Fama-MacBeth Standard Errors Versus Panel Clustering

Cross-sectional asset-pricing data has two directions of correlation at once: stocks in the same month tend to move together (a market-wide shock hits everyone that month), and a single stock's returns tend to be correlated with its own past returns over time (firm-specific persistence). Standard errors that ignore either direction understate the true uncertainty and produce t-statistics that look far more convincing than they should. Two well-known fixes exist — Fama-MacBeth and clustering — and they are not interchangeable; each is built to handle a different one of those two directions well, and can handle the other one badly.

What each method actually assumes

Fama-MacBeth runs a separate cross-sectional regression in each time period, then averages the resulting coefficients across periods and computes a standard error from how much those period-by-period estimates vary over time. This handles cross-sectional correlation within a period automatically and elegantly — since each period's regression only uses that period's cross-section once, the same-month comovement across stocks is baked into how noisy each period's estimate is, and the time-series standard error of the average captures it correctly. What it doesn't handle well is correlation within a stock over time: if a firm's residuals are persistently correlated across months (which is common), Fama-MacBeth standard errors as usually computed can still be too small, because the classic version assumes each period's cross-sectional estimate is independent of the others.

Clustering by firm, run on a single pooled panel regression, does the reverse: it explicitly allows each firm's observations across all time periods to be correlated with each other in any way, producing correct standard errors for that persistence. But clustering by firm alone assumes different firms in the same month are uncorrelated with each other — exactly the cross-sectional comovement Fama-MacBeth handles well. If a market-wide factor moves every stock together in a given month, firm-clustered standard errors on a pooled regression can still be too small.

Worked example. A regression of next-month return on a valuation ratio across 3,000 stocks and 300 months gives a coefficient estimate of 0.008. Plain OLS standard errors (ignoring both directions of correlation) produce a t-statistic of 5.2. Clustering by firm alone drops it to 3.4 — some of that apparent precision was firm-level persistence masquerading as independent observations. Clustering by month alone (equivalent in spirit to Fama-MacBeth here) drops it further to 2.8, since much of the remaining "independence" was actually shared monthly market movement. Double clustering by firm and month brings it to 2.3 — still significant, but a materially more honest number than the 5.2 the naive regression reported.

Which to use

In practice, the safest general approach for panel asset-pricing regressions is double clustering — by firm and by time period simultaneously — because it corrects for both directions of correlation at once without picking a side. Plain Fama-MacBeth is a reasonable default when cross-sectional comovement within a period is the dominant concern and firm-level persistence is mild; plain firm-clustering is reasonable when the reverse is true. When in doubt, or when reviewing someone else's regression table, checking which of these three approaches was used — and whether the choice matches which correlation is actually dominant in the data — is one of the fastest ways to sanity-check a claimed significant result in empirical finance.