Just Trial Once: Ongoing Causal Validation of Machine Learning Models (2502.09467v1)

Abstract: Machine learning (ML) models are increasingly used as decision-support tools in high-risk domains. Evaluating the causal impact of deploying such models can be done with a randomized controlled trial (RCT) that randomizes users to ML vs. control groups and assesses the effect on relevant outcomes. However, ML models are inevitably updated over time, and we often lack evidence for the causal impact of these updates. While the causal effect could be repeatedly validated with ongoing RCTs, such experiments are expensive and time-consuming to run. In this work, we present an alternative solution: using only data from a prior RCT, we give conditions under which the causal impact of a new ML model can be precisely bounded or estimated, even if it was not included in the RCT. Our assumptions incorporate two realistic constraints: ML predictions are often deterministic, and their impacts depend on user trust in the model. Based on our analysis, we give recommendations for trial designs that maximize our ability to assess future versions of an ML model. Our hope is that our trial design recommendations will save practitioners time and resources while allowing for quicker deployments of updates to ML models.

• The paper proposes a methodology to infer or bound the causal impact of new machine learning models using data from prior randomized controlled trials, avoiding the need for new, costly experiments.
• It formalizes conditions and derives tight bounds on a new model's causal impact, showing how existing RCT data, especially from multi-model trials, can verify core assumptions.
• The study recommends designing RCTs with multiple models to improve future evaluations and emphasizes that enhanced model accuracy alone does not guarantee improved causal outcomes.

This paper addresses the challenge of evaluating the causal impact of updated ML models in high-risk domains, like healthcare and criminal justice, using data from prior randomized controlled trials (RCTs). The core problem is that ML models are frequently updated, but running new RCTs for each update is expensive and time-consuming.

The authors propose a methodology to infer or bound the causal impact of deploying new models, even if they weren't part of the original RCT. This approach considers two practical aspects: ML model performance influences user trust (and, consequently, outcomes), and ML predictions are typically deterministic, potentially creating situations where certain model outputs are never observed for particular inputs.

The authors formalize conditions under which data from an existing RCT can be used to bound the causal impact of new models. These conditions include assumptions about the data generating process, performance monotonicity (improvements in model performance don't harm outcomes), the existence of neutral actions (actions that don't depend on model performance), and bounded outcomes. They derive tight bounds on the causal impact of deploying a new model. They show that core assumptions can be checked using data from RCTs that trialed multiple models.

The paper provides a simple estimator for these bounds and a procedure for generating asymptotically valid confidence intervals. A simulation paper demonstrates the framework's ability to provide more rigorous conclusions about model updates compared to simply judging models based on raw performance metrics.

The authors give practical recommendations for pre-trial design and post-trial analysis. They recommend conducting RCTs with multiple ML models and varying performance characteristics to maximize the ability to assess future model updates and emphasizing that improved accuracy alone is not sufficient for deployment, because a more accurate model may have worse outcomes if accuracy is gained by sacrificing performance on some key subpopulation.