When economists apply regression to monthly or quarterly data to estimate structural parameters, they generally either (a) use seasonally adjusted data or (b) use unadjusted data and allow for seasonal shifts in the intercept term. Starting from a more general model than either of these procedures, this paper spells out their implicit assumptions and derives expressions for the specification bias arising from incorrect assumptions. The expected values of the estimates in (a) and (b) are weighted averages of the coefficients in the more general model. The weights are regression coefficients. If parameters do not vary seasonally, procedure (b) yields unbiased estimates. Expected values of estimates from procedure (a) are random variables. If we apply covariance analysis to test for seasonal variation in coefficients, seasonally unadjusted data will yield unbiased estimates; seasonally adjusted data will not. Empirical results obtained from seasonally adjusted data differ substantially from those obtained from unadjusted data.