Causal Inference for Marketing: A Light Introduction

Most marketing measurement asks whether spend and sales move together. Causal inference asks something harder and more useful: if we change spend, what happens to sales? That gap between association and intervention is where a lot of budget decisions go wrong.

The Question Behind Every Measurement Debate

When a dashboard shows that weeks with heavier paid social also had higher revenue, it is tempting to call that proof. It is not. Those weeks may have been peak season. Brand interest may have been rising already. Search demand may have pulled both clicks and sales upward. The chart shows a relationship. It does not tell you what would have happened if you had spent less.

That "what would have happened" question is the heart of causal inference. In marketing terms, it is the difference between:

• Association: people who saw our ads converted more
• Causation: our ads caused conversions that would not have happened otherwise

Attribution systems, platform ROAS, and naive regressions are very good at the first. Budget decisions need the second. Causal graphs make the difference visible: an arrow means "influences," and the path we care about is the one from the intervention to the outcome.

Counterfactuals: The Missing Half of the Story

A causal effect is always a comparison between two worlds: the world where you ran the campaign, and the world where you did not. You only ever observe one of those. The unobserved world is the counterfactual.

Incrementality is not "how many conversions happened after someone saw an ad." It is "how many conversions would have been lost if the ad never ran."

Lift tests make this concrete by creating a control group that approximates the no-ad world. Observational methods like MMM try to reconstruct that world with models and assumptions. Different tools, same underlying question.

Be Precise About What You Want to Estimate

Causal work starts before the model. It starts by naming the estimand: the quantity you actually care about. In marketing, common estimands include:

• Average effect of a campaign: how much did this flight lift sales across the market?
• Effect of a budget change: what happens if we move £500k from TV to paid search?
• Effect for a segment: does retargeting lift new customers, or mostly people who would have bought anyway?

These sound similar, but they are not interchangeable. A model that answers "how correlated is spend with sales?" can still fail to answer "should we spend more?" Clarity on the estimand stops teams from fitting a clever model to the wrong question.

Confounding: Why Smart People Get Fooled

The classic threat to causal claims is confounding: a third factor that influences both the marketing decision and the outcome. In a DAG, that looks like a fork: one common cause pointing into both treatment and outcome.

Familiar marketing examples:

• Seasonality: Christmas drives both media investment and sales
• Intent: high-intent users are more likely to see retargeting and more likely to convert
• Brand strength: strong brands attract more organic demand and often spend more on media
• Promotions: price cuts and advertising often land in the same weeks

Retargeting makes the same structure painfully concrete. Purchase intent sends people into the retargeting pool and makes them more likely to buy. Platform dashboards often read that correlation as ROAS.

If you ignore those common causes, the model may credit marketing for outcomes it did not create, or miss effects that were real but masked. This is why "just regress sales on spend" is rarely a causal strategy. You need a story about what else is going on, and a plan for dealing with it.

Causal Graphs as a Thinking Tool

You do not need heavy maths to start thinking causally. Most measurement debates reduce to a few graph patterns:

• Forks create confounding (seasonality, intent)
• Chains carry mediated effects (TV → awareness → sales)
• What you condition on changes which paths you block or open

Mediation is the trap that catches careful analysts. Brand awareness may sit on the path from TV to sales. That does not make awareness a confounder. It makes it a mediator. If your estimand is the total effect of TV, conditioning on awareness can remove the very mechanism you care about.

The graph makes those choices visible before you argue about coefficients. The practical payoff is discipline: instead of asking "which variables improve model fit?", ask "which variables do we need to adjust for to isolate the effect of this intervention?"

What a lift test changes in the graph

Randomisation is powerful because it cuts the arrows into treatment. In a holdout test, ad exposure is assigned by design, so seasonality and intent no longer decide who gets treated. The backdoor closes, and the contrast between test and control becomes a credible estimate of the causal effect.

Identification Comes Before Estimation

A useful habit from causal inference is to separate three steps:

1. Define the estimand, what effect do we want?
2. Justify identification, under what assumptions can data answer that question?
3. Choose an estimator, experiment, difference-in-differences, regression adjustment, MMM, and so on

Teams often jump straight to step three. That is how you end up with beautiful uncertainty intervals around the wrong quantity. If the identifying assumptions fail, better software will not save you.

How This Shows Up in Everyday Measurement

Once you see the causal question clearly, familiar tools rearrange themselves:

• Lift tests create the counterfactual through randomisation. They are the cleanest answer when you can afford to withhold media.
• MMM estimates the effect of media under assumptions about confounders, carryover, and saturation. It is causal only to the extent those assumptions hold.
• Attribution describes journeys and credit allocation. It can be useful for operational optimisation, but it is usually not a substitute for causal incrementality.
• Platform ROAS often mixes selection effects with true lift. High ROAS can mean efficient targeting of people who were going to convert anyway.

The strongest measurement systems combine these deliberately: use experiments where you need ground truth, use models to generalise across channels and budgets, and treat platform metrics as inputs rather than final answers.

A Practical Starting Point

You do not need a full causal programme to start improving decisions. Begin with three questions every time someone presents a ROI number:

1. What intervention does this estimate correspond to?
2. What is the implied counterfactual?
3. What confounders could make association look like causation?

Those questions alone raise the quality of measurement conversations. They also make it clearer when you need a lift test, when an observational model is good enough, and when a dashboard number should not drive a million-pound reallocation.

Causal inference is not a replacement for marketing judgement. It is a way of making that judgement honest about evidence. In a world where every platform wants credit for every conversion, that honesty is a competitive advantage.