*Medical Technology and Practice Patterns Institute (MTPPI), Bethesda, MD
†VA NY Harbor Healthcare System, New York, NY
Departments of ‡Epidemiology
§Biostatistics, Harvard School of Public Health
∥Harvard-MIT Division of Health Sciences and Technology, Boston, MA
Supplemental Digital Content is available for this article. Direct URL citations appear in the printed text and are provided in the HTML and PDF versions of this article on the journal's Website, www.lww-medicalcare.com.
Supported in part by an AHRQ Grant R21 HS19513 and a NIH Grant R01 HL080644.
Study findings presented at AHRQ 2012 Annual Conference on September 11, 2012 at Rockville, Maryland.
Disclaimer: The data were provided by the US Renal Data System but the interpretation and reporting of these data are the responsibility of the authors and in no way should be seen as the official policy or interpretation.
The authors declare no conflict of interest.
Reprints: Dennis Cotter, MSE, Medical Technology and Practice Patterns Institute (MTPPI), 5272 River Road, Suite 500, Bethesda, MD 20816. E-mail: firstname.lastname@example.org.
Randomized trials found that use of erythropoiesis-stimulating agents to target normal hematocrit (Hct) levels (>39%) compared with 27%–34.5% increases cardiovascular risk and mortality among chronic kidney disease patients. However, the effects of the most widely used Hct target in the past 2 decades, 34.5%–39%, have never been examined.
To compare the effects of 2 Hct target strategies—30.0%–34.5% (low) and 34.5%–39.0% (mid) in a high-risk population: elderly dialysis patients with significant comorbidities.
Observational data from the US Renal Data System were used to emulate a randomized trial in which patients were assigned to either Hct strategy. Follow-up started after completing 3 months of hemodialysis and ended 6 months later. We conducted the observational analogs of intention-to-treat and per-protocol analyses. Inverse-probability weighting was used to adjust for measured time-dependent confounding by indication.
A total of 22,474 elderly patients with both diabetes and cardiovascular disease who initiated hemodialysis in 2006–2008.
Hazard ratios (HRs) and survival probabilities for all-cause mortality and a composite cardiovascular and mortality endpoint.
The intention-to-treat HR (95% confidence interval) for mid versus low Hct strategy was 1.05 (0.99–1.11) for all-cause mortality and 1.03 (0.98–1.08) for the composite endpoint. The per-protocol HR (95% confidence interval) for mid versus low Hct strategy was 0.98 (0.78–1.24) for all-cause mortality and 1.00 (0.81–1.24) for the composite outcome.
Among hemodialysis patients, we did not find differences in 6-month survival or cardiovascular risk between clinical strategies that target Hct at 30.0%–34.5% versus 34.5%–39.0%.
Anemia affects nearly all end-stage renal disease (ESRD) patients and is associated with diminished quality of life, decreased survival, and adverse cardiovascular outcomes.1–3 Dialysis patients receive erythropoiesis-stimulating agents to elevate their hematocrit (Hct) levels. The optimal Hct target is unknown. Four randomized trials in patients with chronic kidney disease have shown that patients targeted to near-normal Hct values [39% for Trial to Reduce Cardiovascular Events with Aranesp Therapy (TREAT),4 39%–45% for Cardiovascular Risk Reduction in Early Anemia Trial with Epoetin-Beta (CREATE),5 40.5% for Correction of Hemoglobin and Outcomes in Renal Insufficiency (CHOIR),6 and 42% for the Normal Hematocrit Trial (NHT)7] had worse clinical outcomes than patients targeted to low Hct values (27%–34.5%)5–7 or placebo.4 For example, in the NHT study, the only trial that included dialysis (vs. predialysis) patients,7 the high Hct arm experienced a 27% increase in mortality compared with the control arm.
Safety concerns and lack of proven benefits for targeting patients to normal Hct levels prompted the Food and Drug Administration (FDA) to recommend that physicians “reduce or interrupt the dose” of epoetin if Hct exceeds 33%. However, until 2011, the majority of ESRD patients in the United States were targeted to a middle Hct range of 34%–39%.8 No randomized trials have compared the FDA-recommended strategy of Hct <33% with the commonly used strategy of Hct 34%–39%, which requires higher epoetin doses. In addition to higher mortality and cardiovascular risks of targeting normal Hct, the cost implications of using higher doses are significant; epoetin is the single largest Medicare drug expenditure: ∼$2 billion annually between 2005 and 2010 and ∼11% of all ESRD costs.9
In this research, we use observational data to compare the effects of a low Hct strategy of 30%–34.5%, similar to the FDA-recommended strategy, with a commonly used mid range Hct strategy of 34.5%–39%. Like the NHT and TREAT trials, we focused on elderly ESRD dialysis patients at high risk for adverse cardiovascular outcomes.
We used observational data from the United States Renal Data System (USRDS) to emulate a randomized clinical trial10 among elderly hemodialysis patients with both diabetes and cardiovascular disease. The USRDS includes 93% of US dialysis patients with Medicare coverage.11 Most claims cover a service period of approximately 1 month (average duration is 24 days).12 We used the USRDS standard analytic files for 2006–2009 that contained variables from the patient, medical evidence, and facility data files. Figure 1 represents the patient selection process.
We considered 2 dynamic treatment strategies for erythropoiesis-stimulating agent use.
- Mid Hct strategy: intravenous epoetin-α to achieve and maintain Hct values between 34.5% and 39.0%; or
- Low Hct strategy: intravenous epoetin-α to achieve and maintain Hct values between 30.0% and 34.5%.
Under both strategies, epoetin dose is (i) increased by at least 10% if previous Hct is below the target range; (ii) decreased by no more than 10% times (previous Hct minus lower end of range) or increased by no more than 10% times (upper end of range minus Hct) if previous Hct is within target range; (iii) decreased by at least 25% (or withheld) if previous Hct is above the target range. For simplicity, we did not consider strategies that vary according to the evolving clinical characteristics of the patients.
Similar to previous randomized controlled trials, the 2 endpoints of interest were all-cause mortality and a composite outcome including death and hospitalization for myocardial infarction, stroke, or congestive heart failure.5,6 Previous studies have verified that ICD-9 codes used to define myocardial infarction, congestive heart failure, and stroke have specificity higher than 90% and sensitivity between 67% and 86%.13–16
Table 1 summarizes the characteristics of the hypothetical randomized trial and how we emulated it using the observational data. Supplemental Digital Content 1A and 1B, http://links.lww.com/MLR/A560, http://links.lww.com/MLR/A561 describe the monthly assignment to each strategy.
To conduct the observational analog of an intention-to-treat, we fit separate pooled logistic models to estimate probability of death and of the composite outcome at each month, conditional on an indicator for treatment strategy (Mid or Low Hct), baseline covariates (see Table 1), and month of follow-up (cubic splines). This model was fit to the expanded dataset that resulted from duplicating the patients whose data were consistent with both strategies at baseline. To adjust for potential selection bias due to censoring by loss to follow-up, we estimated stabilized inverse-probability (IP) weights as previously described.17,18 The weighted outcome models estimate the observational analog of the intention-to-treat average hazard ratio (HR) for the Mid Hct versus Low Hct strategies.
We also conducted the observational analog of a “per-protocol” analysis in which we estimated the effect estimates if all subjects had adhered to their baseline strategy throughout the entire follow-up. To estimate the per-protocol HR of the outcome for Mid Hct versus Low Hct strategy, we fit the above models to the expanded dataset after censoring patients when they deviated from their original strategy. We used stabilized IP weights to adjust for time-dependent selection bias due to this censoring.19 The denominator of the weights was estimated by fitting 2 nested models to the original (nonexpanded) dataset: a logistic regression model to estimate each patient’s probability of not receiving epoetin (8% of the patient-months had zero dose) in nonhospitalized person-months, and a linear regression model to estimate each patient’s density (assumed to be normal) of log epoetin dose among those with nonzero dose at that (and the previous) month. Both models included the covariates listed above for the mortality model plus the time-varying covariates (see Table 1). The numerator of the weights was estimated by fitting to the expanded dataset a logistic regression model for the probability of not deviating from their original strategy conditional on baseline variables and an indicator for strategy. The estimated weights had 99th and 95th percentiles of 90 and 5 for the mortality endpoint, and 75 and 5 for the composite endpoint, respectively. To mitigate the impact of outliers, we truncated the weights to a maximum value of 20. The truncated weights had a mean weight of 1.1 (SD 2.9) for both mortality and composite endpoint analyses. When we truncated the weights at 40, 50, and 100 in multiple sensitivity analyses, the estimates did not materially change.
We also estimated the survival curves under each treatment strategy by using the predicted values of weighted outcome models that additionally included the product “interaction” terms between the treatment strategy indicator and the month variables. Point-wise 95% confidence intervals for all parameter estimates were calculated by nonparametric bootstrap based on 300 full samples.
Of 22,474 eligible patients (Fig. 1), 5395 were classified as following the Mid Hct strategy only, 7439 as following the Low Hct strategy only, and 9640 as following both the strategies. Compared with patients only in the Low Hct strategy, patients in the Mid Hct strategy had higher predialysis and baseline Hct, lower doses of epoetin and iron before baseline, a higher Charlson score, fewer inpatient days before baseline, and were more likely to receive dialysis services from Fresenius (the nation’s largest dialysis chain) (Table 2). Patients following both strategies had shorter inpatient stays before baseline and had the highest average Hct value in the first 3 months of dialysis.
In the expanded dataset with duplicates, there were 2738 deaths (3825 composite events) under the Mid Hct strategy and 2292 deaths (3281 composite events) under the Low Hct strategy. The HRs (95% confidence interval) were 1.05 (0.99–1.11) for death and 1.03 (0.98–1.08) for the composite event in the intention-to-treat analysis, and 0.98 (0.78–1.24) and 1.00 (0.81–1.24), respectively, in the per-protocol analysis (Table 3). In analyses restricted to the 87% of patients with Hct>30% at baseline (considered epoetin responsive patients), the HRs were similar. Unadjusted and partially adjusted estimates were also similar (Supplemental Digital Content 2, http://links.lww.com/MLR/A562). The 6-month risk difference between the Mid and Low Hct strategies was 0.0% in all analyses (Fig. 2).
Our estimates did not change when we restricted our analysis to patients with serum albumin level <3.5 g/dL, who might be expected to have worse outcomes, when IP weights were estimated under a gamma or truncated normal distribution for the log of epoetin dosage, when we used different knot locations for splines of log epoetin dosage and Hct values, when we used cubic splines of iron dose, and when we applied different dosing algorithms to define Hct target strategies (ie, 10%, 15%, 20%, and 25% dose titration rates). To test the robustness of our estimates to longer follow-up, we followed patients through the end of their first year on dialysis (ie, treatment strategies and patient outcomes were evaluated during the 9-mo period after baseline). Intention-to-treat and per-protocol analysis resulted in essentially null estimates, although there was substantial attrition due to artificial censoring in the latter (data not shown).
Previous randomized trials of epoetin therapy—3 in predialysis chronic kidney disease patients4–6 and 1 in dialysis patients7—found increased mortality or no benefits for near-normal (>39%), Hct targets compared with targets <34.5% (Appendix Table). After publication of the TREAT trial, the FDA changed the epoetin label and advised reducing or interrupting the dose of epoetin if the Hct exceeded 33%. Using observational data, we emulated a randomized trial in a high-risk population of elderly dialysis patients with both diabetes and cardiovascular disease assigned to either a Mid (34.5%–39%) or a Low (30%–34.5%) Hct target strategy to be achieved through epoetin therapy. We found no differences in the 6-month mortality and cardiovascular risk between these 2 clinical strategies, which supports the current FDA recommendations for a target Hct level up to 33% in hemodialysis patients.
It has been suggested20 that Hct targets higher than those recommended by the FDA might help some patients (eg, those who respond to epoetin and achieve a high Hct) and harm others (eg, hyporesponsive patients). We found no indication of benefit for a higher Hct target after removing poor responders from our analyses, which supports FDA’s target Hct recommendation for all patients.
Until January 2011, nearly all nephrologists and dialysis providers targeted a Hct level >33%. In 2011, the enhanced ESRD Prospective Payment System bundled epoetin therapy into the dialysis composite rate.21 Preliminary indications suggest that both epoetin use22 and Hct levels have been dramatically reduced after PPS implementation resulting in Hct targets that appear to be more consistent with current FDA recommendations.23,24
We used the largest observational database available for ESRD research. Our study, however, has several limitations. First, the validity of our estimates depends on the assumption that all confounding factors were correctly included in the model.25,26 We used statistical methods that appropriately adjust for measured time-varying confounders, but residual confounding could still remain. Second, claims data are collected primarily for billing purposes, and their quality for research might be suboptimal. However, the information on health outcomes used in our study has been previously validated.27,28 Finally, our approach made it hard to find differences in the intention-to-treat analysis (because most patients were assigned to both strategies), and it was difficult to study a longer follow-up period in the per-protocol analysis because many patients were artificially censored. Additional analysis based on the parametric g-formula might help alleviate these problems.
Our results support FDA’s most recent advisories recommending a Hct target of <33% when treating hemodialysis patients, including those with serious comorbidities. The statistical methods employed here can be applied to USRDS and other observational databases to examine different Hct strategies in specific populations and provide preliminary data for planning randomized clinical trials.
1. Collins AJ, Li S, St Peter W, 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.
2. Lowrie EG, Huang WH, Lew NL, et al. Friedman EA .The relative contribution of measured variables to death risk among hemodialysis patients.Death on Hemodialysis. 1994 .Dordrecht:Kluwer Academic Publishers; 121–141.
3. Ma JZ, Ebben J, Xia H, et al .Hematocrit level and associated mortality in hemodialysis patients.J Am Soc Nephrol. 1999; 10:610–619.
4. Pfeffer MA, Burdmann EA, Chen CY, et al .A trial of darbepoetin-alfa in type 2 diabetes and chronic kidney disease.N Engl J Med. 2009; 361:1–14.
5. Drüeke TB, Locatelli F, Clyne N, et al .Normalization of hemoglobin level in patients with chronic kidney disease and anemia.N Engl J Med. 2006; 355:2071–2084.
6. Singh AK, Szczech L, Tang KL, et al .Correction of anemia with epoetinalfa in chronic kidney disease.N Engl J Med. 2006; 355:2085–2098.
7. Besarab A, Bolton WK, 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.
8. . US Renal Data System .http://www.usrds.org/adr.htm
USRDS 2009 Annual Data Report: Atlas of Chronic Kidney Disease and End-Stage Renal Disease in the United States. National Institutes of Health, National Institute of Diabetes and Digestive and Kidney Disease, Bethesda, MD. Available at: http://www.usrds.org/adr.htm. Accessed April 21, 2012
10. Danaei G, Rodríguez LA, Cantero OF, et al .Observational data for comparative effectiveness research: an emulation of randomized trials of statins and primary prevention of coronary heart disease.Stat Methods Med Res. 2013; 22:70–96.
11. Foley RN, Collins AJ .The USRDS: what you need to know about what it can and can’t tell us about ESRD.Clin J Am Soc Nephrol. 2013; 8:845–851.
13. Goff DC, Pandey DK, Chan FA, et al .Congestive heart failure in the United States: is there more than meets the I(CD code): The Corpus Christi Heart Project.Arch Intern Med. 2000; 160:197–202.
14. Kiyota Y, Schneeweiss S, Glynn RJ, et al .Accuracy of Medicare claims-based diagnosis of acute myocardial infarction: estimating positive predictive value on the basis of review of hospital records.Am Heart J. 2004; 148:99–104.
15. Tirschwell DL, Longstreth WT .Validating administrative data in stroke research.Stroke. 2002; 33:2465–2470.
16. Birman-Deych E, Waterman AD, Yan Y, et al .Accuracy of ICD-9-CM codes for identifying cardiovascular and stroke risk factors.Med Care. 2005; 43:480–485.
17. Zhang Y, Thamer M, Cotter D, et al .Estimated effect of epoetin dosage on survival among elderly hemodialysis patients in the United States.Clin J Am Soc Nephrol. 2009; 4:638–644.
18. Zhang Y, Thamer M, Kaufman JS, et al .High doses of epoetin do not lower mortality and cardiovascular risk among elderly hemodialysis patients with diabetes.Kidney Int. 2011; 80:663–669.
19. Cain LE, Robins JM, Lanoy E, et al .When to start treatment? A systematic approach to the comparison of dynamic regimes using observational data.Int J Biostat. 2010; 6:
20. Solomon SD, Uno H, Lewis EF, et al .Erythropoietic response and outcomes in kidney disease and type 2 diabetes.N Engl J Med. 2010; 363:1146–1155.
22. Hirth RA, Turenne MN, Wheeler JR, et al .The initial impact of Medicare’s new prospective payment system for kidney dialysis.Am J Kidney Dis. 2013;
[Epub ahead of print]
25. Cook NR, Cole SR, Hennekens CH .Use of a marginal structural model to determine the effect of aspirin on cardiovascular mortality in the Physicians’ Health Study.Am J Epidemiol. 2002; 155:1045–1053.
26. Hernán MA, Hernández-Díaz S, Werler MM, et al .Causal knowledge as a prerequisite for confounding evaluation: an application to birth defects epidemiology.Am J Epidemiol. 2002; 155:176–184.
27. Completeness and reliability of USRDS data: comparisons with the Michigan Kidney Registry.Am J Kidney Dis. 1992; 20:84–88.
28. . USRDS .How good are the data? USRDS data validation special study.Am J Kidney Dis. 1992; 20:5 Suppl 2 68–83.