Chemotherapy and Survival Benefit in Elderly Patients With Advanced Non–Small-Cell Lung Cancer

Multivariate analysis of treatment selection and survival incorporated sociodemographic characteristics expected to be associated with access or treatment, baseline health status that might impact chemotherapy tolerance, and tumor characteristics that might affect treatment decisions and survival. Sociodemographic characteristics included age, race, sex, marital status, concurrent enrollment in Medicaid or Medicare Savings Programs (Medicaid/MSP) in the year before diagnosis, urban residence, and median household income measured at the census tract level. Tumor characteristics include histology (adenocarcinoma, squamous cell, large cell, and other), grade (well or moderately differentiated v poorly and undifferentiated), and substage (stage IIIb with effusion v stage IV). Baseline comorbidity was measured using the CCI^20,21 a count of 19 condition indicators, measured during the 12 months before diagnosis, and weighted to capture expected contribution to noncancer mortality. Our health status measures incorporated claims-based indicators for services commonly needed by patients with poor functional or PS. These services, including hospitalization, skilled nursing or long-term care facility stays, home oxygen, other home health services, and claims for walkers, wheelchairs and related equipment, were identified in preliminary analyses of the Medicare Current Beneficiary Survey, based on their association with difficulty in ambulation (see Appendix Table A2, online only, for further details.) Service use was measured during the same time period as the CCI. We created a count of the number of indicators present, with variables for none, one, and two or more.

Univariate and bivariate analyses describe rates of first- and second-line chemotherapy, chemotherapy regimens, and characteristics of the sample. We estimated logistic and multinomial logit regression models to examine the association between patient and disease characteristics and receipt of first-line chemotherapy, and specific regimens (single agent, nonplatinum doublet, relative to platinum-doublet therapy) among those who received chemotherapy.²² For both models, we present the odds ratios and 95% CI. We also present the marginal probability, which is the magnitude of effect for each categoric independent variable relative to its reference group, holding all other variables at their sample means.

We estimated the effect of any chemotherapy and first-line regimen (platinum doublet compared with single-agent therapy) on the probability of 1-year survival and survival duration. Survival was calculated from the diagnosis date (for any chemotherapy) or first treatment date (for comparisons across treatment regimens) to date of death or censoring. Death date is reported in Medicare enrollment files capturing death dates through 2005. Patient survival was censored if patients were alive at the end of the study period.

For each treatment choice, we estimated discrete 1-year survival models using logistic regression to compare patients with and without chemotherapy, and a multinomial logit to compare across treatment regimens. We estimated multivariate nonparametric (Cox) survival models to examine the effect of treatment modality on survival. We tested the models for the proportionality assumptions, and added interaction terms between treatment and time to adjust for identified nonproportional hazards in the model examining effects of any chemotherapy.²³ We used inverse propensity score weighing to adjust for selection into the various treatment subgroups based on observable characteristics.²⁴ Propensity score analysis balances the characteristics of the treatment groups, reducing confounding by factors that affect both treatment and survival. All analyses were performed using Stata statistical package (version 9, StataCorp, College Station, TX). The study was approved by the University of Maryland Baltimore institutional review board.

A total of 21,285 patients met the study eligibility criteria. Overall, 30.1% of patients received chemotherapy, with 25.8% initiating chemotherapy within 90 days. Second-line therapy was administered to 37.1% of those with any chemotherapy use. Patients receiving platinum-doublet first-line regimens were more likely to receive second-line therapy. Figures 1A and 1B summarize the regimens observed for first- and second-line chemotherapy. Two thirds of patients with chemotherapy received a platinum-doublet first-line regimen, with 23.3% receiving a single agent. Among patients receiving a platinum-based regimen, 72% received a platinum-taxane regimen, with the remainder receiving platinum with another agent. The overwhelming proportion of platinum regimens involved carboplatin use. Among those receiving second-line therapy, 59.1% received a single agent. Use of chemotherapy increased during the study period, from 20.4% in 1997 to 27.8% in 2002. Appendix Table A3 (online only) describes time to initiation, duration of, and time between therapy lines for each type of regimen.

Elderly patients with advanced NSCLC are predominantly white, male, married, and older than 75 years (Table 1). Half of patients have a CCI value greater than or equal to 1, and more than one third have at least one proxy indicator of poor PS. Our models indicate that patient demographics, underlying health status, and tumor characteristics determine receipt of chemotherapy. Age is a particularly strong determinant. Those in groups older than 80 years are less likely to be treated than patients in the youngest age group (19% and 13% v 34%, respectively; P < .01). Health status was an independent predictor, as patients with one or two or more poor PS proxy indicators were less likely to be treated than patients with no indicators (21% and 13% v 25%, respectively; P < .01). African Americans and persons enrolled in Medicaid/MSP were also less likely to be treated, while married persons and those in higher income areas were more likely to be treated.

Among patients receiving chemotherapy, some of the factors that affect decisions to receive any treatment also influence whether the patient receives a platinum-based doublet (Table 1). Increased age, comorbidity, and African American race were associated with single-agent therapy. For example, patients older than 80 years were more likely to receive a single agent, relative to the group age 66 to 69 years (43% and 58% v 15%, respectively; P < .01). Males and married persons were less likely to receive single agents. The PS indicators did not distinguish selection of platinum-doublet versus single agent chemotherapy among those who received chemotherapy.

There was a clear survival benefit associated with receipt of some chemotherapy (7.1 v 2.5 months; P < .001). Adjusted 1-year survival was 11.6% without (95% CI, 11.1 to 12.0), and 27.0% with chemotherapy (95% CI, 26.4 to 27.6). Among those receiving treatment, median survival was 7.7 months for platinum doublets versus 5.3 months for single agents (Appendix Table A4, online only, provides median survival by chemotherapy receipt and age group.) When we estimated the effects of chemotherapy use on survival duration controlling for patient and disease characteristics, we found that the hazard of death was reduced by approximately one half for treated patients (hazard ratio [HR], 0.558; 95% CI, 0.547 to 0.569; Table 2, Fig 2A). The assumption of proportional hazards was not supported, so the model was re-estimated with a chemotherapy-time interaction. The estimated HR at time t = 1 month after diagnosis was 0.376 (95% CI, 0.365 to 0.388) with the mortality hazard increasing by 0.056 for every 1-month period. Factors associated with increased mortality risk include increasing age, male, high comorbidity, poor PS, and poorly defined tumor grade. African American race, married, and stage IIIb (v stage IV) were associated with reduced hazard.

When we compared survival for platinum-doublets versus single agents, we found that the survival outcomes were superior for those on platinum-doublet chemotherapy (Table 3, Fig 2C; HR, 0.734; 95% CI, 0.705 to 0.765). The assumption of proportional odds was supported. Being married, nonwhite, and stage IIIb was associated with improved survival, while increased comorbidity and male sex were among the characteristics associated with an increased mortality risk. The adjusted 1-year survival probabilities were 30.1% for platinum doublets (95% CI, 28.9 to 31.4) versus 19.4% for patients receiving single-agent therapy (95% CI, 18.3 to 20.4).

Most Medicare beneficiaries have some source of supplemental insurance coverage beyond Medicare Parts A and B to help pay for cost sharing for Medicare covered and selected uncovered services. Supplemental medical and drug coverage may be provided through a number of sources. The type of insurance will depend on a variety of factors, for example, whether the beneficiary or spouse had employment related insurance before retirement, or whether the beneficiary met income eligibility requirements for publicly subsidized programs.

Estimates from 2002 (the period reflected in the data in this study) reveal that 35% of Medicare beneficiaries had supplemental medical insurance provided by an employer, and that most of those had prescription drug coverage; 17% were enrolled in Medicaid or Medicare Savings Programs and 14% had prescription drug coverage through that source; 15% were enrolled in a Medicare Advantage (MA) plan, and 12% had drug coverage through that source; and 21% had Medigap or other privately purchased plans, and 12% had prescription drug coverage through that mechanism. Overall, 12% had no supplemental medical coverage during the full year, and 18% had no prescription drug coverage all year. An additional 27% had only part-year prescription drug coverage (http://www.kff.org/medicare/upload/Medicare-Chart-Book-3rd-Edition-Summer-2005-Section-3.pdf & http://www.kff.org/medicare/upload/Medicare-Chart-Book-3rd-Edition-Summer-2005-Section-5.pdf). It is important to note that the sources of coverage for prescription drugs have changed substantially with implementation of Medicare Part D in 2006.

The generosity of medical coverage and prescription drug coverage will vary depending on the source. In Table A1, we identify the types of supplemental plans or programs, eligibility requirements, and whether they are measured in the SEER-Medicare data.

There are several implications associated with data limitations in SEER-Medicare. Medicare claims data for beneficiaries enrolled in MA plans are incomplete or missing entirely. Because MA plans are risk bearing, they do not submit detailed claims to the Medicare program. Hence, many SEER-Medicare analyses exclude MA enrollees. The need to exclude MA enrollees may be problematic, to the extent that MA enrollees have different underlying health status, or that MA plan providers may recommend different treatment regimens. Hence, it may not be possible to generalize from estimates generated from the non-MA enrolled patient population to the full Medicare population.

There is a strong association between type of supplemental insurance, socioeconomic status, and health status. In addition, there is often an association between race and socioeconomic status. Analyses that attempt to disentangle the role of socioeconomic status or race on access to care will be limited by the lack of information on supplemental insurance.

As noted in Table A1, SEER-Medicare provides information on state buy-in of Part B premiums, which indicates either Medicaid or Medicare Savings Programs enrollment. Enrollment in one of these programs requires that the beneficiary have income under 120% to 135% of the federal poverty level. However, not all Medicare beneficiaries who are poor are enrolled in one of these programs. Hence, the state buy-in indicator is not a sensitive indicator of poverty.

These limitations are common to Medicare administrative data sets, as the Medicare program does not routinely collect information on supplemental coverage in a centralized location, other than Medicaid. Other sources of data that include information on supplemental insurance held by Medicare beneficiaries may provide important opportunities for data triangulation.

Our objective was to identify measures of service use from insurance claims that would be correlated with functional status limitations in mobility and self care, the key elements of the Eastern Cooperative Oncology Group performance status (PS) scale. We used data from the Medicare Current Beneficiary Survey (MCBS), a panel survey that is nationally representative of community based and institutionalized elderly and disabled Medicare beneficiaries (Adler GS: Health Care Financ Rev 15:153-163, 1994). The MCBS uses a rotating panel design, sampling approximately 5,100 new beneficiaries each year, with the cohort remaining in the survey for up to 4 years, subject to death or loss to follow-up. At any given time the survey includes approximately 15,000 beneficiaries. We selected a single year (2001) for convenience. The survey captures information on demographics, income, insurance, self-reported health behaviors, health status, and self-reported information on functional status, including limitations and dependence in activities of daily living (ADLs) and instrumental activities of daily living (IADLs), and difficulty with activities requiring strength, stamina, or agility (the physical performance measures), measured in the fall of the year. When combining data on community dwelling and institutionalized Medicare beneficiaries, we found that 36% had at least one ADL limitation, 45% reported an IADL limitation, and 33% reported a lot of difficulty or inability to walk one quarter of 1 mile. The MCBS is linked to detailed Medicare Part A and B claims for participants in each survey year, including claims from hospital inpatient, hospital outpatient, physician, laboratory, durable medical equipment, skilled nursing facility, home health agency (HHA), and hospice. The claims provide detailed information concerning date, location, provider, type of service, and associated diagnoses. Specific procedures are reported using International Classification of Diseases 9th revision procedure codes for hospital-based services. Services provided by all other provider groups are coded using Healthcare Common Procedure Coding System, a variant of the Current Procedural Terminology codes.

In our analysis, we employed a single indicator—have a lot of difficulty or unable to walk a quarter of 1 mile—as a proxy measure for poor PS. MCBS participants were stratified according to their responses to this question, grouping those with little to no difficulty together, and grouping those with a lot of difficulty or unable to perform the task. Next, we identified a set of service use categories that we expected to be associated with difficulty with ambulation. This initial set of indicators of poor PS included walking aids (eg, canes, crutches, walkers), wheelchairs and related supplies, commodes, home oxygen and related respiratory therapy supplies, hospital admission, hospital bed, admission to a skilled nursing facility, and any HHA claim. The Healthcare Common Procedure Coding System codes for these services were identified and the indicators for each service category were created using the MCBS claims data. The prevalence for each indicator was estimated for the MCBS, and compared for patients with a lot of difficulty walking compared to those with little to no difficulty walking a quarter of 1 mile. Results are presented in Appendix Table A2.

As indicated in Appendix Table A2, all the poor PS indicators were more common in MCBS patients with difficulty walking compared with those without difficulty. Of those with difficulty, 13.0% had HHA services, 10.4% used oxygen or respiratory therapy supplies, 9.7% used a wheelchair, 6.4% purchased walking assistive devices, compared to 2.7%, 2.0%, 0.9%, and 1.4%, respectively, for those little or no difficulty walking. Differences across the groups were significant for all of the PS measures we tested (P < .001).

Based on the encouraging results related to these indicators, we selected the most prevalent for inclusion in our study of treatment and survival benefit for advanced NSCLC. We subsequently created an indicator for long-term care facility stay based on site of care reported for provider services.

In ongoing research (Grant No. NIH/NCI R21 CA137283-01A1; A.J.D.), we are developing a multivariate model that will predict PS as a function of a broader set of claims based indicators—physician consults, diagnostic tests, durable medical equipment, other services used—with selected interactions with diagnostic indicators. Similar to the Charlson Comorbidity Index commonly used in SEER-Medicare, we expect to produce prediction weights that can be used to generate a predicted PS value for inclusion in analyses of cancer treatment and outcomes.