📈 economics

Detecting and Mitigating Group Bias in Heterogeneous Treatment Effects

This paper introduces a unified statistical framework to detect and mitigate systematic group bias that arises when aggregating machine learning-predicted individual treatment effects, offering closed-form correction methods and analyzing their economic impact on profit-maximizing targeting strategies.

Original authors: Joel Persson, Jurriën Bakker, Dennis Bohle, Stefan Feuerriegel, Florian von Wangenheim

Published 2026-02-25

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Joel Persson, Jurriën Bakker, Dennis Bohle, Stefan Feuerriegel, Florian von Wangenheim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Average" Trap

Imagine you are a chef running a massive restaurant chain. You have a super-smart AI robot (a Machine Learning model) that predicts exactly how much each individual customer will enjoy a new spicy sauce.

The Robot's Job: It looks at every single person's taste buds, diet, and mood to say, "Alice will love this sauce (+10 points), but Bob will hate it (-5 points)." This is called a Heterogeneous Treatment Effect (HTE). It's highly personalized.
The Boss's Job: The restaurant manager doesn't care about Alice or Bob individually. They need to know: "How does this sauce perform in the New York branch? How about in the London branch?" They need to group the data to make business decisions.

The Problem: The paper argues that if you simply take the robot's individual predictions for Alice, Bob, and everyone else in New York, and just add them up to get an average, you might get the wrong answer.

Even if the robot is perfect at guessing what individuals like, the math of "averaging" those guesses can introduce a systematic error (bias) when you look at the group as a whole. It's like trying to measure the average height of a basketball team by averaging the heights of individual players, but accidentally weighing the tall players less because there are fewer of them in your sample.

The Core Concept: The "Group Bias"

The authors call this error Group Bias.

The Analogy: The Noisy Microphone
Imagine the AI robot is a microphone recording a choir.

The Individual Level: The microphone is perfect at hearing every single singer's voice clearly.
The Group Level: The manager wants to know how loud the "Tenor Section" is.
The Glitch: Because the Tenor section has fewer singers than the Soprano section, and the microphone was calibrated to hear the whole room, the Tenors' voices get drowned out or distorted when you try to calculate their specific volume. The math of "averaging" the individual recordings doesn't perfectly reconstruct the true volume of the Tenor section.

The paper shows that this happens even when:

The experiment was random (no cheating).
The AI model is mathematically correct for individuals.
You have a lot of data.

The bias happens because of how you aggregate the data, especially when groups are different sizes or have different "noise" levels.

Part 1: Detecting the Bias (The "Lie Detector")

The authors created a statistical tool to act like a Lie Detector Test for these group averages.

How it works: They compare two numbers for each group (e.g., New York):
1. The Model's Guess: What the AI says the group average is (by summing up individual predictions).
2. The Reality Check: What actually happened in a real-world experiment (A/B test) for that specific group.
The Result: If these two numbers are significantly different, the "Lie Detector" beeps. It tells you, "Hey, your model is lying about the New York group average!"

Part 2: Fixing the Bias (The "Smart Shrinkage")

Once you know the model is biased, how do you fix it?

The Naive Fix (The "Brute Force" Approach):
You might think, "Okay, the model is off by 5 points for New York. Let's just subtract 5 points from every prediction."

The Flaw: This is dangerous. If the model was off by 5 points just because of random noise (bad luck in the data), subtracting 5 points makes the error worse. It's like trying to fix a shaky photo by adding a filter that makes it blurrier.

The Authors' Fix (The "Smart Shrinkage"):
The authors propose a Shrinkage Strategy. Think of this as a Volume Knob that adjusts automatically based on how confident you are.

Scenario A (High Confidence): The data for New York is huge and clear. The model is definitely off by 5 points.
- Action: Turn the knob all the way up. Correct the full 5 points.
Scenario B (Low Confidence): The data for a small town is tiny and noisy. The model looks off by 5 points, but it might just be random noise.
- Action: Turn the knob down. Only correct 1 point. Don't trust the "error" too much because it might be a fluke.

This method balances bias (being wrong on average) and variance (being jittery/unstable). It shrinks the correction toward zero when the data is noisy, preventing you from over-correcting.

Part 3: The Business Impact (The "Profit Trade-off")

The paper also asks: "Does fixing this bias actually help us make money?"

The Analogy: The Targeting Dilemma
Imagine you are a marketer deciding who gets a coupon.

The Rule: "Give a coupon to anyone the AI thinks will buy the product."
The Twist: If you fix the group bias, you might change the AI's predictions slightly.
- Good News: Your reports to the boss (the group averages) are now accurate.
- Bad News: You might accidentally stop giving coupons to some people who would have bought the product, or give them to people who wouldn't.

The Finding:

If the profit margins are tight (you need to be very precise to make money), fixing the bias might actually lower your profits slightly because you are changing the targeting rules based on imperfect data.
However, if you use the Smart Shrinkage method (turning the knob carefully), the profit loss is usually tiny.
The Takeaway: It's often worth fixing the bias to get accurate reports and fair decisions, even if it costs a tiny bit of potential profit, as long as you don't over-correct.

Summary: What Should You Take Away?

Averages can lie: Just because an AI is good at predicting individuals doesn't mean its group averages are right.
Check your math: Always compare your model's group predictions against real-world experimental data (A/B tests) to see if they match.
Don't over-correct: If you find a bias, don't just subtract it blindly. Use a "smart" method that considers how noisy your data is. If the data is shaky, trust the original model a little more.
Fairness vs. Profit: Fixing these biases makes your data more honest and fair, but you have to be careful not to mess up your profit-maximizing decisions in the process.

In short: Trust your AI for individuals, but double-check its group math, and fix it gently.

1. Problem Definition

The paper addresses a critical, often overlooked issue in causal machine learning: Group Bias. While modern machine learning (ML) models can accurately estimate individual-level Conditional Average Treatment Effects (CATEs), organizations typically aggregate these predictions to subgroups (e.g., demographics, markets, risk strata) to derive Group Average Treatment Effects (GATEs) for reporting and decision-making.

The authors demonstrate that aggregating predicted CATEs does not generally recover the true experimental GATE, even when:

The underlying CATE model is correctly specified.
The data comes from a randomized controlled trial (RCT).
The model is unbiased at the individual level.

This discrepancy arises due to two main factors:

Non-collapsibility: For relative effect measures (e.g., risk ratios, lift), the expectation of a ratio is not the ratio of expectations (Jensen's inequality). Simple averaging of relative CATEs yields a biased GATE unless specific weighting is applied.
Regularization and Global Optimization: ML models are typically trained to minimize global loss functions. When groups have unequal sample sizes or signal-to-noise ratios, regularization shrinks predictions toward the global mean. Consequently, aggregating these shrunken predictions introduces systematic bias at the group level, particularly for smaller or noisier groups.

2. Methodology

The authors propose a unified statistical framework consisting of three stages: Definition, Detection, and Mitigation.

A. Definition of Group Bias

The authors define group bias ( $b_g$ ) as the discrepancy between the model-implied GATE and the experimentally identified GATE for a specific group $g$ :
$b_g = E[W \hat{\tau}_f(X) \mid G=g] - E[Y(1) - Y(0) \mid G=g]$
Where:

$\hat{\tau}_f(X)$ is the predicted CATE.
$W$ is a collapsibility weight function (uniform for additive effects; proportional to baseline outcomes for relative effects).
The bias is defined as the difference between the weighted average of predictions and the true experimental effect.

B. Detection Framework

The framework provides a non-parametric, model-agnostic method to detect bias:

Estimation: The bias is estimated as the difference between the model-implied GATE (aggregated predictions with weights) and the experimental GATE (calculated via difference-in-means or regression adjustment on the hold-out data).
Inference: The authors derive an asymptotically normal estimator for the bias. They provide a formal statistical test (Wald test) to determine if the bias is significantly different from zero.
Cross-Group Comparison: The framework allows testing for disparities in bias across different groups (e.g., is Group A more biased than Group B?).

C. Mitigation via Shrinkage

A naive correction (subtracting the estimated bias $\hat{b}_g$ ) is shown to be suboptimal because it ignores estimation variance, potentially amplifying noise in small groups. Instead, the authors propose a shrinkage-based correction:
$\hat{\tau}_{corrected}(X) = \hat{\tau}_f(X) - \gamma_g \hat{b}_g$
Where $\gamma_g \in [0, 1]$ is a shrinkage factor determined by minimizing the expected loss (risk) of the residual bias.

They derive closed-form solutions for optimal $\gamma_g$ under two loss functions:

Mean Error Loss: A binary decision rule. Correct fully ( $\gamma_g=1$ ) only if the bias is statistically significant; otherwise, do nothing ( $\gamma_g=0$ ).
Mean Squared Error (MSE) Loss: A continuous shrinkage factor that balances bias reduction against variance inflation:
$\gamma_g^* = \frac{b_g^2}{b_g^2 + \sigma_g^2}$
This acts as an empirical Bayes estimator, shrinking the correction toward zero when the signal-to-noise ratio is low (i.e., when the group is small or the bias estimate is noisy).

3. Key Contributions

Theoretical Identification: The paper formally identifies that CATE aggregation induces bias even under ideal experimental conditions, linking this to the concept of collapsibility in causal inference.
General Framework: It offers a model-agnostic detection and mitigation framework applicable to both additive and relative effect scales, requiring only sample moments and no retraining of the original ML model.
Optimal Shrinkage Strategies: The derivation of closed-form, risk-minimizing shrinkage factors that adaptively handle estimation uncertainty, preventing the over-correction of noisy groups.
Economic Analysis: The authors analyze the trade-offs of debiasing in profit-maximizing targeting. They show that while debiasing improves causal inference, it can alter targeting decisions and reduce profits if the bias correction shifts users across the decision threshold. They characterize when these trade-offs are significant.

4. Empirical Results

The framework was validated using two large-scale datasets:

Booking.com (A/B Test): Analyzed a marketing intervention across countries of origin.
- Findings: The raw CATE model significantly underestimated GATEs for smaller countries and overestimated them for larger ones due to regularization.
- Mitigation: The risk-minimizing shrinkage strategies (MSE-) successfully aligned predicted GATEs with experimental GATEs, reducing bias dispersion without the noise amplification seen in naive correction.
Criteo Uplift Prediction Dataset: Used for off-policy evaluation of targeting policies.
- Findings: Naive debiasing led to significant profit losses by altering targeting decisions for users near the profit-threshold.
- Insight: Risk-minimizing strategies (MSE- and Mean Error) preserved most of the profit while correcting the bias, demonstrating that "coarse" targeting policies (based on debiased GATEs) can be a viable alternative to granular CATE targeting when bias is a concern.

5. Significance and Implications

For Practitioners: The paper provides a practical toolkit for auditing ML-driven causal models. It shows that simply aggregating individual predictions is insufficient for reliable subgroup analysis.
For Decision Making: It highlights a fundamental tension between granular personalization (using raw CATEs) and reliable causal inference (using debiased GATEs). The authors suggest that organizations may need to accept coarser targeting policies to ensure that reported effects are causal and unbiased.
Methodological Advancement: By treating group bias as a statistical estimation problem with a bias-variance trade-off, the paper moves beyond simple fairness constraints to a rigorous, risk-based approach to bias mitigation in causal machine learning.

In conclusion, the paper establishes that group bias is an inherent feature of aggregating CATEs, provides a robust statistical method to detect and correct it, and offers economic guidance on when and how to apply these corrections in real-world targeting scenarios.