Why Use a Score Test for an IV Effect?
In an instrumental-variable (IV) analysis, a conventional confidence interval can be unreliable when the instrument predicts treatment only weakly. A score-based test offers another way to test a hypothesized causal effect without relying on a precise, approximately normal estimate of the IV coefficient. Its central question is whether the data provide evidence against a specified effect, after accounting for nuisance parameters such as the first-stage coefficients.
This tutorial focuses on the weak-IV-robust score test commonly associated with Kleibergen’s approach. It does not mean that any test labelled “score” is automatically robust to weak instruments. The test’s construction matters: it evaluates the structural-effect restriction while allowing the first-stage parameters to be nuisance parameters, rather than treating an estimated first stage as known.
The coefficient \(\beta\) represents a causal effect only under substantive causal assumptions. In a constant-effect structural model, these include a correctly specified structural relation, consistency, no interference, instrument relevance for point identification, instrument exogeneity, and exclusion of any pathway from \(Z\) to \(Y\) other than through \(D\). With heterogeneous effects, a causal interpretation requires a design-specific estimand and assumptions—for example, monotonicity when interpreting an IV estimand as a local average treatment effect. A weak-IV-robust test addresses uncertainty from weak relevance; it does not verify these causal assumptions.
Hypothesis, Nuisance Parameters, and Score
The null is a value of the structural effect, not a null hypothesis that the instrument is irrelevant. For a proposed value \(\beta_0\), form the transformed outcome \(Y-\beta_0D\). Under the structural null,
The excluded instruments should have no effect on this transformed outcome under the null. In reduced-form language, if the coefficients of \(Z\) in the outcome and treatment equations are \(\gamma\) and \(\pi\), respectively, the restriction is \(\gamma-\beta_0\pi=0\). The first-stage vector \(\pi\) is not assumed known; it is a nuisance parameter.
A score test measures how strongly the data point away from the null restriction, evaluated at estimates that satisfy the null. The nuisance parameters are fitted under that restriction. The score is then adjusted so that variation attributable to estimating those nuisance parameters is not mistaken for evidence against \(H_0\). This adjustment is essential: the first-stage estimates and the structural restriction are statistically related.
In a regular likelihood model, let \(U_\beta\) be the score for the parameter of interest and \(U_\eta\) the score for nuisance parameters \(\eta\), evaluated under the null-constrained fit. Let \(I\) denote the information matrix, partitioned into corresponding blocks. The nuisance-adjusted score and its information are
The score statistic is the squared, information-scaled departure from the null. For a \(q\)-dimensional restriction, its general form is \(U_{\mathrm{eff}}^{\mathsf T}I_{\mathrm{eff}}^{-1}U_{\mathrm{eff}}\). In weak-IV applications, the specialized reduced-form score construction and its covariance estimate are chosen to retain validity when first-stage coefficients are small. The displayed likelihood expression explains nuisance adjustment; by itself it is not a license to use an arbitrary classical score test for IV.
Worked Calculations
Worked Example: Testing a Hypothesized Treatment Effect
Suppose a hypothetical study uses two excluded instruments to estimate the effect of a treatment on a continuous outcome, measured in outcome points. The structural model is \(Y=\beta D+C\delta+u\), with prespecified covariates \(C\). The investigators want to test whether the effect is 3 points.
The hypotheses are \(H_0:\beta=3\) points and \(H_a:\beta\ne3\) points. Under the null, the transformed outcome is \(Y-3D\). If the two instruments have reduced-form coefficient vectors \(\gamma\) and \(\pi\) for \(Y\) and \(D\), the null restriction is \(\gamma-3\pi=0\). This is a two-dimensional restriction, so the score statistic is compared with \(\chi^2_2\).
The study must still justify instrument exogeneity and exclusion, and specify the causal interpretation of \(\beta\). If the first stage is weak, the score test is designed not to require a strong first stage for its null calibration; weak relevance may nevertheless mean that the test has little power. The proposed value is in outcome points per treatment unit, and the null is not a test that the instruments have no effect on treatment.
Worked Example: Computing a Two-Instrument Score Statistic
For a hypothetical analysis of \(H_0:\beta=\beta_0\), suppose the reduced-form score calculation, after fitting nuisance parameters under the null, gives \(S=(2.4,1.0)^{\mathsf T}\). Suppose its estimated covariance matrix is \(\widehat\Omega=\begin{pmatrix}4&0\\0&1\end{pmatrix}\). These are the efficient score and covariance inputs, not raw regression coefficients.
The inverse covariance is \(\widehat\Omega^{-1}=\begin{pmatrix}1/4&0\\0&1\end{pmatrix}\). Therefore,
There are two restrictions, so the reference distribution is \(\chi^2_2\). For two degrees of freedom, the upper-tail probability is \(P(\chi^2_2\ge2.44)=e^{-2.44/2}\approx0.295\). At \(\alpha=0.05\), the critical value is approximately 5.991, and \(2.44<5.991\). The test therefore does not reject this hypothesized effect. This is not evidence that \(H_0\) is true; it says that this test does not find statistically significant evidence against it at the 5% level.
Worked Example: A Score Test Rejects the Null
Consider another hypothetical two-instrument test of a specified structural effect. After null-constrained nuisance fitting, suppose \(S=(3.0,1.5)^{\mathsf T}\) and \(\widehat\Omega=\begin{pmatrix}2&0\\0&1\end{pmatrix}\). The instruments are assumed valid for this illustration, and the covariance estimator is appropriate for the sampling design.
The statistic is
With \(q=2\), the p-value is \(P(\chi^2_2\ge6.75)=e^{-6.75/2}\approx0.0342\). Since \(6.75>5.991\), the null is rejected at the 5% level. In context, the score test finds evidence against the specified causal-effect value under the assumed IV design and model. It does not show that the instrument is valid, nor does it establish which alternative effect value is correct.
Why the Test Is Weak-IV Robust
The key feature is the null-based score construction. The null restriction is evaluated while the first-stage coefficients remain nuisance parameters; the test does not first divide by an estimated first-stage effect and then treat the resulting ratio as approximately normal. That distinction matters because ratios can have highly non-normal sampling distributions when the denominator—the first-stage association—is near zero.
Under the null and the model’s regularity conditions for the reduced-form score, the nuisance-adjusted score has an asymptotic covariance that can be estimated, and its quadratic form has a chi-square limit. The argument does not require the first-stage coefficient to be bounded away from zero. If the instruments contain little information about treatment, the null distribution can still be correctly calibrated while the test has low power: a valid test may fail to distinguish many effect values.
Robustness is conditional on the sampling and error assumptions used to obtain the score covariance. With independent observations and homoskedastic errors, a model-based covariance may be appropriate. With heteroskedasticity, use a heteroskedasticity-robust covariance; with clustered observations, use a cluster-robust construction and enough independent clusters for its approximation to be credible. The number of instruments and tested restrictions must also be handled as specified by the test implementation.
Comparison With Anderson–Rubin Tests
The Anderson–Rubin (AR) test and the score test can evaluate the same structural null, \(H_0:\beta=\beta_0\), but they build their evidence differently. For AR, regress \(Y-\beta_0D\) on the included covariates and excluded instruments, then test whether the excluded-instrument coefficients are jointly zero. In the homoskedastic linear model, the usual AR statistic is the joint regression \(F\)-test; robust versions use the corresponding robust joint test. Under the null, this transformed-outcome restriction holds regardless of how strong the first stage is.
The score test also tests the structural-effect value, but uses a nuisance-adjusted score rather than the AR transformed-outcome joint regression statistic. Its construction can use information differently and can be more powerful in some settings. Neither test is uniformly more powerful in every design, and their statistics and p-values need not agree. For \(q\) instruments, each test evaluates a \(q\)-dimensional restriction, though the precise finite-sample reference distribution depends on the implementation and assumptions.
| Feature | Score-based weak-IV test | Anderson–Rubin test |
|---|---|---|
| Null | \(\beta=\beta_0\), with first-stage parameters treated as nuisance parameters | \(\beta=\beta_0\), tested through the excluded-instrument coefficients for \(Y-\beta_0D\) |
| Core calculation | Nuisance-adjusted score scaled by its covariance | Joint test of excluded instruments in the transformed-outcome regression |
| Weak-instrument validity | Designed to remain valid under weak relevance with the appropriate score construction | Does not require a strong first stage for null validity under its assumptions |
| Power and interpretation | May use information differently from AR; no universal power advantage | Often straightforward to implement and interpret as a transformed-outcome test |
Both tests can be inverted over candidate values \(\beta_0\) to form confidence sets: retain values not rejected by the test. Weak-IV-robust confidence sets may be wide, unbounded, or disconnected. Such a result can reflect limited information, not a computational error. A score test’s rejection or non-rejection at one value should not be turned into a conventional symmetric interval without performing the inversion.
Diagnostics, Limitations, and Common Pitfalls
Before interpreting a score-test result, check that the tested structural model, instrument set, and covariance estimator match the study design. Confirm that the statistic uses the intended nuisance-adjusted score and the correct degrees of freedom. Review first-stage diagnostics to understand the information available, but do not use a weak-instrument diagnostic as a substitute for the score test or treat a strong-looking first stage as proof of instrument validity.
- Weak relevance: the test can be correctly sized yet have little power. A non-rejection may leave a broad range of effect values plausible.
- Invalid instruments: weak-IV robustness does not protect against violations of exogeneity or exclusion. The causal conclusion depends on these assumptions.
- Wrong covariance: using a homoskedastic covariance with heteroskedastic or clustered data can invalidate the reference distribution.
- Wrong degrees of freedom: the chi-square degrees of freedom correspond to the dimension of the tested restriction, usually the number of excluded instruments in the joint test.
- Overinterpreting the p-value: a p-value is not the probability that the null is true, nor a measure of the causal effect’s size.
Common mistake: “The score test is weak-IV robust, so it proves the instrument is valid and a non-significant result confirms the proposed effect.” The correction is that weak-IV robustness concerns the test’s calibration when instrument relevance is weak. It neither tests all instrument-validity assumptions nor confirms the null. Report the tested effect, statistic, degrees of freedom, p-value, covariance method, and the design assumptions that support a causal interpretation.
Check Your Understanding
Answer each question, then compare with the solution immediately below it.
- A study tests \(H_0:\beta=2\) with three excluded instruments. What restriction is tested in reduced-form notation, and what is the usual asymptotic score reference distribution?
Solution: If \(\gamma\) and \(\pi\) are the reduced-form coefficient vectors for outcome and treatment, the restriction is \(\gamma-2\pi=0\). It has three components, so the usual asymptotic reference distribution is \(\chi^2_3\).
- A two-instrument score statistic is 4.00. Is it significant at the 5% level using the chi-square reference distribution?
Solution: The 5% critical value for \(\chi^2_2\) is approximately 5.991. Since \(4.00<5.991\), do not reject at 5%. This does not establish that the null is true.
- Does weak-IV robustness guarantee high power when the first stage is nearly zero?
Solution: No. It concerns validity of the null calibration under weak relevance. A weak first stage can leave the test with little power to distinguish effect values.
- What does an AR test examine after a candidate value \(\beta_0\) is specified?
Solution: It tests whether the excluded instruments jointly predict the transformed outcome \(Y-\beta_0D\), conditional on included covariates.
- Can a weak-IV-robust score test show that an instrument satisfies exclusion?
Solution: No. Instrument validity is a causal design assumption. The score test’s weak-IV robustness does not verify exclusion or exogeneity.