The Real World Effectiveness of Hematopoietic Transplant Among Elderly Individuals With Multiple Myeloma

Abstract

Background:

Hematopoietic stem cell transplant (HSCT) is the preferred treatment for young patients with multiple myeloma (MM), but for older adults there is limited evidence on its effectiveness from clinical trials.

Methods:

We used the Surveillance, Epidemiology, and End Results (SEER)–Medicare database to identify individuals age 66 years and older with multiple myeloma (MM) who were diagnosed between 2000 and 2007. We used traditional multivariable analysis, propensity score–based analysis, coarsened exact matching, and an instrumental variable analysis to compare survival for individuals who did or did not receive an hematopoietic stem cell transplant. Survival was measured by Cox proportional hazard models. All statistical tests were two-sided.

Results:

Patients with MM receiving an HSCT were more likely to be white, married, younger, and have fewer comorbidities. Results from all analytic techniques consistently showed that HSCT statistically significantly improved survival, with hazard ratios (HRs) ranging from 0.531 to 0.608 (traditional multivariable analysis: HR = 0.582, 95% confidence interval [CI] = 0.49 to 0.69; propensity score analysis: HR = 0.572, 95% CI = 0.46 to 0.72; coarsened exact matching: HR = 0.608, 95% CI = 0.49 to 0.76; instrumental variable analysis: HR = 0.531, 95% CI = 0.36 to 0.78, all P values ≤ .001).

Conclusions:

Overall survival has increased among patients with MM receiving HSCT. This finding was consistent across statistical methods, indicating robustness of our findings.

Hematopoietic stem cell transplant (HSCT) is considered the preferred treatment for eligible patients with multiple myeloma (MM) (1,2). However, for older adults there is limited evidence from clinical trials. Currently, some European clinical guidelines do not recommend that patients over age 65 years receive a transplant; in contrast, US guidelines are less restrictive regarding age (3). Recent studies have found that older adults are increasingly receiving HSCTs (4,5). This study aims to provide evidence on the real-world effectiveness of hematopoietic stem cell transplant for elderly patients.

There is little evidence that documents the real-world effectiveness because of the small numbers of individuals who undergo transplantation. These studies generally show favorable results, but are typically based on small phase 2 studies and experience at individual centers, and use a case-control design (6–14). The findings are further limited because of selection bias (transplant candidates are healthier than rejected candidates) and because standard regression methods do not correct for unmeasured confounders.

Recently developed statistical methods can minimize selection bias. Matching techniques, including propensity scores and coarsened exact matching, directly adjust for confounding variables by matching patients with similar exposure to the treatment of interest. These matching techniques can only address selection bias caused by characteristics that are observable in the data. However, instrumental variable analysis controls for both observable and unobservable characteristics by identifying an exogenous variable, known as an instrument, which is correlated with the treatment but not with the outcome. Researchers then capitalize on variations in the instrument to impute an unbiased estimate of the treatment’s causal effect on the outcome.

The objective of this study was two-fold. First, we assessed survival differences resulting from an HSCT among older individuals with MM. Second, we compared the performance of four analytic approaches. To address these aims, this study used observational claims data to compare the survival of MM patients who had an HSCT to survival among those patients who did not. We used traditional multivariable regression, propensity score matching, coarsened exact matching (CEM), and instrumental variable analysis to assess the effectiveness of transplant and the robustness of our results.

Methods

Data

We used the Surveillance, Epidemiology, and End Results (SEER)–Medicare database for this analysis. SEER is a population-based cancer registry that covers 26% of the population and collects information on tumor characteristics and survival, as well as demographic information. Patients in the SEER are linked to their fee-for-service (FFS) Medicare claims. The Medicare database includes data on patients with Medicare Part A (inpatient) and Part B (outpatient), including billed claims and services (15,16). The SEER-Medicare database has been shown to effectively measure surgery, has been extensively used to measure use of surgical procedures, and has previously been used to measure use of HSCT among patients with acute myeloid leukemia (15,17–20). More information on the SEER-Medicare data can be found in previous publications (15,16,21,22).

Clinical and Demographic Characteristics

The institutional review board approval was waived because SEER-Medicare data is deidentified administrative data with no personal identifiers. We selected individuals with MM diagnosed between October 1, 2000 and December 31, 2007 and a valid recorded date of birth (n = 22 287). We chose October 1, 2000 as the start date because Medicare started reimbursing for HSCT at this time and December 31, 2007 as the end date because it allowed us to follow patients for at least two years after diagnosis to assess survival. Patients were required to be between the ages of 66 and 80 years when diagnosed (n = 10 382). Included patients were required to have both Medicare Parts A and B FFS coverage starting one year prior to diagnosis. Managed care enrollees were excluded because Medicare claims do not record their utilization (n = 6831). We then identified patients as having received an HSCT if they had any of the following codes: International Classification of Disease-9 (ICD-9, 41.00, 41.01, 41.04, 41.07, 41.09) or Healthcare Common Procedure Coding System (HCPCS, 38241). We limited our sample to patients who either had an HSCT or who lived at least six months after diagnosis to ensure that they would have lived long enough to have been offered a transplant (n = 4515).

Outcomes and Explanatory Variables

The primary outcome was all-cause mortality, defined as survival time from diagnosis until death or the end of the study period (December 31, 2010). We included age at diagnosis, comorbidity burden, sex, marital status, urban status, and race as explanatory variables. We calculated comorbidity burden using the Klabunde version of the Charlson Index using both inpatient and outpatient claims for 12 months prior to the date of diagnosis (23,24). The Klabunde adaptation modifies the Charlson Index to predict mortality for patients with cancer rather than the general population.

Statistical Analysis

We compared the characteristics of patients who did and did not undergo an HSCT using chi-squared tests for categorical variables and t tests for continuous variables. All statistical tests were two-sided, and P values of less than .05 were considered statistically significant. We then conducted an adjusted analysis using Cox regression for our four methodological approaches: 1) standard multivariable regression in the unmatched cohorts, 2) propensity scored, matched analysis, 3) coarsened exact matching analysis, and 4) the instrumental variable analysis. Cox model proportionality assumptions were assessed using scaled Schoenfeld residuals. For all analyses, we calculated hazard ratios with 95% confidence intervals.

Propensity scoring analysis controls for selection bias caused by observable characteristics. We calculated the propensity score using a logistic regression with the following variables: sex, race, age, comorbidity score, urban status, area level poverty (25), marital status, SEER reporting region, and year of diagnosis. Individuals were matched using a 1:1 nearest neighbor approach with a 0.1 caliper without replacement (26); this approach eliminated 17 individuals who had an HSCT. We performed several sensitivity analyses, varying the caliper and the number of matches.

Coarsened exact matching (CEM) is a relatively new technique (27,28). CEM is like propensity scoring in that it creates a group that did not receive the treatment of interest that is similar to the group that did receive the treatment (27). CEM matches case patients and control patients based on “coarsened” explanatory variable categories; in our analysis these included age, race, sex, marital status, urban status, comorbidity score, and year diagnosed. CEM then creates the explanatory variables strata. If a stratum does not include at least one case patient and one control patient, it is excluded. For this analysis, we performed a 1:1 match and included only one control patient for every case patient. This procedure caused 12 individuals without control patients in the same stratum to be dropped.

The instrumental variable (IV) analysis can adjust for both observed and unobserved characteristics, but depends on construction of an appropriate exogenous instrument. To be considered an instrument, a variable must be correlated with the treatment but not with the outcome. As has been done in many other health services research studies, we used geographic variation as our instrument (29–34). We created the IV in two steps, similar to the instrument used by Hadley et al. and Wright et al. (29,30). In the first step, we used logistic regression to predict the probability of having an HSCT as a function of age, race, sex, marital status, urban status, area level poverty, and comorbidity score. In the second step, for each year and health service area (HSA), we calculated the difference between the observed and predicted number of transplants. This is a valid instrument for several reasons. First, there is statistically significant variation across areas (F = 233, P = .002) and a strong association between receipt of treatment and location (F = 221, P < .001). Second, location should be independent of patient characteristics. We conducted the IV analysis using the two-step residual inclusion method (35) with hazard rates being estimated in the second stage equation. We used bootstrapping to construct standard errors and confidence intervals for the second-stage results.

Results

A total of 4515 patients with MM were diagnosed between October 1, 2000 and December 31, 2007, and 263 (5.8%) individuals had a transplant (Table 1, Figure 1). Median follow-up time was 32 months, and the mean age diagnosed was 72.7 years (range, 66–79 years). Table 1 provides descriptive characteristics for the patients stratified by receipt of a transplant. There were statistically significant differences between individuals who received an HSCT and those who did not. As shown in Table 1, transplanted individuals were more likely to be male, white and married, younger, from lower poverty areas, and to have been diagnosed earlier during the time period.

Table 1.

Characteristics of observational, propensity score–matched and coarsened exact–matched patients

Characteristics	Observational			Propensity score–matched cohort			Coarsened exact–matched
	Transplant, % (n = 263)	Nontransplant, % (n = 4252)	P	Transplant, % (n = 246)	Nontransplant, % (n = 246)	P	Transplant, % (n = 251)	Nontransplant, % (n = 251)	P
Age at diagnosis, y
66–69	62	24	<.001	60	59	.72	61	61	1.00
70–74	32	37		34	35		33	33
75–80	6	39		6	6		6	6
Female	39	47	.01	40	44	.47	39	39	1.00
Race
White	88	75	<.001	87	88	.73	88	82	.24
Black/other	12	25		13	12		12	17
Marital status
not married/ unknown	23	41	<.001	25	33	.08	23	23	1.00
Married	77	59		75	67		78	78
Poverty, lowest to highest
Quartile 1	30	22	<.001	30	29	.90	31	26	.21
Quartile 2	26	23		27	29		26	22
Quartile 3	22	24		21	18		21	26
Quartile 4	15	25		15	18		15	21
Missing	7	6		7	8		7	6
Urban/rural
Big metro	61	54	.12	61	61	.98	61	54	.15
Metro/urban	31	35		32	31		35	39
Less urban/ rural	8	12		8	8		7	7
Charlson Comorbidity Index
0	63	60	.13	62	69	.30	65	64	.38
1	24	20		24	18		23	23
2+	13	20		14	13		12	12
Year diagnosed
2000–2001	24	21	<.001	22	23	1.00	22	22	1.00
2002–2003	27	28		28	27		28	28
2004–2005	32	27		32	33		33	32
2006–2007	17	24		17	17		17	16

* Chi-squared tests were used for categorical variables and t tests for continuous variables. All statistical tests were two-sided.

Figure 1.

Consort Diagram. SEER = Surveillance, Epidemiology, and End Results.

As shown in Table 1, for both matching methods (propensity scoring and coarsened exact matching), matched cohorts of transplanted subjects and subjects who were not transplanted did not differ statistically with respect to any characteristics.

As shown in Table 2, all analytic techniques yielded consistent results, indicating that an HSCT statistically significantly improves survival (Figure 2). As shown in Table 2, the traditional multivariable analysis found that among transplant recipients, survival improved (hazard ratio [HR] = 0.582, 95% confidence interval [CI] = 0.49 to 0.69). Corresponding improvements estimated by the other methods were similar, with a hazard ratio of 0.572 (95% CI = 0.46 to 0.72) for propensity score analysis, a hazard ratio of 0.608 (95% CI = 0.49 to 0.76) for coarsened exact matching, and a hazard ratio of 0.531 (95% CI = 0.36 to 0.78) for IV analysis (all P ≤ .001). All coefficients and standard errors are reported in Supplementary Table 1 (available online).

Table 2.

Adjusted Cox proportional hazards models comparing survival of autologous transplant vs nontransplant stratified by analytic method*

Estimation method	No.	Adjusted HR (95% CI)	P
Traditional regression	4515	0.582 (0.49 to 0.69)	<.001
Propensity score–matched	492	0.572 (0.46 to 0.72)	<.001
Coarsened exact–matched	504	0.608 (0.49 to 0.76)	<.001
Instrumental variable	4515	0.531 (0.36 to 0.78)	.001

* All statistical tests were two-sided. CI = confidence interval; HR = hazard ratio.

Figure 2.

Kaplan-Meier survival curves between propensity score–matched patients who did and did not receive a hematopoietic stem cell transplant (HSCT). Solid line = No HSCT; dashed Line = HSCT. No. = 492; P < .01. All statistical tests were two-sided.

Discussion

This analysis showed that HSCT improves overall survival for older individuals with MM. Previously, the survival benefit of HSCT had been demonstrated primarily among younger individuals, with survival gains similar to those observed in older patients (36). Furthermore, this study overcame many of the methodological challenges that have hindered previous studies by using sophisticated analytic methods and a nationally representative dataset.

This analysis investigated whether in the real world older individuals also accrue the survival benefit associated with HSCT and addressed selection bias and confounding based on observable characteristics using propensity score analysis and CEM. Furthermore, we minimized bias from unobservable characteristics using IV analysis.

In this study, we found that relatively few older individuals (5.8%) received an HSCT, using SEER-Medicare data. This low rate of transplantation may reflect the limited supply of cancer centers that are able to offer HSCT for older individuals. It is unclear if the supply of centers that offer HSCT in SEER reporting regions is generalizable to nonreporting regions, meaning that the national HSCT rate may be lower or higher than our estimate. Despite this limitation, this study suggested that many individuals may benefit if they were to receive a transplant.

Although it is clinically beneficial, this procedure is costly. Based on expected expenditures, HSCT would likely increase initial costs, but decrease costs in the years following transplantation; that is, the high up-front cost may be made up for by lower costs in the long run. The cost-effectiveness literature has indeed found that HSCT is cost-effective, with a cost-effectiveness ratio far under the conventional $100 000 per quality-adjusted life-year upper threshold for value (6,37–41). In related work, we have found that this procedure is cost-effective in this population (42).

This study paid special attention to addressing methodological challenges affecting previous research. First, because this study used a nationally representative dataset (SEER-Medicare), findings can be generalized to the broader US population with MM. Second, we used a variety of analytic approaches to assess the robustness of our results. Our multivariable regression allowed us to control for observable covariates. However, if observable characteristics vary substantially between treatment groups, the regression may not adequately control for their influence (30). Matching methods, such as propensity scoring and CEM, address this situation by balancing observable characteristics among patients who do and do not receive transplants. Finally, the IV methods address bias potentially introduced by unobservable characteristics that influence both treatment and outcome. For example, healthier patients are more likely to live and might be more likely to receive a risky treatment. In this case, even if the treatment does not affect outcome, the analysis results could indicate a positive association between treatment and survival (43). To develop an effective IV, a researcher must identify an instrument that influences treatment choice, but does not directly influence the outcome. Because the instrument acts as an explanatory variable that breaks the causal link between potential confounders and the outcome, any association between the instrument and the outcome must be caused by the treatment. Because all four of the methods used in this paper produced similar effect estimates, our analysis indicates that the result is robust against potential sources of bias.

We acknowledge a series of limitations. We were unable to observe discrete clinical characteristics in the administrative data, such as patient plasma cell burden, specific laboratory values (eg, hemoglobin, calcium, creatinine, and albumin), integrated prognostic markers (eg, beta2-microglobulin), tumor genetic factors (eg, loss of all or part of chromosome 13), or response to induction therapy. We recognize that these clinical factors influence a patient’s eligibility for receiving an HSCT and acknowledge this as a limitation of the study. However, we were able to overcome these unobserved confounders by using instrumental variables. We conditioned on patients living at least six months or receiving a transplant. Because this restriction biases the results towards the null, the actual impact of transplant on survival may be even greater than what we have estimated. We acknowledge that our instrumental variable results depend on our having successfully created an exogenous instrument. It is possible that there is a relationship between HSA and overall survival; however, this relationship is unlikely. Furthermore, location has been used as an instrument in many studies (29–34). Additionally, the instrument passes all statistical tests. The IV analysis allows researchers to overcome confounding caused by unobserved factors, but it complicates interpretation of the results. Specifically, we cannot identify all factors associated with treatments and outcomes to help guide clinical decision-making. Despite these limitations, the findings suggest that older individuals do benefit from HSCT.

In summary, this study found a statistically significant survival benefit from HSCT for older individuals. These findings were robust across a variety of statistical methods. Given the strength of this result, its consistency across methods, and the fact that many patients over the age of 65 years are diagnosed with MM, clinical trials may be warranted to better understand the effectiveness of HSCT in older adults.

Funding

Research reported in this publication was supported by the National Cancer Institute of the National Institutes of Health under award number T32 CA009429 (GLS). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Conflict of Interest

The authors have no conflict of interests.