Only RCTs were included to allow assessment of the highest-quality evidence available. Quasi-randomised trials (using alternate allocation) were excluded.

Participants meeting the DSM (versions III-R, IV, IV-TR, 5, 5-TR) or ICD (versions 9, 10, 11) diagnostic criteria for a feeding or eating disorder were eligible for inclusion. This included those with AN, BN, BED, ARFID, pica, rumination disorder, and OSFED (formerly EDNOS). A standardised assessment of feeding and eating disorder symptomatology was necessary to ascertain diagnoses with the DSM and ICD. There were no restrictions in terms of age (child, adolescent and adult), sex or gender.

LI psychological interventions designed to treat feeding and eating disorders were included. For the purposes of this review, LI psychological interventions were defined as an intervention that (1) utilises self-help materials, (2) is 6 h or less of contact time (with each contact typically ≤ 30 min), and (3) any input is provided by practitioners or supporters who have been specifically trained to deliver the intervention [21]. The intervention had to be eating disorder-specific, and a stand-alone treatment to be included.

Studies were excluded if they evaluated LI psychological interventions integrated with another treatment, such as specialist face-to-face psychotherapy augmented with a LI psychological intervention. Studies were also excluded if the variable under experimental manipulation was not the LI psychological intervention, for example, a LI psychological intervention plus a smartphone app compared to a LI psychological intervention alone. We also excluded studies in which the LI psychological intervention was designed to prevent the onset of feeding and eating disorders. There were no restrictions on recruitment or treatment setting.

Studies comparing a LI psychological intervention against a high intensity psychological treatment, a non-eating disorder specific psychological intervention, or a waiting list control condition were included in the review. Studies comparing two types of the same LI psychological intervention through different delivery formats (e.g., bibliotherapy vs online), and provision of guidance (guided vs unguided) were excluded. We also excluded studies which used a pharmacological treatment as the comparator.

Studies were included only if they reported core eating disorder outcomes at baseline and post-intervention at a minimum. Outcomes had to be assessed with standardised, well-validated measures in order to be comparable across studies. Studies were only included in the meta-analyses if statistics allowing for effect size estimation of core eating disorder outcomes (e.g., binge eating frequency, eating disorder-related attitudes) had been reported. Rates of remission/recovery were only extracted and analysed if definitions were outlined in the original manuscripts.

The primary outcomes were as follows:

Eating disorder psychopathology Operationalised using the most global measure of eating disorder psychopathology reported in each study. The Eating Disorder Examination (EDE), in interview or self-report questionnaire (EDE-Q) format, was prioritised for this analysis [36] due to it being the most widely used measure.

The secondary outcomes were these core eating disorder outcomes at short (< 12 months) and long-term (≥ 12 months) follow-up, as well as drop-out rate and acceptability of the interventions. Qualitative results from measures of treatment acceptability were extracted where available.

The software program, Comprehensive Meta-Analysis version 3 [41], was used for computing and pooling effect sizes. In view of the considerable heterogeneity among the studies, a random effects model was adopted for all meta-analyses. Separate analyses were conducted for studies comparing against high intensity interventions, non-eating disorder-specific interventions and waiting list controls. For trials with more than one LI psychological intervention condition, effect sizes were calculated separately for each intervention.

For continuous outcomes of response (e.g., global EDE score), the effect size indicating the standardised mean difference (SMD) between the two groups at post-test (Hedges’ g) was calculated for each comparison. Hedges’ g was chosen as it adjusts for biases caused by small sample sizes [42]. A negative g favours LI psychological interventions over comparisons. SMDs were transformed into the Number Needed to Treat (NNT), using Kraemer and Kupfer’s [43] formulae. The NTT refers to the number of patients that have to be treated to achieve one additional positive outcome over a comparator. For dichotomous outcomes of response (e.g., abstinence from binge eating), the effect sizes were expressed in terms of the risk ratio (RR), otherwise known as relative risk. The RR is a ratio of the probabilities of achieving remission between two conditions. The RR was chosen because it is easier to interpret than the odds ratio [42]. An RR greater than 1 favours LI psychological interventions over comparisons. We recalculated remission and recovery rates for the intent-to-treat analyses using the number of randomised participants as the denominator of the proportion of remission/recovery; as such, remission and recovery rates in this review may differ from those reported in the original manuscripts. 95% Confidence Intervals (CIs) were calculated for each outcome. Where two or more measures were used per outcome, they were combined and the pooled effect size was calculated so that only one effect size per study was included in the analysis.

A series of subgroup analyses were performed according to the mixed effects model. In this model, studies within subgroups are pooled using a random effects model, while tests for significant differences between subgroups are conducted within the fixed effects model [44]. For continuous variables (e.g., age), meta-regression analyses were used to examine whether there was a significant relationship between the continuous variable and the effect size, as indicated by a regression coefficient (Z value) and associated p values. We aimed to explore the potential moderating effects of the following variables:

Participant age.

Statistical heterogeneity was examined using Cochran’s Q and I² statistics [45]. A significant Q statistic indicates varying effect sizes across studies as well as sample or methodological differences that may contribute to variance. The I² statistic assesses the percentage of variability due to heterogeneity rather than to random error. A value of 0% indicates no observed heterogeneity, whereas scores of 25%, 50% and 75% indicate low, moderate and high heterogeneity, respectively.

Publication bias was examined through visual inspection of a funnel plot, and by using Egger’s regression intercept to test funnel plot asymmetry [46]. We also used Duval and Tweedie's [47] trim-and-fill procedure, which estimates the number of studies that have to be removed to make the funnel plot symmetrical, and then imputes an estimated effect size after publication bias has been taken into account.

As illustrated in the PRISMA flow diagram (see Fig. 1), the search strategy yielded 16,007 articles after the removal of duplicates. Following title and abstract screening, a total of 204 full-text papers were retrieved, of which 171 were excluded because they did not meet the inclusion criteria. Thirty-three RCTs met inclusion criteria for the narrative synthesis, including one cluster-RCT and seven pilot/feasibility RCTs. Nineteen of the studies included in the current investigation were not available in the most recent meta-analyses conducted on this topic [33].Fig. 1

PRISMA flow diagram of study selection

The included studies encompassed 3665 individuals with eating disorders. The majority of studies included participants aged 18 years or older; only one study focused on adolescents (i.e., aged 12–18 years; [38]). The studies were predominantly comprised of females, with ten studies having exclusively female participants. Participant gender was not stated in one study [48]. The majority of studies focused on participants with BED (n = 15) and BN (n = 5), and one study focused on participants with AN [38]. Twelve of the studies included ‘mixed’ samples with a range of eating disorder diagnoses. No studies included participants with ARFID, pica or rumination disorder.

Across the 33 included studies, 39 LI psychological interventions to treat eating disorders were investigated. The most commonly studied treatment modality was CBT (n = 31). Other treatment modalities included Compassion-Focused Therapy (n = 1), DBT (n = 3), FBT (n = 1) and a dissonance-based program (n = 1). Two studies used a LI psychological intervention that combined elements from multiple treatment modalities, such as Compassion Attention and Regulation of Eating Behaviour [49] and Acceptance and Commitment Therapy (ACT)-influenced CBT [50]. The majority of studies delivered the LI psychological intervention with a manual or book via bibliotherapy (n = 28), nine delivered the intervention using an online platform and two studies used a CD-ROM.

Of the 33 RCTs, eight studies compared a LI psychological intervention against a high intensity psychological intervention, nine against a non-eating disorder-specific psychological intervention and 21 against a waiting list control group. High intensity therapies included group CBT (n = 2), individual CBT (n = 2), FBT (n = 1), Interpersonal Psychotherapy (IPT; n = 1), Integrative Cognitive-Affective Therapy (ICAT; n = 1), and a specialist outpatient treatment which combined CBT and IPT (n = 1). Non-eating disorder-specific psychological interventions included self-esteem unguided self-help (n = 1), perfectionism guided self-help (n = 1), behavioural weight loss guided self-help (n = 1), mindfulness-based CBT guided self-help for depression (n = 1), a self-assertion manual (n = 1), behavioural weight loss treatment (n = 1), and usual care or referral to usual care (n = 3).

Although a range of measures were considered appropriate to quantify eating disorder outcomes, most studies administered the EDE [51] or EDE-Q [36] to assess eating disorder psychopathology. Alternative outcome measures included in these analyses were the Eating Disorder Inventory (EDI; [52]) and the Binge Eating Scale (BDI; [53]). Full sample and study characteristics are outlined in Table 1.

Effect size data for the seven studies comparing against a high intensity psychological intervention can be found in Table 3 (7 comparisons). Forest plots of effect sizes on each primary outcome for studies comparing against a high intensity psychological intervention are presented in Additional file 4.1. Effect size data for each subgroup analyses are displayed in Additional file 4.2. See Additional file 4.3 for funnel plots examining publication bias.

For eating disorder psychopathology, the pooled between-group effect size (g) at post-treatment was − 0.13 (95% CI [− 0.30, 0.04], p = 0.13; NNT = 13.51), suggesting low and high intensity psychological interventions were equally efficacious at reducing eating disorder psychopathology. At short-term (< 12 months) follow-up, LI interventions were superior to high intensity interventions at reducing eating disorder psychopathology (n = 4; g = − 0.20; 95% CI [− 0.40, − 0.01], p = 0.04). No indication for publication bias was found (t = 0.56, p = 0.60).

In relation to DSM severity specifier outcomes, there was a small but significant effect in favour of LI psychological interventions when compared to high intensity therapies (g = − 0.15; 95% CI [− 0.31, 0.00], p < 0.05; NNT = 11.11). There was no significant difference between low and high intensity interventions at short-term (n = 4; g = − 0.11; 95% CI [− 0.32, 0.10], p = 0.30) or long-term (≥ 12 months) follow-up (n = 3; g = − 0.12; 95% CI [− 0.32, 0.08], p = 0.22). There was no indication for publication bias (t = 0.84, p = 0.44).

There was an overall effect in favour of high intensity therapies compared with LI interventions on achieving remission and recovery (RR = 0.70; 95% CI [0.56, 0.87], p < 0.01). This means that provision of high intensity therapies increased the chances of remission and/or recovery by around 30%. At short-term follow-up, high and low intensity interventions were comparable in achieving remission and recovery (n = 4; RR = 0.68; 95% CI [0.64, 1.01], p = 0.07). There was no indication for publication bias (t = 0.67, p = 0.55).

Effect size data for the seven studies comparing eating disorder-specific LI interventions against a non-eating disorder specific psychological intervention can be found in Table 4 (8 comparisons). Forest plots of effect sizes on each primary outcome for studies comparing against non-eating disorder specific interventions are presented in Additional file 5.1. Effect size data for each subgroup analyses are displayed in Additional file 5.2. See Additional file 5.3 for funnel plots examining publication bias.

In relation to eating disorder psychopathology, the pooled effect sizes were significantly greater for LI psychological interventions compared to non-eating disorder-specific interventions (g = − 0.35; 95% CI [− 0.49, − 0.22], p < 0.01; NNT = 5.10). These differences were no longer significant at short-term follow-up (n = 3; g = − 0.31; 95% CI [− 0.66, − 0.04], p = 0.08). No indication for publication bias was found (t = 0.42, p = 0.70).

Results also showed that LI psychological interventions had a small but significant effect on DSM severity specifier-related outcomes compared to non-eating disorder specific interventions (g = − 0.22; 95% CI [− 0.34, − 0.09], p < 0.01; NNT = 8.06), but comparable at short-term follow-up (n = 3; g = − 0.15; 95% CI [− 0.39, 0.04], p = 0.12). Visual inspection of a funnel plot indicated that the pooled effect size of studies comparing LI interventions against non-eating disorder specific interventions may have been influenced by publication bias, however Egger’s test was not significant (t = 1.87, p = 0.13). Following adjustment for missing studies using Duval and Tweedie’s (2000) trim-and-fill procedure (3 imputed studies), Hedges g was − 0.16 (95% CI − 0.26, − 0.06; NNT = 11.11).

There was an overall effect in favour of LI psychological interventions compared to non-eating disorder specific interventions on achieving remission and/or recovery (RR = 1.47; 95% CI [1.13, 1.92], p < 0.01), with those who received a LI intervention having an increased chance of remission and/or recovery of 47%. This effect increased and remained significant at short-term follow-up (n = 4; RR = 1.93; 95% CI [1.48, 0.53], p < 0.01). There was no indication for publication bias (t = 0.50, p = 0.64).

Meta-analyses were performed at the post-intervention timepoint only. It was not possible to conduct analyses at follow-up due to trials using a crossover design, nor was it possible to explore the moderating effect of ‘mode of delivery’ as all interventions were self-led. Meta-regression analyses found no significant association between age and effect size on any of the comparisons. Effect size data for the 17 studies comparing against a waiting list control condition can be found in Table 5 (22 comparisons). Forest plots of effect sizes on each primary outcome for studies comparing against waiting list controls are presented in Additional file 6.1. Effect size data for each subgroup analyses are displayed in Additional file 6.2, and funnel plots examining publication bias are in Additional file 6.3.

For eating disorder psychopathology, the pooled effect sizes were moderate, statistically significant, and in favour of the LI psychological intervention (g = − 0.68; 95% CI [− 0.90, − 0.46]; p < 0.01; NNT = 2.70). However, Cochran’s Q-test identified moderately high heterogeneity across these studies (I² = 67; Q = 42, p < 0.01). Meta-regression analyses revealed that the total RoB score had a significant effect on effect size (z = − 2.28, p = 0.02); only two studies with a waiting list condition had a total RoB score of ≤ 4. When considering moderators, there was a significant effect of ‘format of intervention’, with bibliotherapy (n = 8; g = − 0.93, 95% CI [− 1.28 to − 0.58]) superior to online (n = 5; g = − 52; 95% CI [− 0.69, − 0.35]) and CD-ROM interventions (n = 2; g = − 0.12; 95% CI [− 0.46, 0.21]). Subgroup analyses also revealed a moderating effect of ‘type of guidance’, with email guidance (n = 3; g = − 0.82; 95% CI [− 1.09, − 0.54]) more efficacious than online guidance (n = 2; g = − 0.39; 95% CI [− 0.61, − 0.16]). Visual inspection of a funnel plot indicated potential publication bias; however, Egger’s test was not significant (t = 1.97, p = 0.07) and Duval and Tweedie’s [47] trim-and-fill procedure resulted in no imputed studies.

Results showed a moderate effect in favour of LI psychological interventions on DSM severity specifier outcomes compared with waiting list (g = − 0.60; 95% CI [− 0.74, − 0.45], p < 0.01; NNT = 3.05). Meta-regression analyses revealed no significant effect of total RoB score on effect size, and no statistically significant differences among any of the subgroups investigated. A funnel plot indicated that the effect size may have been influenced by publication bias, although Egger’s test was not significant (t = 2.01, p = 0.07). Following adjustment for missing studies using Duval and Tweedie’s [47] trim-and-fill procedure (2 imputed studies), g was − 0.57 (95% CI [− 0.71, − 0.42]).

The effect of LI psychological interventions on achieving remission and/or recovery when compared to waiting list controls was RR = 3.01 (95% CI [1.93, 4.69], p < 0.01). This suggests that individuals who received a LI psychological intervention were 3× more likely to achieve remission and/or recovery than individuals waiting for treatment. However, a meta-regression analysis demonstrated that total RoB score was significantly associated with effect size (z = 1.94, p = 0.05); only one study in this comparison had a total RoB score of ≤ 4. Subgroup analyses found no significant differences between subgroups. A funnel plot indicated that the effect size was influenced by publication bias, which was confirmed by Egger's test (t = 3.02, p = 0.01). After adjusting for missing studies using Duval and Tweedie’s [47] trim-and-fill procedure (5 imputed studies), the RR reduced to 2.41 (95% CI [1.60, 3.62]).