Title: Enhancing Incrementality Measurement in Advertising Amid Privacy-Induced Signal Degradation

Abstract:

While advertising networks rely on randomized lift tests to quantify incrementality, the implementation of privacy-preserving reporting mechanisms often compromises data integrity. These systems introduce various forms of signal degradation, including match-rate attrition, linkability deficits, attribution-window constraints, aggregation-threshold filtering, randomized reporting noise, and signal loss that varies across segments. This study addresses these challenges by framing privacy-constrained advertising measurement as a robust causal decision-making problem.

The proposed framework operates by taking a randomized experiment and an ambiguity set representing privacy-induced degradation, then projecting the fiber of clean, unfiltered experimental scenarios that remain compatible with observations onto an incrementality functional. This process yields decisions that are certified, rejected, or left unresolved. The core contribution is a precise decision frontier. Observations falling outside this boundary allow for uniformly valid certification or rejection of incrementality. Conversely, data points within the frontier lack sufficient information for any methodology to uniformly distinguish between incrementality that exceeds a threshold and that which does not.

Additional theoretical findings include guarantees for finite-sample certification, bounds on sample complexity, and a minimax lower bound demonstrating that signal loss diminishes effective information. The study also explores the tradeoff between reporting granularity and performance. Empirical validation using 2.0 million rows from the Criteo Uplift dataset and 64,000 rows from the Hillstrom email experiment confirms that clean conversion lift is positive in both cases, with values of 0.00112 and 0.00495, respectively.

The analysis reveals that population-level certification remains viable under mild degradation in the Criteo data and even severe degradation in the Hillstrom data. However, when accounting for simultaneous uncertainty and reporting noise, all finite-sample stress tests across both datasets result in unresolved outcomes. Ultimately, this research establishes a decision-theoretic layer for privacy-conscious incrementality measurement, delivering the most robust causal claims possible given the constraints of degraded advertising signals.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC