Causal Inference out of Control: Estimating Performativity without Treatment Randomization

Authors: Gary Cheng, Moritz Hardt, Celestine Mendler-Dünner

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We complement our claims with an analysis of ready-to-use finite sample estimators and empirical investigations. More broadly, our results deriving identifiability conditions tailored to digital platform settings illustrate a fruitful interplay of control theory and causal inference. (...) To complement our study, we propose a two-stage regression estimator and an adjustment formula estimator for estimating PEt from finite samples with theoretical guarantees. We also simulate a recommendation system to empirically test the efficacy of our assumptions at reducing overlap violations. We present more experiments in the Appendix, including ones using real microeconomic data.
Researcher Affiliation Academia 1Stanford University Department of Electrical Engineering 2Max Planck Institute for Intelligent Systems, T ubingen, and T ubingen AI Center 3ELLIS Institute T ubingen. Correspondence to: Gary Cheng <chenggar@stanford.edu>.
Pseudocode No The paper does not contain any sections or figures explicitly labeled as 'Pseudocode' or 'Algorithm'.
Open Source Code No The paper does not provide any explicit statements about open-sourcing the code for the described methodology, nor does it include links to a code repository.
Open Datasets Yes We simulate an online platform with items, recommendations, users, and ratings using Rec Lab (Krauth et al., 2020) (...) We use an avocado time series dataset (Kiggins, 2018)
Dataset Splits No The paper does not specify explicit training, validation, or test dataset splits. It mentions using 'semi-synthetic recommender system data' and an 'avocado time series dataset' but does not detail how these datasets were partitioned for training, validation, and testing.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as GPU models, CPU specifications, or cloud computing instance types.
Software Dependencies Yes For RF-DML, the residualizing procedure uses a random forest model implemented using sci-kit-learn 1.2.0.
Experiment Setup Yes We initiate the EASE recommender by simulating 100 cycles. In the no-update variation, we fix the EASE recommender weights to their initial value, and in the other variation, we continue to update the weights during the following N cycles based on user ratings. (...) We discretize the logged price into two buckets: ( 0.479,0.131],(0.131,0.683] and the logged demand into two buckets (14.539,15.014],(15.014,15.837].