Randomization at the level of primary care practice: use of pre‐intervention data and random effects models