Research map / Probabilistic and scientific modeling / Treatment effects

Just Ramp-Up: Debiasing Regression-based Estimator for A/B Tests under Network Interference

NeurIPS 2026 · Probabilistic and scientific modeling › Treatment effects · approach: Distributional treatment effects · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that pooling ramp-up A/B-test stages with different treatment proportions substantially reduces bias of regression-based GATE estimators (linear regression and a lightweight ChebConv GNN) under network interference.

Problem

Estimating the global average treatment effect (GATE) from A/B tests under network interference, where experiments only expose limited treatment proportions and GATE requires extrapolation to global treatment and control, leading to large bias.

Approach

Pool data from multiple ramp-up stages with different treatment proportions to train a regression estimator; closed-form bias-variance analysis under general linear interference shows bias dominates and merging stages with dispersed proportions reduces it, explained by increased micro-level exposure variation and synergy with cluster-level randomization.

Key results

• Under 2-hop interference with unit-level randomization, GNN bias/MSE falls from -1.401 / 1.963 (single stage at c=0.5) to -0.611 / 0.374 with two stages and to -0.159 / 0.025 with four stages.
• Cluster-level randomization with merged stages reaches MSE 0.008 at t=4 (vs 0.912 for the single-stage baseline under cluster randomization in the repeated-experiment table at t=1).
• Merging repeated experiments at a fixed 50% proportion does not reduce bias (unit bias worsens from -1.402 at t=1 to -1.488 at t=5).

Surprising findings

• Merging repeated experiments at the same treatment proportion does not reduce bias and can slightly increase it, despite more data.
• Adding intermediate ramp-up stages can increase bias; merging only the two extreme stages is often sufficient for linear interference.

Limitations

• Closed-form analysis covers only linear regression under general linear interference with fixed potential-outcome parameters across stages.
• Excludes cross-round (temporal) spillovers and carryover effects; nonlinear and dynamic-graph results are simulation-only without guarantees.
• Under a fixed traffic cap, merging can still leave substantial bias.

Experimental setup

Datasets: FB-Stanford3 (socfb-Stanford3) Facebook network

Benchmarks: FB-Stanford3 simulations under linear-in-means, multihop, multiplicative, quadratic, square-root, and dynamic-graph inte

Metrics: bias standard deviation MSE

Baselines: single-stage experiment (c=0.5) Lagrange interpolation estimator linear regression estimator

Builds on: linear-in-means model Neyman-Rubin potential outcomes ChebConv AIPW

Closest papers by meaning

• Interference Among First-Price Pacing Equilibria: A Bias and Variance Analysis (ICLR 2025 · similarity 0.86)
• Strategic A/B testing via Maximum Probability-driven Two-armed Bandit (ICML 2025 · similarity 0.85)
• Online Experimental Design With Estimation-Regret Trade-off Under Network Interference (NeurIPS 2025 · similarity 0.85)
• Independent-Set Design of Experiments for Estimating Treatment and Spillover Effects under Network Interference (ICLR 2024 · similarity 0.85)
• Higher-Order Causal Message Passing for Experimentation with Complex Interference (NeurIPS 2024 · similarity 0.85)
• Doubly Robust Causal Effect Estimation under Networked Interference via Targeted Learning (ICML 2024 · similarity 0.84)