Survival comparisons between haemodialysis and peritoneal dialysis

Dialysis modality selection for end-stage renal disease patients should not solely be dictated by survival comparisons but also take into account patient preference. Nevertheless, potential mortality differences between dialysis modalities in (subgroups of) patients may contribute to modality choice. Survival comparisons have therefore frequently been made. With one exception, the investigators used an observational study design to study this issue. Observational studies fulfil a valuable role in nephrology research, but their most important drawback is that selection bias by the clinician may occur [1]. Even after adjustment for potential confounders in the statistical analysis, there is usually at least some amount of residual confounding due to unmeasured variables. This may prevent a fair comparison of outcomes between patient groups, something that is usually feasible from well-conducted randomized controlled trials. Almost a decade ago, such a trial with random allocation of dialysis modality was unsuccessful because patient and physician preference turned out to play a crucial role in modality choice [2]. Despite this unsuccessful attempt, a new trial has started in China (trial registration NCT01413074 at clinicaltrials.gov). However, until the results of this trial are presented, mortality in haemodialysis and peritoneal dialysis patients can only be compared based on large-scale observational studies. Even though there is some heterogeneity in the results, these observational studies usually indicated that the mortality risk on haemodialysis treatment compared with that on peritoneal dialysis treatment changes over time, with the lowest relative risk for patients on peritoneal dialysis in the first 2 years of therapy. After these first 2 years, mortality risk increases for those who started on peritoneal dialysis and patient survival becomes similar for haemodialysis and peritoneal dialysis patients, or even somewhat better for patients on haemodialysis [3–5]. This finding could be attributable to residual confounding. For example, it has been postulated that patients who start dialysis urgently are at high risk of death and as they are treated predominantly with haemodialysis. This could induce selection bias in the comparison of mortality between haemodialysis and peritoneal dialysis patients. Couchoud et al. [6] showed in 2007 that mortality risk was significantly increased with 50% among elderly patients (75 years or older) with an ‘unplanned’ start of haemodialysis when compared to patients with a ‘planned’ start suggesting that a comparison between both dialysis modalities would be more balanced after removing the unplanned haemodialysis starts. Quinn et al. [7] recently confirmed the findings of Couchoud et al. by showing that haemodialysis and peritoneal dialysis were associated with similar survival in incident patients starting dialysis electively as outpatients. A planned start is usually not taken into account in survival comparisons of haemodialysis and peritoneal dialysis. Another point frequently not taken into account is whether haemodialysis and peritoneal dialysis are provided in a state-of-the-art manner. For example, in many survival comparisons, the type of vascular access used for haemodialysis is not included in the analyses, whereas Perl et al. [8] recently showed that type of vascular access plays an important role in the relationship between dialysis modality and mortality. They found in a Canadian cohort that starting haemodialysis with a central venous catheter largely explained the higher early mortality risk of haemodialysis. In order to reduce the influence of selection bias and confounding, research groups started to re-assess the associations between dialysis modality and mortality risk in large cohort studies using more advanced statistical methods, in addition to the conventional methods of survival analysis, i.e. Kaplan–Meier and Cox proportional hazards models. Examples of such methods are the use of time-dependent covariates in survival analysis [4], marginal structural models [9, 10] and the use of treatment propensity scores in statistical models by means of adjustment, stratification or matching [4, 11, 12]. In the current issue of Nephrology Dialysis Transplantation, Yeates et al. [13] applied different statistical methods to compare the survival of haemodialysis and peritoneal dialysis patients using data on >35 000 incident dialysis patients from the Canadian Organ Replacement Register. They performed both a standard intention-to-treat analysis and a time-dependent as-treated analysis for which proportional and non-proportional hazards models were built to compare mortality risks between both groups. For the non-proportional hazards analyses, they used a piecewise exponential survival model, using successive 6-month intervals in the first 5 years of dialysis treatment. Contrary to their hypothesis that over time survival had worsened for peritoneal dialysis when compared to haemodialysis treatment, they found that overall adjusted survival remained similar for haemodialysis and peritoneal dialysis even in the most contemporary cohort. After stratification for diabetic status, age and sex, they found that non-diabetic patients in the youngest age group (<45 years) showed survival benefits on peritoneal dialysis treatment, while there was no difference in mortality in the older age groups. However, survival on peritoneal dialysis was worse than on haemodialysis among patients with diabetes, in particular in those with higher age, which confirms the findings from previous studies from the USA and Europe [6, 10, 14]. Yeates et al. used a similar approach as that presented in a recent paper by Mehrotra et al. [10] who performed the same analyses among patients from the United States Renal Data System. In addition, Mehrotra et al. used propensity score weighting and proportional and non-proportional hazards marginal structural models with inverse probability weighting. Mehrotra et al. concluded that in the most recent cohort (2002–2004), there was no significant difference in the risk of death between patients who started on haemodialysis or peritoneal dialysis. In their paper, Yeates et al. state to have re-assessed the survival of peritoneal dialysis versus haemodialysis patients because the characteristics of dialysis patients in Canada changed over time; patients tend to be older and carry higher levels of comorbidity. The authors hypothesized that as a result, survival on peritoneal dialysis as compared to haemodialysis would have worsened in the study period between 1991 and 2007 and in particular in the most contemporary cohort [13]. This turned out not to be the case. This leaves the possibility that the case-mix of peritoneal dialysis patients has changed to a lesser extent than that of haemodialysis patients for example because of the (stricter) patient selection for this modality. Unfortunately, the authors did not make a distinction between haemodialysis and peritoneal dialysis patients when summarizing the trends in demographic and clinical characteristics over time. Therefore, the potential role of (differences in trends of) patient characteristics in the survival comparison could not be evaluated. In addition, in survival studies of dialysis patients, renal transplantation is one of the most common reasons for censoring participants. This may induce substantial bias in the survival comparison because renal transplantation is in fact an event that ‘competes’ with death and is thus a so-called competing risk for mortality. Especially, since in some countries peritoneal dialysis patients are considerably more likely to receive a kidney transplant when compared to haemodialysis patients, the censoring for transplantation is a potentially important source of bias when comparing patient survival. Also, the increasing waiting time for a transplant leaving relatively healthy patients on dialysis for a longer period may have ‘improved’ patient survival on dialysis. Although the authors recognize this problem of competing risks in their discussion section, they did not apply any of the available statistical methods to take informative censoring into account, such as competing risks methods for survival analysis or marginal structural models with weighting for the inverse probability of censoring, as was done in earlier similar studies [9, 10]. In conclusion, the findings of Yeates et al. confirm those of previous studies. Because so far it has turned out impossible to perform randomized controlled trials, any evidence base for this treatment comparison has needed to come from observational studies, despite their drawbacks. Up to now, these observational studies have indicated that peritoneal dialysis is associated with better survival for some specific patient subgroups but with worse survival for other subgroups. This paper and previous papers have shown that also when using sophisticated statistical methods, overall survival on haemodialysis and peritoneal dialysis is similar. Nevertheless, we think that an important next step for the statistical analyses should be the use of competing risk methods to account for the difference in transplant rates between both patient groups (see related article by Yeates et al. [13]). None declared. (See related article by Yeates et al. Hemodialysis and peritoneal dialysis are associated with similar outcomes for end-stage renal disease treatment in Canada. Nephrol Dial Transplant 2012; 27: 3568–3575.)