The Shanken Errors-in-Variables Correction

Testing an asset pricing model like the CAPM typically uses the Fama-MacBeth two-step procedure: first estimate each asset's beta (its sensitivity to a risk factor) from a time-series regression, then run a cross-sectional regression of average returns on those estimated betas to recover the factor's risk premium. The problem is that the betas used in the second step aren't the true betas — they're noisy estimates from the first step — and treating them as if they were known exactly makes the standard errors on the estimated risk premium look smaller than they really are. This is a classic "errors-in-variables" problem: using an imperfectly measured input as if it were perfectly measured.

The Shanken correction adjusts the standard errors from the second-stage regression upward by a factor that depends on how noisy the first-stage beta estimates were and how much the factor itself varies. Concretely, the corrected variance of the estimated risk premium $\hat{\gamma}$ is multiplied by roughly $\left(1 + \frac{\hat{\gamma}^2}{\hat{\sigma}_f^2}\right)$, where $\hat{\sigma}_f^2$ is the factor's own variance — so when the estimated premium is large relative to how variable the factor is, the correction inflates the standard errors more, and when betas were estimated very precisely to begin with, the correction shrinks toward negligible.

Without this adjustment, a researcher testing whether a factor's risk premium is statistically significant can overstate their confidence, because the naive standard error ignores that beta itself is an estimate, not a known constant — a subtle but well-documented source of false-positive factor "discoveries" in cross-sectional asset pricing tests.