Comparing Mortality of Elderly Patients on Hemodialysis versus Peritoneal Dialysis: A Propensity Score Approach

Materials and Methods

Results

The frequency and distribution of important patient characteristics are presented in Table 2. Patients starting RRT on PD were less likely to be black, low SES, or to have undergone major abdominal surgery. Furthermore, PD patients were more likely to have coronary artery disease, congestive heart failure, and mental disease other than depression compared with those starting on HD, whereas the latter group tended to have more peripheral vascular disease, and a significantly higher proportion had history of alcohol and/or substance abuse. Health care utilization patterns were also different between patients starting on PD versus HD.

For the outcomes analysis, we first built a Cox proportional hazards model that contained all demographic covariates, comorbidities, and renal diagnoses listed in Table 1. We omitted the covariates for diabetic nephropathy and hypertensive kidney disease due to their expected high collinearity with the respective comorbidities. The covariate depicting late versus early nephrologist referral was also included. This model revealed that patients who started their RRT with PD had a 24% higher rate of death during the first year of treatment compared with those starting on HD (hazard ratio [HR], 1.24; 95% confidence interval [CI], 1.09 to 1.41). However, we found the modality-time interaction term to be highly significant (P < 0.001), thus violating this assumption of the Cox proportional hazards model. We therefore proceeded with building interval-Cox models, stratifying our analyses into 90-d intervals. Table 3 describes the four interval-cohorts with regard to number of enrollees, person-time accrued, and the number of deaths observed.

Figure 1 shows that patients initially assigned to PD had a 23% higher mortality rate than HD patients during the first 90 d after initiation of RRT (HR_months1–3, 1.23 [1.04 to 1.46]), whereas there was no mortality differential in the second 3-mo period (HR_months4–6, 1.05 [0.76 to 1.45]). The relative death rate thereafter increased again (HR_{months 7 to 9}, 1.28; (0.90 to 1.83]). In the fourth quarter-year of RRT, the mortality rate of PD patients was again higher compared with HD patients. Individuals who had started RRT on PD died at a 57% higher rate than did HD patients during that interval (HR_months9–12, 1.57 [1.05 to 2.36]).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1. Relative death rates of peritoneal dialysis (PD) versus hemodialysis (HD) patients (hazards ratios [95% CI]).

We next attempted to control for center effects in two ways: first, by introducing dummy covariates for each treatment site into the PS model. The PS including center identification did particularly well in distinguishing future PD from HD patients, with a c statistic of 0.82 (29). Second, we adjusted for potential differences in the underlying background death rate among centers by using center as a blocking variable in the proportional hazards models.

The results obtained from applying either or both of these techniques were not substantially different. Controlling for center in either of these ways attenuated the effect estimates slightly (Figure 1). However, confidence intervals were wider than in the models that did not block by center or use PS. Nevertheless, the overall pattern of survival differences remained robust: patients starting on PD died at a 16% higher rate during the first 90 d of treatment (HR_months1–3, 1.16; (0.96 to 1.42]), whereas there was no difference in the second 3-mo interval. Thereafter, those who began dialysis with PD demonstrated a higher mortality again. A Kaplan-Meier actuarial survival plot by modality from the propensity score–matched cohort can be found in Figure 2.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2. Kaplan-Meier survival plot in propensity score matched pairs (n = 1062).

Because the hazard ratios observed in the third and fourth interval were similar, we conducted a post hoc analysis combining those intervals. We found a statistically significant 45% higher mortality rate among patients initially assigned to PD compared with HD in the interval from 7 to 12 mo after treatment assignment (HR_months7–12, 1.45; [1.07; 1.98]). The detailed results for all models assessing the stratified relative interval death rates on PD versus HD in the overall cohort are shown in Table 4.

Subgroup and Sensitivity Analyses

We did not find any statistically significant interactions between patients with or without diabetes (P > 0.30 in all interval Cox models). However, previous reports have suggested that such effect modification may in fact be present (2,6,8–10,20). We therefore conducted separate analyses (including PS and center adjustment) among patients with diabetes and those without a diagnosis of diabetes (Figure 3). Although the finding of higher mortality in the first 90 d among PD starters was present in patients with and without diabetes, we found that the excess death rate observed beyond 180 d of RRT was solely driven by patients with diabetes. In this later time period, mortality was not different between patients without diabetes who started RRT on PD versus HD.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3. Subgroup analysis of patients with versus without diabetes (hazards ratios [95% CI]).

We repeated all analyses on a population subset excluding those patients who switched modality at least once during the first month of RRT (n = 182) and also those for whom we were unable to identify the first treatment modality without making inference from presence of peritoneal dialysis catheters (n = 106). These results were essentially unchanged when compared with the full cohort (Figure 4). We also evaluated the effect of modality choice on survival among those patients who first consulted a nephrologist more than 90 d before onset of RRT and who did not switch during the first month of therapy. In this group of patients, who could be assumed to have made an educated decision and to have had sufficient time to be prepared for treatment, we found an even more pronounced result (Figure 4).

• Download figure
• Open in new tab
• Download powerpoint

Figure 4. Sensitivity analysis (hazards ratios [95% CI]).

Discussion

In this large cohort of elderly ESRD patients followed during their first year on RRT, we found that patients initially assigned to PD had a higher overall mortality rate than patients initially assigned to HD (HR, 1.24; 95% CI, 1.09 to 1.41). We furthermore found that the relative survival pattern between treatment modalities changed over time (Figure 1). Stratifying by 3-mo intervals, we found that patients starting RRT on PD were 16% more likely to die during the first 3 mo of treatment (HR, 1.16 [0.96 to 1.42]). Between 3 and 6 mo after beginning of RRT, we observed similar mortality rates among patients assigned to PD and HD (HR, 1.03; 95% CI, 0.71 to 1.51). However, in the final 6 mo of the first year of RRT, patients initially assigned to PD had a 45% higher mortality rate than HD patients (HR, 1.45 [1.07 to 1.98]). These results were robust in multivariate models that controlled for a large number of demographic, medical, and health care utilization characteristics. Additional bias-reduction was introduced with the use of propensity scores, as well as by controlling for center effects.

This analysis of the death rates of patients starting RRT on HD versus PD was more detailed than previous analyses. To study the relative mortality between these two treatment modalities, we selected from a large database a historic cohort of incident ESRD patients whose kidney disease had developed chronically, rather than occurring acutely. The experiences of those patients were then followed prospectively in light of their first treatment assignment.

Our finding that patients starting RRT on PD were at a higher mortality risk within the first 3 mo of treatment has to our knowledge not been described before. Most earlier studies did not test for or failed to detect violations of the proportional hazards assumption, and thus chose to present effect estimates that were aggregated over time. Alternatively, many analyses, especially in the United States, omitted the first few months of RRT from analysis entirely. Interestingly, we found no difference in survival between patients assigned to PD or HD in the period from 4 to 6 mo after initial treatment allocation. This finding may be explained by depletion of susceptibles: (32) many patients who did not fare well on PD died, and those still alive had accommodated to the treatment.

Previous reports have suggested that mortality between PD and HD patients differed across important patient subgroups, most prominently with regard to age and presence of diabetes (2,6,8–10,20). Results for patients with and without diabetes were very similar in our study when we did not control for PS or center effects (not shown). However, when introducing such techniques, we found the effects of modality choice on mortality among patients with diabetes to be more pronounced but not present for patients free of diabetes. This is consistent with previous findings that demonstrated that patients with diabetes had higher mortality when on PD rather than HD, which was not the case among patients without diabetes (8,9).

A central purpose of this study was to model and analyze the consequences of real-life baseline decision-making. However, we could not automatically assume that the modality used for the first dialysis was in congruence with the decision that had been made by the patient and her/his physician, due to inadequately late nephrologist consultation or to complications during the preparation process. For example, it may have been decided to use HD as the modality for long-term treatment, but at the time of first dialysis the shunt had not sufficiently matured, resulting in this interval being bridged using PD. Other scenarios might go in the opposite direction. To minimize such “misclassification,” we also presented results on a subset of patients who did not switch modality during the first month of RRT, assuming that their initial modality corresponded to their choice for long-term maintenance treatment. For this step, we excluded all first-month switchers as well as those patients for whom we had to make any inference regarding their first treatment modality from the presence or absence of claims for insertion of a PD catheter. In a second step, we took this point even further and eliminated those patients who had not consulted with a nephrologist until ≤3 mo from RRT. Seeing a nephrologist earlier gives the patient a considerably better chance to make an educated decision about which modality to use as well as sufficient preparation time for a smooth initiation of treatment. However, even applying such restriction criteria did not change our findings.

Most previous studies have not assessed treatment modality at baseline, but rather at an arbitrary later point in time, most often 90 or 120 d after first dialysis. In the United States, this is for one practical reason: availability of data that are collected from medical claims. The large nationwide registry of patients with ESRD, the United States Renal Data System (USRDS), is based on the medical claims submitted to Medicare as the predominant payor of ESRD services. However, Medicare does not become a principal payor until day 90 of RRT, which leaves the experiences of ESRD patients during the first 90 d of treatment in obscurity. For the purpose of comparing outcomes of initial therapy, this can be highly problematic, because it is of paramount importance to assess treatment assignment from the first day of treatment. Only if such left-censoring were completely noninformative, i.e., nondifferential with regards to all exposure and confounder characteristics, could an analysis starting at a later point potentially be unbiased. This is usually not the case, and certainly not in the ESRD population as demonstrated here and in earlier work (33). In the present context, such biases would tend to favor PD. First, the relatively more numerous early deaths among PD starters would be ignored, and their more frequent modality switches discarded (33). Second, studies of the survival of HD versus PD patients assessed after day 90 would be conducted in a population that was unrepresentative of all new RRT starters. Such information would overestimate the benefits of PD and understate those of HD. Third, by ignoring the number and direction of earlier modality switches, a substantial proportion of individuals labeled as “HD patients” at 3 mo would be misclassified with regard to their original treatment choice. It is in this way that the present study differs from previous studies that used interval analyses but assessed treatment on day 90 (20).

Two statistical aspects of the present study are also noteworthy. The present analysis went beyond the multivariate modeling and stratification that had been done before by the introduction of PS to further control for nonrandom treatment assignment and by adjusting for center effects. The value of PS lies in their ability to efficiently control for baseline differences between treatment groups using a single scalar (27,28). This is particularly appealing when prediction of treatment assignment is complex and correct model specification would require introduction of multiple interaction or higher-order terms. This may lead to difficulties in interpretability when one wishes to evaluate and estimate main effects. However, although introduction of PS reduces bias introduced by imbalances of observed covariates, it does nothing to reduce bias caused by unobserved confounders.

Prediction of choice of dialysis modality could be expected to be poor in the absence of information on factors known to be important for this decision, such as financial incentives for providers of RRT, educational deficits in the health care team, and/or patient, physician bias, resource availability, social mores, and cultural habits (34). However, it seemed reasonable that some of this predictive power could be captured by including information into the PS model that identified the individual dialysis facility. Indeed, the c statistic increased substantially from 0.64 to 0.82 when facility covariates were added into the model (29). A number of recent articles have shown that various outcomes are highly associated with center-specific factors even after adjustment for individual patient characteristics and that center characteristics can be even stronger predictors of outcomes than individual patient predictors (35–37). This dominance of center-specific factors also expressed itself in the large improvement of the c statistic of our propensity score model when entering dummy covariates for each center. This clearly indicates that center policies regarding modality assignment were important determinants of treatment choice beyond individual patient characteristics. When adding quintiles of estimated PS into the survival models, the results did not change very much in the overall population; we found marginal to moderate attenuation in the effect estimates toward the null value (Figure 1).

Outcomes such as mortality are also correlated with the center where treatment is received. Mortality among ESRD patients is highly variable across centers, but clustered within centers, even when controlling for patient characteristics (36,37). Failure to account for this phenomenon may lead to biased effect estimates and reduced power (23). Various analytical techniques are available to control for such center-effects. We decided to account for such effects in our analysis by blocking on center. Such methodology permits efficient estimation of common relative mortality rates among strata of individuals, but allows baseline rates to vary by center (23).

These findings should be considered in light of the study’s limitations. Although its findings are externally valid for patients aged 66 or older, they are probably not generalizable to the ESRD population <66 yr of age, and their generalizability to other geographic regions is also unclear. The present study may also suffer from shortcomings inherent in all claims database research: missing data, miscoding, as well as residual confounding arising from overly crude categorization of confounding conditions such as comorbidities or socioeconomic status. Furthermore, we do not have any information on clinical and biologic parameters, patient compliance with treatment, or dose of dialysis prescribed or delivered. It is also not clear whether the findings from the early and mid-1990s are representative of current patient outcomes. The techniques of both HD and PD are evolving over time, and associations between outcomes and delivered dose of dialysis were discovered earlier for HD compared with PD (38–40). Although we did not find any significant trends over time, our findings will still need to be confirmed in more recent cohorts of patients.

Finally, the value of this analysis to patients, health care providers, and policymakers depends on whether the modality chosen at baseline reflected actual long-term treatment decisions. We believe that our population selection algorithm and the robustness of our findings after application of stringent criteria regarding timely nephrologist care and early modality switching in sensitivity analyses make a compelling argument that this was actually the case.

We confined our analysis to the description of associations between treatment assignment and mortality in a 1-yr follow-up. The present study stopped short of describing the different causes of deaths, evaluating the effects of center size on outcomes, or providing an analysis beyond the first year of RRT. We found that the population at hand was too small to answer more detailed questions or to assess outcomes further down the road in a meaningful way.

Conversely, there are several characteristics of this study that address problems inherent in past studies with similar objectives. Most important of these is its use of an inception cohort of new starters of RRT, drawn from a variety of settings in an entire state. Thus, the important biases arising from follow-up after a later point in time are not present, and the results are representative of patients receiving care in typical settings. Furthermore, we were able to assemble information on more patient characteristics than had been available in previous studies. From such information we calculated PS, which enabled us to reduce distortions originating from selection bias even further in this observational study. By including information on centers, we improved prediction of treatment assignment considerably and we controlled for variations in treatment practices across institutions (23). Such control for center effects had been infrequent in analyses of mortality in RRT (11). Furthermore, analysis of data stratified by time took account of the fact that mortality hazards between treatment modalities varied over time. This provided important insights into the time course of excess risks among PD patients. Overall, this study complements and confirms earlier analyses that have shown that older patients with diabetes are at an increased mortality risk on PD compared with HD (19,20).

Conclusions

These data suggest that elderly patients who start RRT on peritoneal dialysis are at a higher risk of death both in the very early phase as well as during later courses of treatment, compared with comparable patients who begin their RRT on hemodialysis. We confirm earlier studies that found excess mortality among older patients, especially those with diabetes on PD (19,20). Our findings are based on the experience of a large and typical cohort of incident ESRD patients aged >65 in a large eastern state of the United States rather than of a single center with above-average experience with PD (2,15,22), or from populations with higher overall utilization of PD relative to HD than is seen in the United States that may not be representative of the way renal health care is delivered in this country (4,6,11,17,21). If replicated in other studies, these results may be relevant to the decision of patients nearing ESRD and their physicians in making decisions about which treatment to choose for long-term RRT. Other important considerations for such treatment decision are quality of life, access and availability, and the associated costs. Application of similar approaches to large and nationally representative prospective inception cohorts of ESRD patients seems important to confirm our findings.