Questions: Dynamic Panel Models and Arellano-Bond/Blundell-Bond Estimation

Question 1

A researcher estimates a dynamic panel model of firm investment (y_{i,t} depends on y_{i,t-1}) using fixed effects (within-estimation). The core problem with this approach is:

• A. Fixed effects cannot be applied when the panel is unbalanced or has missing observations
• B. The within-transformation requires subtracting each unit's mean of y, which includes y_{i,t-1} — creating correlation between the regressor and the transformed error (Nickell bias)
• C. Fixed effects removes too much variation, making it impossible to estimate the coefficient on the lagged dependent variable
• D. The lagged dependent variable must be treated as a fixed effect, not a regressor

Answer: (B)

The Nickell bias arises because the within-transformation subtracts each unit's mean, and that mean includes y_{i,t-1} — the very variable being used as a regressor. This induces correlation between the demeaned regressor and the demeaned error, making the FE estimator inconsistent. Crucially, this bias is of order 1/T, not 1/N — it persists even with large samples of units and only disappears if T is large.

Question 2

Arellano-Bond estimation uses y_{i,t-2}, y_{i,t-3}, and further lags as instruments for Δy_{i,t-1} in the first-differenced equation. Why are these lags valid instruments while y_{i,t-1} itself is not?

• A. Lags of y are always valid instruments by the exclusion restriction; y_{i,t-1} is excluded only because it appears directly in the regression
• B. y_{i,t-2} and earlier are correlated with Δy_{i,t-1} but uncorrelated with Δε_{i,t} = ε_{i,t} − ε_{i,t-1}, since they predate both error terms (assuming no serial correlation in the original errors)
• C. The Arellano-Bond procedure automatically selects valid instruments using a data-driven selection algorithm
• D. Further lags are weakly correlated with Δy_{i,t-1}, which makes them safer instruments with less risk of Hausman test failure

Answer: (B)

Instrument validity requires two conditions: relevance (correlated with the endogenous regressor) and exogeneity (uncorrelated with the error). y_{i,t-2} is relevant because y_{i,t-1} depends on its own history. It satisfies exogeneity because it was determined before ε_{i,t} and ε_{i,t-1} occurred — so Cov(y_{i,t-2}, Δε_{i,t}) = 0 under the assumption of no serial correlation in the original errors ε_{i,t}. y_{i,t-1} fails because it contains ε_{i,t-1}, which appears in Δε_{i,t}.

Question 3

The Nickell bias in fixed effects estimation of a dynamic panel model vanishes as the number of cross-sectional units N grows large, just as with other panel estimators.

• T. True
• F. False

Answer: False

The Nickell bias is of order 1/T, not 1/N. Increasing the number of units N does not resolve it — you would need T (the number of time periods per unit) to be large. In typical panels where N is large and T is small (e.g., annual surveys with 5–10 waves), the bias is substantial and FE estimates are inconsistent regardless of sample size.

Question 4

First-differencing a dynamic panel model removes the fixed effects but creates a new endogeneity problem: the first-differenced lagged dependent variable is correlated with the first-differenced error.

• T. True
• F. False

Answer: True

After first-differencing, the regressand is Δy_{i,t} and the lagged regressor is Δy_{i,t-1} = y_{i,t-1} − y_{i,t-2}. The first-differenced error is Δε_{i,t} = ε_{i,t} − ε_{i,t-1}. Since y_{i,t-1} is a function of ε_{i,t-1} (outcomes depend on past shocks), Δy_{i,t-1} and Δε_{i,t} both contain ε_{i,t-1} and are therefore correlated. First-differencing escapes the Nickell bias only to introduce this new endogeneity, which is why instruments are still needed.

Question 5

Why does first-differencing fail to solve the endogeneity problem in a dynamic panel model, even though it successfully eliminates the fixed effects?

Model answer: First-differencing removes the time-invariant individual effect but creates a mechanical correlation between the first-differenced lagged dependent variable and the first-differenced error. Δy_{i,t-1} = y_{i,t-1} − y_{i,t-2} contains y_{i,t-1}, which was itself generated partly by the shock ε_{i,t-1}. The first-differenced error Δε_{i,t} = ε_{i,t} − ε_{i,t-1} also contains ε_{i,t-1}. These shared terms make the regressor and error correlated, violating OLS exogeneity. The fix (Arellano-Bond) is to use lags of y dated t-2 and earlier as instruments, since they predate both ε_{i,t} and ε_{i,t-1}.

This is the central paradox of dynamic panels: the two natural fixes for panel endogeneity (FE and first-differencing) both fail when a lagged dependent variable is present. The Arellano-Bond approach resolves this by exploiting the panel's own history as a source of internal instruments.