The PREVIEW participants and design

The recruitment and design [21], and main results [22] of the PREVIEW intervention (ClinicalTrials.gov NCT01777893) have been reported in detail previously. Adult (25–70 years) men and women with overweight (BMI ≥ 25 kg/m2) and pre-diabetes were recruited from June 2013 to February 2015 via newspaper, radio, and television advertisements and by primary and occupational health care providers. Pre-screening was conducted via telephone and potentially eligible participants (n = 5472, Supplementary Fig. 1) were invited to a screening visit to confirm pre-diabetes according to the criteria of the American Diabetes Association [23]. For more details regarding inclusion/exclusion criteria, see Fogelholm et al [21]. The intervention was conducted similarly in eight countries: Denmark, Finland, The Netherlands, the UK, Spain, Bulgaria, Australia, and New Zealand. The local Human Ethics Committees reviewed the study protocol at each of the intervention centers. All participants provided written informed consent prior to any screening procedures.

The 3-year intervention consisted of two phases (Supplementary Fig. 2). Intervention started with a 2-month weight loss phase using commercial low-energy diet products (The Cambridge Weight Plan®) to achieve daily energy intake of 3.4 MJ [24]. Because the main objectives of the intervention included weight maintenance, ≥8% weight loss was required for continuation to a 34-month weight maintenance phase. Eligible participants (n = 1857) were randomized to follow one of two intervention diets (moderate-protein, moderate-glycemic index (GI) diet aiming at 15 E% of protein, 55 E% of carbohydrate, and GI > 56 or high-protein, low-GI diet aiming at 25 E% of protein, 45 E% of carbohydrate, and GI < 50) and physical activity programs (high-intensity exercise 75 min/week or moderate-intensity exercise 150 min/week).

The behavior change intervention relied on a theory- and evidence-based PREVIEW Behavior Modification Intervention Toolbox (PREMIT) specifically designed for PREVIEW [25]. PREMIT offered a stage-based approach to behavior modification based on Transtheoretical Model [26]. The first 6 months included the active behavior change (learning new skills, frequent group visits), and the remaining 2.5 years was considered a behavior maintenance stage. The PREMIT behavior modification intervention was delivered in group visits organized throughout the intervention with decreasing frequency. Out of total 17 group visits, 10 were organized during the first 6 months (Supplementary Fig. 2). Even though the intervention was group-based, participants were guided within the limits of study diets and physical activity programs to make choices that best suited their personal preferences. For example, they were able to freely choose from variety of foods and exercise alternatives.

The analytical sample of the present study

The present analysis focused on long-term maintenance of changes in eating behavior. The analytical sample included 1311 participants, who attended at least one study visit after 6 months (during the behavior maintenance stage) and provided data on at least one eating behavior (Supplementary Fig. 1).

Participants who were excluded from the analytical sample (n = 912, Supplementary Fig. 1) were younger and had higher BMI and perceived stress levels at baseline and at 6 months (all p < 0.001, Supplementary Table 1) than participants in the analytical sample. The analysis regarding weight reduction success included 962 participants who completed the study. At baseline, completers were older and had lower BMI (both p < 0.001) than late drop-outs (n = 349), who were included in the analytical sample, but did not attend the final study visit. Their perceived stress levels were also lower (p = 0.023).

PREVIEW intervention comprised two different study diets and eating behavior may be related to the composition of diet [27]. However, we have previously reported that there was no difference between the diet groups in changes in eating behavior dimensions [28] and to aid comprehension, we have also shown it in the present study (Supplementary figure 3). Additionally, according to accelerometer data, there was no difference in total physical activity (assessed by counts per min) between the groups [22]. Hence, participants were merged into one group irrespective of original randomization in the present analysis.

Measurements

Only measurements that are relevant to the present analysis are described here. For further information, see the PREVIEW methodology paper [21]. Clinical investigation days were conducted throughout the intervention at the following time-points: baseline, and 2, 6, 12, 18, 24, 36 months. During these visits, anthropometry was performed and participants completed several questionnaires.

Eating behavior and perceived stress

Eating behavior and perceived stress were assessed using widely used and validated psychometric questionnaires: 51-item Three Factor Eating Questionnaire (TFEQ) [4] and 10-item Perceived Stress Scale (PSS) [29]. The questionnaires were self-administered and completed with computer platform during all measurement points except at 18 months.

Three Factor Eating Questionnaire (TFEQ)

Total scores for disinhibition (0–16 points) and hunger (0–14 points) were calculated. The original cognitive restraint scale of TFEQ was further divided to flexible and rigid dimensions (both 0–7 points) according to Westenhoefer et al. [17]. For all four scales, higher scores indicated higher tendency to the given eating behavior. Cronbach’s Alphas were calculated separately for each of the six time-points. For flexible restraint they ranged from 0.65 to 0.72, for rigid restraint from 0.43 to 0.55, for disinhibition from 0.77 to 0.82, and for hunger from 0.81 to 0.84.

Perceived Stress Scale (PSS)

The questionnaire contains 10 items, which are rated from 0 (never) to 4 (very often). Summary scores (range from 0 to 40 with higher scores indicating higher perceived stress) and Cronbach’s Alphas (range from 0.78 to 0.87) were calculated for each of the six time-points.

In addition to using the continuous PSS score at 6 months, we wanted to identify the participants with frequent high stress levels during the behavior maintenance stage, because it is reasonable to assume that prolonged high stress has a stronger effect on behavior. There is no established cut-off for the PSS score to screen for high stress. Hence, we identified the 20% scoring highest on the PSS at baseline, which resulted in a cut-off ≥20 for high stress. A similar relative cut-off approach has been used before [30, 31]. Frequent high stress during the intervention was defined as having high perceived stress at least two out of four measurement points between 6 months and the end of study.

Anthropometry and 3-year weight reduction success

Weight was measured at each time point in a fasting state, with an empty bladder, wearing underwear or other light clothing. A measurement was taken to the nearest 0.1 kg. Height was measured at the screening visit (before baseline) to the nearest 0.5 cm.

Total weight loss during the whole intervention was calculated as percentages of baseline weight ((3-year weight − baseline weight) / baseline weight × 100%). To facilitate the visualization and meaningful interpretation of the results, participants were categorized into three categories according to total weight reduction success after 3 years: (1) Successful, total weight loss >8%, (2) Partially successful, total weight loss 1–8%, and (3) Unsuccessful, total weight loss <1%.

Statistical methods

The descriptive data were shown as mean (SD) or n (%) unless otherwise stated. Normality of the distributions was evaluated visually from histograms. The changes in eating behavior dimensions were analyzed using linear mixed effects models with maximum likelihood estimation. This estimation method uses all available data from all participants despite missing data at some or several time-points and it is as powerful tool to handle missing data as multiple imputation [32]. Main effects were used to analyze whether predictors (perceived stress and weight reduction success) were associated with overall levels of eating behaviors. Interaction term for predictor ∗ time was added to analyze, whether the predictor was associated with change in eating behavior. Nonsignificant interaction terms were omitted from the final reported models. To control for potential confounding, the models were adjusted for fixed effects (age at the time of signing informed consent (in years), sex, intervention diet, and eating behavior and BMI at baseline) and random effects (participant ID and intervention centers). P-values for fixed effects were estimated using Satterthwaite approximation for degrees of freedom [33] and p-values for interactions were derived from ANOVA tables.

Results of mixed models are reported as beta estimates (95% confidence interval, CI). Estimated marginal means and 95% CIs were calculated to visualize the results concerning categorical predictors. Pairwise comparisons were conducted at each relevant time point with Bonferroni adjustment. Levene’s test was used to test homogeneity of variances of eating behavior and equal variances were assumed, because the test indicated similar variances in many of the time-points. To evaluate the effect sizes, standardized beta estimates were calculated for perceived stress as continuous variable, and Cohen’s ds’s were calculated for group comparisons at relevant time-points [34]. Additionally, between-group Cohen’s ds’s were calculated for change in eating behavior in selected time periods.

Statistical analyses were conducted using the statistical program R version 4.0.3 [35] with R Studio. Package lme4 was used to perform linear mixed effects analyses [36], and package lmerTest was used to obtain p-values for fixed effects [37]. The threshold for statistical significance was set at p < 0.05.