Physical activity and sedentary behavior trajectories and their associations with quality of life, disability, and all-cause mortality

Data comes from the three waves of the World Health Organization (WHO) Study on global AGEing and adult health (SAGE) in Mexico. SAGE is a multi-country, longitudinal study, based on nationally representative samples of individuals aged ≥50y. It has been conducted in six countries -- China, Ghana, India, Mexico, Russia, and South Africa – which each bring different geographic distributions, population sizes, income levels (low and medium), and phases in the epidemiological transition. To date, SAGE has three longitudinal measurements in Mexico, details of the study design have been published elsewhere [13]. Briefly, Wave 1 (baseline) was collected between July and September 2009, with a total sample of 2404 respondents ≥50y. Wave 2 data was collected between July and October 2014, with a refreshed sample of 618 new interviews (also ≥50) additional to the remained sample of Wave 1; and Wave 3 from August to November 2017 with 2318 participants (including 610 new interviews). In total, 3277 individuals were interviewed in the three waves. Given that the aim of this study was to identify longitudinal trajectories of the PA and SB, we included participants with at least two measurements. The final analytical sample had 3209 subjects. The response rate (the proportion of those initially invited, i.e., those with baseline measurement) was 76% (Fig. S1 in the Additional file, Appendix 1). We identified baseline differences between the final sample and excluded participants in several analytical variables -- the latter were older and had a higher prevalence of frailty and multimorbidity (p < 0.05).

For all-cause mortality data analysis, sample definition proceeded as follows. The individuals were included if they had two of three complete measurements (censored observations) or had measurements for waves 1 and 2, and their death occurred between waves 2 and 3. Additionally, and given that our cohort includes incorporating new individuals (with rolling admissions), older adults with measurements in waves 2 and 3 were considered with delayed entry. Table 1 shows the different settings and their sample size. Hence, this analysis included a total of 2564 older adults.

Covariates were categorized as follows: sex (1 = female), age, and number of years of formal education. Socioeconomic status (SES) of the household was derived using the WHO standard approach to estimate permanent income from household ownership of durable goods, dwelling characteristics (type of floors, walls, and cooking stove), and access to services such as water, sanitation, and electricity [21]. SES was included as a continuous variable, with higher values indicating better SES. Multimorbidity was included as a dichotomous variable (with/without multimorbidity) and was defined as the presence of two or more chronic non-communicable conditions from the list of nine chronic diseases included in the SAGE study. The operational definitions of these diseases have been published elsewhere [22]. Frailty status was determined using a modified frailty phenotype, based on the criteria proposed by Fried et al. [23], which includes five components: weight loss, exhaustion, low physical activity, slow walking speed and weakness. Details of the previous application of this frailty measurement in the SAGE sample has been published elsewhere [24]. We also used sarcopenia as an additional covariate. In accordance with previous publications, we defined the presence of sarcopenia as having low skeletal muscle mass (reflected by lower skeletal muscle mass index) and either a slow gait speed or weak handgrip strength. Details of the specific algorithms used to define sarcopenia status in the older adult population using the SAGE study sample are published elsewhere [25]. Finally, body mass index (BMI) was calculated using weight (kg) and height (cm) (BMI = Weight [kg] / Height [m2]) and was incorporated into the analysis as a continuous variable.

Baseline characteristics are presented in percentages and means (standard deviation) as appropriate. Health and sociodemographic characteristics related to longitudinal trajectories of PA were compared using Chi-square or ANOVA tests.

We used growth mixture modelling (GMM) to investigate the longitudinal trajectories of PA and SB [26]. GMM is useful since it provides information regarding the growth factors of each different trajectory. The intercept and slope (growth factors) are interpreted as usual in longitudinal modelling: the level of outcome variable when time is equal to zero and the rate of change in the outcome over time, respectively. According to current recommendations [27], we initially specified a single-class latent growth curve model to determine the pattern of change over time. Given the number of available measurements in the SAGE study (i.e., three waves) we examined a linear and a quadratic pattern of change. We specifically applied the GMM with class-specific random intercepts. We also applied an exploratory approach and fitted models with an increasing number of classes to identify the optimal latent class model.

We determined the best model (i.e., the one with the optimal number of classes) based on statistical criteria, parsimony, and interpretability [27]. We considered: (1) the lowest values of the goodness of fit measures - Bayes Information Criteria (BIC), Akaike Information Criteria (AIC) and the sample-size adjusted BIC (aBIC), (2) the following versions of the likelihood ratio tests (LRT): Vuong-Lo-Mendell-Rubin, Lo-Mendell-Rubin adjusted, and Bootstrapped, (3) the way to which the trajectory classes captured distinct and important patterns in the data, and (4) the quality of the model in terms of posterior probability diagnostics, namely the entropy and average posterior probability for each trajectory class.

We analyzed the association between the PA and SB trajectories and the distal outcomes (QoL, disability, and all-cause mortality) using linear mixed-effects regression and Cox proportional hazards models. Given that the Mexico-SAGE study included multiple members from within the same household and had repeated measurements of QoL and disability from the same individual, our data had a three-level hierarchical structure, with measurement occasions at level 1, individuals at level 2, and households at level 3. We then fitted a random intercept models including the subject and household IDs as random effects for QoL and disability outcomes.

For the Cox model, new admissions are left-truncated observations (also known as delayed entry), implying that not all individuals start to be at risk simultaneously, which in turn represents a potential bias [28‐30]. Then, we considered the delayed entry feature in our statistical analysis following the proposal of Lamarca et al. [31] to analyze left-truncated data with the older adult population using age as the time scale. In our study, consider the survival time as the elapsed time from age 50 until the event of interest. Additionally, we used clustered standard errors to account for correlation between repeated measurements within individuals.

All models were adjusted using the covariates described above. Furthermore, the modeling of PA was adjusted by weekly hours of SB, and SB was adjusted by PA (METs/hours per week), and both (PA and SB) for the follow-up time. Regression coefficients, hazard ratios, and 95% confidence intervals were reported.

Models for GMM were estimated in Mplus v8.5 by full maximum likelihood (FML) and robust standard errors to non-normality [32]. To avoid local maxima for the expectation-maximization (EM) algorithm, we estimated the models with 200 random starting values and 100 iterations per set of starting values. According to the guidelines for reporting on latent trajectory studies [27], the code syntax is provided in the Additional file, Appendix 2. We used Stata 17.0 to model the association between PA and SB trajectories and distal outcomes.

This study was conducted following the STROBE guidelines for reporting cohort studies (STROBE checklist is reported in the Additional file, Appendix 3).

The single-class latent growth curve model showed favorable evidence for the linear model (p < 0.01) over the quadratic model (p = 0.99) for both PA and SB. We adjusted models from one to four trajectories. Table 2 provides the information of the Model Selection Criteria for all the models tested. For PA and SB, the three-class model was the best according to the fit indices, entropy and the p-values associated with the three LRT used (Vuong-Lo-Mendell-Rubin, Lo-Mendell-Rubin adjusted, and Bootstrapped). Additionally, we selected the three-class model based on the lower values of AIC, BIC, and aBIC.

Figures 1 and 2 show the trajectories of PA and SB in the three-class model. For PA, the first class was identified as “low-PA-decreasers,” with the lowest baseline PA and the least steep decreasing trajectory. There were 1244 individuals in this class (39% of the sample) with an average baseline PA of 41.2 (SE = 0.5, p-value< 0.01) and a decline rate of − 3.1 (SE = 0.1, p-value< 0.01). The second class, “moderate-PA-decreasers,” had a medium baseline PA and a decreasing trajectory. This class had 593 individuals (18% of the sample) with moderate level of PA in baseline (intercept = 67.8, SE = 0.5, p-value< 0.01) and steeper decline (slope = − 4.4, SE = 0.1, p-value< 0.01). The third class, “high-PA-decreasers,” had a high baseline PA and a steep decreasing trajectory. This class had 1372 participants (43%) with a higher average baseline PA (intercept = 97.6, SE = 0.9, p-value< 0.01) but also with the steepest decreasing slope (− 5.3, SE = 0.1, p-value< 0.01) (Table 3).

For SB, we also identified three classes. The first group, “low-maintainers”, was composed of 2888 OA (90%), with the lowest baseline levels of SB (intercept = 2.1, SE = 0.1, p-value< 0.01) and a relatively stable trajectory (slope = 0.1, SE = 0.01, p-value< 0.01). The second group, “steep-decreasers”, had 64 individuals (2%), with the highest baseline SB (intercept = 13.1, SE = 0.6, p-value< 0.01) and a steep decreasing trajectory (slope = − 1.3, SE = 0.1, p-value< 0.01). Finally, the third group, “steep-increasers”, had 257 OA (8%), medium baseline SB levels (intercept = 4.9, SE = 0.3, p-value< 0.01), and a steeper increasing trajectory (slope = 0.8, SE = 0.1, p-value< 0.01) (Table 3).

Table 4 shows baseline health and sociodemographic characteristics by PA trajectories. In comparison to individuals in classes 1 (low baseline PA and less steep decreasing trajectory) and 2 (medium baseline PA and decreasing trajectory), older adults in class 3 (high baseline PA and steep decreasing trajectory), were younger (p-value< 0.01), mostly male (p-value< 0.01), had more years of formal education (p-value< 0.01) and better SES (p-value< 0.01). They also displayed a significantly lower prevalence of sarcopenia (p-value< 0.01) and multimorbidity (p-value< 0.01) and had the lowest levels of BMI (p-value< 0.01). Regarding SB trajectories, individuals with the worst trajectories (classes 2 and 3) had unfavorable health conditions than those in class 1. The former had a lower quality of life (p < 0.01), greater disability (p < 0.01), and a higher mortality rate (p < 0.01). Additionally, they had higher prevalences of multimorbidity (p < 0.01), frailty (p < 0.01), and sarcopenia (p < 0.01) (Additional file, Appendix 3).

The associations of PA and SB trajectories with the distal outcomes QoL, disability and all-cause mortality are depicted in Table 5. Regarding QoL, the classes with higher baseline PA levels (moderate-decreasers and high-decreasers) had the higher QoL levels (β = 2.90; 95% CI: 1.78;4.01; and β = 2.81; 95% CI: 1.64;3.98, respectively). Meanwhile, the group with the best SB trajectory (low-maintainers) also had the higher levels of QoL compared to the group of steep-increasers (β = − 3.70; 95% CI: − 5.31;-2.09).

The groups of moderate-PA-decreasers and high-PA-decreasers also had lower disability scores compared to the low-PA-decreasers group (β = − 6.11; 95% CI: − 7.39;-4.82; and β = − 7.60; 95% CI: − 8.95;-6.25, respectively). For SB, the steep-decreasers and steep-increasers groups showed worse levels of disability (β = 5.81; 95% CI: 1.94;9.69; and β = 8.81; 95% CI: 6.92;10.70, respectively) in comparison with the low-maintainers group.

For both outcomes (QoL and disability) and exposures (PA and SB) the results show that between-subject differences explain a greater proportion of the variance associated with the outcomes than between-household differences. Specifically, 33% of the variation in QoL is explained by individual differences (ICC = 0.33; 95% CI: 0.30–0.36) compared to 14% by household differences (ICC = 0.14; 95% CI: 0.09–0.20). Meanwhile, for disability, the observed data were 43% for between-subject differences (ICC = 0.43; 95% CI: 0.40–0.46), and 6% for between-household differences (ICC = 0.06; 95% CI: 0.02–0.16).

For all-cause mortality data, the median duration of follow-up was 2002 days (5.5 years), with an interquartile range of 1196 days. In total, 368 deaths were observed, equivalent to a mortality rate of 26.1 per 1000 person-years. Regarding the observed associations, the high-PA-decreasers group had a lower risk of dying than the low-PA-decreasers group (HR = 0.33; 95% CI: 0.24;0.44). Finally, the steep-increasers group in SB had a higher mortality risk than the low-maintainers group (HR = 1.44; 95% CI: 1.05;1.96).

Our study has some strengths. To the best of our knowledge, this is the first study conducted in LMIC to assess the longitudinal trajectories of PA and SB, and their associations with QoL, disability, and all-cause mortality among OA population. Additionally, we used several outcomes that could be highly significant for this population group. There were also some limitations to our study. First, we used self-reported PA and SB data instead of objective measurements, which could have led to an overestimation of PA and underestimation of SB. In particular, the reported time of SB (two hours on average) seems low given that some studies have reported higher levels of SB among older adults [55]. However, previous studies with cross-sectional data (including the six countries in the SAGE study) have reported similar levels in the average number of SB daily hours [56, 57]. Additionally, it is important to note that our SB measurement is based on a single question investigating the time that older adults spend in a sitting position (excluding sleep time). Other SB measurements are based on complete batteries/questionnaires that include different activities not considered in the SAGE study [58, 59]. This bias could cause some older adults to be miscategorized in some of the estimated trajectories. Second, recall and survivor bias can affect the results of epidemiological studies with older adults. It is possible that participants in our study with worse disability and QoL died at younger ages. Therefore, the healthiest individuals would be interviewed, and the associations may have been underestimated. Third, losses to follow-up between waves (15% after Wave 1 and 17% after Wave 2) seem high. Although, the SAGE study has added new individuals in each round to compensate for these losses. Still, when comparing the excluded individuals with those in the analytical sample, the former were older and less healthy. Consequently, the observed associations could be underestimated since younger and healthier individuals tend to have a better quality of life, less disability, and lower mortality rates.