Study Sample
This study used data from three waves of the China Health and Retirement Longitudinal Study (CHARLS 2011–2015), whose design was based on the Health and Retirement Study (HRS) in the US. The CHARLS comprised a nationally representative sample of adults in China aged ≥45 years. The CHARLS sample was obtained using four-stage stratified sampling with the probability-proportional-to-size (PPS) technique [20]. The baseline survey covered 28 provinces, 150 counties/districts and 17,708 respondents from 10,257 households and was conducted between June 2011 and March 2012. Two follow-up interviews were conducted in 2013 and 2015. We restricted our sample to 3,876 respondents who met the following criteria: (1) aged ≥60 years at baseline, (2) completion of all three study waves (2011–2015), (3) no history of diseases with potentially strong effects on cognitive function (e.g., cancer, stroke, memory-related disease) at baseline and (4) rural hukou status at baseline.
Measures
Migrants and Non-migrants
We divided our sample into two groups, migrants and non-migrants, based on the hukou system, which was used to classify rural and urban residents in previous studies [21, 22]. Older adults (≥60 years) with rural hukou status who resided in urban areas during all three study waves (n = 850) were defined as rural-to-urban elderly migrants. Non-migrants were defined as respondents with rural hukou status who resided in rural areas during the three study waves (n = 3,026).
Cognitive Function
Cognitive function was measured using an adapted Chinese version of the Mini-Mental Status Examination (MMSE), which included similar concepts to those used to measure cognitive function in the US Health and Retirement Study (HRS) [23]. According to previous publications [24-26], we divided cognitive function into two dimensions: episodic memory and mental status. We generated an episodic memory score (range: 0–10) as the average of the immediate and delayed recall scores. The mental status score (range: 0–11) was based on the following three items: figure drawing, serial subtraction of 7 from 100 (up to five times) and the ability to identify the date (month, day, year), day of the week and season of the year. The total cognition score, which incorporated both dimensions, ranged from 0 to 21. A higher score indicated better cognitive function.
Psychosocial Factors
Data on the psychosocial factors were obtained from the baseline survey. The psychosocial factors comprised the living arrangement, social attachment and depression. Living arrangement was determined by whether the respondent was coupled, lived with his/her adult children and had provided any care to his/her grandchildren. According to the CHARLS code book, if the respondent reported that he/she co-resided with any adult child, regardless of whether he/she took care of grandchildren, the respondent was classified as ‘lives with children’. Caring for grandchildren was defined as the provision of care to any grandchildren younger than 16 years during the past year, regardless whether the respondent lived with grandchildren.
The measure of social attachment was adapted from the definition provided in the English Longitudinal Study of Aging (ELSA). Specifically, that study divided social attachment into four domains: civic participation, leisure activities, cultural engagement and social networks. To accommodate the Chinese social background of our subjects, we excluded cultural engagement, which was assessed by the frequency with which the participants reported visiting art galleries, museums or exhibitions and attending theatres, concerts, operas and cinemas, from our analysis. Civic participation was defined as the participation in activities associated with a community-related organisation or in volunteer or charity activities. Subjects who reported that they had participated in one of the above-mentioned activities within 1 month before the interview were classified as having civic participation. Participation in leisure activities was defined as playing mah-jong, cards or chess; visiting a community; attending an athletic, social or other type of club; or attending an educational or training course within 1 month before the interview. The domain of social network was restricted to friendships and was defined as interactions with friends within 1 month before the interview. Other core social networks experienced by elders were measured under the domain of living arrangement.
Depression was measured using the 10-item Centre for Epidemiologic Studies Depression Scale (CES-D-10). The total scores ranged from 0 to 30, and a higher score indicated more severe depression.
Covariates
The subjects’ demographic characteristics, socioeconomic and health statuses and health behaviours were considered as covariates in our study. In accordance with prior CHARLS studies and the distribution of educational attainment among older Chinese adults, the subjects were classified into four educational levels: illiterate; some primary school (not completed); finished primary school; and higher than primary school [27, 28]. Household income was defined as the sum of all annual income at the household level and was stratified into three levels (low, medium and high) according to the lower and upper quartiles. Retirement status was dichotomised as retired or not retired. Retirement was defined as a history of employment (including agricultural and non-agricultural work) and a current status of no longer working, while non-retirement was defined as participation in current employment (agricultural and non-agricultural) or no history of such work throughout one’s lifetime.
Health status was assessed according to the number of activities of daily living (ADL) in which the subject experienced disability and chronic disease status. ADLs were determined as the number of activities during which the subject experienced difficulties (range: 0–6). Chronic disease status was determined from self-reported diagnoses. According to previous studies, smoking, alcohol consumption and afternoon napping may affect cognitive function in the elderly [9, 23, 29]. Therefore, we considered these three items as health behaviours. The subjects were categorised as non-smokers, light/moderate smokers (<20 cigarettes per day currently or a history of smoking) or heavy smokers (≥20 cigarettes per day currently). They were further categorised into three alcohol consumption categories: non-drinkers, ≤1 drink per month or >1 drink per month. The subjects were further categorised as non-nappers, short nappers (<30 min), moderate nappers (30–90 min) or extended nappers (>90 min) [29].
Statistical Methods
The characteristics of the sample were described according to sex and migrant status. Continuous variables are reported as means and standard deviations, while categorical variables are reported as percentages. The t-test was used to compare normally distributed continuous variables. The chi-square test was used to compare the nominal variables, namely retirement, living arrangement, social attachment and chronic disease, while the rank-sum test was used to compare the ordinal variables of age group, education level, household annual income, smoking, alcohol consumption and afternoon napping.
We examined differences in the subjects’ cognitive function trajectories using multilevel linear regression analyses in which the follow-up wave was set as the first level (low level) and coded as 0, 1 or 2 to represent the longitudinal term, and individuals were set as the second level (high level). We assumed that individuals would have different baseline levels of cognitive function and different rates of cognitive decline. Therefore, we estimated the random coefficient models. First, we detected the difference in cognitive trajectories between migrants and non-migrants and evaluated the presence of a sex-specific difference by establishing a model that included the interacting terms of time, migration status and sex. Because we identified a significant interaction between sex and the migration status with respect to the total cognition and mental status scores (see Additional file 1: Table S1), we stratified all of the analyses by sex.
Given the observed difference in cognitive function between female migrants and non-migrants, we constructed a series of adjustment models to explore the possible underlying factors. In these adjustment models, the psychosocial factors and covariates were entered at level 2 (interpersonal level). Model 1 was adjusted for the age group and time of follow-up. Model 2 comprised model 1 plus the socioeconomic status, while models 3 and 4 added psychosocial factors. Finally, model 5 included the health status and health behaviours. We used multiple imputation by chained equations (MICE) to impute any missing values. This process was performed using R version 3.4.5 with the ‘mice’ package. We also conducted a sensitivity analysis by running models in which the missing values had not been imputed, and achieved similar results. In all of the analyses, statistical significance was based on a two-tailed P value < 0.05. All of the analyses were performed using R software version 3.4.5.