2.4. Statistical analyses

Analyses were performed in SPSS version 27 (SPSS Inc. Chicago). Reporting followed the Strengthening the Reporting of Observational Studies in Epidemiology guidelines (STROBE) (Supplementary Table 2) (von Elm et al., 2007). The hypotheses and analysis plan were specified before the data were accessed.

Descriptive statistics (n (%) or mean ± SD) were reported for the key variables: snacking frequency at t1 and t2, snacking frequency change, anxiety severity, GAD, age, sex, ethnicity, SES, BMI, disinhibition, cognitive restraint, and flexible restraint. The scores for snacking at t1 and t2 were also split by GAD categories (GAD/non-GAD), to understand the variability in snacking scores according to the presence of clinical anxiety levels at t1.

We modelled change in snacking between the two time points via a linear model with snacking at t2 as the dependent variable, adjusted for baseline snacking at t1. Model 1 included anxiety and snacking at t1 to assess whether anxiety in the model had an impact on snacking at t2 independent of habitual snacking. Model 2 included the potential confounders of age, sex, ethnicity, SES, BMI, and overall cognitive restraint, as the general construct of restraint has been identified as a more relevant confounder than the sub-constructs of flexible and rigid restraint (Kwong et al., 2020; Magklis et al., 2019; Warne et al., 2021). Regression coefficients reported are unstandardised and residuals were inspected visually to check model assumptions.

Linear regression models (adjusted as described above) assessed whether the association between anxiety and snacking frequency was mediated by disinhibition or moderated by flexible restraint, only when a main effect was observed between anxiety severity/GAD and snacking. Using Hayes’ SPSS PROCESS macro (version 3.5) (Hayes, 2017), bias-corrected 95% CIs for all effect (unstandardised beta) coefficients were estimated using bootstrapping with 5000 samples (Bollen & Stine, 1990).

Evidence of mediation was denoted by the pathway ‘a x b’ in Fig. 4, which is the product of pathway ‘a’ (the coefficient of the association between the independent variable; anxiety or GAD, and the mediator; disinhibition) and ‘b’ (the coefficient of the association between the mediator and the dependent variable; snacking). This is seen as the ‘indirect effect’ through which anxiety affects snacking via disinhibition. If the evidence for a direct effect (c’) of anxiety on snacking was removed after the mediator was added into the model, this indicated full mediation, whereas if it remained, this indicated partial mediation (Baron & Kenny, 1986). Moderation was explored via interaction terms between anxiety and flexible restraint and GAD and flexible restraint. Anxiety and flexible restraint were mean-centred prior to creating interaction terms to aid with interpretability (Haldar, Jackard, Turrisi, & Wan, 1990; Montoya, 2019). The simple slopes procedure was used to present the moderation effect graphically, to show how the anxiety-snacking association differed between three categories of flexible restraint (low: mean -1SD, medium: mean, high: mean +1SD) (Aiken et al., 1991).

Model testing disinhibition mediating the association between Generalised Anxiety Disorder (GAD) and snacking (n = 1418).

Mediations are from linear regression models using unstandardised coefficients, adjusted for age, sex, ethnicity, socioeconomic status, BMI and cognitive restraint. c’ pathways represent the direct effect of GAD on snacking when the mediator (disinhibition) is in the model, the product of a and b (a x b) pathway represents the indirect effect where GAD is associated with snacking indirectly through disinhibition.