Influence of trial sample size on treatment effect estimates: meta-epidemiological study