AI Paper Atlas

Research map / Probabilistic and scientific modeling / Treatment effects

Tailoring Strictly Proper Scoring Rules for Downstream Tasks: An Application to Causal Inference

Roman Plaud, Alexandre Perez-Lebel, Antoine Saillenfest, Thomas Bonald, Marine Le Morvan, Gaël Varoquaux, Matthieu Labeau

ICML 2026 · Probabilistic and scientific modeling › Treatment effects · approach: Proximal and deep causal effect estimation · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace log-loss with the tailored IPW loss and canonical mappingIPW estimator with linear propensity model · Kang & SchaferRMSE: 66.90 -> 2.68Improved
replace log-loss with the tailored lossAIPW with linear propensity model · Kang & SchaferRMSE: 46.51 -> 5.07Improved
replace log-loss with the tailored lossMLP propensity model · ACIC 2017RMSE: better in 95 out of 96 configurationsImproved
replace log-loss with the tailored lossXGBoost propensity model · ACIC 2017RMSE: better in 62 out of 96 casesMixed
high selection strength (propensities near 0/1) versus low strengthtailored loss vs log-loss · ACIC 2017relative RMSE improvementImproved
add the variance (1/p^3) term to the bias-only loss (ablation)tailored loss · Kang & SchaferRMSE: IPW 3.39 -> 2.68Improved
add the variance term to the bias-only loss (ablation)tailored loss · IHDP and JobsRMSE: IHDP IPW 0.27 -> 0.34No clear effect
pair tailored loss with its canonical quartic mapping instead of sigmoid (ablation)logistic regression with SGD · Kang & Schafer simulationmax gradient norm: 1024 -> 0.275Improved
use the analytical p-y backward pass instead of autograd (ablation)MLP with tailored loss · synthetic N=10^6, d=100relative training time: 1.14x -> 1.02xImproved
use quartic canonical mapping in XGBoost custom objectiveXGBoost · synthetic N=10^6, d=100relative training time: 1.00x -> 5.03xHurt
use tailored loss instead of log-loss, Jobs datasetIPW with linear propensity model · JobsRMSE: 0.46 -> 0.03Improved
post-hoc trimming versus tailored lossIPW baseline · Jobsmean rank: Trimming 1.70 vs Custom 1.60 (IPW)No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy