Statistical analysis
We calculated the weighted mean difference for continuous data, which were age at onset of psychosis and age at initiation of smoking. For the categorical outcomes of rates of tobacco use in patients and controls, we calculated odds ratios from cross-sectional data and relative risks from prospective data.
We used Stata version 10 with the metan command. We judged a p value less than 0∙05 significant. We used a random-effects model in all analyses because we expected the data to be heterogeneous across studies. We calculated I² values to test for heterogeneity between studies. We deemed I² less than 25% to have low heterogeneity, 25–75% to have medium heterogeneity, and greater than 75% to have high heterogeneity. We also assessed publication bias and selective reporting with Egger’s and Begg’s tests47,48 and by inspecting the symmetry of funnel plots, as recommended in the Cochrane handbook.