Data sources and study population
Our study was based on data obtained from the SRTR, the CMS ESRD program and the Social Security Death Master File (5). The SRTR maintains a database of all candidates for and recipients of solid organ transplants in the United States. Patients on waiting lists for organ transplantation and those who receive organ transplants are followed on a periodic basis with the use of data collection forms completed by organ-transplantation programs and submitted to the Organ Procurement and Transplantation Network. These follow-up data, in addition to data from the network regarding patients on waiting lists and the allocation of organs, are included in the SRTR database. The SRTR supplements information on vital status with data on deaths from the Social Security Death Master File and the Medicare Beneficiary Database maintained by CMS. Data collection by the SRTR is exempt from oversight under the “public benefit or service program” provisions of the Code of Federal Regulations (45 CFR 46.101[b]), as approved by the institutional review board of the Health Resources and Services Administration of the Department of Health and Human Services.
The Social Security Death Master File includes updated information on all participants in the Social Security system. Information on deaths reported to the Social Security system for the administration of the death, disability and retirement benefit programs is kept in the Death Master File database.
CMS maintains a database of all patients treated for ESRD in the United States, that includes information about demographics, treatment, hospitalization and costs for Medicare beneficiaries and other patients with ESRD who have received maintenance renal replacement therapy (RRT). This database also includes records of any changes in vital status or method of renal replacement, including kidney transplantation (6). Our study population included candidates 18 years of age and older who received LT between April 27, 1995 and December 31, 2008 (n = 59 242). Living donor and multiorgan transplants were excluded. The time period was divided into pre-MELD and MELD eras. Patients who received LT before February 28, 2002 were assigned to the pre-MELD era, and those who received LT on or after February 28, 2002 were assigned to the MELD era.
We constructed an analysis file containing information on the baseline demographic and clinical characteristics of the LT recipients who met inclusion criteria. The analysis file was linked to the CMS ESRD database to identify patients who received RRT after transplantation of the liver. The linkage between SRTR data and CMS data was established by matching patient-level sources, finding similarities in patient identifiers such as social security numbers, health insurance claim number, names and nicknames, gender and date of birth. More information on the nature of such linkages is described by Dickinson et al. (6).
In the descriptive analysis, continuous variables were expressed as mean ± standard deviation and categorical variables were expressed as proportions. The primary outcome was new-onset post-LT ESRD, defined as the earliest of initiation of chronic dialysis, wait listing for kidney transplantation and receipt of a kidney transplant ascertained by the CMS 2728 medical evidence form. This form is completed by the patient's dialysis center within 45 days of initiation of chronic dialysis or receipt of kidney transplantation. The listing for renal transplantation was ascertained from the SRTR data. These are objective and well-defined outcomes captured in the CMS ESRD and SRTR databases. Note that the CMS 2728 form is not completed if the patient is in acute renal failure or is expected to recover renal function. Therefore, a patient who requires dialysis for some period of time, after transplantation but recovers their renal function would not be included in the CMS ESRD database. Similarly, patients who receive continuous veno-venous hemodialysis or a few courses of RRT in the immediate posttransplant phase would not be included in the CMS ESRD database. Patients in the acute phase of renal failure are not listed for kidney transplantation and, therefore, would not be in the data set either. Since, we are looking at post-LT ESRD as a primary outcome and because listing of kidney transplant is a surrogate for RRT, we included both initiation of chronic dialysis and listing for kidney transplantation as our outcome. The date of placement on the waiting list for kidney transplantation was tracked for patients with liver transplants in whom, ESRD subsequently developed. Renal replacement modality changes from dialysis to renal transplantation were also recorded in order to identify patients who received a renal transplant from either a living or deceased donor.
Patients were followed from the date of LT to the earliest of new-onset ESRD, death or the end of observation period. The incidence rate was calculated as the number of ESRD events divided by total patient time expressed as patient-years.
Given the available data structure and our objectives, death (before ESRD onset) was treated as a competing risk; i.e. since, post-LT death precludes post-LT new-onset ESRD. Therefore, we modeled the cause-specific hazard of ESRD (7), which can be thought of as the rate of ESRD incidence among patients alive and ESRD-free. An alternative approach would be to model the cumulative incidence of ESRD, which can be thought of as the follow-up time-specific probability of ESRD (acknowledging that ESRD onset cannot occur following death 0. The cumulative incidence of ESRD essentially averages over deceased and surviving patients and, hence, would be most useful for descriptive purposes.
In the first part of the analysis, Cox regression was used to contrast the pre-MELD and MELD eras with respect to ESRD incidence, adjusting for recipient's age, gender, race, diagnosis, height, weight, status, pre-LT hypertension, diabetes, hospital status at LT, previous LT, history of transjugular intrahepatic portosystemic shunt (TIPS), donor age, donor gender, cold ischemia time, local versus (regionally, nationally) shared organ and donation after cardiac death (DCD) and the type of immunosuppression. We fitted one model in contrast to covariate-adjusted ESRD incidence rates between the two eras. We, then fitted a second model that coded the year of transplant as a continuous predictor (to test for a trend over calendar time), but with a change-point at year 2002 (to test for a change in the trend beginning in 2002, the start of MELD-based allocation).
In the second part of the analysis, Cox regression was used to determine the risk factors for new-onset post-LT ESRD among LT recipients. In this case, only patients transplanted in the MELD era who were not on RRT at LT were used, because lab measurements (i.e. creatinine, bilirubin, INR, albumin) were generally not available in the pre-MELD era. The model was adjusted for all covariates listed above, plus bilirubin, international normalized ratio (INR), creatinine, change in creatinine pre-LT (slope), sodium and albumin at LT. The slope of creatinine was estimated using least squares regression, (i.e. the familiar slope estimator from simple linear regression) based on all available creatinine values from the time of wait listing to the time of LT. The MELD score update is a complex process. Serum bilirubin, creatinine and INR are components of the MELD score. For some candidates, it could be one MELD update (creatinine values) between listing and transplant while for others it could be close to 10–15 or greater MELD updates, (each including a creatinine value) available to calculate the slope.
As implied previously, we fitted proportional hazards models to the cause-specific hazard of ESRD. This was carried out using PROC PHREG in SAS v9.2 (SAS Institute, Cary, NC, USA). In particular, the input record for each patient was the covariate; time between LT and the earliest of death, ESRD, loss to follow-up and end of study (with loss to follow-up and end of study, both treated as independent censoring); and an event indicator taking the value 1 for ESRD and 0 for either censoring or death. Note that the grouping of loss to follow-up, end of study and death as event = 0 should not be mistaken for an assumption that death and ESRD are independent. Coding all these events as event = 0 is merely a computational trick to make PHREG compute the risk sets appropriately. Putter et al. provide an excellent summary of related issues in the competing risks setting (8).
For the third component of the analysis, a time-dependent Cox model was used to study the association between post-LT ESRD and mortality, with ESRD (yes/no) coded as a time-dependent binary indicator.
As a subanalysis, we also modeled the rate of death before ESRD onset (i.e. cause-specific hazard of death among patients alive and ESRD-free), again using Cox regression. This model complements the first part of the analysis, in the sense that patients can cease to be alive and ESRD-free via two mutually exclusive events: ESRD onset (modeled in the first part of the analysis) and death before ESRD (the subanalysis).
All statistical analyses were conducted using SAS v9.2 (SAS Institute).