Multi-Perspective Confounder Adjustment

The Backdoor Criterion: Closing Spurious Paths in the Triad

The most common strategy for dealing with confounding is to identify and "block" all non-causal paths between a treatment and an outcome. These non-causal paths are called backdoor paths, and blocking them is the essence of the backdoor criterion.

Let's begin with a clear, visual explanation of confounding paths and how conditioning on a variable can block them, which is the core principle of backdoor adjustment.

The crucial insight is that a backdoor path is a source of spurious association between a treatment $T$ and an outcome $Y$ that is not part of the causal effect of $T$ on $Y$. To get an unbiased estimate of the causal effect, we must "close" every such path by conditioning on a variable that lies on it.

Adapting Backdoor Adjustment to the Triad

In your research, the treatment, outcome, and confounder can reside on different perspectives. Let's consider a core causal question:

• Treatment (T): A specific activity between the wife and bull (e.g., a weekend trip). This is a dyadic-level treatment.
• Outcome (Y): The husband's level of jealousy. This is an individual-level outcome.
• Confounder (U): The couple's pre-existing "relationship security." This is a latent construct that influences all three individuals.

Let's represent this with a DAG:

A simplified DAG illustrating how 'Relationship Security' (U) can confound the effect of a 'Wife-Bull Activity' (T) on the 'Husband's Jealousy' (Y). The red path is a backdoor path.

In this diagram, "Relationship Security" (U) is a common cause. A higher level of security might make the couple more likely to agree to the activity ($U\to T$) and might independently lower the husband's baseline jealousy ($U\to Y$). The path $T\leftarrow U\to Y$ is a backdoor path. If we simply correlate the activity with the husband's jealousy, we will get a biased result because we are confounding the true effect of the activity with the effect of pre-existing security.

To isolate the causal effect of $T$ on $Y$, we must close this backdoor path by conditioning on $U$. In practice, this means measuring "Relationship Security" (perhaps using a validated psychometric scale) and including it as a covariate in our statistical model.

The key challenge in your work is that confounders are rarely this simple. They can be individual traits, dyadic dynamics, or group-level properties, all creating complex backdoor paths that span the six perspectives of the triad. The first step in any analysis is to draw a comprehensive DAG representing all plausible backdoor paths and identify a "sufficient adjustment set"—the minimum set of variables you must condition on to block them all.

The Frontdoor Criterion: Finding a Causal Path Through the Noise

What happens if the confounder, like "Relationship Security" in our example, is unmeasurable or poorly measured? The backdoor criterion fails because we cannot condition on $U$. In this scenario, the frontdoor criterion provides an alternative, albeit more demanding, path to identification.

The frontdoor strategy works by identifying an isolated mediating mechanism. Instead of blocking the spurious path at the back, we identify the full causal path at the front.

Adapting Frontdoor Adjustment to the Triad

Let's revisit our example, but now assume "Relationship Security" ($U$) is unmeasured. We cannot block the backdoor path $T\leftarrow U\to Y$.

A DAG showing how 'Wife's Communication' (M) can satisfy the frontdoor criterion to identify the effect of the 'Activity' (T) on 'Jealousy' (Y) despite the unmeasured confounder 'Security' (U).

To use the frontdoor criterion, we need to find a mediating variable $M$ that satisfies the three conditions. A plausible candidate for $M$ is the Wife's post-activity communication style (e.g., level of detail, emotional tone). Let's check the conditions:

1. $M$ intercepts all directed paths from $T$ to $Y$. This assumes the only way the activity affects the husband's jealousy is through the wife's communication about it. This is a strong assumption, but plausible if direct effects (e.g., the husband observing the activity) are ruled out.
2. There is no unblocked backdoor path from $T$ to $M$. Our DAG shows the path $T\leftarrow U\to Y$, but no path from $U$ to $M$. This assumes that the couple's underlying security ($U$) does not directly influence the wife's communication style ($M$), which is another strong but potentially defensible assumption.
3. All backdoor paths from $M$ to $Y$ are blocked by $T$. The main backdoor path here is $M\leftarrow T\leftarrow U\to Y$. When we condition on the treatment $T$, we block this path at $T$.

If these three conditions hold, we can statistically identify the causal effect of $T$ on $Y$ even with the unmeasured confounder $U$. The calculation involves two steps: first, estimating the effect of $T$ on $M$, and second, estimating the effect of $M$ on $Y$ (while controlling for $T$). Your background in econometrics and mediation analysis makes you well-equipped to understand the mechanics of this two-stage estimation process.

This demonstrates how a variable from one perspective (the wife) can serve as a crucial tool to identify a causal effect involving another perspective (the husband) in the presence of unobserved confounding.

Advanced Strategy: Double Negative Controls

Both backdoor and frontdoor criteria can fail if the required assumptions are too strong. When dealing with complex network-like dependencies, such as those in your triad, unmeasured confounding can be particularly pervasive and difficult to model. For these situations, an advanced technique known as Double Negative Controls (DNC) offers a powerful alternative.

This method, developed for settings with "unmeasured network confounding," is conceptually sophisticated but directly addresses the core statistical challenge of your research. Given your statistics background, understanding its logic is a valuable extension.

The intuition is as follows: if you have an unmeasured confounder $U$, you can correct for its influence if you can find two other variables:

1. A Negative Control Outcome (NCO): A variable that is caused by the confounder $U$ but is not caused by the treatment $T$.
2. A Negative Control Exposure (NCE): A variable that is caused by the confounder $U$ but does not cause the main outcome $Y$.

By observing how $U$ affects these two "proxy" variables, we can triangulate and subtract its confounding effect from the main T-Y relationship.

Let's consider our triadic example again:

• T: Wife-Bull Activity
• Y: Husband's Jealousy
• U: Unmeasured "Couple's Openness"

To apply DNC, we would need to find:

• An NCO: A plausible NCO could be "The couple's willingness to try a new, non-relational activity" (e.g., take a cooking class together). This is likely influenced by their general "Openness" U, but should not be causally affected by a specific wife-bull activity T.
• An NCE: A plausible NCE could be "The bull's self-reported communication skills." This might be correlated with the couple's "Openness" U (people with better skills might be drawn to more open dynamics), but it should not have a direct causal effect on the husband's jealousy Y, other than through the activity itself.

The following paper establishes the formal framework for this method.

The DNC approach is powerful because it does not require us to measure the confounder itself, only to find valid proxies that satisfy the conditional independence assumptions. It is a frontier method well-suited to the complex, interdependent data structure of your triadic research.

Conclusion

In this lesson, we have explored how to move from theory to practice in estimating causal effects under interference. We've seen that while confounding presents a major hurdle, established and modern statistical strategies can be adapted to overcome it.

Key Takeaways:

• Backdoor Adjustment: This is the standard approach to confounding. In the triadic context, it requires a carefully constructed DAG to identify all backdoor paths, which may span multiple individual and dyadic perspectives, and then conditioning on a sufficient set of variables to block them.
• Frontdoor Adjustment: This is a valuable alternative when a key confounder is unmeasured. It works by identifying a mediating variable that isolates the causal pathway from the confounding one. Its success hinges on strong assumptions about the mediating process.
• Double Negative Controls: This is an advanced strategy for when both backdoor and frontdoor approaches are untenable due to pervasive, unmeasured network confounding. It uses proxy variables (NCO and NCE) to mathematically estimate and remove the influence of the unobserved confounder.

These adjustment strategies are powerful, but they all rest on the untestable assumption that we have correctly identified and controlled for all relevant confounders (or found valid proxies). What if our DAG is wrong? In our next lesson, we will confront this uncertainty head-on by learning how to conduct a sensitivity analysis. This will allow us to quantify how robust our causal conclusions are to the potential presence of unmeasured confounding.