The Concentration Parameter and Instrument Strength

Why First-Stage Information Matters

An instrument can satisfy the conditions needed for identification and still provide little information about the endogenous regressor. When the first-stage relationship is weak, the part of the regressor predicted by the instruments contains substantial sampling noise relative to its systematic signal. The resulting 2SLS estimate can be unstable even when the sample is large enough to produce a numerical estimate.

The concentration parameter summarizes that signal-to-noise relationship in a linear first stage. It is useful because it connects the strength of the first stage to the sampling distribution of the first-stage F statistic and helps explain why weak instruments can cause finite-sample problems for IV estimation. It is not itself an observed test statistic, and it does not establish that an instrument is valid.

Definition in the Linear IV Model

Consider one endogenous regressor \(D\), an outcome \(Y\), a vector of included exogenous controls \(W\), and a vector of \(q\) excluded instruments \(Z\). Write the structural equation and first stage as

Y = \beta D + W\gamma + u, \qquad D = W\delta + Z\pi + v.

Here \(\beta\) is the structural effect of interest, \(u\) is the structural error, and \(v\) is the first-stage error. The vector \(\pi\) describes how the excluded instruments predict \(D\). For the standard concentration-parameter and F-statistic relationship below, assume a correctly specified linear first stage, a full-rank instrument design after controlling for \(W\), and conditionally independent, normally distributed first-stage errors with mean zero and constant variance \(\sigma_v^2\). Instrument validity for estimating \(\beta\) additionally requires the relevant exclusion and exogeneity conditions, including that \(Z\) is uncorrelated with \(u\) conditional on \(W\). Relevance alone is not validity.

Let \(M_W=I-P_W\) denote the residual-maker for the columns of \(W\), where \(P_W\) projects onto their span. Thus \(M_WZ\) is the instrument matrix after removing the linear association with the included controls. Conditional on the design, the concentration parameter is

\mu^2 = \frac{\pi' Z'M_WZ\pi}{\sigma_v^2}.

Definition: The concentration parameter \(\mu^2\) is the squared first-stage signal from the excluded instruments, after controlling for \(W\), divided by the first-stage error variance. It is nonnegative; larger values indicate more first-stage information in this model.

The numerator is the squared length of the instrument-predicted component \(M_WZ\pi\). Dividing by \(\sigma_v^2\) expresses that signal relative to the noise in \(D\). The value is conditional on the instrument and control design. It is not the sample partial \(R^2\), although both describe first-stage strength. In particular, \(\mu^2\) depends on the amount of instrument variation and the scale of first-stage noise, not just on a coefficient viewed in isolation.

Connection to the First-Stage F Statistic

In the first-stage regression, the usual F statistic tests \(H_0:\pi=0\), the joint null that the excluded instruments have no first-stage effect. Under the normal, constant-variance model above, conditional on the design and when \(H_0\) is false, this F statistic has a noncentral F distribution:

F \sim F_{q,\,d}(\mu^2), \qquad d=n-\operatorname{rank}(W)-q.

The notation \(F_{q,d}(\mu^2)\) means a noncentral F distribution with numerator degrees of freedom \(q\), denominator degrees of freedom \(d\), and noncentrality parameter \(\mu^2\). Under \(H_0\), the noncentrality parameter is zero, giving the central F distribution. Under an alternative, a larger \(\mu^2\) shifts the distribution toward larger F values. The observed F is random; the concentration parameter describes the distribution that generates it under the specified design and model.

This relationship explains why the concentration parameter is a measure of first-stage information. It also clarifies a common confusion: the observed F statistic is not the concentration parameter. An observed F may be unusually high or low relative to its distribution, and the distribution itself depends on degrees of freedom as well as on \(\mu^2\).

Worked Example: Calculating Concentration With One Instrument

Suppose a hypothetical study has one excluded instrument \(Z\), with \(n=200\) observations. After removing the intercept and other included controls, suppose the instrument values have sum of squares \(Z'M_WZ=200\). The first-stage coefficient is \(\hat{\pi}=0.15\) units of \(D\) per instrument unit, and the first-stage error variance is \(\sigma_v^2=0.36\) square units of \(D\). For this calculation, treat the stated coefficient and variance as the model values.

With one instrument, the concentration parameter is

\mu^2 = \frac{\pi^2 Z'M_WZ}{\sigma_v^2} = \frac{(0.15)^2(200)}{0.36} = \frac{4.50}{0.36} = 12.5.

The value \(12.5\) summarizes the conditional first-stage signal relative to first-stage noise. It is not an observed F statistic and, by itself, does not establish that the instrument is valid. If the same design and coefficient were observed in four times as many independent observations, so that the residualized sum of squares were \(800\), the parameter would be

\mu^2 = \frac{(0.15)^2(800)}{0.36} = 50.

The greater value reflects more accumulated first-stage information, not a change in the per-unit first-stage coefficient.

A Short Derivation: Why Sample Size Accumulates Information

The first-stage model after removing \(W\) is \(M_WD=M_WZ\pi+M_Wv\). The systematic component has squared length \(\pi'Z'M_WZ\pi\). Under the homoskedastic model, the error variance per observation is \(\sigma_v^2\), so the systematic component measured in error-variance units is precisely \(\mu^2\). This is the source of the noncentrality parameter in the joint test of \(\pi=0\).

To see how the parameter changes with sample size, suppose observations are independent, the instrument distribution and first-stage coefficients remain fixed, and the first-stage variance remains \(\sigma_v^2\). Then the residualized instrument cross-product typically grows proportionally to \(n\). Consequently \(\mu^2\) also grows proportionally to \(n\). This is the strong-identification large-sample setting: the first-stage signal accumulates as observations are added.

A contrasting sequence lets the first-stage coefficient shrink as the sample grows. For a single instrument with residualized sum of squares proportional to \(n\), if \(\pi=c/\sqrt{n}\), then

\mu^2 \ \propto\ \frac{n(c/\sqrt{n})^2}{\sigma_v^2} = \frac{c^2}{\sigma_v^2}.

The concentration parameter stays bounded rather than increasing without limit. In this weak-instrument setting, adding observations does not make the first-stage signal dominate the noise in the usual way. This is one reason large-sample approximations for 2SLS can be unreliable when the first stage is weak.

How Weak Signal Affects 2SLS

After partialling out the included controls, let \(P_Z\) denote projection onto the residualized excluded-instrument space. The 2SLS estimation error for a single endogenous regressor can be written as

\hat{\beta}_{2SLS}-\beta = \frac{D'M_WP_ZM_Wu}{D'M_WP_ZM_WD},

where the expression uses the original variables with the controls partialled out. Equivalently, if \(D\) and \(u\) are already residualized with respect to \(W\), it is \((D'P_Zu)/(D'P_ZD)\). The denominator measures the instrument-predicted variation in \(D\); the numerator reflects the sample association between that predicted variation and the structural error.

When first-stage signal is strong, systematic prediction of \(D\) contributes substantially to the denominator. When signal is weak, the denominator is more affected by first-stage sampling variation. A sample association between the instruments and \(u\), which can occur by chance even if instruments are exogenous in the population, can then have a large influence on the ratio. Across repeated samples, the resulting finite-sample distribution can be skewed and concentrated toward the OLS estimate under common endogeneity patterns. This is a general tendency, not a guarantee about the direction of bias in every design or sample.

Worked Example: A Weak Denominator Amplifies Sampling Noise

Consider a hypothetical centered sample with one instrument \(z\). Let \(z'z=4\), and suppose the residualized first-stage fitted component is \(0.10z\). Thus the fitted component has squared length \(0.10^2(4)=0.04\). Take the structural effect to be \(\beta=2\), and suppose the realized sample has \(z'u=0.02\). This sample covariance is possible by chance even when the population instrument is exogenous.

In the just-identified centered case, the IV estimate is the ratio of the instrument-outcome covariance to the instrument-treatment covariance. Since \(D=0.10z+v\) and the fitted component is the part relevant to the instrument covariance, the corresponding sample ratio is

\hat{\beta}_{IV} = \beta+\frac{z'u}{z'D} = 2+\frac{0.02}{0.10(4)} = 2+\frac{0.02}{0.40} = 2.05.

The modest sample covariance \(z'u=0.02\) changes the estimate by \(0.05\) outcome units per unit of \(D\). If the fitted first-stage component were instead \(0.02z\), with the same \(z'z\) and \(z'u\), the change would be

\frac{0.02}{0.02(4)} = 0.25, \qquad \hat{\beta}_{IV}=2.25.

This calculation illustrates sensitivity of the ratio to a small first-stage denominator. It is not a calculation of expected bias: one realized sample cannot determine the repeated-sampling bias, and the example does not imply that every weak-instrument estimate moves in the same direction.

Worked Example: The F Statistic Is Random Around Its Strength-Dependent Distribution

Suppose a hypothetical first stage has \(q=2\) excluded instruments, \(n=100\), one included intercept, and residualized instrument cross-product matrix

Z'M_WZ = \begin{pmatrix} 100 & 0\\ 0 & 100 \end{pmatrix}, \qquad \pi= \begin{pmatrix} 0.12\\ 0.08 \end{pmatrix}, \qquad \sigma_v^2=0.16.

The concentration parameter is

\mu^2 = \frac{100(0.12^2)+100(0.08^2)}{0.16} = \frac{1.44+0.64}{0.16} = 13.

The denominator degrees of freedom are \(d=100-1-2=97\). Under the stated normal, homoskedastic model, the first-stage F statistic therefore has distribution \(F_{2,97}(13)\). The expected value of this noncentral F distribution is

E(F) = \frac{97}{97-2}\frac{2+13}{2} = \frac{97}{95}(7.5) \approx 7.658.

This is a distributional expectation, not the observed F statistic for a particular sample. A realized F can differ from \(7.658\), and this expected value is not a universal threshold for judging strength. The calculation depends on the assumed model and degrees of freedom.

Interpretation, Diagnostics, and Limits

The earlier material on partial \(R^2\) describes the fraction of residual variation in \(D\) explained by the excluded instruments after controlling for \(W\). The concentration parameter complements that measure by incorporating the scale of first-stage noise and the amount of residualized instrument variation. Neither summary should be confused with a test of instrument validity, which concerns the relationship between the instrument and the structural error and cannot be established by a strong first stage.

In applied work, \(\mu^2\) is generally not known because the population first-stage coefficients and error variance are unknown. Researchers calculate sample diagnostics, such as the observed first-stage F statistic and partial \(R^2\), as evidence about strength. Under the classical homoskedastic model, the noncentral-F connection explains how such an observed statistic relates to a strength-dependent distribution. Outside that model, the relationship may not apply as stated; the form of the first-stage test and its variance assumptions matter.

Several limitations are important. First, the parameter is conditional on a specified design and linear first stage; it is not a model-free measure. Second, a large value does not verify exclusion or exogeneity. Third, a single observed diagnostic is noisy and cannot identify the unknown concentration parameter exactly. Finally, with multiple endogenous regressors or more complicated error structures, strength assessment requires additional methods; the scalar expression here should not be applied as though it fully describes those settings.

Common Mistakes

Calling the observed F the concentration parameter. The F is a random statistic; \(\mu^2\) is the noncentrality parameter that governs its distribution under the stated model.
Reading a large concentration parameter as proof of validity. It concerns relevance and first-stage information, not whether the instrument is independent of the structural error or satisfies exclusion.
Comparing coefficients without considering scale. A first-stage coefficient alone does not determine strength; instrument variation and the first-stage error variance also enter \(\mu^2\).
Assuming more observations always solve weak identification. Under a fixed nonzero first-stage coefficient, information typically accumulates with \(n\). Under a sequence of shrinking coefficients, \(\mu^2\) can remain bounded as \(n\) increases.
Treating finite-sample bias as a fixed direction in every sample. Weak instruments make estimates sensitive and can create bias toward OLS in common settings, but an individual estimate can move in either direction.

Key takeaway: The concentration parameter measures first-stage signal relative to noise, conditional on the design. Its size helps explain the sampling behavior of IV estimators, but it is not an observed statistic, a universal cutoff, or a test of instrument validity.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A hypothetical single-instrument design has \(Z'M_WZ=150\), \(\pi=0.10\), and \(\sigma_v^2=0.25\). Calculate \(\mu^2\).
With \(q=3\), \(n=120\), and \(\operatorname{rank}(W)=2\), what are the numerator and denominator degrees of freedom for the classical first-stage F distribution?
Why can an observed first-stage F statistic not be interpreted as the concentration parameter?
If the first-stage coefficient shrinks proportional to \(1/\sqrt{n}\) while residualized instrument variation grows proportional to \(n\), what happens to the concentration parameter?
Does a large concentration parameter establish that an instrument is valid? Explain.

Solutions. (1) \(\mu^2=(0.10^2)(150)/0.25=1.50/0.25=6.00\). (2) The numerator degrees of freedom are \(q=3\); the denominator degrees of freedom are \(120-2-3=115\). (3) The observed F is a random sample statistic. Under the classical model, its distribution has noncentrality parameter \(\mu^2\); sampling variability means the realized F need not equal that parameter. (4) The squared coefficient is proportional to \(1/n\), which cancels the order-\(n\) growth in residualized instrument variation, so \(\mu^2\) remains bounded, assuming the error variance is fixed. (5) No. A large value indicates first-stage information under the specified model. Validity also requires appropriate exogeneity and exclusion conditions involving the structural error.