Differencing Dynamic Panels And Deriving Moment Conditions

Why Differencing Creates Dynamic-Panel Moments

A dynamic panel includes a lagged outcome among its regressors, so the outcome at one date carries information about outcomes at later dates. If the model also contains an unobserved individual effect, that effect is correlated with the lagged outcome. First-differencing removes the time-invariant effect, but it also transforms the error. The timing of the transformed error determines which earlier outcome levels can provide valid orthogonality conditions.

The objective here is to derive those conditions, not to construct an estimator. The central result is that, under suitable assumptions, an outcome level dated at least two periods before a differenced equation is orthogonal to that equation’s differenced error. The outcome level from just one period earlier generally is not.

The assumptions concern both the errors and the initial conditions. A useful sufficient set is: \(u_{it}\) has mean zero; the idiosyncratic errors are serially uncorrelated, so \(E[u_{it}u_{is}]=0\) for \(t\ne s\); the errors are orthogonal to \(\eta_i\); and the initial outcome \(y_{i1}\) is orthogonal to the subsequent errors. A stronger, familiar way to express the timing restriction is that each \(u_{it}\) has conditional mean zero given \(\eta_i\), \(y_{i1}\), and the errors dated before \(t\). These assumptions rule out feedback from past outcomes or the individual effect into the current innovation.

Serial uncorrelatedness by itself is not a complete initial-condition assumption. If an initial outcome is correlated with later errors, an earlier outcome level can inherit that correlation, and the moment conditions below need not hold. The moment statements are population restrictions: they say that expected products are zero, not that every observed product must equal zero.

First-Differencing the Equation

Define the first difference of a variable as \(\Delta y_{it}=y_{it}-y_{i,t-1}\), and similarly \(\Delta u_{it}=u_{it}-u_{i,t-1}\). Subtract the model at \(t-1\) from the model at \(t\). The effect \(\eta_i\) appears in both equations and therefore cancels:

The first-differenced equation has no individual effect. Its right-hand side still contains a lagged change in the outcome, and that regressor can be correlated with the differenced error. The key is to identify outcome levels with dates early enough to avoid both error terms in \(\Delta u_{it}=u_{it}-u_{i,t-1}\).

Deriving the Valid Lagged-Level Moments

At date \(t\), the differenced error contains \(u_{it}\) and \(u_{i,t-1}\). An outcome level dated \(t-2\) is determined by the individual effect, the initial outcome, and errors dated no later than \(t-2\). Under the stated timing and orthogonality assumptions, it is orthogonal to both errors in the difference. The same argument applies to any still earlier outcome level.

A short derivation makes the timing explicit. For \(s\ge2\), \(y_{i,t-s}\) is dated no later than \(t-2\). Since \(\Delta u_{it}=u_{it}-u_{i,t-1}\), linearity of expectation gives:

The two zero terms follow because the earlier outcome is orthogonal to both \(u_{it}\) and \(u_{i,t-1}\): neither error has entered that outcome, and the assumptions rule out correlation through the initial condition or individual effect. This argument relies on the model’s timing restrictions as well as serially uncorrelated errors.

Why the First Lag Fails

The failure of the immediately preceding level is a direct consequence of the error timing. The level \(y_{i,t-1}\) contains the innovation \(u_{i,t-1}\), and the differenced error subtracts that same innovation. Under the stated assumptions, the covariance is

which is nonzero when the idiosyncratic error has positive variance. The first lag is therefore generally invalid, while the second and deeper lags satisfy the derived moments under the assumptions. This distinction is about dates relative to the differenced error, not about whether a variable is called a “lag.”

Differencing also changes the serial dependence of the errors. Even if the level errors are serially uncorrelated and have common variance \(\sigma_u^2\), adjacent differenced errors share one level error with opposite signs:

Thus serially uncorrelated level errors do not imply serially uncorrelated differenced errors. This is another reason to derive instrument validity from the original error timing rather than assume that differencing leaves all error properties unchanged.

Assumptions, Interpretation, and Common Pitfalls

The derived moments are useful only when their assumptions are credible. If the level errors have serial correlation, an earlier outcome may be correlated with \(u_{i,t-1}\), even when its date makes it appear safely remote. If the initial outcome is correlated with later innovations, the same problem can arise. In either case, the zero covariance step in the derivation is not justified and the stated moments may fail.

A frequent mistake is to say, “The individual effect has been differenced out, so every lagged outcome is a valid instrument.” Differencing removes \(\eta_i\), but does not remove the overlap between \(y_{i,t-1}\) and \(u_{i,t-1}\). The correction is to inspect the transformed error term and exclude the outcome dated \(t-1\); under the required timing assumptions, use levels dated \(t-2\) or earlier for the differenced equation.

A second pitfall is to treat the moment condition as a statement about each individual or each observed sample product. \(E[y_{i,t-s}\Delta u_{it}]=0\) is an expectation restriction. Individual products can be positive or negative, and their finite-sample average can differ from zero. The condition supplies a basis for later estimation procedures; it is not a claim that realized data satisfy an exact identity.