AI & ML
A mechanistic characterization of missingness at random
arXiv stat.TH0 views
arXiv:2611.22022v1 Announce Type: cross Abstract: The missingness-at-random (MAR) assumption underlying many missing-data methods describes abstract statistical independences, whereas practitioners reason more naturally about the causal mechanisms that generate missingness. To bridge this gap, we provide a mechanistic characterization of MAR in terms of paths in a directed acyclic graph (DAG) modified to allow certain edges to cease to operate when an incomplete variable is missing, a property termed "context-specific causal irrelevance (CSCI)". Under a faithfulness condition, MAR holds if and only if the modified DAG contains no MAR-violating paths. Without CSCI, this condition reduces to a previously established d-separation condition, VMAR; thus, MAR without VMAR requires specific kinds of CSCI, as can occur with monotone dropout. These results shift the burden on investigators from passively invoking MAR to justifying, on substantive grounds, the specific mechanisms it requires. We then consider identification of a target law p(v) of variables V by a MAR-assuming analysis, e.g., multiple imputation, that includes complete auxiliary variables Z (variables not in V). Barring a perfect cancellation, such an analysis recovers p(v) if and only if MAR holds with respect to V and Z jointly. We show that if there exists any set of auxiliaries such that (V, Z) obeys MAR, then the complete candidate auxiliaries that are ancestors of the target variables or their missingness indicators form a sufficient set. Contrary to standard recommendations, not all candidate auxiliaries should be used: even complete auxiliaries can introduce MAR violations that would not otherwise occur. We illustrate these results with simulations and an analysis of sexual identity change in the Stockholm Public Health Cohort, where our results indicate that a candidate auxiliary, income, should be excluded; including it attenuated estimates toward the null.
Source:https://arxiv.org/abs/2611.22022PreviousIdentifiability of latent variable models with flexible latent distributionsNextThe Price of Hard Coverage Constraints in Online Conformal Prediction
Related posts
BenchAudit – Auditing LLM benchmarks for leaks, data reuse, and harness gaps
0Hacker News New
Guillotine DJ – Announcement Trailer
0Hacker News New
Researchers hijack Hugging Face account for popular Kokoro AI model
0Hacker News New
Automated LLM Fine-Tuning: GPU Selection, Dataset Analysis and Feedback
0Hacker News New
Human-written Launch HN posts get more upvotes than AI-generated ones
0Hacker News New

Parallels Desktop review
0TechRadar