Study population
We used data from the Cooperative Health Research in the Region of Augsburg (KORA)-Age study, which is a part of the regional KORA research platform for population-based health research in Germany. The KORA research platform consists of population-based surveys and their follow-up studies. The KORA-Age study is a follow-up of participants ≥65 years from four independent cross-sectional samples who completed health surveys conducted between 1984 and 2001 [19]. A population-representative selection of participants from population registries in the city of Augsburg along with two adjacent counties (total population in 2016: 668,500) in the federal state of Bavaria took place [20].
At Age1 (t0, in 2008), the eligible study population consisted of 5986 individuals born before 1944. The first follow-up, Age2 (t1, in 2011/12), only included individuals from a sex- and age-stratified subsample of Age1 participants (n = 1079) with 100 people per stratum (males and females in five age groups, i. e. ten strata). Of those, 822 participated in medical examinations and completed a telephone interview (response rate: 84.3%) consisting of validated questions on, e. g., sociodemographic characteristics, PA, morbidity, and utilization of health care services [21]. Proxies (e. g., informal caregivers) were interviewed if the participant was unable to answer the questions (n = 29 [3.5% of participants]). For the second follow-up (Age3, t2, in 2016), the total sample of Age1 and individuals from the four cross-sectional samples who were born before 1951, and thus aged ≥65 years in 2016, were invited to participate. This resulted in an eligible study population of 6051 at Age3 (t2). Of those, 4083 participated in telephone interviews and questionnaires (response rate: 67.5%; completed by proxies: n = 191 [4.7% of participants]).
Since Age1 (t0) did not assess information on utilization of LTC, only Age2 (t1) and Age3 (t2) were considered for analyses. For the main analysis, we used individuals from t1 (n = 822) and t2 (n = 4083) [see Fig. 1].
Approval for the KORA-Age study was obtained from the Ethics Committee of the Bavarian Medical Association. Individuals agreed to participation with informed consent. Further details on data collection, study design, sampling, and response rates are described elsewhere [22, 23].
Measurement and operationalization of physical activity
According to the World Health Organization (WHO), PA is defined as “any bodily movement produced by skeletal muscles that requires energy expenditure” [5]. PA is an umbrella term for different forms of movement [24]. In this study, we investigated “exercise” and “walking” as types of PA. In the following lines, relevant information about PA according to the “Physical Activity Questionnaire Reporting Checklist” from Nigg et al. is reported [24].
“Exercise” is a planned, repetitive, purposeful, and structured subcategory of PA with the objective to improve or maintain one or more components of physical fitness [5, 7]. In our study, it was assessed with these two questions addressing duration of exercise: “How often do you exercise during winter?” and “How often do you exercise during summer?”. Response categories for those questions were (1) “regularly more than or equal to two hours per week”; (2) “regularly more than or equal to one, but less than two hours per week”; (3) “less than one hour per week”; and (4) “no exercise”. As operationalized by Karl et al. [25], the two responses for summer and winter were initially combined as one variable with the values “high exercise” (= (1) in both summer and winter), “moderate exercise” (= combinations for summer and winter of “(1) + (2)”, “(2) + (2)”, or “(1) + (3)”), “low exercise” (= combinations for summer and winter of “(1) + (4)”, “(2) + (3)”, “(2) + (4)”), and “no exercise” (= combinations for summer and winter of “(3) + (3)”, “(3) + (4), “(4) + (4)”). To facilitate the analyses and interpretation of results, we further dichotomized these values into “high or moderate exercise” (hereafter called “exercise”) and “no or low exercise” (hereafter called “no/low exercise”). Additional file 1 illustrates a detailed differentiation of the categories.
“Walking” was assessed with the following question addressing duration of walking: “On a typical weekday, how much time do you spend walking? For example, going for a walk, on the way to work or shopping?”. Possible response categories were (1) “more than or equal to one hour”; (2) “more than or equal to half an hour, but less than one hour”; (3) “more than or equal to a quarter of an hour, but less than half an hour”; (4) “less than a quarter of an hour”; or (5) “not applicable” (= no walking due to, e. g., using a wheelchair). To facilitate the analyses and interpretation of results, we further dichotomized these values into “high or moderate walking” (= (1) or (2), in the following: “walking”) and “no or low walking” (= (3), (4), or (5), in the following: “no/low walking”). “Walking+exercise” was applied to individuals doing both “exercise” and “walking”.
The domain of PA – with the exception of walking while “going to work” representing occupational- or transport-based PA – was mainly leisure-time [24]. However, in Germany most people ≥65 years are retired, which allows focusing on “leisure-time PA”. Recall periods were “a typical (summer/winter) season” (exercise), and “a typical weekday” (walking) [26].
To address important components of PA, frequency (number of sessions per week), intensity (walking, moderate-intensity exercise, vigorous-intensity exercise, strength training), and time (average duration of an individual session per week; unit of measurement: “minutes per week”) were assessed in the subpopulation at t2. Questions, response options, and the calculation of the amount of PA per week can be found in Additional file 2.
Measurement of utilization of long-term care
LTC is defined as support with daily activities for people who experience decline in self-care on a long-term basis (> three months) [27]. Daily activities consist of activities of daily living (ADLs) (e. g., dressing, bathing) and instrumental activities of daily living (IADLs) (e. g., cooking, cleaning) [28]. In Germany, there are three main forms of assistance in LTC for community-dwelling adults: formal support with ADLs, formal support with IADLs, and informal support with both ADLs and IADLs [29, 30]. Study participants were asked if they had received LTC due to their health status within the last three months [21]. Types could be (1) a home nursing service (i. e., formal support with ADLs); and/or (2) paid services for household support (i. e., formal support with IADLs); and/or (3) support from friends, family members, or neighbors (i. e., informal support with both ADLs and IADLs). If participants answered “yes” for at least one of the three types, the variable “utilization of LTC” (yes/no) was coded as “yes”. As all individuals were community-dwelling, settings for LTC were either community- or home-based.
Covariates
Covariates related to LTC and PA were identified based on Andersen’s Behavioral Model of Health Services Use (ABMHS) [18, 31,32,33,34], the GNPAR [35], and common correlates of PA in adults [36].
Identified sociodemographic factors were age, sex, education, living arrangement, and income. Age was the participant’s age on the interview date. Sex was defined as the biological distinction of “female” or “male”. Education was comprised of school education, education at university, and vocational training. It was expressed in years (8–17 years). Living arrangement was divided into living “alone” or “not alone”. Income was defined as self-perceived income sufficiency (subjective income). In older adults, this is a common approach to express individuals’ personal evaluations of the relationship between wealth or objective income and their expenses [37]. Participants were asked if, on average, their income was enough to support them until the end of the month. The responses were divided into “sufficient” or “scarce/insufficient”.
Identified health-related factors were falls, multimorbidity, disability score, and Body Mass Index (BMI). Falls were reported as “≥ 1 fall” or “no falls/unknown” within the last year. We used participant self-reports to calculate their Charlson Comorbidity Index [38]. The index considers 13 types of chronic conditions: lung, heart, joint, kidney, gastrointestinal, liver, neurological, and eye diseases; stroke; diabetes mellitus; cancer; hypertension; and HIV [39]. In our study, multimorbidity was defined as the sum of reported chronic conditions ranging from 0 to 13. The Stanford Health Assessment Questionnaire Disability Index (HAQ-DI) was used to measure disability [40]. The HAQ-DI analyzes impairments in IADLs and ADLs. It consists of 20 questions about physical function in eight domains: dressing and grooming, hygiene, eating, standing up, walking, reach, grip, and activities [40]. Its responses range from 0 (no difficulty) to 3 (unable to perform). The highest score in a domain was taken as the domain’s score. The mean of all eight domains constituted the HAQ-DI, which was reported as a continuous value. Participants’ height and weight were measured at the study center through consistent and validated measurement methods (daily calibrated scales; stadiometer) at t1 and were assessed using self-reports following detailed instructions from trained telephone-interviewers at t2. From these values, their BMI in kg/m2 was calculated.
Statistical analyses
We assessed participants’ characteristics at both follow-up timepoints, dropouts, PA values, and comparisons with the GNPAR using descriptive statistics. To investigate the association of different types of PA (walking, exercise, walking+exercise) with utilization of LTC at t1 or t2 (i. e., cross-sectional analysis with repeated measurements), we applied a generalized estimating equation (GEE) logistic model with an unstructured correlation matrix. The model accounts for repeated measurements and their intra-subject correlation [41]. As the study focused on population-averaged effects and not individual, subject-specific changes, we did not apply mixed models, which would have been an alternative for examining intra-subject correlation [42].
In the first step, we analyzed the association of distinct types of PA with utilization of LTC in the entire cohort (n = 4812 observations: sum of t1 (n = 800) and t2 (n = 4012)). As the existence of differences between sexes for utilization of health care services and lifestyle habits is well-known [17, 18, 34], in the second step we applied two sex-stratified models: one for females (n = 2499 observations: t1: n = 395, t2: n = 2104) and one for males (n = 2313 observations: t1: n = 405, t2: n = 1908). As disability (expressed as, e. g., “poor health” or “impaired general physical functioning” [43,44,45,46]) is one of the major barriers to PA, we applied the sensitivity analysis (SA) SA1 to each model (entire cohort, females, males). The SA1 included only observations without disability, defined as an HAQ-DI < 0.5 [47, 48]. For the general model (entire cohort) and the observations without disability (SA1), we assessed the interaction of sex with types of PA through the calculation of the respective interaction terms.
In all models, we compared observations with utilization of LTC to those without utilization of LTC. The covariates sex and education did not change from t1 to t2 and thus were considered fixed variables. All other covariates were treated as time-dependent.
Individuals with either missing values in all types of LTC (= transformation variable “utilization of LTC” (outcome)) or in both “exercise” and “walking” (= transformation variable “PA” (exposure)) were excluded through listwise deletion [49, 50]. This resulted in a final sample size of n = 800 at t1 and n = 4012 at t2. Missing values in subdomains of the variables “utilization of LTC” (t2: home nursing service and paid services for household support (n = 1); assistance of family members, friends, or neighbors (n = 1)) or “PA” (t1: walking (n = 2); t2: exercise (n = 2); walking (n = 8)) were imputed through single stochastic regression imputation using logistic regression with the fully conditional specification method [51]. This imputation strategy is based on the assumption that missing values are missing at random, meaning that they are conditionally independent from the unobserved value, hence the underlying missing data pattern is arbitrary [52, 53]. To test our model’s robustness, we conducted SA2. It excluded observations with missing values in the above-mentioned subdomains of outcome or exposure.
Regarding covariates, twenty missing values (2.5%) at t1 (multimorbidity (n = 9), BMI (n = 7), income (n = 4)), and 126 missing values (3.1%) at t2 (BMI (n = 53), income (n = 35), multimorbidity (n = 27), falls (n = 5), disability score (n = 2), living arrangement (n = 2), education (n = 2)) were identified. We imputed binary variables using single stochastic regression with the fully conditional specification method and continuous variables using predictive mean matching [51]. We based imputation of all missing values mainly on the models’ covariates (|correlation coefficient| > 0.4) [49].
We tested for multicollinearity of covariates in all models (threshold: |r| ≤ 0.8). We calculated odds ratios (OR) and 95% confidence intervals (CI). In all analyses, results with p-values ≤0.05 were considered statistically significant. We performed all statistical analyses using SAS software, release 9.4 (SAS Institute, Cary, NC).