Introduction
Standard causal inference typically focuses on "hard" interventions—setting a variable $X$ to a fixed constant value, denoted by the $do(x)$ operator. However, many real-world policies are better described as "shift interventions." Instead of asking "what if everyone received 10mg of a drug?", we ask "what if everyone received 5mg more than they currently do?" These "soft" interventions are context-dependent, as the new value depends on the natural, observed value. This paper formalizes the identification and estimation of these effects, particularly in the presence of unobserved confounders.
Defining Shift Interventions: SITs vs. SIPs
The authors distinguish between two primary classes of shift interventions:
- Shift Interventions on the Treated (SITs): These are defined relative to the naturally observed values of the treatment. They are conceptually linked to the "effect of treatment on the treated" (ETT).
- Shift Interventions as Policies (SIPs): These are defined recursively. In a multi-treatment setting, a shift in the first treatment affects the second treatment relative to what it would have been under the first shift. This framework aligns with dynamic treatment regimes.
A key insight is that SITs and SIPs diverge mathematically. Shifting based on "what actually happened" (SIT) versus "what would have happened under the policy" (SIP) requires different identification conditions and formulas.
Graphical Identification
To determine when these effects can be calculated from observed data, the authors utilize Directed Acyclic Graphs (DAGs) for fully observed systems and Acyclic Directed Mixed Graphs (ADMGs) for models with hidden variables.
The paper provides "sound and complete" algorithms for identification. A significant finding is the "Recanting Witness" problem: SITs are generally harder to identify than SIPs. Even with full data, complex path dependencies in multi-treatment scenarios can render SITs unidentifiable, a challenge similar to identifying the ETT in standard causal analysis.
Estimation and Empirical Results
Beyond identification, the authors derive an Efficient Influence Function (EIF) for single-treatment cases. This enables "doubly robust" estimation, where the causal effect estimate remains unbiased if either the treatment assignment model or the outcome model is correctly specified.
Simulation and Real-World Application
- Simulation: Synthetic data confirmed the double robustness of the proposed estimator.
- ICU Readmission: Using EHR data from over 5,000 cardiac surgery patients, the authors estimated the effect of increasing ICU stays by 24 hours. They found that this shift did not significantly reduce readmission ("bounceback") rates. The analysis suggested that readmission risk was driven by baseline health factors rather than the duration of the stay itself.
Conclusion and Future Directions
While the paper establishes a rigorous foundation for shift interventions, challenges remain. The current efficient estimator is limited to single-treatment cases, and handling continuous shifts without discretization (to maintain "positivity") is an area for future research. Nonetheless, shift interventions offer a more practical and policy-relevant framework for decision-makers than traditional fixed-value interventions.
Summary updated