Objective To compare the prevalence of polycystic ovary syndrome (PCOS) phenotypes and obesity among patients detected in referral versus unselected populations. Design Systematic review and meta-analysis. Setting Not applicable. Patient(s) Thirteen thousand seven hundred ninety-six reproductive-age patients with PCOS, as defined by the extended Rotterdam 2003 criteria. Intervention(s) Review of PUBMED, EMBASE, and Cochrane Library, 2003–2016. Only observational studies were included. Data were extracted using a web-based, piloted form and combined for meta-analysis. Main Outcome Measure(s) PCOS phenotypes were classified as follows: phenotype A, clinical and/or biochemical hyperandrogenism (HA) + oligo-/anovulation (OA) + polycystic ovarian morphology (PCOM); phenotype B, HA+OA; phenotype C, HA+PCOM; and phenotype D, OA+PCOM. Result(s) Forty-one eligible studies, reporting on 43 populations, were identified. Pooled estimates of detected PCOS phenotype prevalence were consequently documented in referral versus unselected populations, as  phenotype A, 50% (95% confidence interval [CI], 46%–54%) versus 19% (95% CI, 13%–27%);  phenotype B, 13% (95% CI, 11%–17%) versus 25% (95% CI, 15%–37%);  phenotype C, 14% (95% CI, 12%–16%) versus 34% (95% CI, 25–46%); and  phenotype D, 17% (95% CI, 13%–22%) versus 19% (95% CI, 14%–25%). Differences between referral and unselected populations were statistically significant for phenotypes A, B, and C. Referral PCOS subjects had a greater mean body mass index (BMI) than local controls, a difference that was not apparent in unselected PCOS. Conclusion(s) The prevalence of more complete phenotypes in PCOS and mean BMI were higher in subjects identified in referral versus unselected populations, suggesting the presence of significant referral bias.