Causal Identification under Interference: The Role of Treatment Assignment Independence

This paper demonstrates that standard causal identification methods (such as IV, RDD, and DiD) still identify valid "average direct effects" under arbitrary interference, provided that treatment assignments across units remain independent, and proposes a sensitivity analysis framework to quantify the impact of violating this independence.

The Gist

Imagine you are trying to study the effect of a new fertilizer on how tall plants grow.

The Old Way: The "Lonely Plant" Assumption

In traditional science, researchers often use the "Individualistic Treatment Response" (ITR) assumption. This is like assuming every plant in your garden is a hermit. You assume that if you give Plant A fertilizer, its growth depends only on that fertilizer. You assume Plant A doesn't care what Plant B, C, or D are doing. This makes the math easy: you just compare fertilized plants to unfertilized ones and call it a day.

The Problem: The "Social Butterfly" Reality

But in the real world—whether it's plants, economics, or medicine—nothing is a hermit. This is called Interference.

• In the garden: If you fertilize Plant A, it might grow so big that it shades Plant B, or its roots might steal nutrients from Plant C.
• In the economy: If you give a job-training program to one person, they might take a job that would have gone to someone else (competition), or they might tell their friends about it (information sharing).

When "interference" happens, the old math breaks. If you use the "Lonely Plant" formulas in a "Social Butterfly" world, your results will be biased. You might think the fertilizer is amazing, but really, you're just seeing the side effects of plants crowding each other out.

The Paper’s Big Discovery: The "Secret Handshake"

The authors of this paper asked: "If we know the world is social, can we still use our old math formulas without them lying to us?"

They found that the math can still work, but only if one specific condition is met. They call this Conditional Assignment Independence (CAI).

The Analogy: The Secret Handshake
Imagine a group of people at a party. You want to know if wearing a red hat makes people more outgoing.

• Interference: If I wear a red hat, I might start talking to you, which changes your behavior.
• The "Secret Handshake" (CAI): The math stays accurate as long as the decision to wear a red hat isn't part of a "secret handshake" or a coordinated plan.

If people choose their hats completely independently (even if they interact once they are wearing them), the old math formulas actually identify a very specific, useful thing called the Average Direct Effect. This is the "pure" effect of the hat on you, stripped of the chaos caused by everyone else's hats.

However, if there is a secret handshake—meaning people coordinate their hat colors (e.g., "If you wear red, I'll wear red too")—then the old math becomes a mess of confusion and bias.

The "Stress Test": How Much Can We Trust the Results?

Since researchers often don't know if a "secret handshake" is happening in their data, the authors created a Sensitivity Analysis.

Think of this like a Stress Test for a Bridge.
The researchers say: "We assume there is no secret handshake (no coordination). But let's pretend there is. How much coordination would it take to make our bridge collapse (make our scientific conclusion wrong)?"

They created a tool (an R package called caisensitivity) that tells a researcher: "Your results are very strong; you'd need a massive, highly coordinated conspiracy of hat-wearers to prove you wrong," or "Your results are fragile; even a tiny bit of coordination would flip your conclusion upside down."

Summary in Three Sentences

1. The Problem: Most economic studies assume people act in isolation, but in reality, people's actions affect one another (interference).
2. The Solution: The old math formulas still work to find "pure" individual effects, provided that people's decisions to participate aren't being coordinated behind the scenes.
3. The Tool: The authors provided a way for scientists to "stress test" their findings to see how much hidden coordination would be required to make their results invalid.

Technical Summary

This paper, "Causal Identification under Interference: The Role of Treatment Assignment Independence" by Owusu and Márquez, addresses a fundamental gap in causal inference: what happens to standard empirical identification strategies when the "no-interference" (or Individualistic Treatment Response, ITR) assumption is violated?

Below is a detailed technical summary of the paper.

---

1. The Problem: The Failure of ITR

In most economic and social science research, researchers use identification strategies—such as Selection-on-Observables (SOO), Instrumental Variables (IV), Regression Discontinuity (RDD), and Difference-in-Differences (DiD)—under the assumption that a unit's outcome depends only on its own treatment. This is the ITR assumption.

However, in many real-world settings (e.g., job training programs, infectious disease spread, or market competition), interference occurs: the treatment of one unit affects the outcomes of others (spillovers). When interference is present, the standard "Average Treatment Effect" (ATE) is no longer well-defined by the usual formulas. Researchers often continue to use these formulas while remaining "agnostic" about the structure of interference, which risks producing biased or uninterpretable results.

2. Methodology: A Finite-Population Framework

The authors adopt a finite-population potential-outcomes framework where a unit $i$’s outcome $Y_i$ can depend arbitrarily on the entire vector of treatment assignments $D = (D_1, \dots, D_N)$.

The core of their methodology is the distinction between two types of independence:

1. Interference in Outcomes: The dependence of $Y_i$ on the treatments of others ($D_{-i}$).
2. Dependence in Assignment: The dependence of unit $i$’s treatment ($D_i$) on the treatments of others ($D_{-i}$).

The authors introduce a critical new condition: Conditional Assignment Independence (CAI).

• Assumption 3 (CAI): $D_{-i} \perp \!\! \perp D_i \mid W_i$.
  This assumption states that, conditional on observed covariates, a unit’s treatment assignment is independent of the treatment assignments of all other units.

3. Key Contributions and Results

A. Characterization of Identification Formulas

The paper’s primary theoretical contribution is proving that standard ITR-based formulas do not necessarily "break" under interference; rather, they transform into different causal objects depending on the assignment mechanism.

• Selection-on-Observables (SOO):
  • If CAI holds, the standard formula identifies the Average Direct Effect (ADE)—the effect of a unit's own treatment holding others' treatments fixed.
  • If CAI fails, the formula identifies the ADE plus a bias term ($B_{ADE}$) that depends on how the distribution of others' treatments shifts when unit $i$ is treated vs. untreated.
• Instrumental Variables (IV):
  • If the instrument is independent across units, the Wald ratio identifies the Local Average Direct Effect (LADE).
  • If instruments are dependent, the Wald ratio is biased by the systematic difference in others' instruments across instrument values.
• Regression Discontinuity (RDD):
  • Under a sharp design, if the running variables are independent across units, the discontinuity identifies the ADE. Otherwise, it is confounded by shifts in the distribution of others' treatments near the cutoff.
• Difference-in-Differences (DiD):
  • If CAI holds, the DiD estimator identifies the Average Direct Effect on the Treated (ADTT). If CAI fails, the estimator is biased by the interaction between untreated trends and the differing distributions of others' treatments between groups.

B. Sensitivity Analysis Framework

Since CAI is not testable from observed data (it is a restriction on the unobserved assignment mechanism), the authors propose a novel sensitivity analysis procedure for the SOO design.

• They model departures from CAI by allowing the variance of the number of treated units within covariate strata to deviate from the binomial (independent) benchmark by a factor $\xi \geq 1$.
• They define a robustness value ($\xi^*$): the smallest degree of assignment dependence required to overturn a statistically significant result.

4. Empirical Evidence and Simulations

• Monte Carlo Study: The authors demonstrate that the IPW (Inverse Probability Weighting) estimator is unbiased for the ADE when CAI holds, but becomes increasingly biased as the correlation between treatment assignments ($\rho$) increases.
• Simulation (LaLonde Data): They show that the sensitivity of the Fisher P-value to CAI violations depends heavily on the distribution of outcomes. If outcomes are "heavy-tailed" (contain outliers), the inference is much more sensitive to assignment dependence than if outcomes are Gaussian.
• Application: Applying the method to the LaLonde job-training data, they show how a researcher can quantify how much "peer interference" in treatment take-up would be needed to invalidate their findings.

5. Significance

The paper provides a rigorous theoretical "safety net" for applied economists. Its significance lies in three areas:

1. Reinterpretation: It tells researchers that their estimators are still targeting meaningful causal objects (ADEs) even if spillovers exist, provided that treatment assignment is relatively independent across units.
2. Warning: It provides a formal mathematical explanation for why "spillover" settings often lead to biased estimates if the assignment mechanism is endogenous to the social network or group structure.
3. Tooling: By providing a sensitivity analysis framework and an R package (caisensitivity), it moves the literature from qualitative concerns about interference to quantitative assessments of robustness.