How I frame causal questions

A simple checklist for turning vague research questions into answerable causal questions: intervention, outcome, population, time, and counterfactual.
Published

February 5, 2026

A vague question produces vague evidence.

The easiest way to improve a causal analysis is often to improve the question.

Here is the framing I use when I want a causal question that is actually answerable.

The five-part causal question

A well-formed causal question usually specifies:

  1. Intervention: what is being changed?
  2. Outcome: what is being measured?
  3. Population: for whom?
  4. Time window: when do we expect effects?
  5. Counterfactual: compared to what?

Let’s walk through each part.

1) Intervention: what changes?

Not “exposure” in the abstract — a change you could imagine implementing.

Bad: - “Does education affect income?”

Better: - “What is the effect of completing secondary school (vs not completing) on income?”

Even better: - “What is the effect of a policy that increases secondary school completion by X% on income at age 30?”

You don’t always need the last level of detail, but you do need something that implies a real comparison.

2) Outcome: what exactly is measured?

Outcomes are often underspecified.

Bad: - “Does the program reduce poverty?”

Better: - “Does the program reduce the probability of being below the poverty line?”

Even better: - “Does the program reduce the probability of being below the poverty line 12 months after enrollment?”

Also ask: is the outcome measured consistently, and does measurement itself change after the intervention?

3) Population: for whom?

Effects are not universal constants.

Bad: - “Does remote work increase productivity?”

Better: - “Among software developers in mid-sized firms, does remote work increase measured output?”

Sometimes the biggest difference between studies is not methods — it’s population.

4) Time window: when do we expect effects?

Effects vary over time. Many arguments silently assume “immediate and permanent.” That’s rarely true.

Ask: - What is the short-run effect? - What is the long-run effect? - Are there adaptation dynamics?

In policy settings, time is a major confounder: trends, shocks, and seasonality can look like treatment effects.

5) Counterfactual: compared to what?

This is the heart of causal inference.

Every causal estimate is a comparison between: - what happened under the intervention - what would have happened without it

So you must define the “without it.”

Examples: - no policy change - alternative policy - delayed adoption - business-as-usual care

A surprising number of disagreements are really disagreements about counterfactuals.

The “estimand sentence”

When I want to be precise, I write a one-sentence estimand:

“Among [population], what is the effect of [intervention] versus [comparison] on [outcome] measured over [time window]?”

If you can’t write this sentence, your study design will likely be unstable.

A worked example

Vague: - “Do nutrition protocols improve ICU outcomes?”

Framed: > “Among adult ICU patients, what is the effect of implementing an early enteral nutrition protocol (vs standard practice) on 28-day mortality and ICU length of stay, measured during the admission and up to 28 days after ICU entry?”

Now you can design: - what counts as protocol exposure - what comparison is fair - what timing matters - what confounding is plausible

Why this matters

Framing does three things immediately:

  1. It forces you to be honest about what you can claim.
  2. It tells you what data you need.
  3. It makes your identification strategy easier to evaluate.

The fastest way to improve your causal inference is to treat question framing as a first-class step, not a formality.

If you want, send me a question you’re working on and I’ll rewrite it into a well-formed causal question plus 2–3 plausible counterfactuals.