Background: Anaemia is a common problem experienced by critically-ill people. Treatment with erythropoiesis-stimulating agents (ESAs) has been used as a pharmacologic strategy when the blunted response of endogenous erythropoietin has been reported in critically-ill people. The use of ESAs becomes more important where adverse clinical outcomes of transfusing blood products is a limitation. However, this indication for ESAs is not licensed by regulatory authorities and is called off-label use. Recent studies concern the harm of ESAs in a critical care setting. Objectives: To focus on harms in assessing the effects of erythropoiesis-stimulating agents (ESAs), alone or in combination, compared with placebo, no treatment or a different active treatment regimen when administered off-label to critically-ill people. Search methods: We conducted a systematic search of the Cochrane Central Register of Controlled Trials (CENTRAL), MEDLINE, Embase, PsycINFO via OvidSP, CINAHL, all evidence-based medicine (EBM) reviews including IPA and SCI-Expanded, Conference Proceedings Citation Index- Science, BIOSIS Previews and TOXLINE up to February 2017. We also searched trials registries, checked reference lists of relevant studies and tracked their citations by using SciVerse Scopus. Selection criteria: We considered randomized controlled trials (RCTs) and controlled observational studies, which compared scheduled systemic administration of ESAs versus other effective interventions, placebo or no treatment in critically-ill people. Data collection and analysis: Two review authors independently screened and evaluated the eligibility of retrieved records, extracted data and assessed the risks of bias and quality of the included studies. We resolved differences in opinion by consensus or by involving a third review author. We assessed the evidence using GRADE and created a 'Summary of findings' table. We used fixed-effect or random-effects models, depending on the heterogeneity between studies. We fitted three-level hierarchical Bayesian models to calculate overall treatment effect estimates. Main results: Of the 27,865 records identified, 39 clinical trials and 14 observational studies, including a total of 945,240 participants, were eligible for inclusion. Five studies are awaiting classification. Overall, we found 114 adverse events in 33 studies (30 RCTs and three observational studies), and mortality was reported in 41 studies (32 RCTs and nine observational studies). Most studies were at low to moderate risk of bias for harms outcomes. However, overall harm assessment and reporting were of moderate to low quality in the RCTs, and of low quality in the observational studies. We downgraded the GRADE quality of evidence for venous thromboembolism and mortality to very low and low, respectively, because of risk of bias, high inconsistency, imprecision and limitations of study design. It is unclear whether there is an increase in the risk of any adverse events (Bayesian risk ratio (RR) 1.05, 95% confidence interval (CI) 0.93 to 1.21; 3099 participants; 9 studies; low-quality evidence) or venous thromboembolism (Bayesian RR 1.04, 95% CI 0.70 to 1.41; 18,917 participants; 18 studies; very low-quality evidence). There was a decreased risk of mortality with off-label use of ESAs in critically-ill people (Bayesian RR 0.76, 95% CI 0.61 to 0.92; 930,470 participants; 34 studies; low-quality evidence). Authors' conclusions: Low quality of evidence suggests that off-label use of ESAs may reduce mortality in a critical care setting. There was a lack of high-quality evidence about the harm of ESAs in critically-ill people. The information for biosimilar ESAs is less conclusive. Most studies neither evaluated ESAs' harm as a primary outcome nor predefined adverse events. Any further studies of ESA should address the quality of evaluating, recording and reporting of adverse events.