Tracing the Whole System

Opening Scene

A plumber confirming that a valve genuinely affects a downstream fixture’s pressure still has a further, genuinely useful question to ask: does the valve affect pressure directly, or does it work by first changing flow rate in an intermediate pipe, which then changes pressure downstream? Both are real causal effects. But understanding the actual path — direct, indirect, or some mix of both — changes what you’d do to intervene on the system most effectively.

In Plain English

Mediation analysis decomposes a total causal effect into a direct effect (the cause acting on the outcome without passing through any specific intermediate variable) and an indirect effect (the cause acting on the outcome by first changing a mediating variable, which then changes the outcome). Understanding this decomposition matters practically: if an effect is entirely indirect, intervening directly on the mediator might achieve the same result more efficiently than intervening on the original cause.

The Old Way

Before formal mediation analysis existed, understanding a causal pathway’s internal structure was often left informal:

A business might know that a training program improves sales, without knowing whether it works by improving product knowledge, by improving confidence, or by some mix of both mediating factors.
A public health study might know that an intervention reduces a disease’s spread, without decomposing how much runs through behavior change versus how much runs through other mechanisms.
A researcher might establish that A causes C, without any formal method for testing whether it happens directly or through an intermediate step B.

In each case, knowing that an effect exists fell short of understanding how it actually operates.

What’s Changing (and Why AI Is the Reason)

Formal statistical mediation analysis, building directly on the causal graph framework from Article 5, now lets analysts rigorously decompose a total effect into its direct and indirect components, rather than leaving the internal mechanism unexamined.
This decomposition has genuine, practical value for intervention design: if most of a valuable effect runs through a specific mediator, that mediator itself might be a more efficient, more directly actionable intervention target.
As causal machine learning has matured, more flexible, less restrictive methods for estimating mediation effects have become available, extending beyond the classical linear models this approach originally relied on.

The Metaphor, Fully Extended

Behind the Wall	Mediation Analysis Concept
A valve’s effect on downstream pressure, direct or through an intermediate pipe	A cause’s effect on an outcome, direct or through a mediating variable
Tracing the exact path water takes through the system	Tracing the exact causal path an effect takes through a system
Deciding whether to adjust the original valve or the intermediate pipe	Deciding whether to intervene on the original cause or the mediating variable
Understanding the full plumbing path, not just confirming a connection exists	Understanding the full causal path, not just confirming an effect exists

For Beginners: What to Actually Do

Practice asking, for any established causal effect, “does this work directly, or does it pass through some intermediate step?”
Learn the basic distinction between a mediator (something the cause changes, which then changes the outcome) and a confounder (something that independently affects both).
Get comfortable with the idea that decomposing an effect into direct and indirect components has real, practical implications for where to intervene.

For Practitioners and Leaders: The Deeper Layer

Use mediation analysis to identify more efficient intervention points, when a valuable effect runs substantially through an identifiable mediator.
Invest in mediation analysis specifically for complex interventions where understanding the mechanism, not just confirming the outcome, has real strategic value.
Recognize mediation analysis relies on its own set of causal assumptions, built on the same causal graph reasoning covered in Article 5.

Quick Recap

Mediation analysis decomposes a total causal effect into direct and indirect components running through a mediating variable.
Understanding this decomposition has practical value for identifying the most efficient intervention point.
It builds directly on the causal graph framework, requiring its own explicit assumptions about the system’s structure.
Modern causal machine learning has extended mediation analysis beyond its classical, more restrictive linear origins.

Where This Fits in the Series

Article 14 covered tracing an effect’s full path through a system. Article 15 covers what happens when that system’s plumbing itself changes mid-analysis.