Causal-driven attribution (CDA): Estimating channel influence without user-level data

AI summary

Overview

Research area: Marketing attribution and causal inference for digital advertising (stat.ML), sitting at the intersection of causal discovery, structural causal modelling, and privacy-preserving analytics.

Technical level: Advanced. The paper assumes familiarity with directed acyclic graphs, structural causal models, and time-series causal discovery (PCMCI).

Scope (1 sentence): The paper proposes and evaluates on synthetic data a Causal-Driven Attribution (CDA) framework that estimates directional channel influence on conversions using only aggregated impression-level time series, without user identifiers or click-path tracking.

What This Paper Is About

Marketers need to know which advertising channels actually cause conversions, but most attribution methods require user-level path data that has become largely unavailable because of privacy regulation (GDPR, CCPA, Apple's iOS14 changes) and platform restrictions. The authors describe this as a "privacy-attribution paradox": the need for sophisticated attribution grows at the same time that data accessibility shrinks. The paper's goal is to recover both the direction of causal influence between channels and the magnitude of each channel's contribution to conversions using only aggregated impression-level data.

Key Contributions

1. A theoretical approach integrating Structural Equation Modelling (SEM) into a temporal causal discovery framework for simulating marketing attribution.
2. A novel Causal-Driven Attribution (CDA) framework that operates entirely on aggregated impression-level data, running without user identifiers or click-path tracking.
3. A synthetic data generation process built on causal DAGs with baseline impressions, growth rates, conversion rates, random noise, and cross-channel influence terms, created specifically so that a ground-truth causal graph and effect values exist for validation.
4. A demonstration that directional channel relationships can be recovered from observational time series alone using PCMCI, with quantified accuracy under both a known and a predicted causal graph.

The authors state that to the best of their knowledge, no existing model can fully reveal causal relationships among touchpoints, including the direction of influence between channels and their impact values, using only aggregated data.

Main Findings

• Accuracy with the true causal graph: CDA achieves an average relative RMSE of 9.50% when supplied with the true causal graph.
• Accuracy with the predicted causal graph: When the causal structure is instead discovered from the data, CDA achieves an average relative RMSE of 24.23%. The authors describe this as meaningful signal recovery even under structural uncertainty.
• Cross-channel interdependencies are captured: The framework recovers relationships between channels, not only channel-to-conversion effects, because cross-influence terms make channels act as confounding variables — which the authors describe as essential for causal inference.
• Privacy-preserving attribution is feasible: The model produces interpretable attribution insights without any reliance on user identifiers or click-path tracking.
• No suitable real dataset was found: The authors report that after nearly two years of exploration, they could not find a dataset enabling comparison between traditional multi-touch attribution and their approach, and that all existing publications on advertising and causal learning rely on only two publicly available datasets, neither of which meets the requirements of this work.
• Note on scope of evidence: The evaluation reported in the provided content is based on large-scale synthetic data designed to replicate real marketing dynamics; real-world empirical validation is not reported.

Methodology in Plain English

The authors could not find real marketing data with a known ground truth, so they built a simulator instead, and then tested their method on the simulated data.

Step 1 — Build a causal map. They start from a Directed Acyclic Graph: nodes are marketing channels (the example DAG uses Affiliates, Facebook, TikTok, Google Ads, YouTube) and the conversion outcome; arrows point in the direction of causation and no loops are allowed. Nodes are grouped into causal layers labelled S³ (Affiliates), S² (Facebook, TikTok, Google Ads), S¹ (YouTube), and S⁰ (conversion). Green arrows in their figures indicate positive weights (putative amplification/halo) and red arrows indicate negative weights (putative substitution/crowd-out).

Step 2 — Turn the map into equations. Each channel's impressions follow a linear trend with a baseline, a growth rate, and random noise. Channels are then allowed to influence each other concurrently (no lag) through weight terms w_ij, so that channel i's activity depends on its own trend plus the current activity of the other channels. Conversions per channel are modelled as a linear function of channel impressions, and total conversions equal baseline (non-marketing) conversions plus the sum of channel-driven conversions.

Step 3 — Simulate. User-defined inputs specify the channels and their activities, the time period (for example 365 days), a baseline conversion range, a per-channel conversion-rate range, and a parent–child influence weight range. Outputs are daily or weekly impression and conversion time series plus the causal graph with impact values. The dataset schema is 365 daily observations across n channels, with an index, total conversions, and one impression column per channel.

Step 4 — Discover the causal structure from data. In reality businesses only have time series. Instead of expensive expert-driven lift tests and A/B experiments, the authors apply PCMCI, an automated causal discovery procedure based on graphical causal modelling and conditional independence. PCMCI first runs a condition-selection step to identify relevant conditioning sets for each variable from lagged candidates, removing variables that are conditionally independent of the target. The authors note PCMCI scales well to large time series and handles linear, nonlinear, and time-delayed dependencies.

Step 5 — Estimate effects. A Structural Causal Model then quantifies how much each channel contributes to conversions, giving both direction and magnitude.

Why This Matters

Impact on research. The work reframes attribution from a path-reconstruction problem to a causal-structure-recovery problem, and it argues that the field needs a new family of attribution methods built on aggregate rather than individual data. It also contributes a reusable simulation framework with ground truth, addressing a documented scarcity of suitable public advertising datasets. The paper distinguishes itself from prior aggregate-data work such as Kireyev et al. (2016), which combined time-series analysis with Granger causality — the authors note that Granger causality is not true causal inference, since it only shows that past values of one series improve prediction of another.

Real-world applications:

• Budget allocation under privacy restrictions: marketers can estimate incremental channel contributions when persistent identifiers are unavailable.
• Impression-first platforms: channels such as short-form video, influencers, connected TV, and impression-first entertainment platforms that do not generate clicks can be measured, since the method uses impressions rather than click paths.
• Measurement for smaller or emerging channels: the paper notes that existing frameworks break down for smaller or emerging channels that do not provide persistent user identifiers.
• Cost-sensitive advertisers: the method is described as computationally lightweight and accessible to advertisers without complex data engineering, unlike heavy API integrations or deep learning models with large computational requirements.

Industry relevance. The paper is motivated by practitioner realities: media platforms each track only their own touchpoints, creating what the authors call structural opacity; last-click attribution became the industry default despite mounting criticism; and prior work (Reiley et al., 2012) found that click-based models misattribute impact by crediting clicks from users who would have converted anyway. The authors also cite evidence that impressions without website visits can generate sales, and that media exposure can increase organic branded searches that never appear in click sequences.

Future Directions

• Validation on real marketing data. The study relies on synthetic data because no suitable dataset with ground truth was found; testing whether CDA transfers to real settings is the central open question, especially given the authors' own caveat that real marketing data cannot be perfectly simulated because awareness and behaviour involve psychological factors.
• Reducing sensitivity to causal discovery error. The gap between 9.50% relative RMSE with the true graph and 24.23% with the predicted graph points to improving the discovered structure as a direct path to better attribution.
• Comparing against traditional multi-touch attribution. The authors originally aimed to compare traditional MTA methods with CDA but could not obtain data permitting that comparison, leaving the benchmark open.
• Extending the simulation assumptions. The current generator uses concurrent (no-lag) cross-channel effects, linear trends, and acyclic structure; whether these assumptions can be relaxed while preserving identifiability is not resolved in the content provided.

Target Audience

This paper is most useful to quantitative marketing researchers and causal inference methodologists, data scientists and analytics leaders in advertising or growth teams who must measure channel performance under privacy constraints, and academic readers interested in applying temporal causal discovery (PCMCI) and structural causal models to aggregated marketing time series. Practitioners looking for a drop-in deployment recipe may find it advanced, since it presumes comfort with DAGs and structural causal modelling.

Note on completeness: The provided paper content is truncated partway through the PCMCI description in Section 3.2.1. The detailed results, model validation, and discussion sections referenced in the paper's outline are not included in the content supplied, so only the headline accuracy figures stated in the abstract are reported above. Any benchmark comparisons, validation details, or secondary metrics that appear in those later sections are not reported here because they are not present in the available content.

Authors’ abstract

Attribution modelling lies at the heart of marketing effectiveness, yet most existing approaches depend on user-level path data, which are increasingly inaccessible due to privacy regulations and platform restrictions. This paper introduces a Causal-Driven Attribution (CDA) framework that infers channel influence using only aggregated impression-level data, avoiding any reliance on user identifiers or click-path tracking. CDA integrates temporal causal discovery (using PCMCI) with causal effect estimation via a Structural Causal Model to recover directional channel relationships and quantify their contributions to conversions. Using large-scale synthetic data designed to replicate real marketing dynamics, we show that CDA achieves an average relative RMSE of 9.50% when given the true causal graph, and 24.23% when using the predicted graph, demonstrating strong accuracy under correct structure and meaningful signal recovery even under structural uncertainty. CDA captures cross-channel interdependencies while providing interpretable, privacy-preserving attribution insights, offering a scalable and future-proof alternative to traditional path-based models.