Study design and cohort development
This study was approved by the Research Board and the Hospital Ethics Board at St. Boniface Hospital in Winnipeg, Manitoba. We used the Canadian Organ Replacement Registry (CORR) as the source cohort for this analysis. CORR is a validated ESRD registry that captures data on all dialysis patients in Canada including demographics, death, dialysis modality, comorbidities and transplantation who supply written consent [22, 23]. We included all adult patients starting hemodialysis between January 1, 2001 and December 2009 who survived greater than 90 days. Patients with missing co-morbidity information were excluded (N =1,408). The final analytic cohort included 28,971 (Women 11,792 (40.7%), Men 17,179 (59.3%)). All data was de-identified, retrospective from the CORR and thus individual participant consent was not required.
Definitions
Co-morbid illnesses included a history of coronary artery disease, acute myocardial infarction, diabetes mellitus, pulmonary edema, peripheral vascular disease, malignancy, hypertension medication usage, current cigarette smoker, lung disease, any serious illness and stroke. Serious illness was defined as any illness that could shorten life expectancy to less than 5 years. Causes of ESRD included vascular disease, diabetes mellitus, glomerulonephritis, interstitial disease, polycystic kidney disease, obstruction, other and unknown. Provinces and territories were categorized as geographic regions as follows: Atlantic (New Brunswick, Nova Scotia, Prince Edward Island, Newfoundland), Central (Ontario), Prairies (Alberta, Saskatchewan, Manitoba, Nunavut, Northwest Territories), Pacific (British Columbia, Yukon). Any pre-dialysis care was defined as contact with a Nephrologist for 30 or more days prior to hemodialysis initiation. Distance to centre was calculated as the direct linear distance in kilometres between a patients postal code from their primary residence at dialysis initiation to the nearest dialysis provider. Comorbidities and laboratory data were ascertained at the onset of ESRD. The outcome of interest was all-cause mortality and follow-up for outcomes was until December 31st, 2009.
Statistical analyses
Continuous variables of interest were summarized as mean or medians with standard deviation or inter-quartile range as appropriate. Differences in baseline characteristics were determined by student’s t-test for continuous variables and chi-square or the Mann–Whitney test for dichotomous variables.
To assess the relationship between the sexes and mortality, we examined both traditional cause-specific cox proportional hazards model and the modified risks regression model of Fine and Gray to account for the competing risk of transplantation [24]. Models were adjusted for demographics, co-morbidities, body mass index (BMI), distance from centre, pre-dialysis care, cause of ESRD, geographic region, serum hemoglobin and albumin. The competing risk accounted for in the Fine and Gray models was renal transplantation.
To assess whether age was an effect modifier for mortality among the sexes, an age X sex interaction term was added separately to the traditional cox and competing risks models. In separate models, age was entered both as a continuous variable and categorical variable (age < 45, 45–55, 55–65, 65–75, 75–85, 85+ years). Categories were created based on the number of death events and individual models were created per age stratum with men as the referent. As there are few kidney transplants with increasing age, age categories were collapsed in the competing risk models.
Multiple imputation was employed for missing values with a random draw from the predictive distribution from an imputation model repeated ten times [25]. The proportion and covariates with missing data that was imputed included BMI (8.2%), distance from centre (2.0%), cause of ESRD (2.9%), geographic region (0.2%), serum hemoglobin (15.4%) and albumin (21.5%). An iterative Markov chain Monte Carlo (MCMC) method was used and pooled estimates of 10 rounds of imputation reported. Imputed and non-imputed models were compared and as there were no substantive changes in point estimates, the pooled imputed results of 10 rounds of imputation were reported. Analyses were performed using PASW Version 18 and the Fine and Gray analyses were performed using R. All hypothesis tests were two sided with statistical significance to find as having a P value of <0.05.