The Potential Outcomes Framework: How It Applies to Marketing Attribution

The Potential Outcomes Framework: How It Applies to Marketing Attribution

The potential outcomes framework (also called the Rubin Causal Model) defines a causal effect as the difference between two states: the outcome when a person receives a treatment, and the outcome when they do not. In marketing, "treatment" means ad exposure, and the framework answers whether showing someone an ad actually changed their behavior or whether they would have converted regardless.

This is the statistical backbone behind every credible incrementality measurement, and understanding it explains why most marketing attribution tools get causation wrong.

The Fundamental Problem

For every customer, two potential outcomes exist:

• Y(1): What happens if the customer sees your ad (the treated outcome)
• Y(0): What happens if the customer does not see your ad (the control outcome)

The causal effect for that individual is Y(1) - Y(0).

Here is the problem: you can only ever observe one of these outcomes. If a customer saw your Meta retargeting ad and then purchased, you know Y(1) = purchase. But you will never know Y(0) for that specific person. Would they have bought anyway after abandoning their cart? You cannot rewind time and find out.

This is called the fundamental problem of causal inference, and it is not a technical limitation that better tracking can solve. It is a logical impossibility. No pixel, no cookie, no server-side tracking integration can tell you what would have happened in the alternative scenario.

Multi-touch attribution pretends this problem does not exist. It observes that the customer saw the ad and purchased, and assigns credit. The potential outcomes framework forces you to acknowledge that observation alone is not causation.

How the Framework Connects to Marketing

Let us translate the abstract notation into concrete marketing terms.

Scenario: Meta Prospecting Campaign

You run a Meta prospecting campaign targeting lookalike audiences. In one week, 10,000 people in your target audience see the ad, and 200 of them purchase.

What traditional attribution reports: 200 conversions, $30,000 in revenue, 3x ROAS.

What the potential outcomes framework asks:

• Of those 200 purchasers, how many had Y(1) = purchase AND Y(0) = purchase? These people would have bought regardless. The ad did not cause their conversion. They might have discovered your brand through organic search, a friend's recommendation, or a previous touchpoint.
• How many had Y(1) = purchase AND Y(0) = no purchase? Only these represent the ad's true causal effect. This is incremental revenue.

If 60 of the 200 purchasers would have bought anyway, the true incremental revenue is not $30,000 but $21,000. The true ROAS is 2.1x, not 3x. That 30% gap between reported and real ROAS is money being misattributed, and it compounds across every channel and campaign.

The Three Assumptions That Make Estimation Possible

Since we cannot observe both potential outcomes for the same individual, the framework relies on three assumptions to estimate causal effects from data:

1. SUTVA (Stable Unit Treatment Value Assumption)

One person's treatment does not affect another person's outcome. In marketing, this is approximately true: whether Customer A sees your ad does not directly change whether Customer B purchases. (Network effects and word-of-mouth are exceptions, but the assumption holds well enough for most paid media analysis.)

2. Positivity (Overlap)

Every individual must have a non-zero probability of being in both the treatment and control groups. If 100% of visitors to your site see your retargeting ad, there is no control group, and you cannot estimate the counterfactual. This is why holdout testing and geo-lift experiments are valuable: they deliberately create control groups.

3. Ignorability (Unconfoundedness)

Conditional on observed characteristics, treatment assignment is independent of potential outcomes. In plain English: after accounting for everything you know about a customer (their browsing history, demographics, purchase history), whether they saw the ad is unrelated to their inherent likelihood of purchasing.

This assumption is the hardest to satisfy in marketing. Platforms like Meta and Google Ads deliberately target people who are likely to convert, which means treatment assignment is correlated with potential outcomes. Smart causal inference methods address this through techniques like propensity score adjustment, instrumental variables, and regression discontinuity designs.

From Theory to Practice: Estimation Methods

The potential outcomes framework is the conceptual foundation. The actual estimation happens through specific statistical methods, each suited to different data conditions:

Randomized Experiments

The gold standard. Randomly assign people to treatment (see ad) and control (do not see ad), then compare average outcomes. The randomization ensures ignorability holds by construction.

In e-commerce, this takes the form of:

• Geo-lift tests: Turn off ads in randomly selected geographic regions
• Holdout experiments: Withhold ads from a random subset of users
• Conversion lift studies: Platform-run experiments (Meta, Google) that measure incremental lift

Observational Methods

When experiments are not feasible, observational methods estimate potential outcomes from non-experimental data:

• Propensity score matching: Match treated individuals with untreated individuals who have similar characteristics, then compare outcomes.
• Inverse probability weighting: Reweight the data so that the treated and control groups become comparable.
• Difference-in-differences: Compare changes in outcomes before and after a campaign, relative to a control group that did not receive the campaign.
• Causal forests: Machine learning methods that estimate heterogeneous treatment effects across subgroups. See our guide to causal forests and ATEs.

Model-Based Approaches

• Bayesian structural time-series models: Estimate what revenue would have been without a campaign using historical patterns and external predictors. Google's CausalImpact package is a well-known implementation.
• Marketing mix modeling: Regress revenue on marketing inputs to estimate each channel's contribution. Modern Bayesian MMM produces posterior distributions over potential outcomes.

What This Means for Your Attribution Stack

If your current attribution tool does not reason about potential outcomes, it is not measuring causation. Here is a diagnostic:

Tools like Triple Whale and Northbeam operate in the right column. They track touchpoints and model attribution paths, but they do not estimate what would have happened without the ad. That makes them fundamentally unable to identify wasted spend.

Causality Engine operates in the left column. Its methodology is built on the potential outcomes framework, using a combination of Bayesian causal models, incrementality testing, and observational causal inference to estimate the counterfactual for every campaign.

Why This Matters for E-commerce Brands

For beauty brands, fashion brands, supplements brands, and other DTC verticals on Shopify, the practical implication is straightforward: the potential outcomes framework is the reason your true ROAS is almost certainly different from what your dashboards show.

Brands that adopt causal attribution consistently find:

• 30-40% of ad spend is allocated to campaigns with near-zero incremental impact
• Retargeting ROAS is overstated by 3-5x because it targets customers who were already going to buy
• Prospecting campaigns are often undervalued because they create demand that other channels later claim credit for

The potential outcomes framework does not just improve measurement. It changes which decisions you make, and those decisions compound into millions of dollars over time.

Apply the Framework to Your Data

Understanding potential outcomes is the first step. Applying it to your own campaigns is where the value lives. Book a demo to see how Causality Engine estimates the counterfactual for every campaign in your stack, or explore our Shopify attribution guide for a hands-on walkthrough of causal attribution methodology.

Ready to measure what your ads actually cause? Start your free trial and get incremental ROAS numbers within 48 hours.