The evidential value of research on cognitive training to change food‐related biases and unhealthy eating behavior: A systematic review and p‐curve analysis

Cognitive bias modification (CBM), which retrains implicit biases towards unhealthy foods, has been proposed as a promising adjunct to improve the efficacy of weight loss interventions. We conducted a systematic review of research on three CBM approaches (i.e., cue‐specific inhibitory control, approach bias modification, and attentional bias modification) for reducing unhealthy eating biases and behavior. We performed a p‐curve analysis to determine the evidential value of this research; this method is optimally suited to clarify whether published results reflect true effects or false positives due to publication and reporting biases. When considering all CBM approaches, our results suggested that the findings of CBM trials targeting unhealthy eating are unlikely to be false positives. However, only research on attentional bias modification reached acceptable levels of power. These results suggest that CBM interventions may be an effective strategy to enhance the efficacy of weight loss interventions. However, there is room for improvement in the methodological standards of this area of research, especially increasing the statistical power can help to fully clarify the clinical potential of CBM, and determine the role of potential moderators.


Summary
Cognitive bias modification (CBM), which retrains implicit biases towards unhealthy foods, has been proposed as a promising adjunct to improve the efficacy of weight loss interventions. We conducted a systematic review of research on three CBM approaches (i.e., cue-specific inhibitory control, approach bias modification, and attentional bias modification) for reducing unhealthy eating biases and behavior. We

| INTRODUCTION
The prevalence of obesity has rapidly grown worldwide and is expected to be on the rise in the near future. 1,2 Excess weight is related to an increased likelihood of suffering a wide range of physical illnesses, as well as emotional distress and related mental health disorders. 3 Lifestyle and weight management interventions based on dietary and physical activity counseling, delivered either in isolation or in combination with pharmacological or cognitivebehavioral interventions, are the most widely used treatments for excess-weight. 4,5 For people with severe obesity, when other approaches do not lead to significant weight loss, bariatric surgery is the treatment of choice. 6 However, the success of these interventions is highly variable and there is mixed evidence regarding their long-term benefits. 4,7,8 The development of less invasive and more sustainable weight-loss interventions is thus a priority for public health systems.
Obesity is a multifactorial and heterogeneous condition, but consistent evidence suggests that overeating of energy-dense food is the main driver of the current obesity problem 1,3 (henceforth, we use the term energy-dense food to refer to highly palatable foods with excessive caloric content and poor nutrient density, including ultraprocessed foods). Energy-dense food stimulates the brain's reward circuit and may promote persistent habits. 9,10 However, current interventions for obesity do not directly target habit formation and modification mechanisms. This may be one of the reasons for their limited efficacy in the long run. In this context, cognitive bias modification (CBM) interventions have emerged as promising add-on treatments to remodel eating habits and thereby promote long-lasting changes in eating behavior. 11,12 In addition, they can be easily implemented in eHealth applications for computers and smartphones, which facilitates their tailoring to individual patients' characteristics 13 and broadens the population that can be reached in a cost-effective way. 14 An additional advantage of CBM is that they reduce the invasiveness and potential side effects of some of surgery or medication-based treatments.
CBM interventions are grounded in dual-process models which posit that choice behavior is determined by two separate, but interconnected systems. The terminology to refer to both systems varies across different models, 15 but all of them assume that one of these systems is unconscious, automatic, and impulsive, while the other is conscious, deliberate, and reflective. To ensure clarity and consistency, we will follow Strack and Deutsch's terminology 16 and refer to these systems as impulsive and reflective, respectively. While the former is fast, relatively rigid, requires no higher order cognitive resources and relies on associative processes, the latter processes information slowly, is flexible but resource-demanding, and influences decisions by weighting the value and probability of potential consequences. 16 CBM interventions primarily target the impulsive system. 12 In the domain of eating behavior, they aim to help individuals overcoming the influence of food cues that signal availability of energy-dense food.

| Cognitive bias modification interventions
Cognitive bias modification (CBM) encompasses different interventions aimed at modifying eating behavior by retraining food-related biases. 11,12 More specifically, they target two key biases, namely, "go" or approach responses. These responses are triggered by energydense food stimuli (i.e., approach/response bias) and the automatic attentional capture by such stimuli (i.e., attentional bias). CBM includes cue-specific inhibitory control (henceforth, INH), approach bias and attentional bias modification training (APP and ATT respectively). 11 INH training is intended to override response biases produced by appetitive stimuli, by extinguishing associations between energydense food stimuli and motor-response (i.e., "go") tendencies. Note the difference with general inhibitory control training, 12 which aims to improve overall response inhibition capabilities by using affectively neutral stimuli. INH training tackles the impulsive system while general inhibitory control training focuses on the reflective system. Two explanatory mechanisms have been proposed for the efficacy of INH.
On the one hand, food-related stimuli may end up producing inhibition by themselves rather than a "go" response. 17 On the other hand, such stimuli may lose incentive value. 18 Most studies on this type of training have used modified versions of the Go/No-go and Stop Signal tasks, where unhealthy food stimuli are massively paired with stopping signals to favor behavioral inhibition. 12 The goal of APP is to modify impulsive approach tendencies towards affective stimuli. In the food domain, this training usually implies either associating unhealthy food stimuli with avoidancerelated words and healthy food stimuli with approach-related words (i.e., implicit association training), 19 or practicing approach and avoidance movements in response to healthy and unhealthy visual food cues, respectively (i.e., approach and avoidance training). 20 The latter is designed to emulate the action of pulling and pushing away food as it occurs in real-life environments.
Finally, ATT aims to reduce the attention-grabbing effect of energy-dense food stimuli. 21 For this purpose, people are trained to withdraw their attention from such stimuli and directed it towards healthy foods or neutral stimuli, usually by means of a modified dotprobe task. 22 Such training is believed to produce lasting changes in attentional processing and, therefore, in affective and overt behavioral responses to environmental cues.

| Aims of the present study
Two recent systematic reviews have thoroughly examined the benefits of CBM interventions. 12,23 Both reviews concluded that CBM may be effective in modifying some automatic food-related processes (i.e., food-related biases) and clinical outcomes (e.g., weight loss), but they also raised concerns about the existence of null-findings and inconsistencies across studies. Complementarily, the most recent meta-analyses on this topic suggest that inhibitory control training (including both general and INH) and ATT may have a small but significant effects in changing eating behavior, while APP is not effective to this end 24 (see Box 1 for an overview of methodological aspects and main findings of these studies). However, several questions remain unexplored. These meta-analyses 24,25 made no distinction between cue-specific and general inhibitory control, which prevents unraveling the distinctive benefits of both interventions. Furthermore, both meta-analyses used a popular method known as "trim-and-fill" to correct for publication bias, but the results of these analyses are inconsistent. One of them suggests that there is no publication bias in research on CBM, 24 while the other one casts doubts at this respect. 25 Given that trim-and-fill returned a significant bias-corrected effect size in the meta-analysis conducted by Yang et al., 24 these authors concluded that the observed effects must be real, that is, they cannot be solely due to the selective publication of significant findings. Unfortunately, trim-and-fill is known to undercorrect for publication bias 26 and shows alarmingly high false-positive rates. 27 Given the shortcomings of this method, at least some of the conclusions of Yang et al. 24 might be premature and should be confirmed with alternative methods. The main goal of the present study is to test whether the effects of CBM interventions, both considered together and separately, are reliable using p-curve analysis, a state-of-the-art method that can handle publication bias more effectively than trim-and-fill. 26 The p-curve method has not been used in previous meta-analyses of CBM interventions, possibly because it is based on statistical information that is rarely coded in systematic reviews and meta-analyses.
Unlike standard meta-analytic methods, p-curve analysis does not rely on effect sizes but on the crucial statistic of each experiment. 28 These crucial statistics are often different from the information used to compute effect sizes, and coding them is sometimes a challenging task.
Because of this, most meta-analyses typically explore publication and reporting biases using methods (such as trim-and-fill or regressionbased methods) that do not require this information and can be directly applied to effect sizes instead. Note also that p-curve analysis is not aimed at computing a bias-corrected estimate of the average effect size (although it can be adapted for this purpose), 26 but to test the hypothesis that the statistically significant results of a set of studies are not false positives. In other words, p-curve does not replace, but rather complements the analyses typically reported in standard meta-analytic reviews.
An additional advantage of p-curve analysis over other methods is that it also returns a bias-corrected estimate of the average power of the studies, which can be used to assess whether the sample sizes of the studies are sufficiently large. Such analysis is worthwhile, since much of the available evidence on CBM comes from pilot or proof-ofconcept studies 11 that may not be sufficiently powered to detect the effects examined in this literature. Thus, our study can contribute to assess the quality of previous research on this topic.

| METHOD
We conducted a systematic search for intervention studies using different types of cognitive bias modification training (CBM) among participants with healthy or excess weight and/or dysfunctional excessive eating patterns. We updated the literature search of a previous systematic review on cognitive training and neuromodulation techniques in unhealthy eating and obesity, that is, Forcano et al. 23 We used the same search terms with respect to CBM and added the studies identified in this updated search to those that had already been examined in that systematic review. We followed the PRISMA guidelines for systematic review and meta-analysis protocols. 29

| Information sources and search strategy
The literature search was conducted in February 2021, covering the interval time since the previous review, 23 that is, in January 2017.
We examined PubMed and SCOPUS databases with the following search terms: Approach bias, Attentional bias, Cognitive bias or Response Inhibition; and Modification or Training, and Body mass index, BMI, weight, obesity, food consumption, food choice, food valuation, or food craving. Search results were assessed for inclusion. After removing duplicates, abstracts were screened and those articles which clearly did not meet the inclusion criteria were removed. Afterwards, the full-text of the remaining articles were examined. In addition, citation list from the selected articles were scrutinized for potential inclusion of further studies. This procedure was made by the first author. When there was some ambiguity in the articles to be selected, a consensus was reached between first and senior authors.

| Data collection
A key point in p-curve analysis is the selection of contrasts of interest to be included from each study. The selection should focus on the statistical contrasts that are most directly related to the main hypothesis and should also depend on the studies' design. In all cases, we followed the guidelines for p-curve analysis provided by Simonsohn, Nelson, and Simmons. 26 When two or more contrasts were eligible for the analysis, we selected the first one presented by the authors for the main analysis, and the second one for a robustness test.
P-curve analysis ignores nonsignificant results. However, to avoid introducing any bias in our selection, we selected statistical contrasts based on the predictions of the authors and the experimental designs of the studies, ignoring whether or not they turned out to be statistically significant. For exploratory studies without clearly defined hypotheses, we applied the same method. As our selection was based on the focal hypothesis made by the authors, we did not distinguish between different types of outcome measures, that is, between near transfer outcomes (i.e., response, approach, and attentional biases as directly addressed in the training protocols) and far transfer outcomes (i.e., outcomes that are intended to be indirectly modified such as eating behavior and its proximal determinants).
For the sake of transparency, the reader can find a p-curve disclosure table at https://osf.io/cdqj5/, where we describe in detail each selection and also justify any departure from the recommended guidelines. Broadly, five studies did not report sufficient information to include the key contrasts of the main hypotheses in our analysis. In those cases, we selected information from the following hypothesis (or aim, in the case of exploratory studies) that was available. In seven studies, the crucial eligible statistics were not reported for any of the main hypotheses, and thus we selected the results that, in our opinion, deviated less from the p-curve analysis guidelines (e.g., an omnibus test comparing the experimental group vs. two control groups instead of a comparison between just two groups). In the disclosure table, we also explain the rationale behind our selection of crucial contrasts for complex designs not covered in the p-curve guidelines (e.g., studies with more than three factors).

| Data analysis
All the analyses reported below were carried out with the p-curve web application (http://www.p-curve.com/app/). P-curve analysis is based on the fact that the distribution of p values is different when the null hypothesis is true and when the alternative hypothesis is true.

| Search results
The results of the updated literature search are displayed in Figure 1. The initial search yielded 636 entries. Four additional articles were identified by other sources (e.g., inspecting reference lists).
After screening the titles and abstracts of all the entries, 38 full texts were examined, but 17 were excluded because they did not meet the inclusion criteria. Thus, 21 articles were added for analyses to those already identified by Forcano et al. 23 Among the 45 selected articles, 13,17,19,20,21, 12 included multiple studies, we therefore  Table 1, while a detailed description of the main characteristics of these studies is shown in Table S1 of the Supporting Information. BMI:

| Quantitative results
Normal weight a 23 Excess weight a 6 Several BMI ranges 20

BMI nonreported 12
Specific characteristics of the samples: Frequent consumers or people who like a specific palatable food b

13
People who experience craving for a specific palatable food b 4 Uncontrolled eaters 1 Unsuccessful/restrained/chronic dieters 6 People who wished to lose weight/change dietary behavior 5 People who underwent a bariatric surgery 1 No special characteristics 31 a Studies which did not specify BMI range, where classified according BMI mean. b In 4 studies included in Forcano et al., 23 frequent consumers and people who experienced craving were also people who like to reduce their eating behavior (2 studies), who usually lose control over eating (1 study) or were restraint dieters (1 study).
F I G U R E 2 Results of p-curve analyses for all studies F I G U R E 3 Results of p-curve analyses segregate by training type. INH, cue-specific inhibitory control training; APP, approach bias modification; ATT, attentional bias modification for APP and INH power estimates were substantially lower: 36% (CI: 8%,72%) and 66% (CI: 33%,87%) for the main and robustness test of APP, respectively, and 54% (CI: 26%,77%) and 22% (CI: 5%,54%) for INH ( Figure 3, middle and lower panel). Overall, these results suggest that only research on ATT reaches acceptable power levels. Additional analyses of the distribution of effect sizes, detailed in the Supporting Information, suggest that the higher statistical power of ATT studies is probably due to the fact that, overall, the effects explored in this literature tend to be larger than those explored in studies on INH and APP.

| DISCUSSION
We performed a p-curve analysis to examine the evidential value of research on cognitive bias modification training (CBM) for foodrelated biases, which are proximal factors of unhealthy eating. 70 We Testing moderation effects, that is, two, three or four-way interactions, increases the need for larger sample sizes. 76 Our findings claim for methodological improvement of research on all CBM interventions in general, but on INH in particular.
Unlike previous findings of Yang et al. 24 and Aulbach et al., 25 our results suggest that current research on APP has evidential value.
Their analyses focused on far transfer outcomes and found no evidence of the efficacy of APP in this context. Conversely, here, we focused on the main hypotheses tested by the investigators, rather than examining intervention effects separately by outcome type. That is, we analyzed APP effects on food-related approach biases (i.e., near transfer outcomes), when the investigators focused on such hypotheses. Although, we cannot rule out the possibility that the evidential value of this research is explained by changes in near transfer outcomes, our study supports the value of further examining the direct or indirect benefits of APP on eating-related outcomes. In this regard, one of the largest studies on the effectiveness of APP to modify implicit biases in alcohol use disorder showed that changes in approach tendencies mediated the reduction of relapse rates found at 1-year follow-up. 77 Furthermore, the benefits of APP in alcohol use disorder increased as a function of the training dose. 78  At one end of that continuum, impulsivity, that is, the proneness to act without sufficient forethought, is considered a vulnerability factor that increases the likelihood of giving in to the temptation to eat energy-dense food. Attentional bias for food stimuli may spur craving, hindering the ability of impulsive individuals to reduce its influence on eating behavior. 81 At the opposite end, compulsivity, namely, feeling compelled to repeatedly act in a certain way despite being aware that such acts do not align with intended goals, 82 is a maintenance factor of overeating that implies maladaptive habit formation and difficulties in flexibly adapting behavior to avoid undesired consequences. 81 One key cognitive component of compulsivity is attentional disengagement. 83 The inability to shift attention away from food-related stimuli could be especially disadvantageous in the context of heightened attentional salience of such stimuli (i.e., attentional bias). Given current evidence suggesting that people with obesity show greater attentional bias and difficulties to disengage from food-related stimuli (see Kakoschke et al., 84 for a review of compulsive eating behavior), ATT could be especially useful to break such vicious circle. Thus, ATT may address key processes underlying impulsive and compulsive overeating, which are necessary to gain control over food-related stimuli, thereby reducing their imperative influence.
At this point, a question may arise about how to implement CBM interventions in clinical settings. In some of the largest and bestdesigned trials conducted with clinical populations (i.e., substance use disorders) to date, CBM interventions were applied as adjunctive rather than stand-alone treatments, yielding small but significant effects on recovery rates and relapse prevention. 85,86 CBM can be easily implemented within existing treatment approaches and services. They can complement dietary interventions and/or psychological therapies, which tap into reflective processes (e.g., health goals), by focusing on more implicit/impulsive processes.
In our view, based on the findings of this study, the field of CBM in eating behavior is mature enough to go one step further and adopt higher methodological standards. That is, given the existing preliminary evidence, it is time to make an effort to establish its benefits and determine its active mechanisms via gold standard methods. 79,87 In this regard, preregistering studies may help researchers develop analytic plans in advance and not deviate from them, thus decreasing the likelihood of making strong inferences on the basis of unpowered and unplanned, post-hoc exploratory analyses. In this sense, our findings align with the call by Boffo et al. 88 to adhere to strict methodological standards in research on CBM to increase its robustness and clear up their clinical applications both as add-on and stand-alone treatments.

ACKNOWLEDGMENTS
We thank Emily Giddens for his assistance in revising the English Investigador). Funding agencies had no role in the study design, collection, analysis or interpretation of the data, writing the manuscript, or the decision to submit the paper for publication.