Weak Instruments and First-Stage Diagnostics

Instrumental-variables (IV) regression is the standard fix when a regressor is correlated with the error term — you replace it with an "instrument" that's correlated with the regressor but not directly with the outcome. The catch is that this trick only works cleanly if the instrument is a strong lever on the regressor. If the instrument barely moves the regressor at all, IV doesn't just get noisier — it gets systematically biased in a way that can be worse than not using an instrument at all. Checking instrument strength is not optional due diligence; it's the difference between a usable estimate and a number that looks precise but means nothing.

An analogy: a loose steering wheel

Picture trying to figure out how much turning a car's steering wheel affects its direction, but the only data you have is from a wheel that's mostly disconnected from the axle — turning it 90 degrees barely nudges the wheels. Any tiny bit of play or friction elsewhere in the linkage will dominate your reading, and you'll draw wildly wrong conclusions about the true steering ratio. A strong instrument is a tightly connected wheel: turn it, and the axle reliably responds, so whatever relationship you observe between wheel angle and direction reflects the real mechanism, not noise in the linkage.

The diagnostic, one piece at a time

IV regression has two stages. The first stage regresses the endogenous regressor $X$ on the instrument $Z$ (plus controls):

$$X_i = \pi_0 + \pi_1 Z_i + v_i .$$

Here $\pi_1$ measures how strongly the instrument moves the regressor; a weak instrument means $\pi_1$ is close to zero relative to its noise. The standard diagnostic is the first-stage F-statistic, which tests whether $\pi_1 = 0$ can be rejected — in plain English, whether the instrument reliably moves the regressor at all, using only this one equation.

$$F = \left( \frac{\hat{\pi}_1}{\text{SE}(\hat{\pi}_1)} \right)^2 .$$

This squares the first-stage t-statistic: it's large when the estimated effect of the instrument is large relative to its own uncertainty. The widely cited rule of thumb (Stock & Yogo) is that $F$ below about 10 signals a weak instrument, at which point the second-stage IV estimate can be badly biased toward the plain OLS estimate — precisely the biased number IV was meant to avoid — and its reported standard errors understate the true uncertainty.

Worked example 1: a clean strong-instrument case

A researcher wants the effect of a firm's leverage on its stock return, using an instrument — say, an exogenous change in a local tax rule that affects some firms' borrowing costs but not their returns directly. The first-stage regression of leverage on the tax-rule instrument gives $\hat{\pi}_1 = 0.85$ with $\text{SE}(\hat{\pi}_1) = 0.10$. The F-statistic is $(0.85/0.10)^2 = 72.25$, well above 10 — the instrument moves leverage strongly and reliably, so the second-stage IV estimate can be trusted at face value.

Worked example 2: a weak-instrument warning sign

Now suppose a different candidate instrument — same-industry average leverage in a distant region — gives $\hat{\pi}_1 = 0.04$ with $\text{SE}(\hat{\pi}_1) = 0.02$. The F-statistic is $(0.04/0.02)^2 = 4$, well below 10. Even if the second-stage regression spits out a seemingly precise-looking coefficient with a small standard error, that precision is illusory: simulations show that with $F=4$, the IV estimate's true sampling distribution is badly skewed and can be centered far from the truth. The correct response isn't to report the number with a caveat — it's to find a stronger instrument or fall back to a design that doesn't require one.

What this means in practice

Any quant running IV — instrumenting order-flow toxicity, corporate actions, or macro shocks with something plausibly exogenous — should report the first-stage F-statistic alongside the main result, not just the second-stage coefficient. A high F-stat is a prerequisite for trusting the point estimate at all, not a nice-to-have robustness check. When instruments are weak, better options are finding a stronger instrument, using weak-instrument-robust confidence intervals (e.g., Anderson-Rubin), or reporting that identification has failed rather than presenting a biased number with a false sense of precision.