Socioeconomic Factors and Memory Decline

The analytical subsamples included only data from core members aged 50 and older in each study at their baseline wave (see Supplementary Figures S1 and S2 for the corresponding flow charts). The associations between each socioeconomic marker and memory decline over up to 8-year follow-up from Waves 5 to 9 in ELSA and Waves 1 to 4 in CHARLS were examined by employing a coordinated analysis of linear mixed models (maximum likelihood estimation, unstructured covariance), which accounts for between- and within-subjects variability across repeated measures, taking into consideration that the same individuals’ measures are correlated. A “time” variable was generated to represent the follow-up from Waves 5 to 9 in ELSA and between Waves 1 and 4 in CHARLS. The time in the study has been created for each study to account for the period between waves, and every unit indicates a 1-year increase in follow-up time (range from 0 to 8 years). Memory change was modeled as a linear function of time measured from the baseline wave until the end of the study period. Random effects for the intercept and slope were fitted for each individual, allowing participants to have different scores at baseline and rates of change in memory. The slopes were adjusted for the baseline memory, as the rate of decline might strongly depend on this. Independent analyses were conducted for each marker of SES and memory change over time within each cohort. To test whether memory trajectories differed between participants, we included in the model the SES marker, covariates, time, baseline memory, time × SES marker, and time × covariates. Unstandardized coefficients and 95% confidence intervals (CIs) for baseline memory (intercept) and linear change (slope) were presented for each of the two cohorts from fully adjusted models. Missing observations were assumed to be missing at random (Little & Rubin, 2002 ), and model assumptions were verified by examining residuals computed from the predicted values. Both linear and quadratic effects were tested, but the linear model showed a better fit based on the Bayesian Information Criteria (Raftery, 1986 (link); Raftery et al., 2012 (link)). Baseline cross-sectional sample weights were used for each cohort analysis to ensure that the sample is representative of the general population. Four supplementary analyses were conducted. In the first analysis, we retested the association for each SES marker while we mutually adjusted for the other two markers. The second analysis presented a sex-stratified investigation for education and urbanicity because these two factors showed a significant sex interaction in CHARLS. The third supplementary analysis examined the rates of memory decline in three subset populations samples matched for baseline memory: subset sample 1 (baseline memory scores <9), subset sample 2 (baseline memory scores 10–12), subset sample 3 (baseline memory 13+). The fourth supplementary analysis explored a more detailed categorization of education within each cohort to better understand the country-level differences in each country’s educational system and its relationship with memory change (see Supplementary Material). All analyses were performed using STATA version 16. The manuscript was written following STROBE guidelines.