A retrospective study of all first transplant recipients age older than 17 years between 1988 and 1999 that had at least 2 years of follow-up information (data available through December 2001) was conducted to determine the predictive nature of certain patient diagnostics. The independent renal function parameters used included creatinine level (treated as continuous and categorical), the relative difference in 1/creatinine level from 6 months to a year, and calculated GFR levels. Outcomes of death-censored graft loss, patient death, and overall graft loss were measured as the response in the models. Observed patients were included that possessed follow-up information for a given time period (either 2 years or 7 years), and baseline levels at 6 months and 1 year were both tested as explanatory variables. Those patients with creatinine levels greater than four at these periods were excluded from the study, as their immediate health-related concerns were thought to confound their study outcome. For the purpose of calculating GFR estimates, values for height and weight were imputed, using the Markov Chain Monte Carlo methodology (11) with height (if nonmissing) and weight (if nonmissing) as predictors stratified by race and gender. The analyses that incorporated clearance levels were calculated as:
were repeated with and without imputation to verify the results (12). Values of weight and height that were thought to be miscoded were treated as missing (<20 kg and <80 cm, respectively), these observations represented less than 0.1% of all eligible records. The endpoints addressed were patient death, death-censored graft loss, and overall graft loss. Logistic regression was used to analyze the likelihood of particular outcomes for a given follow-up period. Output from the logistic regression was further utilized to generate predictive diagnostics and ROC plots. Univariate logistic regression was performed for creatinine levels (treated as both continuous and categorical), change in 1/creatinine level from 6 months to 1 year, and GFR:
Groupings for Scr (mg/dL), when treated as categorical, were (0−1.2, 1.2 ≥€ 1.4, 1.4 ≥€ 1.6, 1.6 ≥€ 1.8, 1.8 ≥€ 2.1, 2.1 ≥€ 2.5, 2.5 ≥€ 4.0). An alternative to the formula estimating filtration rate that was proposed by the MDRD Study was also tested (13):
Glomerular filtration rate levels were expressed per 1.73 m2 of body surface area; body surface areas were calculated utilizing the formula by Lam and Leung (14):
The area calculation for the ROC curve was made by the Trapezoidal Rule (15). In addition to the area under the ROC curve as a prediction diagnostic, we examined the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for specific renal function levels. The PPV was calculated as the conditional probability that the event of interest occurred given the at-risk diagnostic was present at the baseline time period, and in a similar fashion, NPV was calculated as the conditional probability that the outcome of interest did not occur given that the diagnostic baseline level was not designated as at-risk. Multivariate logistic regression was performed for the same outcomes; covariates used in the model were donor and recipient demographics, primary diagnosis, dialysis time before transplant, HLA-matching, cold ischemia time, year of transplant, PRA level, recipient CMV status, cadaveric/living donor, antiproliferative medication at transplant, and immunosuppressive medication at baseline. Medication regimens were not implemented in the multivariate models for the 7-year follow-up periods, as the majority of patients from that time period were on Azathioprine and Sandimmune in addition to steroids. Beyond the scope of the overall variable used in the model, we examined distinct levels of applicable diagnostics as to their predictive power for overall graft loss. In particular, we examined two potential surrogate endpoints for the creatinine level that have been suggested by the transplant community of 1.6 mg/dL and 1.8 mg/dL, and a –30% change in 1/creatinine that has also been proposed for this purpose. Statistical analysis was performed using SAS software v8.02 (Cary, NC).