NDT Advance Access originally published online on January 31, 2006
Nephrology Dialysis Transplantation 2006 21(6):1652-1662; doi:10.1093/ndt/gfk095
© The Author [2006]. Published by Oxford University Press on behalf of ERA-EDTA. All rights reserved. For Permissions, please email: journals.permissions@oxfordjournals.org
Original Articles: Dialysis and Transplantation
Anaemia and mortality in haemodialysis patients: interaction of propensity score for predicted anaemia and actual haemoglobin levels
Tricia L. Roberts1,
Robert N. Foley1,2,
Eric D. Weinhandl1,
David T. Gilbertson1 and
Allan J. Collins1,2
1 Chronic Disease Research Group, Minneapolis Medical Research Foundation and 2 University of Minnesota, Minneapolis, Minnesota
Correspondence and offprint requests to: Robert N. Foley, MB, Chronic Disease Research Group, 914 South 8th Street, Suite S-253, Minneapolis, MN 55404. Email: RFoley{at}cdrg.org
 |
Abstract
|
|---|
Background. Haemoglobin levels in haemodialysis patients could
represent unknown comorbidities, more severe levels of known
comorbidities, as well as therapeutic choice. Thus, integrating
factors predictive of anaemia with actual haemoglobin levels
might improve prognostic discrimination.
Methods. We retrospectively studied 93 087 patients who started haemodialysis between 1998 and 2000. Clinical and treatment factors from months 4 through 9, derived from Medicare claims, were used to develop propensity scores for anaemia (mean haemoglobin <11 g/dl). Tertiles of propensity scores were interacted with five levels of actual mean haemoglobin to form 15 groups, ranging from low (anaemia) probability with (mean) haemoglobin <10 g/dl to high probability with haemoglobin 
13 g/dl. Cox proportional hazards regression evaluated mortality and first hospitalization among these groups.
Results. The anaemia propensity score improved overall prognostic discrimination. Propensity score adjustment significantly improved prediction of mortality (P<0.0001) after covariate adjustments including haemoglobin. For mortality, the highest and lowest adjusted hazard ratios (AHR) appeared in these groups, respectively: high probability with haemoglobin <10 g/dl (AHR 1.64 [1.54, 1.75], P<0.0001), and low probability with haemoglobin 12 to <13 g/dl (AHR 0.79 [0.74, 0.85], P<0.0001). Higher haemoglobin levels were associated with lower mortality even after propensity score adjustment. Similar patterns resulted for first hospitalization; however, the interaction was significant only for hospitalization (P = 0. 0212).
Conclusions. Integrating factors predictive of anaemia improves overall prognostic discrimination. Propensity score adjustment refines the prognostic association of haemoglobin levels in haemodialysis patients.
Keywords: anaemia; dialysis; hospitalization; mortality; propensity score
|
Introduction
|
|---|
Anaemia is a common feature of end-stage renal failure, and
erythropoietin deficiency is one of several causative factors
[
1]. The observation, however, of an inverse relationship between
haemoglobin levels and erythropoietin doses challenges the hypothesis
that erythropoietin deficiency is the sole cause of anaemia
in haemodialysis patients [
2]. Resistance to the action of erythropoietin
may also be important in the pathogenesis of anaemia in this
population [
3]. Inflammation in particular has been implicated
in recent years [
4–11]. Therefore, it seems implausible
that the presence of anaemia in a given patient is fully explicable
by therapeutic choice.
Many illnesses considered to be ‘non-inflammatory’ are associated with higher than expected levels of inflammatory cytokines. It is plausible that the presence of anaemia in one of two otherwise identical patients, with a given comorbid condition, could reflect the presence of a more severe level of that comorbidity. Similarly, all things being equal, a greater severity of anaemia could indicate the presence of unmeasured comorbidity. We hypothesized, therefore, that an approach integrating factors predictive of anaemia with actual haemoglobin levels might improve prognostic discrimination.
|
Subjects and methods
|
|---|
Objectives
In incident dialysis patients, our objectives were:
- to evaluate the factors associated with haemoglobin levels <11.0 g/dl. In particular, we wished to enumerate, in the form of a propensity score, whether individual patients had low, moderate, or high probabilities of anaemia;
- to determine whether anaemia propensity scores modulate the association between actual haemoglobin levels and outcomes, with mortality and hospitalization as primary and secondary outcomes, respectively;
- to determine whether integrating anaemia propensity scores with actual haemoglobin levels alters prognostic discrimination. Specifically, we studied whether (a) adjusting for propensity score and (b) interacting anaemia propensity scores with actual haemoglobin levels altered prognostic discrimination and
- to determine whether higher actual haemoglobin levels are associated with decreased mortality (and hospitalization) after adjustment for propensity score.
Data sources and patient population
We retrospectively studied Medicare patients who started haemodialysis between 1 January 1998 and 31 December 2000. Months 4 through 9 inclusive of dialysis therapy (the entry period) were used to define patient characteristics; follow-up began after 9 months. Patients were included in the study if they survived the entry period, had Medicare as a primary payer (MPP) throughout the entry period, had valid data for date of birth, sex, race, ethnicity and primary cause of end-stage renal disease, and had at least one erythropoietin claim and one urea reduction ratio claim during the entry period. Patient counts after each exclusion were as follows: all incident patients with MPP at the start of the entry period, n = 142 324; survived the entry period, n = 117 666; with MPP through the end of the entry period, n = 116 596; with valid demographic data, n = 112 025; with at least one erythropoietin claim, n = 96 091; and with at least one urea reduction ratio claim, n = 93 087.
Medicare institutional outpatient claims provided data on haematocrit value, epoetin (EPO) dose and urea reduction ratio. Patient demographics were obtained from the Identification and Medical Evidence portions of the Renal Beneficiary Utilization System of the Centers for Medicare & Medicaid Services. International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes were used to define comorbid conditions during the study entry period, using the codes shown in Appendix 1.
Analysis
Multivariate logistic regression was used to determine factors associated with mean haemoglobin levels <11 g/dl during the entry period. Regression coefficients from this model were then used to calculate a probability of a mean haemoglobin level <11 g/dl for each patient. The overall distribution of probabilities was arbitrarily divided into tertiles. Fifteen groups were formed by combining the three tertiles of anaemia probability and the following five categories of mean haemoglobin: <10, 10 to <11, 11 to <12, 12 to <13 and 13 g/dl. Cox proportional hazards regression evaluated mortality during follow-up, which started after the end of the entry period and ended at the earliest of: transplantation, switch to peritoneal dialysis, loss to follow-up, 31 December 2002, 1 year of total follow-up. The time to the first day of the first hospital admission in the follow-up period was handled similarly, with the following provisions: (1) for patients with a hospitalization that spanned the end of the entry period, follow-up began on the day of discharge; (2) patients with a hospitalization that spanned the entire follow-up period were excluded (n = 17) and (3) follow-up was censored when Medicare was no longer the primary payer.
Mortality and hospitalization models were compared to evaluate improvement in prognostic discrimination by adjustment for propensity score and by the interaction of anaemia propensity scores with mean haemoglobin levels. These models included: (1) mean haemoglobin levels but not anaemia propensity score, (2) mean haemoglobin levels and anaemia propensity score as separate variables and (3) the interaction of mean haemoglobin levels with anaemia propensity score. The likelihood ratio test compared models (1) and (2) to test propensity score, and models (2) and (3) to test the interaction. Also, to illustrate the models’ performance, observed and predicted survival probabilities were compared for models (2) and (3) (Appendix 4). Data splitting was used as a method of internal validation of the models [12], and the results support their predictive ability.
With the exception of EPO dose, the covariates that were included in the logistic regression model to obtain the propensity scores were also included in the Cox proportional hazards regression models; EPO dose was omitted as an adjustment variable because of collinearity with anaemia propensity score (correlation coefficient between propensity score tertiles and EPO dose quartiles = 0.65). Cox regression models that stratified by quartiles of EPO produced risk for mortality and hospitalization similar to those seen without stratification, and are not reported here.
Analyses were performed using SAS for Windows, version 8.2 (SAS Institute, Cary, NC).
|
Results
|
|---|
Patient characteristics
Table 1 displays baseline patient characteristics for the overall
population. In addition, a comparison of those with and those
without anaemia is shown. Using multiple logistic regression,
factors associated (
P<0.05) with haemoglobin <11 g/dl
were higher EPO dose, greater numbers of hospital days and catheter
insertions, urea reduction ratio <65%, younger age, female
sex, black or other race (other than white, black, Asian, or
native American), the absence of diabetes mellitus or absence
of cerebrovascular accident/transient ischaemic attack and the
presence of these comorbid conditions: congestive heart failure,
peripheral vascular disease, malignancy, peptic ulcer disease,
connective tissue disease, dementia and acquired immune deficiency
syndrome.
Figure 1 plots anaemia propensity scores against mean
haemoglobin. The tertiles of propensity score were <0.16,
0.16–0.30 and >0.30. Using simple linear regression
to predict haemoglobin from propensity score in the full dataset
(
n = 93 087), propensity score explained 11.8% of the variation
in observed haemoglobin (
R2 = 0.118).
View larger version (12K):
[in this window]
[in a new window]
|
Fig. 1. Anaemia propensity score (probability of anaemia) vs mean haemoglobin level. Random sample of 2500 patients. Linear regression line shown.
|
|
Overall mortality and first hospitalization rates were 257 and
1282 per 1000 patient-years, respectively. Abbreviated mortality
and hospitalization associations from Cox proportional hazards
regression models are shown in
Table 2, and full results in
Appendices 2 and
3. The association between higher haemoglobin
levels and decreased mortality and hospitalization risk appeared
even after adjustment for propensity score.
Figures 2 and
3 illustrate the interaction of mean haemoglobin level and anaemia
propensity score, when patients with moderate anaemia propensity
scores and mean haemoglobin levels between 11 and 12 g/dl were
chosen as the reference group. The highest and lowest adjusted
hazard ratios (AHR) were seen in these groups: for mortality,
high probability with haemoglobin <10 g/dl (AHR 1.64 [1.54,
1.75],
P<0.0001) and low probability with haemoglobin 12
to <13 g/dl (AHR 0.79 [0.74, 0.85],
P<0.0001); for first
hospitalization, high probability with haemoglobin <10 g/dl
(AHR 1.34 [1.29, 1.39],
P<0.0001) and low probability with
haemoglobin
13 g/dl (AHR 0.89 [0.84, 0.93],
P<0.0001). As
mean haemoglobin decreased, the increase in hospitalization
risk appeared steepest among patients with high probability
of anaemia. However, the relationship between mean haemoglobin
and mortality appeared generally consistent among propensity
score groups.
View this table:
[in this window]
[in a new window]
|
Table 2. Abbreviated results of cox proportional hazards regression models for mortality and first hospitalization
|
|
View larger version (12K):
[in this window]
[in a new window]
|
Fig. 2. AHR for mortality, by mean haemoglobin level and probability of anaemia. Error bars represent 95% CI. Reference group is mean haemoglobin 11 to <12 g/dl and moderate probability. Analysis adjusts for age, sex, race, ethnicity, number of catheter insertions, number of hospital days, urea reduction ratio and 13 comorbid conditions.
|
|
View larger version (11K):
[in this window]
[in a new window]
|
Fig. 3. AHR for first hospitalization, by mean haemoglobin level and probability of anaemia. Error bars represent 95% CI. Reference group is mean haemoglobin 11 to <12 g/dl and moderate probability. Analysis adjusts for age, sex, race, ethnicity, number of catheter insertions, number of hospital days, urea reduction ratio and 13 comorbid conditions.
|
|
The likelihood ratio tests (
Table 3) showed that the interaction
of mean haemoglobin and anaemia propensity score was significant
only for first hospitalization (
P = 0.5366 for mortality;
P = 0.0212 for hospitalization). Therefore, the risk of anaemia
modulated the association between mean haemoglobin levels and
hospitalization, but not mortality. The likelihood ratio test
also showed that adjustment for propensity score significantly
improved prediction of mortality and hospitalization (
P<0.0001
for each) after adjustment for mean haemoglobin and other covariates.
View this table:
[in this window]
[in a new window]
|
Table 3. Likelihood ratio tests of anaemia propensity score and the interaction of mean haemoglobin and anaemia propensity score
|
|
|
Discussion
|
|---|
We found that integrating anaemia propensity scores with actual
haemoglobin levels improved overall prognostic discrimination.
Adjusting for anaemia propensity score significantly improved
prediction of mortality and first hospitalization, while the
interaction of mean haemoglobin level and anaemia propensity
score improved prognostic discrimination only for hospitalization.
Higher observed haemoglobin levels were associated with lower
mortality and hospitalization, even after adjustment for probability
of anaemia. Adjustment for propensity score may refine the prognostic
association of haemoglobin levels. These findings suggest that
such an approach could be used as a marker for unmeasured comorbidity
in observational studies.
Some studies have used indices of EPO dose to haemoglobin level as a marker of therapeutic response [13–16]. Typically, such a ratio is calculated with EPO dose as the numerator and haemoglobin level as the denominator. Such an approach might be appropriate in experimental studies, especially when comparing the response to a single EPO dose, ideally from a fixed starting haemoglobin level. We chose not to use such an approach for several reasons. First, combining two variables, each of which has an infinity of possible values, must necessarily discard informational quanta, or bytes. Also, we felt that such an approach does not mirror clinical practice, which integrates: starting haemoglobin level, an aspiration to reach a target haemoglobin level, sequential dose changes (when achievement does not match aspiration) and previous therapeutic response patterns, at the level of individual patients. For example, a patient maintaining a long-term haemoglobin level of 5 g/dl and an unchanged EPO dose of 1000 units/week has the same EPO-to-haemoglobin ratio as a patient whose haemoglobin falls from 12 g/dl to 8 g/dl on 2000 units/week or rises from 8 g/dl to 12 g/dl on 2000 units/week. Intuitively, these three instances might be described as therapeutic nihilism, therapeutic resistance and therapeutic response, respectively.
To date, there is no evidence from randomized trials that haemoglobin levels have differential effects on mortality in dialysis patients. The range recommended by the National Kidney Foundation Kidney Disease Outcome Quality Initiative has largely relied on observational studies [17–24]. In contrast, several intervention trials have failed to demonstrate improvements in ‘hard’ outcomes when haemoglobin levels are increased from 10–11 g/dl to 13–14 g/dl [25–27]. The first of these studies used death or non-fatal myocardial infarction as primary study outcome. No statistically significant differences were apparent when event rates were compared according to random haemoglobin target assignment, even though haemoglobin levels were extremely well separated. In contrast, a clear, indirect relationship was present between outcome rates and the actual haemoglobin levels achieved [25]. Given that the object of a controlled trial is to make therapeutic intent dependent on chance, this study may be evidence that factors beyond therapeutic intent contribute to the relationship between anaemia and outcomes in haemodialysis patients. In addition, it has been suggested that high doses of EPO are themselves intrinsically toxic, particularly when haemoglobin levels are low [28]. EPO treatments are determined by haemoglobin response in a semi-algorithmic fashion. If this is the case, the hypothesis that EPO is intrinsically toxic becomes difficult to prove in an observational setting.
Initial prognostic comparability is a fundamental issue to address when comparing two factors. Only random assignment can generate groups that naturally tend to balance both known and unknown prognostic factors. That being said, many questions cannot be readily addressed using experimental designs. In observational studies, every effort must be made to maximize adjustment for known prognostic factors and to minimize non-random treatment assignment. Typically, most studies use prognostic indicators as adjustment variables in multivariate regression models. Propensity scores have been suggested to lower the likelihood of non-random treatment assignment [29,30].
Our study has several limitations. It is retrospective. Exposures reflected clinical need and were not collected at predetermined intervals. Comorbidity assessment was based on administrative claims. Confirmation that higher anaemia propensity scores reflect greater inflammatory activity was not possible. However, we feel that our study also has positive features. The patient and event numbers were high, so that relatively precise risk estimates could be generated. While we found that both the severity of anaemia and the propensity for anaemia were associated with worse outcomes, we can in no way claim cause-and-effect relationships. Only randomized controlled trials can test the hypothesis that changing haemoglobin changes outcomes, after factoring in propensity scores. We fully endorse the need to perform definitive trials in haemodialysis patients. Definitive trials in dialysis patients are both difficult and time consuming. It is likely that observational data will continue to inform much of our clinical practice. Novel analytical approaches may help to increase the alignment between association and causation.
|
Appendix 1.
|
|---|
|
Appendix 2.
|
|---|
|
Appendix 3.
|
|---|
|
Appendix 4.
|
|---|
|
Acknowledgments
|
|---|
The authors thank Dana D. Knopic and James Kaufmann, PhD, for
assistance with manuscript preparation and editing, respectively.
Conflict of interest statement. This study was supported by an unrestricted research grant from Amgen Inc., Thousand Oaks, California. Drs Foley and Collins have received consultant fees from Amgen.
|
References
|
|---|
- Adamson JW, Eschbach J, Finch CA. The kidney and erythropoiesis. Am J Med 1968; 44: 725–733[Medline]
- Coladonato JA, Frankenfield DL, Reddan DN et al. Trends in anemia management among US hemodialysis patients. J Am Soc Nephrol 2002; 13: 1288–1295[Abstract/Free Full Text]
- Stivelman JC. Resistance to recombinant human erythropoietin therapy: a real clinical entity? Semin Nephrol 1989; 9: 8–11[Medline]
- Goicoechea M, Martin J, de Sequera P et al. Role of cytokines in the response to erythropoietin in hemodialysis patients. Kidney Int 1998; 54: 1337–1343[CrossRef][Web of Science][Medline]
- Gunnell J, Yeun JY, Depner TA, Kaysen GA. Acute-phase response predicts erythropoietin resistance in hemodialysis and peritoneal dialysis patients. Am J Kidney Dis 1999; 33: 63–72[Web of Science][Medline]
- Horl WH, Jacobs C, Macdougall IC et al. European best practice guidelines 14–16: inadequate response to epoetin. Nephrol Dial Transplant 2000; 15 [Suppl 4]: 43–50
- Barany P. Inflammation, serum C-reactive protein, and erythropoietin resistance. Nephrol Dial Transplant 2001; 16: 224–227[Free Full Text]
- Kato A, Odamaki M, Takita T et al. High blood soluble receptor p80 for tumour necrosis factor-alpha is associated with erythropoietin resistance in haemodialysis patients. Nephrol Dial Transplant 2001; 16: 1838–1844[Abstract/Free Full Text]
- Kim JK, Park BS, Park MJ et al. The predictive parameters of erythropoietin hyporesponsiveness in patients on continuous ambulatory peritoneal dialysis. Korean J Intern Med 2001; 16: 110–117[Medline]
- Nassar GM, Fishbane S, Ayus JC. Occult infection of old nonfunctioning arteriovenous grafts: a novel cause of erythropoietin resistance and chronic inflammation in hemodialysis patients. Kidney Int Suppl 2002; 80: 49–54[Medline]
- Stenvinkel P, Barany P. Anaemia, rHuEPO resistance, and cardiovascular disease in end-stage renal failure; links to inflammation and oxidative stress. Nephrol Dial Transplant 2002; 17 [Suppl 5]: 32–37
- Altman DG, Royston P. What do we mean by validating a prognostic model? Stat Med 2000; 19: 453–473[CrossRef][Web of Science][Medline]
- Kalantar-Zadeh K, McAllister CJ, Lehn RS et al. Effect of malnutrition-inflammation complex syndrome on EPO hyporesponsiveness in maintenance hemodialysis patients. Am J Kidney Dis 2003; 42: 761–773[CrossRef][Web of Science][Medline]
- Keven K, Kutlay S, Nergizoglu G, Erturk S. Randomized, crossover study of the effect of vitamin C on EPO response in hemodialysis patients. Am J Kidney Dis 2003; 41: 1233–1239[CrossRef][Web of Science][Medline]
- Aviles B, Coronel F, Perez-Garcia R et al. Anemia management in haemodialysis. EuCliD database in Spain. Nefrologia 2002; 22: 555–563[Medline]
- Reddan DN, Frankenfield DL, Klassen PS et al. Regional variability in anaemia management and haemoglobin in the US. Nephrol Dial Transplant 2003; 18: 147–152[Abstract/Free Full Text]
- Foley RN, Parfrey PS, Harnett JD et al. The impact of anemia on cardiomyopathy, morbidity, and mortality in end-stage renal disease. Am J Kidney Dis 1996; 28: 53–61[Web of Science][Medline]
- Madore F, Lowrie E, Brugnara C et al. Anemia in hemodialysis patients: Variables affecting this outcome predictor. J Am Soc Nephrol 1997; 8: 1921–1929[Abstract]
- Locatelli F, Pisoni RL, Combe C et al. Anaemia in haemodialysis patients of five European countries: association with morbidity and mortality in the Dialysis Outcomes and Practice Patterns Study (DOPPS). Nephrol Dial Transplant 2004; 19: 121–132[Abstract/Free Full Text]
- Locatelli F, Conte F, Marcelli D. The impact of haematocrit levels and erythropoietin treatment on overall and cardiovascular mortality and morbidity–the experience of the Lombardy Dialysis Registry. Nephrol Dial Transplant 1998; 13: 1642–1644[Free Full Text]
- Ma J, Ebben J, Xia H, Collins A. Hematocrit level and associated mortality in hemodialysis patients. J Am Soc Nephrol 1999; 10: 610–619[Abstract/Free Full Text]
- Xia H, Ebben J, Ma J, Collins A. Hematocrit levels and hospitalization risks in hemodialysis patients. J Am Soc Nephrol 1999; 10: 1309–1316[Abstract/Free Full Text]
- Collins AJ, Li S, St.Peter WL et al. Death, hospitalization, and economic associations among incident hemodialysis patients with hematocrit values of 36 to 39%. J Am Soc Nephrol 2001; 12: 2465–2473[Abstract/Free Full Text]
- Ofsthun N, Labrecque J, Lacson E, Keen M, Lazarus JM. The effects of higher hemoglobin levels on mortality and hospitalization in hemodialysis patients. Kidney Int 2003; 63: 1908–1914[CrossRef][Web of Science][Medline]
- Besarab A, Bolton JK, Browne JK et al. The effects of normal as compared with low hematocrit values in patients with cardiac disease who are receiving hemodialysis and epoetin. N Engl J Med 1998; 339: 584–590[Abstract/Free Full Text]
- Foley RN, Parfrey PS, Morgan J et al. Effect of hemoglobin levels in hemodialysis patients with asymptomatic cardiomyopathy. Kidney Int 2000; 58: 1325–1335[CrossRef][Web of Science][Medline]
- Furuland H, Linde T, Ahlmen J et al. A randomized controlled trial of haemoglobin normalization with epoetin alfa in pre-dialysis and dialysis patients. Nephrol Dial Transplant 2003; 18: 353–361[Abstract/Free Full Text]
- Cotter DJ, Stefanik K, Zhang Y, Thamer M. Improved survival with higher hematocrits: where is the evidence? Semin Dial 2004; 17: 181–183[CrossRef][Medline]
- D'Agostino RB, Jr. Propensity score methods for bias reduction in the comparison of a treatment to a non-randomized control group. Stat Med 1998; 17: 2265–2281[CrossRef][Web of Science][Medline]
- Winkelmayer WC, Glynn RJ, Mittleman MA et al. Comparing mortality of elderly patients on hemodialysis versus peritoneal dialysis: a propensity score approach. J Am Soc Nephrol 2002; 13: 2353–2362[Abstract/Free Full Text]
Received for publication: 20.12.04
Accepted in revised form: 4. 1.06
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
|
|
 
Y.-F. Lin, V.-C. Wu, W.-J. Ko, Y.-S. Chen, Y.-M. Chen, W.-Y. Li, N.-K. Chou, A. Chao, T.-M. Huang, F.-C. Chang, et al.
Residual Urine Output and Postoperative Mortality in Maintenance Hemodialysis Patients
Am. J. Crit. Care.,
September 1, 2009;
18(5):
446 - 455.
[Abstract]
[Full Text]
[PDF]
|
|
|
|
|
|
|
A. Ishani, H. Guo, D. T. Gilbertson, J. Liu, S. Dunning, A. J. Collins, and R. N. Foley
Time to target haemoglobin concentration (11 g/dl) risk of hospitalization and mortality among incident dialysis patients
Nephrol. Dial. Transplant.,
August 1, 2007;
22(8):
2247 - 2255.
[Abstract]
[Full Text]
[PDF]
|
|
|
Top
Abstract
Introduction