Common Correlated Effects Estimation

Why Augment a Panel Regression?

A panel regression can be misleading when an unobserved shock affects many units at once. For example, a change in financing conditions may influence each unit’s outcome and its regressors, but with a different strength for each unit. If that common influence is omitted, a regressor’s coefficient can partly reflect the shock rather than its own relationship with the outcome.

Common correlated effects (CCE) estimation addresses this problem without requiring the researcher to estimate each common factor directly. It augments the regression with cross-sectional averages of the outcome and regressors. Under suitable conditions, these averages provide proxies for the factor space. The method is useful when units may respond heterogeneously to common shocks and regressors may be correlated with those shocks.

A common-factor representation helps explain the method. One useful specification is

Here \(f_t\) is an \(r\)-vector of unobserved common factors; \(\lambda_i\) and \(\Lambda_i\) are unit-specific factor loadings; \(d_t\) contains observed common controls, such as deterministic terms; and \(\varepsilon_{it}\) and \(v_{it}\) are idiosyncratic components. The loadings allow the same factor to affect different units with different magnitudes and signs. The central difficulty is that \(f_t\) can be correlated with \(x_{it}\), so treating it as part of an ordinary regression error can invalidate the slope estimate.

Why Cross-Sectional Averages Proxy for Factors

Average the unit-level equations across \(i\). If the common factors affect a non-negligible share of units, the averages inherit their time variation. In a simplified factor-driven regressor equation, for example, \(\bar x_t\) contains the term \(\left(N^{-1}\sum_i\Lambda_i\right)f_t\). The average of the idiosyncratic components may become small as \(N\) grows when those components are sufficiently weakly dependent across units. The outcome average also carries common-factor variation through its loadings and through the factor-driven regressors.

The key requirement is not that an average equal a factor. Rather, the matrix of factor loadings in the averages must contain enough independent information to span the relevant factor space. If \(r\) factors matter, the population loading matrix linking \(\bar z_t\) to \(f_t\) must have rank \(r\). Then the averages can represent the factors up to a nonsingular linear transformation, with remaining terms that become negligible under the asymptotic conditions.

CCE Regressions and Their Targets

For each unit, the CCE regression includes the observed regressors, any observed common controls, and the cross-sectional averages. A standard unit-specific form is

The coefficient \(\delta_i\) is a nuisance coefficient: its purpose is to absorb the common variation represented by the averages, not to give a structural interpretation to each average. The slope of substantive interest is \(\beta_i\). Estimating a separate regression for each unit yields \(\hat\beta_i\); the CCE mean-group estimator is the unweighted average of those estimates:

This estimator allows the slopes to differ across units. Its target is the cross-unit average slope, provided the units and asymptotic framework support that interpretation. A pooled CCE specification instead imposes a common slope \(\beta_i=\beta\) and estimates a shared slope while allowing the common-factor controls—and, in appropriate specifications, unit intercepts or other nuisance terms—to absorb unit-specific features. The choice between these approaches is substantive: pooling is not merely a computational shortcut if slope heterogeneity is present.

Factor-Proxy Argument: A Short Justification

The logic can be stated as a linear-algebra argument. Suppose the vector of cross-sectional averages has the representation \(\bar z_t=A f_t+r_t\), where \(A\) is a loading matrix and \(r_t\) is a remainder that becomes negligible as \(N\) grows. If \(A\) has full column rank \(r\), there is a left inverse \(L\) such that \(LA=I_r\). Consequently,

As the remainder becomes small, linear combinations of the averages approximate the factors. Since the CCE regression includes the averages themselves, it also includes their linear combinations. The component \(\lambda_i'f_t\) can therefore be represented, up to a small remainder, by a linear combination of the included averages. This explains why the augmentation can control for common factors without estimating \(f_t\) or the factor loadings directly.

This argument depends on both parts of the condition. If \(A\) lacks full column rank, no left inverse can recover all \(r\) factor directions. If the remainder does not become negligible, averages may retain substantial idiosyncratic noise rather than acting as reliable proxies. And even with adequate proxies, the slope is not identified if \(x_{it}\) has no variation left after the controls are partialled out.

Conditions for Credible CCE Estimation

CCE consistency is an asymptotic result, not a guarantee from adding averages mechanically. The precise assumptions depend on the specification and whether the target is a unit-specific slope, a mean-group slope, or a pooled common slope. The main conditions have a common interpretation:

• Relevant factors are represented. The factors driving the outcome and the regressors must be captured by the factor space spanned by the averages. The average-loading matrix must have sufficient rank.
• Cross-sectional averages are informative. The factors must affect enough units, and idiosyncratic components must be sufficiently weakly dependent across units for their averages to become negligible or otherwise well behaved.
• There is usable time-series variation. After controlling for the averages and any included common terms, the regressors must retain enough variation to identify the slopes. Exact or asymptotic collinearity prevents identification.
• The asymptotic regime supports the approximation. Many CCE arguments rely on a large cross-section \(N\) so the averages proxy for factors, and a sufficiently informative time dimension \(T\) for the unit regressions. The required joint growth conditions depend on the result being used.
• The slope restriction matches the target. CCE mean group allows heterogeneous slopes; pooled CCE imposes a common slope. Consistency for one target does not imply consistency for the other when the restriction is wrong.

The augmentation handles dependence generated by the represented common factors; it does not automatically correct every form of dependence or endogeneity. Strong spatial structure not captured by the averages, missing common drivers, measurement problems, or feedback that violates the regression’s identifying assumptions may still cause difficulty. CCE is a factor-robust estimation strategy under stated conditions, not a substitute for specifying a credible model.

Common Mistakes and Practical Checks

A frequent mistake is to treat \(\bar y_t\) and \(\bar x_t\) as observed factors and interpret their regression coefficients as factor loadings. They are proxies built from observed panel variables; their coefficients are nuisance controls, and generally do not identify the underlying factors or loadings. A second mistake is to assume that including averages guarantees consistency. If their loading matrix does not span all relevant factors, or if the regressor is collinear with the augmentation, the intended slope may not be identified consistently.

Before interpreting a CCE estimate, check which slope target is being estimated, how the averages are constructed, whether the factor-proxy rank condition is plausible, and whether enough within-unit time variation remains after augmentation. Also distinguish the CCE estimator from subsequent inference: estimating factors or accounting for uncertainty in factor estimation can require additional procedures.