Year : 2007 | Volume
: 18 | Issue : 1 | Page : 24--30
Comparing Outcomes on Peritoneal and Hemodialysis: A Case Study in the Interpretation of Observational Studies
Simon J Davies
Institute for Science and Technology in Medicine, Keele University; Department of Nephrology, University Hospital of North Staffordshire, Stoke-on-Trent, United Kingdom
Simon J Davies
Department of Nephrology, University Hospital of North Staffordshire Princes Road, Hartshill, Stoke-on-Trent ST4 7LN
|How to cite this article:|
Davies SJ. Comparing Outcomes on Peritoneal and Hemodialysis: A Case Study in the Interpretation of Observational Studies.Saudi J Kidney Dis Transpl 2007;18:24-30
|How to cite this URL:|
Davies SJ. Comparing Outcomes on Peritoneal and Hemodialysis: A Case Study in the Interpretation of Observational Studies. Saudi J Kidney Dis Transpl [serial online] 2007 [cited 2020 Oct 21 ];18:24-30
Available from: https://www.sjkdt.org/text.asp?2007/18/1/24/31841
The fact that more than one type of dialysis modality exists, specifically peritoneal (PD) and hemodialysis (HD), is in itself, a cause for celebration. It should be seen as increasing options for patients who are already considerably disadvantaged in life, providing choice to support differing lifestyles as well as potentially enabling longer periods of renal replacement, especially when used sequentially and collaboratively.  All too often, however, these two modalities are simply viewed in a competitive light. There are reasons for this of course; both good and bad, depending on whether it is the patients' or the nephrologists' prejudices that are the driving force. We do need to know that patients are making choices that do not disadvantage themselves in survival terms.
Recent studies examining why patients choose one modality over another have found that, in general, this choice is governed by lifestyle and circumstances. Indeed, most patients decline randomization to modality: no less than 96% of those who were in a position to choose in the Netherlands based NECOSAD study.  Both this study and the CHOICE study found a number of factors associated with opting for HD or PD.  PD patients were more likely to be relatively young, less co-morbid, enjoy better health status, wish to work, be better educated, not live alone, live further from their HD unit and to have had a better preparation for renal replacement therapy. Their initial survival should be better and it is; the question is, therefore, is it better enough? Or put another way, is it safe to say that modality is a lifestyle choice or should it be tempered by survival concerns?
Studies that have attempted to answer this question are fraught with analytical difficulties. In fact, the history of these analyses over the last 10 to 15 years can be seen as a case study of these problems. Thus, the purposes of this article are to review the literature comparing survival according to initial dialysis modality and at the same time introduce the reader to the pitfalls in analysing observational comparative data. It will also try to answer the question!
The Perfect Study?
The architects of the NECOSAD study are to be congratulated in their attempt to do the perfect study.  Designed as a randomized control trial with survival and quality adjusted life-years as the endpoints, the problem with recruitment meant that in the end it was severely underpowered, with just 38 patients. Compared to the largest observational studies of almost half a million patients, this is clearly a problem. In this context, a significant difference, as judged by a P value, has little value as the result cannot be generalized in any meaningful way.
Nevertheless, it is worth looking a little closer at the outcomes, if only to see how this informs our interpretation of observational studies. Immediately, we see one of the problems in comparing treatment modalities that have a complementary role in patient management; patients switch therapies. Put transplantation, as a third modality into the picture, then it becomes even more complex. In NECOSAD, the proportion of patients who initiated treatment on a modality different to that to which they were randomized, the likelihood of switching dialysis modality or being transplanted differed by modality. Again, the numbers are too small to be certain that this would be a phenomenon that could be generalized, but it does influence the analysis. This problem can be dealt with by comparing the outcomes using an intention to treat versus an 'as treated' approach. These authors did this and they found that whilst there appeared to be a 5-year survival advantage to patients randomised to PD, this disappeared when the 'as treated' analysis was performed. In other words, the randomized patients tended to reflect what is seen in real life; it is difficult to reduce decision making to a single event in the life of a patient with chronic disease in which many events occur and different therapies are available.
Pitfalls of Observational Studies
So far we have identified at least two of the problems that beset observational studies when comparing treatments: the important differences that exist between individuals when opting for their chosen modality, and the problem of therapy switching. There are others however. These include the type of study design; for example, prevalent versus incident cohort, lead-time bias, informative censoring, nonproportionality of risk with time and subgroup analysis [Table 1]. Each of these issues, with examples from the literature [Table 2] will be considered in turn.
Prevalent Versus Incident Study Populations.
In 1995, Bloembergen et al. published a study concluding that mortality in PD patients was 19% higher than HD patients.  If true across the board, and this data was corrected for age and diabetic status, then this would be of great concern. They used a point prevalent design. This means that they looked at who was on dialysis on the first day of each year for 1987-89. Any patient who lived less than a year, by falling between these dates would be missed, thus giving a bias against the modality associated with better shortterm outcomes. Taking the same patients and time period Collins et al. re-analyzed this data using the prospective incident cohorts.  They found, using the same adjustments for age and diabetes, that now there was a survival advantage for most patients treated with PD [Table 2]. The only sub-group for whom this was not the case, which accounted for 18% of the total patient number, were diabetic women over the age of 55 years. This direct comparison of data analysis techniques shows very clearly how important study design is and the need to interpret prevalent studies with extreme care.
Non-proportionality of Risk
Meanwhile, the Canada Organ Replacement Registry (CORR), a comprehensive, countrywide, prospective national database reported overall, a superior survival on PD compared to HD .  However, it was clear from this analysis that the relative advantage of PD over HD was not uniform, but concentrated within the first two years. In other words, the relative risk-benefit of PD and HD differs with time spent on the therapy. This phenomenon has been described in most subsequent incident cohort studies. For example, the Danish Registry, again a whole nation comprehensive database with carefully documented comorbidity, including whether or not patients were listed for transplantation as an independent covariate, found an early relative advantage to PD that was almost identical to the CORR study.  Another common theme to these studies is that the relative advantage to PD seems to be shorter lived in the US and also for older diabetic patients. The relevance of this will be discussed later. From an analytical viewpoint, it is important that the correct tool is used, specifically avoiding the Cox proportional hazards model that assumes risk of dropout is proportional in the two modalities under comparison.
Lead-time Bias and Informative Censoring
What might be the reason for this early PD advantage and what statistical pitfalls could be occurring? Clearly, one reason might be that PD and HD patients are very different. These studies all adjust for the known predictors of mortality, but there are factors we do not measure so easily that might be important. Approaches to this will be discussed later. There are other factors that need to be taken into account, however. PD patients might benefit because they are more likely to commence dialysis in a planned fashion and thus start earlier. This would result in what is termed lead-time bias. Effectively, this means that different patient groups are commencing the study at different points along the survival curve that characterizes renal failure. As we know from most studies that residual kidney function influences survival, this must be taken into account. This can be done by correcting for the level of residual kidney function at the start of dialysis, as was done in the NECOSAD analysis.  However, this group also corrected for longitudinal changes in residual renal function (RRF). Several studies, including NECOSAD (nonrandomized patients) have found that RRF is relatively preserved in PD and may indeed be one of the explanations of the relative early survival benefit. Alternatively, it can be seen as essential to good outcomes in PD resulting in a disproportionate switch of modality in anuric patients to HD. This confounding process is called informative censoring, and it occurs when a clinical factor that influences survival in its own right changes with time and causes dropout in a different manner in the two modalities being compared. It can be avoided to some extent by using 'as treated' rather than 'intention to treat' approaches but to get to the truth more sophisticated statistical approaches are required. 
Returning to the NECOSAD analysis of non-randomized patients, this was one of the few incident cohort patients not to find an early survival advantage to PD. This is in part due to lack of power, as the pattern of relative risks with time (early advantage PD, late advantage HD, [Table 2]) was the same as seen in much larger registry studies. It could, however, also have been affected by their decision to correct not only for RRF at the start of PD, which would remove an unfair bias in favor of PD, but also their correction for continued preservation of RRF over time, which might be seen as a positive attribute of the modality and therefore, a fair and important bias towards PD. It would be interesting to see a re-analysis of this data separating out the two effects.
In 2003 Stack et al. and Ganesh et al. published two rather similar analyses of patients in the same incident cohort of US patients, suggesting that in addition to concerns over PD use in older diabetic patients, there was evidence that patients with either coronary artery disease or heart failure had worse outcomes on PD ,  In doing subgroup analyses such as these, it is
important to recognize that there are modality-age-co-morbidity interactions that should be taken into account. When the analytical approach taken by these authors was compared with that which does take these interactions into account, as in the larger cohort study by Vonesh, a better fit of the data was obtained.  The picture is in a sense less complicated than these studies would suggest. Essentially, being young with less co-morbidity, especially diabetes, is associated with better outcome in PD whereas being older and diabetic, the reverse tends to be true. As Foley has stated in his excellent review of this topic,  'analytical edifices built on subgroups have shaky foundations; the most credible findings come from analyses that include all subjects, with adjustment for the factors later used to define subgroups.'
Correcting for Differences.
As indicated at the start of this article, patients who choose to do PD rather than HD are different in many ways. To correct for these differences, statisticians take measurable factors that are known to predict survival, such as age and comorbidity, and adjust for these in their analyses. Unfortunately however well we measure co-morbidity, this only accounts for about 15-20% of the variance observed in survival. Could other factors, such as education and less clearly defined socioeconomic factors explain the rest of this variance, or is it unpredictable? This is of course the power of randomized studies; they should equally distribute the un-measurable determinants of outcome. Another approach to this problem is to use a technique called propensity scoring. The idea behind this is that by identifying the measured covariates that are associated with choosing a therapy at the start of the study, the unmeasured covariates will correlate with these. This is obviously a major assumption. It was used by the CHOICE study in their analysis, which also for the first time included many important factors such as employment status, education, distance from the unit and marital status.  The similarity between the findings using propensity scoring and traditional measurable covariate analysis in this study encourages us to think that these various factors do indeed correlate. There are many problems with this study in terms of comparison of survival on PD and HD which include lack of power, systematic over-sampling of PD patients (which was not offered as a modality choice in half of the units), the inclusion of lab results obtained after the start of treatment (effectively using the future to predict the past) and perhaps most importantly a finding that is at complete odds with the rest of the literature; that the PD patients who did the least well compared to HD were the those who were younger and not diabetic. Whilst the authors should be congratulated on the detail of the study related to choice and the use of propensity scoring, there is a danger that this study will be overinterpreted when it comes to the survival analysis that was significantly flawed.
What Can We Conclude?
Looking at the overall picture, as summarized in [Table 2], the important message is that survival, whether commencing renal replacement with PD or HD is very similar. The pattern that emerges is that starting with PD makes sense, when preferred for life-style reasons and that greater care needs to be taken when using PD in older, diabetic patients. It is also important not to overestimate the clinical relevance of a relative risk ratio in this context. For example, if the typical survival length of a co-morbid, elderly diabetic patient on dialysis is 24 months, a relative increased mortality risk of 1.15 only translates into 3-4 months difference in survival. Statistical differences will be found with very large numbers but these are unlikely to outweigh the clinical differences, lifestyle preferences and thus quality of life issues for the individual patient.
|1||Van Biesen W, Vanholder RC, Veys N, Dhondt A, Lameire NH. An evaluation of an integrative care approach for end-stage renal disease patients. J Am Soc Nephrol 2000; 11(1):116-25.|
|2||Jager KJ, Korevaar JC, Dekker FW, et al. The effect of contraindications and patient preference on dialysis modality selection in ESRD patients in The Netherlands. Am J Kidney Dis 2004;43(5):891-9.|
|3||Jaar BG, Coresh J, Plantinga LC, et al. Comparing the risk for death with peritoneal dialysis and hemodialysis in a national cohort of patients with chronic kidney disease. Ann Intern Med 2005;143(3):174-83.|
|4||Korevaar JC, Feith GW, Dekker FW, et al. Effect of starting with hemodialysis compared with peritoneal dialysis in patients new on dialysis treatment: a randomized controlled trial. Kidney Int 2003;64(6):2222-8.|
|5||Bloembergen WE, Port FK, Mauger EA, Wolfe RA. A comparison of mortality between patients treated with hemodialysis and peritoneal dialysis. J Am Soc Nephrol 1995;6(2):177-83.|
|6||Collins AJ, Hao W, Xia H, et al. Mortality risks of peritoneal dialysis and hemodialysis. Am J Kidney Dis 1999;34(6):1065-74.|
|7||Fenton SS, Schaubel DE, Desmeules M, et al. Hemodialysis versus peritoneal dialysis: a comparison of adjusted mortality rates. Am J Kidney Dis 1997;30(3):334-42.|
|8||Heaf JG, Lokkegaard H, Madsen M. Initial survival advantage of peritoneal dialysis relative to haemodialysis. Nephrol Dial Transplant 2002;17(1):112-7.|
|9||Termorshuizen F, Korevaar JC, Dekker FW, Van Manen JG, Boeschoten EW, Krediet RT. Hemodialysis and peritoneal dialysis: comparison of adjusted mortality rates according to the duration of dialysis: analysis of the Netherlands cooperative study on the adequacy of dialysis 2. J Am Soc Nephrol 2003;14(11):2851-60.|
|10||Vonesh EF, Greene T, Schluchter MD. Shared parameter models for the joint analysis of longitudinal data and event times. Stat Med 2006;25(1):143-63.|
|11||Stack AG, Molony DA, Rahman NS, Dosekun A, Murthy B. Impact of dialysis modality on survival of new ESRD patients with congestive heart failure in the United States. Kidney Int 2003;64(3):1071-9.|
|12||Ganesh SK, Hulbert-Shearon T, Port FK, Eagle K, Stack AG. Mortality differences by dialysis modality among incident ESRD patients with and without coronary artery disease. J Am Soc Nephrol 2003;14(2):415-24.|
|13||Vonesh EF, Snyder JJ, Foley RN, Collins AJ. The differential impact of risk factors on mortality in hemodialysis and peritoneal dialysis. Kidney Int 2004;66(6):2389-401.|
|14||Foley RN. Comparing the incomparable: hemodialysis versus peritoneal dialysis in observational studies. Perit Dial Int 2004;24(3):217-21.|