Effects conditional on post-treatment events generated by independent mechanisms

1. Problem Statement

In both observational studies and randomized controlled trials (RCTs), researchers frequently encounter post-treatment events (e.g., dropout, non-adherence, death) that occur after the assignment of treatment ($A$) but before the measurement of the outcome ($Y$).

• The Challenge: Conditioning on these post-treatment variables ($D$) to isolate specific causal effects (e.g., direct effects or effects among survivors) often leads to collider stratification bias. Even in RCTs, if $D$ is affected by $A$ and shares unmeasured common causes ($U$) with $Y$, conditioning on $D$ opens a non-causal path between $A$ and $Y$, rendering standard conditional comparisons uninterpretable.
• Current Limitations: Existing causal estimands that account for post-treatment events, such as Conditional Separable Effects (CSE) and Survivor Average Causal Effects (SACE), typically require the measurement of the unmeasured confounders ($U$) or proxy variables to achieve identification. In many practical settings, measuring $U$ is impossible.

2. Methodology and Core Assumptions

The authors propose a novel identification strategy that relies on the concept of "Independent Mechanisms" rather than the measurement of confounders.

Causal Framework

The study utilizes a Finest Fully Randomized Causally Interpreted Structured Tree Graph (FFRCISTG) model. The standard setup involves:

• $A$: Randomized treatment.
• $D$: Post-treatment event (e.g., death, non-adherence).
• $Y$: Outcome (undefined if $D$ occurs).
• $U$: Unmeasured common causes of $D$ and $Y$.

Key Assumptions

The identification of causal effects without measuring $U$ relies on decomposing the treatment and the post-treatment event generation process:

1. Treatment Decomposition (Assumptions 4 & 5): The treatment $A$ is conceptualized as having two independent components:

   • $A_D$: A component affecting the post-treatment event (e.g., anesthetic side effects causing death).
   • $A_Y$: A component affecting the outcome directly (e.g., the surgery itself affecting quality of life).
   • In observed data, $A = A_D = A_Y$, but hypothetical interventions can set them discordantly.

2. Independent Mechanisms (Assumptions 1, 2, & 3):

   • Mediation: The treatment affects the post-treatment event $D$ only through a specific sub-component $D_A$ (mediated by $A_D$).
   • Independence: The mechanism generating $D_A$ (the treatment-induced event) is statistically independent of the mechanism generating $D$ via other causes (represented by $U$).
   • Determinism: If the specific treatment-induced event $D_A$ occurs, the general post-treatment event $D$ occurs ($D_A=1 \implies D=1$).

3. Multiplicative Survival Model: These assumptions imply that the probability of survival ($D=0$) factorizes into a function of the treatment and a function of the unmeasured confounders:
   $$Pr(D^{a}=0 | U=u) = Pr(D=0 | U=u, D_A=0) \times Pr(D_A=0 | A=a)$$
   This multiplicative structure is the mathematical key to the identification result.

3. Key Contributions and Results

A. Identification of Conditional Separable Effects (CSE)

The primary contribution is Proposition 1, which states that under the "Independent Mechanisms" assumptions, the Conditional Separable Effect can be identified using only the observed data among those who did not experience the post-treatment event ($D=0$).

• The Estimator:
  $$CSE(a') = E(Y | A=1, D=0) - E(Y | A=0, D=0)$$
• Significance: This result is counter-intuitive because standard DAG theory suggests conditioning on a collider ($D$) induces bias. However, the authors prove that under the multiplicative survival model, the distribution of the unmeasured confounder $U$ remains invariant across treatment arms conditional on survival. Thus, the "naive" conditional comparison is actually a valid causal estimate of the effect of the $A_Y$ component while holding $A_D$ fixed.

B. Identification of Survivor Average Causal Effect (SACE)

The authors extend this logic to the SACE (the effect among "always-survivors"). By adding a monotonicity assumption (Assumption 6) and a cross-world independence assumption (Assumption 7), they show that the SACE is also identified by the same functional:
$$SACE = E(Y | A=1, D=0) - E(Y | A=0, D=0)$$

C. Relaxation of Assumptions

The paper provides Proposition 3 (Appendix D), showing that identification holds even if there is a measured variable $L$ that causes $D_A$, $D$, and $Y$, provided $L$ is adjusted for in the estimation. This makes the method applicable to settings where some, but not all, confounding is measured.

4. Illustrative Examples

The authors validate their theory through three distinct scenarios:

1. Plastic Surgery (Truncation by Death):

   • Scenario: Surgery ($A$) improves quality of life ($Y$) but carries a risk of death ($D$) due to specific anesthetic reactions ($D_A$).
   • Application: If anesthetic reactions are independent of patient comorbidities ($U$) that affect both death and quality of life, the difference in quality of life among survivors directly estimates the effect of the surgery itself, free from the bias of the anesthetic risk.

2. Birth Weight Paradox:

   • Scenario: Maternal smoking ($A$) is associated with low birth weight ($D$) and infant mortality ($Y$). Low birth weight infants of smokers have lower mortality than low birth weight infants of non-smokers (the paradox).
   • Application: The authors argue that if "low birth weight due to smoking" ($D_A$) is mechanistically independent of other causes of low birth weight/mortality ($U$), the association between smoking and mortality among high-birth-weight infants ($D=0$) is a valid estimate of the direct effect of smoking on mortality, resolving the paradox without needing to measure all confounders.

3. Smoking Cessation (Non-adherence):

   • Scenario: Comparing e-cigarettes vs. nicotine patches. Non-adherence ($D$) is driven by side effects ($D_A$).
   • Application: If side effects (e.g., skin irritation from patches) are independent of other reasons for non-adherence, the effect of the treatment assignment among adherers estimates the effect of the delivery system (e.g., patch vs. e-cig) independent of the side-effect profile.

5. Significance and Implications

• Overcoming the "Unmeasured Confounder" Barrier: This work challenges the dogma that conditioning on post-treatment events always requires measuring confounders to avoid bias. It demonstrates that specific structural assumptions about the mechanism of the post-treatment event can substitute for data on confounders.
• Practical Utility: It provides a rigorous justification for analyzing "per-protocol" or "survivor" populations in RCTs when the post-treatment event is driven by a specific, isolated mechanism (e.g., a specific drug side effect) that is unrelated to the patient's underlying health status.
• Theoretical Bridge: It connects separable effects, principal stratification, and instrumental variable analysis, showing that under specific "independent mechanism" conditions, simple conditional expectations yield complex causal estimands.
• Caveat: The validity of the results hinges entirely on the plausibility of the "Independent Mechanisms" assumption. In many real-world scenarios, treatment-induced events may correlate with patient characteristics ($U$), making this assumption difficult to justify without strong subject-matter knowledge.

In summary, Piccininni and Stensrud provide a powerful theoretical framework allowing researchers to interpret conditional associations in the presence of post-treatment events as causal effects, provided the event is generated through mechanisms independent of the unmeasured causes of the outcome.