Under assumptions MLR.1-MLR.5, the variance of an OLS estimator is 2

V ar(j ) = SSTj (1 Rj2) for j = 1, . . . , k

Is there a way to decrease the variance of the estimator j without collecting more data? If yes, is there a cost or trade-off to doing so