Double Clustering by Firm and Date

A regression of stock returns on some firm characteristic, run on a panel of many stocks over many months, has correlation running in two directions at once: a given stock's returns are correlated with its own past returns (firm-level persistence), and all stocks' returns in a given month are correlated with each other (market-wide shocks hit everyone the same month). Clustering standard errors by firm alone fixes the first problem but not the second; clustering by month alone fixes the second but not the first. Double clustering — clustering by firm and by date simultaneously — fixes both at once, and has become close to a default choice for this kind of panel regression in empirical finance.

How it works, in outline

The mechanics build the same underlying idea (cluster-robust variance estimation) three times and combine them: once clustering by firm, once clustering by date, and once clustering by the combination of firm-and-date (which for panel data with one observation per stock-month is just each individual observation). Adding the firm-clustered and date-clustered variance estimates together and then subtracting the individual-observation variance estimate gives the two-way clustered standard error — a formula that avoids double-counting the correlation that both single clusterings would otherwise credit twice.

The practical effect is usually to widen standard errors (and shrink t-statistics) relative to using either single-direction clustering alone, because double clustering is acknowledging more of the true correlation structure in the data rather than ignoring one direction of it. A coefficient that looked comfortably significant under firm-only or date-only clustering can become marginal, or lose significance entirely, once both directions are accounted for — which is exactly why checking which clustering method underlies a reported result matters before trusting it.

A concrete example

A cross-sectional regression of next-month stock return on a valuation ratio, run across 2,000 stocks and 240 months. Clustered by firm only, the coefficient's t-statistic is 3.1. Clustered by month only, it's 2.6. Double-clustered by firm and month, it drops to 2.0 — still significant at conventional levels, but noticeably weaker than either single-direction clustering suggested, because both a persistent firm effect and a shared monthly market effect were present in the data and only the double-clustered estimate accounts for both simultaneously.