Data source

We analyzed longitudinal data from Rounds 1–9 (2011–2019) of the National Health and Aging Trends Study (NHATS). NHATS is a prospective cohort designed to investigate trends in disability, physical functioning, and aging in older adults in the United States [13, 14]. The study conducts annual in-home interviews for a nationally representative sample of over 8,000 Medicare beneficiaries aged 65 and older. While at the time of writing NHATS data were available through Round 12 (2022), we only included data through Round 9 because the performance-based measures used to identify frailty (grip strength and walking speed tests) were not captured in Round 10 (2020) due to the COVID-19 pandemic. Our study included all older adults who completed the primary NHATS data collection instrument (the Sample Person interview) in 2011. We excluded older adults with missing cancer or diabetes history in Round 1 (N = 8) and those with a history of skin cancer (N = 648). Those with a history of skin cancer were excluded due to the high prevalence of non-aggressive skin cancers that behave differently than other cancers, which could skew results for those with a cancer history. Since NHATS does not differentiate between malignant melanoma and other skin cancers in their public use files, those with malignant melanoma were also excluded.

Frailty

Frailty was measured annually from Rounds 1–9 using the Fried frailty phenotype [4]. The Fried frailty phenotype defines frailty as a clinical, aging-related syndrome based on the presence of five clinical signs or symptoms: exhaustion, low physical activity, shrinking, slowness, and weakness. We assessed the five frailty phenotype components in NHATS using self-reported and performance-based measures based on definitions from prior research (e.g., slowness and weakness are based on grip strength and walking trials, respectively) [11]. Consistent with the original definition, older adults who met definitions for three to five frailty phenotype components were considered “frail”, those with one to two components were considered “prefrail”, and those with zero were considered “robust” [4].

Cancer and diabetes history

Cancer history was determined based on responses to the following question during the initial NHATS interview: “Has a doctor ever told you that you had cancer”? Although NHATS collects additional details on the specific type of cancer, these are not available in the public use files and were not available for our analysis. Individuals with a history of diabetes during the initial NHATS interview were identified through their responses to the survey question: “Has a doctor ever told you that you had diabetes?” We stratified all our analyses using the following categories: neither condition, cancer only, diabetes only, or both cancer and diabetes.

Other variables

Death is determined in NHATS by attempted follow up and confirmation from a proxy respondent or family member [13]. We identified several covariates in the NHATS data to account for differences in demographic and clinical characteristics by cancer and diabetes history and to account for potentially informative missing data. These variables included socioeconomic status (wealth and educational attainment), demographics (age, self-reported race and ethnicity, gender, metro/non-metro residence, and marital status), BMI, use of mobility devices, and prior history of myocardial infarction, heart disease, high blood pressure, arthritis, recent falls, osteoporosis, lung disease, stroke, depression, and anxiety [13, 15, 16].

Statistical analysis

This study received expedited approval from the Wake Forest School of Medicine Institutional Review Board (#IRB00108630). All analyses were conducted using SAS Version 9.4 (SAS Institute Inc., Cary NC) and R Version 4.2.3.

Missing data

Fifteen percent of individuals were missing frailty information in Round 1. We accounted for missing frailty and covariate data using multiple imputation with fully conditional specification (also called multiple imputation with chained equations) [17, 18]. This missing data approach fits a series of iterative predictive models to multiply impute each variable with missing data. Our imputation models included the outcome (frailty) and stratification variables (cancer and diabetes history), demographics (age, race, gender), socioeconomic status (wealth, educational attainment), other comorbidities, history of falls, and use of mobility devices. We generated ten multiply imputed datasets using ten burn-in iterations. All statistical analyses were conducted within each multiply imputed dataset and then combined to generate a single parameter estimate using Rubin’s rule [19]. Since cancer and diabetes history had a low degree of missingness (N = 8), we excluded those with missing cancer and diabetes information, rather than imputing these variables.

Standardization

We accounted for differences in the age, race, and gender distributions across cancer and diabetes strata using standardization via standardized mortality ratio (SMR) weighting [20]. Older adults with a history of both cancer and diabetes served as the standard, as the primary goal of our study was to draw inference to those with comorbid cancer and diabetes. SMR weights were calculated separately within each multiply imputed dataset. All statistical analyses incorporated both the SMR weights and the NHATS survey sampling weights to draw inference to Medicare beneficiaries over 65 years of age in 2011.

Baseline prefrailty and frailty prevalence differences

We estimated the difference in the prevalence of prefrailty and frailty during Round 1 for older adults with cancer only, diabetes only, a history of both cancer and diabetes, or neither condition using an SMR- and survey-sample weighted multinomial regression model with the frailty phenotype as the dependent variables (robust was the reference group) and cancer and diabetes strata as the explanatory variable. Older adults without a history of cancer or diabetes were the reference group for comparison.

Multistate model

We described longitudinal frailty state transitions across the three frailty phenotype categories (robust, prefrail, and frail) and death between NHATS rounds using an interval-censored, nonparametric multistate model [21]. We implemented the multistate model using a two-step approach. First, the risk of mortality over time was estimated using an SMR- and survey-sample weighted Kaplan-Meier estimator (using the svykm function in R) to account for right censoring due to loss-to-follow-up in NHATS [22]. Then, the proportion of individuals within each frailty phenotype state among those alive in each round was non-parametrically calculated based on the observed data. The multistate models were fit separately within the four cancer and diabetes strata.

Cumulative incidence of prefrailty and frailty

We estimated the nine-year cumulative incidence of (1) prefrailty or frailty and (2) frailty using SMR- and survey-sample weighted Aalen-Johansen estimators that accounted for death as a competing risk and censoring due to loss-to-follow-up [23, 24]. Those who were robust in Round 1 of NHATS served as the study population in the analysis of the cumulative incidence of prefrailty or frailty, and those who were robust or prefrail in Round 1 were included in the analysis of the cumulative incidence of frailty. These study populations were selected to represent the populations at risk for prefrailty and frailty, respectively. 95% confidence intervals were estimated based on the 2.5th and 97.5th percentiles from 1,000 bootstrapped samples [25].

We conducted a sensitivity analysis that considered a composite outcome of frailty or death over the nine-year period. The time to the first occurrence of frailty or death was considered the event. The cumulative incidence of frailty or death, among those who were robust or prefrail at baseline, was estimated using an SMR- and survey-sample weighted Kaplan-Meier estimator, stratified by cancer and diabetes history.