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