Confounding and the Backdoor Criterion

An analyst notices that hedge funds with more Bloomberg terminals per trader show higher returns, and wonders whether buying more terminals causes better performance. But fund size drives both: bigger funds buy more terminals and have more resources to generate returns through other channels. Terminal count and returns move together not because one causes the other but because a third thing, fund size, causes both. That third thing is a confounder, and the question of exactly which variables to control for to strip its influence out — no more, no fewer — is answered by the backdoor criterion.

An analogy: a false trail with two starting points

Picture two trails that both start at the same trailhead (fund size) and end at two different lookout points (terminals, returns). If you only ever see hikers arriving at the lookouts, the two lookouts will seem connected — busy days at one coincide with busy days at the other — purely because they share a starting point, not because one lookout feeds hikers into the other. To see whether the lookouts are directly connected by their own trail, you need to control for the shared trailhead: compare only hikers who started from the same point. Confounding is exactly this kind of spurious connection through a shared origin, and "closing the backdoor" means blocking that shared-origin path in your analysis.

The criterion, one symbol at a time

Draw the causal structure as a DAG (see directed acyclic graphs for causality): nodes are variables, arrows are direct causal effects. A backdoor path from treatment $D$ to outcome $Y$ is any path connecting them that starts with an arrow pointing into $D$ (as opposed to the "front door" path where $D$'s own arrows lead forward to $Y$). A set of variables $Z$ satisfies the backdoor criterion relative to $(D, Y)$ if:

1. No variable in $Z$ is a descendant of $D$, and
2. $Z$ blocks every backdoor path between $D$ and $Y$.

$$P(Y \mid do(D=d)) = \sum_{z} P(Y \mid D=d, Z=z)\, P(Z=z)$$

In plain English: if you can find a set of variables that intercepts every non-causal, backward-pointing route from treatment to outcome, then averaging the treatment-outcome relationship within each value of those variables and re-weighting by how common that value is, recovers the true causal effect — exactly what you'd get from an experiment. $do(D=d)$ denotes intervening to set $D=d$, as opposed to merely observing $D=d$.

Worked example 1: closing the backdoor by hand

Fund size $S$ causes both terminal count $T$ and returns $R$; suppose there's no other link. The backdoor path is $T \leftarrow S \rightarrow R$. Controlling for $S$ blocks it. Say among small funds ($50m), average terminal count is 5 and average return is 6%; among large funds ($500m), average terminal count is 20 and average return is 9%. Pooled naively, terminals and returns look strongly related (5 terminals → 6%, 20 terminals → 9%). But suppose that within small funds, going from 3 to 7 terminals moves average return from 6.1% to 5.9% — no real effect — and within large funds, going from 15 to 25 terminals moves it from 9.1% to 8.9% — again no real effect. Conditioning on $S = \{Z\}$ reveals the true within-stratum effect of terminals on returns is roughly zero; the pooled correlation was entirely confounding by fund size.

Worked example 2: which set actually works

Suppose the DAG is: $\text{Experience} \to \text{Adopts New Strategy}$, $\text{Experience} \to \text{Returns}$, and separately $\text{Market Regime} \to \text{Adopts New Strategy}$, $\text{Market Regime} \to \text{Returns}$, with Experience and Market Regime independent of each other. Two backdoor paths exist: through Experience and through Market Regime. $Z = \{\text{Experience}\}$ alone leaves the Market Regime backdoor path open — bias remains. $Z = \{\text{Market Regime}\}$ alone leaves the Experience path open. Only $Z = \{\text{Experience}, \text{Market Regime}\}$ blocks both simultaneously and satisfies the criterion. This is why "control for the obvious variable" is often not enough: you need every distinct backdoor path blocked, not just one.

What this means in practice

Every regression that "controls for confounders" is implicitly attempting to satisfy the backdoor criterion, whether or not the analyst draws the DAG explicitly. Drawing the graph first is valuable precisely because it separates the mechanical question ("does my control set block all backdoor paths?") from the intuitive but unreliable habit of throwing in every available variable. The criterion also clarifies what not to control for: variables that are effects of treatment, or that create new spurious paths, should be excluded — see colliders and collider bias and bad controls and post-treatment bias.