Sample
Data for this analysis were derived from the German Ageing Survey (DEAS – “Deutsches Alterssurvey”) which was conducted as a nationwide representative cross-sectional and longitudinal study of the community-dwelling population aged older than 40 years in Germany. The survey is funded by the German Federal Ministry for Family Affairs, Senior Citizens, Women, and Youth (BMFSFJ). Since 2002, the study is organized by the German Centre of Gerontology (DZA – “Deutsches Zentrum für Altersfragen”). The cross-sectional baseline samples have been set up on a recurring four-year basis starting in 1996 and ending in 2014. These baseline samples were stratified by gender, age as well as region (Eastern/Western Germany) and randomly selected from population registers within municipalities. If a written permission was given, participants from the baseline samples were re-contacted to take part in additional waves. Data collection for panel assessments were performed in 2002, 2008, 2011, 2014. The sixth wave (2017) is currently conducted.
The objective of the DEAS is to provide a unique database that enables researchers to study the multifaceted living conditions of the German population aged older than 40 years as well as the consequences of aging on individuals and processes of social transition. Therefore, participants were asked questions regarding their socio-demographic circumstances, their living conditions as well as different topics related to aging by using oral face-to-face interviews. These personal interviews took place at the participants’ homes and were carried out by trained interviewers organized as Computer-Assisted Personal Interview (CAPI). At the end of the interview, respondents were asked to fill out a standardized questionnaire that covered, for example, subjective attitudes as well as psychological topics and health.
Due to data availability in the outcome measures and our variable of interest, we restricted the analyses to the second (2002), third (2008) and fourth (2011) waves. Differences between the sample sizes were due to different recruitment approaches. In 2002, 2008 and 2011, individuals who had participated in previous waves were asked for a re-interview. New study participants were solely included in 2002 and 2008. The overall sample size in 2002 was 5194 individuals (of whom 3670 were new and 1524 were re-interviewed) and in 2008 8200 individuals (of whom 6205 were new and 1995 were re-interviewed). In 2011, 4854 individuals took part in the survey interviews (of whom 0 were new and 4854 were re-interviewed). Response rates were 38% in 2002, 38% in 2008 and 56% in 2011. In total, the response rates of the DEAS were consistent with other German surveys, but rather low compared to other European studies on aging [25]. Time or health restrictions as well as refused re-participation were the most common reasons for missing follow-up data [26]. More details of the DEAS have been reported elsewhere [27]. The study was conducted in accordance with the ethical standards of the Helsinki Declaration. Small incentives were provided to all individuals who participated.
Variables
Dependent variables
The use of health care services was measured retrospectively by means of inpatient and outpatient services for 12 months preceding the interview. Outpatient services were assessed by the number of GP visits as well as the number of specialist visits (house calls included). The variable specialists comprise several medical specialties, which are reported in Additional file 1. The corresponding number of GP and specialist visits was measured as “never”, “once”, “2–3 times”, “4–6 times”, “7–12 times”, or “more often” (open answer). Following Bock et al. [28], it was recoded as “never” = 0; “once” = 1; “2–3 times” = 2.5; “4–6 times” = 5; “7–12 times” = 9.5; and “more often” = 13. Regarding the inpatient sector, the number of days in hospital was assessed. The question on hospitalization was treated as a binary variable (1 = “at least one night in hospital”; 0 = “not one night in hospital”).
Mental health services were not investigated separately in our study, as their percentage of usage is quite small (Neurologist/Psychiatrist services in 2002: 6.49%, in 2008: 5.84%, in 2011: 10.42%). Furthermore, visits to psychotherapists and radiologists were not considered due to reasons of data availability. In our study, we did not differentiate whether the type of ward was psychiatric or somatic in the hospital.
Independent variables
Our key explanatory variable was voluntary work in groups and organizations. The respondents were asked if they execute an honorary office in the groups or organizations in which the person is a member. Response choice of the volunteer variable was binary (1 = “yes”; 0 = “no”).
To investigate the study hypothesis, various variables were chosen to be included in the regression models as an alternative way to potentially explain the relationship between volunteering and health care use among older adults. Variables were selected based on the theoretical framework by Andersen’s behavioral model, preceding research results as well as theoretical interest [29, 30].
The Andersen’s behavioral model is one of the most widely acknowledged models to identify determinants that could plausibly be associated with health care use [30]. It categorizes three basic individual components of health care use: Predisposing factors (socio-demographic and health-related belief characteristics, such as age, gender, education and health attitude), enabling resources (such as income and social insurance status) as well as need for health care (perceived and evaluated state of health) [29].
Regarding predisposing factors age, gender (male/female), marital status (married, living together with spouse, others [married, living separated from spouse; divorced; widowed; never married]), employment status (working; retired; other: not employed) and educational level were included. Educational level was categorized according to the International Standard Classification of Education (ISCED-97) [31] scale with three categories: low (ISCED 0–2), medium (ISCED 3–4) and high (ISCED 5–6).
Enabling resources covered the (log) monthly equivalent net income in Euro (according to the new OECD equivalence scale) and self-rated accessibility of doctors and pharmacies (1 = “there are enough doctors and pharmacies in the vicinity”; 0 = “there are not enough doctors and pharmacies in the vicinity”).
Regarding need factors, self-rated health and morbidity were measured. Self-rated health was quantified by using five states ranging from 1 = “very good” to 5 = “very bad”. To assess morbidity, the number of chronic diseases was recorded which was adapted from the Charlson Comorbidity Index [32]. Furthermore, lifestyle factors, such as current smoking status (1 = “currently smoking”; 0 = “currently not smoking”) and self-reported body mass index (BMI) were included. BMI thresholds were classified according the World Health Organization (WHO): underweight (BMI < 18.5 kg/m2), normal weight (18.5 kg/m2 ≤ BMI < 25 kg/m2), overweight (25 kg/m2 ≤ BMI < 30 kg/m2), and obese (BMI ≥ 30 kg/m2) [33].
Statistical analyses
To estimate the impact of volunteering in groups and organizations on the use of health care services, FE regressions were used. As a special feature, FE regression models allow for the association between time-constant factors and the explanatory variables. Under the assumption of strict exogeneity, FE regressions lead to consistent estimates [34]. In contrast, techniques such as the pooled ordinary least square (POLS) or the random effects (RE) would lead to inconsistent estimates when unobserved factors and the explanatory variables are correlated. Time-constant unobserved factors, such as gender or genetic disposition, are a widespread topic, especially in social sciences that need to be considered within modern research [35].
Our choice towards the FE specification has been confirmed by performing the Hausman test [36]. It basically investigates whether there is an association between the unobserved time-constant factors and the explanatory variables. The null hypotheses (they are not associated) were rejected for all outcome measures. Therefore, FE regressions were favored against RE regression models and used within this analysis (see Additional file 2).
By removing all time-constant factors, FE estimates are based solely on changes within individuals over time (intra-individual changes). As a consequence, the FE estimator is not biased by time-constant unobserved heterogeneity. It is also called “within-estimator” and enables to estimate causal effects by comparing intra-individual changes (with certain restrictions). Since all between-unit variation is eliminated, time-constant variables cannot be estimated in FE regression analysis. Nevertheless, these variables can be used for descriptive purposes and in terms of moderator variables in sensitivity analysis [35].
To estimate the predictors of GP and specialist visits, we conducted a FE Poisson regression, which is a commonly used model for measuring count-data [34]. As suggested by Stock and Watson [37], cluster-robust standard errors (SE) were used in order to avoid substantial underestimation of the true standard errors. To estimate the predictors of the binary outcome variable hospitalization, a conditional FE logistic regression was applied [38].
For sensitivity analysis, the main model was extended by including an interaction term composed of level of educational status and volunteering (volunteering x education). The underlying idea was that the impact of volunteering on the use of health care services might differ by educational level [39]. As there is evidence that in Germany formal volunteering is more likely performed by men [40], we further tested whether gender-specific links between volunteering and the outcome variables exist. Therefore, we included a respective interaction term (volunteering x gender).
The proportion of missing values was smaller than 3% for all explanatory variables, except for the variable income which had less than 6% missing values. To achieve statistical significance, explanatory variables need to reach a p-value smaller than 0.05. Statistical analyses were conducted using Stata 14 [41].