This note considers a paradox arising in the least squares estimation of linear regression models in which the error terms are assumed to be iid and possess finite r-th moment, for r.in.[1,2).