This paper shows that a bootstrap Bartlett adjustment to the likelihood ratio statistic may be used to estimate p values, with errors of order n - 3 2 , for tests of linear restrictions in the normal linear regression model. The results help to explain the encouraging performance of the bootstrap in some simulation evidence reported recently by Rocke (1989) in the case of the seemingly unrelated regressions model. Bartlett-type adjustments to the Wald and Lagrange multiplier tests are also investigated.