Reducing Bias in Estimates of Linear Models by Remeasurement of a Random Subsample