Study design and setting
The SAGE was designed as a multi-wave panel study representative of the population aged 50 and older, with a smaller cohort of respondents aged 18–49 for comparative purposes but our study only involved older adults (50 years and older). The SAGE is a crossectional survey conducted in six LMICs. The objective of SAGE was to generate valid, reliable, and comparable information on a range of health and well-being outcomes of public health importance in adult and older adult populations [19]. SAGE wave 1 was conducted between 2007 and 2010 in six LMICs namely: China (2008–2010), Ghana (2008–2009), India (2007–2008), Mexico (2009–2010), the Russian Federation (2007–2010) and South Africa (2007–2008). The sampling design was a multi-stage stratified cluster sampling. The details of the sampling technique have been described elsewhere [19]. Briefly, each country was stratified into mutually exclusive strata. In the first stage of sampling, enumerations areas (EAs), referred to as clusters, were drawn from each stratum using probability proportional to size. In the second stage, households with at least one person aged 50 and older within each EA were randomly selected and interviewed. A smaller sample of adults aged 18–49 years were also selected. Ghana, India, Mexico, and Russia used the Wave 0 (2002-2004) sampling frame and reinterviewed at least 50 percent of the Wave 0 respondents. China used a new sampling frame based on a national health surveillance system, and South Africa did not collect follow-up interviews but used the same Wave 0 sampling frame. The average response rate for the individual in Wave is as follows: China (93%), Ghana (81%), India (68%), Russian Federation (83%), Mexico (53%) and South Africa (75%). The low response rate in Mexico could partly be attributed to the short time available for the fieldwork which did not allow sufficient time for multiple revisits if the respondent was not at home at the initial visit [20].
Data collection procedures
The field data collection was conducted in the six LMICs using a standardized survey instrument. The interviewers and their field supervisors were given a more rigorous and in-depth and training on the content of the questionnaire and translation protocols. The interviews were completed using either a computer-assisted personal interview, paper, and pencil format or both. The SAGE household questionnaire consists of a household roster and modules about the dwelling, income, transfers in and out of the household, assets, and expenditures. The individual questionnaire has modules on sociodemographic factors: age; marital status; education; ethnicity/background; religion; language spoken; area of residence; employment and education of parents; childhood residence, migration, health and its determinants, disability, work history, risk factors, chronic conditions, caregiving, subjective well-being and quality of life, health care utilization and health systems responsiveness, health functioning, chronic conditions, and health care utilization, anthropometric measurements (height, weight, waist and hip circumferences), blood pressure measures and a blood sample via finger prick, and performance tests including near and distant vision, a timed 4-m walk, grip strength, lung function, and cognition. The global positioning system was used to record the location coordinates of every household in the study. The household and the individual data were linked based on unique household and individual identifiers.
Outcome measures
The primary outcome of interest was the proportion of participants who self-reported higher QoL. The participants were asked to rate their overall QoL using five-point Likert scale response categories (very bad=5, bad=4, moderate=3, good=2 and very good=1). The very bad, bad, and moderate quality of life categories were merged as “lower/poorer quality of life” and were coded as 1, whereas the good and very good categories were merged as “higher quality of life” and were coded as 0. This method of re-categorizing the outcome as binary has been used elsewhere [21]. Although DD was the primary outcome variable of interest when investigating social determinants of DD, it was the main exposure of interest when we investigated the causal effect of DD on QoL. The presence of major depressive disorder was based on the International Classification of Diseases, Tenth Revision (ICD 10) diagnostic criteria, and was derived from an algorithm that took into account respondent reporting symptoms during the previous 12 months. Depression was diagnosed when the participants had a minimum of four depressive symptoms after admitting to experiencing at least one of these: depressed mood, loss of interest and enjoyment, and reduced energy leading to increased tiredness and diminished activity as listed in ICD-10 DCR (F32) and lasting most of the day and almost every day for at least two weeks. We included those who have previously been diagnosed with depression (either on treatment or not) to the prevalent cases obtained from the ICD10 diagnostic criteria. We considered covariates including sociodemographic factors, economic instability, education, social cohesion, neighborhood and adult built environment, chronic health conditions, healthcare factors, lifestyle factors, injuries, disability, oral health, and deaths in the household 24 months preceding the survey (Figure 1, and Appendix 1 of the supplementary material).
Statistical analysis
For each country, we summarized the background characteristics of participants using proportions for categorical variables and mean for continuous variables. To prepare the data for analyses, we pooled the six SAGE countries’ data and the standard sampling weight was de-normalized. To de-normalize the adult sampling weight, we divided the standard sampling weight of the adults aged 50+ by the sampling fraction of adults aged 50+. The sampling fraction of adults aged 50+ is the ratio of the total number of adults interviewed in the survey over the total number of adults aged 50+ in the country at the time of the survey. The de-normalization of the sampling weight became necessary because we studied different countries with different population sizes at different time points of the survey. Besides, de-normalization makes it possible to generate an unbiased estimate and draw a valid conclusion of the true impact of DD on QoL in the six LMICs.
The total number of adults aged 50+ interviewed within this period was 35,164 and distributed as follows: China-13,408; Ghana-4,305; India-7,108; Mexico-2,309; Russian Federation-3,763; and South Africa-3,842. Using items response theory (rating scale models), we constructed latent variables such as mobility, self-care, pain and discomfort, cognition, interpersonal activity, sleep and energy, and vision based on some observed characteristics. The Rao-Scott chi-square test (a design-adjusted version of the Pearson chi-square test) and a design-adjusted one-way analysis of variance were used to test the relationship between each covariate and how they vary among the six countries.
To identify factors independently associated with the prevalence of DD, we used the double selection Least Absolute Shrinkage and Selection Operator Poisson regression model (DSLASSOPM). This model was appropriate because it obtained unbiased and efficient estimates by addressing the problems of multicollinearity that arose from a large number of highly correlated predictor variables. Based on the literature and the availability of data or variables in the secondary datasets, we identified 58 variables a priori and were grouped into nine main domains namely: sociodemographic, educational, economic, neighborhood, lifestyle, social cohesion, healthcare, chronic conditions, and injuries/deaths in the last 12 months preceding the survey. To estimate the effect of a particular domain on DD, we adjust for the remaining eight domains. We assumed that the variables selected from DSLASSOPM are the only known independent predictors of DD and therefore applied weighted dominance regression analysis [22] to determine the relative importance of these predictors. Dominance analysis is an ensemble method that ranks the predictors in terms of importance by aggregating results across multiple models. The general dominance statistics were derived as a weighted average marginal/incremental contribution to the overall fit statistic that a predictor variable makes across all models in which the predictor variable was included.
To quantify the effect of DD on QoL, we used inverse probability weighting Poisson regression adjustment model (IPWRA). One of the important characteristics of IPWRA is double robustness (i.e., even if one of the models (exposure or outcome) is misspecified, the estimator is still consistent). To achieve the goal of reducing bias in the differences in covariate distributions between subjects with DD and those without DD, we chose well-matched samples of the older adults with DD and control groups (individuals without DD). The reduction in bias in terms of covariate imbalance between the two groups increases the likelihood of estimating the causal effect of DD on QoL. To satisfy the assumption of ignorable treatment assignment, we included in the matching procedure, all variables are known to be related to both DD and QoL [23, 24]. This assumption states that in the absence of the covariates studied, including DD, there are no unobserved factors that could influence the QoL.
Meta-analysis was used to obtain a single estimate of the effect of DD on QoL from the six LMICs. We achieved this by computing a weighted average of the studies’ individual estimates of DD effect on QoL. Random-effects meta-analysis models using R DerSimonian and NJCct Laird [25] estimation of the random-effects variance was used. This approach incorporates an estimate of between-study variation (heterogeneity) in the weighting. We further applied the Knapp–Hartung adjustment to the overall effect size standard error. All analyses were performed in Stata 16 MP (StataCorp, College Station, Texas, USA) and all estimates were adjusted for complex survey design characteristics (sampling weight, stratification, and clustering).