Appendix I: Joint Statistical Model
The relationship between change in eGFR and risk of graft failure was assessed using the joint modeling framework.1 Within the joint modeling framework, 2 models are fit to the data simultaneously: 1 for the longitudinal process (change in eGFR) and 1 for the event process (time-to-death-censored graft failure and time to all-cause graft failure).1
The joint model consists of 2 submodels: 1. A longitudinal linear mixed effects (random intercept-slope) model for modeling patterns of change in eGFR (i.e., the “true” eGFR process); and 2. Cox’s hazards model for modeling time to death-censored graft failure and time to all-cause graft failure. Joint estimation of the submodels is achieved by assuming they are correlated via individual-level random effects.1 The random effects in linear mixed effects model (the intercept and slope) are assumed to be multivariately normally distributed with mean 0 and unstructured covariance matrix.
The default specification of the joint model assumes the risk of an event at a moment in time depends on the current “true” value of the biomarker at that moment. This assumption may be reasonable; however, it is more likely the risks of death-censored graft failure and all-cause graft failure are related to dynamic characteristics of the longitudinal trajectory of eGFR, such as individual slopes. To accomplish this, the survival submodel was modified to include a slope parameterization of the eGFR process. This parameterization is of interest as it postulates that patients who show a steeper decease in eGFR (i.e., individual slope) are more likely to experience the event.
With Cox’s hazards model, the baseline hazard function is left completely unspecified to avoid the impact of miss-specifying the distribution of survival times. However, within the joint modeling framework, leaving the baseline hazard function unspecified may lead to underestimation of the standard errors of the parameter estimates.1 To overcome this issue, the piece-wise constant baseline hazard function was used.1
For modeling change in eGFR over time (the longitudinal process), prognostic factors included donor type (deceased versus living donor), donor age (per year increase), length of time from kidney transplantation to diagnosis of active AMR (per month increase), C4d positive stain at time of diagnosis of active AMR (yes versus no/unknown), and anti-HLA DSA positive (Class I only and Class II only versus both Class I and Class II). These factors were chosen for inclusion in the Linear Mixed Effects model their potential prognostic significance. For modeling time to death-censored graft failure and all-cause graft failure (the event process), prognostic factors included Linear Mixed Effects model-based prediction of eGFR at time of diagnosis of active AMR (i.e., baseline eGFR) and individual change in eGFR per month (slope parameterization).
Residual diagnostics were used to assess model assumptions. For the linear mixed effects submodel, subject-specific residuals were used to check the homoscedasticity and normality assumptions while marginal residuals were used to assess misspecification of the mean response structure.
Joint models were used to provide conditional survival predictions (death-censored free survival and graft survival) beyond 1 year following diagnosis of active AMR for a set of hypothetical eGFR dynamic characteristics (slopes). Predictions were used to determine change in eGFR over a 1-year time horizon (e.g., 1-year postdiagnosis of AMR) associated with a clinically meaningful difference in graft survival beyond 1 year. The 1-year time frame was chosen for the eGFR profile because this is the length of follow up that is typically used for pivotal trials. The set of hypothetical individual scenarios was based on the distribution of the individual slopes estimated from the longitudinal submodel. For each individual profile, the conditional survival beyond 1 year was estimated (e.g., at 5 years postdiagnosis of active AMR).
Survival rates for individuals with a stable eGFR profile were provided as the basis for quantifying a
clinically meaningful difference in death censored and overall graft survival due to a progressive decrease in eGFR from baseline at 1 year. The hypothetical individual scenarios were used to address the following question: how do changes in the rate of eGFR affect graft survival prognosis?
Appendix II: Definition of study visit windows for summarizing Estimate GFR in the First Year Postdiagnosis
Nominal Study Visit Laboratory Testing Window
Study Visit Window (Lab Date – AMR Diagnosis Date)
Day 0
(Diagnosis of active AMR)
-30 to +3 days Day -30 to Day 3
Month 1 (Day 30) ± 7 days Day 23 to Day 37
Month 2 (Day 61) ± 7 days Day 54 to Day 68
Month 3 (Day 91) ± 10 days Day 81 to Day 101
Month 6 (Day 183) ± 14 days Day 169 to Day 197
Month 9 (Day 274) ± 14 days Day 260 to Day 288
Month 12 (Day 365) ± 30 days Day 335 to Day 395
Month 15 (Day 456) ± 30 days Day 426 to Day 486
Month 18 (Day 548) ± 30 days Day 518 to Day 578
Month 21 (Day 639) ± 30 days Day 609 to Day 669
Month 24 (Day 730) ± 30 days Day 700 to Day 760
Month 27 (Day 821) ± 30 days Day 791 to Day 851
Month 30 (Day 913) ± 30 days Day 883 to Day 943
Month 33 (Day 1004) ± 30 days Day 974 to Day 1034
Month 36 (Day 1095) ± 30 days Day 1065 to Day 1125
Above pattern is repeated for measurements beyond Month 36
± 30 days Above pattern is repeated for measurements beyond Month 36 Abbreviations: AMR, antibody-mediated rejection.
Supplementary Tables
Table S1: Joint Modeling Results of the Estimated GFR Longitudinal Process and Risk of All-cause Graft Failure Composite Endpoint (Supplementary analysis set)
A. Linear Mixed Effects Submodel for the Longitudinal eGFR Processa
B. Cox’s Hazards Submodel for Time to Death-censored Graft Failure Composite Endpoint
Abbreviations: eGFR = estimated glomerular filtration rate; DSA = donor-specific antibody; AMR = antibody mediated rejection. aEstimated GFR restricted to the first 12 months postdiagnosis of active AMR
Linear Mixed Effects Model for Longitudinal eGFR Process
Comparison Estimate Standard Error
P-value
Variable
Intercept 50.717 7.831 <0.0001
Time Per month -0.753 0.106 <0.0001
Prognostic factors
Site Barcelona vs
Wisconsin cohort
-0.414 5.880 0.944 Manitoba vs
Wisconsin cohort
11.555 3.830 0.003
Donor type Deceased vs Living -0.886 3.052 0.772
Donor age Per year increase -0.312 0.108 0.004
Length of time from kidney transplantation to AMR diagnosis
Per month increase -0.028 0.033 0.400 C4d positive stain at time of AMR
diagnosis
Yes vs no/unknown -0.376 4.389 0.932 Anti-HLA DSA Class at time of AMR
diagnosis
Class I only vs Class I and Class II
5.727 4.254 0.178
Class II only vs Class I and Class II
7.249 3.550 0.041
Event Process Comparison Hazard
ratio
95% Confidence Interval
P-value
Variable
Estimated GFR (predicted value) at time of diagnosis
Per unit increase
0.910 0.875, 0.947 <0.0001
Slope Per month increase
in eGFR 0.170 0.067, 0.429 <0.0001
References
1. Rizopoulos D. Joint Models for Longitudinal and Time-to-Event Data: With Applications