When the Plumbing Changes Mid-Repair

Opening Scene

A plumber tracing a long-term pressure problem across many months faces a genuinely harder version of the confounding problem from Article 4: the household’s water usage patterns — the confounding factor — don’t stay fixed. They shift with the seasons, with occupancy changes, with new appliances installed partway through the investigation. A method built to control for a single, fixed confounder can badly mislead when the confounder itself is a moving target.

In Plain English

Time-varying confounding occurs when a confounding variable changes over time and is itself affected by earlier treatment — creating a genuinely trickier problem than the static confounding covered in Article 4, because standard adjustment methods can introduce new bias when the confounder is also a consequence of the treatment being studied. This shows up commonly in situations with repeated treatment decisions over time, like ongoing medical treatment or a sequence of business interventions.

The Old Way

Before this had formal causal inference language, longitudinal treatment effects were often estimated with methods that didn’t properly account for this feedback loop:

• Medical research on ongoing treatments historically struggled with the fact that a patient’s changing health status both affects and is affected by treatment decisions over time, a genuinely circular relationship standard methods weren’t built to handle.
• Business analysis of repeated interventions — an ongoing marketing campaign, say — often failed to account for how earlier campaign decisions affected later customer characteristics, which then influenced later campaign decisions.
• Standard regression methods, applied naively to this kind of longitudinal data, could introduce new bias rather than removing it, a genuinely counterintuitive and important finding.

Specialized methods emerged specifically because naively extending static confounding adjustment to this dynamic setting doesn’t work.

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

1. Methods like marginal structural models, developed specifically for time-varying confounding, now provide statistically valid ways to estimate causal effects in these genuinely more complex, longitudinal settings.
2. As organizations increasingly make sequences of related decisions over time — a series of interventions, an ongoing personalization strategy — time-varying confounding has become directly, practically relevant well beyond its original clinical research context.
3. Reinforcement learning and sequential decision-making methods in machine learning increasingly intersect with this exact problem, since an AI agent making repeated decisions over time faces precisely this same feedback structure.

The Metaphor, Fully Extended

Behind the Wall	Time-Varying Confounding Concept
A household’s usage patterns shifting across the months of an investigation	A confounding variable changing over the course of a study
Earlier repairs affecting later usage patterns, which affect later repair decisions	Earlier treatment affecting a later confounder, which affects later treatment decisions
A method built for a single fixed confounder failing on this moving target	Standard confounding adjustment failing on a time-varying confounder
Specialized long-term tracing methods built specifically for a shifting system	Specialized methods like marginal structural models built specifically for time-varying confounding

Quick Recap

• Time-varying confounding occurs when a confounder changes over time and is itself affected by earlier treatment decisions.
• Standard confounding adjustment methods can introduce new bias rather than removing it in this setting.
• Specialized methods like marginal structural models were developed specifically to handle this more complex case.
• This structure appears directly in sequential decision-making and reinforcement learning contexts.

Where This Fits in the Series

Article 15 covered a confounder that won’t hold still. Article 16 covers the risk that remains even after every known confounder has been carefully accounted for.