A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
The Arellano-Bond estimator addresses Yᵢₜ = αYᵢₜ₋₁ + X'ᵢₜβ + αᵢ + εᵢₜ by first-differencing to eliminate αᵢ, then using lagged Yᵢₜ as instruments for ΔYᵢₜ₋₁. This is a dynamic panel GMM estimator consistent as N → ∞ with T fixed, addressing the Nickell bias of FE with lagged dependent variables.
The Arellano-Bond estimator solves a problem that arises the moment you include a lagged dependent variable in a panel model. You already know from studying lagged-dependent-variable regression that Yᵢₜ₋₁ on the right-hand side creates endogeneity if there are unobserved individual fixed effects αᵢ. The natural fix — fixed effects (within) estimation — makes things worse, not better. Demeaning or first-differencing removes αᵢ, but the demeaned lagged dependent variable is mechanically correlated with the demeaned error: this is the Nickell bias, which does not vanish as the sample grows unless T → ∞. For typical panels with large N and small T (many firms or countries, few time periods), the Nickell bias is severe.
The Arellano-Bond insight is to use first-differencing to eliminate αᵢ, then find instruments for the differenced lagged dependent variable from within the dataset itself. First-differencing gives ΔYᵢₜ = αΔYᵢₜ₋₁ + ΔX'ᵢₜβ + Δεᵢₜ. The problem is that ΔYᵢₜ₋₁ = Yᵢₜ₋₁ − Yᵢₜ₋₂ is correlated with Δεᵢₜ = εᵢₜ − εᵢₜ₋₁ because Yᵢₜ₋₁ depends on εᵢₜ₋₁. But notice that Yᵢₜ₋₂ is a valid instrument: it is correlated with ΔYᵢₜ₋₁ (directly, since ΔYᵢₜ₋₁ = Yᵢₜ₋₁ − Yᵢₜ₋₂), and uncorrelated with Δεᵢₜ as long as εᵢₜ is not serially correlated. For longer panels, even more lags are available as instruments, building up a potentially large instrument matrix.
The estimation framework is Generalized Method of Moments (GMM). With many potential instruments, GMM combines them efficiently into a single estimator by minimizing a quadratic form in the moment conditions. Arellano-Bond (also called difference-GMM) uses levels of lagged Y as instruments for the first-differenced equation. A refinement called the Blundell-Bond system-GMM estimator additionally exploits the levels equation, using lagged differences as instruments — this is particularly valuable when Y is close to a random walk and the Arellano-Bond instruments become weak.
Applying Arellano-Bond requires two diagnostic tests. The Sargan/Hansen test checks instrument validity — whether the full set of instruments is jointly exogenous. A rejection suggests some instruments are invalid (correlated with the error), often because the original error is serially correlated. The Arellano-Bond AR(2) test checks for second-order serial correlation in the first-differenced residuals; finding AR(2) would imply the original errors are serially correlated, which would invalidate the instrument construction. A well-specified model should show AR(1) but not AR(2) in the differenced residuals, and pass the Sargan/Hansen test with a large p-value.
A practical limitation is instrument proliferation: as T grows, the instrument count grows quadratically, leading to a large instrument matrix that can produce biased Hansen test statistics and finite-sample problems. Applied practitioners often limit the lag depth manually (using only the first two or three lags as instruments) to keep the instrument count manageable relative to the number of groups N. Arellano-Bond is the standard tool for dynamic panel settings in macroeconomics (growth regressions), corporate finance (capital structure dynamics), and labor economics (wage dynamics) — anywhere the theory predicts past outcomes directly influence current ones and the panel has large N, small T.