Xuxi Zhang

Summary (PCS) and Mental Component Summary (MCS), both ranging from 0 (lowest) to 100  (highest level of health). 18, 22 Covariates  Various socio‐demographic characteristics were assessed at baseline and incorporated as  covariates 23, 24 , including age (in years), sex and country. Education level concerned the highest  level of education the participant completed and was categorized according to the 2011  International Standard Classification of Education (ISCED) into primary or less (ISCED 0‐1),  secondary or equivalent (ISCED 2‐5) and tertiary or higher (ISCED 6‐8). Living situation was  categorized into living with others (‘with partner, no child’, ‘with partner and children’,  ‘without partner, with children’ or ‘in a household shared with others’) or not living with  others. With respect to life style, three aspects were measured. Firstly, three items of the  AUDIT‐C measured high‐risk alcohol use on a scale ranging from 0 (lowest risk) to 12 (highest  risk) 25 . A score of 4 or more in men and a score of 3 or more in women indicate hazardous  drinking or active alcohol use disorders. 25  Secondly, one item on exercise assessed the  frequency of a person being engaged in activities that require low or moderate energy (once  a week or less versus more than once a week). Thirdly, one item on smoking assessed whether  a person smoked. Finally, multi‐morbidity was defined as having at least two of 14 common  chronic conditions 26 , including heart attack, hypertension, diabetes, stroke, high blood  cholesterol, asthma, arthritis, osteoporosis, chronic lung disease, cancer or malignant tumor,  stomach or duodenal ulcer, Parkinson’s disease, cataract and hip or femoral fracture. 27 Statistical analyses  In order to examine mean differences in PCS and MCS scores between frail and not frail groups,  effect sizes were estimated by dividing the difference in mean scores between subgroups by  the largest SD. Cohen’s effect sizes (d) were used for the interpretation of relevant differences:  0.20 ≤ d <0.50 was considered a small difference; 0.50 ≤ d < 0.80 was considered a moderate  difference; d ≥ 0.80 was considered a large difference. 28 To control for the cluster effect of countries we performed multilevel linear regression models  as well as multivariate linear regression models, but found similar results (data not shown).  Hence, we chose three multivariate linear regression models to investigate the independent  contribution of frailty on HRQoL. PCS and MCS scores were included as the dependent variable.  The first model regarded only frailty, physical, psychological or social frailty as determinant  ( crude model ). The second model additionally included the covariates as determinants  ( adjusted model ). To explore the contribution of the three domains of frailty on HRQoL, the  third model included all three domains of frailty and the covariates as determinants ( full  model ). Regression diagnostics included tests for linearity between the determinants and  dependent variables and tests for normality of residuals with kernel density plots. Variance  inflation factors were adopted for tests of multicollinearity. No violation of basic assumptions  for regression and no multicollinearity problems were found.  Finally, we assessed interactions between frailty as well as three domains of frailty and socio‐ 2 31 Association between frailty and HRQoL