Causal Forests and Average Treatment Effects in Marketing Attribution

Causal Forests and Average Treatment Effects in Marketing Attribution

A causal forest is a machine learning method that estimates how the effect of a treatment (such as an ad exposure) varies across different subgroups of a population. The average treatment effect (ATE) is the mean difference in outcomes between the treated group and the control group across the entire population. Together, these concepts power the next generation of marketing attribution, moving beyond click tracking to measure what ads actually cause.

What Is an Average Treatment Effect?

The average treatment effect answers a deceptively simple question: on average, how much did this marketing campaign change the outcome compared to doing nothing?

Formally, ATE is defined as:

ATE = E[Y(1)] - E[Y(0)]

Where Y(1) is the outcome when a person is exposed to the ad, and Y(0) is the outcome when they are not. The fundamental problem is that you can never observe both for the same individual. This is what statisticians call the "fundamental problem of causal inference," and it is the reason marketing attribution is so difficult.

Traditional multi-touch attribution sidesteps this problem by ignoring it entirely. It tracks clicks and assigns credit based on rules (last-click, linear, time-decay) without ever asking whether the ad caused the conversion. A customer who would have purchased anyway gets attributed to whatever channel they touched last.

The ATE framework forces a different question: across all the people who saw this Meta ad, how many additional purchases occurred that would not have happened without the ad? That is the incremental impact, and it is the only number that should inform budget decisions.

From Average Effects to Heterogeneous Effects

The ATE gives you one number for an entire campaign. But marketing effects are rarely uniform. A Meta prospecting campaign might have a strong incremental impact on new-to-brand customers and near-zero impact on existing customers who would have purchased anyway. A Google Ads brand campaign might lift conversions among price-sensitive segments while being irrelevant to loyal buyers.

This is where heterogeneous treatment effects (HTE) become essential. HTE analysis breaks the average into subgroup-specific effects, answering: for whom does this campaign work, and for whom is it wasted?

Estimating HTEs accurately is exactly what causal forests were designed to do.

How Causal Forests Work

Causal forests, introduced by Athey and Wager (2018), extend the random forest algorithm to estimate treatment effects rather than predict outcomes. Here is how they work at a conceptual level:

Step 1: Honest Splitting

A standard random forest splits data to minimize prediction error. A causal forest splits data to maximize the difference in treatment effects between the two resulting groups. If customers in one leaf node show a 5% lift from ad exposure while customers in another show a 0.2% lift, the tree has discovered a meaningful distinction.

The "honest" part means the forest uses separate subsets of data for building the tree structure and estimating treatment effects within each leaf. This prevents overfitting and produces valid confidence intervals.

Step 2: Forest Aggregation

Like a standard random forest, a causal forest builds hundreds or thousands of trees on bootstrapped samples and averages their estimates. This reduces variance and produces stable treatment effect estimates for each individual or subgroup.

Step 3: Conditional Average Treatment Effects

The output is a conditional average treatment effect (CATE) for every observation. You can aggregate these CATEs by any dimension: by audience segment, by geographic region, by device type, by time of day. The average of all CATEs gives you the ATE. Subgroup averages reveal where the campaign truly drives value.

Why This Matters for Marketing Attribution

Consider a practical example. A fashion brand spends $200,000/month on Meta Ads across prospecting, retargeting, and brand campaigns. Traditional last-click attribution says retargeting has a 6x ROAS while prospecting has a 1.8x ROAS. The obvious move is to shift budget from prospecting to retargeting.

But a causal forest analysis tells a different story:

• Retargeting ATE: 0.3% incremental conversion lift. Most retargeted users were already going to buy. The high ROAS is an illusion created by targeting high-intent shoppers.
• Prospecting ATE: 2.1% incremental conversion lift among lookalike audiences in the 25-34 age bracket, but only 0.1% lift among 45+ audiences.

The actionable insight is not "spend more on retargeting." It is "increase prospecting spend on 25-34 lookalikes and cut prospecting spend on 45+ audiences." That level of precision is impossible with aggregate MMM alone and invisible to click-based attribution.

This is the methodology underlying Causality Engine's incrementality layer. By estimating heterogeneous treatment effects across campaigns, audiences, and creatives, it identifies exactly where marginal dollars create real value and where they are subsidizing conversions that would have happened anyway.

Causal Forests vs. Other Methods

Causal forests are particularly powerful in marketing contexts because:

1. They handle observational data. You do not need to run a controlled experiment to estimate effects, though experiments improve precision. When combined with techniques like propensity score matching, causal forests can produce credible estimates from the data you already have.

2. They scale to many covariates. Marketing data is high-dimensional: device, location, time, creative variant, audience segment, purchase history, weather, seasonality. Causal forests automatically discover which dimensions drive treatment effect heterogeneity.

3. They provide uncertainty estimates. Each CATE comes with a confidence interval, so you know whether a detected difference is statistically meaningful or just noise.

Practical Considerations

Causal forests are not magic. Several conditions must hold for the estimates to be trustworthy:

• Overlap: There must be variation in who gets treated and who does not. If 100% of a subgroup sees the ad, the forest cannot estimate what would have happened without it. This is related to the potential outcomes framework requirement of positivity.
• Unconfoundedness: The method assumes that, conditional on observed covariates, treatment assignment is as good as random. In marketing, this is approximated through rich covariate sets and validated through incrementality tests.
• Sample size: Estimating heterogeneous effects requires more data than estimating averages. Brands with fewer than 10,000 monthly conversions may need to rely on ATE estimates rather than granular CATEs.

How Causality Engine Uses Causal Forests

Causality Engine applies causal forest methodology as part of its causal inference stack. When you connect your Shopify store and ad accounts, the platform:

1. Builds a covariate set from transaction history, ad exposure data, and customer attributes
2. Estimates CATEs across campaigns and audience segments
3. Aggregates results into channel-level and campaign-level incremental ROAS metrics
4. Surfaces the budget reallocation moves that maximize incremental revenue

The result is attribution that tells you not just what happened, but what your ads actually caused. For brands comparing tools, see how this approach differs from Triple Whale and Northbeam, which rely on click-based or modeled attribution without a causal inference layer.

Start Measuring Causal Impact

If you are spending $50K or more per month on ads and making budget decisions based on platform-reported ROAS, you are almost certainly misallocating spend. Book a demo to see causal forest-powered attribution in action, or read our Shopify attribution guide for a deeper look at the methodology.