Many-Instrument Bias in Two-Stage Least Squares

Why Instrument Count Matters

Two-stage least squares (2SLS) uses instruments to isolate variation in an endogenous regressor that can support causal estimation. Adding instruments can seem helpful: each additional instrument may explain more of the treatment or exposure. But a larger instrument set also gives the first stage more freedom to fit the observed sample. It can fit not only genuine instrument-driven variation but also random variation in the endogenous regressor that is correlated with the structural error. As a result, finite-sample 2SLS can be biased toward the ordinary least squares (OLS) estimate.

This is a finite-sample problem, not a claim that a valid instrument becomes invalid when other valid instruments are added. Population validity concerns the relationship between the instruments and the structural error. Many-instrument bias concerns how a sample projection behaves when there are many instrument directions relative to the information in the sample.

Definition: In a model with one endogenous regressor and \(K\) excluded instruments, many-instrument bias refers to the tendency of finite-sample 2SLS to be pulled toward OLS as the number of instruments becomes large relative to the sample size or the available first-stage signal. The tendency is especially important when instruments are weak. It is a bias tendency under specified conditions, not a guarantee about the direction of every realized estimate.

The Projection Mechanism

Consider a linear structural equation \(y=\beta x+u\), where \(x\) is endogenous and \(u\) is the structural error. Let \(Z\) be the matrix of excluded instruments, with \(K\) linearly independent columns after any included controls have been accounted for. For clarity, the equations below are written after residualizing the variables on those controls. Define the projection matrix \(P_Z=Z(Z'Z)^{-1}Z'\). The fitted values from the first stage are \(\hat{x}=P_Zx\).

The 2SLS estimator can be written as \(\hat{\beta}_{2SLS}=(x'P_Zy)/(x'P_Zx)\). Substituting the structural equation gives the exact identity

\hat{\beta}_{2SLS}-\beta = \frac{x'P_Zu}{x'P_Zx}.

Instrument validity means that the instruments satisfy the relevant population conditions, including exclusion from the structural error and sufficient relevance for identification. In a simple unconditional formulation, exclusion is \(E[Z_i u_i]=0\). This does not mean that the realized sample cross-product \(Z'u\) must equal zero. Even with valid instruments, a particular sample can have nonzero chance correlations.

To see why instrument count matters, write the first-stage model as \(x=Z\pi+v\), where \(Z\pi\) is the instrument-driven component and \(v\) is the part not explained by the instruments. Endogeneity means that \(v\) may be correlated with \(u\). In a sample, \(P_Zv\) is generally nonzero: the instrument space can project some of the noise \(v\) onto its fitted values. If \(v\) and \(u\) are correlated, this projected noise can also carry information about \(u\), even though the instruments satisfy the population exclusion condition.

The more independent directions in \(Z\), the more sample variation can be projected onto the first-stage fitted values. This is the overfitting mechanism. It is analogous to fitting a flexible prediction model: in-sample fit can improve while some of the improvement comes from chance features of this sample rather than stable signal.

A Projection-Based Bias Approximation

A simple derivation makes the role of \(K\) concrete. Suppose the instruments are treated as fixed, the model is homoskedastic, and, conditional on \(Z\), \(v\) and \(u\) have mean zero with \(E[v_i u_i\mid Z]=\sigma_{vu}\) and \(E[v_i^2\mid Z]=\sigma_v^2\). These are convenient conditions for this illustration; the resulting approximation is not an exact finite-sample formula for all designs.

Using \(x=Z\pi+v\), the numerator of the 2SLS estimation error is

x'P_Zu=\pi'Z'u+v'P_Zu.

Under the stated conditional mean assumptions, the first term has conditional expectation zero. For the second term, write it as a sum of products \(v_i(P_Zu)_i\). The diagonal elements of \(P_Z\) contribute the within-observation covariance \(\sigma_{vu}\); off-diagonal terms have expectation zero under independent sampling. Therefore,

E[x'P_Zu\mid Z] = \sigma_{vu}\operatorname{tr}(P_Z) = K\sigma_{vu}.

The last equality follows because a projection matrix has trace equal to its rank, here \(K\). Similarly, the expected denominator is

E[x'P_Zx\mid Z] = \pi'Z'Z\pi+K\sigma_v^2.

The first term is the instrument-driven signal in the sample, and the second is projected first-stage noise. Replacing the expectation of a ratio by a ratio of expectations gives the useful schematic approximation

\operatorname{Bias}(\hat{\beta}_{2SLS}) \ \approx\ \frac{K\sigma_{vu}}{\pi'Z'Z\pi+K\sigma_v^2}.

This replacement is an approximation because, in general, the expectation of a ratio is not the ratio of its expectations. Still, the expression displays the mechanism: more instrument directions increase the projected-noise contribution, while strong instrument-driven signal can reduce its influence.

For comparison, in this same stylized model the leading OLS bias is approximately \(n\sigma_{vu}/(\pi'Z'Z\pi+n\sigma_v^2)\). When the first-stage signal dominates the noise terms in both denominators, the ratio of the 2SLS bias to the OLS bias is approximately \(K/n\). This motivates the familiar \(K/n\) relative-bias heuristic: under weak-instrument and homoskedastic approximations, relative bias tends to grow with the number of instruments as a share of the sample size. It is a guide to the direction and scale of the concern, not a universal correction factor. With very weak signal, substantial heteroskedasticity, or a different design, the approximation can be poor.

Worked Examples

Worked Example: Comparing Instrument Counts

Suppose a hypothetical study has \(n=500\) observations. Consider two otherwise comparable instrument sets, one with \(K=5\) instruments and the other with \(K=80\). Assume, only for this approximation, that the instruments are valid, the model is homoskedastic, and the weak-instrument relative-bias heuristic is appropriate. Suppose the OLS estimate has a bias of \(0.30\) outcome units per treatment unit.

For \(K=5\), the instrument-to-sample ratio is \(5/500=0.01\). The heuristic gives a 2SLS bias of about \(0.01(0.30)=0.003\) outcome units per treatment unit.

For \(K=80\), the ratio is \(80/500=0.16\), giving an approximate bias of \(0.16(0.30)=0.048\) outcome units per treatment unit.

These calculations do not predict the exact bias in either sample. They illustrate how, under the stated approximation, increasing the instrument count can make the 2SLS bias a larger share of the OLS bias—even if all instruments are individually valid.

Worked Example: A Sample Projection Can Move Toward OLS

Consider a small, purely hypothetical sample represented in four residualized coordinates, with no included controls. Let the endogenous regressor be \(x=(2,3,2,1)'\), and let the structural error be \(u=(1,-1,-1.3,0)'\). Suppose the structural slope is \(\beta=2\), so \(y=2x+u\). Compare using the first instrument coordinate alone with using the first two coordinate instruments together. This is an algebraic illustration of a sample projection; these four realized values alone cannot establish population validity.

With the first coordinate instrument, the projection retains only the first coordinate. Thus \(x'P_Zu=2(1)=2\), and \(x'P_Zx=2^2=4\). The 2SLS estimation error is \(2/4=0.5\), so the estimate is \(2+0.5=2.5\).

With both coordinate instruments, the projection retains the first two coordinates. Now \(x'P_Zu=2(1)+3(-1)=-1\), while \(x'P_Zx=2^2+3^2=13\). The estimation error is \(-1/13\approx-0.07692\), giving an estimate of about \(1.923\).

For comparison, OLS has estimation error \(x'u/(x'x)\). Here \(x'u=2-3-2.6=-3.6\) and \(x'x=4+9+4+1=18\), so the OLS error is \(-3.6/18=-0.2\), and the OLS estimate is \(1.8\). Adding the second instrument moved this particular 2SLS estimate from \(2.5\) toward \(1.8\). This example demonstrates how sample projections can change the estimate; it does not show that adding an instrument must move every estimate toward OLS.

Worked Example: More First-Stage Fit Is Not Necessarily More Signal

Suppose a hypothetical data-generating process has a true structural slope of \(1.00\) outcome units per treatment unit and an OLS bias of \(+0.40\) in those units. A set of \(K=3\) valid instruments produces a first-stage \(R^2\) of \(0.10\). A larger set of \(K=30\) instruments produces an \(R^2\) of \(0.42\) in the same \(n=120\) observations. Assume the weak-instrument relative-bias heuristic is reasonable for illustration.

For the smaller set, \(K/n=3/120=0.025\). The approximate 2SLS bias is \(0.025(0.40)=0.010\), so the corresponding approximate estimate is \(1.00+0.010=1.010\).

For the larger set, \(K/n=30/120=0.25\). The approximate bias is \(0.25(0.40)=0.100\), giving an approximate estimate of \(1.100\). Although the first-stage \(R^2\) rose from \(0.10\) to \(0.42\), that increase alone does not show that the added instruments contribute stable, exogenous signal. Some of the additional in-sample fit can reflect noise. The numerical bias values are illustrative approximations, not estimates recoverable from the reported \(R^2\) values alone.

Interpretation and Practical Implications

Many-instrument bias matters because a 2SLS estimate can look precise or seem to benefit from a strong in-sample first stage while still being pulled toward the confounded OLS relationship. The key issue is not simply whether each instrument passes a validity argument. The number of instrument directions, the sample size, the strength of their combined signal, and the dependence between the first-stage disturbance and the structural error all affect the finite-sample behavior.

The \(K/n\) heuristic is most useful as a warning signal, not a decision rule. It does not tell an analyst to multiply an estimate by a correction factor, nor does it establish that instruments are invalid. Nor can a high first-stage \(R^2\) by itself distinguish true predictive signal from overfitting. Instrument validity remains a substantive and design-based question; finite-sample bias is a separate estimation concern.

Common Mistake

Wrong reasoning: “Every instrument is individually valid, and adding instruments increased first-stage \(R^2\), so the 2SLS estimate must be less biased.”
Correction: Population validity does not prevent chance sample correlations between the projected first-stage noise and the structural error. Adding instruments increases the dimension of the space used to fit \(x\), which can increase overfitting and finite-sample bias. A higher in-sample \(R^2\) is not proof that the extra fit comes from useful population signal.
Wrong reasoning: “The \(K/n\) approximation gives the exact bias for any sample.”
Correction: It is a heuristic derived under restrictive approximations. Its accuracy depends on the design, instrument strength, error structure, and the relative amount of signal and noise.

Key takeaway: With many instruments, the first stage can fit more of the endogenous regressor’s sample noise. Under common finite-sample approximations this can pull 2SLS toward OLS, even when the instruments satisfy population validity conditions. Instrument count and in-sample fit should therefore be interpreted alongside substantive validity and the available first-stage signal.

Check Your Understanding

Try these questions before reviewing the explanation above.

Why does population instrument validity not require the realized sample cross-product \(Z'u\) to equal zero?
In the projection derivation, what does \(\operatorname{tr}(P_Z)=K\) represent?
Under the \(K/n\) heuristic, what happens to approximate relative bias if \(n\) stays fixed and \(K\) increases?
Why can adding instruments raise first-stage \(R^2\) without providing stronger population signal?
What assumptions make the bias calculation a heuristic rather than an exact formula?