The data are from Cohorts 6–8 of the Medicare Health Outcomes Survey (HOS), an annual national survey of Medicare beneficiaries voluntarily enrolled in Medicare Advantage (MA) private health plans (we were limited to three cohorts because of resource constraints, and the questionnaire switched versions beginning in Cohort 9, so we selected earlier years to ensure consistency in the measures). All managed care plans with MA contracts are required to participate in the HOS, which is used mainly for quality improvement programs, reports, and monitoring and improving health outcomes. When the data were collected, about 13 % of the Medicare population was enrolled in MA plans (Gold et al. 2010).

MA plans include persons aged 65+ and a small number of persons aged <65 with disabilities. Surveys are mailed to a randomly selected cohort of MA plan enrollees each spring, with a follow-up survey 2 years later. Non-respondents are surveyed by telephone. For large plans, a simple random sample of 1,000 members are surveyed; for smaller plans with fewer than 1,000 members, all members are surveyed. HOS eligibility requires only continuous enrollment in the same plan for at least 6 months and no end-stage renal disease. Approximately 160,000 individuals are sampled for each cohort, and about 105,000 respond to the complete survey in each cohort.

This study obtained an HOS limited data set including dates of death from the Centers for Medicare & Medicaid Services and an exemption from the investigators’ institutional review boards.

The analysis sample includes all individuals aged 65+ from HOS Cohorts 6–8 whose baseline surveys were collected between 2003 and 2005. HOS members under 65 were excluded because they are eligible for Medicare due to disability and represent a specialized population. Individuals were also excluded from the analysis if they had missing values for any of the HRQOL-4 variables, marital status, or health indicators (n =96,583). Proxy respondents were excluded (n = 38,473) because perception of HRQOL varies by reporter (Albrecht and Devlieger 1999; Andresen et al. 2001; Ellis et al. 2003). The total analysis sample was 191,001.

The primary dependent variable is mortality status. Short-term mortality is defined occurring within 90 days of interview; long-term mortality is the status within the maximum follow-up length, slightly more than 2.5 years (analysis of 30 days and 1 year is available upon request). The time between the baseline survey and an individual’s last observation—either date of death or the end of follow-up—was used in all Cox survival models described below. The exact date of death on the HOS is identified from the Medicare Enrollment Database and Social Security Administration Master Death File.

The primary independent variables are the core CDC Healthy Days measures (HRQOL-4), which have been included on the HOS since 2003. These items have undergone cognitive testing and have demonstrated content, construct, criterion, and predictive validity, test–retest reliability, and internal consistency (Andresen et al. 2001; Moriarty et al. 2003; CDC 2000). The items include GRSH, physically unhealthy days (PUDs), mentally unhealthy days (MUDs), and days with activity limitations (ALDs). GSRH is assessed on a five-point Likert scale (“excellent,” “good,” “fair,” “poor,” or “bad”). PUDs are defined by the question “Now thinking about your physical health, which includes physical illness and injuries, for how many days during the past 30 days was your physical health not good?”; MUDs by “Now thinking about your mental health, which includes stress, depression, and problems with emotions, for how many days during the past 30 days was your mental health not good?”; and ALDs by “During the past 30 days, for about how many days did poor physical or mental health keep you from doing your usual activities, such as self-care, work, or recreation?”

In the analysis, GSRH was used as a categorical variable with “excellent” as the reference group. PUDs, MUDs, and ALDs were grouped into 0-, 1- to 10-, 11- to 20-, and 21- to 30-day outcomes, with 0 day as the reference group, following past precedent (Dominick et al. 2002).

All analyses adjusted for age, sex, race/ethnicity, marital status, household income categories, educational status, and Spanish survey language. In addition, a count of 12 self-reported chronic diseases (hypertension or high blood pressure; stroke; emphysema, asthma or chronic obstructive pulmonary disease; Crohn’s disease, ulcerative colitis, or inflammatory bowel disease; arthritis of hip, knee, hand or wrist; diabetes, high blood sugar, or sugar in the urine; congestive heart failure; any cancer other than skin cancer; angina pectoris or coronary heart disease; a myocardial infarction or heart attack; sciatica; other heart conditions, such as problems with heart valves or rhythm), six activities of daily living (ADLs; 0–2, 0=no difficulty, 1=has difficulty, 2=unable to do the activity), back pain (1–5, 5=greatest pain), smoking frequency (never, sometimes, or every day), and indicators for treatment of breast, colon, lung, and prostate cancer were included in all models. Survey mode and MA plan type indicators were also included.

Cox proportional hazards analysis of the risk of mortality as a function of time to death since baseline, HRQOL, and other covariates were run as both bivariate and multivariate models. Cox models (Cox 1972) are a semi-parametric survival model for estimating the rate of mortality based on the time to event. Censored data, in which mortality is not observed, enter the model naturally. The time variable in all models was the number of days from baseline until death, if observed, or the censoring date. The Cox model assumes proportional hazards over time, implying that the hazard ratio for a given covariate does not change over time. This assumption was tested by interacting HRQOL variables with time.

Two approaches were used to assess the relative importance of the three HDs compared to GSRH for predicting mortality. Predictive validity was assessed using the C-index (18), which ranges from 0.5 to 1.0, and reflects the proportion of paired outcomes in which differing survival times are correctly predicted by the model. To compare the relative predictive power of different factors, the C-index was estimated for several models: (1) age; (2) age, GSRH; (3) age, HDs; and (4) age, GSRH, HDs. DeSalvo et al. (2005) and Quinten et al. (2009) also used similar approaches to assess the performance of HRQOL for predicting mortality outcomes. In addition, the full model (age, GSRH, HDs) was run with and without each individual term to compare the difference between the full model and each reduced form.

Relative importance was also assessed for the same models by the proportion of explained variation (PEV; Schemper 1993; Schemper and Stare 1996; Heinze and Schemper 2003). PEV measures the amount of variation in mortality (0–100 %) attributable to independent variables relative to the total variation in mortality. The partial PEV for each factor of interest reflects the PEV with and without the factor of interest.

Among the 191,001 respondents in this analysis, mean age was 75 years, 10 % were non-White, 58 % were female, and 58 % were married (Table 1). Within 90 days of the baseline survey date, 718 respondents (0.4 %) died. Over the maximum observation time of slightly more than 2.5 years, 13,414 (7 %) had died. Table 2 shows responses for each of the HRQOL-4 items. Compared to all persons in the HOS sampling frame, including non-respondents, the analysis sample is slightly younger (among those aged 65 and older) and has fewer minorities, more married persons, and higher education levels. All of these factors are associated with conditioning on non-proxy respondents and non-missing data. A detailed table of comparisons is available from the authors upon request.

At the maximum follow-up time, age was the strongest predictor of mortality, with an estimated hazard ratio (HR) of 1.09 (P <0.0001) at age 65 that increased with age (Table 3). The HRs for each category of the four HRQOL variables were also highly significant (P <0.001), were greater than 1, and increased with each decrement in HRQOL. For any level of HDs (1–10, 11–20, and 21–30), the HRs were smallest for MUDs and largest for ALDs, ranging from 1.23 for 1–10 MUDs to 3.65 for 21–30 ALDs. Most other covariates had the expected relationship with mortality, such as increased HRs for males, persons with more chronic diseases, for those with ADLs, and smokers (all P <0.0001). Married respondents and persons with higher income (both P <0.0001) had decreased HRs. HRs were <1.0 for those responding in Spanish (P = 0.004) and for races other than White and Black (P =0.001), and HRs exceeded 1.0 for telephone respondents (P <0.0001).

For short-term (90-day) survival, HRs increased with worse categories of HRQOL for GSRH, PUDs, and ALDs. HRs for PUDs and ALDs were significant at 11–20 days (P <0.05) and 21–30 days (P <0.001), roughly 1.4 and 1.7 for each measure, with confidence intervals overlapping across comparable levels of the two variables (second and third columns of Table 3). HRs for “fair” and “poor” GSRH were statistically significant (P <0.001) and larger than HRs for any levels of PUDs or ALDs. Contrasting with the other measures and the bivariate results, the HR for the only significant level of MUDs was <1, indicating a decreased risk of mortality (P <0.05) with 11– 20 days of “not good” mental health in the past 30 days relative to no days of “not good” mental health. The relationship for other control variables was largely as expected, although fewer of these variables were significantly associated with mortality relative to the bivariate analysis. However, back pain score is associated with decreased risk of mortality (P <0.001) compared to increased risk in the bivariate analysis.

Over the maximum follow-up, several additional relationships that were not significant over 90 days were statistically significant. HRQOL results were consistent with the 90-day period, although the HRs are closer to 1.0 in most cases except for MUDs, indicating that the proportional hazards assumption may not have been met. Likewise, worse HRQOL corresponded to larger HRs, except for MUDs. All included categories of GSRH were significantly associated with mortality (P <0.05 or better). All categories of ALDs were significant at P <0.01 or better, but only 21–30 days of PUDs was significant (P <0.001). At 90 days, greater MUDs were associated with a decreased risk of mortality, but only 21–30 days was statistically significant (P <0.01).

To test the proportional hazards assumption of the Cox model, time interactions with the HRQOL variables were included in an additional multivariate model of maximum follow-up length (Table 4). Except for “fair” GSRH, all interactions were statistically significant for the same HRQOL terms that were significant in the corresponding model without time interactions in Table 3. Between short- and long-term follow-up, the combined HR for “poor” GSRH decreased from 7.1 to 3.5 (P <0.05). For PUDs, only 21–30 days was significantly associated with increased mortality (P <0.05), and the HR declined from 1.5 to 1.1 over the same period. The HRs for ALDs decreased from 1.6 to 1.1 for 11–20 days (P <0.05) and from 1.9 to 1.2 for 21–30 days (P <0.01). The only significant interaction for MUDs, 11–20 days (P <0.05), moved in the opposite direction over time, with HR increasing from 0.8 to 1.1 by 2.5 years. All other covariates in Table 3 were also included in these models. For the time interactions at 90 days only, interactions between GSRH and time were significant.

Using the total C-index, the HRQOL variables added more information to total fit in the short-term model, while age added more information over the long-term model (Table 5, first panel). The added contribution of the four HRQOL variables ranged from 0.12 to 0.17 (19–27 % relative increase) at 90 days compared to only 0.07–0.08 (10–11 %) at 2.5 years. For overall cohorts and both follow-up lengths, GSRH alone added more to the fit than all three of the HD measures.

The relative contributions of age, GSRH, and the three HDs were assessed in terms of the average change in PEV or individual contributions to the full model C-index (Table 5, bottom panel). GSRH was the single largest predictor for the short-term model of 90 days and the second most important predictor in all cohorts at 2.5 years. Age was the most important predictor over the long-term model of about 2.5 years and the second most important predictor in the 90-day model. The relative importance of all three HD measures, compared to GSRH and age, was lower in both the short- and long-term models, whether assessed by PEV or C-index. In the three healthy days measures, the long-term results display a clear pattern in both the PEV and C-index results: ALDs are most predictive of long-term mortality, PUDs the next most predictive, and MUDs provide almost no information. In the short-term models, the pattern is not as consistent. The PEV shows ALDs to be most predictive of the three, while MUDs and PUDs vary in rank depending on the cohort. The C-index results show PUDs more predictive in Cohorts 7–8 and ALDs in Cohort 6, with MUDs always at the bottom.