Background For patients with heart failure, there is an inverse relation between body mass index (BMI) and mortality, sometimes called the obesity-paradox. However, the relationship might be either U- or J-shaped and might differ between patients with reduced (HFrEF) or preserved left ventricular ejection fraction (HFpEF). We sought to investigate this further in a dose–response meta-analysis of published studies. Methods PubMed and Embase from June 1980 to April 2017 were searched for prospective cohort studies evaluating associations between BMI and all-cause mortality in patients with HFrEF (LVEF < 40%) or HFpEF (LVEF ≥ 50%). Summary estimated effect sizes were obtained by using a random-effects model. Potential non-linear relationships were evaluated by using random-effects restricted cubic spline models. Results Ten studies were identified that included 96,424 patients of whom 59,263 had HFpEF (mean age 68 years of whom 38% were women) and 37,161 had HFrEF (mean age 60 years of whom 17% were women). For patients with HFpEF, the summary hazard ratio (HR) for all-cause mortality was: 0.93 (95% CI 0.89–0.97) per 5 units increase in BMI ( I 2  = 75.8%, p for heterogeneity = 0.01 and Begg’s test, p  = 1.0, Egger’s test, p  = 0.29) but the association was U-shaped ( p for non-linearity < 0.01) with the nadir of risk at a BMI of 32–33 kg/m 2 . For patients with HFrEF, the summary HR for all-cause mortality was: 0.96 (95% CI 0.92–0.99) ( I 2  = 95%, p for heterogeneity < 0.001 and Begg’s test, p  = 0.45, Egger’s test, p  = 0.01). The relationship was also U-shaped ( p  < 0.01), although ‘flatter’ than for HFpEF, with the nadir at a BMI of 33 kg/m 2 . Conclusions For patients with heart failure, the relation between BMI and mortality is U-shaped with a similar nadir of risk for HFpEF and HFrEF at a BMI of 32–33 kg/m 2 . Whether interventions that alter weight in either direction can alter risk is unknown.