Compared to What? Baselines and Metrics for Counterfactual Prompting

• URL: http://arxiv.org/abs/2605.01048v1
• Date: Fri, 01 May 2026 19:23:33 GMT
• Title: Compared to What? Baselines and Metrics for Counterfactual Prompting
• Abstract summary: We observe prediction flip rates on MedQA of 14.9% when we surgically change patient gender.<n>We propose a framework in which we compare differences observed under target interventions to those induced by paraphrasing inputs.<n>We find that these effects largely dissipate when we account for general model sensitivity.
• Score: 39.56472929066589
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Counterfactual prompting (i.e., perturbing a single factor and measuring output change) is widely used to evaluate things like LLM bias and CoT faithfulness. But in this work we argue that observed effects cannot be attributed to the targeted factor without accounting for baseline ``meaning-preserving'' modifications to text that establish general model sensitivity. This is because every counterfactual edit is a compound treatment that bundles the variable of interest with incidental surface-form variation; this violates treatment variation irrelevance. We observe prediction flip rates on MedQA of 14.9% when we surgically change patient gender. However, this is statistically indistinguishable from the flip rates induced by simply paraphrasing inputs (14.1%). In this case, it would therefore be unwarranted to conclude that the LLM is especially sensitive to patient gender. To account for this and robustly measure the effects of targeted interventions, we propose a framework in which we compare (via statistical testing) differences observed under target interventions to those induced by paraphrasing inputs. We then use this framework to revisit a analysis done on the MedPerturb dataset, which reported evidence of model sensitivity to patient demographics and stylistic cues. We find that these effects largely dissipate when we account for general model sensitivity, with only 5 of 120 tests reaching statistical significance. Applying the same framework to occupational biography classification, we detect clearly significant directional gender bias, showing that the framework identifies real directional effects even when they are small. We evaluate a range of metrics -- aggregate, per-sample distributional, and regression -- and find that per-sample metrics are dramatically more powerful than aggregate metrics and regression powerfully and uniquely characterizes effect direction and magnitude.

Related papers

• Hit Selection Using SSMD-Based Machine Learning Performance Metrics in High-Throughput Screening Assays [0.0]
  We introduce a model-based framework that derives sensitivity, specificity, and area classification metrics from the strictly standardized mean difference (SSMD)<n>Unlike classical statistical power, which approaches 1 as sample size grows regardless of how small the true non-zero difference between group means is, the SSMD-derived sensitivity converges to a finite population value.<n>We demonstrate the utility of this framework in a hepatitis C virus primary screen comprising approximately 22,000 single-replicate measurements.
  arXiv  Detail & Related papers  (2026-08-06T20:33:48Z)
• Beyond Scalar Scores: Exploring LLM-based Metrics for Clinical Significance Evaluation in Radiology Reports [49.5225801722164]
  Existing metrics obscure this requirement by reducing report quality to a medically ungrounded scalar.<n>We study this boundary using ReEvalMed benchmark as testbed and evaluate metric-level clinical significance.
  arXiv  Detail & Related papers  (2026-06-17T08:10:30Z)
• Assessing the robustness of heterogeneous treatment effects in survival analysis under informative censoring [50.164756034797136]
  Dropout is common in clinical studies, with up to half of patients leaving early due to side effects or other reasons.<n>When dropout is informative, it introduces censoring bias, because of which treatment effect estimates are also biased.<n>We propose an assumption-lean framework to assess the robustness of conditional average treatment effect estimates in survival analysis when facing censoring bias.
  arXiv  Detail & Related papers  (2025-10-15T10:51:17Z)
• Estimating Misreporting in the Presence of Genuine Modification: A Causal Perspective [11.438018798893113]
  In settings where ML models are used to inform the allocation of resources, agents might have an incentive to strategically change their features to secure better outcomes.<n>We propose a causally-motivated approach to identify and quantify how much an agent misreports on average by distinguishing deceptive changes in their features from genuine modification.
  arXiv  Detail & Related papers  (2025-05-29T19:06:30Z)
• Causal Inference from Text: Unveiling Interactions between Variables [20.677407402398405]
  Existing methods only account for confounding covariables that affect both treatment and outcome. This bias arises from insufficient consideration of non-confounding covariables. In this work, we aim to mitigate the bias by unveiling interactions between different variables.
  arXiv  Detail & Related papers  (2023-11-09T11:29:44Z)
• Causal Dynamic Variational Autoencoder for Counterfactual Regression in Longitudinal Data [5.766786445424951]
  Current methods for estimating treatment effects over time assume that all confounders are observed or attempt to infer unobserved ones.<n>Our approach focuses on unobserved adjustment variables, which specifically have a causal effect on the outcome sequence.<n>Our proposed Causal Dynamic Variational Autoencoder (CDVAE) is grounded in theoretical guarantees concerning the validity of latent adjustment variables.
  arXiv  Detail & Related papers  (2023-10-16T16:32:35Z)
• Sharp Bounds for Generalized Causal Sensitivity Analysis [30.77874108094485]
  We propose a unified framework for causal sensitivity analysis under unobserved confounding. This includes (conditional) average treatment effects, effects for mediation analysis and path analysis, and distributional effects. Our bounds for (conditional) average treatment effects coincide with recent optimality results for causal sensitivity analysis.
  arXiv  Detail & Related papers  (2023-05-26T14:44:32Z)
• A Notion of Feature Importance by Decorrelation and Detection of Trends by Random Forest Regression [1.675857332621569]
  We introduce a novel notion of feature importance based on the well-studied Gram-Schmidt decorrelation method. We propose two estimators for identifying trends in the data using random forest regression.
  arXiv  Detail & Related papers  (2023-03-02T11:01:49Z)
• Counterfactual Invariance to Spurious Correlations: Why and How to Pass Stress Tests [87.60900567941428]
  A spurious correlation' is the dependence of a model on some aspect of the input data that an analyst thinks shouldn't matter. In machine learning, these have a know-it-when-you-see-it character. We study stress testing using the tools of causal inference.
  arXiv  Detail & Related papers  (2021-05-31T14:39:38Z)
• Efficient Causal Inference from Combined Observational and Interventional Data through Causal Reductions [68.6505592770171]
  Unobserved confounding is one of the main challenges when estimating causal effects. We propose a novel causal reduction method that replaces an arbitrary number of possibly high-dimensional latent confounders. We propose a learning algorithm to estimate the parameterized reduced model jointly from observational and interventional data.
  arXiv  Detail & Related papers  (2021-03-08T14:29:07Z)
• Increasing the efficiency of randomized trial estimates via linear adjustment for a prognostic score [59.75318183140857]
  Estimating causal effects from randomized experiments is central to clinical research. Most methods for historical borrowing achieve reductions in variance by sacrificing strict type-I error rate control.
  arXiv  Detail & Related papers  (2020-12-17T21:10:10Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.