Research map / Probabilistic and scientific modeling / Treatment effects
Causal Inference out of Control: Estimating Performativity without Treatment Randomization
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.
Gives identifiability conditions for estimating the performative effect of platform actions on user consumption from observational data by modeling temporal dynamics, without assuming randomized actions, with finite-sample estimators and simulation experiments.
Problem
Estimating the causal effect of algorithmic platform actions on user consumption from observational data, where deterministic ML-driven actions violate overlap and confounding from past consumption is high-dimensional.
Approach
Model the platform as a controller in a time-aware dynamical system; exogenous consumption shocks at consecutive time steps propagate through a responsive platform action, restoring overlap of the previous action and the previous consumption state so the performative effect is identifiable without randomization.
Key results
- In the nonparametric case, two consecutive consumption shocks (M=2) suffice for identifiability of the performative effect, while M<2 can be unidentifiable.
- In the linear case, with rollouts of length K>=2 and a full-row-rank platform action, a single consumption shock suffices for identifiability; K=1 is insufficient.
- On RecLab simulation, the adjustment formula estimator with K=1 gives TE1 of 4.74 for EASE versus a ground truth of about 4.76, and coverage fails at K=3 (4-5 undefined terms of 27).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EASE rather than the random recommender | RecLab Topic, K=1, N=751 · RecLab Topic | estimated TE1: 4.58 -> 4.68 | Helped |
| Use EASE rather than the random recommender | RecLab Topic, K=2, N=751 · RecLab Topic | estimated TE1: 4.56 -> 4.69 | Helped |
| Fix EASE recommender weights at their initial value instead of continuing to update them (ablation) | EASE recommender, RecLab Topic, K=1, N=751 · RecLab Topic | estimated TE1: 4.68 -> 4.65 | No clear effect |
| Fix EASE recommender weights at their initial value instead of continuing to update them (ablation) | EASE recommender, RecLab Topic, K=2, N=751 · RecLab Topic | estimated TE1: 4.69 -> 4.66 | No clear effect |
| Increase the confounder window from K=2 to K=3 (ablation) | Random recommender, N=751 · RecLab Topic | fraction of undefined terms: 0 / 9 -> 4 / 27 | Hurt |
| Increase the confounder window from K=2 to K=3 (ablation) | EASE (no update), N=751 · RecLab Topic | fraction of undefined terms: 0 / 9 -> 5 / 26 | Hurt |
| Increase the confounder window from K=2 to K=3 (ablation) | EASE, N=751 · RecLab Topic | fraction of undefined terms: 0 / 9 -> 4 / 27 | Hurt |
| Increase the confounder window from K=2 to K=3 (ablation) | RecLab Topic, N=751 · RecLab Topic | estimated TE1 | No clear effect |
Surprising findings
- Even multiple consumption shocks do not suffice for identifiability in the standard time-agnostic causal model, while two consecutive shocks suffice in the time-aware model.
- Under the Markovian (K=1) setting, estimates are comparable to K=2, suggesting the Markovian assumption does not hinder accuracy.
Limitations
- Linear results and nonparametric results rely on structural assumptions (time-invariant dynamics, separable structural equations, Markovian dynamics) that may not hold in practice.
- Overlap coverage deteriorates rapidly with rollout length K in the adjustment formula estimator due to dimensionality.
- Validation is limited to semi-synthetic simulation and appendix experiments; the nonlinear full-rollout extension is left to future work.
Experimental setup
Datasets: RecLab simulated recommender data real microeconomic time-series data (appendix)
Benchmarks: RecLab Topic environment
Metrics: estimated treatment effect TE1 fraction of undefined terms (coverage) bootstrap confidence intervals estimation error
Baselines: EASE random recommender
Builds on: Pearl do-calculus performative prediction predicting-from-predictions RecLab
Closest papers by meaning
- On Online Experimentation without Device Identifiers (ICML 2024 · similarity 0.87)
- Transferring Causal Effects using Proxies (NeurIPS 2025 · similarity 0.86)
- A Practical Approach to Causal Inference over Time (AAAI 2025 · similarity 0.86)
- Off-policy estimation with adaptively collected data: the power of online learning (NeurIPS 2024 · similarity 0.84)
- Optimistic Algorithms for Adaptive Estimation of the Average Treatment Effect (ICML 2025 · similarity 0.84)
- Active Treatment Effect Estimation via Limited Samples (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.