Eating versus skipping breakfast has no discernible effect on obesity-related anthropometric outcomes: a systematic review and meta-analysis

Registration

Our study was registered with the PROSPERO international prospective register of systematic reviews (CRD42016033290) on 21 JAN 2016. The initial registration limited papers up to the registration date; however, because of the time between initial registration and this manuscript, the search was updated to 02 JAN 2020 (see Search and review strategy, below). Earlier versions of this work were published as abstracts for the American Society for Nutrition’s Annual Meeting and Scientific Sessions^9,10.

Inclusion and exclusion criteria

Inclusion criteria were:

Studies were excluded if:

Search and review strategy

Our first search was completed on 20 JAN 2016, the search refreshed on 26 JAN 2017, and the search finalized on 02 JAN 2020, with results from prior searches being deduplicated from subsequent searches.

In all search phases, searches were executed by using the application programming interfaces (APIs) of AltHealthWatch, CINAHL, Proquest Theses and Dissertations Global, PsycInfo, and Scopus using R (version 3.5.2). The following was used to search Scopus, with analogous search strategies adapted for the other databases:

Search results across databases were compared for duplication, including by title, abstract, and PubMed ID. Studies with titles and abstracts addressing animals that did not also include words related to human subjects were excluded programmatically. Titles and abstracts were then coded independently by at least two authors for inclusion/exclusion criteria. If both authors excluded a study for violation of any inclusion or exclusion criterion, it was excluded; if at least one did not exclude it, the paper was passed on for full text review.

Meta-analysis

All data and code used to estimate effect sizes and meta-analyses are provided as Extended data at https://doi.org/10.5281/zenodo.4041392¹². Additional details are included as comments within the code, including exact approaches to estimating each effect size within a study.

Effect sizes comparing breakfast eating versus skipping on each outcome were calculated for each study. Each effect size was calculated as a difference-in-difference in the native units of the outcome (e.g., kg for weight). Only outcomes for which there was more than one effect size were meta-analyzed: body weight, BMI, body fat percentage, fat mass, lean mass, fat free mass, adipose tissue mass, waist circumference, waist:hip ratio, fat mass index, sagittal abdominal diameter, and lean tissue mass. Lean mass, fat-free mass, and lean tissue mass were meta-analyzed together as ‘lean mass’; fat mass and adipose tissue mass were meta-analyzed together as ‘fat mass’. Total body water percentage and muscle mass are both reported only in Ogata et al.¹³; although muscle mass as an outcome was excluded, Ogata et al. also reported lean mass, which is captured in the pooled lean mass analysis.

Farshchi et al.¹⁴ reported pre and post means and standard deviations separately for each treatment period in a two-arm cross-over design. Although the unbiased estimate of the difference-in-difference was calculable from the pre and post means in each condition, the lack of information on the correlation of change within or between conditions precluded us from directly calculating the variance of the effect. We requested summaries from the authors, but the authors informed us they no longer had the raw data given that the paper was published in 2005. Thus, within-condition and between-condition correlations had to be estimated. Sievert et al.⁸ used a correlation coefficient of 0.3 for post-only values. We chose to estimate within-period change scores based on the within-condition correlation coefficients we estimated from Geliebter et al.¹⁵ because Geliebter et al. had all values needed to estimate within-condition, pre-post correlation coefficients. All correlation coefficients from Geliebter were greater than 0.99. Effect sizes were estimated for each outcome. Because Farshchi et al. reported no statistically significant results for any outcome, any statistically significant estimates were recalculated using the largest within-condition correlation that resulted in non-significant effect sizes. This approach may underestimate the variance, which would provide the study more weight in the meta-analysis; however, the leave-one-out analysis described below gives Farshchi the lowest weight possible.

Geliebter et al.¹⁵ reported three conditions: skipping, corn flakes, and oat porridge. We used the recommended method of the Cochrane Handbook, which is to “combine multiple groups that are eligible as the experimental or comparator intervention to create a single pair-wise comparison”¹⁶. Because we were interested in breakfast eating versus breakfast skipping, the two breakfast conditions were pooled together.

Leidy et al.¹⁷ also reported three conditions: skipping, a normal protein breakfast, and a high protein breakfast. We requested summaries from Leidy et al., who graciously provided us with separate group means and standard deviations for the changes. We used the recommended method of the Cochrane Handbook to combine breakfast conditions as described above.

Neumann et al.¹⁸ reported three conditions: skipping, high carbohydrate breakfast, and high protein breakfast. Again, we used the method recommended by the Cochrane Handbook to combine breakfast conditions. Neumann et al. reported individual-level data in their supplementary table. While reviewing the values in their supplement, we found some results to be implausible. We reached out to the authors, who clarified several subjects’ data. For our analysis, we used the updated values, and include the final updated dataset in our supplement.

Schlundt et al.¹⁹ reported follow-up data at 6 months, but the methods descriptions were unclear as to whether the interventions to eat or skip breakfast were continued past the 12-week intervention. Authors were contacted about this detail and for additional outcomes data at 12 weeks that were either not directly reported or reported as no significant strata (i.e., habitual breakfast eaters or skippers) or treatment effects; the authors informed us they no longer had the raw data given the study was published in 1992. We therefore chose to only use the change in body weight data from 12-weeks. Independent effect sizes were estimated for habitual breakfast eaters and habitual breakfast skippers.

Dhurandhar et al.²⁰ reported body weight for the completers-only analysis in their paper. Because they registered their study as also measuring BMI, and because of the mention of an intention to treat analysis, we contacted the authors (one of whom, DBA, is a coauthor on the present meta-analysis), who provided us with summary data. Note that they also had a third group, in which participants received no specific breakfast eating or breakfast skipping recommendations; we limited our analysis to the intention to treat analyses of the breakfast eating and breakfast skipping groups. Independent effect sizes were estimated for habitual breakfast eaters and habitual breakfast skippers.

LeCheminant et al.²¹ were contacted for estimates of change over time for data in their Table 3. The authors graciously provided estimates of change within each group for each outcome. The data used herein, as shared by the authors, differs slightly from their publication because of increased precision and because of a reporting error in which percent body fat did, in fact, have a small but non-significant increase in the no breakfast group. This error does not change the results of their study, but the corrected values are used herein.

Ogata et al.¹³, Betts et al.²², and Chowdhury et al.²³ effect sizes were calculated with routine equations.

Meta-analyses were calculated using the metafor package(version 2.1-0) in R. Each of 12 independent effects sizes (10 papers; 2 stratified by baseline habit) were included in each analysis as possible, depending on which outcomes were reported in which studies. Random effects analyses were calculated; no fixed effects analyses were calculated because design heterogeneity made the assumption of effect sizes being part of a homogenous distribution tenuous. The adjustment by Knapp and Hartung²⁴ was used given the relatively small number of effect sizes. Two effect sizes were derived from separate papers of the Bath Breakfast Project (BBP; Betts et al. and Chowdhury et al.). Because these were independent samples (normal or with obesity) we treated them as independent even though they came from the same overarching study. Similarly, although there is plausibly some correlation amongst effect sizes calculated within the habit strata in Dhurandhar et al. and Schlundt et al. by nature of being part of the same overarching study, we treated the effect sizes as independent.

Leave-one-out analysis was used as a sensitivity analysis to investigate the influence of any single study for each outcome, in which each study was omitted from the dataset at a time, and then the meta-analysis was recalculated.

Effect estimates are displayed as mean differences with 95% confidence intervals in the native units of the outcome. I² (%) and p-values for tests of heterogeneity are also reported. No multiple-comparison corrections are applied within or among outcomes. There are few effect sizes (k=12), there is substantial design heterogeneity (e.g., study length, types of breakfast, different populations), and there is statistical heterogeneity in several outcomes; therefore, funnel plot asymmetry is not presented because visual estimation of asymmetry is unreliable for small k²⁵, the test is underpowered for small k²⁶, and any association between effect size and variance may plausibly be explained by study design or other factors rather than just publication bias²⁶.

We calculated a secondary analysis comparing habitual breakfast eaters versus habitual breakfast skippers, as defined by the authors of the published papers. This subgroup analysis completed the preliminary analyses we initially published as abstracts^9,10. If baseline habits were not reported, or if authors reported that the baseline habits of the participants were mixed, the study arms were excluded. The subgroup analysis was based on the same effect size estimates generated for the primary analysis. The same functions were used in R, except for modeling the outcome variable as a function of both assigned condition and baseline breakfast habits. Additional methods can be found in the subgroup analysis document and code located in https://doi.org/10.5281/zenodo.4041392.

Risk of bias

Risk of bias was assessed independently by two investigators (MMBB/JEM for all but Ogata 2019 and MMBB/AWB for Ogata 2019) using Cochrane’s Risk of Bias Tool²⁷. Given that the interventions are obvious to participants (eating versus skipping breakfast), we only coded blinding of personnel, and readers should be aware of the risk of non-blinded interventions. We do not use the approach of assigning a binary risk of bias to an entire study (e.g., if one criterion is high risk in a study, the entire study is considered high risk); however, we provide the individual ratings for each article and readers can apply such an approach if they wish.

PRISMA diagram

The search results are shown in the PRISMA diagram in Figure 1. The results of each of the three phases of the search are shown.

Figure 1. PRISMA diagram.

Three searches were undertaken. For searches 2 and 3, the numbers in parentheses represent unique results to that search. *Several ‘papers from other sources’ were identified in prior searches, but those papers were captured by the third search.

Inclusion table

Ten papers were included with 12 effect sizes (see Table 1 for descriptions). Briefly, of the 10 studies included: six were conducted in the United States, three in the United Kingdom, and one in Japan; two were cross-over RCTs and eight were parallel arm RCTs; length ranged from 6 days to 16 weeks; five provisioned some or all foods and five were recommendations for dietary consumption; two stratified on baseline eating or skipping habits, two included only habitual breakfast eaters, three included only habitual breakfast skippers, two reported mixed baseline habits, and one did not specify baseline habits; four reported race/ethnicity of participants; four included females only, one included males only, and five included both females and males. For breakfast definitions, dietary compositions, and timing, see Table 1 and Figure 2. Breakfast definitions and timing of consumption varied amongst the studies included and ranged from highly controlled and prescribed to broad recommendations (Figure 2).

Figure 2. Schematic of breakfast versus skipping timing and patterns.

The top section outlines the patterns for the included studies; the middle section shows a few examples of studies we did not classify as eating versus skipping breakfast that are explained further in the ‘Notable Exclusions’ section and in Table 3; and the bottom is a legend for the figure. ‘Inferred eating window’ represents the times we inferred that participants were permitted or recommended to consume food as reported in the papers; ‘specified eating window’, ‘breakfast eating window’, and ‘assigned eating times’ were reported by the authors in either absolute or relative times (e.g., number of hours since waking). For more details for the included studies, see Table 1.

Meta-analyses of anthropometric outcomes

Figure 3 shows a composite forest plot that includes all meta-analyzable, obesity-related, anthropometric outcomes. In all cases, the 95% confidence intervals included the null of no differences between skipping and eating breakfast (frequently interpreted as “not statistically significant”). Table 2 shows the numerical estimates of the values displayed in the forest plots. Therefore, no discernible effects of breakfast eating or breakfast skipping on body weight (kg), BMI (kg/m²), body fat percentage (%), fat mass (kg), lean mass (kg), waist (cm), waist:hip ratio, sagittal abdominal diameter (cm) and fat mass index (kg/m²) were found in these primary analyses.

In the secondary analysis comparing habitual breakfast eaters versus habitual breakfast skippers (as defined by the authors of the published papers), forest plots and a summary table can be found in the subgroup between group forest plots document and subgroup between group summary table located in https://doi.org/10.5281/zenodo.4041392. Briefly, there was no discernible effect of stratification of baseline habits on four of the outcomes: body weight (4 effect sizes for habitual breakfast eaters; 4 effect sizes for habitual breakfast skippers), body fat percentage (2 and 2 effect sizes, respectively), fat mass (1 and 2 effect sizes, respectively), or lean mass (1 and 2 effect sizes, respectively). Insufficient study arms existed to test differences for waist circumference, waist:hip ratio, sagittal abdominal diameter, and fat mass index. Only BMI was statistically significant (2 and 4 effect sizes, respectively). The negative estimate of –0.36[-0.65,-0.07] BMI units indicates that individuals assigned to conditions consistent with their baseline habits had lower values than those assigned opposite of their habits. That is, habitual breakfast skippers had lower BMI when skipping breakfast and habitual breakfast eaters had lower BMI when eating breakfast, compared to skippers eating and eaters skipping.

Figure 3. Composite forest plot of seven meta-analyzable anthropometric outcomes.

Sagittal abdominal diameter and fat mass index were only included in the two papers from the Bath Breakfast Project (Betts et al. and Chowdhury et al.), and are not plotted here; outcomes of muscle mass and total body water percent were only included in Ogata et al., and so no meta-analyzable estimate was possible. See Table 2 for the numerical values of these seven analyses, plus the sagittal abdominal diameter and fat mass index. Studies without point estimates and confidence intervals within an outcome indicates that the study did not report that outcome. 95% confidence intervals for individual studies and for the width of the diamond representing the summary estimate are presented. Horizontal dotted lines for the summary of the meta-analyses represents the 95% prediction interval. For the column ‘Habit’: e, habitual eaters; s, habitual skippers; u, unspecified or mixed. Values to the left of 0 indicate that the breakfast condition had a greater decrease in the outcome relative to the skipping condition. For all outcomes except lean mass, values to the left would therefore traditionally be considered 'favoring' the breakfast condition (e.g., lower body weight)

Risk of bias

Risk of bias varied by study (Figure 4). Two studies had low risk of bias across all categories: Dhurandhar 2014 and Ogata 2019¹³. Two studies, Betts 2014²² and Chowdhury 2016²³, were coded as high risk of bias for the criterion of blinding participants and personnel because the authors clearly indicated that personnel were not blinded. Given that the interventions are obvious to participants (eating versus skipping breakfast), we only focus on blinding of personnel, and readers should be aware of the risk of non-blinded interventions. On the other hand, many of the categories in the risk of bias in each study were unclear, and it is therefore uncertain whether the risk was high or low.

Figure 4. Risk of bias assessment.

Each included paper was assessed for risk of bias using the Cochrane Risk of Bias tool. Given that the interventions are obvious to participants (eating versus skipping breakfast), we only coded blinding of personnel, and readers should be aware of the risk of non-blinded interventions.

Sensitivity analysis: Leave-one-out analysis

The leave-one-out analysis is shown in Figure 5. Little difference is noted among the analyses, with substantial overlap of confidence intervals in all cases. When considering statistical significance (i.e., confidence intervals that do not include 0), leaving Farshchi et al.¹⁴ out of the analysis results in significantly greater BMI in the breakfast conditions than the skipping conditions. When Leidy et al.¹⁷ is excluded, fat mass is greater in the breakfast than the skipping conditions. Waist:hip ratio is centered on zero with no estimable confidence interval when Chowdhury et al.²³ is left out because the other three estimates are all 0.00. We reiterate that none of these summaries took multiple comparisons into account.

Figure 5. Leave-one-out analysis.

Within each column, the diamond represents the meta-analytic summary estimate when leaving out the study in a particular row. Row and column combinations without diamonds represent outcomes that are not reported for that particular study. *The waist:hip ratio had no estimable confidence interval because the three remaining estimates were all 0.00. Sagittal abdominal diameter and fat mass index were only included in the two papers from the Bath Breakfast Project (Betts et al. and Chowdhury et al.), and therefore a leave-one-out analysis would include only a single study; outcomes of muscle mass and total body water percent were only included in Ogata et al., and so a leave-one-out analysis is not possible.

Notable exclusions

Notable exclusions are located in Table 3. Broad areas to note are the lack of a skipping group for comparison to breakfast groups, intervention periods that were less than 72 hr in duration, studies that had the comparison of interest but did not measure body weight, and studies whose primary purpose did not isolate breakfast eating versus breakfast skipping, such as time restricted feeding and shift in consumption periods. Two examples of the latter include Wehrens et al.,²⁸ who shifted all meals by 5 hours (as well as not being in a randomized order), to extreme time restriction of Halberg et al.²⁹ who assigned only breakfast or dinner (Figure 2).

In this meta-analysis, our included studies were all conducted in adults/adolescents, but, as noted in Table 3, there have been several related studies conducted in children; however, none of the studies in children had a true skipping group. For instance, Rosado et al.³⁰ had a control group with no intervention, which is not equivalent to assigning children to skip breakfast. Similarly, Powell et al.³¹ did have a group that was assigned to consume a slice of orange as an attention placebo control, but again the children were not assigned to otherwise skip breakfast.

Summary

The causal effect of eating versus skipping breakfast on obesity-related anthropometric outcomes was non-significantly different from zero across body weight, BMI, body fat percentage, fat mass, lean mass, waist circumference, waist:hip ratio, sagittal abdominal diameter, and fat mass index. Our results largely match our prior analyses^9,10, as well as the analysis of body weight conducted by Sievert et al.⁸.

The choices of inclusion/exclusion criteria, adjustments, and assumptions to use when meta-analyzing data can influence the interpretation of results, so we highlight some of our choices here. Our choice of including studies greater than 72 hours in duration (with the shortest actually included being 6 days) should avoid including experiments measuring only transient differences from hydration, glycogen, or other very acute physiological responses. Furthermore, this approach allowed us to collate all relevant experiments and enables readers to reanalyze our results with their preferred study duration cutoff should the reader so choose. We chose to define breakfast eating versus breakfast skipping according to the intentions of the original study authors where possible. This was primarily done because a common recommendation is to eat breakfast to lose weight without any other qualifications. In doing so, we included studies that 1) may or may not have been on a weight loss background, 2) included a range of weight categories, and 3) included a variety of breakfast compositions, timings, and definitions. Another limitation that was well state by one of our reviewers: “Many of the studies included in this and other SR [systematic reviews] are small, and are often conducted in participants with a mixed usual breakfast habit who are not overweight or attempting to limit their energy intake or reduce their body weight.”

While we cannot rule out that there may be some statistically significant combination of studies, subgroups, splitting-versus-pooling of different breakfasts, or different imputation strategies, we note that the results are fairly consistently centered near zero. However, Sievert et al.(⁸) and Bonnet et al.(¹¹) both concluded small but statistically significant differences in favor of breakfast skipping. Each review included a different subset of studies, predominantly driven by duration of studies included. For one specific point of comparison, Sievert et al. used a different imputation strategy than we did for Farshchi et al. We estimated the correlations based on the estimates from Geliebter et al., while constraining the interval to the statistically non-significant results. Bonnet et al. did not include Farshchi et al. because it was too short (2 weeks) for their cutoff of 4 weeks.

Our leave-one-out analyses, produced only two values that became statistically significantly different in favor of skipping breakfast: BMI when Farshchi et al. was excluded. The subgroup analysis of baseline breakfast habits included few effect sizes to estimate interaction effects between baseline breakfast habits and breakfast assignment. We could test for differences in 5 of the outcomes. Of these, BMI showed a significant interaction between baseline habits and assigned condition. The statistical significance of these secondary estimates should be considered within the following contexts. In the leave-one-out analyses, the 95% confidence intervals did not differ substantially from other leave-one-out analyses. In the subgroup analyses, few effect sizes were available for comparison. There were 4 treatment effect estimates each for habitual breakfast eaters and habitual breakfast skippers for body weight, while there were only 2 and 4, respectively, for BMI. Some outcomes had only 1 and 2 effect estimates for eaters and skippers, respectively. In all cases, we did not adjust for multiple comparisons because of their exploratory nature. Given the small departure of the confidence intervals from 0, it is likely these would no longer be statistically significant if multiple comparisons were considered. Even if effects turned out to be non-zero, the 95% confidence and prediction intervals of the outcomes include effect sizes of low clinical significance, and thus further work would be needed to determine if non-zero effects are actually clinically meaningful.

Despite this relative consistency in summary effect sizes, we note that there was substantial design heterogeneity. The length of studies, for instance, varied substantially. To be confident in effects of obesity-related interventions, longer term studies are desired. However, the need for longer-term studies is often to guard against concluding that early effects (weeks to months) will result in sustained weight loss over months to years. Given the overall null findings herein, suggesting a need for longer studies would serve to test whether these relatively acute null findings reflect long-term adaptations to establishing breakfast habits. In addition, some have argued that it is not merely eating versus skipping breakfast that is important, but rather that the type of breakfast matters (c.f., Leidy et al. 2016⁷). Such an argument does not invalidate the question asked or the findings of this meta-analysis, however. If, for instance, a breakfast of a particular characteristic is what influences weight – be it fiber content, protein, energetic load, timing from waking, or others – then the question would not be whether eating versus skipping breakfast matters; rather, research would need to test the effects of that particular breakfast versus comparator groups, whether those comparator groups be different breakfasts or no breakfast at all.

We clarify that our results are limited to obesity-related anthropometric outcomes. As stated previously, “Just because breakfast consumption may not have a statistically significant effect on weight does not make breakfast a bad recommendation”⁵⁶, nor does it necessarily make it a good recommendation. Our results do not inform whether eating versus skipping breakfast is of value for blood glucose control (c.f.,⁵⁷, cardiometabolic risk factors (c.f.,^11,15), school performance (c.f.,⁵⁸, or other outcomes; nor do our results inform the effects of eating versus skipping breakfast as part of a multicomponent behavioral/intensive lifestyle intervention or time restriction paradigm (e.g., early vs late time-restricted feeding). The inference of the results of this study are limited to the broad and simplistic recommendation to eat or skip breakfast to affect anthropometric outcomes.

Underlying data

All data underlying the results are available as part of the article and no additional source data are required.

Extended data

Zenodo: Supplemental files for "Eating versus skipping breakfast has no discernible effect on obesity-related anthropometric outcomes: a systematic review and meta-analysis.". http://doi.org/10.5281/zenodo.4041392¹².

This project contains the following extended data:

Reporting guidelines

Zenodo: PRISMA checklist for "Eating versus skipping breakfast has no discernible effect on obesity-related anthropometric outcomes: a systematic review and meta-analysis". http://doi.org/10.5281/zenodo.4041392¹².

Data are available under the terms of the Creative Commons Attribution 4.0 International license (CC-BY 4.0).