Longitudinal eGFR data in the estimation of missing baseline eGFR in patients with glomerular disease.

In observational studies of glomerular disease, baseline estimated glomerular filtration rate (eGFR) at biopsy is commonly defined using the first available measurement within six months. However, eGFR may change rapidly around biopsy, and values obtained distant from the biopsy may not reflect true baseline kidney function. When no eGFR is available, baseline eGFR are often treated as missing and imputed, which may introduce bias if the imputed values do not accurately represent baseline kidney function. To date, no studies have evaluated the accuracy of different imputation strategies for baseline eGFR in glomerular disease. Given the availability of repeated eGFR measurements in many glomerular disease cohorts, we hypothesized that incorporating longitudinal eGFR data into a mixed-effects modeling framework would improve the accuracy of imputing missing baseline eGFR. Using a population-based glomerular disease cohort in British Columbia, Canada, we defined true baseline eGFR as the closest value within 15 days of biopsy. We compared the closest eGFR within varying time windows to the true baseline eGFR using mean absolute error. Predicted baseline eGFR was generated using a mixed-effects model incorporating repeated eGFR measurements. We evaluated the accuracy of single and multiple imputation by comparing imputed values to the true baseline eGFR. Among 2 874 patients, error increased as the time window for selecting the closest eGFR from the biopsy date widened. In single imputation, predicted baseline eGFR based on the mixed-effects model consistently showed greater accuracy than the closest observed values. In multiple imputation, incorporating predicted baseline eGFR improved both the accuracy and precision of imputed values. In conclusion, using eGFR values distant from the time of biopsy may misclassify baseline kidney function. Incorporating longitudinal eGFR data through mixed-effects modeling improves both single and multiple imputation, enhancing accuracy and precision and potentially reducing bias in observational glomerular disease studies.