Weak Identification in Linear Instrumental Variables

Why a Relevant Instrument May Still Be Weak

An instrument can be relevant in the population and still provide little useful information in a particular sample. The key is that relevance is not merely a yes-or-no property for practical inference: its size matters relative to sampling noise. When the first-stage relationship is small, a larger sample may be needed to distinguish it from noise, while the reduced-form relationship used to learn about the treatment effect may also be difficult to detect.

This distinction matters because the IV estimate is a ratio: it divides the reduced-form relationship between the instrument and outcome by the first-stage relationship between the instrument and treatment. If the denominator is imprecisely estimated near zero, a seemingly ordinary ratio can have a highly irregular sampling distribution. Weak identification is therefore a problem of limited information about the parameter, not simply a label for an instrument that fails a particular test.

Setup: First Stage, Reduced Form, and Identification

Consider a linear model with one endogenous treatment \(D\), one instrument \(Z\), and an outcome \(Y\). Any included covariates are understood to have been adjusted for; equivalently, the following relationships can be expressed using variables residualized on those covariates. For clarity, take the variables to be centered and write the structural equation as \(Y=\beta D+u\). The coefficient \(\beta\) is the causal parameter under the model's structural and instrument assumptions.

The covariance condition is stated explicitly because a zero uncentered product moment, \(E[Zu]=0\), does not by itself imply \(\operatorname{Cov}(Z,u)=0\) unless additional mean assumptions hold. Here the variables are centered, and validity is expressed directly as covariance orthogonality. In applied models with covariates, the corresponding condition applies to the adjusted instrument and structural error.

The identification relationship follows from substituting the structural equation into the reduced form. It is useful to derive it rather than treat it as a separate assumption.

If \(\pi\ne 0\), then \(\beta=\gamma/\pi\) is identified from these population relationships. If \(\pi=0\), the instrument provides no first-stage variation with which to separate \(\beta\) from the structural error using this moment. But \(\pi\ne0\) alone does not ensure precise estimation: the slope could be so small that the sample contains little information about its value.

Weak Identification as a Signal-to-Noise Problem

A first-stage slope is informative only in relation to how precisely it can be estimated. For illustration, suppose the first stage is \(D=\pi Z+v\), with centered variables, independent observations, and a homoskedastic first-stage error \(v\) with standard deviation \(\sigma_v\). In a large sample, the standard error of the estimated first-stage slope is approximately \(\sigma_v/\sqrt{n\operatorname{Var}(Z)}\). Thus its signal relative to its standard error is approximately

This expression shows why a nonzero \(\pi\) can be weak in finite samples: a small slope, a small sample, or substantial first-stage noise can all leave the signal hard to distinguish from zero. Increasing \(n\) can improve the signal-to-noise ratio, but a very small first-stage slope may require a large increase in sample size.

The reduced form carries related information. Under validity, \(\gamma=\beta\pi\). If \(\beta\) is fixed while \(\pi\) is small, the reduced-form slope is also small. Under the linear first stage \(D=\pi Z+v\), the reduced-form regression error is \(\beta v+u\), not generally \(u\) alone. With independent observations, a homoskedastic reduced-form error orthogonal to \(Z\), and \(\sigma_{RF}=\operatorname{SD}(\beta v+u)\), the large-sample standard error of the reduced-form slope is approximately \(\sigma_{RF}/\sqrt{n\operatorname{Var}(Z)}\). Its signal-to-noise ratio is therefore approximately \(\sqrt{n}|\beta\pi|\sqrt{\operatorname{Var}(Z)}/\sigma_{RF}\). The reduced form may consequently be hard to distinguish from zero as well.

These approximations describe how information scales; they do not define a universal cutoff between “weak” and “strong.” Weak identification means that the available first-stage and reduced-form information is insufficient for the usual approximations to the IV estimator to work reliably. In an asymptotic sequence, the issue is especially clear when the first-stage slope approaches zero as the sample grows, for example when \(\pi_n=c/\sqrt{n}\) for a fixed constant \(c\). In that case, the first-stage signal need not grow without bound, so the usual large-sample approximation based on a precisely estimated denominator can fail.

What Makes the IV Ratio Unstable

In a sample, the just-identified IV estimator can be written as \(\hat{\beta}_{IV}=\hat{\gamma}/\hat{\pi}\), where \(\hat{\gamma}\) and \(\hat{\pi}\) are the sample reduced-form and first-stage slopes. When \(\hat{\pi}\) is estimated precisely away from zero, the ratio can often be approximated using ordinary large-sample reasoning. When the first stage is weak, sampling variation can make \(\hat{\pi}\) close to zero, even if the population value is not exactly zero. Dividing by such a denominator can produce very large estimates or a distribution that is far from the familiar bell-shaped approximation.

This is why a weak first stage can make conventional confidence intervals and tests for \(\beta\) unreliable. The problem is not simply that the estimate is “noisy” in the usual way: the uncertainty in the denominator affects the shape of the estimator's distribution. A point estimate can look plausible while the data provide little stable information about the underlying parameter. The appropriate response is to assess the strength of identification and use inference designed for weak identification when warranted; the particular diagnostics and procedures are separate topics.

Weak, Underidentified, and Invalid Are Different

These terms describe distinct problems. Underidentification is the population failure of the available instrument variation to identify the parameter in the specified model. Weak identification concerns limited information: the population relationship may be nonzero, but it is too small relative to sampling noise for reliable conventional inference. Invalidity is a failure of an instrument assumption, such as \(\operatorname{Cov}(Z,u)=0\) or the exclusion restriction. A large first-stage slope does not repair invalidity, and validity does not guarantee a strong first stage.

Common Mistake and Practical Interpretation

A common mistake is to reason, “The first-stage coefficient is not exactly zero, so the instrument is strong and the IV estimate is identified well.” The correction is that a nonzero population slope establishes relevance, not adequate sample information. One must consider the magnitude of the first-stage relationship, sample size, variation in the instrument, and the noise in the first stage. Even then, strength says nothing by itself about validity.

A related mistake is to treat a small reduced-form estimate as proof that the causal effect is small. Since \(\gamma=\beta\pi\) under validity, a small reduced form may reflect a small \(\beta\), a weak first stage, or both. Conversely, a large estimated IV ratio does not establish a large effect if it is produced by dividing by a poorly estimated denominator. Interpret the estimate together with the identification information and the instrument assumptions, rather than as a standalone ratio.