SHARE is an (approximately) biennial survey focusing on people older than 50 years from several European countries. It uses computer-assisted personal interviews (CAPI), probability sampling, and the participation rates (for Wave 1) ranged from 38–74%, depending on the country (8). It has been reviewed by an ethics committee and all participants have given informed consent. We refer the reader to Börsch-Supan et al. (9) for more details on SHARE.

The study uses data from Waves one (2004) and two (2007). The only exclusion criterion is missing values for the outcome (GP advice for exercise frequency). The final sample consists of N=21,703 participants from 14 European countries including: Austria (5.90%), Germany (9.93%), Sweden (6.47%), Netherlands (7.90%), Spain (7.14%), Italy (9.04%), France (7.98%), Denmark (6.28%), Greece (4.73%), Switzerland (5.98%), Belgium (11.11%), Czech Republic (7.36%), Poland (6.85%), and Ireland (3.33%).

The frequency of advice for exercise was measured with the following question: “How often does your general practitioner tell you that you should get regular exercise?” (Never; At some visits; At every visit). This question appears only in Waves 1 and 2 in the form of a paper pencil drop-off questionnaire.

The covariates introduced aimed at adjusting for medical, demographic, behavioural, and socioeconomic factors that may influence the GP’s decision to advise (3,4,5,10). Medical covariates included: for physical health, 14 indicators of previous diagnoses, namely, heart attack (including myocardial infarction or coronary thrombosis or any other heart problem including congestive heart failure), high blood pressure or hypertension, high blood cholesterol, diabetes or high blood sugar, arthritis, osteoporosis, Parkinson disease, hip fracture or femoral fracture, stroke, chronic lung disease, asthma, cancer, stomach or duodenal ulcer, and cataracts; for mental health the EURO-D geriatric depression scale (11); for disability, the number of mobility limitations, arm function, and fine motor limitations. Lastly, the categorical Body Mass Index (BMI) was also included.

Demographic covariates (other than gender) included: age, country, employment status (retired, employed or self-employed, unemployed, permanently sick or disabled, homemaker, other), and marital status (married and living together with spouse, registered partnership, married living separated from spouse, never married, divorced, widowed). Behavioural covariates included: Smoking (yes, currently smoke; never smoked daily for at least one year; no, I have stopped) and alcohol consumption in the last six months (≤2 times a month, 1–4 days a week, ≤5 days a week). Socioeconomic covariates included: education (<=primary, <= upper secondary, tertiary) and financial distress (with great difficulty, with some difficulty, fairly easily, easily).

Characteristics between genders were analysed with standardised differences (std.diff.) using the R package “stddiff”. Ordered logistic regressions were used to obtain odds ratios (OR) and 95% compatibility intervals (CI). Additionally, since OR can be misleading regarding the effect on the probability scale, average marginal effects, that is, the average difference (between genders) in probability of belonging to an outcome category, were calculated.

Six modes were employed. Model 1 is the unadjusted model; Model 2 adjusts for the medical covariates; Model 3 additionally adjusts for behavioural risks (smoking, drinking); Model 4 additionally adjusts for the demographic factors (excluding country); Model 5 additionally controls for socioeconomic factors; and Model 6 additionally controls for country in order to provide the no-pooling estimate.

The few missing values (<2%), mainly on BMI, EURO-D, and financial distress, were imputed using SHARE’s own imputations. Continuous variables (i.e., age, disability, depression) were modelled with a quadratic term. Cluster robust standard errors were used at the household level. The α=0.05 level was used for statistical significance.