Confounding — The Central Villain of Causal Inference

If you remember one thing about why causal conclusions go wrong, make it confounding. A confounder is a lurking variable that influences both the treatment and the outcome, manufacturing a correlation that looks causal but isn't. It's the reason observational data is so treacherous, the reason randomized experiments are the gold standard, and the single concept that explains most spurious causal claims in business, science, and ML. This post makes confounding concrete and shows why it's so hard to defeat.

By Pratik Dhanave · July 19, 2026 · 7 min read

#Causal Inference #Confounding #Bias #Statistics #Data Science

Post 2 reduced causal inference to “make the treated and untreated groups comparable.” This post is about the main thing that makes them not comparable: confounding. It’s the villain the rest of the series is organized against — experiments prevent it, observational methods try to adjust it away, causal graphs help you spot it. Understanding confounding deeply is most of understanding causal inference.

What a confounder is

A confounder is a variable that causally influences both the treatment and the outcome. Its presence creates a correlation between treatment and outcome that is not due to the treatment causing the outcome — a fake causal signal.

The canonical examples make it vivid: - Ice cream and drowning correlate, but ice cream doesn’t cause drowning. Hot weather is the confounder: it causes both more ice-cream sales and more swimming (hence drownings). Control for weather and the correlation vanishes. - A feature and retention: users who adopt your new feature retain better — but engagement is the confounder. Already-engaged users both adopt new features and retain; the feature may have caused nothing. The treated (adopters) and untreated groups differ in engagement, so they’re not comparable (post 2), and the retention gap reflects the pre-existing difference, not the feature. - A drug and recovery in observational data: healthier or wealthier patients may both receive the drug and recover more — so the drug looks effective when the confounder (health, wealth) drives both.

The structural signature is always the same: the confounder sits “upstream,” with causal arrows pointing to both the treatment and the outcome, so it makes them move together without any treatment→outcome link. This is exactly why the treated and untreated groups end up different in ways that matter — the confounder sorted units into treatment and affected their outcomes. Confounding is the formal name for “the groups weren’t comparable,” and it’s the most common reason causal claims from observational data are wrong.

Why confounding is so dangerous

Confounding is uniquely treacherous for reasons worth internalizing:

It produces confident, plausible, wrong conclusions. The correlation is real and the causal story is believable (“of course the feature boosts retention”), so confounded results don’t look like errors — they look like insights. This is what makes confounding so much more dangerous than noise: noise looks like noise, but confounding looks like a finding.
You must adjust for confounders to get the right answer — but only confounders. The fix is to “control for” or “adjust for” the confounder (compare treated and untreated units with the same weather, the same engagement level), which removes its contaminating influence. But this requires knowing what the confounders are and measuring them.
The ones you don’t know about are the real problem. You can only adjust for confounders you’ve identified and measured. Unmeasured (or unknown) confounders are the nightmare of observational causal inference: you can’t adjust for a confounder you didn’t think of or don’t have data on, and it silently contaminates your result with no warning. You can never prove no unmeasured confounder exists — which is precisely why observational causal claims are always open to the objection “but what about some confounder you missed?” This unfalsifiable doubt is the permanent weakness of observational methods.
Over-adjusting is also a trap. You must adjust for confounders but not for everything — controlling for a variable on the causal path from treatment to outcome (a mediator), or for a “collider,” can introduce bias rather than remove it. Knowing which variables to adjust for (and which to leave alone) is subtle and is exactly what causal graphs (post 5) are for.

So confounding isn’t just “a thing to watch for” — it’s a problem that produces believable falsehoods, requires knowing and measuring the right variables, can never be fully ruled out in observational data, and punishes both under- and over-adjustment. That combination is why causal inference is hard.

How the field fights confounding

The entire structure of causal inference is a set of strategies against confounding, which is why it’s the organizing villain of the series:

Randomization eliminates it (post 4). If you randomly assign treatment, then treatment is — by construction — unrelated to any confounder, known or unknown, because random assignment can’t be influenced by weather, engagement, health, or anything else. This is the magic of randomized experiments: they defeat all confounders at once, including the ones you never thought of, without needing to measure them. It’s why the RCT is the gold standard and why A/B tests are so valued — randomization sidesteps the entire confounding problem.
Adjustment reconstructs comparability (post 6). When you can’t randomize, you measure the confounders and statistically adjust for them — stratification, regression, matching, propensity scores — comparing like with like. This works only for measured confounders and rests on the (unprovable) assumption that you’ve captured them all (“no unmeasured confounding”), which is the central assumption and central vulnerability of observational causal inference.
Causal graphs make it visible (post 5). Drawing the assumed causal structure as a graph lets you see the confounding paths and determine exactly which variables to adjust for (and which not to), turning confounding from an intuition into a rule you can check.
Natural experiments exploit accidental randomization (post 7). When nature or policy assigns treatment in a way unrelated to confounders, you can sometimes get randomization-like credibility without a controlled experiment.

The unifying insight: causal inference is the science of dealing with confounding. Randomize it away if you can; carefully adjust for it if you can’t; use graphs to know what to adjust for; and stay honest that in observational data, unmeasured confounding is a doubt you can reduce but never eliminate. Every method and every caveat in the rest of the series traces back to this one villain.

The throughline: a confounder causally influences both the treatment and the outcome (hot weather → ice cream and drowning; engagement → feature adoption and retention), manufacturing a real but non-causal correlation — it’s the formal reason the treated and untreated groups aren’t comparable (post 2). It’s uniquely dangerous because it produces confident, plausible, wrong conclusions (looks like a finding, not an error), can only be adjusted away if you know and measure it (unmeasured confounders are unfixable and never fully ruled out — observational causal inference’s permanent weakness), and punishes over-adjustment too (mediators, colliders). The field is organized against it: randomization eliminates all confounders at once (post 4 — the gold standard), adjustment reconstructs comparability for measured confounders under the unprovable “no unmeasured confounding” assumption (post 6), causal graphs show which variables to adjust (post 5), and natural experiments exploit accidental randomization (post 7). Causal inference is the science of dealing with confounding.

Key takeaways

A confounder causally influences both the treatment and the outcome (hot weather → ice cream and drownings; engagement → feature adoption and retention), creating a real correlation that isn’t causal — the formal reason treated and untreated groups aren’t comparable (post 2) and the most common cause of wrong causal claims.
It’s uniquely dangerous because it yields confident, plausible, wrong conclusions — a confounded result looks like an insight, not an error (unlike noise).
The fix is to adjust for confounders (compare like with like) — but only for true confounders, and only ones you know and measure; unmeasured/unknown confounders can’t be adjusted for, silently contaminate results, and can never be proven absent — the permanent weakness of observational causal claims.
Over-adjustment is also a trap: controlling for a mediator (on the causal path) or a collider introduces bias — knowing which variables to adjust for is subtle and is what causal graphs (post 5) are for.
The field is organized against confounding: randomization defeats all confounders at once (post 4, even unknown ones, without measuring them — the gold standard), adjustment reconstructs comparability for measured confounders under “no unmeasured confounding” (post 6), graphs reveal what to adjust (post 5), natural experiments exploit accidental randomization (post 7) — causal inference is the science of dealing with confounding.