Questions: Dynamic Panel Models and System GMM Estimation

Question 1

A researcher estimates a dynamic panel model — one that includes a lagged dependent variable — using standard fixed-effects (within) estimation. Why is this estimator inconsistent?

• A. The within estimator only works for balanced panels without lagged regressors
• B. The lagged dependent variable is mechanically correlated with the individual fixed effect, violating the strict exogeneity required for within estimation
• C. Dynamic panels require at least 100 time periods for the within estimator to converge
• D. The within estimator cannot accommodate both time-varying regressors and lagged outcomes simultaneously

Answer: (B)

The within estimator demeans the data to remove fixed effects αᵢ, producing consistent estimates only when regressors are strictly exogenous — uncorrelated with αᵢ. In a dynamic panel, yᵢ,t-1 contains past y values, which were generated by the same fixed effect αᵢ. After demeaning, yᵢ,t-1 remains correlated with the transformed error term. This violates strict exogeneity, making the within estimator inconsistent. The bias (Nickell bias) is especially severe in short panels where T is small.

Question 2

A researcher uses Arellano-Bond difference GMM to study a highly persistent outcome (estimated ρ ≈ 0.95). The estimates are very noisy with enormous standard errors. The most likely explanation is:

• A. The Arellano-Bond estimator requires normally distributed residuals, which may be violated
• B. With ρ close to 1, lagged levels are weak instruments for the differenced equation — the instrument relevance condition is nearly violated
• C. The number of time periods is too large relative to the number of entities
• D. First-differencing removes the cross-sectional variation needed to estimate persistent processes

Answer: (B)

When ρ is close to 1 (high persistence), the differenced variable Δyᵢ,t-1 = yᵢ,t-1 − yᵢ,t-2 is close to zero — y barely changes period to period. Consequently, the level yᵢ,t-2 is only weakly correlated with Δyᵢ,t-1, making the lagged-levels instruments weak. Weak instruments produce large standard errors and potentially biased estimates. Blundell-Bond (system GMM) addresses this by adding the levels equations back to the system, improving instrument relevance for persistent variables.

Question 3

In the Arellano-Bond estimator, the level yᵢ,t-2 is a valid instrument for Δyᵢ,t-1 in the differenced equation because it is correlated with Δyᵢ,t-1 but uncorrelated with Δεᵢ,t, provided errors are not serially correlated.

• T. True
• F. False

Answer: True

This is the core instrument validity argument. Relevance: yᵢ,t-2 directly enters Δyᵢ,t-1 = yᵢ,t-1 − yᵢ,t-2, so they are correlated. Exogeneity: Δεᵢ,t = εᵢ,t − εᵢ,t-1 contains only errors at time t and t-1; if errors are not serially correlated, the earlier level yᵢ,t-2 is uncorrelated with these. The AR(2) test on differenced residuals checks precisely this assumption — second-order serial correlation would invalidate the t-2 instruments.

Question 4

A Hansen test p-value near 1.0 (extremely high) in a system GMM estimation is strong evidence that most instruments are valid and the model is correctly specified.

• T. True
• F. False

Answer: False

A p-value near 1.0 is actually a warning sign of instrument proliferation. When the number of instruments approaches or exceeds the number of entities, the Hansen test loses statistical power and becomes unable to detect invalid instruments — it produces suspiciously high p-values regardless of instrument validity. The rule of thumb is to keep instrument count below the number of entities. A 'too perfect' Hansen statistic should increase suspicion, not confidence.

Question 5

Why can't standard fixed-effects (within) estimation be applied to a dynamic panel model, and what is the fundamental mechanism Arellano-Bond uses to achieve consistent estimation?

Model answer: The within estimator subtracts each entity's mean to eliminate fixed effects αᵢ. In a dynamic panel, the lagged dependent variable yᵢ,t-1 depends on past y values, which were generated by the same αᵢ — so after demeaning, the transformed regressor is still correlated with the transformed error. This violates strict exogeneity and produces the Nickell bias, which is large when T is small. Arellano-Bond instead first-differences the model (Δyᵢt = ρΔyᵢ,t-1 + ...), which fully eliminates αᵢ. But now Δyᵢ,t-1 is endogenous (it contains εᵢ,t-1 which also appears in Δεᵢ,t). The solution is to instrument Δyᵢ,t-1 with lagged levels yᵢ,t-2, yᵢ,t-3, ... — these are relevant (correlated with the differenced regressor) and exogenous (uncorrelated with the differenced error under no serial correlation). GMM efficiently combines all these moment conditions.

The key insight is that first-differencing and instrumental variables solve two different problems in sequence: FD removes the unobserved heterogeneity, and IV handles the endogeneity created by the dynamic structure. Standard FE tries to solve the first problem but fails on the second because demeaning (unlike differencing) creates a correlation between the transformed regressor and transformed error that persists in short panels.