasible; otherwise the quasi-experimental design whose 
assumptions you can best defend.
4. Estimate with appropriate methods; use doubly-robust approaches where available.
5. Run sensitivity analyses and placebo/falsification tests (effects where none should exist).
6. Report the estimate with uncertainty, the assumptions it rests on, and what would overturn it.

```text
Causal analysis checklist:
[ ] Treatment, outcome, estimand precisely defined
[ ] DAG drawn and critiqued; adjustment set derived from it
[ ] Identification strategy named with its key assumption
[ ] No adjustment for mediators/colliders
[ ] Falsification tests run (placebos, pre-trends)
[ ] Sensitivity to unmeasured confounding reported
```

## Common pitfalls

- **Controlling for everything**: adjusting for mediators (kills the effect) or colliders (creates 
bias). The DAG decides, not the kitchen sink.
- **Unstated assumptions**: every observational method assumes something untestable. Name it 
explicitly.
- **No falsification**: never testing where the method should find nothing. Placebos are cheap 
credibility.
- **Causal language from predictive models**: a regression coefficient is not a causal effect 
without a causal design.
- **Ignoring the estimand**: "the effect" of what, on whom, compared to what? Define it before 
estimating.
- **Fragile identification**: an IV with a weak first stage or a DiD with shaky parallel trends. 
Weak designs produce confident-looking noise.