Panel Econometrics & Time Series
Nickell Dynamic Panel Fixed-Effects Bias Formula (Short-T Audit)
Quantify exact Nickell (1981) downward Fixed-Effects bias in dynamic AR(1) panels y_it = alpha_i + rho*y_{i,t-1} + u_it to decide when Arellano-Bond GMM is required.
Why Fixed-Effects demeaning biases lagged dependent variables downward in short-T panels
In a dynamic panel y_{i,t} = alpha_i + rho * y_{i,t-1} + u_{i,t} with large N and fixed time dimension T, subtracting unit means y_bar_i eliminates alpha_i but injects shock u_{i,t-1} (inside y_{i,t-1}) AND -u_bar_i (inside demeaned errors), creating O(1/T) negative correlation! Nickell (1981) derived both the leading approximation -(1 + rho) / (T - 1) and the exact finite-T probability limit of Bias(rho_FE) for |rho| < 1.
Evaluate a 5-wave longitudinal survey (T = 5 in B3) with true persistence rho = 0.50 (B2): the naive leading term -(1 + 0.5) / 4 = -0.3750, and Nickell’s exact finite-T formula evaluates A_T = 1 - (1 - 0.5^5) / (5 * 0.5) = 49/80 = 0.6125, yielding exact Bias = -49/148 = -0.3311 — dragging expected rho_FE down from 0.5000 to 0.1689! Even at pure white noise rho = 0, Hurwicz-Nickell bias equals -1/T = -0.2000; expanding T to 50 shrinks exact bias to -0.0306, explaining why short-T panels require Arellano-Bond first-difference IV/GMM.
| Panel Horizon & True rho | Exact Nickell Bias | Expected Biased plim(rho_FE) |
|---|---|---|
| Short Panel (T = 5, rho = 0.50) | -0.3311 (-49 / 148) | 0.1689 (Loses 66% of Persistence!) |
| White-Noise Hurwicz Case (T = 5, rho = 0) | -0.2000 (-1 / T) | -0.2000 (Spurious Mean Reversion) |
| Long Macro Panel (T = 50, rho = 0.50) | -0.0306 | 0.4694 (Small O(1/T) Distortion) |
Exact Finite-T Nickell (1981) Fixed-Effects Bias Formula
Handle rho = 0 seamlessly (where (1 - rho^T)/(1 - rho) reduces to T, so A_T = (T - 1)/T and Bias = -1/T) and guard against T < 2 or non-stationary |rho| >= 1.
=LET(rho_0, B2, t_per, B3, IF(OR(t_per < 2, ABS(rho_0) >= 1), "Require T>=2 and |rho|<1", IF(ABS(rho_0) < 1E-12, ROUND(-1 / t_per, 4), LET(a_fac, 1 - (1 - POWER(rho_0, t_per)) / (t_per * (1 - rho_0)), num_b, -((1 + rho_0) / (t_per - 1)) * a_fac, den_b, 1 - ((2 * rho_0) / ((1 - rho_0) * (t_per - 1))) * a_fac, ROUND(num_b / den_b, 4)))))Expected Biased Within-FE Coefficient plim(rho_FE) = rho + Bias Formula
Flag severe short-T bias whenever ABS(Bias) >= 0.05 so analysts switch from Within-FE to Anderson-Hsiao or Arellano-Bond lagged-level instruments.
=IF(ISTEXT(E3), E3, ROUND(B2 + E3, 4))Try the example
Use the original values or change them to test your understanding.