Why do meta-analysis of small trials get the wrong answer?