What is causal inference?
Most marketing data shows what moved together. Spend rose and sales rose. People who saw the ad bought more. Causal inference asks the harder question: would those sales have happened anyway?
The answer rests on a counterfactual, the outcome that would have happened without the marketing. Nobody can observe it directly, so every causal method is a way of estimating it from something that can be observed.
What are the main causal inference methods?
| Method | How it estimates the counterfactual | Key assumption |
|---|---|---|
| Randomized test | A random group doesn't get the marketing | The groups were truly random and stayed separate |
| Difference-in-differences | Compares the change in treated areas with the change in similar untreated ones | Both would have moved in parallel without the campaign |
| Synthetic control | Builds a weighted blend of untreated areas that tracked the treated one before launch | That blend would have kept tracking it |
| Propensity score matching | Compares people who saw the ad with similar people who didn't | Every factor that drove exposure was measured |
| Regression discontinuity | Compares people just above and just below a cutoff, such as a spend threshold | People near the cutoff are otherwise alike |
| Instrumental variables | Uses something that changed exposure but not buying directly | The instrument affects sales only through exposure |
Most of these estimate an average treatment effect: how much the marketing changed the outcome, on average, for the people or places it reached.
What is an example of causal inference?
Retargeting that looked like magic
A retargeting campaign shows a conversion rate five times higher than prospecting. The team plans to move budget into it.
A causal look says to wait. Retargeting reaches people who already visited the site, and they were far more likely to buy anyway. That's selection bias. A holdout test that withheld ads from a random slice of visitors found the true lift was a small fraction of the reported one.
Correlation vs. causation in marketing
| Correlation | Causal estimate | |
|---|---|---|
| The question | Did these move together? | Did this make that happen? |
| Typical source | Dashboards and attribution | Tests and causal models |
| Common trap | Confounding and selection bias | Assumptions that don't hold |
| Safe to move budget on | Rarely alone | Yes, within its limits |
Why does causal inference matter?
Budget decisions are causal decisions. Moving money assumes the channel caused the sales it's credited with, and attribution can't tell the difference between causing a sale and being present for one.
- Confounding. Something else drove both spend and sales, such as a season or a promotion.
- Selection bias. The audience reached was already more likely to convert.
- Reverse causality. Sales drove spend, as when budgets follow demand.
When a randomized test is possible, it's usually the cleanest answer. When it isn't, a quasi-experiment with written-down assumptions beats a correlation presented as an effect.
How does Alli do this?
Alli by PMG's Scenario Planner pairs a media response model built on historical sales with a brand equity model built on current signals. Two-stage estimation separates the direct effect of marketing from the accumulated value of brand, so teams can see what media drove directly and what brand strength contributed.
Forecasts are modeled estimates with their assumptions in view, so teams can compare options before money moves and keep their own tests alongside the model.
Explore Alli Scenario PlannerCausal inference FAQ
What is causal inference in simple terms?
It's working out what caused an outcome by comparing what happened with a credible estimate of what would have happened otherwise.
Why isn't attribution causal?
Attribution assigns credit to the touchpoints present before a sale. It can't tell whether the sale would have happened without them, so it often credits channels that reach people who were already going to buy.
What is selection bias in marketing?
It's when the people who get the marketing differ from those who don't in ways that affect buying. Retargeting audiences, loyal customers, and high-intent searchers are common examples.
Do you need a randomized test for causal inference?
Not necessarily. Methods like difference-in-differences and synthetic control can estimate effects without randomization, but they rest on assumptions a randomized test doesn't need.
Last updated Back to the glossary