Download The Impact of Past Behavior on Health Behaviors: A Dual-Process Approach and more Exercises Psychology in PDF only on Docsity!
The Reasoned Action Approach Applied to Health Behavior: Role of Past Behavior and Test of Some Key Moderators Using Meta-Analytic Structural Equation Modeling
Martin S. Haggera,b,c,d*†, Juho Poletb†, Taru Lintunenb
aHealth Psychology and Behavioral Medicine Research Group, School of Psychology, Faculty of Health Sciences, Curtin University, Perth, Australia bFaculty of Sport and Health Sciences, University of Jyväskylä, Jyväskylä, Finland cSchool of Applied Psychology, Menzies Health Institute Queensland, Griffith University, Mt. Gravatt, Australia dDepartment of Physical Education, Hong Kong Baptist University, Hong Kong
*Correspondence: Martin S. Hagger, Health Psychology and Behavioral Medicine Research Group, School of Psychologyy, Faculty of Health Sciences, Curtin University, GPO Box U1987, Perth, WA6845, Australia, email: martin.hagger@curtin.edu.au †These authors contributed equally to the manuscript
Full citation: Hagger, M. S., Polet, J., & Lintunen, T. (2018). The reasoned action approach applied to health behavior: Role of past behavior and test of some key moderators using meta- analytic structural equation modeling. Social Science & Medicine. doi: 10.1016/j.socscimed.2018.07.
Acknowledgements and Conflict of Interest Statement
Martin S. Hagger’s contribution was supported by a Finland Distinguished Professor (FiDiPro) award (Dnro 1801/31/2105) from Business Finland, and a Kennedy Y. H. Wong Distinguished Visiting Professorship from Hong Kong Baptist University. The funders had no role in the study design, the collection, analysis and interpretation of data, writing the article, and in the decision to submit for publication. Martin S. Hagger and Juho Polet contributed equally to this manuscript so it should be considered jointly first-authored. Martin S. Hagger, Juho Polet, and Taru Lintunen declare they have no conflict of interest.
Abstract Rationale : The reasoned action approach (RAA) is a social cognitive model that outlines the determinants of intentional behavior. Primary and meta-analytic studies support RAA predictions for multiple health behaviors. However, including past behavior as a predictor in the RAA may attenuate model effects. Direct effects of past behavior on behavior may reflect non-conscious processes while indirect effects of past behavior through social cognitive variables may represent reasoned processes. Objective : The present study extended a previous meta-analysis of the RAA by including effects of past behavior. The analysis also tested effects of candidate moderators of model predictions: behavioral frequency, behavior type, and measurement lag. Method : We augmented a previous meta-analytic data set with correlations between model constructs and past behavior. We tested RAA models that included and excluded past behavior using meta- analytic structural equation modeling and compared the effects. Separate models were estimated in studies on high and low frequency behaviors, studies on different types of behavior, and studies with longer and shorter measurement lag. Results : Including past behavior attenuated model effects, particularly the direct effect of intentions on behavior, and indirect effects of experiential attitudes, descriptive norms, and capacity on behavior through intentions. Moderator analyses revealed larger intention-behavior and past behavior-behavior effects in high frequency studies, but the differences were not significant. No other notable moderator effects were observed. Conclusion : Findings indicate a prominent role for non-conscious processes in determining health behavior and inclusion of past behavior in RAA tests is important to yield precise estimates of model effects.
Keywords: habit; non-conscious processes; theory of planned behavior; dual-process theories; behavioral frequency
Introduction
Epidemiological research has consistently identified pervasive links between regular participation in health-related behavior (e.g., physical activity, healthy eating, not smoking, drinking alcohol only in moderation) and reduced risk of chronic disease (e.g., cancers, cardiovascular disease, diabetes) and adaptive health outcomes (Ford et al., 2011; Li et al., 2018). The development of an evidence base of the psychological factors associated with participation in health behaviors has been identified as a priority in behavioral medicine (Conner & Norman, 2015; Sheeran et al., 2017b). The evidence may inform practice by identifying the salient constructs to target in behavioral interventions (Kok et al., 2016). Theories of social cognition have been at the forefront of research on psychological correlates of health behavior as they provide a fundamental understanding of behavior and the associated processes. Prominent among these theories is the reasoned action approach (RAA; Fishbein & Ajzen, 2009). The RAA is a generalized, belief-based theory that identifies sets of personal, social, and control-related factors that impact social behavior and the mechanisms involved. McEachan et al.’s ( 2016 ) meta-analysis of studies applying the RAA in health behavior contexts provided generalized support for its predictions. In the current article, we extended McEachan et al.’s findings by examining the role of past behavior as an indirect predictor of RAA variables in health behavior contexts. In addition, we tested effects of candidate moderators on RAA relations: behavioral frequency, behavior type, and time lag between measures of RAA constructs and behavior. We expected our analysis to advance previous findings by illustrating that inclusion of past behavior in the RAA provides important information on the extent to which health behavior may be subject to non-conscious processes. In addition, we expected studies on target behaviors that are likely to be performed more frequently, therefore offering greater opportunity for habit formation, to have larger effects of past behavior, and smaller effects of intentions, on behavior. We also expected effects of RAA constructs on behavior to be smaller in groups of studies with longer measurement lag.
The Reasoned Action Approach
The RAA is a social cognitive model that has been applied to predict and explain behavior in multiple contexts (Fishbein & Ajzen, 2009). The RAA is a more comprehensive version of its precursor, the theory of planned behavior (TPB; for reviews see Armitage & Conner, 2001; Hagger et al., 2002; Rich et al., 2015), and distinguishes between specific subcomponents of the TPB constructs. Intention is a central construct of the RAA. It reflects the extent to which individuals are motivated to perform a given behavior and is conceptualized as the most proximal antecedent of behavior. Intentions are a function of three sets of belief-based constructs: attitude, subjective norms, and perceived behavioral control (PBC). Attitudes are an individual’s overall positive or negative evaluation of a behavior. The attitude construct is differentiated into experiential and instrumental subcomponents that reflect the affective and utilitarian functions of attitudes, respectively (e.g., Lawton et al., 2009). Subjective norms reflect the individual’s beliefs that significant others want them to participate in the behavior. Subjective norms comprise injunctive and descriptive subcomponents (e.g., Sheeran & Orbell, 1999), which describe social pressures to perform the behavior and beliefs in the extent to which the behavior is typical or normal, respectively. PBC reflects the extent to which the behavior is under the individual’s control. PBC is differentiated into autonomy and capacity subcomponents (e.g., Terry & O'Leary, 1995), which reflect perceptions of control over doing the behavior and perceived confidence in doing it, respectively. Intentions are proposed to mediate relations between the belief-based constructs and behavior.
Predictive validity of the individual subcomponents of the RAA constructs has been supported in primary studies and meta-analytic reviews. Studies distinguishing between instrumental and experiential attitudes indicate that both subcomponents predict intentions directly, but also revealed direct effects of experiential attitudes on behavior. Although not previously hypothesized, the direct effects were interpreted as indicative of more spontaneous, non-conscious effects of affective responses on action (Conner et al., 2015; Lawton et al., 2009). Research has demonstrated significant effects of both normative components on intentions, with effects sizes for descriptive norms tending
other social cognitive variables on intentions and behavior; and (d) it leads to a significant increase in the amount of variance in behavior accounted for by the model (Ajzen, 2002; Brown et al., 2017; Hagger et al., 2016; Hagger et al., 2002; Hamilton et al., 2017; Norman & Cooper, 2011; Protogerou et al., 2018).
Theorists have suggested that direct effects of past behavior on behavior in social cognitive models reflect implicit or non-conscious processes that affect behavior beyond an individual’s awareness (Dombrowski & Luszczynska, 2009; Hamilton et al., 2017; Wood et al., 2014). This perspective is consistent with dual process theories of cognition and behavior, which predict that participation in health behaviors is a function of two processes: a reasoned, deliberative processes, that tend to be represented by the belief-based constructs in social cognitive theories like the RAA, and an implicit, impulsive process, that can be represented by past behavior (Hagger, 2016; Sheeran et al., 2013; Strack & Deutsch, 2004). Past behavior effects may model a number of different types of non-conscious processes. For example, Ouellette and Wood ( 1998 ) suggest that direct effects of past behavior on behavior independent of effects of other social cognitive constructs may reflect habits. A habit is defined as an action or behavioral tendency that it enacted spontaneously, with little conscious awareness or reflection, in response to a set of associated conditions or contextual cues. Research has indicated that contextual factors implicated in the development of habits, such as opportunity for frequent repetition of the behavior in the presence of a stable environment or facilitating conditions, determine whether behavior is predominantly determined by intentions or past behavior (Ouellette & Wood, 1998; Wood, 2017). Large effects of past behavior on behavior relative to intentions may be indicative of a habitual action, while larger effects of intentions on behavior relative to past behavior may be indicative of an action controlled by reasoned processes.
Inclusion of past behavior in social cognitive models may also attenuate effects of social cognitive constructs on behavior. This likely reflects previous or ‘habitual’ decision making; that individuals have processed information about the behavior in a similar way previously. Such previous
decision making is likely captured by measures of the social cognitive constructs in the RAA, and can be modeled by past behavior. In such cases, variance the social cognitive constructs attributable to previous decision making that is shared with behavior is also shared with past behavior. In the absence of past behavior, effects of these variables may give a misleading indication of the extent to which the behavior is determined by constructs representing reasoned processes that lead to action (Ajzen, 2002). Inclusion of past behavior in social cognitive theories is, therefore, important to provide precise estimates of the unique effects of theory constructs on behavior.
Effects of past behavior may also reflect effects of other unmeasured constructs on behavior that represent implicit, non-conscious determinants of behavior. For example, direct effects of past behavior may reflect effects of implicit cognitions or behavioral scripts, constructs that represent stored sets of information developed over time that initiate behavior beyond an individual’s awareness (Abelson, 1981; Ajzen, 2002). These implicit behavioral determinants have proposed to be separable from habits (Ouellette & Wood, 1998; Wood, 2017). The inclusion of past behavior as a predictor of intentions and behavior in tests of social cognitive models may, therefore, provide important information on the extent to which the behavior is controlled by non-conscious processes.
It is important to note that inclusion of past behavior may diminish the predictive validity of social cognitive models. Constructs in social cognitive models should account for effects of past behavior on subsequent behavior, that is, behavioral stability. However, if past behavior completely accounts for the effects of the social cognitive models on behavior, or renders them relatively trivial by comparison, then model constructs are effectively redundant as a means to explain behavior. It also has serious implications for the effectiveness of the model as a means to inform interventions aimed at changing behavior. If the social cognitive constructs assumed to be manipulable through intervention do not predict behavior, then affecting change in the constructs using intervention techniques will have no concomitant effect on behavior. In cases where past behavior is the dominant predictor and social cognitive constructs account for very little variance in behavior, it may be
Specifically, we aimed to extend McEachan et al.’s meta-analysis by augmenting their original correlation matrix among RAA constructs to include past behavior. Next, we tested the RAA model that excludes past behavior (see Figure 1) and compare model effects with the augmented model that includes past behavior (see Figure 2) using meta-analytic structural equation modeling. Overall, we expected our analysis to provide new data on whether RAA effects identified in McEachan et al.’s meta-analysis hold in the presence of past behavior, and provide insight into the process by which past behavior relates to subsequent behavior in health behavior contexts.
We also aimed to test effects of three candidate moderators of RAA model effects within the meta-analysis: behavioral frequency, behavior type, and time lag between social cognitive and behavioral measures. Specifically, we aimed to test whether frequency of performance of the target behavior moderated the effects of intentions on behavior, and the effects of past behavior on intentions and the other model constructs. Consistent with previous theory and research on the role of past behavior in social cognitive models (Ouellette & Wood, 1998; Wood, 2017), we hypothesized that behaviors affording greater opportunity to be performed frequently (e.g., physical activity, eating healthily) are more likely to be acquired as habits, and, as a consequence, the behavior is more likely to be controlled by non-conscious, automatic processes. In contrast, behaviors that are likely to be performed less frequently (e.g., blood donation, screening attendance) offer fewer opportunities for habit formation, and actions are, therefore, more likely to be determined by intentional, reasoned processes. We therefore predicted that the effect of intentions on behavior would be smaller, and the effect of past behavior on behavior as a proxy for habit, would be larger, in studies on behaviors that tend to be performed frequently. It is important to note that behaviors are likely to be habitually controlled not only when individuals have high opportunity to perform them frequently, but also when they are performed in stable contexts (Ouellette & Wood, 1998). Although frequency and context stability are often associated, segregating behaviors in terms of their opportunity to be
performed frequently without considering context stability is a relatively crude means to identify conditions that influence habitual control over behavior.
We also aimed to test whether model effects varied across different behaviors. Following McEachan et al.’s ( 2016 ) original analysis, we tested whether RAA predictions were consistent across studies with target behaviors that offer protection from risks to health (e.g., physical activity, healthy eating) and studies on target behaviors that are risky for health (e.g., smoking, drinking alcohol). Although they did not make specific predictions, McEachan et al. suggested that experiential attitudes were more likely to have larger effects on risk behavior when compared to protection behaviors given that participation in these behaviors might be determined by affective-related beliefs than instrumental beliefs. Their analysis supported this prediction with larger effects for experiential attitudes, but they also found larger effects for instrumental attitudes and descriptive norms. McEachan et al.’s moderator analyses were confined to univariate relations between model constructs. We aimed to extend this analysis by examining moderator effects in a full test of the RAA in which all model paths are estimated simultaneously. In addition, some behaviors, such as physical activity and dietary-related behaviors, were frequently tested in the set of studies used by McEachan et al., so we aimed to test whether or not model effects were consistent across specific behaviors.
McEachan et al. also tested effects of time lag between measures of RAA model constructs and behavior on individual correlations between the constructs. We aimed to conduct a similar analysis, but in the context of the full structural model. Consistent with previous meta-analytic research adopting social cognitive models indicating that behavioral prediction over time tends lead to smaller effects (Hagger et al., 2002; Mankarious & Kothe, 2015; McEachan et al., 2011), we expected effects of social cognitive factors, intentions, and past behavior on behavior to be larger when behavior is measured in close proximity to the social cognitive variables. However, we did not expect the RAA to be entropic, that is, for the proposed model effects to wane over time to zero, partly because temporal correspondence in measures tends to maintain prediction, but also because there are
data from four samples defined by ethnicity (Caucasian, African-American) and length of follow-up (two-week, two-month), we included all four samples in our analysis while McEachan et al. included only the two-month samples. Collection of studies and data extraction were conducted between May and August 2017.
Next, we tested relations among RAA constructs using meta-analytic structural equation modeling (MASEM) using the metaSEM package (Cheung, 2015) in R (R Development Core Team, 2017 ). A typical method used to analyze multiple relations among constructs in social cognitive models is to adopt a univariate approach that involves first correcting correlations among variables in the model across studies using meta-analysis and then using the average bias-corrected correlation matrix as input in a multiple regression or path analysis testing the model. This method has been used in many previous studies (e.g., Hagger et al., 2016; Hagger et al., 2017b; Ng et al., 2012) including McEachan et al. ( 2016 ) in their analysis of the RAA. However, this procedure has been identified as problematic because it necessitates finding a common sample size to estimate the standard errors, such as the harmonic mean of the sample size across studies, and assumes that the correlation matrix is a covariance matrix, which likely leads to inaccuracies in the resulting standard errors, confidence intervals, and chi-square values of the model (Cheung, 2015). MASEM offers an alternative that directly addresses the inherent limitations of the univariate approach. The analysis is conducted in two stages. In the first stage, correlation matrices from individual studies are pooled to form a common correlation matrix using random effects meta-analysis. The analysis produces a pooled correlation matrix representing the average correlation matrix among study variables in the population, with a sampling variance-covariance matrix that represents the known precision estimates of each correlation in the pooled matrix. The first stage yields zero-order correlations corrected for bias using a random effects meta-analytic model ( r+ RE) among study constructs across studies with 95% confidence intervals. In addition, statistics to evaluate heterogeneity in the effect sizes are also provided: the τ^2 statistic, the Q statistic, and the I^2 statistic and its 95% confidence interval.
Statistically significant τ^2 and Q values with I^2 values exceeding 25% with wide confidence intervals are considered indicative of substantive heterogeneity. We also computed conventional fixed- and random-effects meta-analytic estimates for each RAA effect using the metafor package in R for comparison.
In the second stage of the analysis, the pooled correlation matrix is used to estimate the proposed model. Two models were tested: a model testing the hypothesized pattern of effects among study constructs consistent with hypotheses from the RAA (see Figure 1) and a model augmented to include past behavior (see Figure 2). Fit of the proposed model with data from the meta-analysis was evaluated using multiple goodness-of-fit indices: the model goodness-of-fit chi-square, the comparative fit index (CFI), the Tucker-Lewis index (TLI), the standardized root mean square of the residuals, and the root mean error of approximation (RMSEA). A non-significant chi-square, CFI and TLI values that approach or exceed .95, a SRMSR value of less than .008, and a RMSEA value of .005 or less, indicate good fit of the model with the data (Hu & Bentler, 1999). Effects among model constructs were evaluated based on the likelihood based confidence intervals about model parameter estimates. We evaluated whether inclusion of past behavior in the model attenuated effects of RAA constructs on intentions, and of intentions on behavior, by computing the confidence interval of the difference in the parameter estimates across the models (Schenker & Gentleman, 2001). To the extent that the interval does not include zero, we have confirmation of a statistically significant difference in the parameter estimates across models.
We evaluated effects of behavior type, and time lag moderators on RAA model effects by estimating the models separately in groups of studies on behaviors classified as high and low frequency, studies on specific behaviors that had been tested with high regularity, studies on health risk and health protection behaviors, and studies with longer and shorter time lag between social cognitive and behavioral measures. Classification of studies into high ( k = 65) and low ( k = 16) frequency was based on Ouellette and Wood’s ( 1998 ) criterion. Although classification of studies into
size derived from a meta-analysis may deviate from the true effect due to selective biases in the sample of studies in the analysis. A principal source of bias may be ‘publication bias’ caused by a preponderance of studies in the sample with disproportionately large effect sizes relative to the sample size. Regressing effect size on study precision provides an estimate of the extent of bias and an estimate of the effect size corrected for bias. Two methods are used: the precision effect test (PET) and the precision effect estimate with standard error (PEESE). The PET is the linear regression of study effect size on its standard error estimate, weighted by the inverse variance. The intercept of the model provides an unbiased estimate of the true mean effect size. However, the PET may underestimate the true mean effect size when there is evidence of a non-zero effect. The intercept derived from a non-linear (quadratic) regression of study effect size on the square of its standard error, again weighted by the inverse variance, the PEESE, has been shown to provide a more precise estimate of the true mean effect in cases where there is evidence of a non-zero effect. We computed PET and PEESE estimates, with t -test for bias, and statistical significance of the bias-corrected effect from zero to provide an indication of selective bias in each estimate using the PETPEESE function in R (Carter et al., 2017). A full table including study characteristics, extracted data, and moderator coding can be accessed online from the Open Science Framework (OSF) project for this study: https://osf.io/w4hfr.
Results
Zero-order parameter estimates and bias statistics
Zero-order averaged bias-corrected correlations ( r +RE) among RAA model constructs including past behavior from the first stage of the MASEM analysis are presented in Appendix B (supplemental materials), with bias-corrected correlations from conventional random- and fixed-effects meta- analysis included for comparison. The bias-corrected correlations from the MASEM analysis were all significantly different from zero and non-trivial in size ( r +RE range = .158 to .564). Heterogeneity statistics revealed moderate-to-high heterogeneity according to the I^2 statistic. Values for the Q -
statistic across studies also indicated substantial heterogeneity for the models excluding ( Q (1244) = 6742.459, p < .001) and including ( Q (1440) = 7666.848, p < .001) past behavior. As correlations from the MASEM analysis and conventional meta-analytic models indicated positive non-zero effects, the PEESE regression test was taken as an estimate of selection bias. The test revealed significant bias in the majority of the effects. However, the bias-corrected PEESE estimates did not alter conclusions with respect to whether effects were different from zero and their overall size. Analysis scripts and output files for the metaSEM analysis, conventional meta-analysis, and PET- PEESE analysis with funnel plots can be accessed online from the study OSF project: https://osf.io/w4hfr.
Meta-analytic structural equation models
Goodness-of-fit statistics indicated that the second stage MASEM models of the RAA that excluded (Figure 1) and included (Figure 2) past behavior exhibited adequate fit with the data (see Appendix C, supplemental materials). Standardized parameter estimates with 95% likelihood-based confidence intervals from the MASEM analyses for both models are presented in Table 1. Patterns of effects for the model excluding past behavior were consistent with those found by McEachan et al. in their regression analyses, albeit with some variation in the magnitude of the effects. Specifically, we found statistically significant, positive direct effects of experiential and instrumental attitudes, injunctive and descriptive norms, and capacity on intentions. We also found significant and positive direct effects of experiential attitudes, capacity, and intentions on behavior. Effect sizes were small (β < .150), with the exception of the effects for experiential attitude, descriptive norms, and capacity on intentions, and intentions on behavior. In addition, experiential and instrumental attitudes, injunctive and descriptive norms, and capacity had significant and positive indirect effects on health behavior through intentions, consistent with hypotheses. Indirect effects for instrumental attitudes, and injunctive and descriptive norms were relatively small and trivial (β < |.075|) (Seaton et al., 2010),
However, it is important to note that these observed differences were not statistically significant across moderator groups. In addition, we found significantly larger effects of past behavior on experiential attitudes, injunctive norm, and descriptive norm in studies on low frequency behaviors. There were also trends toward larger effects of past behavior on autonomy and capacity in studies on low frequency behaviors.
For the behavior type moderators, the models estimated on dietary behaviors and health risk behaviors did not converge due to small numbers of studies and positive non-definite covariance matrices. The model on health protection behaviors was therefore compared to the full sample, effectively a sensitivity analysis testing whether our conclusions on model effects would change with the exclusion of studies on health risk behaviors. The model estimated on studies on physical activity behaviors was compared to models on other behaviors. We found no differences in model effects for studies on health protection behaviors relative to the overall sample (Appendix E). Effects of past behavior on experiential and instrumental attitudes, and injunctive norms were smaller in studies on physical activity behavior relative to studies on other behaviors. However, there were significant differences or observed trends in the total effects of past behavior and intention on behavior across behavior type (Appendix F). Finally, there were no significant differences or observed trends in model effects across the measurement lag moderator (Appendix G). It seems that model effects were consistent regardless of time gap between measures, indicating that the model effects were consistent in studies with shorter- and longer-term behavioral follow-up.
Discussion The purpose of the present study was to extend McEachan et al.’s ( 2016 ) meta-analysis of the RAA for health behaviors by including past behavior as a predictor in the model. The analysis aimed to evaluate the extent to which past behavior predicts prospectively-measured health behavior beyond the RAA constructs – effects that likely model habitual, non-conscious processes on health behavior. We also set out to examine the effects of candidate moderators of RAA model effects: behavior
frequency, behavior type, and time lag between measurement of social cognitive constructs and behavior. We updated McEachan et al.’s database of correlations among the RAA constructs with correlations between these constructs and past behavior, and extracted data on behavioral frequency, behavior type, and measurement lag. MASEM was used to test the proposed RAA models excluding and including past behavior across studies. Results for the model excluding past behavior replicated the pattern of effects reported in McEachan et al.’s original analysis. Inclusion of past behavior resulted in significant attenuation of the direct effect of intentions on behavior, and the indirect effects of experiential attitudes, descriptive norms, and capacity on behavior through intentions. The indirect effects of past behavior through the social cognitive variables in the model and intentions accounted for a relatively modest proportion of the total effect of past behavior on behavior. We found few statistically significant differences or observed trends in model effects when testing the model in studies across levels of the moderator variables.
Current findings point to the pervasive effect of past behavior as a determinant of health behavior, which has two important ramifications for the interpretation of RAA predictions. First, it means that regardless of any variability in effect sizes due to potential moderators (e.g., context, behavior, or population), the observed effects among RAA constructs in health behaviors are likely to be non-zero. Furthermore, the effect of intentions on behavior is likely to be lower when past behavior is included in the model. In other words, researchers testing the model would expect to find these patterns of relations regardless of variations in study characteristics. This is consistent with previous meta-analytic research examining past behavior effects in social cognitive models (Hagger et al., 2016; Hagger et al., 2002), and recent research applying the RAA to health behaviors (Conner et al., 2017). Second, an important implication of these findings is that tests of the RAA in the absence of past behavior are likely to yield inflated estimates of the averaged effects of intention on health behavior, and indirect effects of RAA constructs acting through intentions, and may lead to misleading conclusions regarding the size of the effects. This has important ramifications for the
