Data sources and study design
This current study was a secondary analysis of the China Health and Retirement Longitudinal Study (CHARLS), which is a national prospective cohort collecting a wide range of social and economic data, personal health information for geriatric and health policy research (http://charls.pku.edu.cn/). A total of 17 705 participants from 150 counties or districts within 28 provinces in China were recruited in the demographic background survey at baseline (visit 1: 2011-2012), and followed up every two years at visit 2 (2013-2014), visit 3 (2015-2016), and visit 4 (2017-2018). Details of the cohort design have been described previously16. The CHARLS study was approved by the institutional review board of Peking University (IRB00001052-11015). Written informed consent was obtained before participation. At each visit, the trained staff conducted face-to-face interviews to collect the sociodemographic characteristics, medical history, health behavior, cognitive function and depressive status using standardized questionnaire.
Of 17 705 participants at baseline, we excluded 832 individuals without age and sex information, or younger than 45 years. Then, 5 970 and 6 733 participants lacking cognitive function data or PA intensity information were excluded, as the activity investigation was limited to a randomly selected subgroup in the CHARLS study. We excluded 47 individuals of Alzheimer's disease, brain atrophy or Parkinson's disease, and 84 individuals with history of stroke attack. Finally, 4 039 participants were included for this current analysis, and 2 319, 2 184, and 1 557 of them provided cognitive function data during follow up at visit 2, visit 3, and visit 4, respectively. The flow chart of this study was shown in Figure 1.
Cognition measurement
In accordance with previous studies17-19, the cognitive performance was tested by two cognition measures in this study: episodic memory and mental intactness. The episodic memory reflects an individual’s ability to immediately repeat ten Chinese words just read to them in any order (referred as immediate recall) and to recall the same words four minutes later (referred as delayed recall). The episodic memory score is the mean value of immediate and delayed recall scores, and ranges from 0 to 10. The mental intactness reflects the mental status based on several questions of the Telephone Interview of Cognitive Status (TICS) battery, including serial subtraction of 7 from 100 (up to five times), the date (month, day, year and season), the day of week, and the ability to redraw a picture shown to the individual. Answers to these questions are summed into the mental intactness score ranging from 0 to 11. The global cognition score is calculated as the sum of episodic memory score and mental intactness score, which ranges from 0 to 21.
Definition of PA intensity
PA was quantified via a modified short form of the physical activity questionnaire 20. Participants were asked to report the number of days and typical spent time-per-day for three activity types (mild, moderate and vigorous) during the previous week. Participant were classified into ‘none’, ‘mild’, ‘moderate’ and ‘vigorous’ PA intensity, according to whether they had the corresponding activity type on three or more days per week and for at least 10 minutes at every time. The vigorous activities make breathe much harder than normal and include heavy lifting, digging, plowing, aerobics, fast bicycling, and cycling with a heavy load; the moderate activities make you breathe somewhat harder than normal and include carrying light loads, bicycling at a regular pace, mopping the floor, doing Taiji and walking fast; spend walking in a usual week; the mild activity refers to waking at work, at home, walking to travel from place to place, and any other walking that you might do solely for recreation, sport, exercise, or leisure. Participants without any of these activity types or not meeting the relevant standards were classified as ‘none’ PA intensity.
Covariates
Baseline measurements of age, sex, education level, marital status, residence location, BMI, smoking, drinking, self-reported health conditions and medication use, and depression status were included as covariates in the current study. Educational level was categorized as ‘primary education’, ‘secondary education’, and ‘third education’. Marital status included ‘married’ and ‘others’. Residence location included ‘urban’ and ‘rural’. Smoking status was defined as ‘never smoking’, ‘current smoker’, and ‘former smoker’. Drinking status was defined as ‘current drinking more than once per month’, ‘current drinking once or less than once per month’, and ‘no current drinking’. Self-reported health conditions included the diagnosis of hypertension, diabetes, dyslipidemia, and heart diseases (heart attack, coronary heart disease, angina, congestive heart failure, or other heart problems). Medication use included the use of anti-hypertensive and anti-diabetic drugs. Depressive symptoms were assessed using the 10-item version of the Epidemiologic Studies Depression Scale (CES-D), and a score of ≥12 indicated the presence of depressive symptoms21.
BMI was calculated as weight (in kilograms)/height^2 (in metre squared), and and grouped into underweight (BMI: <18.5 kg/m2), normal weight (BMI: 18.5-23.9 kg/m2), overweight (BMI: ≥24.0 kg/m2) according to the overweight and obesity standard22.
Statistics analysis
Baseline characteristics are presented as the mean (standard deviation, SD), median [interquartile range, IQR] or number (percentage), as appropriate. To show the distribution differences, we compared the scores of global cognition, episodic memory and mental intactness among those with different PA intensity by Kruskal-Wallis test, both with the population without PA and the whole population as reference.
To investigate the association of PA intensity and cognitive performance at baseline, we performed adjusted analyses using the regression models for the individual-level factors: model 1 was adjusted for age; model 2 was adjusted for age, sex, BMI, education level, marital status, residence location, health conditions, smoking, current drinking and CES-D score. To validate the findings, we did multiple sensitivity analyses. First, 37.5% (1 515 of 4 039) of total data items were missing at residence type (n of missing: 1 397), health conditions (n of missing: 16 for hypertension, 63 for dyslipidemia, 26 for diabetes, 21 for heart diseases) or CES-D assessment (n of missing: 103). Thus, we repeated the analyses using the imputed data by multiple imputation of chained equations method. We created five imputed data sets and pooled the results. Second, to account for the imbalanced self-select probability into different PA intensity group, we applied a multinomial propensity score weighting procedure using the tree-based regression model. The covariate set of age, sex, BMI, education, marital status, residence and CES-D score, which were significantly associated with cognitive performance in the analyses, were considered in this weighting procedure. The ‘es.mean’ and ‘ks.mean’ parameters were chosen as the stopping rule, which refer to the mean of absolute standardized mean difference and the mean of Kolmogorov-Smirnov statistic for measuring the balance across the covariates. The max number of trees was set as 10000 iterations. The balance measures of interest corresponding with iterations were shown in Figure S1 to ensure that the parameter was reasonable. Then, we analyzed the effect of PA intensity in certain subgroups, according to age (<60 years and ≥60 years), sex (male and female), smoking, drinking, BMI level and depression status (CES-D <12 and CES-D ≥12).
To investigate the long-term effect of PA intensity on cognitive performance during follow up, we analyzed the association of PA intensity at baseline with multiple measures of cognition scores using mixed effect model. The main effect of PA intensity and visit time, together with the interaction term, were fitted in the mixed model. Mixed effect model offers a better way to deal with missing data, and subjects with more missing values have a weaker impact on parameter estimation. The same covariates were adjusted as in model 2, and the individual difference was considered as random effect term in the analyses. Furthermore, the subgroup analyses were performed as mentioned above. All the analyses presented above were conducted using packages of ‘mice’, ‘twang’, ‘survey’, and ‘lmerTest’ by R software (version 4.1.0).
Data Availability
The CHARLS dataset is freely available to all researchers in related fields on request. Researchers can gain access to the data (http://charls.pku.edu.cn/). And the datasets used and/or analysed in this current study are available from the corresponding author (Dr. Lixin Tao) on reasonable request.