The association between various biological, behavioural and psychosocial factors and type 2 diabetes mellitus in Africa: a systematic review and meta-analysis


 Background

Type 2 diabetes mellitus (T2DM) is a significant public health concern in many African countries. While the determinants underpinning T2DM are likely to be Africa-specific, knowledge of these risk factors is largely derived from developed countries. This is the first systematic review and meta-analysis to include biological, behavioural and psychosocial risk factors for T2DM in Africa.
Methods

Relevant scientific databases were searched, and data were extracted from 66 studies. Fifty- nine studies reported unadjusted data and were analysed using Comprehensive Meta-Analysis (CMA) software version 2.0. The Odds ratios (OR) and their 95% CIs for the associations between BMI indices (overweight, obesity), central obesity (waist circumference, waist to hip ratio), behavioural (alcohol, fruit and vegetable consumption, smoking), physical inactivity, and psychosocial factors (stress, anxiety, and depression) and T2DM were calculated using a random-effect model. Moderator effects of age, language-spoken, sub-regions and urban/rural location was assessed.
Results

A number of risk factors were associated with T2DM including, BMI-based definitions of obesity [OR = 3.22, 95% CI: 92.73, 3.80)], overweight [OR = 2.22, 95% CI: (1.90, 2.58)], or overweight/obesity [OR = 2.58, 95% CI: (1.76, 3.78)]; Central obesity as measured by waist circumference [OR = 2.51, 95% CI: (1.86, 3.37)], or waist to hip ratio [OR = 2.03, 95% CI: 1.51, 2.72]; psychosocial factors which includes stress [OR = 2.2, 95% CI: (1.46, 3.31)], depression [OR = 2.42, 95% CI: (1.14, 3.31)] and anxiety [OR = 2.05, 95% (CI:1.0, 4.18)] and physical inactivity [OR = 1.88, 95% CI: (1.53, 2.23)]. Current smoker [OR = 1.13, 95% CI: (0.84, 1.49)], alcohol consumption [OR = 1.10, 95% CI: (0.82, 1.47)] and inadequate fruit and vegetable consumption [OR = 0.81, 95% CI: (0.57, 1.16)] were not associated with T2DM. Locality (urban/rural), language spoken, and sub-region (East/West Africa) did not significantly moderate the associations between the risk factors and T2DM.
Conclusion

Obesity (defined by BMI) is most strongly associated with T2DM. Overweight, waist circumference and waist to hip ratio, physical inactivity, psychosocial risk factors defined as stress, depression, anxiety were all significantly associated with T2DM. These findings add novel meta-analyses of associations between diverse individual risk factors and T2DM within the African context.


Introduction
Diabetes Mellitus (DM) is a major contributor to global mortality and morbidity. Since the 1980s, Africa has seen an emergence of T2DM as an important non-communicable disease (NCD) increasingly threatening the health, political and socio-cultural framework of the continent's population [1,2]. Across Africa, T2DM prevalence varies signi cantly within and between countries, geographical location (e.g. urban vs rural) [5,6] and sub-regions (e.g. East vs West Africa) [4], as well as countries of common or different spoken language (e.g. French vs. English speakers) [3]. Comparative gures show that while the prevalence of diabetes in other continents is expected to increase by between 15% (Western Paci c) and 84% (South East Asia) from 2017 to 2045, the prevalence in Africa is predicted to increase by 156% within the same period [4]. As such, Africa is on the path to bear the most signi cant burden of T2DM epidemiology including Disability-Adjusted Life Years (DALY ), premature death and mortality in the decades ahead [5].
In sub-Saharan Africa (SSA), 69% of people with T2DM aged 20 to 79 are undiagnosed [6]. This gure is substantially greater than the global gure of 50% [6]. As such, the majority of Africans with T2DM are diagnosed only after presenting with substantial health complications. This delay in diagnosis, coupled with a lack of access to high-quality healthcare, results in loss of productivity, morbidity, premature death [7], and mortality [8]. The Global Burden of Disease study suggests that all-age total DALY lost due to NCDs including diabetes in SSA, increased by 67% between 1990 (90·6 million [95% UI 81·0-101·9]) and 2017 (151·3 million [133·4-171·8]) [5]. Broader knowledge of the relationship between modi able risk factors and T2DM underpinning this burden and the escalating trend is based predominantly on studies involving Europid and other non-African populations [9,10]. Previous studies suggest that the relationship between risk factors such as body weight indicators (e.g. obesity & waist circumference) and T2DM varies across different populations [9,[11][12][13]. The unique racial composition, culture, socioeconomic factors, dietary patterns, political structures, geography and environment in Africa mean that Africaspeci c studies of risk factors for T2DM are important. Understanding the African-speci c relationships between potentially modi able risk factors (biological, cultural, behavioural and psychosocial) and T2DM is of critical importance to primary prevention efforts [2,14,15].

Design
This study follows the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines in reporting of this systematic review and meta-analysis [53]. Figure 1 depicts the PRISMA ow chart. The protocol for this study has been previously published [3] and is registered with PROSPERO (Registration number CRD42016043027).

Search Strategy
We conducted a database search in English of published quantitative data only. The initial search was conducted in November 2017 with an updated search in February 2019. The search included studies that focused on diabetes and a range of biological, behavioural and psychosocial risk factors within the African context. Please see Additional le 1 for a list of search terms. The search covered scienti c databases including Global Health, PsycINFO, CINAHL, MEDLINE, Psychology and Behavioural Science. Reference lists of included articles were searched for additional articles.

Inclusion and Exclusion Criteria
The study considered articles that reported empirical research, were conducted among populations living in Africa and published in English from 1 January 2000 to 16 February 2019. We included quantitative studies from cross-sectional, longitudinal, cohort and case-control that reported associations between T2DM as an outcome measure and the risk factors of interest. Studies were included if T2DM was measured following the 1999 World health organisation (WHO) and the American Diabetes Association (ADA) criteria. The risk factors examined include body weight indicators which include BMI indices [overweight (OV), obesity (OB) and overweight/obesity (OV/OB)], and central obesity indices (waist circumference (WC) and waist to hip ratio (WHR)]. Others include physical activity (PA), current alcohol consumption (ALC), current smoking (SMO) and psychosocial risk factors (stress, anxiety, and depression).
We excluded studies with outcome measure as type 1 diabetes, impaired glucose tolerance, impaired fasting glucose and gestational diabetes.
For the independent risk factors, we excluded socioeconomic status as a potential risk factor due to its direct link to both non-modi able (e.g. age, sex, and ethnicity) and modi able risk factors (e.g. SMO, ALC, diet, and PA) as opposed to T2DM [54]. We also excluded hypertension and dyslipidaemia as T2DM may causally affect hypertension and dyslipidaemia whereas the opposite relationships are unlikely to be causal [55,56].
Reviews, editorials, commentaries, and letters were excluded. Studies that reported relevant risk factors and T2DM, but not the associations between them, were excluded. Associations with risk factors treated as the outcome measures in multivariate analyses adjusted for various covariates were also excluded.

Screening
The search yielded 39,561 studies and results were imported into Endnote Version 9. Duplicates were removed and 30,713 studies remained. The rst reviewer (AI) assessed study eligibility for inclusion by initially screening the titles and abstracts. For an article to pass the rst stage of screening, the title, and abstract needed to mention diabetes and any of the risk factors. Five hundred and four studies were retained, and full text obtained for the second stage. Another 8 articles were found in the reference lists.
At this stage, the articles were further considered if three reviewers (AI, YP, and AC) considered that their abstract showed that the paper may contain quantitative data and an association between the risk factors and T2DM. The observed average agreement between reviewers on the screening at this stage was 95.9%. After completing this stage, a total of 177 studies were retained for the third stage and screened against all criteria including an appropriate measure of associations by the rst reviewer and results con rmed by a second reviewer (YP). Final decisions to exclude studies at this stage were discussed by two reviewers (AI and YP) and disagreements were resolved by consensus with third (AC) and fourth (CS) reviewers. The agreement between reviewers at this stage was 97.5% with 66 studies remaining for data extraction [16-22, 24, 25, 27-34, 57-94] and including both unadjusted and adjusted data (e.g. adjusted for covariates such as age and family history). Only unadjusted data were included in the meta-analysis.

Data extraction and management
The nal 66 studies were reviewed, extracted and coded by the one author (AI), and extracted data were double-checked by YP. We developed a Microsoft Excel spreadsheet and an accompanying manual for this purpose. We then extracted data from studies that reported an unadjusted bivariate association between the risk factors and T2DM as well as studies that adjusted for different covariates. Information extracted from each article includes study design, number of studies (k), year of publication, study participants (n), risk factors (e.g. OB) and associations with T2DM for effect size calculations [95]. Others include age, geographical locations (urban, rural), countries and sub-regions, language spoken, sample size, and odds ratio (OR) with con dence intervals (CI).
The de nition of T2DM was based on WHO or ADA criteria, and included a fasting blood glucose (FBG) of ≥ 7.0 mmol/L (126 mg/dL), or 2-hour glucose level of ≥ 11.1 mmoL/L (200 mg/dL) during a 75 g oral glucose tolerance test (OGTT) or the glycated haemoglobin (HbA1c) value of C 6.5% [96].
De nitions of smoking, alcohol and fruit and vegetable consumption adopted for this study have been used extensively in previous reviews [97][98][99][100] and include: Current smoking status (yes or no, including former smokers. Current alcohol intake (either i. yes or no; or ii. current drinkers vs. never/non-drinkers). Fruit and vegetable consumption (inadequate vs adequate consumption) and physical activity (vigorous vs low physical). For psychosocial risk factors, the reference group was no stress, depression or anxiety vs. currently having these conditions. Due to the small number of studies for these individual risk factors (stress, anxiety, depression) they were aggregated into a summary variable for psychosocial risk factors.

Data Integration, coding and analysis
The metrics used in this meta-analysis in measuring associations between the risk factors and T2DM were ORs and 95% CIs, employed as a measure of the effect size. Other metrics were converted to ORs including cross-tabulation (2 × 2) of events and non-event (T2DM/ risk factor × T2DM/no risk factor) as well as Chi-squared. Comprehensive Meta-Analysis (CMA) was used for all conversion, calculations as well as coding the data [101]. Where ORs were lower than 1 (i.e. risk factor was protective), ORs were reverse coded by calculating 1/ORs for consistency and ease of interpretation.
Variation in sample sizes was accounted for by calculating weighted effect sizes, giving larger samples more weight. Most studies provided multiple homogenous associations between risk factors and T2DM which were not independent. In resolving the issue of dependence, we used the averaging method to calculate a single effect size per study [102] (i.e., the shifting-unit-of-analysis approach [102]). This approach not only ensured that each study contributes exactly a single independent association per analysis but was more advantageous in retaining as much data as possible [102,103]. Since the study aims to generalise the ndings, the random effect model was used to calculate the overall effect sizes. For moderator analysis, mixed-effect models were a more conservative approach to test for different moderator levels. Moderation analyses were conducted separately for all risk factors.
Heterogeneity and moderator analysis Q and I 2 statistics were used to assess effect sizes of study heterogeneity [104]. The Q statistics measured the existence of heterogeneity, while the I 2 statistics assessed the percentage of study variability due to heterogeneity instead of by chance [104,105]. I 2 values of approximately 50% and ≥ 75% were considered moderate and high heterogeneity respectively [104,105]. Moderation analysis was used to explain the heterogeneity and interaction effects of the risk factors and T2DM. The strength and directions of these associations were tested using moderators such as study participants' age, subregions (East/West Africa), Locality (urban/rural dwelling) and lingua franca (French or English speakers). A moderator variable was analysed if it had at least two levels that were examined in ve or more studies [95]. This is based on the established minimum threshold widely used in various metaanalyses [95,104,106,107].

Vote-counting analysis
This study utilised unadjusted data only and was complemented by summarising the adjusted data using the vote-counting methods. Vote counting is a method of synthesising evidence from multiple estimates by comparing the number of studies with positive results against the number of studies with negative results. However, it does not take into consideration sample size, study quality and effect size [108]. The most common variables adjusted for in this study included hypertension, IFG, age, family history of diabetes, gender, socioeconomic status, education, and marital status [109].

Publication bias analyses
Using the CMA program, we assessed publication bias in three different ways. First, we checked for bias by producing funnel plots and physically examined any evidence of symmetry. Second, we examine the intercept for signi cance of the sample of studies for statistical evidence of bias using Egger's test [110]. Thirdly, due to the propensity to publish studies with statistically signi cant results relative to the nonsigni cant ones, thus "the le drawer problem [111], the Rosenthal's Failsafe number was used to calculate the number of non-signi cant unpublished (or missing) studies needed to be included to the meta-analysis to change statically signi cant results to non-signi cant one [112]. The criterion suggests that, the value of the Failsafe N should be greater or equal to 5 times the number of included studies in a meta-analysis. The Rosenthal's failsafe N was also used to decide whether the effect is an artefact of bias or by chance. The Duval and Tweedie trim and ll methods were then used to adjust for detected bias or unpublished (or missing) papers to assess what the effect size would have been given no bias existed [113,114].

Descriptive
An initial search generated a list of 30,713 publications after duplicates were removed. A total of 66 studies met the inclusion criteria, with 59 of them reporting unadjusted data and hence included in the analyses. Tables 1, 2, and 3 provide descriptive statistics of study, risk factor, and participant level characteristics respectively. The overall sample size was 181,204 participants. There was a high variation in sample sizes across the 66 studies. Study sample sizes ranged from 90 to 45,767 participants, with 37% of articles reporting a sample between 50-500, while 20% of studies had a sample size between 501-1000 and 35% included a sample size between 1001 to 5000. Nine percent of studies had a sample size between 5001-45,767. The included 66 studies were comprised of 328 individual unique associations. Of these associations, 206 were unadjusted from 59 studies. Two articles reported two different studies each in one paper [20,93]. These were comparative studies from multiple countries in Africa. We differentiated these studies by adding study 1 and 2 to each reference. The 66 articles were published between 2000 and 2019. Of these studies, 35 articles were published between 2016 to 2019 (53%), 23 studies between 2011 to 2015 (35%), and 8 studies between 2000 to 2010 (12%). All studies were published in academic peer-reviewed journals. Many of the studies were conducted among West African populations (39%). Of these, 21% were published in Nigeria, and others from Ghana (12%), Senegal (5%), and Burkina Faso (2%). This is followed by East Africa, where 38% of the total studies were published. Among these countries, 11% were from Ethiopia, or otherwise Uganda (8%), Kenya (9%), Malawi (2%), Mozambique (2%), Tanzania (5%), North Sudan (2%) and South Sudan (2%). The majority of the studies used a cross-sectional design (86%). Only 4 studies (6%) [21,66,79,115] reported longitudinal data and 6% were case-control studies [16,61,116,117]. More studies were conducted in

Publication bias analyses
The publication bias analysis shows that, apart from BMI-OV/OB, funnel plots for the BMI indices and the other risk factors were comparatively symmetrical. However, the intercept for Egger's regression test was   (29,154). Three of the signi cant associations were between smoking and T2DM and one was between alcohol and fruit and vegetable consumption respectively.

Discussion
We conducted the rst systematic review and meta-analysis focusing on a range of biological, behavioural as well as psychosocial factors and their associations with T2DM in Africa. Consistent with other systematic reviews and meta-analyses [95,[127][128][129], publication outputs have increased over time [130]. The greatest number of studies were from West and East Africa. The present study showed that all body weight indicators (both BMI and central obesity de ned), physical activity and psychosocial factors were signi cantly associated with T2DM with associations of varying strength. Obesity (de ned by BMI) was found to have the strongest association with increased odds for T2DM of more than 3-fold that of any other body weight measures including waist circumference.
These general observations are consistent with some [131], but not all [132] prior meta-analyses. Findings from the present study are consistent with the study by Vazquez et al. [131] among populations from Mauritius, USA (including African Americans), Asia, and Europe. They are also mostly consistent with the study by Kodama et al. [132] among European subjects, however that study did suggest a stronger association between waist circumference and diabetes than that between obesity (de ned by BMI) and T2DM. Waist circumference is a stronger indicator of intra-abdominal visceral fat than BMI, and closely linked to insulin resistance and hyperinsulinemia [133]. Interestingly, our ndings of a stronger association between BMI and T2DM has also been documented in other pathophysiological studies in African using methods such as dual energy X-ray absorptiometry [134,135]. Studies in South African women suggest that, for the same BMI, African women have less central fat, but greater peripheral fat accumulation than Caucasian women [133][134][135]. Sumner et al. [136] also show that increasing waist circumference results in less visceral adipose tissue among African-American and African women than Caucasian women. Although these ndings were predominantly from studies among women, they may partly explain why the strength of the association between obesity (de ned by BMI) and T2DM is stronger in this study than waist circumference.
The independent role of central obesity in insulin resistance in populations of African descent must not be discounted, however. For example, while detailed examination of the overall body weight indicators shows comparatively larger effect size for the relationship between combined BMI indices (BMI-OV, BMI-OV/OB, and BMI-OB) (OR = 2.59) and T2DM compared to the combined central obesity indices (WC and WHR) (OR = 2.24), a strong and independent association between central adiposity and T2DM was evident [137]. This may suggest that either WC or BMI alone could be used as risk factors for T2DM among Africans. The use of a measuring tape alone to assess WC may be appealing in a setting where resources are minimal [138]. A further issue which may have affected our ndings in relation to associations between waist circumference and T2DM is the high heterogeneity (I 2 = 87) of studies, which could not be explained by the moderation analysis. In addition, although meta-analyses have suggested a stronger association between BMI and diabetes in women than men [131,139], this could not be assessed in the present study due to too few studies reporting ndings separately for females and males.
Our analysis of physical activity indicates an almost two-fold increase in odds of T2DM for those who are inactive. After adjusting for publication bias, the magnitude remained the same. These ndings are congruent with previous meta-analyses among populations from China, the USA, and Australia that have explored associations of medium to vigorous physical activity and T2DM [140]. While the present study focused on vigorous physical activity and T2DM, other reviews included alternative physical activity measures such as walking, leisure-time activity, resistance activity, occupational activities, low, moderate and vigorous-intensity activity [141]. These measures are important within the African context since physical activity patterns in Africa are somewhat different from industrialised countries [142]. Low to moderate activities are most prevalent within Africa [142]. Although beyond the scope of this paper, an indepth comparative examination of these differences is required within the African context.
Psychosocial factors in this study were found to increase the odds of T2DM by more than 2-fold with a combined OR = 2.15. These ndings are consistent with various meta-analytic reviews in non-African populations that explored relationships between psychosocial factors and diabetes [143]. A study by Smith et al. [97] in North Americans (includes Whites, African Americans, Hispanic and Chinese), European, Middle-Eastern and Asian populations found that anxiety increased the odds of diabetes by almost one and a half fold, while that of Ali et al. [144] among a population from the USA, Europe (the Netherlands, Finland, and Italy) and Iraq found that depression increased the odds of diabetes by almost one and a half fold. In the present study, however, the small sample size of depression (k = 4), stress (k = 2) and anxiety (k = 2) may have limited power to detect the association with T2DM. As such these ndings should be treated as preliminary and interpreted with caution. Again, a ne-grained longitudinal study examining the individual psychosocial risk factors (stress, depression, and anxiety) in the African context is required given evidence of differential associations with T2DM.
Findings of the lifestyle risk factors showed that in this African sample, associations between fruit and vegetable intake, alcohol consumption, and smoking and T2DM were not signi cant. However, various studies among non-African populations have shown otherwise. A meta-analysis of longitudinal studies among populations from the United States, Japan, United Kingdom, Germany, Israel, and the Scandinavian countries, showed that active smoking is associated with an increased risk of type 2 diabetes [100]. Similarly, a meta-analysis of longitudinal studies among populations from Europe, USA, Australia, Korea and Japan showed a U-shaped relationship between alcohol consumption and diabetes for both men and women, with a greater protective effect of moderate consumption observed for women [145]. The non-signi cant ndings of the present study may be partly due to the simpler de nitions used in many African studies. Assessing consumption levels based on a report of "Yes" or "No" is likely to not be precise enough to detect an effect, particularly where the association is not strong, or there is a Ushaped association [145,146].
Studies on diet quality show that adherence to the appropriate diet can improve insulin sensitivity and glycaemic control. The current study had data on fruit only and vegetable consumption only, which are protective factors for diabetes. However, only two studies presented associations for fruit and vegetable intake combined, and no study had data on consumption of discretionary, typically ultra-processed, unhealthy foods, which are strongly related to higher body weight and diabetes [147,148]. As such, we were unable to con rm the link between diet and T2DM among the African populations. There is a clear need for further studies (preferably longitudinal) assessing the complex association between diet and the incidence of Type 2 diabetes in Africa [149].
These ndings were supported by the adjusted data synthesised by the vote-counting analyses, speci cally with regards to the more signi cant results including all body weight indicators. Most moderation analyses were non-signi cant except for the combined central obesity indices which were moderated by locality (East and West Africa). However, no other moderator was signi cant. This implies that the risk of acquiring T2DM is independent of geographical location (urban/rural) and spoken language in Africa.

Strengths and Limitations
There are several limitations associated with this study that should be considered when interpreting the ndings. First, only studies published in English were included and thus may under-represent studies predominantly from the Arabic, Portuguese and French-speaking countries in Africa. The use of only unadjusted data may constitute a limitation due to challenges posed by adjusted covariate for systematic reviews and meta-analysis [127]. However, previous systematic reviews and meta-analyses reported non-signi cant ndings between unadjusted bivariate and those adjusted for different covariate [132,150]. Further studies can also explore moderations using meta-regression. In this study, various diabetes de nitions were used among the included studies. However, it is unlikely that these differences in de nition would have had any impact on the results as the aim of this study was to determine the associations between the risk factors and T2DM, and not the optimal cut-point for each risk factor [138].
The absence of study quality assessments in this study is another potential limitation. Using a critical appraisal tool for meta-analytic studies exploring risk factors and T2DM have been rare, due to challenges posed by observational studies with diverse methodologies. It is also possible different study qualities from individual studies may affect the association between risk factors and T2DM. For example, the observed association between physical activity and diabetes in Africa may have been in uenced by using different measurement approaches as well as other confounders used in individual studies. In particular, the different instruments used may constitute a limitation as in the case of psychosocial factors where no two studies used the same tool. Nevertheless, the in uence of unknown residual confounders can also not be ignored in individual studies (particularly the lifestyle risk factors) as this may affect the results. There is an urgent need for quality data, particularly on lifestyle risk factors. In sum, there is urgent need of longitudinal studies with a probability sampling technique to systematically measure both risk factors and T2DM, particularly alcohol, smoking, psychosocial factors, unhealthy food, and a range of relevant risk factors in Africa.
Our study has many strengths which include the systematic nature of the review and the use of comprehensive meta-analytic methods, which have not been used to determine the strength of association between risk factors and T2DM in Africa previously. Finally, the large sample size of the study provides high power and precision in our estimates.

Conclusion
This study constitutes the rst comprehensive systematic review and meta-analysis of the association between a range of biological, behavioural, and psychosocial risk factors and diabetes in Africa to date.
These ndings add novel meta-analyses of associations between diverse individual risk factors and T2DM. Risk factors including obesity (de ned by BMI), overweight (de ned by BMI), and overweight/obesity (de ned by BMI); central obesity (de ned by waist circumference and waist to hip ratio), physical activity, psychosocial risk factors (stress, depression, anxiety) were all signi cantly associated with T2DM. The present study shows that, unlike for studies of European populations, obesity, as de ned by BMI, was the factor most strongly associated with T2DM.
This research provides an update on the associations between a range of factors and diabetes in Africa. This is a growing eld of research as con rmed by the increased rate of publications over time. While this metanalysis has identi ed gaps in the literature, we hope that this metanalysis provides renewed

Declarations
Ethics approval and consent to participate Not applicable.

Consent for publication
Not applicable.
Availability of data and materials All data used generated or analysed during this study are available from the corresponding author on reasonable request.