source: arxiv statistics ml: data-poisoning audits for causal effect estimation
level: research
observational causal studies often combine data from many sources, which makes them open to attacks where someone adds carefully chosen fake records to shift the reported treatment effect. this work introduces a data-poisoning audit for augmented inverse-probability-weighted estimation. the analyst sets a list of possible fake records, a limit on how many can be added, and capacities for each source. an adversary then picks records to maximize the change in a chosen direction.
with preprocessing and nuisance models held fixed, the authors propose a greedy scan that finds the exact worst-case effect shift for any number of added records. to handle the fact that nuisance models are usually refitted after data changes, they derive a total-influence score. this score combines a record's direct impact with its indirect effect through changes in propensity and outcome models. they also provide a conservative bound for the fully refitted scenario.
the method helps analysts understand how vulnerable their causal estimates are to data poisoning. by computing worst-case shifts under different budgets, they can assess robustness before publishing results. the approach is designed for settings where data are pooled from multiple sites, a common practice in modern data science. it gives a practical tool for auditing sensitivity without requiring full retraining for every possible attack.
why it matters: it gives data scientists a way to measure how much a causal estimate could be manipulated by adding fake records, helping to build trust in pooled observational studies.
source: arxiv statistics ml: data-poisoning audits for causal effect estimation