Exposure Mappings and Interference

1. The Challenge of Interference

In the context of our vee triad (A: pivot; B and C: wings), any intervention or "treatment" applied to one member is likely to have spillover effects on the others. For example, if we design an intervention to improve B's sense of security, the resulting changes in B's behavior will almost certainly affect the satisfaction of A and C.

This is known as interference. In the potential outcomes framework, it means an individual's potential outcome depends not just on their own treatment status, but on the complete vector of treatments across all units in the system. Let's denote the treatment for each individual as $T_i$, where $i\in \{A,B,C\}$. The full treatment vector is $\mathbf{T}=(T_A,T_B,T_C)$. Without SUTVA, the potential outcome for person A is not simply $Y_A(T_A)$, but rather $Y_A(\mathbf{T})$.

If our treatment is a simple binary choice (e.g., attending a workshop vs. not), there are $2^3=8$ possible treatment vectors for the triad. This means each individual has 8 distinct potential outcomes. For a system with $N$ individuals, this explodes to $2^N$ potential outcomes, a situation often called the "curse of dimensionality" in causal inference. Estimating effects becomes practically impossible. We need a way to simplify this problem without resorting to the unrealistic assumption of no interference.

2. Defining the Exposure Mapping

The solution is to create a model of interference. We do this by defining an exposure mapping. An exposure mapping, $f(\cdot )$, is a function that reduces the complexity of the full treatment vector $\mathbf{T}$ into a smaller, more manageable set of "exposure conditions," which we'll denote by $d$.

The core idea is to make a weaker, more plausible assumption than "no interference." Instead of assuming $Y_i(T_i,{\mathbf{T}}_{-i})=Y_i(T_i,{\mathbf{T}}_{-i}^{\prime })$ for any two different treatment assignments to the other individuals ${\mathbf{T}}_{-i}$ and ${\mathbf{T}}_{-i}^{\prime }$ (which is the no-interference assumption), we assume that potential outcomes are equivalent only for treatment vectors that map to the same exposure condition. That is, if $f_i(\mathbf{T})=f_i({\mathbf{T}}^{\prime })$, then we assume $Y_i(\mathbf{T})=Y_i({\mathbf{T}}^{\prime })$.

The following reading provides a formal introduction to interference and the concept of an exposure mapping. It is a foundational text in this area of causal inference.

3. Constructing an Exposure Mapping for the Vee Triad

Let's apply the logic from the reading to our vee relationship. Imagine we are testing a daily mindfulness practice as a binary treatment $T_i\in \{0,1\}$ for each member $i\in \{A,B,C\}$. We want to understand its effect on, say, person B's daily reported satisfaction, $Y_B$.

Based on relationship theory, a plausible interference model might be that B's satisfaction depends on:

1. Whether B themselves practices mindfulness ($T_B$).
2. Whether B is "supported" by at least one partner who also practices mindfulness ($T_A=1$ or $T_C=1$).

This is precisely the structure of the example in the reading you just completed. We can define an exposure mapping $d_B=f_B(\mathbf{T},\theta )$ for person B that generates four distinct states:

• Direct + Indirect Exposure ($d_{11}$): B practices mindfulness ($T_B=1$), and at least one partner does ($T_A=1$ or $T_C=1$).
• Isolated Direct Exposure ($d_{10}$): B practices mindfulness ($T_B=1$), but neither partner does ($T_A=0$ and $T_C=0$).
• Indirect Exposure ($d_{01}$): B does not practice mindfulness ($T_B=0$), but at least one partner does ($T_A=1$ or $T_C=1$).
• No Exposure ($d_{00}$): No one in the triad practices mindfulness ($T_A=T_B=T_C=0$).

By defining this mapping, we have made a substantive scientific claim: we are hypothesizing that it is the presence of a treated partner that matters, not the number or identity of treated partners. Under this assumption, the eight potential outcomes $Y_B(T_A,T_B,T_C)$ collapse into four more manageable potential outcomes: $Y_B(d_{11}),Y_B(d_{10}),Y_B(d_{01}),Y_B(d_{00})$. This makes the problem of estimating causal effects tractable.

4. Exposure Mappings within Linear Models

The concept of an exposure mapping can also be expressed implicitly through a statistical model. Given your background in statistics, viewing this through the lens of a linear model can be particularly insightful. A correctly specified model for an outcome under interference must include terms that capture the influence of neighbors. These terms effectively define the exposure mapping.

The following paper presents a framework for using linear models for causal inference under network interference. It demonstrates how interference can be handled as a model specification problem.

This linear model approach offers a different but equally valid way to specify an exposure mapping. Instead of discrete categories, the "exposure" is a continuous variable representing the weighted influence of one's partners. For our vee triad, the adjacency matrix $G$ could represent the emotional closeness or time spent between partners, providing the weights for the spillover effects. For instance, the influence of the pivot (A) might be weighted more heavily than that of the other wing (C) on B's outcome.

The paper also shows how this can be extended to handle multiple interference mechanisms by including sums over different network matrices, $G_k$. This is particularly relevant for complex relationships where interference might operate through different channels (e.g., emotional, logistical, sexual).

Conclusion

In this lesson, we have moved from a simple network description to a causally-aware model of the vee triad. You have learned that interference is the rule, not the exception, in such systems, and that it poses a fundamental challenge to causal inference.

The key takeaways are:

• Interference violates the SUTVA and leads to a combinatorial explosion of potential outcomes, making naive causal analysis intractable.
• An exposure mapping, $d_i=f_i(\mathbf{T},\theta )$, is a formal, researcher-defined function that simplifies the full treatment vector into a manageable number of meaningful exposure conditions.
• This is a crucial modeling assumption about the nature of spillover effects. It makes the estimation of causal effects possible by reducing the dimensionality of the problem.
• We can define exposure mappings explicitly (e.g., creating discrete categories like "direct + indirect exposure") or implicitly through the specification of a statistical model (e.g., including a weighted sum of neighbors' treatments in a regression).

You now have a tool to formally state your assumptions about how influence flows through the triad. This is a prerequisite for defining precisely what you want to measure.

In our next lesson, we will use this concept of exposure mappings to redefine fundamental causal quantities. We will move beyond the simple "average treatment effect" and define specific estimands like direct effects, indirect (spillover) effects, and total effects for an individual operating within this interconnected network.