Laboratory‐based interventions targeting food craving: A systematic review and meta‐analysis

This systematic review and meta‐analysis aimed to quantify the effects of laboratory‐based interventions targeting specific mechanisms of food craving, to identify moderators of effects, and to qualitatively summarize findings. The study was conducted and reported according to the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) guidelines. Sixty‐nine studies were included in the quantitative synthesis, and separate meta‐analyses were conducted for the outcomes self‐reported craving and objective food intake. Results show small to medium positive effects across specific craving interventions on both outcomes. Effect sizes were partly moderated by intervention type. The most effective intervention regarding food intake was in sensu cue exposure. For subjective craving, the most robust evidence was found for beneficial effects of cognitive regulation strategies (ie, reappraisal, suppression, and distraction). Results further indicate that training inhibitory control through behavioral inhibition might be more effective than approach‐avoidance training when considering its effect on subjective craving and food intake. People with external eating habits, overeating, or loss‐of‐control eating might benefit from these types of specific craving interventions. Future research should focus on long‐term effects, transferability, and effectiveness in clinical samples.


| INTRODUCTION
Craving is defined as a strong and seemingly irresistible desire to consume a particular substance, such as a specific food or group of foods. 1 Food craving not only is a core feature of bulimia nervosa (BN) and binge-eating disorder (BED) but is also found in nonclinical samples with external, emotional, or restrained eating and is related to overweight and obesity. [2][3][4][5][6] A recent meta-analysis has shown that craving predicts food intake and weight gain with a medium effect size (ES). 7 A high-powered study has furthermore shown that an early reduction in food craving is a reliable outcome predictor in pharmacological weight-loss trials. 8 Different theoretical models are useful to understand the emergence, generalization, pervasiveness, and urgency of craving. Thereby, both operant and respondent learning processes are of relevance, but also cognitive regulation and control, attentional deployment, and automatized behavior tendencies. The taste of food is known to be a primary reinforcer 9 ; thus, the rewarding properties of the intake of good-tasting food are learned by operant conditioning. Through association learning in the orbitofrontal cortex, the sight or smell of food can become a secondary reinforcer. 9 Thus, craving can be elicited by a variety of cues in the environment, which is known as cue-induced craving 10,11 and leads to a sensitization of associated brain circuits towards these stimuli. Through this process of "incentive sensitization," 10 associated cues become more salient and are processed with increased attentional resources. 12,13 These attentional biases then contribute to the maintenance of craving and the appearance of disordered eating behaviors. 14 Contrary to conditioning models, the elaborated intrusion (EI) theory of desire 15 states that externally triggered intrusive thoughts are only experienced as craving if the intrusion is cognitively elaborated. Accordingly, the genesis of craving is subdivided into two processing steps. First, intrusive thoughts are triggered by environmental cues, which are then followed by a cognitive elaboration process. This differentiation is relevant since interventions based on the EI theory of desire try to target the second processing step by interrupting the cognitive elaboration of cue-triggered craving. 15 By contrast, interventions based on learning theories intervene on the first step by interrupting stimulus-response associations.
There is still the question of why some people may be more susceptible to cue-elicited craving and incentive sensitization than others.
Here, interindividual differences in executive functions may come into play. It is known through cross-sectional and experimental studies that reduced inhibitory control is associated with higher craving, more loss of control over eating, and faster detection of food-related cues. [16][17][18] Furthermore, people with problematic eating patterns show automatic action tendencies towards food, in that they react faster in response to food stimuli than to neutral stimuli when compared with healthy controls. 19,20 Other studies have shown that people with higher craving have a comparably higher tendency to approach food cues 21,22 and this approach bias is related to increased food intake. 17 Therefore, in trying to modify craving and its effects on eating behavior, it appears to be important to target inhibitory control capacities, attentional deployment, and approach/avoidance behavior.
On the basis of these etiological models of craving, specific interventions addressing these different mechanisms of action have been developed to target craving. The first group of interventions tackling food craving appearing in the scientific literature involved cue exposure with response prevention, 11 which is based on the classical conditioning learning theory. Learned associations between stimulus (eg, sight of food) and response (food intake) are sought to be uncoupled by extinction learning. Another group of interventions implement topdown control over craving through cognitive regulation strategies, including reappraisal, acceptance, suppression distraction, or imaginative techniques. 15 A way to target the acquired salience of food stimuli is attentional bias modification (ABM), in which participants are implicitly taught to center their attention not on food stimuli, but rather on some kind of neutral stimulus, which is usually realized through modified dot-probe or antisaccade tasks. ABM tends to reverse incentive sensitization of food stimuli by reducing selective attentional processing of food cues; it usually targets very early attentional processes that may not easily be modified using top-down control strategies. Further interventions targeting executive functions include training to increase inhibitory control. This group of interventions often uses modified Go/No-Go or Stop-Signal tasks in order to train motor inhibitory control when confronted with food stimuli.
Finally, neurofeedback and biofeedback training may increase different facets of self-regulation 23 and thus be helpful for top-down control of craving.
In all these domains, a range of studies has been published, but there has been no systematic review or meta-analysis with regard to the effectiveness of interventions targeting food craving. While we know from recent meta-analyses 24,25 that there are efficacious treatments for craving-related eating disorders, these treatments typically include a wide range of different interventions. Notably, it is usually difficult to infer specific mechanisms of action from randomized controlled trials that evaluate manualized treatment programs. This is especially the case when the modulation of one specific symptom (such as craving) is the focus of interest. Therefore, the present study explicitly aimed to evaluate treatments on the basis of specific mechanisms underlying craving, rather than treatment programs including various interventions. More specifically, the aims of this article were (1) to systematically summarize research with regard to interventions targeting either mechanisms of classically conditioned cue reactivity and/or enhancing top-down control over craving through cognitive, imagery-based, or behavioral training, (2) to quantify effects of craving interventions, and (3) to investigate possible moderators related to sampling and intervention characteristics.

| METHOD
This systematic review and meta-analysis was planned, conducted, and reported in accordance with the evidence-based Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. 26

| Search strategy and study selection
Relevant studies were identified using the following electronic data-  Figure S1] of the study selection process).

| Data extraction and preparation
The following data were extracted from the studies and summarized into tables: first author and year of publication, sample size, participant characteristics (Table S1: diagnoses, eating behavior, age, gender, and body mass index), intervention (type, number of sessions, and type of control intervention), means (M) and standard deviations (SD) of the primary outcomes food intake (Table S2), and craving (Table S3). For food craving, type of self-report measure (ie, visual analogue scale/Likert scale asking for either "craving" or "desire to eat," Food Craving (Chocolate) Questionnaire-State or Trait, 32 and number of sessions. For the outcome craving, type of self-report questionnaire was included post hoc as an additional possible predictor for ES differences* (see the supporting information for more details regarding data extraction).

| Risk of bias and quality assessment
Assessment of the quality of each study was carried out using items A-F of the Effective Public Health Practice Project Quality Assessment Tool (EPHPP 36,37 ), along with two additional criteria regarding data analysis: a priori power analysis and quality/completeness of statistical data reporting. The following criteria were rated: selection bias, study design, confounders (differences between groups prior to intervention), blinding, collection methods, and dropout according to the EPHPP Dictionary. Studies received a total score of "strong" if there was no "weak" rating for any of these criteria, "moderate" with one "weak" rating, and "weak" if there were two or more "weak" component ratings. Statistical data reporting was rated as "complete" if the M and SD or standard errors were reported, as "sufficient" where available from graphs or sent upon request, and as "insufficient" if there were no data available for ES calculation and no response was received from the authors. Power analysis was rated with regard to whether it was done "a priori," "post hoc," or not done ("none").

| Statistical data analysis
Statistical analyses were computed with RStudio 39 version 1.2.5001 using the "compute.es" 40 and "metafor" 41 packages. Given the expected ES variation between studies, the meta-analysis was based on a random effects model using Hedges' g. Estimated mean ES, its 95% confidence intervals (CIs), and its 95% prediction intervals (PIs) were calculated and tested for significance for both of the outcome variables separately. Assessment of inconsistency was carried out by calculating the I 2 value and its CI; this index of heterogeneity ranges from 0% to 100%, indicating low (25%), moderate (50%), and high (75%) heterogeneity. 42 If heterogeneity was moderate or high, metaregressions were calculated for the dummy coded ("psych" 43 ) categorical variables study type (BS-vs WS-design), intervention type (or, if type was nonsignificant, intervention category as a superordinate *Questionnaire type was not originally included as predictor but was added during the revision process. We would like to thank the reviewer for this important suggestion of including questionnaire type as an additional predictor into our regression models. intervention categorization), control type, sample type, and questionnaire type and for the continuous variables mean age, gender (% female), and number of sessions. Meta-regression analyses were based on mixed-effects models using restricted maximum likelihood estimation. Model significance was tested using F-tests; significance of regression coefficients was tested on a t-distribution using the Knapp-Hartung 44 method that is recommended for random effects. 45,46 For categorical variables, the most frequent category was used as the reference category. Using a forward selection process, predictors explaining heterogeneity in the first step were then introduced into multiple regression models. The Akaike information criterion (AIC) was used for model selection; lower values indicate a better model fit. 46 To assess for possible publication bias, asymmetry of funnel plots was visually inspected and Egger's tests 47 were calculated. Sensitivity analyses were conducted with regard to the effects of the exclusion of outliers. Outliers were detected in accordance with Viechtbauer and Cheung 48 using the "metaoutliers" function of "altmeta" 49 ; studies are considered an outlier if the standardized residual is >3. Furthermore, adjusted ES was calculated using the trim and fill method described by Duval and Tweedie. 50 3 | RESULTS

| Risk of bias and quality assessment
Overall, individual study quality was rated as "strong" in 32 studies, "moderate" in 22 studies, and "weak" in 25 studies (see Table S4). The main reasons for low study quality were no control/report of dropout and the use of unvalidated data collection tools (eg, visual analogue or Likert scales). Some studies did not control for confounders (such as differences in state hunger between study groups) sufficiently, which was another reason for weak study quality ratings. Twenty-eight studies did not report on the statistical power of their experimental design, which may lead to bias because of insufficient power.
In the funnel plot for the outcome craving, a slight asymmetry towards higher positive effects in smaller studies was visible, but not significant (P = .26). In separate analyses for BS-and WS-design studies, no asymmetry was observed (z < −0.06, P values > .54), and the trim and fill adjusted ES did not differ from the ES estimated from published studies. This was also true for the funnel plot of studies concerning food intake: Visually, it appeared that there was an asymmetry, whereby smaller studies had higher positive ES, but Egger's test was not significant (z = 1.17, P = .24), and trim and fill analyses did not indicate an expected change in ES because of publication bias (see Figures S2-S4). only five studies with children and three studies with adolescents were found (see Table S1). First, for the effect of study type, it was found that the crossover study did not have a significantly different ES than studies with a randomized BS-design. Next, it was tested whether the potential moderators related to sample (gender, mean age, and sample type) and intervention characteristics (number of sessions, intervention type, and control type) explained any of the variance.

| Food intake
While gender, mean age, sample type, and number of sessions did not explain any of the variance, control type and intervention type had some predictive value with intervention type explaining the largest amount of variability (see Table 1   F I G U R E 1 Forest plot for the outcome food intake with random effects models for subgroups of different intervention types. Note: Subgroups were compared in a meta-regression using intervention type as a dummy coded predictor. Intervention type was a significant predictor with F 7,46 = 2.64, P < .05. The intervention type "in sensu cue exposure" shows the highest summary effect size, which is significantly higher than the reference category of "inhibitory control training" (t = 2.58, P < .05). AAT, approach-avoidance training; ABM, attentional bias modification; CE, cue exposure; CI, confidence interval; ICT, inhibitory control training; N_A, sample size active (intervention) group; N_C, sample size control (sham) group; RE, random effects in control groups and 17 crossover studies with n = 899) and five uncontrolled studies were included.

| Food craving
Of the controlled studies, six ESs were categorized to vivo cue exposure, 53 Table 2). Therefore, BS-and WS-design studies were then analyzed in separate meta-analyses.  Table 2 for statistical parameters and Figure 2 for a forest plot). In an additional (unplanned) analysis, questionnaire type was found a significant  Table 2 for statistical parameters). Taking into account intervention type as an intervening factor, heterogeneity was reduced to I 2 = 30.61% (CI, 0.00-60.07). Compared with the reference category, which was reappraisal with an estimated ES of g = 0.89, distraction had a significantly smaller effect on self-reported craving with an estimated ES of g = 0.34 (B = −0.55, t = −3.08, P < .05). However, the summary effect of distraction in comparison with control conditions was also positive and significant (P < .05). The ES for suppression, although descriptively smaller than for reappraisal, did not significantly differ from the reference category (B = −0.15, t = −0.64, P = .53).

| Sensitivity analyses
Sensitivity analyses were calculated with regard to the exclusion of outliers (see the supporting information for more details). For the food intake meta-analysis, one study was defined an outlier, 85 exclusion of this study led to a slightly reduced but still significant ES for food intake. For WS-design studies with the outcome craving, there was an outlier in the reappraisal subgroup, 95 which reduced the ES of reappraisal as intervention type and of the summary ES of this meta-T A B L E 1 Statistical parameters for meta-regressions concerning the outcome food intake Since intervention type was found a significant predictor, each intervention type was compared with the reference category "inhibitory control training." The intervention type "in sensu cue exposure" showed the highest summary effect size, which was significantly higher than the reference category "inhibitory control training" (t = 2.58, P < .05). All other intervention types did not significantly differ from the reference category (ts < 1.67, Ps > .10). e Control type was coded as follows: 1, no intervention/waiting list control; 2, control/sham intervention; 3, active control intervention (with an attenuating effect on craving); and 4, increase intervention (active intervention with a craving increasing effect).
analysis. However, overall exclusion of outliers yielded only slightly smaller values, which would not change the interpretation of effects.

| Qualitative synthesis of studies excluded from the main analysis
With regard to food intake, one study showed a positive effect of a 2-week ICT for participants with overweight or obesity. 120 In a second study, where cue exposure training was compared with appetite awareness training, overweight children ate less snack food after eight sessions of cue exposure than after appetite awareness training. 121 A study investigating the effects of distraction or acceptance on self-reported craving did not find the intervention effects to differ from the control condition. 122 The fourth study investigated the effects of a six-session neurofeedback training and found a significant change on two subscales of the Food Craving Questionnaires-State Study type was coded as between-versus within-subjects designs. Since there was a significant difference (within-subjects studies had higher effect sizes than between-subjects studies), separate meta-analyses and meta-regressions were calculated for the two study types. b Missing values were imputed by the mean (for between-subjects studies: for mean age three studies and for percent female one study out of 41; for within-subjects studies: for mean age one study out of 26).
f Control type was coded as follows: 1, no intervention/waiting list control; 2, control/sham intervention; 3, active control intervention (with an attenuating effect on craving); 4, increase intervention (active intervention with a craving increasing effect). g Type of questionnaire used for assessment of subjective craving was coded as follows: 1, state questionnaire; 2, trait questionnaire; 3, one-item measure (VAS/Likert scale); 4, mixed. Since questionnaire type was found a significant moderator, each subtype was compared with the reference category "oneitem measure." Studies using single-item measures (ie, visual analogue scales or Likert scales) had higher effect sizes than studies using state questionnaires for the assessment of food craving (t = −2.78, P < .05). Effect sizes of studies using trait questionnaires or different (mixed) measures for the assessment of craving did not significantly differ from studies using single-item measures (t < 1.02, P > .32). h The results of the moderator analysis "percent female" indicate that studies with a lower percent of females had higher effect sizes. According to the model, studies with 50% females would have a mean effect size of g = 1.12, with each 1% more female participants; effect sizes would decrease by 0.01 point.
F I G U R E 2 Forest plot for the outcome subjective food craving (between-subjects design studies) with random effects models for subgroups of different intervention types. Note: Subgroups were compared in a meta-regression using intervention type as a dummy coded predictor. However, intervention type was not a significant predictor with

| DISCUSSION
The aims of this article were to give a systematic summary of the literature regarding interventions to regulate craving in disordered and healthy eating behavior, to meta-analyze the effects of these interventions on craving and food intake, and to determine the influence of intervention-and sampling-related moderators. Separate meta-analyses were calculated for the outcomes food intake and subjective craving (craving studies were further divided into studies using WSvs BS-designs due to methodological considerations).
Overall, this meta-analysis shows small but significant overall effects of laboratory-based interventions on subjective craving and food intake. Moderator analyses show that effects partly depended on the type of intervention used, the percentage of females included, and the type of questionnaire used to assess craving. For subjective craving, WS-designs had higher ES overall than BS-designs. This may be partly explained by reduced error variance in WS-designs, but could also be because all studies with WS-designs were coded to the same intervention category, namely, cognitive regulation strategies.
Thereby, the highest effects on subjective craving were found for reappraisal and suppression; distraction led to significantly smaller effects compared with reappraisal.
With regard to food intake, the current evidence indicates that in sensu exposure to food cues might have the highest benefit; ie, imagining the consumption of high caloric food might help to eat less when exposed to the same food. Importantly, all included studies regarding in sensu exposure were conducted by the same research group 57 or were replications of these studies, 56,58 and all involved student populations whose weight and eating habits were not further defined.
Therefore, as the study by Missbach et al 56 showed that self-control resources are necessary for in sensu exposure to be effective, the effects should be investigated in people with problematic eating behavior, with weight/shape concerns, with low inhibitory control capacities and/or experiencing high craving. Another question regards ecological validity, since these studies were all conducted in controlled laboratory settings and effects might differ in settings that are more naturalistic.
F I G U R E 3 Forest plot for the outcome subjective food craving (within-subjects design studies) with random effects models for subgroups of different intervention types. Note: Subgroups were compared in a metaregression using intervention type as a dummy coded predictor. Intervention type was found a significant predictor with F 2,23 = 4.77, P < .05. The intervention type "reappraisal" shows the highest summary effect size that was the reference category and significantly higher than "distraction" (t = −3.08, P < .05); the difference to "suppression" was not significant (t = −0.64, P = .53). AAT, approach-avoidance training; ABM, attentional bias modification; CE, cue exposure; CI, confidence interval; ICT, inhibitory control training; N_A, sample size active (intervention) group; N_C, sample size control (sham) group; RE, random effects Contrary to in sensu, in vivo cue exposure did not show a significant overall effect on food intake. One recent study, which found a null effect of in vivo cue exposure on food intake, also failed to find a supportive effect of an additive ICT to cue exposure. 51  As there were quite different effects of reappraisal on subjective craving depending on study design, it is worth discussing this in detail.
In studies with WS-designs, reappraisal led to a reduction in craving as compared with a control condition, overall with a high summary ES, which was not found in BS-design studies. In the first type of studies, participants were typically instructed to either change their perspective on the food cue itself (eg, by imagining it is spoiled or by concentrating on other aspects such as color) or on their craving (eg, by changing their perspective from the short-term to long-term consequences of eating each food). These strategies seem to have positive effects on self-reported craving, as assessed directly after doing the instructed task. Concerning BS-studies, interventions coded as cognitive regulation by reappraisal had somewhat different approaches: One study used a 60-minute cognitive restructuring intervention, which contained the aforementioned strategies, but was not as specific as in the WS-studies, and craving was measured after a period of 7 days. 91  Although, in line with the PRISMA guidelines, 26 eligibility criteria and study questions were predefined in a review protocol, the protocol was not preregistered, which should be done in future studies so the research community has insight into the planning and realization of a review. Furthermore, although there were no statistically significant indicators, publication bias cannot be excluded. Visually, there was slight asymmetry in the funnel plots, which might indicate that studies with higher positive ES have a higher probability of publication than studies with negative ES. There may be unpublished data in the field, which were not considered in the current meta-analysis since only published studies were included and no attempts were made to detect unpublished studies (such as contacting researchers/professional societies in the field).
Limitations at the study level were assessed through a quality assessment tool. It showed that many studies did not control for or report dropouts, failed to control sufficiently for confounders, or used unvalidated data collection tools (eg, visual analogue or Likert scales).
The use of these single-item measures might have led to an overestimation of effects (meta-regressions show differences in ES depending on type of measurement), which is problematic for reliability of results. Craving was mostly assessed subjectively through selfreport, which may have led to socially desired answers (especially in crossover trials, where trials with and without the application of certain regulation strategies are compared, this risk is quite high). Furthermore, subjective craving does not always lead to pathological eating behavior such as binge eating. 5 Therefore, it may be expedient to use more objective measures in addition to self-report, such as salivary reaction.
Furthermore, most studies conducted to date only involved a single intervention session and tested effects on craving and/or food intake immediately afterwards. Both clearly limit the clinical implications that can be drawn from these studies regarding the benefits of the respective interventions. In order to draw more meaningful clinical conclusions, studies should investigate whether participants are able to transfer the learned strategies into their everyday life. This could be done using mobile assessment methods such as ecological momentary assessment. Research should also increasingly focus on samples with clinically relevant eating disorders or other disorders affecting the regulation of food intake. In particular, individuals with binge-eating, external, or loss-of-control eating may benefit from treatments targeting the regulation of craving. Since obesity is a growing problem globally, is associated with serious health consequences, 130 and high caloric food is omnipresent in our modern society, understanding the processes underlying the regulation of craving and developing appropriate interventions that can be easily applied and transferred to everyday life is essential.
In summary, this meta-analysis shows small but positive effects of psychological interventions on both subjective (self-reported craving) and objective (food intake) outcome measures. For food intake, the most effective intervention was in sensu cue exposure followed by reappraisal, with ICT also having a significant positive effect. For subjective craving, results differed depending on the study design, but overall cognitive regulation strategies (ie, reappraisal, suppression, and distraction) showed the most robust positive effects. Since there is evidence for a positive effect of ICT but not AAT on food intake, training inhibitory control through behavioral inhibition might be more effective than training avoidance of food stimuli.