Databases (PubMed, Embase, CINAHL, ZETOC, NHS evidence, Clinical trials.gov and SIGLE) were searched using the combined search terms “type 2 diabetes” and/ or “intervention” and “black” or “African” or “Caribbean”. In addition, references of the journals and reference text books were consulted. The searches were done between November and December 2010, May to June 2011 and between September 24^th and October 6^th 2011 (Figure 1).

The population of interest was adult participants of African descent in developed countries which includes “African-Americans”, “African-Caribbean”, “Non-Hispanic blacks” and “self-described blacks”. Studies were restricted to those published in English from the developed

English speaking countries such as the USA and the UK The interventions of interest were non-pharmacological including educational programs. The primary outcome of interest was a measure of difference in HbA1C with the standard deviation.

This included randomised controlled trials (RCTs) or prospective cohort studies published between January 1^st 1995 and October 6^th 2011. The participants, males or females, had type 2 diabetes according to the WHO criteria and were 18 years or older (adults). Trials including participants with diabetes under age 18, persons with type 1 diabetes, known LADA, MODY and those with gestational diabetes were excluded. Also excluded, were interventions in which less than 90% of participants were of African descent. Feasibility studies for an intervention were also excluded, but pilot studies were included if the outcome data were appropriate, and if the study was not duplicated.

Critical appraisal checklists were adapted from Cochrane Collaboration tool for assessing risk of bias and the quality of each study. Information from all included studies was extracted and tabulated. Where data were insufficient for a particular study, authors were contacted for clarification.

Data were extracted using a RevMan data collection form according to the selection criteria, and subsequently analysed. The information extracted included the study information (author and year), participant characteristics and number of participants recruited. The duration of intervention was 6 months where possible. Drop-outs, as well as type of intervention and duration of the intervention/study duration were also extracted with every effort to minimise missing data. The outcome measures of HbA1C difference were calculated for each trial before synthesising the data.

Standardised mean difference was used to describe the data. The mean difference for the intervention and the control were calculated separately and the standard deviation within the groups calculated for missing data. The effect size calculated was from the net difference in mean HbA1C between the intervention and the control groups. Hedges unbiased estimate was used to correct the overestimated value of the mean difference in studies with small sample sizes. Sd (within) was calculated from sq root [25].

RevMan software was used for Meta-analysis of HbA1C results. Testing for heterogeneity in the data by the chi-square test and Forest plot of the trials was included in the review. RevMan statistics were also confirmed by calculation.

All of the studies were comparisons of interventions within an African American population and the effectiveness at improving outcomes including HbA1C. The length of trials ranged from 6 months to 36 months. The sample participant sizes ranged from 64 to 4138 with mean age of 56 - 62 years. Participants were either all or at least 94% African Americans. All selected studies were from the USA. There were over 64% females in every group. The studies took place in urban African American communities, with the exception of Skelly et al. and Anderson-Loftin et al., which were located in rural southeast and South Carolina. Control groups included a 2 × 2 factorial design [31] and a three group experimental design [33]. The studies were of an intervention compared to usual care, conventional care or minimal intervention. Comparisons between the intervention and control group baseline group characteristics were not statistically different within each study. The one exception was that of Anderson-Loftin et al. [27].

In one trial [29] HbA1C was not the primary outcome and in the remaining trials, HbA1C was not the only outcome measured; BMI/weight, lipids and blood pressure were other outcomes measured in the remaining trials. Methods used to measure the unit of the outcome HbA1C varied and included high pressure liquid chromatography, Glycaffin column method, immunoturbidimetric method, and glucometer; none of the methods accounted for haemoglobin variants. HbA1C was measured in % mg/dl; IFCC mmol/mol units were not found in the literature. Proportion of patients reaching target HbA1C was only reported in few studies so this was

abandoned for meta-analysis.

The trials mainly targeted patient education. One study focused on exercise and dietary self-management. There were no multi-component interventions. The interventions used both individual or group designs—3 group interventions, 4 individual and 2 mixed interventions. There was one provider intervention and 8 patient interventions; the trial interventions used the expertise of diabetes educators, nurses, community health workers, lay health educators and counselors and, less often, physicians.

A total number of 3329 participants at baseline were divided into the intervention and control groups. Patients lost to attrition varied between 8% and 23% - 44%; for the latter there was a baseline difference between the intervention and control groups; all of the studies commented on attrition or addressed incomplete data. Using the modified Cochrane Collaboration tool for quality score, allocation concealment was described in one study [27], blinding in four studies [27,29,31,33], and adequate sequence generation in four [27,30,32,33]. Two studies were considered high quality, 6 studies were considered fair quality and 1 study considered inadequate/poor quality (Table 2).

Relevant HbA1C data were available for six of the studies for meta-analysis. The effect measure of standardised mean difference of the random effects model is presented in the forest plot below (Figure 2). The pooled results reveal high heterogeneity, that is, I² = 92% of the variation between the studies is due to heterogeneity p < 0.00001; most of the observed variance is real, indicating that these studies are too different to compare. The random effects pooled standard mean difference favoured the intervention −0.66 (−1.15, −0.17). The test for overall effect produced a p = 0.009. Sensitivity analysis was done with removal of Gary study (Figure 3). This study was an outlier with weight of 18.1% (42.9% in the fixed method analysis) with sample size over 200 in each study arm; the other studies had samples sizes that were more comparable. After sensitivity analysis the heterogeneity was moderate to high I² = 69% p < 0.01. The Chi-square test was still greater than the degrees of freedom at 12.99 versus 5 indicating heterogeneity. The random effects pooled standard mean difference still favoured the intervention after sensitivity analysis −0.48 (−0.81, −0.16). Also after sensitivity analysis the test for overall effect produced a p = 0.009.