Triadic Data: Unpacking Complex Mediation-Moderation

From Causal Story to Identifiable Model

Before you can test a hypothesis, you must first establish that the causal effect you're interested in is identifiable. In essence, identifiability means that it is theoretically possible to isolate the true causal effect from all the spurious correlations and biases that might exist in your data, assuming you had an infinite amount of it. Your background in causal analysis will make these formal concepts quite familiar, but applying them to interdependent relationship data requires careful consideration.

Directed Acyclic Graphs (DAGs) are the primary tool for this task. They force us to be explicit about our causal assumptions and, in return, provide a rigorous method for assessing identifiability. A foundational text by Felix Elwert on graphical causal models offers a clear framework for this.

Let's apply these principles to a concrete example from your research. Consider the second-stage moderated mediation hypothesis:

The indirect effect of a wife-bull activity (X) on the husband's jealousy (Y) through the wife's guilt (M) is weaker when husband-wife communication (W) is high.

The causal path is X → M → Y, and the moderator W influences the M → Y link.

To assess identifiability using a DAG, you must ensure you can satisfy the adjustment criterion for each component of the indirect effect. This means you must be able to block all non-causal ("backdoor") paths without blocking the causal path of interest or creating new bias by conditioning on a collider.

1. Identifiability of the a-path (X → M): You must be able to isolate the effect of the activity on the wife's guilt. This requires controlling for any common causes of both the activity and her guilt (e.g., a pre-existing level of marital satisfaction that influences both the types of activities chosen and her propensity for guilt).
2. Identifiability of the b-path (M → Y): You must be able to isolate the effect of the wife's guilt on the husband's jealousy, while controlling for X (the activity). This is the step most vulnerable to the bias Elwert describes. If there is any unmeasured confounder U that affects both the wife's guilt and the husband's jealousy (e.g., a shared but unspoken relationship anxiety), then the mediator M becomes a collider on the path X → M ← U → Y. When you control for M in your analysis (which is necessary for any mediation model), you condition on this collider, opening a spurious path between X and Y through U. This biases your estimate of the direct effect and, consequently, the indirect effect.

The interdependence of the triad amplifies this risk. A variable representing the bull's disposition could be a confounder (U) that affects both the wife's guilt (M) and the husband's jealousy (Y), making the M → Y path difficult to identify.

From Identifiable Model to Testable Hypothesis

Once you have a theoretically identifiable model, the next step is to determine if it is testable with practical statistical tools. Your goal is to find a statistical model that accurately represents the paths and dependencies of your DAG. Given the non-independence of your triadic data, standard regression is insufficient. This is where specialized models for dyadic and group data are indispensable.

David A. Kenny's suite of web-based applications provides an excellent entry point into specifying and testing these models. These tools are built upon the frameworks we've been discussing, such as the Actor-Partner Interdependence Model (APIM).

A Workflow for Your Triadic Research

These resources provide a clear path from a causal diagram to a statistical test. Here is a workflow for assessing a complex hypothesis in your research context:

1. State the Hypothesis and Draw the DAG:
Be precise. For example: "The effect of the wife's guilt (M) on the husband's jealousy (Y) is moderated by the bull's communication style (W), but only for a specific activity (X)." Draw this out, including all three individuals and plausible confounders.

2. Assess Identifiability with the DAG:
Go through your DAG systematically.

• Are there backdoor paths between your IV, mediator, and DV?
• Do you have observable variables that can block these paths?
• Have you inadvertently created collider bias by your choice of control variables or by focusing on a specific sub-population (e.g., only triads that stay together for more than a year)?
• Specifically, is there a risk of unmeasured confounding between the mediator and the outcome (the M-Y link)? Be honest about this. In social science, this path is almost always the weakest link in identification. A sensitivity analysis, which we covered earlier, would be essential here.

3. Determine the Appropriate Statistical Framework:

• Is the question dyadic? If your pathway is entirely contained within one dyad (e.g., Husband-Wife), you can use a dyadic model.
  • Distinguishability: The roles (e.g., husband vs. wife) are different. Use a tool like DINGY to confirm this statistically. You must use a model for distinguishable dyads, which allows actor and partner effects to differ for each member. The notes in the APIMeM resource show how to adapt the tool for this.
  • Model Specification: For a conditional indirect effect, you would need to combine the logic of APIMeM and APIMoM. This typically requires using a general Structural Equation Modeling (SEM) software package (like lavaan in R, which these tools use in the background) to build a custom model.
• Is the question triadic? If your pathway crosses all three members (e.g., bull's behavior -> wife's emotion -> husband's emotion), a dyadic model is insufficient.
  • Model Specification: The Group APIM (GAPIM) framework is the most appropriate starting point. You can specify a model where, for instance, the husband's jealousy is predicted by:
    • His own traits (actor effect).
    • An aggregate of the wife's and bull's traits/behaviors (partner/others effect).
    • Interactions between these components.
  • A full moderated mediation model in this context would be a highly advanced application, likely requiring a custom SEM specification that extends the logic of GAPIM.

The key insight is that while packaged tools like APIMeM are fantastic for standard dyadic questions, your unique triadic hypotheses will likely require you to specify a custom SEM. Your statistical background provides you with the necessary foundation to do this. The concepts and structures embedded in Kenny's tools provide the exact templates you need to build from.

Conclusion

This lesson has provided a bridge between abstract causal theory and concrete statistical practice. We've established a formal process for evaluating whether your research questions about complex emotional dynamics can be rigorously answered.

Key Takeaways:

• Identifiability precedes testability. You must first use a causal model (like a DAG) to argue that your effect of interest can be isolated from bias.
• The primary threat to identifying mediated effects is unmeasured confounding between the mediator and the outcome, which creates endogenous selection bias when you control for the mediator.
• Interdependence is not a bug; it's a feature to be modeled. Specialized models like APIM and GAPIM are designed to handle non-independence in dyadic and group data.
• For distinguishable roles (husband, wife, bull), you must use models that allow their causal pathways to differ. For fully triadic causal chains, a Group APIM (GAPIM) or a custom-built SEM is the necessary framework.

You now possess the complete analytical toolkit to design your capstone research project. You can define a causal question, visualize its structure, select an appropriate advanced statistical model, and now, assess whether that model is both identifiable and testable.

In our final module, you will synthesize all these skills. You will select and justify the specific experimental or observational design, the dyadic and triadic models, and the data collection plan necessary to construct a coherent causal narrative answering your primary research question.