Setting Reasons for variation in BCG protection against tuberculosis are not well understood. We therefore estimated the protection provided by BCG vaccination against tuberculosis and examined reasons for variation in different settings. Methods We systematically searched English and non-English articles in 10 databases from inception to May 2009 as well as sources such as Google Scholar and trial registers to October 2009. Search terms included tuberculosis, tubercle bacill*, M. tuberculosis, M. bovis, M. africanum, M. canetti, M. microti and M. tuberculosis. Terms for the intervention included BCG Vaccine, BCG, BCG Vacc*, BCG Imm*, bacillus calmette. Two reviewers independently screened studies for inclusion. Data (including results and domains of study quality) were extracted by one reviewer and checked by another. Pooled analyses were based on random and fixed effects meta-analysis. Heterogeneity was quantified using tau-squared (t2) statistics. Random-effects meta-regression was used to investigate associations of study characteristics with intervention effect estimates. Results Of 21,030 references we identified 21 trials reporting tuberculosis outcomes. Protection against tuberculosis disease was variable (t2=0.292). BCG efficacy was high at higher latitude or when BCG was given only to infants (62%, 95% CI 48–72%) or to children after strict screening for tuberculin sensitivity (72% 95% CI 60–81%) Protection against meningeal and miliary tuberculosis combined was higher in infants (90%) and children with stringent testing respectively (92%). The observed protective effect of BCG did not differ by the strain of BCG vaccine used. Stringency of tuberculin testing to exclude those infected or sensitised to mycobacteria explained a good proportion of the observed variation with latitude (t2=0.292 in the null model, 0.100 in the bivariate model). In a multivariable meta-regression analysis there was strong evidence that infant vaccination or stringent prior tuberculin testing before BCG in childhood was associated with increased efficacy (p=0.02) after taking into account latitude and a measure of study quality-diagnostic detection bias. Conclusion BCG protection against tuberculosis varies between settings to an extent which cannot be attributed to chance alone. More efficacious results were seen in studies of individuals screened using stringent criteria (to exclude those already sensitized to mycobacteria), and those at a greater latitude from the equator.