ORE Open Research Exeter TITLE Effective Techniques in Healthy Eating and Physical Activity Interventions: A Meta-Regression AUTHORS Michie, S; Abraham, C; Whittington, C; et al. JOURNAL Health Psychology DEPOSITED IN ORE 03 October 2013 This version available at http://hdl.handle.net/10871/13754 COPYRIGHT AND REUSE Open Research Exeter makes this work available in accordance with publisher policies. A NOTE ON VERSIONS The version presented here may differ from the published version. If citing, you are advised to consult the published version for pagination, volume/issue and date of publication 1 IDENTIFYING EFFECTIVE TECHNIQUES Effective techniques in healthy eating and physical activity interventions: A meta- regression Susan Michie, Centre for Outcomes Research and Effectiveness, University College London, UK Charles Abraham, University of Sussex, UK Craig Whittington, University College London, UK John McAteer, University College London, UK Sunjai Gupta, Department of Health, England Corresponding author Professor Susan Michie, Centre for Outcomes Research and Effectiveness, Department of Psychology, University College, Gower Street, London WC1E6BT. [email protected] Author Note We thank Professors Paschal Sheeran and Robert West and Drs Andrew Prestwich and David French for helpful comments on earlier drafts of the paper, and Dr Philippa Davies and Anna Davies for help in literature searching and data extraction. 2 Abstract Objective Meta-analyses of behavior change (BC) interventions typically find large heterogeneity in effectiveness and small effects. This study aimed to assess the effectiveness of active BC interventions designed to promote physical activity and healthy eating and investigate whether theoretically-specified BC techniques improve outcome. Design Interventions, evaluated in experimental or quasi-experimental studies, using behavioral and/or cognitive techniques to increase physical activity and healthy eating in adults were systematically reviewed. Intervention content was reliably classified into 26 BC techniques and the effects of individual techniques, and of a theoretically-derived combination of self-regulation techniques, were assessed using meta-regression. Main Outcome Measures Valid outcomes of physical activity and healthy eating. Results The 122 evaluations (N = 44,747) produced an overall pooled effect size of 0.31 (95% CI 0.26 to 0.36) (I2 = 69%). The technique, “self-monitoring”, explained the greatest amount of among-study heterogeneity (13%). Interventions that combined self-monitoring with at least one other technique derived from control theory were significantly more effective than the other interventions (0.42 versus 0.26). Conclusion Classifying interventions according to component techniques and theoretically-derived technique combinations and conducting meta-regression enabled identification of effective components of interventions designed to increase physical activity and healthy eating. 3 Key words: physical activity, healthy eating, behavior change, self-regulation, meta- regression, meta-analysis, systematic review 4 Introduction Interventions designed to change health-related behaviors generally include many components and typically produce small effects in meta-analyses, but with large heterogeneity in effectiveness (e.g. National Institute for Health and Clinical Excellence, 2007; Dishman & Buckworth, 1996; Grimshaw et al., 2004). This limits the potential for understanding how intervention content relates to effectiveness and, consequently, the inferences that can be drawn regarding optimal design and the content of future behavior change interventions. Recent guidance has called for new methods to evaluate the effects of “complex” interventions (Craig et al., 2008). This study aimed to assess the utility of classifying the content of behavior change interventions into component techniques and applying meta-regression to identify effective individual techniques and theoretically-derived combinations of techniques. To address this aim, we focused on interventions designed to increase physical activity and healthy eating because these are key change targets in the context of the growing obesity epidemic, one of the most serious health risk factors in both the developed and developing world (World Health Organisation, 2002). We further focused on active interventions that engaged participants in the process of behavior change, rather than passive interventions such as simply providing information or advice. Self management approaches, involving people in their own change, have had considerable success among those with long term illnesses (e.g., Lorig, Ritter & Plant, 2005), can initiate change within other groups (Bandura, 2000; Gupta, 2005). Active interventions have also been found to be more effective than passive interventions in other areas (Albarracín, et al., 2005) and, because of the sustained behaviour changes necessary to translate dietary and physical activity into health benefits, self regulatory processes are likely to be central to health-enhancing change, recommending active engagement of participants. Yet, despite the potential of active, self management approaches, there is little guidance on which techniques are important to the effectiveness. 5 Two methodological advances have enhanced our capacity to learn from intervention evaluations. First, reliable methods of specifying component techniques (e.g., Abraham & Michie, 2008) and, second, use of meta-analysis and meta-regression to identify the effects of individual techniques, and combinations of techniques, across studies (e.g., Albaraccin et al., 2005). In the current study we combined these tools in an investigation of effective change techniques included in healthy eating and physical activity interventions. Repeated calls have been made for precise specification of what makes one behavior change intervention more effective than another and how this can be understood theoretically (e.g., Rothman, 2004). In the current study, we used a reliable taxonomy of 26 techniques to identify intervention content. Reliability checks have shown that independent coders can reliably judge whether or not published intervention descriptions in papers or manuals indicated inclusion of each technique (Abraham & Michie, 2008). If we are to understand, not only what works, but how interventions work, it is necessary to understand the causal mechanisms hypothesised to explain intervention effects (Michie & Abraham, 2004; Michie, Johnston, Francis, Hardeman & Eccles, 2008). Interventions have been found to be more effective if they involve techniques that behavior change theory predicts would act synergistically (Albaraccin et al., 2005). Carver and Scheier’s (1981; 1982) control theory specifies action control processes underpinning behavioral regulation. The theory proposes that setting goals, monitoring behavior, receiving feedback and reviewing relevant goals in the light of feedback are central to self management and behavioral control. Therefore, while we examined which of 26 change techniques would be most strongly associated with intervention effectiveness, we hypothesized that interventions which included five self-regulation techniques derived from control theory would be more effective than other techniques. These were (1) prompt intention formation or goal setting, (2) specify goals in relation to particular contextualized actions, (3) self monitoring of behavior, (4) feedback on performance, and (5) review previously-set goals. These 6 techniques may act additively or synergistically; the number of studies required to detect the latter is substantially greater than the former. Previous studies have employed meta-analysis to assess whether the presence or absence of particular techniques is associated with effectiveness. For example, Albarracín et al. (2005) showed that 10 techniques (e.g., provision of factual information and attitudinal arguments) could be reliably identified in published descriptions of interventions designed to promote condom use, and that inclusion of some of these (e.g., provision of attitudinal arguments) was associated with greater effectiveness, while inclusion of others (e.g., threat-inducing arguments) was not. Noar, Benac and Harris (2007) showed that eight targeted theoretical constructs could be reliably identified in reports of tailored print interventions designed to promote health behaviors, and that inclusion of some of these constructs (e.g., attitudes, self-efficacy) was associated with greater effectiveness. Two (social norms and behavioral intentions) were not associated with effectiveness and one (perceived susceptibility) was associated with decreased effectiveness. Despite the impressive scope of these meta-analytic reviews, they have shortcomings. First, only 10 distinct techniques and eight constructs, respectively, were considered. The need for more comprehensive categorization of intervention content is evidenced from reviews of interventions in other behavioral domains (e.g., Webb & Sheeran, 2006). In addition, Albaraccín et al. (2005) used within-group change over time as the criterion of effectiveness as opposed to behavior change observed in an intervention group relative to changes observed in a matched no-intervention control. This allows inclusion of many more datasets but is a less rigorous criterion of effectiveness because the benefits of controlling for techniques within the control conditions are lost. In addition, both these reviews usedmeta-analysis and/or univariate regression rather than multivariate meta-regression to synthesize the evidence. While meta-analysis provides a technique for combining data from separate studies to arrive at pooled effect size estimates, meta-regression provides a means of assessing both single and multiple predictors of effect size from variables 7 derived from individual studies, while weighting the regression so that precision of study results is properly accounted for (Sutton & Higgins, 2008). The present systematic review applied a reliable taxonomy of behavior change techniques and meta-regression to analyse the effect of individual intervention techniques and the effect of combining five theoretically-derived self-regulation techniques. Method Search strategy and results We searched MEDLINE, EMBASE, PsychINFO, the Cochrane library (Cochrane Central Controlled Trials Register and the Health Technology Assessment database), AMED (Allied and Complementary Medicine Database) and HMIC (Health Management Information Consortium) databases between 1990–2008 for peer-reviewed journal articles written in English. Three search filters were used, one for interventions targeting physical activity/healthy eating, one for study design and one to exclude those with chronic diseases. Studies were also sought from experts in the field, identified by the British Psychological Society’s Division of Health Psychology experts list. Inclusion criteria specified interventions which recruited adults’ (18 years or over) in order to increase their levels of physical activity or healthy eating, used experimental or quasi- experimental designs (that is, controlled trials and interrupted time series designs) and outcome measures that were objective, standardised or validated self-report measures. Inclusion criteria also specified that interventions had to use cognitive or behavioral change strategies so that, for example, interventions consisting only of the provision of information were excluded. The following were excluded: interventions aimed at pregnant or recently post-natal women, amateur or professional athletes, those already engaged in a another intervention such as dietary, slimming or fitness programs, and interventions targeting those not living in the free-population or those exclusively targeting participants with physical or mental health problems. Studies targeting the 8 general population, with a small proportion exhibiting physical or mental health problems, were included if members of that sub-set were assessed as being healthy enough to participate by a physician. This strategy identified 34,769 references (physical activity [PA] = 13,870; healthy eating [HE] = 20,899). After excluding duplicates, 28,440 references remained (PA = 10,859 (including 22 papers recommended by experts in the field); HE = 17,581). In a sample of 300 titles screened independently by two reviewers, there was 100% agreement on inclusion/ exclusion. One thousand and forty one studies identified as potentially relevant were further screened by abstract to assess suitability for inclusion (PA = 472; HE = 569). One hundred abstracts were screened independently by two reviewers, with 85% agreement on inclusion. Disagreements were resolved through discussion and consulting a third reviewer and where uncertainly remained the full paper examined. After screening by abstract, full text papers were obtained for 270 articles (PA = 156; HE = 139). Where there was insufficient statistical or intervention information (N = 17), authors were contacted (35% responded). Detailed evaluation according to the inclusion criteria resulted in a final set of 139 studies. Of these, 38 were excluded from the meta-analysis (see supplementary material, Table S1), leaving 101 papers reporting 122 evaluations (PA = 69; HE = 53). Data extraction In evaluations of PA interventions reporting multiple outcome measures, the most general or comprehensive measure was selected (e.g., exercise level, energy expenditure). For studies of healthy eating, measures of good and/or poor diet were extracted. There was a significant correlation (r = .91, p < .001) between the “good diet” and “poor diet” measures1, consequently, an average effect size from each study was used for the meta-analysis. For studies reporting more 1 For an initial set of 18 studies that reported both good and poor diet measures. 9 than one measure of fat intake, total fat intake (grams per day or % energy from fat) was preferred over saturated fat intake or kcal consumption, since a certain kcal consumption may reflect a more or less healthy diet. For studies reporting the percentage of participants consuming five fruit or vegetable servings per day in addition to the number of fruit and vegetable servings per day, the latter was selected. Effect sizes were indexed as the standardised mean difference (the difference between two means divided by their pooled standard deviation) with Hedge’s correction for small sample size (Hedges & Olkin, 1985). For studies that reported continuous data, the effect size was computed from means and standard deviations (adjusted for baseline differences if reported), or, if these data were not reported, from the sample size and p-value from an appropriate between-groups t- or F- test. For studies that only reported dichotomous data, the log odds ratio was converted into a standardised mean difference using the meta-analysis software. For cluster RCTs, where the study had used an appropriate analysis to account for the effect of clustering, the results of the analysis were used to estimate the effect size. Where the analysis did not properly take account of clustering, we calculated an effective sample size using the following formula: N (effective) = (k x m) / (1+ (m - 1) * ICC, where k indicates the number of clusters; m, the number of observations per cluster; and ICC, the intracluster correlation coefficient (Shojania et al., 2006). We imputed unreported ICCs based on an empirically derived value of 0.05 (Elley, Kerse, Arroll & Robinson, 2003; Elley, Kerse, Chondros & Robinson, 2005). When results were reported only as significant, p = .05 was assumed, and when only as non-significant, p = .50 was assumed. Where data were reported from multiple time points, outcomes or evaluations, an average effect size was used (we explored the effect of doing this using a series of sub-group analyses, but found little difference between sub-groups; data not reported but available from the authors). Where there were two interventions compared in one study and both met the inclusion criteria, we chose the intervention with the greatest effect (because we were exploring determinants of effectiveness). Where a single