Optimal Experiments for Partial Causal Effect Identification
Authors: Tobias Maringgele, Jalal Etesami
Summary
The authors address a resource allocation problem in causal inference: when you can only partially identify a causal effect from observational data (meaning you get bounds, not a point estimate), which experiments should you run to tighten those bounds most effectively given a budget? They formalize this as the "max-potency" problem, where potency measures the worst-case guaranteed improvement in bound width, and show it's computationally hard (NP-hard via reduction from knapsack). They introduce two graphical pruning rules that together eliminate 50-88% of candidate experiments without any computation, and demonstrate the approach on health data (selecting experiments to estimate physical activity's effect on diabetes).
Main takeaways:
- When causal effects are only partially identifiable (you get bounds, not exact values), choosing which experiments to run to tighten those bounds is NP-hard
- "Epistemic potency" measures how much an experiment is guaranteed to narrow the bounds in the worst case
- Two pruning rules based on the causal graph structure alone can eliminate most useless experiments before doing expensive computation: a path-interception rule and an identifiability check
- On benchmark causal graphs, these rules prune 50-88% of candidate experiments on average
- The framework is general enough to apply to real observational datasets (demonstrated on health survey data)
Relevance
Not related to my work on persona installation or conditional behavior. This is about selecting interventions to learn causal structure, whereas I'm studying how behaviors are installed through different training procedures (fine-tuning, prompting, steering).
Abstract
arXiv:2605.06993v1 Announce Type: new Abstract: Causal queries are often only partially identifiable from observational data, and experiments that could tighten the resulting bounds are typically costly. We study the problem of selecting, prior to observing experimental outcomes, a cost-constrained subset of experiments that maximally tightens bounds on a target query. We formalize this as the max-potency problem, where epistemic potency measures the worst-case reduction in bound width guaranteed by an experiment, and show that this problem is NP-hard via a reduction from 0-1 knapsack. Building on the polynomial-programming framework of Duarte et al. (2023), we give a general procedure for evaluating epistemic potency in discrete settings. To control the super-exponential search space, we introduce two graphical pruning criteria that depend only on the causal graph and the query: a novel path-interception rule that exploits district structure to certify zero potency in linear time, and an identifiability check based on the ID algorithm. On Erdos-Renyi random graphs and 11 bnlearn benchmark networks, the two criteria together prune 50-88% of candidate experiments on average without solving a single polynomial program. For the general subset search, we show that ID-pruned experiments are combinatorially inert, yielding a super-exponential reduction in the number of subsets evaluated. We close with an end-to-end demonstration on observational NHANES data, selecting optimal experiments for estimating the effect of physical activity on diabetes.