Study Design and Procedures
This study contributes to the validation of a theory-driven framework that guided the implementation of the SMART2D trial (an adaptive implementation trial to improve self-management and to promote a healthy lifestyle among people at risk of or living with T2D in Uganda and South Africa and a feasibility implementation trial in Sweden)[4]. The study used cross-sectional baseline data collected from two rural districts in eastern Uganda, a peri-urban township in the Western Cape in South Africa, and two socio-economically disadvantaged districts of Stockholm in Sweden.
Study Settings
The rural Ugandan population was characterized by a collectivist society with low levels of migration and a high proportion working in agriculture. The urban South African population was characterized by national and international migrant workers with a relatively high unemployment rate. Participants of this setting reported that frequent migration hindered them to build strong community ties. The urban Swedish population consisted of a high proportion of immigrants (approx. 60%) with a diversity in culture and ethnic background living in a society where health and lifestyle are individualized. All three populations were socio-economically disadvantaged in several aspects, but with a sharper socio-economic inequality in the South African setting. More details about the social and built environment, the health system and the population of the study sites can be found elsewhere[4].
Study Participants, Sampling and Recruitment
Study participants were considered eligible if they had resided in one of the study sites for at least six months; were aged 30 – 75 years; had not been previously diagnosed with T2D for longer than 12 months (for the Ugandan and South African site) or 5 years (for the Swedish site); and had a confirmation of prediabetes or diabetes. Pregnancy and serious mental disability were exclusion criteria. In Uganda, 712 participants were recruited by trained field research assistants approaching households in the study area in a random manner. In South Africa, 566 participants were recruited from two community health centers located in the township upon referral by a health care worker. In Sweden, 147 participants were recruited through screening in public spaces and facility-based screening in two primary health centers. Consenting participants were screened through a fasting plasma glucose test in Uganda, a random plasma glucose test in South Africa and the Finnish Diabetes Risk Score (FINDRISC) in Sweden, except for diabetes patients recruited directly from the health facility in the Swedish setting. Confirmation of diabetes or prediabetes was done using a fasting plasma glucose test in Uganda and South Africa and through an HbA1c test in Sweden. More details about the selection criteria and the recruitment process can be found elsewhere for the Ugandan setting[17], the South African setting[20] and the Swedish setting[21].
Data Collection
A questionnaire was administered by trained field workers and included socio-demographic items, PA- and motivation-related scales, anthropometric and biochemical measurements. Data were collected between January 2017 through December 2017 in Uganda, between August 2017 and November 2018 in South Africa and between June 2017 and January 2019 in Sweden.
Measures
Identified regulation towards physical exercise was assessed through the Treatment Self-Regulation Questionnaire for people with diabetes. This scale has been widely used to test PA self-regulation and studies have reported adequate reliability[22]. Guided by factor loadings identified in the study by Levesque et al.[22], four items were selected to measure identified regulation (see Additional file 1). Participants responded to each item on a 5-point Likert-type scale ranging from 1 (strongly disagree) to 5 (strongly agree).
To measure perceived social support, an adapted version of the scale for participation and involvement of family members and friends in PA was used[23]. This scale has been used and validated in a variety of contexts[23]. Five items of the initial measure were selected based on their presumed cross-cultural adaptability and factor loadings in previous studies (see Additional File 1). Participants responded to each item on a 4-point Likert-type scale ranging from 1 (never) to 4 (more than once a week). Perceived social support shows conceptual parallels with perceived relatedness and the same scale was used by others to measure perceived relatedness[24]. To emphasize the concept of perceived support among the study participants, we introduced the questions with the following statement: “We want to understand to what extent people close to you (friends, family or relatives) have helped you to do physical activity”.
Barrier self-efficacy (or self-regulatory efficacy) corresponds to the perceived capability to maintain PA given various conditions or impediments (i.e. barriers). Six items were adapted from the health-specific self-efficacy scale developed by Schwarzer et al. (2007) (see Additional File 1). Barriers included in the original questionnaire were modified to barriers relevant to the study contexts. Participants responded to each item on a 5-point Likert-type scale ranging from 1 (strongly disagree) to 5 (strongly agree). Self-efficacy and perceived competence have shown to be correlated[25], but a conceptual difference needs to be acknowledged[18]. Unlike self-efficacy, perceived competence encompasses the concept of personal effectance, or the perceived need to effect change and attain valued outcomes[18].
PA was measured through: 1) self-reported frequency of vigorous PA; and 2) self-reported frequency of moderate PA. Initial questions from the World Health Organisation "STEPS" survey[26] were contextually adapted (see Additional File 1). Different measures were chosen since associations with SDT constructs may depend on the intensity of PA[6]. More detail on the motivational and PA measures can be found in a previous study on SDT by De Man et al.[17].
HbA1c was measured using capillary blood samples obtained with a Point-of-care HbA1c Analyzer Cobas b101 (Roche Diagnostics) with the respective test and control reagents.
Contextual Adaption
All measures were translated into the local language of the study populations (i.e. Lusoga, Swedish, Arabic, Somali and isiXhosa), and adapted to the context based on inputs from a team of local research assistants. Measures were then back translated to English and adjustments made where necessary to ensure that the meaning of the questions was not lost. Local validity was ensured through piloting in a non-study area, training of data collectors (e.g. through mock interviews), and minimizing inter-interviewer variability.
Data Analysis
The current state-of-the-art approach to compare mean levels and associations of latent constructs across different settings is multigroup structural equation modelling (MGSEM)[27]. A major condition to apply this technique is measurement invariance of constructs across different settings[27]. Measurement invariance supports the idea that subjects of different subgroups have a similar understanding and give a similar meaning to the items of a latent construct. Testing for measurement invariance was based on subsequent steps imposing additional constraints to the models. Before testing for measurement invariance, separate measurement models were assessed for each construct and country separately. Confirmatory factor analysis (CFA) was used to assess the loadings of the item indicators on the latent variables (i.e. identified regulation, barrier self-efficacy, and social support) and the goodness of fit of these measurement models. Subsequently, simultaneous analysis of equal form (i.e. configural invariance), equivalence of factor loadings (i.e. metric invariance) and equivalence of intercepts (i.e. scalar invariance) was conducted across countries through MGSEM.
In case these measurement models would yield an acceptable fit and were shown invariant, the fit of the hypothesized structural equation model (SEM) was assessed per country separately. Finally, to test if the associations between the constructs across the three countries were similar, we compared the difference in model fit between a model constraining these associations as similar across settings and a model without constraining a specific association. In case model fit was not significantly worse between these nested models, we concluded that that specific association was similar across settings. Model fit was evaluated based on multiple indices, including root-mean-square error of approximation (RMSEA) corrected for nonnormality[28], with target values as proposed by Hu and Bentler[29]: the comparative fit index (CFI) ≥ 0.95, Tucker-Lewis index (TLI) ≥ 0.95, the root mean square error of approximation (RMSEA) ≤0.06, and the standardized root mean square residual (SRMR) ≤0.08. Criteria used to assess the difference in fit between nested models included a difference in CFI bigger than -0.01 combined with a difference in RMSEA smaller than 0.015 or a non-significant scaled χ-square difference test[30]. Since items’ distributions departed from normality, we used maximum likelihood estimation with robust standard errors and a Satorra-Bentler scaled test statistic[31]. Covariates were added to address potential sources of confounding based on theory and identified through directed acyclic graphs. Education and body mass index (BMI) were included to adjust the motivational constructs. Age, sex, occupation, BMI and education were included to adjust the PA outcomes[15]. Data were analyzed using R software with the packages “lavaan” and “semTools”. To study the link between PA outcomes and HbA1c, a linear regression model was used controlling for the following covariates: age, sex, BMI and reported intake of oral antidiabetic medication[32].
Missing Data
Missing data for the Ugandan site varied from 0.0-1.3% per variable, for the South African site from 0.0-1.4% and 2.0-10.2% for the Swedish site. The variable responding to the question: “are you currently on any oral hypoglycemic agents?” was missing more frequently: 49.4 %, 51.2 % and 14.3% among participants in the Uganda, South African and Swedish site respectively. BMI was missing among 16.3% of the participants in the Swedish site. Multivariate imputation by chained equations with predictive mean matching was used to handle the missing data under a missing at random assumption. Rubin's rules were used to pool point and SE estimates across 30 imputed data sets. The procedure was done using the “Mice” package in R. For the variable regarding oral hypoglycemic treatment, a sensitivity analysis was ran ignoring this variable.
Ethics Approval
The study was approved by the ethics committees in each of the respective countries (details were masked). Informed consent was obtained from all individual participants included in the study.