This paper considers unit root regressions in data having simultaneously extensive cross-section and time-series variation. The standard least squares estimators in such data structures turn out to have an asymptotic distribution that is neither $O_p(T^{-1})$ Dickey-Fuller, nor $O_p(N^{-1/2})$ normal and asymptotically unbiased. Instead, the estimator turns out to be consistent and asymptotically normal, but has a nonvanishing bias in its asymptotic distribution.