Long-Panel Asymptotics And Within Estimation

Why the Time Dimension Matters

In a dynamic panel, the lagged outcome is correlated with the unit-specific effect. The within transformation removes that effect, but it also makes the transformed lagged outcome correlated with the transformed error. This is the source of the familiar finite-\(T\) bias in dynamic fixed-effects estimation. Unlike a bias that disappears merely by adding more units while holding \(T\) fixed, this bias can shrink as the time dimension grows.

The phrase “the bias shrinks with \(T\)” needs qualification. A leading bias formula depends on the asymptotic regime and on the assumptions about the panel. In particular, the standard leading formula below is for a large-\(N\) limit at fixed \(T\), followed by an expansion as \(T\) increases. It is not a universal finite-sample formula for every \(N\).

The stationarity and error assumptions make the calculations below transparent; they are not innocuous details. For example, persistent common shocks, serial correlation in \(u_{it}\), nonstationary initial conditions, or a different dynamic specification can change the bias and the appropriate inference. The tutorial focuses on the coefficient \(\rho\), not on adding regressors or specifying a broader model.

Within Estimation and Its Large-\(N\) Limit

For each unit, let \(\bar{y}_i=T^{-1}\sum_{t=1}^{T}y_{it}\), \(\bar{y}_{i,-1}=T^{-1}\sum_{t=1}^{T}y_{i,t-1}\), and \(\bar{u}_i=T^{-1}\sum_{t=1}^{T}u_{it}\). The within estimator applies unit demeaning and regresses demeaned \(y_{it}\) on demeaned \(y_{i,t-1}\):

Because \(y_{it}=\alpha_i+\rho y_{i,t-1}+u_{it}\), subtracting unit means removes \(\alpha_i\). The resulting transformed error is \(u_{it}-\bar{u}_i\), which is generally correlated with \(y_{i,t-1}-\bar{y}_{i,-1}\). Thus removing the fixed effect does not make the within regression an ordinary regression with an exogenous regressor.

When \(N\) grows with \(T\) fixed, independent units allow the numerator and denominator, each divided by \(N\), to converge to their unit-level expectations. The ratio of those limits is a probability limit, not generally the expectation of the finite-\(N\) ratio. Expanding that probability limit for large \(T\) gives the standard leading result:

Why the Leading Bias Is of Order \(1/T\)

A short derivation shows both the direction and the scale of the effect. Write \(x_{it}=y_{i,t-1}\). In the within regression, the numerator of \(\hat{\rho}_W-\rho\) is the sum of demeaned \(x_{it}\) times demeaned \(u_{it}\). For one unit, its expectation is \(\operatorname{E}[\sum_t(x_{it}-\bar{x}_i)(u_{it}-\bar{u}_i)]\). The untransformed terms have zero expectation because \(x_{it}\) is dated before \(u_{it}\). The remaining contribution is \(-T\operatorname{E}[\bar{x}_i\bar{u}_i]\), or equivalently \(-T^{-1}\operatorname{Cov}(\sum_t x_{it},\sum_t u_{it})\).

Under the stationary AR(1) model, \(x_{it}\) contains past errors with geometrically declining weights. Summing the covariances across dates therefore gives \(\operatorname{Cov}(\sum_t x_{it},\sum_t u_{it})=T\sigma^2/(1-\rho)+O(1)\). Consequently, the expected demeaned numerator for one unit is \(-\sigma^2/(1-\rho)+O(T^{-1})\): it is of constant order in \(T\), not of order \(T\).

The denominator for one unit is the sum of squared demeaned lagged outcomes. Stationarity implies that its leading term is \(T\operatorname{Var}(y_{i,t-1})\). Since \(\operatorname{Var}(y_{i,t-1})=\sigma^2/(1-\rho^2)\), that denominator is \(T\sigma^2/(1-\rho^2)+O(1)\). Dividing the two leading terms yields

The bias is negative for \(-1<\rho<1\). For positive persistence, it pulls the within estimate downward. The calculation also explains the order \(1/T\): the demeaning-induced numerator is of constant order, while the regressor’s sum of squares grows in proportion to \(T\). The ratio-of-expectations reasoning here applies to the large-\(N\) probability limit; it does not justify replacing the expectation of a finite-sample random ratio by a ratio of expectations.

What Changes When \(N\) Is Fixed?

The preceding result is deliberately stated with its asymptotic order made explicit. With fixed \(N\), the within estimator is still consistent as \(T\) grows under suitable stationarity, ergodicity, and nondegeneracy conditions. But its finite-\(T\) expected bias is not generally given by the large-\(N\) leading formula. At fixed \(N\), the estimator is a random ratio, and the numerator and denominator are dependent. Taking the expectation of that ratio introduces a correction that is not captured by the ratio of their expectations.

For example, in a one-unit stationary Gaussian AR(1) calculation, the leading finite-\(T\) bias can be \(-(1+3\rho)/T\), rather than the large-\(N\) expression \(-(1+\rho)/T\). The point is not to substitute that one-unit expression into other settings; it is to recognize that the regime matters. A finite-\(N\) panel, a large-\(N\) probability limit, and a joint growth of \(N\) and \(T\) are distinct asymptotic questions.

Long-Panel Within Estimation and Dynamic-Panel GMM

Dynamic-panel GMM is commonly motivated by panels with many units and a relatively short time dimension. Difference-GMM uses moment conditions based on lagged levels, while system GMM adds level-equation moments under additional assumptions. Those methods can address the short-\(T\) dynamic fixed-effects problem, but their performance depends on the validity and strength of the moments and on instrument choices. Earlier material on instrument proliferation and finite-sample GMM inference is relevant when applying them; a large instrument set is not automatically an advantage.

When \(T\) is sufficiently long, within estimation becomes more attractive because its leading dynamic bias shrinks with \(T\), and it avoids constructing a large lag-instrument set. This is a comparison of approaches, not a rule that within estimation is always preferable for a long panel. The result above depends on a correctly specified, stationary dynamic model and suitable error assumptions. If these fail, increasing \(T\) does not repair misspecification or invalid assumptions.

The dimensions also matter for inference. If units are independent, adding units supplies more independent cross-sectional information. Adding time observations supplies additional within-unit information, but observations from the same unit are linked by the dynamic process. Therefore, “more observations” is not by itself a complete description of precision: one must state which dimension grows and what dependence assumptions justify the standard errors. The later subject of panel inference treats dependence structures and covariance estimation in detail.

Common Mistakes and Practical Checks

• Using the large-\(N\) bias formula at fixed \(N\) without qualification. The leading formula \(-(1+\rho)/T\) is for a specified sequential limit. A finite-\(N\) expectation involves the dependence between a random numerator and denominator.
• Assuming that adding units removes the dynamic fixed-effects bias. At fixed \(T\), the large-\(N\) probability limit can remain biased. More units mainly sharpen estimation around that limit under cross-unit independence.
• Treating the leading expansion as an exact correction. The omitted \(O(T^{-2})\) terms matter more when \(T\) is modest. Do not mechanically add \((1+\hat{\rho})/T\) to an estimate and call the result unbiased.
• Concluding that a long panel guarantees valid inference. Long \(T\) can reduce this particular bias under the model; it does not establish stationarity, rule out serial or cross-sectional dependence, or validate the chosen standard errors.
• Choosing between within and GMM from dimensions alone. The time and unit dimensions guide the asymptotic approximation, but moment validity, instrument strength, model specification, and the intended inference also matter.