Participants
Participants (N = 118, Mage = 35.7 years, SD = 11.8; 53 men, 65 women), living in London, UK, were members of the general public recruited by a market-research recruitment company. All participants received a flossing intervention, and the final 80 participants recruited received a vitamin C tablet interventionFootnote 1 (Mage = 35.1 years, SD = 11.8). The study was explained verbally and in a written information sheet. All participants provided informed, written consent.
To facilitate the investigation of habit formation rather than bolstering existing habits, inclusion criteria were: typically floss no more than twelve times a month (i.e. three times a week) at recruitment; ‘sometimes’, ‘rarely’ or ‘never’ take vitamin C tablets; ‘definitely’, ‘probably’, or ‘maybe’ willing to try to floss and take vitamin C tablets more frequently.
Design and procedure
This intervention study used a longitidinal single-group design. Behavioural repetition, habit and all self-report measures were recorded via online questionnaires every four weeks for a 16 week period, resulting in five timepoints (T0-T4). The vitamin C intervention took place at baseline (T0), and the flossing intervention four weeks post-baseline (T1). The interventions were at different timepoints to avoid competition between behaviours at the initiation phase, due to potential self-control or memory limitations [22]. As the study was investigating the impact of (unmodified) covariates on the habit formation process, a control group was unnecessary, therefore all participants received the habit formation intervention. There were home visits at T0, T1 and T4. The study procedure is outlined in Table 1. The study received institutional ethical approval.
Interventions
Participants were provided with floss and vitamin C tablets at T0. The intervention techniques are specified according to the Behaviour Change Techniques Taxonomy v1 [23].
Vitamin C intervention
The online vitamin C intervention was delivered at T0, embedded within the study questionnaire. Information was presented about the function of vitamin C, and possible benefits of vitamin C supplements (‘information on health consequences’). To encourage engagement with intervention materials, participants were asked the extent to which they think they could achieve each benefit through taking vitamin C tablets. (These responses formed the ‘perceived benefits’ variable.) Participants were instructed to record precisely when in their routines they would take vitamin C tablets (‘implementation intentions’).
In order to boost the intervention, within the T1 questionnaire participants were asked three multiple choice questions about benefits of vitamin C, then given correct answers and explanations (‘health consequences’). Also at T1, participants were asked when they take their vitamin C tablet, whether this was a good time for them to take it, and whether they wanted to try taking it at a more convenient time (‘coping planning’, ‘reviewing behavioural goals’). They were asked if they forgot to take the tablet because they could not see it, and whether they wanted to move it to a more visible place (‘restructuring physical environment’).
Flossing intervention
This occured at T1, in an individual session with the researcher, lasting 30–40 min. Participants were given an information leaflet, (also explained orally by the researcher) describing positive ‘health consequences’ and ‘social consequences’ of flossing, and instructions on how and when to floss.Footnote 2 Participants were guided in forming ‘implementation intentions’ and specified when and where to floss [24], based on their own personal routines. This was written on the leaflet, and read aloud, along with a pledge to floss every night, to establish ‘commitment’ and a ‘behavioural contract’.
Self-report measures
All measures were reported on a seven-point Likert scale (1 = strongly disagree, 7 = strongly agree) unless indicated otherwise, with the stem for each behaviour of ‘flossing my teeth in the evening’ and ‘taking a vitamin C tablet every day’.
Behaviour
At T0, participants reported baseline monthly frequency for both target behaviours. For flossing, participants were asked if they had ever flossed their teeth regularly before. At T1-T4, participants reported the number of times they had flossed in the evening, and had taken their vitamin C tablet in the past week. (Potential response options: 0–7 days.)
Habit
Habit was measured using the Self Report Behavioural Automaticity Index (SRBAI) [25], a reliable and valid subscale of the Self-Report Habit Index [26]. As the measurement was specifically of automaticity, the key component of habit, the habit concept will be indexed by “automaticity” throughout the Results section. For each behaviour, four items (e.g. ‘I do automatically’) followed the stem. (Combining all timepoints, flossing α = 0.98, vitamin C α = 0.98.) An option was added to the SRBAI (“N/A, I never floss my teeth in the evening/take vitamin C tablets”) to minimise misuse of the “neither agree nor disagree” response by participants who never perform the behaviour [27]. The “N/A” responses were assigned an automaticity score of zero, or treated as missing when there was a possibility of dormant habits (i.e., stored habit associations with the potential to elicit behaviour, but which rarely manifest in performance due to lack of exposure to associated cues) [28]. Dormant habits were deemed possible for participants who might have performed the target behaviour regularly before the intervention, which was determined from responses in the baseline questionnaires.Footnote 3 Following the intervention, participants were judged to have been re-exposed to the cues, thus giving any dormant habits the opportunity to be manifested, so after the interventions, the “N/A, I never do behaviour X” response was assigned an automaticity score of zero.
Context stability
Participants were asked whether they perform the target behaviours ‘in the same place every time’ and ‘at the same point in my routine every time’ [29]. As in the SRBAI, an “N/A, I never…” option was included to minmise mid-point responding, and was treated as missing for participants with potential dormant habits. (Flossing α = 0.94, vitamin C α = 0.94.)
Intention
Intention was measured using two items: “I aim to...” and “I intend to...”. (Flossing α = 0.92, vitamin C α = 0.91.).
Reward measures
For flossing, only the reward construct of pleasure was measured. For vitamin C, reward constructs were measured for: pleasure, intrinsic motivation, perceived utility and perceived benefits of the behaviour.
Pleasure measured how pleasant participants find the behaviour (e.g. is something I like a lot-dislike a lot). The intrinsic motivation measure was adapted from the exercise-specific BREQ-2 (Behavioural Regulation in Exercise) measure [30], and assessed identification (e.g. “…is important to me”), integration (e.g. “…is part of the way I have chosen to live my life”) and intrinsic motivation (e.g. “…is something I enjoy”). These were weighted as + 1, + 2 and + 3 respectively to calculate an overall score for autonomous motivation [30]. Perceived utility measured how generally useful participants think the behaviour is (e.g. very beneficial-very harmful). Perceived benefits (measured at T0, T1 and T4 only) measured the extent to which participants feel they could achieve specfic benefits from taking vitamin C tablets (e.g. reduction in length and severity of colds). This was measured using six items, on a 5-point Likert scale (I can definitely/definitely not achieve this). (For all constructs α > 0.79.) See Additional file 1: Appendix 1 for the full list of self-report measures.
Statistical methods
Paired t-tests assessed whether the interventions had a significant effect on behaviour and automaticity, by comparing scores at the point of intervention administration (T0 for vitamin C, T1 for flossing) with scores at T4. (The participants assigned a missing initial automaticity score due to the potential for dormant habits could not be included in the t-test comparing pre-intervention with T4 automaticity, however they were included in all other analyses. Excluding those with potential dormant habits (i.e. those who may have simply reactivated dormant habits, as opposed to forming new habits,) from the comparison of initial and final habit scores, allowed for a conservative estimate of the effect of the intervention.) Structural Equation Modelling (SEM) was used to investigate dynamic predictors of habit formation, with separate models (comprising the same basic form) constructed at each timepoint. It was not possible to test a longitudinal model of such complexity, as the assumptions required would be too unrealistic (e.g. no unmeasured confounding across time periods, and having correct model specifications for all the additional time relationships between the variables). The difficulty of meeting assumptions would be exacerbated by the necessity of mediating pathways in order to test the second hypothesis. Therefore, it was deemed statistically appropriate to conduct separate models for each timepoint. Each reward construct was tested individually (i.e. without other reward variables present in the model), to assess the first hypothesis of whether each variable affects habit formation.
The models were constructed to reflect known predictors of habit (as indexed by automaticity), and to address the hypothesis testing the mechanisms by which reward affects habit. The model is shown in Fig. 1. The basis of the model was that behaviour influences automaticity, and both automaticity and behaviour are influenced by their value at the previous timepoint. Another pathway allows behaviour to be influenced by automaticity at the previous timepoint. Pathways from reward to behaviour (and via intention to behaviour) were included to test for an effect of rewards on automaticity mediated by behaviour. An interaction term was created between reward and behaviour, and included as a predictor of automaticity, to test whether reward moderates the behaviour-automaticity relationship. A direct path from reward to automaticity was included for completeness, and to avoid an overly prescriptive model.
Intention and context stability were included, with paths to behaviour and automaticity. Interaction terms were created between behaviour and both intention and context stability, and included as predictors of automaticity. This controlling of theoretically expected covariates allows investigation of the mechanism of habit formation. The reward, intention and context stability variables were those from the same timepoint as the behaviour and automaticity outcomes, as intention should be measured close to the point of behaviour [31].
Behaviour data between T1 and T3 and context stability data was missing for 38 participants due to data collection problems, however, this was assumed to be missing at random (therefore not leading to bias in the results). Maximum Likelihood estimation was used to account for missing data. Each model’s goodness of fit was assessed using the Comparative Fit Index (CFI) and the Coefficient of Determination (CD, comparable to R-squared), with values close to one indicating a good fit.