Arellano-Bond Tests For Serial Correlation

Why Test Differenced Residuals?

Difference-GMM removes time-invariant individual effects from a dynamic panel equation, but differencing also changes the serial dependence of its errors. Consequently, a residual correlation diagnostic for the differenced equation must be interpreted differently from one applied to an equation whose errors are assumed serially uncorrelated. In particular, first-order correlation in the differenced residuals is usually expected; second-order correlation is the more consequential warning for the standard lagged-level instruments.

Consider a dynamic panel model with an individual effect and idiosyncratic error. For simplicity, write its dynamic component as \(y_{it}=\alpha y_{i,t-1}+\mu_i+u_{it}\), where \(i\) indexes individuals and \(t\) indexes time. First differencing gives \(\Delta y_{it}=\alpha\Delta y_{i,t-1}+\Delta u_{it}\), with \(\Delta u_{it}=u_{it}-u_{i,t-1}\). The Arellano–Bond tests use residuals from this differenced equation to assess their serial correlation by lag.

Why First-Order Correlation Is Expected

Suppose the level errors \(u_{it}\) have mean zero, are serially uncorrelated within each individual, and have a common variance \(\sigma^2\). Then \(\Delta u_{it}\) and \(\Delta u_{i,t-1}\) share the level error \(u_{i,t-1}\), with opposite signs. Their covariance is

Each differenced error has variance \(2\sigma^2\), so their correlation in this special homoskedastic case is \(-\sigma^2/(2\sigma^2)=-1/2\). The precise correlation need not be \(-1/2\) when variances differ over time, but differencing still creates a shared component. By contrast, \(\Delta u_{it}\) and \(\Delta u_{i,t-2}\) contain no common level error. If the level errors are serially uncorrelated, their covariance is zero.

Test Statistic and Interpretation

Let \(\hat e_{it}\) denote the residual from the fitted differenced equation. For a chosen lag \(m\), form the within-individual products \(\hat e_{it}\hat e_{i,t-m}\) for all time pairs available in the estimation sample. The residual-based moment is their average, or a suitably weighted version when the available time pairs vary. Under the null of no serial correlation at lag \(m\), its population counterpart is zero.

A convenient way to express the test is as a standardized residual moment. Let \(\hat Q_m\) be the aggregate residual-product moment and let \(\widehat{SE}(\hat Q_m)\) be its estimated standard error. Then

The large-sample reference distribution under the null is standard normal. For a two-sided test at significance level \(\alpha\), reject \(H_0\) if \(|z_m|>z_{1-\alpha/2}\), or equivalently if the two-sided p-value \(2\Phi(-|z_m|)\) is less than \(\alpha\). In applied Difference-GMM, the standard error must reflect the panel structure and the estimation of the model. It is not generally appropriate to treat every time-pair product as an independent observation.

The test is a diagnostic for a specified lag, not an estimate of an autoregressive coefficient in the residuals. A significant result indicates evidence against zero residual correlation at that lag under the test’s assumptions; a nonsignificant result means the data do not provide sufficient evidence to reject that null. It does not prove that the null is true.

Why Second-Order Correlation Matters

The standard Difference-GMM moment conditions use lagged levels of the outcome as instruments for the differenced equation. For example, \(y_{i,t-2}\) can instrument \(\Delta y_{i,t-1}\) in the equation dated \(t\), provided it is uncorrelated with \(\Delta u_{it}\). When level errors are serially uncorrelated, the differenced error \(u_{it}-u_{i,t-1}\) is uncorrelated with errors dated \(t-2\) and earlier. Under the additional timing and initial-condition assumptions needed for the instrument, this supports the relevant lagged-level moment condition.

Second-order correlation in differenced errors signals a potential failure of the underlying no-serial-correlation assumption for level errors. For instance, if level errors are serially correlated, \(u_{i,t-1}\) may be correlated with \(u_{i,t-2}\). Since \(y_{i,t-2}\) reflects earlier shocks, that dependence can make \(y_{i,t-2}\) correlated with \(\Delta u_{it}\). The standard lagged-level instrument may then be invalid. A significant second-order test is therefore a warning to scrutinize the specification and instrument moments; it is not, by itself, a complete diagnosis of which assumption failed.

Reading the Diagnostic Pattern

The first- and second-order tests answer separate questions and should be reported separately. A common pattern under the standard Difference-GMM assumptions is rejection of the first-order null and failure to reject the second-order null. The first result is compatible with the mechanical dependence created by differencing; the second is the result needed to avoid evidence of the particular higher-order dependence that threatens the usual lagged-level moments.

Failure to reject the second-order null is not proof of instrument validity. The test may have limited power, and instrument validity also depends on assumptions about timing, the model specification, and other sources of dependence. Likewise, rejecting the second-order null should prompt investigation rather than an automatic claim that a particular instrument is invalid. The test is most useful as one diagnostic within a clearly stated model and set of moment assumptions.

Common Mistakes and a Testing Workflow

The most consequential mistake is treating any rejection as equally damaging. First-order correlation is generally expected under serially uncorrelated level errors, whereas second-order correlation is the more direct warning for the standard lagged-level instruments. Another mistake is interpreting a large p-value as proof that residuals are uncorrelated. The correct conclusion is only that the test did not detect evidence against its null at the chosen level.