Diet Behavior Change Techniques in Type 2 Diabetes: A Systematic Review and Meta-analysis

Dietary behavior is intricately linked to type 2 diabetes and has become an increasingly complex phenomenon to understand and change. There is a long association between diet and the pathogenesis of type 2 diabetes. A recent study suggested that reduced risk of type 2 diabetes was strongly associated with dietary factors such as greater intake of fruit, vegetables, legumes, nuts, whole grains, and long-chain fats and a lower intake of sugar-sweetened beverages (1), trans fat, processed/red meats, and sodium and a moderate alcohol intake (2). Dietary factors have also been linked to the highest proportion of deaths in type 2 diabetes, stroke, and heart disease (3). There is a need to identify factors associated with effective clinical outcomes in dietary interventions (4–6). Identifying effective behavior change techniques (BCTs) in successful dietary approaches to type 2 diabetes management may help to refine and improve the scalability of successful approaches to changing dietary behavior. A BCT is an observable, replicable, and irreducible component of an intervention designed to alter or redirect causal processes regulating behavior, such as “feedback” or “self-monitoring” (7). The objective of this systematic review and meta-analysis was to identify dietary BCTs, intervention features, and specific diets associated with changes in HbA_1c and body weight in type 2 diabetes.

This systematic review and meta-analysis followed a registered protocol (http://www.crd.york.ac.uk/PROSPERO/display_record.asp?ID=CRD42016042466). A PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) checklist was created and PRISMA review guidelines were followed (Supplementary Data Table S1).

1. Randomized controlled trials (RCTs) of any duration using a dietary intervention published in peer-reviewed journals between 1 January 1975 and 12 April 2017 were included.

2. RCTs with a comparison arm or control group that constituted usual care were included. Usual care could include typical diabetes dietary treatment such as recommended by the American Diabetes Association or carbohydrate exchange–type diets.

3. Human participants were older than 18 years of age with clinically confirmed type 2 diabetes at time of recruitment.

4. The primary clinical outcome measure was HbA_1c. However, studies reporting HbA_1c results as a secondary outcome measure were also included.

5. Randomized crossover trials were included if relevant outcome data were reported for both intervention and control groups prior to subjects crossing over to the other diet.

1. RCTs of diabetes prevention or RCTs in populations at risk for type 2 diabetes

2. RCTs that used pharmacological agents exclusively to treat type 2 diabetes

3. RCTs that included supervised physical activity

4. RCTs that targeted multiple chronic diseases, gestational diabetes, or type 1 diabetes

5. Studies not reported in English

6. Studies that focused exclusively on supplement or micronutrient use

The following databases were searched using a Boolean combination of keywords and MeSH terms: Cochrane Library, CINAHL, Embase, PubMed, PsycINFO, and SCOPUS (Supplementary Data Table S2). Search terms were developed following the protocol of an earlier review (8) and using a series of sensitivity analyses of terms, cross-checking results against identified reference criteria. Additional records identified from other sources such as reference lists of relevant reviews, studies with multiple intervention arms in an earlier review, and all included studies were searched for additional sources. The original search was conducted on 22 February 2016 and repeated on 12 April 2017.

Articles were initially screened by two research team members (K.A.C. and R.M.) based on titles and abstracts and then full texts of the remaining articles (see Fig. 1). A third member of the review team (H.L.G.) oversaw any disagreements on search results and had the final say on included studies. Interrater agreement by the Cohen κ for the full-text search results was 0.82.

Studies that aimed to control or change environment were classified as studies where all food or the majority of food was provided to participants. Studies in this category could also be described as studies with high internal validity. Studies that aimed to change behavior were classified as studies where participants were instructed or educated about diet changes by dietitians or health care professionals; they included a variety of diets and no food was provided. Studies in this category could also be described as having high external validity.

Low-carbohydrate diets were classified as studies where carbohydrate intake of <130 g/day was recommended (9). Low-fat diets were classified as studies where dietary fat intake of <30% was recommended (10). High-protein diets were classified as studies where protein intake of >2 g/kg/day was recommended (11).

Data extraction was carried out by one member of the team (K.A.C.), and relevant information was stored in Excel file templates. All data regarding HbA_1c, weight loss, intervention features, BCTs, fidelity coding, and risk of bias was checked by another member of the research team (R.M.). All corresponding authors were contacted by e-mail (where contact details were available, n = 50 of 54) using a standardized template to request additional information. The response rate was 34% (over a period of 6 weeks).

The Cochrane Collaboration risk of bias tool was used to assess methodological quality (12). Assessment criteria are applied to seven aspects of trials to yield an appraisal of “low risk,” “high risk,” or “unclear risk” of bias. Studies were independently assessed by two members of the review team for risk of bias and methodological quality (K.A.C. and R.M.). Assessment results were discussed and agreed upon after the first 10 studies and again after the first 20 studies. Interrater reliability was calculated and discrepancies were discussed after each round. Treatment fidelity was assessed according to the five categories proposed by Bellg et al. (13). Each category and subcategory was assigned a score of yes, no, or unclear. RCTs were independently assessed by two members of the review team (K.A.C. and H.L.G.) and results were discussed and assessments repeated following discussion of the first 10 and first 20 articles. Interrater reliability is based on the final 34 studies.

The BCT taxonomy v1 of Michie et al. (7) was used to identify and code BCTs related to diet only that were identified in each study. A list of all 93 BCTs and their descriptions is available (http://www.bct-taxonomy.com) (7). A BCT was only coded when it was explicitly mentioned in the intervention methodology. BCTs were coded separately for intervention and control groups. A coding rubric was developed by three authors (K.A.C., R.M., and H.L.G.) to guide the coding process (Supplementary Data Table S3). All included studies were coded independently by two authors (K.A.C. and R.M.) immediately after training in the use of the Michie et al. taxonomy. A third master coder (H.L.G.) independently assessed the coding results and arbitrated any disagreements. Cohen κ and prevalence-adjusted bias-adjusted κ calculations were used to establish intercoder reliability of BCTs present and absent. A BCT must have been used in at least three studies for inclusion in the moderator analysis (14). Both coders discussed coding practices and results after coding the first study and again after independently coding the remaining 53 studies. All available information, including study manuals, protocols, and earlier methodology papers, was also used to code each study. Rationale for classification of intervention features such as mode of delivery, provider, intensity/frequency of intervention, and other variables included is documented in an earlier review carried out by the review team (8). Intensity was defined as the number of total or face-to-face contacts with intervention personnel and frequency was defined as the average number of weeks between contacts.

We defined an HbA_1c reduction of ≥0.3% (3.3 mmol/mol) as clinically significant, which follows the precedent set by others (14,15). Meta-analyses were conducted using RevMan (v5.3) on the primary outcome measure of HbA_1c and the secondary outcome of body weight, where sufficient data were available. We recorded changes in outcomes at 0–3, 3–6, 6–12, and 12–24 months. The use of these time points allowed a greater number of studies to be included. In studies that reported data from multiple time points, we used the time point closest to the end of the intervention for analysis. Mean differences and SDs in the differences for HbA_1c and weight loss between baseline and the different time points were calculated. SDs for missing data were calculated from SE, P, and t values, where reported, using the Cochrane guidelines or were calculated using a correlation (method documented in an earlier review by this research team) (8). Correlations of 0.75 (HbA_1c) and 0.98 (weight loss) were used to calculate missing SDs following a sensitivity analysis of correlations and reported SDs. Statistical significance of the moderator and meta-analyses was set at P < 0.05. In all cases the meta-analyses used the raw mean difference and a random effects model to calculate results. Meta-analyses were carried out on both interventions that aimed to change behavior only and interventions that aimed to change environment only. Further meta-analyses were carried out on HbA_1c and weight loss at 0–3, 3–6, 6–12, and 12–24 months. Meta-analyses were also carried out on different diet types.

A series of moderator analyses were carried out to investigate the association between BCTs/intervention features and clinical outcome results (HbA_1c) where the presence or absence of a certain BCT or intervention feature in certain studies was associated with changes in HbA_1c. The moderator analyses reported the standardized mean difference in outcomes, using Comprehensive Meta-Analysis (CMA) software. The random effects analysis was used to conduct all moderator analyses. For every moderator variable (BCT and intervention feature), we calculated the point estimate, 95% CIs, Q statistic, and P value.

As a result of the large effect size observed in the control group, a series of subgroup analyses were carried out in an attempt to elucidate the true effect size of intervention groups compared with control groups. Further subgroup sensitivity analyses were carried out on true control groups, excluding studies where the control group had >1 contact with a dietitian, and an additional subgroup analysis removed studies with a control group reduction in HbA_1c of ≥0.3% (3.3 mmol/mol). Moderator analyses of BCTs associated with reductions in HbA_1c were also carried out on studies that used interventions aimed at changing behavior primarily or environment primarily and studies that reported results at 3 months only.

The inclusion criteria were met by 54 studies (16–69). Summary characteristics of included studies are outlined in Supplementary Data Table S4. Average age of participants was 57.7 ± 4.7 years for intervention groups and 58.1 ± 5.1 years for control groups. For intervention and control groups, respectively, mean duration of diabetes, where reported, was 7.6 ± 3.3 years and 7.3 ± 3.1 years, mean baseline HbA_1c was 8.1% (69.8 mmol/mol) ± 1% and 8.12% (69.9 mmol/mol) ± 1%, and BMI was 32.1 ± 4.1 kg/m² and 31.8 ± 4 kg/m². Six of the included studies were carried out in a community center setting, 12 studies did not report the setting, one study was online, one study was in a hotel setting, and all remaining studies (n = 34) were carried out in a clinical or academic setting. All participants in the 54 studies were classified as having type 2 diabetes; however, diagnostic criteria for HbA_1c varied among included studies from a minimum of 5.5% (47.4 mmol/mol) (34) to a maximum of 12% (103.4 mmol/mol) (23,57). The mean percentage of dropouts per study was lower in the intervention groups (12.6%) than in control groups (16.4%). Studies with low–glycemic index and high-protein diets reported the lowest number of dropouts (mean percentage of dropouts 1% and 1.8%, respectively), with meal replacement studies reporting the highest dropout rate (mean percentage of dropouts 28%) (Supplementary Data Table S5).

In the assessments of risk of bias, 63% of the studies were classified as “unclear,” while 34% were “low” across all seven variables and only 2% were classified as “high” risk of bias (Supplementary Data Tables S6a and S6b). Raters agreed on 78% of risk of bias assessments following initial assessment and came to agreement on the remainder through discussion. Treatment fidelity results are documented in Supplementary Data Table S7. Overall, the reported use of treatment fidelity strategies was very low, with 78% assessed as having not used a treatment fidelity strategy. The most widely used treatment fidelity strategy was in the subcategory of “monitoring and improving enactment of treatment skills,” where 68.5% of all studies reported use of “ensure participant use of behavioral skills.” Raters agreed on 76.5% of assessments and resolved remaining disagreements through discussion and arbitration.

Meta-analysis of interventions that aimed to change behavior (n = 39) showed an overall reduction in HbA_1c of 0.32% (3.5 mmol/mol) (95% CI −0.40, −0.23; P < 0.00001), while interventions that aimed to change or control the environment showed an overall reduction in HbA_1c of 0.5% (5.5 mmol/mol) (95% CI −0.65, −0.34; P < 0.00001) (Fig. 2). Sensitivity analysis removing studies with a >0.3% (3.3 mmol/mol) reduction in HbA_1c in the control group (n = 21) increased further the observed effect size on HbA_1c, with behavioral interventions showing a reduction of 0.32% (3.5 mmol/mol) (95% CI −0.41, −0.24; P < 0.00001) compared with 0.66% (7.3 mmol/mol) (95% CI −0.88, −0.45; P < 0.00001) for environment-controlled studies (n = 11) (Supplementary Data Table S8).

Studies included in this review focused on different dietary approaches: low carbohydrate (n = 9), low fat (n = 16), high protein (n = 5), meal replacements (n = 4), low glycemic index (n = 3), medical nutritional therapy (n = 2), Mediterranean (n = 2), and others (n = 13). There was considerable heterogeneity in the diets used in control groups. There was a “true” control group in 28 studies, where no additional intervention support or contact was provided. In 16 studies, American Diabetes Association or American Heart Association guidelines were applied to control groups, with varying degrees of intervention support provided (Supplementary Data Table S9). The duration of interventions carried out ranged from 4 weeks to 2 years. In 21 studies, there was an additional minor physical activity component.

Studies using meal replacements and high-protein diets were associated with the greatest reductions in HbA_1c (0.56% [6.2 mmol/mol] and 0.5% [5.5 mmol/mol], respectively). Low-carbohydrate diets showed a greater reduction in HbA_1c (0.44% [4.8 mmol/mol]) than low-fat diets (0.40% [4.4 mmol/mol]) or low–glycemic index diets (0.09% [1 mmol/mol]) (Supplementary Data Table S10).

Meta-analyses showed differences in HbA_1c between intervention and control groups at different time points, presented graphically in Fig. 3. Combining all studies and all time points in one overall meta-analysis (n = 59, 54 studies) showed a reduction in HbA_1c of 0.35% (3.8 mmol/mol) (95% CI −0.43, −0.28; P < 0.00001) (Supplementary Data Table S11). Heterogeneity as measured by I² was 62%, 44%, 38%, and 68% at 0–3, 3–6, 6–12, and 12–24 months, respectively. Sensitivity analysis comparing data at exactly 3, 6, 12, and 24 months to data at 0–3, 3–6, 6–12, and 12–24 months using a larger dataset (n = 54) showed no significant differences.

The difference in body weight loss between intervention and control groups was 2.34 kg (95% CI −2.99, −1.69; P < 0.00001), 2.94 kg (95% CI −3.92, −1.97; P < 0.00001), 2.27 kg (95% CI −3.32, −1.21; P < 0.0001), and 2.14 kg (95% CI −3.34, −0.93; P = 0.0005) at 0–3, 3–6, 6–12, and 12–24 months, respectively (Supplementary Data Table S12). Combining all studies and time points revealed a reduction in body weight of 2.41 kg (95% CI −2.96, −1.86; P < 0.00001) (Supplementary Data Table S13). Heterogeneity as measured by I² was 84%, 93%, 88%, and 27% at 0–3, 3–6, 6–12, and 12–24 months, respectively.

A total of 42 distinct BCTs were applied in the intervention groups, 7 of which were reported only once. The number of BCTs used in a single RCT ranged from 3 (25,35,45) to 17 (41). The five most frequently occurring BCTs were “instruction on how to perform a behavior” (n = 54), “credible source” (n = 45), “self-monitoring of behavior” (n = 37), “monitoring of behavior by others without feedback” (n = 32), and “social support (unspecified)” (n = 24) (Table 1).

Control group BCTs were coded separately, and 28 different BCTs were identified, with the number of BCTs used in a single study ranging from 0 (49,60) to 15 (44) (Supplementary Data Table S14). Interrater agreement as determined by the Cohen κ was 0.7 after coding 44 studies and 0.68 after all 54 studies were coded. A breakdown by category of BCTs used (Supplementary Data Table S15) and BCTs not used (Supplementary Data Table S16) was also carried out.

The original moderator analysis showed no BCTs coded for diet behavior were associated with >0.3% (3.3 mmol/mol) reduction in HbA_1c (Supplementary Data Table S17). Subgroup analysis of interventions using only true control groups showed that the BCTs “social comparison” (0.52% [5.7 mmol/mol], P = 0.012) and “feedback on behavior” (0.365% [4 mmol/mol], P = 0.046) were associated with clinically and statistically significant reductions in HbA_1c (Supplementary Data Table S18). Subgroup analysis of interventions excluding studies with control group pre- to poststudy change of >0.3% (3.3 mmol/mol) in HbA_1c also showed the BCT “feedback on behavior” (0.34% [3.7 mmol/mol]) associated with clinically significant reductions in HbA_1c (Supplementary Data Table S19). Subgroup analysis of BCTs reporting outcome changes at 3 months showed that the BCT “problem solving” (0.63% [6.9 mmol/mol]) was associated with clinically significant reductions in HbA_1c (Supplementary Data Table S20). Moderator analysis was not carried out at 12 months because insufficient data were available.

Subgroup analysis of interventions aimed at changing behavior showed that the BCTs “feedback on behavior” (0.52% [5.7 mmol/mol], P = 0.007) and “adding objects to the environment” (0.39% [4.3 mmol/mol]) were associated with clinically significant reductions in HbA_1c (Supplementary Data Table S21). Subgroup analysis of interventions aimed at changing the dietary environment showed that the BCT “problem solving” (0.5% [5.5 mmol/mol]) was associated with a clinically significant reduction in HbA_1c (Supplementary Data Table S22).

The original moderator analysis showed no intervention feature was associated with a clinically significant reduction in HbA_1c (Supplementary Data Table S23). Subgroup moderator analysis excluding studies with control group pre- to poststudy change of >0.3% (3.3 mmol/mol) in HbA_1c showed that the only intervention feature associated with a clinically significant reduction in HbA_1c was the use of a theoretical model or framework (0.33% [3.6 mmol/mol]). Other intervention features associated with reductions in HbA_1c were a higher frequency and number of both total and face-to-face contacts with intervention personnel (Supplementary Data Table S23).

These findings suggest that changing or controlling dietary environmental factors may be more effective than strategies to change dietary behavior in attempting to reduce HbA_1c in adults with type 2 diabetes. High-protein diets and meal replacement programs produced the greatest reductions in HbA_1c. A clinically significant difference in HbA_1c at 0–3, 3–6, and 6–12 months was reported when all dietary approaches were combined in meta-analyses. Weight loss occurred but diminished over time. Moderator analyses identified four BCTs—“problem solving,” “feedback on behavior,” “adding objects to the environment,” and “social comparison”—and the intervention feature “use of theory” that were associated with clinically significant reductions in HbA_1c.

Diets where the environment was changed or controlled (e.g., where all food was provided) were more than twice as effective in reducing HbA_1c than diets using behavioral change interventions. This observation was consistent when a range of different foods were provided, including high-protein (29,30,48), meal replacement (21,27,50,57), low-carbohydrate (51,62), low-fat (51,55,63), Mediterranean (33), Korean traditional (35), vegetarian (36), and partial formula or partial low-calorie diets (52). These studies represent a more internally valid approach compared with studies aimed at changing behavior; however, successful externally valid interventions are required in order to change diet in a real-world setting. It has been suggested that environmental changes to social, built, and food environments, in addition to individual behavioral changes, are required in order to adopt a healthy diet and lifestyle (70). Changing the environment has been identified as one of the overall theoretical themes associated with changing behavior, particularly in the longer term (71).

In regard to the type of diet, our finding of a modest but statistically significant reduction in HbA_1c at 3, 6, and 12 months needs to be interpreted with caution, as a range of different diets were combined in an effort to elucidate the most effective BCTs and intervention features. The observed reduction in HbA_1c is greater than in previous reports with fewer (n = 20) studies of high-protein, low-carbohydrate, and low–glycemic index diets. Our data also indicate that the use of meal replacements and high-protein diets results in the greatest reduction in HbA_1c, with low-carbohydrate diets showing a greater reduction in HbA_1c than low-fat diets. However, meal replacement interventions also had the highest dropout rate, suggesting these interventions may not be externally valid or the most acceptable approach for participants. The average number of dropouts per study was the lowest for the high-protein diets at 1.8%, suggesting that it was the diet most acceptable to participants. The overall meta-analysis showed an overall weight loss of 2.41 kg, with the greatest decrease observed at 6 months (2.94 kg) but diminishing over time—a pattern consistent with previous work (4).

Beyond the BCT “adding objects to the environment,” the three other BCTs “social comparison,” “feedback on behavior,” and “problem solving” all have strong theoretical foundations and have been shown to be efficacious in other studies (5,8,72,73). Given that the use of more BCTs was not associated with greater effectiveness, the pattern of application or fidelity of use of BCTs may be of greater importance. Of the 93 BCTs in the taxonomy of Michie et al. (7), 51 BCTs were not found in any of the 54 reviewed studies. BCTs most frequently used came from the categories “feedback and monitoring,” “shaping knowledge,” “goals and planning,” “comparison of behavior,” and “social support.” The BCTs “behavioral contract” and “commitment” were not used in any of the included studies. The studies reviewed focused almost exclusively on reflective motivation, suggesting that deployment of a wider range of BCTs needs to be investigated in changing dietary behavior and improving HbA_1c and body weight in adults with type 2 diabetes.

The only intervention feature in the moderator analysis that was associated with clinically significant reductions in HbA_1c was the “use of theory/model” to inform interventions. Similar findings have been reported, in which dietary behavior interventions in cancer prevention were more effective when informed by theory (74). However, fidelity of the use of theory was not reported in the studies included in our review or other reviews (74), and descriptions of use of theory varied considerably from “integrated concepts from different theories” (46) to “behavior modification treatment used principles from the modern learning theory” (37) and “group educational classes were based on the social cognitive theory” (21). The social cognitive theory (75) was the only theory reported more than once (21,46,64).

Our findings might suggest that higher frequency and greater number of contacts are associated with greater reductions in HbA_1c, which is consistent with our previous systematic review of combined diet and physical activity interventions (8), although this may arise from more intervention content being delivered. However, we cannot be sure because fidelity was so poorly reported in almost all categories and in all studies apart from one subcategory of “monitoring and improving enactment of treatment skills,” where 68.5% of studies reported use of fidelity. This subcategory was coded “yes” when the intervention description reported that subjects carried out a 3- or 7-day food record and that it was reviewed by the dietitian. The criteria (13) for intervention fidelity assessment do not take account of the extent of use of each category, which is particularly relevant in assessing participant adherence to dietary programs, i.e., enactment. Low levels of enactment make it difficult to assess the efficacy of interventions.

These findings from large, well-controlled dietary interventions have potentially important implications for type 2 diabetes management and suggest that interventions aimed at changing the dietary environment warrant further scrutiny. It is impractical from a treatment perspective to provide food and control the food environment as a scalable solution to type 2 diabetes treatment in the community. However, this finding does provide evidence that changing or altering the food environment or using highly internally valid interventions is efficacious. Providing foods at the beginning of a program or intervention might be an effective strategy to help people manage their diabetes, followed by instruction on how to choose, shop for, and prepare these foods, gradually weaning them off of reliance on foods provided. We would suggest that future studies look at the economic ramifications of changing the food environment from policy, marketing, and farming perspectives. Individual behavior change efforts might benefit from increased awareness of the dietary environment and exertion of greater control over one’s dietary environment. For example, the individual could be guided through an audit of their current home food environment (stored food supplies) and inappropriate food would be removed to eliminate potential impulse food consumption of inappropriate foods. If we really want to change diet in real-world settings, we also need to find BCTs associated with successful externally valid dietary interventions.

Future studies ought to quantify intervention fidelity, which would allow the identification of more effective BCTs. The use of video or audio recordings of consultations with dietitians and other health care professionals may help to better define the range of BCTs being deployed in any given intervention.

A strength of our study is the use of the most recent and comprehensive taxonomy of behavior change techniques available (7). We have provided a comprehensive analysis of fidelity categories and subcategories as well as detailed subgroup meta- and moderator analyses. Limitations of our study are the heterogeneity in the dietary interventions and the different diagnostic criteria for type 2 diabetes, which are likely to have resulted in variable effects on HbA_1c and weight in patients with type 2 diabetes. However, we think that this heterogeneity in the interventions and in their efficacy is likely to have increased random error rather than bias, making our findings even more compelling. The quality of reporting of BCTs in different studies varied considerably and was poor overall. This is particularly true of fidelity measures, and any conclusions must be tempered with the recognition of significant inter- and intrastudy variability in adherence to intervention protocols. As study protocols do not always stipulate each BCT used, BCTs are likely to have been underreported.

In conclusion, this systematic review and meta-analysis provides evidence that changing dietary environment may be more important than focusing on dietary behavior in type 2 diabetes treatment. More robust reporting of content, fidelity, and frequency of BCTs and intervention fidelity, as well as better alignment of intervention design with behavior change theory, would be helpful in refining interventions so that they are more efficacious.

Funding. The authors thank the Irish Research Council for Science, Engineering and Technology for funding this project (Project ID GOIPG/2013/873).

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. K.A.C., G.ÓL., F.M.F., L.R.Q., K.A.M.G., and H.L.G. formulated the research question and defined the search terms. K.A.C. carried out the electronic searches. K.A.C. and R.M. carried out the search process, the methodological assessment, and the BCT coding. H.L.G. guided the BCT coding process and acted as a master coder. K.A.C. and H.L.G. carried out the fidelity assessment. K.A.C. carried out the moderator analysis and the meta-analysis. All authors were involved in writing and reviewing the final manuscript.