Boundary Cases Of The Random-Effects Transformation

Why the Boundary Cases Matter

Random-effects estimation applies a transformation that subtracts part of each unit’s time average from its observations. How much is subtracted depends on the relative sizes of the unit-effect and idiosyncratic error variances. When the unit effect contributes almost no variance, the transformation should do little; when it dominates, the transformation should remove almost all unit-level averages. These limiting cases clarify how random-effects estimation relates to pooled OLS and the within estimator.

The limits describe the behavior of the transformation and estimator, not a diagnostic test of the random-effects assumptions. In particular, a transformation that is close to within demeaning does not by itself establish that the unit effect is correlated with the regressors or that a fixed-effects interpretation is required.

Model and Quasi-Demeaning Parameter

Consider a balanced panel with \(N\) units and \(T\) observations per unit. Write the static panel model as

Here \(y_{it}\) is the outcome, \(x_{it}\) is a vector of regressors, \(\beta\) is the slope vector, \(a_i\) is the time-invariant unit effect, and \(u_{it}\) is the idiosyncratic error. The variance-components assumptions used here are that \(a_i\) has mean zero and variance \(\sigma_a^2\), \(u_{it}\) has mean zero and variance \(\sigma_u^2>0\), and the unit effects and idiosyncratic errors are mutually uncorrelated. The idiosyncratic errors are uncorrelated across time and units. For the usual random-effects interpretation, the unit effects are also assumed uncorrelated with the regressors at all dates; the transformation alone does not guarantee this condition.

For any unit-level variable \(z_{it}\), let \(\bar z_i=T^{-1}\sum_{t=1}^{T}z_{it}\) denote its time average. The random-effects quasi-demeaning transformation is

The same transformation is applied to the outcome and every regressor. Because \(\sigma_u^2>0\) and \(\sigma_a^2\geq 0\), the parameter satisfies \(0\leq\theta<1\) for finite variance components. A value near zero leaves observations nearly unchanged; a value near one nearly subtracts each unit’s full time average.

Why This Value of \(\theta\) Appears

The formula can be justified by considering the covariance of a unit’s error vector. Write \(e_i=(a_i+u_{i1},\ldots,a_i+u_{iT})'\). Its covariance matrix is

The unit-mean direction, represented by \(\mathbf{1}\), has error variance \(\sigma_u^2+T\sigma_a^2\). Any direction that contrasts observations within the unit has variance \(\sigma_u^2\), because the common effect \(a_i\) cancels in a within-unit contrast. Let \(P=\mathbf{1}\mathbf{1}'/T\) denote projection onto the unit-mean direction. The transformation matrix \(I_T-\theta P\) leaves within-unit contrasts unchanged and multiplies the unit-mean direction by \(1-\theta\). Thus the transformed variance in that direction is \((1-\theta)^2(\sigma_u^2+T\sigma_a^2)\). Choosing \(\theta\) to make it equal to the within-direction variance gives

This equalization is the basis of the quasi-demeaning transformation for random-effects generalized least squares. It also makes the limiting cases transparent: as the variance contribution along the unit-mean direction changes, the amount of that direction retained in the transformed data changes.

Boundary Case 1: Negligible Unit-Effect Variance

If \(\sigma_a^2=0\), the unit effect has no variance contribution. Substitution into the formula gives

The transformation then leaves every observation unchanged: \(z_{it}^{*}=z_{it}\). The transformed regression is the original pooled regression, so random-effects GLS reduces to pooled OLS under this boundary case. If \(\sigma_a^2\) is positive but small relative to \(\sigma_u^2\), \(\theta\) is small and the transformation is close to pooled OLS. The exact equality with pooled OLS applies at zero unit-effect variance; “close” at a small positive variance is an approximation, not an identity.

Boundary Case 2: Unit-Effect Variance Dominates

For fixed positive \(\sigma_u^2\) and fixed \(T\), as \(\sigma_a^2/\sigma_u^2\) grows without bound,

Consequently, \(z_{it}^{*}\) approaches \(z_{it}-\bar z_i\), the within-unit demeaned observation. The transformed regression therefore approaches the within regression for slopes identified by within-unit variation. A regressor that is constant over time for each unit has \(x_{it}-\bar x_i=0\), so its coefficient cannot be identified by the exact within regression. The intercept is likewise removed by full demeaning.

This is a limit as the variance ratio grows; it is not the same as saying that a finite, very large \(\sigma_a^2\) makes \(\theta\) exactly one. Nor does the limiting algebra establish the random-effects exogeneity condition. It describes what the transformation does as the unit-mean direction receives progressively less weight relative to within-unit contrasts.

Intermediate Values Connect the Two Estimators

Between the boundaries, the transformed regression retains some unit-level mean variation and removes some of it. The following example makes that interpolation explicit. With an intercept \(c\), the transformed equation has intercept \((1-\theta)c\), since a constant transforms to \(1-\theta\). At \(\theta=1\), this intercept vanishes.

Limits and Interpretation

For fixed \(T\), the parameter depends on the variance ratio through

This expression summarizes the transition. A zero ratio gives \(\theta=0\); an increasingly large ratio makes \(\theta\) approach one. Increasing \(T\) also increases \(\theta\) for a fixed positive variance ratio: the unit-mean component accumulates the common effect across more observations, so the GLS transformation removes a larger fraction of the unit average. This does not mean the transformed data become fully demeaned for a fixed finite \(T\) and finite variance ratio.

The endpoints are useful as conceptual checks on a calculation and as guides to what the transformed regression emphasizes. Near zero, unit-level differences remain largely in the data, as in pooled OLS. Near one, persistent differences across units contribute little to identifying slopes, and within-unit changes dominate. The two estimators may nevertheless differ for reasons beyond this interpolation, including differences in identifying variation and in the assumptions needed for interpretation.

Common Mistakes

• Treating a near-boundary value as the boundary itself. A value such as \(\theta=0.90050\) still subtracts less than the full unit mean. Only the limit \(\theta\to1\) is full demeaning.
• Reversing the variance ratio. A larger \(\sigma_a^2/\sigma_u^2\) raises \(\theta\); a larger \(\sigma_u^2\) relative to \(\sigma_a^2\) lowers it. This follows directly from the expression \(1-1/\sqrt{1+T\sigma_a^2/\sigma_u^2}\).
• Claiming that the boundary proves an assumption. A transformation near within demeaning does not prove that the unit effect is correlated with regressors. The parameter describes variance components, not the validity of an exogeneity condition.
• Assuming every coefficient survives full demeaning. A time-invariant regressor becomes zero after subtracting its unit mean, so its coefficient is not identified from within-unit variation alone.