**2. Methods**

#### *2.1. Patient Population*

All RTRs (aged ≥ 18 years) who were at least 1-year post-transplantation were approached for participation in the current study during outpatient clinic visits between 2001 and 2003. All RTRs were transplanted in the University Medical Center Groningen (Groningen, the Netherlands). The study has been described in detail previously [19]. Among 847 RTRs approached for participation, 606 RTRs agreed to participate and were included. All patients provided written informed consent and the study protocol was approved by the local medical ethical committee (METc 2001/039). The study protocol adhered to principles of the Declaration of Helsinki and the Declaration of Istanbul. The co-primary endpoints of the study were all-cause and cardiovascular death. Cause of death was obtained by linking the number of the death certificate to the primary cause of death as coded by a physician from the Central Bureau of Statistics according to the International Classification of Diseases, 9th revision (ICD-9; https://icd.codes/icd9cm). CV death was defined as deaths in which the principal cause of death was cardiovascular in nature, using ICD-9 codes 410 to 447. Secondary endpoint constituted death-censored graft failure (DCGF). DCGF was defined as return to dialysis or re-transplantation. For the current analyses, we excluded RTRs who did not have plasma samples available for measuring EPO levels (*n* = 14) and RTRs who used exogenous EPO (*n* = 13) due to positive interference in EPO measurement, resulting in 579 RTRs eligible for analyses. Median follow-up time from inclusion to endpoint was 7.0 (interquartile range (IQR), 6.2 to 7.4) years. Data on the co-primary and secondary endpoints were available in all 579 participants. There was no loss-to-follow-up in the current study.

#### *2.2. Data Collection*

Relevant donor, recipient, and transplant characteristics at baseline were extracted from the Groningen Renal Transplant Database, as described in detail previously [19]. Information on medical history and medication use was obtained from patient records. Participants' height and weight were measured with participants wearing light indoor clothing without shoes. Blood pressure was measured according to a strict protocol as previously described [19]. Alcohol consumption and smoking behavior were recorded using a self-reported questionnaire. Smoking behavior was classified as never, former, or current smoker.

#### *2.3. Laboratory Procedures*

Blood samples were drawn during the next outpatient clinic visit after agreeing to participate. Blood was drawn in the morning after an 8–12 h overnight fast, and all measurements were performed in samples of the same timepoint. In plasma EDTA samples frozen at −80◦C, we measured plasma EPO levels using an immunoassay based on chemiluminescence (Immulite, Los Angeles, CA) [20]. We measured plasma total FGF23 levels with a human FGF23 (C-terminal) enzyme-linked immunosorbent assay (ELISA; Quidel Corp., San Diego, CA, USA) with intra-assay and interassay coefficients (CVs) of variation of <5% and <16% in blinded replicated samples, respectively [21]. The total FGF23 immunometric assay uses two antibodies directed against di fferent epitopes within the C-terminal part of FGF23, and as such the assay detects both the intact hormone as well as C-terminal cleavage products, and therefore measures total FGF23 levels. We measured plasma ferritin levels using an electrochemiluminescence immunoassay (Modular analytics E170, Roche diagnostics, Mannheim, Germany). Renal function was determined by estimating GFR by applying the Chronic Kidney Disease Epidemiology Collaboration equation [22]. Proteinuria was defined as urinary protein excretion ≥ 0.5 g/<sup>24</sup> h in 24-h urine collection. Serum cholesterol was measured using standard laboratory procedures. Serum creatinine was assessed using a modified version of the Ja ffé method (MEGA AU 510; Merck Diagnostica, Darmstadt, Germany). Erythrocytosis was defined as hemoglobin level higher than 16.0 g/dL for women, and higher than 16.5 g/dL for men [23].

#### *2.4. Statistical Analyses*

Data were analyzed using IBM SPSS software, version 23.0 (SPSS Inc., Chicago, IL), R version 3.2.3 (Vienna, Austria) and STATA 14.1 (STATA Corp., College Station, TX). Data are expressed as mean ± standard deviation [SD] for normally distributed variables and as median (25th–75th interquartile range (IQR)) for variables with a skewed distribution. Categorical data are expressed as numbers (percentages). Co-linearity was tested by means of variance inflation factor (VIF) calculation, with a VIF score of lower than 5 indicating no evidence for co-linearity. We used Cox proportional hazards regression analysis to investigate the association between EPO levels and prospective outcomes. Assumptions of proportionality in Cox regression analyses were checked using Schoenfeld residuals plots and checking nonsignificance of covariates and with the global test (Model 1; EPO with death and CV death; *P* > 0.30 for global test). In these Cox regression analyses, we adjusted for potential confounders based on univariable associations or for factors of known biologic importance. We adjusted for age, sex, body surface area (BSA), eGFR, proteinuria, time since transplantation, presence of diabetes, systolic blood pressure (SBP), total cholesterol, and use of calcineurin inhibitors, proliferation inhibitors, and angiotensin-converting enzyme (ACE)-inhibitors and angiotensin II-receptor blockers (ARBs) (Model 1). We subsequently adjusted for potential mediators in the pathway between EPO and death, i.e., hemoglobin levels (Model 2); for ferritin (Model 3), high-sensitive C-reactive protein (hs-CRP) (Model 4), and finally for total FGF23 (Model 5). Due to skewed distribution, EPO, ferritin, hs-CRP, and total FGF23 were natural log-transformed. We repeated the Cox regression analyses between EPO and outcomes with EPO levels being divided in quartiles. Furthermore, we generated Kaplan–Meier curves to visually show the e ffect of increased risk of death and cardiovascular death while being in the highest EPO quartile. A log-rank test for trend was used to compare rates of death across quartiles. We also assessed the association between FGF23 levels and prospective outcomes adjusting for all potential confounders according to Model 1 and including EPO. To reflect the contribution of covariates in the di fferent Cox regression models, we generated Supplementary Tables S1–S4 showing the strength of covariates in univariable and multivariable models. To allow comparability between the hazard ratios (HR) of covariates, HR of continuous variables in Supplementary Tables S1–S4 are shown as expressed per SD. As sensitivity analysis, we assessed the prevalence of di fferent etiologies of CKD in total cohort and across EPO quartiles, and we adjusted the association between EPO and all-cause and cardiovascular death for etiology of CKD. Subsequently, we calculated the percentage of change in HR before and after adjustment for FGF23. Percentage change in HR was calculated as—(HR without adjustment – HR with adjustment)/(HR without adjustment – 1) × 100% [24]. Hereafter, we performed mediation analyses with the methods as previously described by Preacher and Hayes, which are based on logistic regression [25,26]. These analyses allow for testing significance and magnitude of mediation on the association between EPO and outcomes [25,26]. Overall, 0.4% of demographic data were missing and these data were imputed using regressive switching [27]. Five datasets were multiply-imputed, and results were pooled and analyzed according to Rubin's rules [28]. In all analyses, a two-sided *p*-value < 0.05 was considered significant.
