Longitudinal tests of the theory of planned behaviour: A meta-analysis

ABSTRACT

In a meta-analysis of longitudinal analyses of the theory of planned behaviour, we tested a series of extended or auxiliary theory-consistent hypotheses: construct stability, theory predictions within and between occasions, consistency over time or stationarity in theory effects and reciprocal effects among constructs. We also tested the effects of moderators on theory effects: measurement lag, health behaviour type (protection, risk) and specific health behaviours (alcohol, dietary and physical activity). A systematic search identified 87 studies eligible for inclusion. Meta-analytic structural equation models supported construct stability and theory effects within and between occasions. Only the perceived behavioural control–intention effect exhibited stationarity. We found little evidence of reciprocal effects, and theory effects were small after accounting for reciprocal effects. We observed theory-consistent effects for the behaviour-type moderators, but no variation in model effects for the measurement lag moderators. Findings advance knowledge of the correlates of intentional behaviour and associated processes over time.

KEYWORDS:

• Social cognition theory
• attitude theory
• reciprocal effects
• stability
• panel design

Acknowledgements

We thank Daniel J. Phipps for his assistance with data collection and collation.

Author Declarations

The authors received no funding for this research.

The authors report there are no competing interests to declare.

Data availability statement

Data files, data analysis scripts and output are available online: https://osf.io/xfjq7/

Disclosure statement

No potential conflict of interest was reported by the authors.

Supplementary Material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/10463283.2023.2225897

Notes

¹ It should be noted that the type of change referred to here relates to the temporal stability of constructs and their liability to change over time due to variation in extraneous sources of information. This should be distinguished from the extent to which the constructs are subject to change through, for example, the introduction of manipulations or techniques designed to affect subsequent change in these constructs, namely their pliability.

² Note that “cause” here refers to direction in effects rather than change through the experimental manipulation or extraneous influence on one theory construct and its concomitant effect on another, see Liska (1984).

³ A spreadsheet providing full details of studies including full sample demographics, detailed description of constructs measured and target behaviours, operationalisation of the behaviour, and moderator coding is provided online: https://osf.io/xfjq7/

⁴ Other moderators were coded including likelihood of the target behaviour to be formed as a habit and the complexity of the target behaviour (see Hagger et al., 2023), but as these analyses were not directly germane to our longitudinal analysis, we have instead reported them in the online supplemental materials (see Appendix I). Other moderator variables considered were behaviour measure type (self-report vs. non-self-report) and other specific behaviours. However, studies adopting non-self-report measures of behaviour, or targeted other specific behaviours (e.g., smoking, entrepreneurialism, safe-sex behaviour, sleep-related behaviours, and learning behaviours), numbered very few (k < 10), or had empty cells in their pooled correlation matrix of study constructs, or both, precluding model estimation.

⁵ A spreadsheet providing full details of study characteristics and moderator coding is available online along with analysis scripts and output for inter-rater agreement analyses where relevant:https://osf.io/xfjq7/

⁶ One study reported collecting data on multiple occasions, some with a measurement lag of four weeks or fewer, and others with a lag of greater than four weeks (Wanberg et al., 2005). Data from this study were therefore included in both the proximal and distal groups of the measurement lag moderator.

⁷ We also coded an additional behaviour type moderator variable, categorising studies into those that targeted one-off and those that targeted repeated behaviours. This moderator almost exactly mirrored the likelihood of habit formation moderator coded for our ancillary analysis (see Appendix I) and, given the strong conceptual basis for the latter, superseded this moderator. The coding for the one-off vs. repeated behaviour moderator variable is included in the data file available online: https://osf.io/xfjq7/

⁸ One study reported within-study effect sizes for alcohol consumption, dietary behaviour, and physical activity (Norman et al., 2018). Data from this study were therefore represented in all three categories of our specific behaviour moderator variable.

⁹ Data files and analysis code and output for the inter-rater agreement analyses are available online: https://osf.io/xfjq7/

¹⁰ Correlations from the multi-level multivariate meta-analysis model unadjusted for covariates are presented in the online supplemental materials (see Table E1, Appendix E).

¹¹ Parameter estimates and variability statistics for all of the meta-analytic structural equation models unadjusted for covariates are presented in the online supplemental materials (see Appendices F, G, and K).

¹² Full results of the longitudinal model are presented in Table F1, Appendix F (online supplemental materials).

¹³ Full results of the panel model are presented in Table G1, Appendix G (online supplemental materials).

¹⁴ Results of these difference tests are available online: https://osf.io/xfjq7/

¹⁵ As ancillary analyses, we also examined effects of three other moderator variables on model effects: likelihood of habit formation (high, low), behavioural complexity (high, low), and type of behaviour (health behaviour, non-health behaviour. The latter moderator was pre-registered. Full descriptions, rationale, results, and discussion of these moderator analyses appear in Appendix I of the online supplemental materials.

¹⁶ In a closed-system cross-lagged model with perfect stabilities, variance shared between model constructs within occasions reflected in, for example, a correlation or direct effect between them, would be unchanged over time, and the model would exhibit perfect stationarity in these effects. However, where stabilities are imperfect, within occasion effects will decline in value over time – such an isolated stability model is, therefore, entropic. Cross-lagged effects between the constructs involved in the within-occasion effects, however, can serve to maintain stationarity in these effects to the extent that the cross-lags account for the shortfall in shared variance attributed to the imperfect stabilities. For further details, the reader is directed to the lucid treatment of this topic by Hertzog and Nesselroade (1987).