Study sample
Participants were drawn from a large ongoing population-based cohort study, the WHO Study on global AGEing and adult health (SAGE) in Shanghai. Details concerning the SAGE have been previously described [17]. Briefly, SAGE is a longitudinal study on the health and well-being of adults aged 50 and older in six low- and middle-income countries (LMICs): China, Ghana, India, Mexico, Russian and South Africa. In China, the study was constructed including wave 1, implemented in 2009/10, wave 2 in 2014/15 and wave 3 in 2018/19. We enlarged the sample size of SAGE in Shanghai, China to obtain a sub-state representative sample using the same multistage clustered sampling method and survey assessment. In particular, wave 2 served as the baseline and wave 3 as the follow-up in this study, as they contain a more comprehensive set of assessments. A longer follow-up, through December 31, 2021, was additionally conducted to ascertain the participants’ survival status. At baseline (2014/15), 5402 community dwellers aged 50 and older were recruited from five districts of Shanghai, China, and included in the analysis for mortality. After 4 years, 5077 subjects (325 had died) were invited to undergo the follow-up assessment, while 1592 were excluded (1334 did not return, 52 declined, and 206 had an unrecognized disability or hospitalization status); leaving 3485 participants eligible for the analysis for disability and hospitalization. A flow chart of the participant selection process is shown in Fig. 1. Comparisons of the non-responders with respondents in terms of baseline age, sex, and frailty status were conducted (Additional file 1) and results suggested that the issue of the representativeness should not represent a potential bias, although the response rate was 68.6%.
Frailty Scales
The four frailty scales were briefly described below. An overview of all items constructed in each scale can be found in Additional file 2. In particular, to maximize the use of available data, a scale was included in subsequent analyses if no more than 20% of all items were missing [13]; meanwhile, missing items for FP, FRAIL, and TFI were imputed with 0, whereas no substitution procedure was required for FI because of its distinctive derivation method used in this study.
Frailty index (FI). Following a standard procedure [18], we created a 35-item FI, which differed from our previous publication [19] due to the use of different data. Even so, it has been suggested that an index with 30–40 variables is sufficiently accurate for predicting adverse outcomes [18, 20]. The included variables were converted as a certain proportion of the deficit. For each participant, these deficits were summed up and then divided by the total possible deficit to derive the FI, with a value of 0.20 or greater defined as frail.
Frailty phenotype (FP). The FP was constructed using an adapted phenotypic definition based on the criteria of five components proposed by Fried et al. [5]: slowness, weight loss, low grip strength, exhaustion, and low physical activity. It has been previously operationalized in SAGE [21, 22], and the same criteria were applied in this study. Likewise, participants were classified as frail if 3 or more criteria were present.
FRAIL scale. We used an adaption of the FRAIL scale [23], which considers deficits accumulated in five domains: fatigue, resistance, ambulation, illness, and loss of weight. FRAIL has not been explored in SAGE before. Briefly, fatigue, resistance, and ambulation were assessed using separate self-reported questions. Participants were classified as illnesses if they had 5 or more out of 9 self-reported chronic diseases. The weight loss criterion was ascertained based on the lowest quintile of BMI. Individuals with 3 or more deficits were recognized as frail.
Tilburg Frailty Indicator (TFI). The TFI, developed as an integral conceptual model of frailty, comprises two subscales [24]. One subscale addresses the determinants of frailty such as socio-demographics, the latter addresses the level of frailty across physical (8 items), psychological (4 items) and social domains (3 items), and is used in this study, yet it has not previously been explored in SAGE. Items were assessed using self-reported questions as well as several functional tests. Theoretical scores of the TFI range from 0 to 15, with a score of 5 or greater defining frailty [25].
Outcome Measures
Outcome measures were disability, hospitalization at 4 years, and 4- and 7-year all-cause mortality.
Disability was assessed both during 2014/15 and 2018/19 using eight activities of daily living (ADL) tasks (moving around, bathing, dressing, maintaining appearance, getting up from lying down, eating, toileting and controlling urine) [26]. For each ADL task, participants were asked, “Do you have difficulty in” performing the task in the preceding 30 days? The response was in a Likert scale format ranging from “none” to “extreme/cannot”. Respondents were considered as disability if they reported severe or extreme difficulties in performing at least one of the eight tasks listed above; then, the onset of new disability was defined as a newly identified disability during 2018/19. Spending overnight in hospital at least once, ascertained by asking “have you been hospitalized in the prior four years?” during 2018/2019, was regarded as hospitalization. 4- and 7-year all-cause mortality was determined by linking data to the Shanghai Death Registry during 2018/2019 and on December 31, 2021, respectively.
Covariates
Using the literature on disability, hospitalization, and mortality in older adults as a guide, commonly cited risk factors were selected as potential covariates and then identified in the dataset. Hence, covariates include age, sex, marital status (partnered [married/cohabiting], not partnered [separated/divorced/widowed or never married]), educational level achieved (no education, less than primary, primary, secondary or higher), smoking status (never smoked, current smoker or former smoker) and body mass index (BMI). Measured height and weight were used to calculate a standard BMI (calculated as weight in kilograms divided by height in meters squared).
Statistical analysis
Descriptive statistics were presented as either mean (standard deviation) or n (%), with comparisons between four different outcome groups using t-test/Wilcoxon rank-sum test or chi-square test, as appropriate. Logistic regression models were measured to investigate the association of dichotomized frailty status (frail, non-frail (reference)) identified by each scale with adverse outcomes, with results reported as odds ratios (ORs) and 95% confidence intervals (CIs). All regression models were performed unadjusted and then adjusted for the same covariates above (fixed model). For each outcome, a receiver operator characteristic (ROC) curve based on the continuous scores of each scale was created and the area under the curve (AUC) was calculated with their corresponding 95% CIs to assess the unadjusted predictive ability. AUCs were compared using Wilcoxon tests. Frailty prevalence, sensitivity, and specificity for each scale and each outcome were also calculated using the proposed cutoffs, and those above or below the proposed values. We used the following acceptable minimum thresholds: ≥0.60 for AUC [27], ≥ 0.8 for sensitivity [28] and ≥ 0.6 for specificity [28]. Statistical analyses were performed using the SAS software (version 9.4, SAS Institute, Inc., Cary, NC), and a 2-sided p < 0.05 was considered statistically significant.