Causal Inference from Competing Treatments

Ana-Andreea Stoica, Vivian Y. Nastl, Moritz Hardt

TL;DR

This paper introduces a novel framework that merges causal inference with game theory to solve a common problem in digital experimentation: how to estimate the "true" effect of a treatment when you are competing with others for a subject's attention. The authors prove that by strategically bidding to maximize a specific "sample value" metric, researchers can reach a Nash Equilibrium that minimizes statistical estimation error even in highly competitive environments.

What Problem Does This Paper Address?

In a perfect world, a Randomized Controlled Trial (RCT) gives a researcher full control over who receives a treatment. In the real world—especially in online advertising—you aren't alone. Multiple advertisers are competing for the same "slots" on a user's screen.

This competition creates a position effect: an ad shown at the top of the page (Rank 1) is significantly more effective than the same ad shown at the bottom (Rank 10). If an advertiser wants to know the "true" causal effect of their ad (defined as the effect if it were shown in the top rank), they face two interconnected hurdles:

  1. Statistical: How do you combine noisy data from different ranks to get one accurate estimate?
  2. Game-Theoretic: How should you bid on users, knowing that other advertisers are also bidding strategically to improve their own estimates?

Standard causal inference tools usually ignore the competition, while standard game theory (like auction theory) usually assumes agents want to maximize clicks or revenue, not the accuracy of a scientific estimate.

Methods

The authors build a game-theoretic model where $k$ "admins" (advertisers) compete for $n$ subject slots.

  • The Decay Model: They assume the treatment effect decays by a known factor $\alpha_r$ as the rank $r$ increases.
  • The Objective: Unlike traditional auctions where the goal is profit, the "utility" here is the negative of the Mean Squared Error (MSE). In other words, an admin "wins" if they get the most accurate estimate of the causal effect.
  • The "Sample Value" Proxy: Because minimizing MSE is mathematically difficult to solve in a game, the authors introduce a tractable proxy called Sample Value ($S$), defined as the weighted sum of samples across ranks: $S = \sum n_r \alpha_r^2$.
  • Theoretical Analysis: They use Le Cam’s method to establish a minimax lower bound on estimation error and then prove that maximizing Sample Value is a near-perfect proxy for minimizing error.

Datasets

This is a theoretical paper and does not utilize empirical datasets. The authors use a stylized "online advertising" scenario as a running example to ground their mathematical proofs.

Key Insights and Ideas

  • The Bridge: The paper’s biggest "aha moment" is the formal link between statistical precision and bidding strategy. It treats "information" as the currency being auctioned.
  • Sample Value as Effective Sample Size: The term $\sum n_r \alpha_r^2$ essentially represents the "effective" number of Rank-1 samples an advertiser has collected. This insight simplifies a complex statistical problem into a straightforward resource allocation problem.
  • Avoiding the "Bidding War": The authors find that in equilibrium, rational agents actually prefer to avoid competition. Instead of everyone fighting for the top spot on a few users, it is often better to spread a budget thinly across many users to ensure a steady stream of "cleaner" (less contested) data.
  • Equilibrium Robustness: They demonstrate that a Nash Equilibrium (a state where no one wants to change their strategy) in the simpler "Sample Value" game is also an approximate Nash Equilibrium in the much more complex "Estimation Error" game.

Key Results

  • Minimax Optimality: The authors prove that the best possible estimation error any admin can achieve is proportional to $1/S$.
  • Existence of Pure Nash Equilibrium: They prove that a pure Nash Equilibrium always exists for the Sample Value game. In this state, admins split their budgets uniformly and maximize their "reach" while minimizing overlap with competitors.
  • Convergence: As budgets grow larger relative to the number of competitors ($B = \omega(k^2)$), the gap between the tractable proxy and the true error objective vanishes, leading to a near-perfect (o(1), o(1))-approximate Nash Equilibrium.

Limitations and Future Work

  • Known Decay Factors: The current model assumes the decay parameters ($\alpha_r$) are known. While the authors discuss how these might be estimated from historical data, the game theory assumes they are fixed constants.
  • Homogeneous Subjects: The model treats all subjects as identical placeholders. In reality, some users are more "valuable" for an experiment than others (e.g., they are more likely to click).
  • Simple Auction Rules: The paper uses a specific probabilistic allocation rule. Future work could explore how these results hold up under more common auction formats like Generalized Second Price (GSP) or VCG auctions.
  • Price of Anarchy: The authors suggest that future research should investigate the "Price of Anarchy"—how much total estimation accuracy is lost across the whole system because admins are acting selfishly rather than cooperating.

Summary updated

BibTeX
@inproceedings{stoica2024causal,
  author = {Ana-Andreea Stoica and Vivian Y. Nastl and Moritz Hardt},
  title = {{Causal Inference from Competing Treatments}},
  booktitle = {International Conference on Machine Learning},
  year = {2024},
  url = {https://openreview.net/forum?id=HZ6lrZzB02},
}
Figure