Study Variation in Estimating GFR from Creatinine, Cystatin C, Age, and Sex

Background: The Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equations estimate GFR using serum creatinine and/or cystatin C, age, and sex. These equations were created by pooling datasets from multiple research studies and populations into a single dataset. We investigate implications of study variation on the relationship of GFR with its predictor variables in CKD-EPI datasets. Methods: We combined data used to develop and validate the 2021 CKD-EPI equations (11,737 individuals in 23 studies). For GFR equations based on creatinine or cystatin C, random effects modeling was used to partition measured GFR (mGFR) into five components: the average predictor-GFR relationship across studies, predictable variation due to differences in studies of CKD populations vs. non-CKD populations, non-predictable variation across studies, within-study prediction errors, and measurement error. Results: Overall, the average predictor-GFR relationship across studies did not substantially differ from the pooled 2021 CKD-EPI equations, which yield estimates intermediate between CKD and non-CKD populations. Still, considerable variation persisted beyond differences by CKD population (Figure). For lower creatinine/cystatin C levels, 50-70% of the GFR estimation error can be attributed to between-study differences, whereas 30-50% of the error is due to estimating GFR from predictor variables within a given study, after accounting for mGFR measurement error. For higher levels, 15-37% is due to between-study differences and 63-85% is due to within-study errors. Conclusion: Our results largely corroborate the CKD-EPI equations, which remain a pragmatic solution for widespread GFR estimation. Single study GFR equations may underestimate external prediction error by ignoring study level variability, especially in non-CKD populations. Funding: NIDDK SupportForest plot of CKD-EPI equation coefficients and population-averaged and study-specific coefficients from the random effects models for creatinine and cystatin C.