Inference After Estimating Common Factors

Why Factor Estimation Changes Inference

A panel regression can have a precisely estimated coefficient conditional on the observed regressors and still have an understated standard error. One reason is that outcomes across units may share unobserved shocks. Another is that the common factors and their loadings are estimated rather than known. Inference must address both sources of uncertainty: the dependence left by common shocks and the extra variation introduced by estimating the common component.

This tutorial focuses on inference after common factors have been estimated. It distinguishes a factor-adjusted covariance calculation from common-correlated-effects (CCE) inference. The methods are related in purpose, but they do not make identical assumptions or use identical regressions. Neither method automatically solves every dependence problem: the remaining errors, the panel's asymptotic regime, and the factor-estimation procedure all matter.

Model and Sources of Uncertainty

Consider a panel with units \(i=1,\ldots,N\) and periods \(t=1,\ldots,T\). A linear model with \(r\) common factors is

Here \(y_{it}\) is the outcome, \(x_{it}\) is a vector of observed regressors, \(\beta\) is the coefficient vector of interest, \(f_t\) is an \(r\)-vector of common factors, \(\lambda_i\) is unit \(i\)'s loading vector, and \(\varepsilon_{it}\) is the idiosyncratic error. The factor component \(\lambda_i'f_t\) induces dependence across units at a given time whenever units load on the same factor. The factor and loading scales are not separately identified without a normalization; inference concerns the regression coefficient and fitted common component, not an arbitrary factor rotation.

There are two distinct uncertainty questions. First, even if \(f_t\) and \(\lambda_i\) were known, common factors would affect the variance of a coefficient estimator unless their contribution to its score were removed or accounted for. Second, in practice the factors and often the loadings are estimated. Their estimation error can affect the coefficient estimate and its covariance. A covariance calculation that conditions on the estimated factor values as if they were fixed addresses the first question only insofar as its residual covariance model is adequate; it does not, by itself, account for the second.

How a Common Shock Enters the Variance

A short derivation shows why cross-sectional independence is consequential. For illustration, suppose \(x_{it}\) is scalar and treated as fixed, and consider the no-intercept least-squares estimator that ignores the factors:

Write \(D=\sum_{t,i}x_{it}^{2}\). Substituting the model gives

If the scalar factors are independent across time with variance \(\sigma_f^2\), and independent of the idiosyncratic errors, then the common-factor part contributes

This follows because the coefficient on \(f_t\) in the numerator is \(\sum_i x_{it}\lambda_i\), and independent factors have zero cross-time covariances. If the idiosyncratic errors are also independent with variance \(\sigma_\varepsilon^2\), their contribution is \(\sigma_\varepsilon^2/D\). A covariance estimate that treats all \(NT\) observations as independent captures neither the common-factor term nor, generally, other serial or cross-sectional dependence.

Factor-Adjusted Covariance Inference

A factor-adjusted analysis estimates the common component and uses the adjusted regression to estimate \(\beta\). One common route is to estimate latent factors from the panel, for example by principal components, and then include the estimated factors in the regression. Another uses an estimated common component to adjust the outcome or residuals. In either case, the covariance calculation must match the estimating procedure: a sandwich covariance based on adjusted scores can account for remaining heteroskedasticity or dependence only to the extent its score aggregation and covariance estimator permit.

For a regression with estimated factors \(\hat f_t\), the plug-in calculation treats those factors as fixed regressors. It can be a useful first-order approximation in settings where factor estimates are sufficiently accurate and their effect on \(\hat\beta\) is asymptotically negligible. That conclusion requires conditions; it is not automatic. If factor estimation error affects the coefficient's first-order distribution, the covariance needs an additional correction. The same issue applies to estimated loadings and to uncertainty about the estimated common component.

An analytic correction can incorporate the factor-estimation influence into the coefficient's variance. A practical alternative, when justified by the design and asymptotics, is a resampling procedure that repeats the factor estimation, loading estimation, adjustment, and coefficient estimation in every resample. Re-estimating the factors matters: resampling only the final regression residuals while keeping the original factors fixed does not reproduce uncertainty arising from factor estimation. The resampling scheme must also preserve the relevant dependence structure; an ordinary independent-observation bootstrap is not automatically valid for a dependent panel.

Comparing Factor Adjustment with CCE Inference

CCE estimation handles unobserved common influences differently. Rather than first estimating latent factors, a CCE regression augments the unit-level regression with cross-sectional averages of the outcome and regressors, such as \(\bar y_t=N^{-1}\sum_i y_{it}\) and \(\bar x_t=N^{-1}\sum_i x_{it}\). Under its identifying conditions, these averages span or proxy the relevant common-factor space as \(N\) grows. Unit-specific CCE regressions can allow heterogeneous factor loadings and heterogeneous regression coefficients; pooled variants impose additional restrictions appropriate to the target parameter.

CCE inference is based on the augmented regression's scores and a covariance estimator suited to the panel's remaining dependence and asymptotic regime. It does not estimate and then plug in a particular latent factor series in the same way as principal-components factor adjustment. However, cross-sectional averages are themselves data-derived regressors, and their use does not make inference assumption-free. Validity relies on adequate factor proxies, sufficient cross-sectional information, identification, and a covariance method appropriate to residual dependence. A CCE covariance estimate and a factor-adjusted covariance estimate can therefore differ even when both aim to account for common shocks.

Choosing and Reporting the Inference

The central practical question is not simply whether a factor adjustment was performed. It is whether the reported covariance estimate reflects the complete estimation procedure and the dependence that remains. Before interpreting a confidence interval or test, report the factor or proxy method, the factor dimension or augmentation, the covariance estimator, the relevant clustering or time-dependence treatment, and the asymptotic justification.

A plug-in factor-adjusted standard error is most defensible when the factor-estimation contribution is asymptotically negligible for the coefficient under the stated growth regime and regularity conditions. When that is not justified, use a covariance correction or a suitable procedure that re-estimates the common component. CCE can be attractive when cross-sectional averages proxy the factor space under its conditions, but those conditions and the covariance treatment still require scrutiny. The two approaches are alternatives with distinct assumptions, not interchangeable labels for the same calculation.

• Mistake: “The estimated factors explain common shocks, so an ordinary regression standard error is valid.”
• Correction: The adjustment must be paired with a covariance estimator valid for the remaining errors, and factor-estimation uncertainty must be negligible or included.
• Mistake: “CCE uses cross-sectional averages, so it automatically handles every form of cross-sectional dependence.”
• Correction: CCE relies on those averages adequately proxying the relevant common factors under specified conditions; unmodeled dependence or unsuitable asymptotics can still invalidate inference.