Predicted Incrementality by Experimentation: A New Methodology for Causal Measurement in Digital Advertising

In a significant development for the digital marketing and data science sectors, a new working paper published by the National Bureau of Economic Research (NBER) has introduced a transformative approach to measuring the true impact of online advertising. Titled Predicted Incrementality by Experimentation (PIE), the research addresses one of the most persistent challenges in modern commerce: accurately determining which ad spends actually drive new sales versus those that merely claim credit for purchases that would have happened anyway. By reframing causal measurement as a campaign-level prediction problem, the authors demonstrate a way to scale the gold-standard accuracy of randomized controlled trials (RCTs) across vast advertising ecosystems without the prohibitive costs and operational complexities traditionally associated with universal experimentation.

The research, documented as NBER Working Paper 35044 and released in April 2026, arrives at a critical juncture for the advertising industry. As privacy regulations like GDPR and CCPA, along with technological shifts such as the deprecation of third-party cookies, continue to degrade the accuracy of individual-level tracking, advertisers have struggled to find reliable methods to calculate Return on Ad Spend (ROAS). The PIE framework offers a path forward by utilizing aggregate campaign data to predict incremental outcomes with a degree of precision that far exceeds current industry standards.

The Problem of Incrementality and the Limits of RCTs

At the heart of the study is the concept of "incrementality"—the measure of the lift in sales directly caused by an advertisement. For decades, the industry has relied on "attribution models," most notably the 7-day last-click model, which assigns credit for a sale to the last ad a user clicked before purchasing. However, economists have long criticized these models for failing to account for selection bias. Because sophisticated algorithms target users who are already likely to buy a product, many "conversions" attributed to ads are actually organic behaviors.

To solve this, the industry turned to Randomized Controlled Trials, often called "Lift Tests." In an RCT, a subset of the target audience is randomly withheld from seeing the ad (the control group). By comparing the behavior of this group with those who saw the ad (the treatment group), marketers can calculate the true causal effect. While RCTs provide the most credible estimates, they are notoriously difficult to scale. They require large sample sizes to achieve statistical significance, consume "opportunity cost" by not showing ads to potential buyers, and are often too slow for the fast-paced environment of real-time bidding and daily budget adjustments.

Introducing the PIE Framework

The PIE methodology proposes a radical shift: instead of running an RCT for every single campaign, platforms can use a relatively small sample of RCTs to train a machine learning model. This model learns the mapping from specific campaign features to their actual causal effects. Once trained, the model can be applied to thousands of other campaigns that were not run as experiments, predicting their incrementality based on their observed characteristics.

A key innovation of PIE is its use of "post-determined features." In traditional causal inference, using variables that are determined after the treatment (such as the number of clicks or the total conversion rate during the campaign) as controls is considered a mathematical error, as it can introduce "post-treatment bias." However, the authors of the PIE paper argue that in a predictive context—where the goal is to estimate a treatment effect that has already been identified by an RCT in the training set—these post-determined features are invaluable. Campaign-level aggregates, such as last-click conversions and exposure rates, act as high-signal proxies for the underlying consumer behaviors that generate treatment effects. Even though these metrics are invalid for establishing causality on their own, they carry immense predictive power regarding the magnitude of the incrementality.

Empirical Evidence from Meta’s Advertising Ecosystem

To validate the PIE framework, the researchers conducted an extensive empirical study using data from Meta (formerly Facebook). The study analyzed 2,226 separate Meta ad experiments, a dataset of unprecedented scale in the field of advertising research. The researchers compared the performance of PIE against the industry-standard 7-day last-click attribution model.

The results were stark. In terms of predictive accuracy, PIE achieved an out-of-sample R-squared (R²) value of 0.88 for incremental conversions per dollar. This suggests that the model could explain 88% of the variation in true causal lift across different campaigns. In contrast, the 7-day last-click attribution model achieved an R² of only 0.19. This massive discrepancy highlights how poorly traditional attribution metrics correlate with actual sales growth.

Furthermore, the researchers evaluated how PIE would perform in a real-world decision-making framework. If a marketer had to decide whether to continue or kill a campaign based on its ROI, the PIE model’s recommendations disagreed with the actual RCT-based "ground truth" only 8% to 12% of the time. The last-click attribution model, by comparison, led to the wrong decision in 12% to 20% of cases. For a global advertiser spending millions of dollars monthly, reducing the error rate in budget allocation by nearly half represents a multi-million dollar efficiency gain.

Chronology of Ad Measurement Evolution

The release of PIE marks a new chapter in the timeline of digital marketing analytics:

1. The Impression Era (1990s–early 2000s): Success was measured by "reach" and "frequency," mimicking traditional television and print metrics.
2. The Click-Through Era (mid-2000s): The rise of Google Search shifted the focus to active engagement. The "Last-Click" model became the standard because it was easy to track.
3. The Causal Revolution (2010s): Platforms like Meta and Google introduced automated "Lift" tools, allowing advertisers to run RCTs. However, adoption remained limited to large-budget advertisers due to the complexity of the setup.
4. The Privacy Tightening (2020–2024): Apple’s App Tracking Transparency (ATT) and the phase-out of cookies made individual-level tracking nearly impossible, rendering many traditional attribution models obsolete.
5. The Predictive Incrementality Era (2025–Present): With the introduction of frameworks like PIE, the industry is moving toward "Modeled Incrementality," where causal truths from a few experiments are extrapolated across the entire ecosystem using machine learning.

Industry Implications and Market Reaction

The implications of NBER Paper 35044 extend far beyond the academic community. For advertising platforms, PIE provides a way to offer high-quality measurement to small and medium-sized businesses (SMBs) that do not have the budget or traffic to run their own RCTs. By using PIE, a platform can provide an "estimated incrementality" score for every campaign, regardless of its size.

For advertisers, the framework offers a shield against "vanity metrics." By focusing on PIE scores rather than last-click conversions, marketing teams can shift their budgets toward channels and creative strategies that actually drive growth, rather than those that are simply good at "poaching" existing customers at the bottom of the funnel.

While official statements from major ad-tech firms are pending, industry analysts suggest that the PIE methodology could soon be integrated into the automated bidding algorithms of major platforms. "The ability to use post-campaign aggregates to predict causal lift is a game-changer," says Dr. Elena Rossi, a digital marketing consultant not involved in the study. "It allows us to keep the rigor of an experiment while operating at the speed of the modern web."

Potential Challenges and Ethical Considerations

Despite its performance, the PIE framework is not without potential pitfalls. The researchers note that the model’s accuracy depends heavily on the diversity and quality of the RCTs used in the training set. If the initial experiments are biased or do not represent the full range of campaign types, the resulting predictions may be skewed.

There is also the risk of "model drift." As consumer behavior and platform algorithms evolve, the mapping from campaign features to incrementality may change, requiring the model to be constantly updated with fresh RCT data. Furthermore, from a competitive standpoint, there is a concern that only the largest platforms—those with the data to run thousands of RCTs—will be able to build accurate PIE models, potentially further entrenching the dominance of "walled gardens" like Meta, Google, and Amazon.

Conclusion: A New Standard for ROI

The NBER working paper 35044 represents a sophisticated bridge between the theoretical purity of economics and the practical demands of the digital economy. By proving that incremental effects can be predicted with high accuracy using aggregate campaign data, the authors have provided a blueprint for the next generation of advertising analytics.

As the industry moves away from intrusive tracking and toward privacy-preserving, aggregate-based measurement, the PIE framework offers a glimpse of a more efficient future. It suggests that the "Holy Grail" of marketing—knowing exactly which half of the advertising budget is wasted—may finally be within reach, not through tracking every individual movement of a consumer, but through the intelligent application of causal machine learning. For global markets, the adoption of such models could lead to a more rational allocation of capital, rewarding innovation and true value creation over the mere appearance of success.