Interpreting J Tests And Diagnosing Moments

What the J Test Can—and Cannot—Tell You

An overidentification test asks whether a set of moment restrictions is jointly compatible with the data after the model parameters have been estimated. Hansen’s J test summarizes that compatibility in one statistic. Its value can provide evidence against the joint restrictions, but it does not name which restriction, if any, is responsible. That distinction matters in practice: a small p-value is a reason to investigate the model and its moments, not a ready-made verdict about a particular instrument or equation.

The test is defined for a model with more moment conditions than unknown parameters. Write the \(q\)-dimensional moment vector for observation \(i\) as \(g_i(\theta)\), where \(\theta\) is a \(k\)-dimensional parameter. Let \(\bar g_n(\theta)=n^{-1}\sum_{i=1}^n g_i(\theta)\). The maintained population restriction is that, at the true parameter \(\theta_0\), the expectation of every component is zero.

For independent observations, the long-run covariance matrix \(S\) is the covariance matrix of \(g_i(\theta_0)\). With serial dependence, it is the long-run covariance of the moment sequence, estimated accordingly. The chi-square result assumes a correctly specified joint null, identification (the moment Jacobian has rank \(k\)), a nonsingular \(S\), appropriate regularity and moment conditions, and consistent covariance estimation. The number of degrees of freedom is the number of restrictions in excess of the parameters, \(q-k\).

The statistic measures the size of the fitted sample moments relative to their covariance. Large values are more at odds with the joint null. The p-value is the probability, assuming the null and the reference approximation, of obtaining a J statistic at least as large as the observed one. It is not the probability that the null is true. At significance level \(\alpha\), reject \(H_0\) when the p-value is less than \(\alpha\); otherwise, do not reject.

Why the Reference Distribution Has \(q-k\) Degrees of Freedom

A short derivation clarifies why fitting the parameters matters. Let \(G=E[\partial g_i(\theta_0)/\partial\theta']\) be the \(q\)-by-\(k\) Jacobian, with rank \(k\). Under the null, the central limit theorem gives \(\sqrt n\,\bar g_n(\theta_0)\) an asymptotic normal distribution with mean zero and covariance \(S\). Estimating \(\theta\) shifts the moment vector along the \(k\) directions represented by \(G\). Efficient GMM chooses the parameter shift that makes the remaining covariance-weighted moment discrepancy as small as possible.

Here \(a\) represents a local change in the estimated parameter, and \(P_H=H(H'H)^{-1}H'\) is the projection onto the column space of \(H\). Minimization removes the component of \(U\) in that \(k\)-dimensional space. The residual projection \(I_q-P_H\) has rank \(q-k\). A standard normal vector’s squared length after projection onto a space of that rank has a chi-square distribution with \(q-k\) degrees of freedom. This is the asymptotic reason the fitted J statistic tests only the overidentifying restrictions.

If \(q=k\), there are no excess restrictions and \(q-k=0\): the J test cannot assess overidentification. Identification of the parameters alone does not create an overidentification test.

Interpreting Nonrejection

A nonrejection means the observed joint discrepancy is not sufficiently large, at the chosen significance level, to reject the moment restrictions. It does not establish that every moment is valid, prove the structural model correct, or show that the instruments are exogenous. A test may have limited power, particularly in a finite sample or when violations are small. The conclusion should therefore be stated as a lack of sufficient evidence against the joint restrictions under the test’s assumptions.

Interpreting Rejection—and the Limits of Attribution

Rejection indicates that the fitted moments are jointly too far from zero relative to their estimated covariance to be readily explained by sampling variation under the null. It is evidence against the joint restrictions as maintained, subject to the assumptions behind the test. Possible explanations include invalid moments, a misspecified structural model, a parameter or functional-form problem, or a failure of an assumption used for estimation or inference. The J statistic alone does not distinguish among them.

Nor can its single quadratic form identify a particular invalid moment. The moments may be correlated, so the inverse covariance matrix combines their discrepancies rather than assessing each one in isolation. Further, the parameter estimate is chosen using the moment system: discrepancies that can be reduced by changing the estimated parameters are absorbed in the fitting process. What remains is a joint discrepancy in the \(q-k\) directions not used to identify those parameters. Many different patterns of component-level errors can produce the same J value.

Targeted Moment Diagnostics

After a rejection, diagnostics should be tied to a specific scientific or economic rationale and to a clearly defined alternative set of restrictions. One common approach compares nested moment sets. Suppose a base set contains \(q_0\) moments and a larger set adds \(r\) moments. Both sets use the same \(k\) parameters, and the base set identifies them. A difference-in-J (or C) test evaluates whether the added moments contribute restrictions that remain compatible with the data, conditional on the base set.

The conditional wording is essential. A C-test rejection is evidence against the added restrictions if the base restrictions and the testing assumptions are maintained; it does not independently prove the base moments are valid. If the base set is itself misspecified, the conditional interpretation may fail. The test also needs compatible weighting and covariance estimation. Subtracting unrelated J statistics that use incompatible constructions does not automatically yield a valid C test.

A disciplined diagnostic sequence is useful when the joint test rejects:

Descriptive examination of individual fitted moments can also help locate patterns worth investigating. But a large-looking component is not, by itself, a formal test of that moment’s validity: it may be correlated with other components, affected by parameter estimation, and selected after looking at the same data. If many subsets are tested, the chance of at least one apparently small p-value rises unless the analysis accounts for multiplicity or the diagnostics are treated as exploratory.

Common Pitfall and Practical Interpretation

Wrong reasoning: “The J test rejected, so instrument three is invalid.” The test’s null concerns the joint set of overidentifying restrictions, and the statistic does not isolate a component. Correction: report rejection of the joint null, then use a justified targeted test—such as a nested C test—to investigate a specified added block, while stating the assumptions that make the comparison meaningful.

The converse overstatement is also common: “The J test did not reject, so all instruments are valid.” Nonrejection is not proof of validity. It means the test did not provide sufficient evidence against the joint restrictions at the chosen level, given its power, reference approximation and maintained assumptions. In either case, report the test statistic, degrees of freedom, p-value, significance level, and a conclusion in the context of the moment restrictions being tested.