The Relationship between Health Insurance and Mortality for Cancer Patients: Medicare Advantage versus Fee-For-Service Medicare

Compared to traditional fee-for-service Medicare (FFS), private Medicare Advantage (MA) plans offer additional health insurance coverage but restrict access to medical providers. This study measured how MA enrollment, relative to FFS enrollment, may influence mortality for cancer patients. The study used linked data from the Surveillance, Epidemiology, and End Results Program and Medicare administration (SEER-Medicare) including diagnoses between 2006 and 2011 at all four major cancer sites (breast, colorectal, lung, prostate). The key innovation of the study was to measure and account for variation in prescription drug coverage between MA and FFS cancer patients. Among cancer patients with Part D coverage, MA enrollment was associated with modestly increased mortality. The estimated relationships were statistically distinguishable from zero for lung cancer and (in most model specifications) colorectal cancer. The findings are consistent with a hypothesis that restricted provider access may reduce health outcomes for patients who already have a serious illness.

excluded from the statistical analyses. The final sample for the statistical analysis thus included 181,256 cancer patients, 50.7 percent of whom were covered by FFS at diagnosis.

Mortality Measures
The outcome of interest was mortality. Cancer diagnoses in the sample ranged from 2006 through 2011 and mortality (taken from the Medicare administrative data) follow-up lasted through 2013, so each patient had at least two-years of follow-up and some had up to 8 years.
The study used multiple measures of mortality, each with its own advantages and disadvantages. The simplest mortality measure was an uncensored indicator for death by any cause within two years of cancer diagnosis. Since MA patients tend to be healthier than FFS patients, even conditional on patient characteristics and chronic conditions (see Newhouse, Price, McWilliams, Hsu, & Mcguire, 2015; for an excellent summary of the Medicare selection literature), the statistical model would ideally control for patients' non-cancer comorbidities. Unfortunately, however, this information is not available. The SEER registries do not collect information on comorbidities and the Medicare claims data only includes FFS patients. Because differential non-cancer health is not addressed, estimates from this measure should be considered a lower bound.
In the absence of comorbidity measures, information about cause of death was used to (at least partially) control for selection related to non-cancer health. The SEER registries use algorithms that process cause of death from death certificate data (Note 1). If a cancer patient dies from something unrelated to their cancer, an all-causes mortality measure considers that the same as it does a death caused by cancer.
Cancer-caused mortality measures, on the other hand, consider the two causes of death to be different.
The second mortality measure used in the study was an indicator for death caused by cancer within two years of cancer diagnosis. This measure implicitly, and likely inaccurately, assumes that all patients who died from something unrelated to cancer would not have died from cancer within two years of cancer diagnosis. Because of this, estimates from this measure should be considered an upper bound.
The two other mortality measures used were similar to the two discussed above, except placed in a framework of a hazard model in order to address censoring. They are discussed in the Cox Proportional Hazards: Methodology and Results section.

Key Independent Variables
The independent variable of interest was MA enrollment during the year of cancer diagnosis. While patients can switch between MA and FFS over time, switching is rare. Over the same time frame as this study, Lissenden (2018) found that a cancer diagnosis induced more switches to FFS and less switches to MA in the year after cancer diagnosis, but no detectable change in switching behavior in the year of cancer diagnosis. This is because switching was generally not allowed within a calendar year in these years. Thus, enrollment in the year of cancer diagnosis reflects a decision the patient made prior to being diagnosed with cancer. Sensitivity models that defined MA enrollment from the year prior to cancer diagnosis or excluded all switchers (at any point before or after cancer diagnosis) produced results that were similar to the preferred models.
A key confounder related to both MA enrollment and cancer mortality is prescription drug coverage.
Prescription drug coverage decreases mortality for Medicare (cancer and other) patients (Gowrisankaran, Town, & Barrette, 2011). Prescription drug coverage also varies significantly between MA and FFS patients; the vast majority of MA patients have Medicare Part D included in their benefits but many FFS patients have alternative or no prescription drug coverage. Relative to previous studies measuring the relationship between MA enrollment and cancer mortality, the key advantage of this study is that prescription drug coverage is at least partially observed. In particular, it is observed whether or not each patient had Medicare Part D coverage when they were diagnosed with cancer. Over 91% of the MA cancer patients in the study had Medicare Part D, but fewer than 45% of FFS cancer patients did. Many of the patients without Medicare Part D coverage may have had alternative prescription drug coverage, but this is unobservable in the data. In other words, conditional on having Part D coverage, MA coverage was associated with a 2.9 percentage point (9 percent) higher chance of dying within two years after a lung cancer diagnosis.

Linear Regression: Methodology and Results
The goal of this section is to measure how MA enrollment relates to mortality conditional on observable patient characteristics that may influence mortality. In particular, a linear regression was used with control variables for health service area (HSA), age band, sex, race, ethnicity, marital status, measures of socioeconomic status, year of diagnosis, and measures of cancer site and severity.
Non-cancer comorbidities were unobserved, but as discussed in the previous section, the use of both all-causes (lower bound) and cancer-caused mortality (upper bound) outcomes help to understand the implications of any resulting bias.
There were two socioeconomic variables measured based on the census tract in which the patient lived at the time of their diagnosis; the percent of residents without a high school degree and the percent of residents with a college degree. The other control variables were all categorical. The HSAs were defined using county of residence and the mapping from the National Cancer Institute, which is meant to represent service areas for cancer treatment. There were several SEER variables used to measure cancer severity: summary stage, cancer grade, and, for breast cancers, estrogen and progesterone receptivity. Sensitivity models that measured cancer severity more granularly, using SEER's derived American Joint Committee on Cancer 6 th edition stage groupings, produced results that were similar to the preferred models.     # Hospitals with Cancer Programs 9 (7) 6 (6)

Results
Tables 5A (all-causes mortality) and 5B (cancer-caused mortality) report the linear regression estimates.
The estimates from the model that includes the control variables are the preferred estimates. The estimates for all-causes mortality (Table 5A)

Cox Proportional Hazards: Methodology and Results
As shown in the previous section, the choice between all-causes or cancer-caused mortality as the outcome has a small but clinically meaningful impact on the estimates. This is presumably due to bias that results from MA cancer patients having more non-cancer comorbidities (or, more generally, poorer non-cancer health) than FFS cancer patients. In the absence of data on non-cancer comorbidities, the preferred approach to address this bias is to use a Cox proportional hazards model that treats non-cancer deaths as censoring events. Unlike with a linear regression model, which cannot address right-censoring, no assumption is needed regarding future survival for patients who die for reasons unrelated to their cancer.

Methodology
Unlike the linear regression that modeled the probability of death within two years after cancer diagnosis, the hazard rate modeled in the Cox proportional hazards model is the probability of death in the n th month after cancer diagnosis given survival through the (n-1) st month. The Cox proportional hazards model is the simplest and most popular hazard model. It does not impose any assumptions regarding the underlying hazard rate. Instead, it only assumes that the relationship between the key variable(s) and the hazard rate is proportional to survival duration. Schoenfield residuals are commonly used to statistically test this assumption and the Schoenfield test for this study revealed no evidence against the proportionality assumption.
Similar to fixed effects in a linear regression, covariates can be used as stratifiers in a linear regression.
Stratifiers do not have a coefficient that is estimated and need not have a true relationship with the hazard rate this is proportional to survival duration. All of the variables that were used as fixed effects in the linear regression model (age, year of diagnosis, HSA, and cancer severity) were used as stratifiers in the Cox proportional hazards model.

Results
Tables 6A (all-causes mortality) and 6B (cancer-caused mortality) report the estimates from the Cox proportional hazards models. Again, the estimates from the models that include the control variables are the preferred estimates. For lung and prostate cancers, the estimates from the model without any    For lung cancer patients, MA enrollment was associated with an 8.9 percent increased hazard of dying from any cause and a 12.6 percent increased hazard of dying from cancer. For colorectal cancer patients, MA enrollment was associated with an 11.5 percent increased hazard of dying from any cause and a 17.8 percent increased hazard of dying from cancer.

Limitations
This study had several limitations. Though the statistical analysis attempted to minimize selection bias,  (Baker, Phillips, Haas, Liang, & Sonneborn, 2004;and Rizzo, 2005). It is thus possible that MA enrollment may improve pre-diagnosis cancer care.
However, it is important to note that routine cancer screening is not recommended or common for lung cancer.
Second, the SEER-Medicare data is not a nationally representative sample. However, SEER regions contain over one-fourth of the US population. They are also similar to non-SEER regions in terms of cancer incidence by age and race (Kuo & Mobley, 2016), and in terms of MA penetration rates over time. Third, due to data constraints, this study was unable to adjust for quality of life. Fourth, this study was not able to examine any heterogeneity within types of MA plans.

Discussion
Conditional on having Medicare Part D and controlling for patient characteristics, MA enrollment was found to be associated with increased mortality for patients diagnosed with cancer. The most convincing statistical evidence was for lung cancer, followed by colorectal cancer. These two cancers, and especially lung cancer, are much more deadly than breast and prostate cancer on average.
Like any observational study, the potential for selection bias due to unobserved factors must be taken seriously. However, a rich set of controls was used and the estimates (particularly for lung cancer cancer) were not sensitive to those controls. Additionally, a large risk selection literature implies that any selection bias is most likely to understate any effect of MA enrollment to increase mortality. If the results of this study are driven by selection bias, it is a new selection bias that has not yet been documented in the literature.
This study was unable to confirm a mechanism for the observed positive association between MA enrollment and mortality for cancer patients. One hypothesis, based on the trade-off of enhanced coverage but restricted provider access in MA plans compared to FFS, is that restricted provider access in MA plans increases mortality for patients with particularly deadly cancers that may benefit from access to particular specialists. Examining the implications of restricted provider access for vulnerable populations is a promising area for future work, particularly now that MA encounter data have become available to researchers (Note 4). In studies that were not restricted to cancer patients, evidence indicates that MA patients are admitted to lower-quality hospitals (Friedman & Jiang, 2010) and lower-quality nursing homes (Meyers, Mor, & Rahman, 2018) than FFS patients.