Weak Identification Versus Underidentification

Two Different Identification Problems

An instrumental-variables model can fail to identify its structural coefficients in two importantly different ways. The instruments may fail to provide enough independent first-stage variation to identify all endogenous regressors. Alternatively, they may provide variation in every required direction, but so little relative to sampling noise that estimates and conventional inference are unreliable. The first problem is underidentification; the second is weak identification.

The distinction matters because the questions a diagnostic must answer are different. A rank diagnostic asks whether the instruments provide enough independent directions of relevance at all. A strength diagnostic asks whether the available directions are informative enough in the sample for reliable estimation and inference. A full-rank first stage does not, by itself, guarantee strong identification.

The First-Stage Rank Condition

Write the structural equation as \(y=X\beta+W\gamma+u\), where \(X\) is an \(n\times K\) matrix of endogenous regressors and \(W\) contains included exogenous controls. Let \(Z\) be an \(n\times L\) matrix of excluded instruments. After accounting for \(W\), write the population first stage as

Here, \(\Pi\) records how the excluded instruments predict the endogenous regressors, conditional on the included controls. Equivalently, define \(\tilde Z\) and \(\tilde X\) as the variables after removing their linear projections on \(W\). If \(E[\tilde Z\tilde Z']\) is nonsingular, the population first-stage coefficient matrix is

The rank condition for identifying all \(K\) endogenous-regressor coefficients is \(\operatorname{rank}(\Pi)=K\). This requires at least \(L\ge K\), but having enough instruments by count is not sufficient: their first-stage effects must span \(K\) independent directions. The rank condition addresses relevance and identification, not instrument validity. Exclusion and exogeneity remain separate assumptions.

For one endogenous regressor, \(K=1\), the condition is simply that at least one excluded instrument has a nonzero conditional first-stage coefficient. With several endogenous regressors, relevance of each instrument considered separately does not ensure that the instruments distinguish all the regressors’ independent combinations.

Why Rank Failure Prevents Unique Identification

The rank condition can be understood from the IV moment equations. Suppose the instruments are valid, so their population moments with the structural error are zero, and suppose the controls have been accounted for. The reduced-form instrument-outcome moments are then related to the structural coefficients by

When \(E[\tilde Z\tilde Z']\) is nonsingular, multiplying by its inverse reduces the identifying relationship to \(\Pi\beta\). If \(\operatorname{rank}(\Pi)<K\), there is a nonzero vector \(a\in\mathbb{R}^K\) with \(\Pi a=0\). Consequently, for any scalar \(t\),

So the instrument-outcome moments cannot distinguish \(\beta\) from \(\beta+ta\). At least one structural-coefficient direction is not uniquely identified. Conversely, when \(\Pi\) has full column rank, the map from \(\beta\) to \(\Pi\beta\) is one-to-one: \(\Pi\beta_1=\Pi\beta_2\) implies \(\Pi(\beta_1-\beta_2)=0\), and the null space contains only zero, so \(\beta_1=\beta_2\). This establishes why the rank condition separates underidentification from the possibility of identification. It does not establish that identification is strong enough for accurate finite-sample inference.

Full Rank Can Still Mean Weak Identification

Rank is a yes-or-no population property. Strength also depends on how large the relevant first-stage signal is relative to its sampling uncertainty. In models with multiple endogenous regressors, the least-informative direction matters: some combinations of regressors can be predicted well while another is predicted very poorly. The numerical size of a coefficient matrix alone is not a scale-free measure of strength, since rescaling an instrument or regressor changes its entries. Strength diagnostics must account for sampling uncertainty and the model’s scaling.

Choosing the Right Diagnostic and Reading Failures

The diagnostic question should match the problem being considered. A rank or underidentification diagnostic evaluates whether the first-stage system has enough independent instrument relevance to identify all endogenous regressors. A strength diagnostic evaluates whether the identified directions are sufficiently informative for the intended estimation and inference. The first-stage F statistic and its variants are strength diagnostics covered separately; they should not be treated as interchangeable with a rank test.

In practice, the population matrix is unknown, so analysts use sample estimates and statistical tests. A sample matrix that appears full rank does not prove that the population matrix is well separated from rank deficiency: small sampling changes may produce a full-rank estimate even when the data provide little evidence for the weakest direction. Conversely, failure to reject a rank-deficiency null is not proof that the true rank is deficient. Interpret a rank test as evidence about a specified null, not as a direct measurement of strength.

Several patterns can alert an analyst to trouble:

• There are fewer excluded instruments than endogenous regressors. The rank condition cannot hold because \(\operatorname{rank}(\Pi)\le L<K\).
• First-stage effects across endogenous regressors are nearly proportional. The matrix may be full rank in exact arithmetic, but one direction can be weak.
• One endogenous regressor, or a particular combination of them, has little instrument-predicted variation after controls are accounted for.
• Conclusions change sharply with small changes in specification or sample. This can be a warning of weak information, though it is not by itself a formal diagnosis.

A useful sequence is to check whether the model and instrument set make the rank condition possible, assess evidence about underidentification, and then examine strength using diagnostics appropriate to the number of endogenous regressors and the error structure. Neither rank nor strength diagnostics verify exclusion or exogeneity; those assumptions require substantive justification and their own assessment.

Common Mistake and Practical Limits

Wrong reasoning: “There are two instruments for two endogenous regressors, and every first-stage coefficient is nonzero, so both regressors are identified strongly.” This conclusion does not follow. Nonzero entries can still form proportional columns, producing a rank-deficient matrix. Even if the columns are independent, one direction may be so imprecisely estimated that identification is weak.

Correction: First ask whether the coefficient matrix has \(K\) independent columns. Then ask whether the least-informative direction is strong enough, relative to sampling uncertainty, for the planned inference. These are separate questions, answered by different diagnostics.

The rank condition is a necessary identification check within the stated IV setup, not a guarantee of a useful estimate. It relies on the instrument set and controls being defined appropriately, and full rank does not imply valid instruments, precise estimates, or trustworthy conventional confidence intervals. When instruments are weak, conventional 2SLS estimates and inference may behave poorly even though the formal rank condition holds. The appropriate response depends on the model and research goal; no single rank result resolves every inferential concern.