The Exclusion Restriction and Instrument Validity

A researcher wants the causal effect of margin-account access on a trader's returns, and uses the trader's zip code's local bank-branch density as an instrument, reasoning that more nearby bank branches makes it easier to get approved for a margin account, which affects returns only through margin access. This works only if bank-branch density has no other channel to returns — but branch density also correlates with local wealth, which affects returns through completely different channels (better information access, wealthier social networks, other credit products) that have nothing to do with margin accounts. If any such side-channel exists, the instrument is invalid, no matter how strong its correlation with margin-account access is — this is the exclusion restriction, and it's the assumption instrumental variables lives or dies on.

An analogy: a courier who's only supposed to deliver one package

Imagine hiring a courier whose sole job is to move package A (the treatment) from sender to recipient, and you want to measure how much A's arrival affects the recipient's later behavior. If the courier, unbeknownst to you, is also handing the recipient package B (some other influence) on the same trip, then anything you observe about the recipient's behavior can't be cleanly attributed to package A alone — some of it might be B's doing. The exclusion restriction is the requirement that the courier delivers only the one package you're studying; if there's a second, unaccounted-for delivery riding along, your whole measurement is contaminated and there's no way to see it from the recipient's behavior alone.

The assumption, one symbol at a time

For instrument $Z_i$, treatment $D_i$, and outcome $Y_i$, potential outcomes $Y_i(d, z)$ denote what $Y_i$ would be under treatment level $d$ and instrument value $z$. The exclusion restriction states:

$$Y_i(d, z) = Y_i(d, z') \quad \text{for all } z, z'$$

In plain English: fixing the treatment level $d$, the outcome doesn't depend on which value the instrument took — the instrument only ever reaches $Y$ by first changing $D$, never directly, and never through some other back-channel. Combined with relevance ($Z$ actually predicts $D$) and the ignorability of $Z$ itself, this is what lets the simple ratio in the Wald/IV estimator (see the local average treatment effect) isolate a causal effect of $D$ on $Y$ rather than a mix of $D$'s effect and the instrument's own direct effect.

Worked example 1: a violation you can reason about, even without testing it

Suppose local bank-branch density genuinely predicts margin-account approval (a valid first stage — relevance holds), but it also independently correlates with local financial literacy programs, which improve trading decisions through channels with nothing to do with margin access. The IV estimate of margin access's effect on returns would then partly reflect financial literacy's direct effect, smuggled in through the correlation between branch density and literacy programs. No F-statistic or first-stage regression can detect this — it's invisible in the data actually used to estimate the model, because financial literacy isn't measured. The only way to catch it is domain reasoning: does branch density plausibly affect returns through any channel other than margin access? Here, plausibly yes, and that alone should disqualify the instrument, regardless of how strong the branch-density-to-margin-access correlation is.

Worked example 2: a case built to satisfy exclusion — and its remaining risk

Now suppose the researcher instead uses a firm-specific policy change — a temporary system glitch that randomly locked some clients out of margin-account applications for two weeks — as the instrument. This is more defensible: the glitch plausibly affects returns only by delaying margin access. Suppose relevance holds strongly and the resulting LATE estimate is +1.8% return per quarter. Even here, exclusion could still fail subtly — if the same glitch also delayed other account features (say, real-time quote feeds) that independently affect returns, exclusion is violated again. Even a seemingly "clean" instrument requires accounting for every system it touched, not just the one you intended to study.

What this means in practice

Unlike relevance, which can be checked directly with a first-stage regression (see weak instruments and first-stage diagnostics), the exclusion restriction is fundamentally untestable from the data used to fit the model — it's an assumption about the absence of channels you haven't observed, defended with subject-matter argument, not a statistic. The strongest instruments in applied causal work (natural experiments, policy glitches, weather shocks) earn their credibility precisely because their designers can plausibly rule out alternative channels one by one, not because any test confirmed validity.