The Taiwan Longitudinal Study on Aging (TLSA), which began in 1989, was conducted to investigate the impacts of socioeconomic factors on the health and emotional wellbeing of older adults . That study recruited a nationally representative sample of community residents aged ≥50 years, who were then followed-up every 3 to 4 years. This study included participants who had completed the 1999, 2003, and 2007 TLSA surveys (n = 2807) to construct FP/FI trajectories, and excluded respondents who were lost to follow-up or died during this period from final analyses. Detailed information about the TLSA is provided by the Taiwan Health Promotion Administration . The study was reviewed and approved by the Institutional Review Board (IRB) of Taipei Veterans General Hospital (No. 2021-05-023CC) and was conducted in accordance with the principles of the Declaration of Helsinki.
Construction of variables
Due to incomplete TLSA data on walking speed, grip strength and body weight, we used surrogate variables, as other published questionnaire-based studies have done. Exhaustion was defined by the same two questions from the Center for Epidemiologic Studies Depression Scale (CES-D) that the original FP definition used . Handgrip strength and gait speed were assessed as in the Nagi questionnaire , by replies to survey items “difficulty in picking up or twisting using your fingers” and “can you walk 200–300 m?”, respectively, on a four-point Likert scale; respondents who answered “very difficult” and/or “can’t do it at all” were defined as having “weakness” and/or “slowness”. BMI ≤18.5 was designated as the cut-off for defining ‘body weight loss’. Physical activity was measured as the sum of the weighted score calculated from the intensity and frequency of leisure time physical activities; participants engaging in moderate-intensity activity every day, 1–2 times/week, and less than once/week, scored 4, 2, and 0.8, respectively. Those with low-intensity or sedentary activity at the same respective frequencies, scored 2, 1, and 0.4, or 1, 0.5 and 0.2. Men with summed weighted score < 3 or women scoring < 2 were categorized as having low physical activity . Exhaustion, weakness, slowness, low physical activity, and weight loss (BMI ≤18.5) were re-coded by values of 1 or 0, corresponding with presence or absence of each condition, respectively. FP score was calculated as the summary score of these five conditions, ranging from 0 to 5. Supplementary Table 1 lists the original FP definition and the corresponding variables in the 1999, 2003 and 2007 TLSA questionnaires.
We designed a frailty index according to the standard procedure  and held a consensus meeting of geriatricians to decide which deficits to include. Similarly to the frailty index developed by Rockwood et al. , a total of 72 variables were selected; these encompassed: health status and comorbidities (17 items); mobility, activities of daily living (ADL) and instrumental ADL (IADL) (22 items); cognitive function (10 items); psychological status (10 items); stress (4 items) and life satisfaction (6 items); and the sensory domain (3 items). Supplementary Table 2 lists the selected variables and corresponding FI values. All variables were re-coded by values ranging from 0 to 1; 0 or 1 indicated the absence or presence of each deficit respectively, while 0.5 indicated intermediate status. Likewise, variables that were scored on four- or five-point Likert scales were assigned corresponding ordinal values (0, 0.33, 0.66, 1 in four-point Likert scale and 0, 0.25, 0.5, 0.75, 1 in five-point Likert scale), with larger values indicating more severe impairment. An individual’s FI was calculated by dividing the sum of their assessment scores for deficit items by the maximum possible score.
Baseline demographics included age, sex, education, marital status, urbanization of residential area, alcohol consumption (more than once/week), tobacco smoking status (current), difficulty meeting living expenses (answered “some” or “much” difficulty on a four-point Likert scale). Medical history encompassed good self-rated health (answered “very good” or “good” on a five-point Likert scale) and physician-diagnosed morbidities, including the total number documented, and hypertension, diabetes, heart disease, stroke, cancer, chronic lung disease, arthritis, peptic ulcer disease, hepatobiliary disease, hip fracture, cataract, chronic kidney disease (including renal stones) and gout. Mobility assessment included: 1) squatting; 2) standing for 15 min; 3) standing for 2 h; 4) raising both hands over head; 5) grasping objects with fingers; 6) lifting 11–12 kg; 7) running for 20–30 min; 8) walking 200–300 m; 9) climbing 2–3 flights of stairs. ADL was assessed by: 1) taking a bath; 2) dressing; 3) eating; 4) getting up from bed; 5) moving around the house; 6) toileting. IADL was evaluated by: 1) buying personal items; 2) managing money; 3) taking public transportation on one’s own; 4) doing physical work at home; 5) doing light tasks at home; 6) making phone calls. Any difficulty with each item would score 1 point. Summed scores ranged from 0 to 9 points for mobility impairment, and 0 to 6 points each for ADL and IADL impairment. Cognitive function was evaluated using the Short Portable Mental Status Questionnaire (SPMSQ) , which was only performed in participants ≥65 years old in 1999. Owing to discrepancies between the questionnaires administered in 1999, 2003 and 2007, we selected eight questions that were included in all three of these waves, with a higher score indicating better performance. Depressive mood was evaluated with the CES-D 10-item Likert score, in which higher scores (0–30) indicate more depressive symptoms. Sensory assessments included vision, hearing, and oral intake; answering “very poor” or “poor” on a five-point Likert scale was defined as impairment. Social participation was defined as engagement in social activities including religious, political, or trade union groups, voluntary work, or educational classes, etc.
For major time-dependent variables potentially associated with frailty change, including comorbidity, self-rated health, body mass index, mobility impairment, depressive symptoms (defined by CES-D), visual impairment, hearing impairment, oral intake difficulty, meeting living expenses, social participation, cognition (defined by SPMSQ, for age over 65), group-based trajectory modeling was applied to ascertain longitudinal groupings. For example, we identified four trajectories for number of comorbidities: 1) stable low; 2) stable moderate; 3) gradual increase; and 4) stable high. To maintain stability of the regression model given the limited sample size, we pooled the stable moderate, gradual increase, and stable high groups to facilitate comparison with the stable low group. Supplementary Table 3 shows the major time-dependent variable groupings.
Disability was defined as institutionalization or needing a special caregiver at home to assist with ADL in the 1999, 2003 and 2007 TLSA surveys. Participants without disabilities in 1999 or 2003 but who had become disabled by 2007, were classed as having incident disability.
FP and FI trajectories were constructed using group-based trajectory modeling, which assumes that participants represent a mixture of groups that each have distinctive biological trajectories . To select the best model, we followed the procedures recommended by Nagin et al. . To make trajectory models, we determined the number of trajectories, followed by the shapes, by testing the polynomial order including linear and quadratic terms for each group. The Bayesian Information Criterion (BIC) index was adopted to assess models’ goodness of fit. When two models were compared, a value of 2 x ΔBIC (BICcomplex model − BICsimple model) greater than 10 indicated strong evidence favoring the more complex model . Moreover, each trajectory group was required to include greater than 5% of the total study sample. Thus, the best-fitting model was considered to be that with the highest BIC index and in which all trajectories met the pre-specified prevalence criterion. After classifying the trajectory groups, every participant was given posterior probabilities for each group and assigned to that with the highest probability. Nagin et al. proposed that average posterior probabilities should exceed 0.7 for each group, since this is indicative of reliability and good intragroup homogeneity .
To compare baseline characteristics between different FP and FI groups, analyses of variances (ANOVA) was used for continuous variables and X2 or Fisher exact test for categorical variables. P-values for trends were presented individually.
To investigate the major determinants of frailty change, we stratified the study sample into four groups based on FP and FI trajectories: Group 1) favorable FP + favorable FI change; Group 2) favorable FP + unfavorable FI change; Group 3) unfavorable FP + favorable FI change; Group 4) unfavorable FP + unfavorable FI change.
Multinominal logistic regression was applied to investigate relationships between baseline demographics, comorbidities, major time-dependent variables and FP/FI trajectories. In Model 0, we tested each variable by adjusting age, sex, education, baseline FP score and FI, while Model 1 included age, sex, education, baseline FP and FI score, and baseline demographics and comorbidities with p-value < 0.1 in Model 0. Model 2 was further adjusted by major time-dependent variables (Table 2 and Supplementary Table 4).
Binominal logistic regression adjusted by age, sex, education, and change of comorbidities, was used to explore the hypothetical association between FP/FI trajectory groups and incident disability in 2007.
Analyses were performed using SAS, version 9.4 (SAS Institute, Inc., Cary, NC) and SPSS, version 24.0. (Armonk, NY: IBM Corp).