Several studies have examined the possible association between late referral to a nephrologist and mortality on maintenance dialysis in patients with chronic kidney disease (1–8). Other analyses have measured the effect of LR on outcomes such as timely vascular access creation (2,9–18), modality choice for renal replacement therapy (RRT), and technique survival (19,20) and health care costs (21). These analyses have made the implicit assumption that the association between outcomes and the timing of referral was direct, controlling for all other variables in the respective underlying multivariate models. Typically, the variables contained in such multivariate models were of demographic nature (age, race, gender, socioeconomic status) or related to comorbid conditions. It was implied, but not proven, that the favorable associations in patients referred early rather than late were a consequence of the nephrologist’s contribution to their care. However, it is also possible that late nephrologist referral is merely a proxy for reduced access to adequate health care overall. If this were the case, some or all of the improved outcomes associated with early referral might instead simply be the result of better access to the healthcare system or an overall higher quality of care independent of nephrologist input.
Propensity scores are a tool that can correct for nonrandom exposure assignment and have been shown to reduce such potential bias in observational studies (22). The propensity score is the expected probability of receiving one treatment over another for a given patient, based on that patient’s baseline characteristics. Typically, a logistic regression model of the actual treatment received is fit to the data, and the propensity score for each patient is then estimated. Numerous covariates as well as higher order terms and covariate interactions can thus be condensed into a single scalar variable, the propensity score (23,24). Several studies have demonstrated that propensity scores can reduce bias in observational studies to balance observed baseline covariates in two treatment groups. Once estimated, propensity scores can be used for analyses of outcomes as covariates in multivariate models, as matching factors, or to define strata for separate analyses. The use of propensity scores in analyzing the association between timing of first nephrologist referral and mortality during the first year of RRT offers a novel approach to compensate for potential differences in healthcare access and utilization among late referral versus early referral patients.
Materials and Methods
The present analysis carried forward our previous findings of reduced survival on RRT in patients with delayed first referral to a nephrologist (8). We studied a large population of incident peritoneal and hemodialysis patients enrolled in the Medicaid, Medicare, or Pharmaceutical Assistance for the Aged and Disabled programs in the state of New Jersey who had progressed chronically rather than acutely to end-stage renal failure (ESRD) (n = 3014) and required chronic renal replacement therapy. To restrict the cohort in this fashion, we studied only patients who had been diagnosed with a renal disease >1 yr before first dialysis. We excluded patients who received only a single dialysis and survived >1 mo thereafter or who received a limited series of dialysis treatments and survived >2 mo. This approach defined an inception cohort of 3014 patients starting dialysis between 1991 and mid-1996. Timing of nephrologist referral was dichotomized; patients who saw a nephrologist >90 d before their first chronic dialysis were labeled as early referrals (ER), and all others were considered late referrals (LR). All traceable identifiers were removed before analysis to protect patient confidentiality. The study was approved by the institutional review board of our institution.
We calculated propensity scores using a logistic regression model, with timing of referral (LR versus ER) as the outcome studied. To build the propensity score model, we selected patient characteristics and healthcare utilization indicators in the 4 to 12 mo before the initiation of dialysis (Figure 1). We also included a number of higher order terms and covariate interactions (Tables 1 and 2). Model fit and predictive power were assessed using the c statistic. A c statistic of 1 indicates perfect prediction, whereas a value of 0.5 reflects random chance (25).
;)
Figure 1.:
Periods of covariate ascertainment.
Table 1: Covariates used to estimate propensity score for timing of nephrologist referral (≤90 d versus >90 d before first renal replacement therapy)
Table 2: Patient characteristics by timing of first nephrologist referral
For the outcomes model we built a multivariate Cox proportional hazards model. Here, death was the outcome of interest; patients were censored at the earliest of 365 d after first dialysis, renal transplantation, loss to follow-up, and end of study period. The covariates that were entered into this outcomes model were assessed in the year before first dialysis (Figure 1). Those covariates were age (continuous), gender, race (White/Black/other), socioeconomic status, and a number of baseline comorbid conditions. The original models had adjusted for underlying renal disease or for a limited set of nonrenal comorbidities (hypertension, coronary artery disease, congestive heart failure, diabetes malignancy), but not for differences in healthcare utilization (8). The baseline model used here differs from the one published previously in that it contains a large number of comorbid conditions rather than renal diagnoses to control for confounding (8). Propensity scores were then used in two ways (23). First, we entered the propensity scores (in quintiles) into the same model for which the initial findings were first described. Second, we ran the same model within a population of 1039 pairs of patients from the late referral group and the early referral group who were matched by their propensity score. The propensity matching was conducted using the “greedy match” macro (26). We then sought to determine whether such control over the level of healthcare utilization changed the association between late referral and mortality in the first year of maintenance dialysis. For sensitivity analyses, we determined whether a patient had initiated dialysis on peritoneal dialysis or hemodialysis. Among those on hemodialysis, we created two variables indicating whether a peripheral vascular access procedure was performed (a) before first dialysis and (b) >14 d before onset of hemodialysis. The specific algorithms used to identify patients who started on peritoneal dialysis and who underwent surgery for peripheral vascular access before RRT among starters on hemodialysis have been published in detail previously (18).
Results
Of the 3014 patients who comprised the cohort, 1429 patients died during the first year of dialysis. Twenty-one patients received a renal transplant during the first year and were censored. The remaining 1564 patients completed the 1-yr follow-up or were censored at the end of the database (June 30, 1996); no patient was lost for follow-up during the period of study. Table 3 shows the number of patients who died, were transplanted, or were censored by timing of referral and time period. The conventional model without inclusion of propensity scores indicated that patients who first saw a nephrologist ≤90 d before onset of dialysis had a 36% higher mortality rate compared with those who had their first nephrologist visit earlier (Hazards ratio [HR] = 1.36; 95% confidence interval [CI], 1.22 to 1.51) (8). Other characteristics that were associated with increased mortality were congestive heart failure (HR = 1.86; 95% CI 1.62 to 2.14), cancer (HR = 1.14; 95% CI, 1.01 to 1.29), cerebrovascular disease (HR = 1.15; 95% CI, 1.02 to 1.30), peripheral vascular disease (HR = 1.24; 95% CI, 1.09 to 1.41), liver disease (HR = 1.59; 95% CI, 1.20 to 2.10), chronic obstructive pulmonary disease (HR = 1.28; 95% CI, 1.14 to 1.43), and drug or alcohol abuse (HR = 1.50; 95% CI, 1.24 to 1.81). We then built the logistic regression model to estimate the likelihood of late referral (i.e., the propensity score) for each patient, using all covariates listed in Table 1. The c statistic of the propensity score model indicated moderate prediction of late versus early referral (c= 0.683). When introducing quintiles of propensity scores into the otherwise identical Cox proportional hazards model, the effect estimate of LR versus ER on mortality was quite similar; patients referred late were now at a 31% higher risk of death compared with those referred >90 d before onset of chronic dialysis (HR = 1.31; 95% CI, 1.17 to 1.46). All other parameter estimates remained nearly identical.
Table 3: Number of individuals at baseline, censored, and deaths by time interval and referral status
Propensity scores were successfully matched on all 1039 LR patients to one patient each from the ER pool (n = 2078). A Kaplan-Meier plot of patient survival confirmed that patients who actually were referred late had a lower survival in the first year of maintenance dialysis compared with those who were referred to a nephrologist earlier, even in comparison with comparable patients who had the same expected likelihood of late referral based on their demographic characteristics and clinical and healthcare utilization experience (Figure 2). Similarly, the corresponding Cox proportional hazards model revealed a 40% higher mortality rate during the first year among LR compared with ER (HR = 1.40; 95% CI, 1.23 to 1.59). Including the healthcare utilization covariates in the outcomes model as well did not change the results (see Table 4).
Figure 2.:
Kaplan-Meier plot of actuarial survival by timing of referral in propensity score matched population (n = 2078).
Table 4: Effect estimates of the late versus early nephrologist referral association
We next examined the time course of the observed mortality difference. On the basis of the Kaplan-Meier survival curve, the excess mortality rate among LR versus ER patients appeared limited to the first few months of RRT. We tested the validity of this impression by building two interval Cox models: one from onset of dialysis to 90 d of dialysis, the second one evaluating patients who survived 90 d of dialysis until 365 d after onset of maintenance dialysis. Indeed, the detrimental effect of late referral was limited to the first 3 mo, during which patients who were referred to a nephrologist late had a 75% excess mortality compared with those who were referred early (HR = 1.75; 95% CI, 1.48 to 2.08). After day 90 of maintenance dialysis, the death rate between those two groups was not different (HR = 1.03; 95% CI, 0.84 to 1.25).
We speculated whether the superior 90-d survival of patients who were referred early was attributable to modality choice or to vascular access creation among those starting on hemodialysis. A formal test of such a proposition is a significance test of an interaction between timing of referral and first treatment modality (peritoneal dialysis versus hemodialysis) in the full cohort, and between timing of referral and vascular access surgery before first dialysis among those initiating treatment on hemodialysis. Within the propensity-matched cohort, we found that baseline modality choice was not an effect modifier of the association between LR and 90-d mortality. The interaction term was nonsignificant (P = 0.84), and the main effect of LR was only slightly attenuated (HR = 1.66; 95% CI, 1.19 to 2.32). Similarly, we did not find a significant interaction between timing of referral and vascular access surgery before onset of dialysis on early mortality. While peripheral vascular access placement before first dialysis was associated with a survival benefit over the first 90 d of dialysis (HR = 0.66; 95% CI, 0.47 to 0.93), the test for effect modification remained nonsignificant (P = 0.66). Again, the effect estimate for the LR covariate remained materially unchanged. In a sensitivity analysis, we created a covariate that indicated whether a peripheral vascular access was created >14 d before first dialysis, thus making it more likely that the peripheral vascular access was used for the first dialysis. However, the results remained nearly identical to the ones described above.
Discussion
The purpose of this study was to test whether the first-year mortality benefit enjoyed by patients who first saw a nephrologist >90 d before dialysis might have been simply a surrogate for better healthcare access overall, rather than the benefit of subspecialist care. The introduction of propensity scores that included parameters of healthcare utilization did not materially change the estimate of this effect (HR = 1.36 → 1.31). Evaluating a cohort of pairs matched on propensity score also did not change the original finding materially (HR = 1.36 → 1.40). In all cases, late referral remained an important predictor of mortality in the first year of maintenance dialysis, even after such means of bias reduction (see Table 3).
These data indicate that late nephrologist referral is not simply a proxy for overall poor healthcare access/utilization, but that there is probably a specific and substantial contribution by nephrologists during preparation for dialysis that reduces early mortality.
A second finding is that the detrimental association of delayed nephrologist referral with mortality has its effect during the first 90 d of dialysis but not thereafter. This seeming disparity between the current finding and previous work is probably because the proportionality assumption was either not tested or not found to be significant (1–8). Of these studies, only the recent article by Jungers et al. (7) provides a Kaplan-Meier plot depicting survival by length of nephrologist care before RRT. Their plot also indicates that the effect of later referral on survival is most pronounced during the first months of RRT. This pattern of early survival benefit after RRT can be explained by the concept of “depletion of susceptibles” (27), a form of survival bias. That is, those patients who are vulnerable to the effects of suboptimal preparation for RRT die at an excess rate, whereas those who are more robust are more likely to survive. Moreover, after depletion of those “susceptibles” the survival experience of the remaining subjects is indistinguishable from those who saw a nephrologist in a timely fashion. This concept would also suggest that associations between mortality and timing of first nephrologist referral or duration of nephrologist care are driven solely by the excess mortality of late referrals within the first few months of RRT.
We hypothesized that the benefits of early referral might be mediated through more educated decisions regarding modality choice and via timely creation of peripheral vascular access. When testing these hypotheses formally, we found that neither modality choice nor presence of a peripheral vascular access before onset of dialysis modified the association between LR and 90-d survival, indicating that other measures of care are instrumental in providing the survival benefit among ER. The current analysis does not permit us to define these other components of nephrology care that are probably associated with a reduction in medium or long-term mortality. However, on the basis of extrapolation from processes of care in prevalent ESRD patients, it is likely that interventions such as improved BP control, treatment of anemia, improved nutrition, managing derangements in divalent ion metabolism, etc. may contribute to this putative benefit (28).
Limitations of the present study include the use of claims data, which may be compromised by patient and data miscoding, missing data elements, and misclassification. An a priori assumption is made that these limitations occur in a random rather than systematic way, supporting the internal validity of the study. The overall similarity of our findings to those reported by others spanning a large period of time and multiple geographies suggest that these biases are not a major confounder. Similarly, it is likely that there is residual confounding arising from limited characterization of the severity of comorbid conditions. Additional clinical and laboratory information is unavailable that could enhance the robustness of our models and could serve as explanatory variables. For example, it would be of interest to have information on the rate of decline of residual renal function for inclusion into the propensity score model. Using such information, the predictive power of the model would likely be increased, which would lead to further bias reduction. The moderate prediction of the propensity score model (c= 0.683) indicates that several cofactors that contribute to the referral decision remained unobserved, thus leaving the possibility of residual confounding by such unobserved data.
Because this study captured information on patients who were older and/or indigent, further work will be needed to determine their generalizability to younger and less indigent populations or to populations in other geographic regions. Also, the results of this study do not apply to patients with acute renal failure or those patients who are not diagnosed with chronic kidney disease until shortly before the need for RRT.
These findings indicate that early referral to a nephrologist remains important in influencing the course of patients with chronic kidney disease progressing to ESRD, even after adjusting for other patterns of healthcare utilization during the predialysis phase. Furthermore, it seems that the deleterious effects of late referral occur within the first few months of RRT. These findings suggest important research and policy implications. If late patient referral does occur, extensive corrective interventions may have greater benefit if offered early after resuming RRT. Further research is needed to identify those at highest risk within the late referral population, and appropriate models of care delivery for the first months of RRT need to be developed and tested in this vulnerable population.
This study was supported by grants from the Agency for Health Care Research and Quality (HS-09398) and the National Institute on Aging (R03-AG-1839501).
1. Khan IH, Catto GRD, Edward N, Macleod AM: Death during the first 90 days of dialysis: A case-control study. Am J Kidney Dis 25: 276–280, 1995
2. Campbell JD, Ewigman B, Hosokawa M, Van Stone JC: The timing of referral of patients with end-stage renal disease. Dial Transpl 18: 660–685, 1989
3. Eadington DW: Delayed referral for dialysis. Nephrol Dial Transpl 11: 2124–2126, 1996
4. Khan IH, Catto GR, Edward N, MacLeod AM: Chronic renal failure: Factors influencing nephrology referral. QJM 87: 559–564, 1994
5. Ellis PA, Reddy V, Bari N, Cairns HS: Late referral of end-stage renal failure. QJM 91: 727–732, 1998
6. Roubicek C, Brunet P, Huiart L, Thirion X, Leonetti F, Dussol B, Jaber K, Andrieu D, Ramananarivo P, Berland Y: Timing of nephrology referral: Influence on mortality and morbidity. Am J Kidney Dis 36: 35–41, 2000
7. Jungers P, Massy ZA, Nguyen-Khoa T, Choukroun G, Robino C, Fakhouri F, Touam M, Nguyen AT, Grünfeld JP: Longer duration of predialysis nephrological care is associated with improved long-term survival of dialysis patients. Nephrol Dial Transplant 16: 2357–2364, 2001
8. Avorn J, Bohn RL, Levy E, Levin R, Owen W, Winkelmayer WC, Glynn RJ: Nephrologist care and mortality in patients with chronic renal insufficiency. Arch Intern Med 162: 2002–2006, 2002
9. Ratcliffe PJ, Phillips RE, Oliver DO: Late referral for maintenance dialysis. Br Med J 288: 441–443, 1984
10. Muirhead N, Blyndal K: Potential cost savings of planned dialysis start [Abstract]. J Am Soc Nephrol 6: 553, 1995
11. Ifudu O, Dawood M, Homel P, Friedman EA: Excess morbidity in patients starting uremia therapy without prior care by a nephrologist. Am J Kidney Dis 28: 841–845, 1996
12. Sesso R, Belasco AG: Late diagnosis of chronic renal failure and mortality on maintenance dialysis. Nephrol Dial Transplant 11: 2417–2420, 1996
13. Levin A, Lewis M, Mortiboy P, Faber S, Hare I, Porter EC, Mendelssohn DC: Multidisciplinary predialysis programs: Quantification of their impact on patient outcomes in two Canadian settings. Am J Kidney Dis 29: 533–540, 1997
14. Arora P, Obrador GT, Ruthazer R, Kausz AT, Meyer KB, Jenuleson CS, Pereira BJG: Prevalence, predictors, and consequences of late nephrology referral at a tertiary care center. J Am Soc Nephrol 10: 1281–86, 1998
15. Stehman-Breen CO, Sherrard DJ, Gillen D, Caps M: Determinants of type and timing of initial permanent hemodialysis vascular access. Kidney Int 57: 639–645, 2000
16. Goransson LG, Bergrem H: Consequences of late referral of patients with end-stage renal disease. J Int Med 250: 154–159, 2001
17. Astor BC, Eustace JA, Powe NR, Klag MJ, Sadler JH, Fink NE, Coresh J: Timing of nephrologist referral and ateriovenous access use: The CHOICE study. Am J Kidney Dis 38: 494–501, 2001
18. Avorn J, Winkelmayer WC, Bohn RL, Levin R, Glynn RJ, Levy E, Owen W: Delayed nephrologist referral and inadequate vascular access in patients with advanced kidney failure. J Clin Epi 55: 711–716, 2002
19. Winkelmayer WC, Glynn RJ, Levin R, Owen W, Avorn J: Late referral and modality choice in end-stage renal disease. Kidney Int 60: 1547–1554, 2001
20. Winkelmayer WC, Glynn RJ, Levin R, Mittleman MA, Pliskin JS, Avorn J: Late nephrologist referral and access to renal transplantation. Transplantation 73: 1918–1923, 2002
21. Schmidt RJ, Domico JR, Sorkin MI, Hobbs G: Early referral and its impact on emergent first dialyses, health care costs, and outcome. Am J Kidney Dis 32: 278–283, 1998
22. Rosenbaum PR, Rubin DB: The central role of the propensity score in observational studies for causal effects. Biometrika 70: 41–55, 1983
23. D’Agostino RB Jr: Tutorial in biostatistics. Propensity score methods for a bias reduction in the comparison of a treatment to a non-randomized control group. Statist Med 17: 2265–2281, 1998
24. Joffe MM, Rosenbaum PR: Invited commentary: Propensity scores. Am J Epidemiol 150: 327–333, 1999
25. Ash AS, Shwartz M: Evaluating the performance of risk-adjustment methods: Dichotomous outcomes. In: Risk Adjustment for Measuring Health Care Outcomes, 2
nd edition, edited by Iezzoni LI, Chicago, IL, Health Administration Press, 1997
26. Parsons LS: Reducing bias in a propensity matched-pair sample using greedy matching techniques. Available at:
http//www2.sas.com/proceedings/sugi 26/p214–26.pdf. Accessed on 2/19/02
27. Yola M, Lucien A: Evidence of the depletion of susceptibles effect in non-experimental pharmacoepidemiologic research. J Clin Epidemiol 47: 731–737, 1994
28. Obrador GT, Pereira BJ: Early referral to the nephrologist and timely initiation of renal replacement therapy: A paradigm shift in the management of patients with chronic renal failure. Am J Kidney Dis 31: 398–417, 1998
