Protocol and registration

The review protocol and statistical analysis plan was preregistered and is available on the Open Science Framework at the following link: https://doi.org/10.17605/OSF.IO/U38Z4. The review protocol was updated from a previously registered protocol. All the protocol revisions are documented in the supporting information.

Eligibility criteria

Studies were included if they met the following inclusion criteria:

1. - Published in one of the following behaviour analysis journals: Analysis of Verbal Behavior, Behavior Analysis in Practice, Education and Treatment of Children, Journal of Applied Behavior Analysis, Journal of the Experimental Analysis of Behavior, The Behavior Analyst, Perspectives on Behavior Science, and The Psychological Record.
2. - Were systematic reviews or meta-analyses or combined systematic review and meta-analyses.
3. - Contained the phrase “systematic review” or “meta-analysis”/ “meta-analytic” or some variant thereof, in either the title or abstract of the study.

Studies were excluded if they were described as a critical review, narrative review, literature review, scoping review, or were a primary study reporting original data. There were no restrictions on the condition, domain, assessment, intervention or exposure being studied, or language of publication; however, we excluded reviews in which the focus was on non-human animal studies. No restrictions on publication date were specified.

In total, seven journals (Perspectives on Behavior Science was, up until 2018, titled The Behavior Analyst) were selected for inclusion as they represent all publications by the Association for Behavior Analysis International (ABAI) and the Society for the Experimental Analysis of Behavior (SEAB), with the exception of Behavior and Social Issues. We selected these publications for review as ABAI and SEAB are two of the most prominent professional membership organizations in behaviour analysis and therefore might be said to represent flagship behaviour analytic publications [16]. In addition, these journals are made available to those that are Board Certified Behavior Analysts (BCBAs), and members of the ABAI and are therefore likely to be key sources that practicing behaviour analysts use to support evidence-based decision making. Behavior and Social Issues was excluded because its primary focus on cultural systems analysis falls outside the scope of the intervention-focused literature appraised in the present review. Finally, we conducted a scoping exercise using the planned search protocol to estimate the likely yield of systematic reviews and meta-analyses. The projected volume was compatible with our screening and extraction capacity, so we adopted this set of journals as the sampling frame for the present study. The included journals had impact factors ranging from 0.8 to 3.8: Journal of Applied Behavior Analysis (JABA; IF = 3.0), Perspectives on Behavior Science (IF = 3.8), Journal of the Experimental Analysis of Behavior (JEAB; IF = 1.9), Behavior Analysis in Practice (IF = 1.7), Education and Treatment of Children (IF = 1.4), The Analysis of Verbal Behavior (IF = 0.8), and The Psychological Record (IF = 0.8).

Database searches

Searches within each of the specified journals were initially undertaken in March 2023 using APA PsycINFO via the ProQuest electronic database. The search was updated in December 2024 to capture the full publication record for 2023. We selected this date to ensure complete coverage for a full year for all journals (for the purpose of detecting trends over time). The following search string was used to identify eligible studies: publication(“Education and Treatment of Children”) AND tiab(“meta-analy*” OR “systematic review”). We adapted the search string for each journal by replacing the publication specifier with the respective journal name. Once duplicates had been removed, the list of studies was recorded on an Excel spreadsheet. To determine eligibility, two graduate-level reviewers independently reviewed the titles and abstracts in accordance with the study inclusion criteria and classified papers as include, exclude, or uncertain. The reviewers consulted the full text of the paper for all studies classified as either include or uncertain. Any disagreements were resolved via discussion with a third member of the research team.

Quality assessment

We assessed all included studies using the AMSTAR 2 and R-AMSTAR instruments. The AMSTAR 2 consists of 16 items that are evaluated with “yes” or “no” (1, 3, 5, 6, 10, 13, 14, 16); with “yes”, “partial yes”, or “no” (2, 4, 7, 8, and 9); or with “yes”, “no”, or “no meta-analysis conducted” (11, 12, and 15). A summary of each item is presented in the results; however, see the appendix for a full description of each item. Items 2, 4, 7, 9, 11, 13, 15 are deemed to be “critical” as they have a substantial impact on the confidence regarding the validity of a review and its conclusions [18]. Scoring an item as “yes” indicated that a particular review had met the item criterion.

The developers of the AMSTAR 2 explicitly recommend that individual item ratings are not combined to create an overall score [18]. We therefore employed the R-AMSTAR instrument developed by Kung et al. [19] as a means of quantifying overall methodological quality. The R-AMSTAR is based on the original AMSTAR 11-item tool; however, each of the items has been broken down into sub-criteria which a review either meets or does not meet, with some items marked “not applicable” when a meta-analysis was not conducted. We anticipated a high proportion of reviews containing SCEDs, therefore we used a version of the R-AMSTAR that is better suited to evaluating reviews and meta-analyses that include SCEDs [21]. For every review, we coded whether the study met each of the 44 criteria in the instrument (four criteria from each of the 11 items). If a study explicitly met an individual criterion, we gave it a score of 1, otherwise we assigned it 0 (e.g., for ‘no’, ‘not reported’ or ‘not applicable). We then calculated the total score by summing the scores of all criteria that were met. For reviews that included a meta-analysis, the maximum achievable score was 44, for reviews that did not include a meta-analysis the maximum achievable score was 41. For the purposes of descriptive reporting, scores were converted to a percentage of the maximum applicable score for each study, allowing comparison of overall adherence across reviews regardless of whether a meta-analysis was conducted. In the statistical analysis, this variation in the number of applicable items was accommodated within the model. Each item of the R-AMSTAR was modeled as an individual Bernoulli trial, with the differing denominators handled through the model structure rather than through score standardisation. The item-level criteria for the modified R-AMSTAR are shown in the results section.

Two researchers independently scored all the included studies with both instruments. When there was a disagreement with respect to scoring, the item was independently reviewed and determined by a third member of the research team. For both primary measures, inter-rater reliability was assessed using overall percentage agreement and Gwet’s AC1. For items where ‘Not Applicable’ was a valid response (Item 9), or where data were missing, these were treated as a distinct third category to ensure that disagreements regarding item applicability were captured in the reliability estimate. For the R-AMSTAR, overall inter-rater agreement was calculated as 81.7% and mean Gwet’s AC1 was 0.73. For the AMSTAR 2 ratings, inter-rater agreement was 88.8%, and Gwet’s AC1 was 0.86. The items that generated the most common disagreement between raters were whether the adequate level of study detail was included (Item 8) whether satisfactory risk of bias assessment was undertaken in the primary studies (Item 9)

Study characteristics

The following data were extracted for each paper where it was available: (a) author and year of publication); (b) country of institutional affiliation of all authors (c) intervention characteristics (if appropriate) (d) primary outcome; (e) the number of included studies; (f) type of experimental designs included (if appropriate); (g) participants characteristics; and (h) presence of a PRISMA statement. In addition to these items, all of which were specified in our preregistration, we also recorded whether the review, (i) covered an intervention or procedure that was applied in nature (i.e., had social significance for the recipients). Data from each paper were independently extracted by two members of the research team. Where we encountered a discrepancy between the two reviewers, a third person reviewed the item and came to a decision. We calculated total Interobserver Agreement (IOA) for items (c) through (i), by dividing number of agreements by the total number of items scored. This resulted in an overall IOA score of 82.8%. Consistent with the approach adopted in systematic reviews that report percentage agreement, this exceeded the commonly used threshold of 80% [25].

Outcome variables and planned analysis

• To determine the methodological quality of the included systematic reviews we used the AMSTAR 2 [18] to score each of the included reviews. For each of the 16 items, we provided a narrative synthesis.
• Using the AMSTAR 2, we rated the included reviews as “High”, “Moderate”, “Low” and “Critically Low” in terms of their methodological quality. We also present the number of “Critical Items” (from those eligible items) scored as “No”. Eligibility for items relevant to a meta-analysis was determined by whether the review included a meta-analysis (yes/no).
• We also calculated a score for each study using an adapted version of the Revised AMSTAR (R-AMSTAR) [19]. The adaptations include the scoring amendments detailed by Jamshidi et al. [21].

Statistical analysis

In order to examine change in scores over time we used a hierarchical logistic regression modelling approach. Regression analysis has been used in previous studies to explore changes in R-AMSTAR adherence across time [21]. Here, as specified in our preregistered analysis plan we employed a Bayesian regression framework. In a Bayesian approach, prior information about the parameters in the model (e.g., priors) are specified before the observed data are incorporated. Combining these prior parameter distributions with the observed data yields a posterior distribution for each parameter. Statistical inferences are then drawn by examining the central tendency and spread of the posterior distributions. Specifying priors allows reasonable assumptions to be incorporated into the model, constraining estimates to plausible ranges [26]. We used mildly informative, regularising priors, which encode skepticism toward extreme values and help prevent over-fitting to outliers in the data.

Each R-AMSTAR item was coded dichotomously (1 = criterion met; 0 = not met). To accommodate the variation in numbers of eligible items (e.g., meta-analysis-specific items that were not applicable), we modeled each item as one Bernoulli trial and allowed random intercepts for both study and journal which accounted for the varying number of applicable items across studies. The model specification was as follows: let i index items, j studies, and k journals where j[i] and k[i] denote the study and journal associated with item response i. With y_i ∈{0,1} and mean-centered publication year x_i,

(1)

Thus, the model includes (1 | study) and (1 | journal) random intercepts and a single population-level slope for year (i.e., average yearly improvement across studies). We used weakly informative priors: Normal(0, 3) for the intercept (allowing a wide range on the probability scale), Normal(0, 0.25) for the year effect, and Exponential (1) for the standard deviations. Before fitting to data, we examined whether our priors implied reasonable outcomes by conducting prior-only simulations (i.e., simulating from the prior predictive distribution). We also performed generative simulations to verify that the model recovered (known) parameters. Specifically, we created synthetic data sets with the same hierarchical structure (Bernoulli outcomes with random intercepts for study and journal, and a fixed effect of centered year), varying the true year slope across multiple scenarios (e.g., none/moderate/strong). We chose this analytic approach for two reasons: First, hierarchical models employ partial pooling, whereby estimates for individual journals are informed both by journal-specific data and by the overall distribution of effects across journals. This is particularly beneficial when the number of studies nested within individual journals is small, as it prevents the model from over-relying on sparse local data while still allowing journal-level variation to be estimated. Second, the Bayesian framework affords greater flexibility in handling convergence with small sample sizes, a consideration that was particularly pertinent given that, at the point of preregistration, the sample sizes nested within individual journals were not yet known.

The models were fitted using the Stan computational framework (http://mc-stan.org) via the brms package [27] accessed through the statistical computing software R [28]. Model estimation was performed using Markov Chain Monte Carlo via the No-U-Turn Sampler [29]. The models were run with four parallel chains with 2000 iterations including a 1000 iteration warm up, resulting in 4000 post-warmup samples. The quality of the models was assessed by inspecting the chains to ensure mixing, convergence and stationarity. We verified that the number of effective samples was above 200 and that Rhat values were approximately 1. The data and code used to perform all the statistical analysis is provided at the following Open Science Framework repository: https://doi.org/10.17605/OSF.IO/QRGEW.

The figures were produced using the dplyr [30] and ggplot2 [31] packages. We also used the marginaleffects package [32] to generate the marginal effect estimates. To represent the central tendency and spread of the posterior distributions we report the posterior mean estimates and their corresponding 95% credible intervals (CI).

Deviations from preregistration

We deviated from the planned analyses in two ways. Following the recommendations of Willroth and Atherton [33], we describe each deviation, its rationale, and likely consequences for the study’s conclusions.

The first deviation was the omission of the planned AMSTAR 2 regression model. We had initially intended to use both AMSTAR 2 and R-AMSTAR scores as outcome variables in separate regression analyses. Upon closer consideration of AMSTAR 2, we determined that aggregating items into a total score was inappropriate, as the instrument’s developers explicitly recommend against quantitative summation for inferential purposes [18]. This decision was methodological rather than data-contingent. The deviation is unlikely to have biased our conclusions: R-AMSTAR was explicitly designed for quantitative interpretation and has been used for this purpose in prior work [21], while the AMSTAR 2 findings are reported in full descriptively and converge with the R-AMSTAR results.

The second deviation was the omission of the planned subgroup comparison between reviews involving exclusively neurodiverse participants (autistic individuals or those with intellectual or developmental disabilities) and those involving exclusively neurotypical participants. In the final dataset, very few reviews met these criteria, rendering the planned comparison statistically uninformative. Although this deviation narrows the scope of inference, it is unlikely to have materially affected our primary conclusions regarding methodological quality and temporal trends.

Study characteristics

Of the 64 systematic reviews we identified, almost half were published in Education and Treatment in Children (43.8%). Eleven were published in the Journal of Applied Behavior Analysis (17.2%), and nine in both Behavior Analysis in Practice (14.1%) and Perspectives on Behavior Science (14.1%). The remaining 11% of reviews were published in The Psychological Record (k = 3), The Analysis of Verbal Behavior (k = 2) and Journal of the Experimental Analysis of Behavior (k = 2). The included studies were published between 1988 and 2023, with a median publication year of 2020. There was a notable increase in publications from 2018 onward, with 67.2% (k = 43) of the studies published in the period between 2018 and 2023. Authors from institutions in the USA were represented in the majority of studies (86.6%, k = 58). Canadian institutions were represented in four studies, while institutions from Wales, Australia, South Korea, and New Zealand each contributed two studies. The remaining studies included authors from institutions in Northern Ireland, Republic of Ireland, India, Belgium, and Colombia.

A wide variety of topics were covered (see supplementary materials for details of individual studies) including intervention (79.7%; k = 51) and non-intervention-focused (20.3%; k = 13) topics. An example of a non-intervention review, Odum et al. [34], reviewed delay discounting literature, whereas an intervention review example included Hawken et al., [35], who reviewed the evidence for the Check-in/Check-out intervention in schools. Of the eligible reviews, (64.1%; k = 41), involved autistic participants, or individuals with intellectual or developmental disability (this included individuals described as having a learning disability, Down Syndrome, or Pervasive Developmental Disorder); however, only 11 reviews (17.2%) comprised samples drawn exclusively from these populations. Some of the reviews did not provide information about the study participant demographics. For example, a review on publication practices for which participant information was not applicable [36], and three papers in which participant information was not explicitly reported [34,37,38].

The number of included studies within the reviews ranged from 6 to 200 (median = 23) and 25% of the studies involved a meta-analysis (k = 16). Of the research designs included in the primary studies 78.1% (k = 50) reported using single case experimental designs, and 46.9% (k = 30) included studies with some type of group-based design or analysis (e.g., Randomised controlled trial (RCT); quasi-experimental group design, within-group, correlational). Six studies (9.4%) either did not report primary study designs, reported information was unclear, or this information was not applicable given the topic of the review. PRISMA guidelines were used by 39.1% (k = 25) of the included reviews. More than three quarters of the reviews did not include any information about a funding source for the review (76.6%). Of those that mentioned a funder, 17.2% reported funding for the review and 6.2% reported no funding.

Quality assessment: AMSTAR 2

The AMSTAR 2 results are shown in Fig 2 and are included in full in the supplementary materials. The bar chart (left) shows the percentage of eligible systematic reviews that met each AMSTAR-2 assessment criterion, calculated only from reviews where the criterion was applicable. Items are ordered by agreement level, with the lowest agreement items at the top of the plot. The heatmap (right) displays individual review responses for each criterion, with studies ordered by the cumulative number of ‘Fully met’ and ‘Partially met’ responses, from the highest to lowest (from left to right). This visualisation reveals both overall patterns of methodological quality across the different items of the AMSTAR 2 for the sample of studies as a whole, but also the study-specific variations in adherence.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. AMSTAR 2 ratings for all included reviews.

Green indicates “Yes” (criterion fully met), orange indicates “Partial” (criterion partially met), red indicates “No” (criterion not met), and grey indicates “Not Applicable” (criterion not relevant for that review; for example, if no meta-analysis was performed). The items on the left are presented in order of adherence (or partial adherence) to the AMSTAR 2 from lowest (at the top) to highest. Note. PICO = participant, intervention, comparison, outcome; RoB = risk of bias; RCT = randomised controlled trial; meta = meta-analysis.

https://doi.org/10.1371/journal.pone.0350142.g002

In only three of the AMSTAR 2 items (item 5, 6, and 16) did more than 50% of the reviews fully meet the criterion. The most commonly adhered items were Item 6 (79.7%; Authors performed data extraction in duplicate) followed by Item 5 (64.1%; Authors performed study selection in duplicate) and then Item 16 (54.7%; Authors reported any potential sources of conflict of interest, including any funding they received for conducting the review) The items that performed least well were as follows: 1) the presence of a review protocol (Item 2; partially met in 3.1% of reviews); 2) listing and justification for excluded studies (Item 7, fully met criterion in 1.6% of studies); 3) Assessing Risk of Bias for studies involving RCTs (Item 12, fully met criterion in 3.1% of studies); 4) Performing a fully comprehensive search strategy (Item 4, partially met in 6.2%); and 5) Reporting on sources of funding in the primary studies (Item 10, not met in any study).

The overall methodological quality of the included studies was rated according to criteria specified in Shea et al., [18]. For a determination of high confidence in the results a review should have either no or one non-critical weakness. For a rating of moderate confidence, a review should have no critical flaws. For a high confidence rating, a review should have no more than one non-critical weakness. Finally, for a methodological quality rating of ‘critically low’, a review contains two or more critical flaws. All studies in our review received a rating of “critically low” indicating that each contained more than one critical flaw (with or without non-critical weaknesses).

The most common critical flaws across the studies were Item 2 (0%; Protocol registered before commencement of the review), Item 4 (0%; Adequacy of the literature search), Item 11 (0%; Appropriateness of meta-analytical methods; overall score from both RCTs and Non-RCTs) and Item 9 (0%; Risk of bias from individual studies being included in the review; overall score from both RCTs and non-RCTs). No reviews scored ‘yes’ on any of these items. The critical items that were adhered to most often were Item 15 (40%; Assessment of presence and likely impact of publication bias), followed by Item 13 (31.2%; Consideration of risk of bias when interpreting the results of the review), and finally Item 7 (1.6%; Justification for excluding individual studies).

Quality assessment: R-AMSTAR

We scored each review against the 44 criteria specified in the R-AMSTAR (41 for studies without a meta-analysis). Table 1 details the percentage of reviews that met each criterion on an item-by-item basis (see supplementary materials for study-level scoring). For each review we then calculated the total number of R-AMSTAR criteria that had been met and divided that by the total number of eligible items. This gave us a percentage score for each study. This study-level raw percentage data can be seen in Fig 3A.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. R-AMSTAR compliance across systematic reviews, and Bayesian hierarchical model estimates of change over time.

(A) R-AMSTAR compliance rates by publication and year. Each point represents an individual systematic review, with colours indicating the publication source; the y-axis shows the percentage of R-AMSTAR criteria met by each review. (B) Predicted probability of R-AMSTAR item adherence over time. The solid pink line shows the population-average predicted probability of meeting individual R-AMSTAR criteria as a function of mean-centred publication year, derived from a Bayesian hierarchical logistic regression model with random intercepts for studies and journals; the darker and lighter pink ribbons show the 50% and 95% credible intervals, respectively, and the open circles represent observed (raw) study-level adherence proportions. (C) Posterior distribution of the average marginal effect (AME) of a one-year increase on R-AMSTAR adherence, expressed in percentage points per year. The density represents the posterior distribution of the yearly change in probability of meeting individual R-AMSTAR criteria; the solid black line marks the mean AME (1.08 percentage points per year) and the dotted lines mark the 95% credible interval (2.5th and 97.5th percentiles of the posterior). (D) Model fit assessment: predicted versus observed R-AMSTAR adherence. Studies are ordered from lowest (left) to highest (right) observed adherence proportion; open circles represent observed study-level proportions, filled pink circles show model-predicted values, and pink vertical lines indicate 95% credible intervals around predictions. Note: BAP = Behavior Analysis in Practice; ETC = Education and Treatment of Children; JABA = Journal of Applied Behavior Analysis; JEAB = Journal of the Experimental Analysis of Behavior; POBS = Perspectives on Behavior Science; PR = Psychological Record; TAVB = The Analysis of Verbal Behavior.

https://doi.org/10.1371/journal.pone.0350142.g003

We then used the R-AMSTAR scores to answer our second research question: Has adherence to methodological and reporting standards in systematic review improved across time? The multilevel logistic regression revealed a positive temporal trend in R-AMSTAR adherence (β = 0.05, 95% CI: [0.02, 0.07]), indicating that the log-odds of meeting each criterion increased by 0.05 per year. This corresponds to an odds ratio of 1.05 (95% CI: [1.02, 1.07]), or approximately a 5% increase in the odds of compliance per year. To express this effect in more interpretable terms, the average marginal effect was 1.08 percentage points per year (95% CI: [0.59, 1.58]), meaning that, across studies, the probability of meeting R-AMSTAR criteria increased by approximately 1.08 percentage points annually. Fig 3B displays the model-predicted probability of meeting R-AMSTAR criteria as a function of publication year.

Fig 3C shows the full posterior distribution of the average marginal effect (AME). The entire posterior distribution lies above zero, which provides robust evidence for improvement in systematic review quality over the study period. We observed moderate variability between journals in their average quality levels (SD = 0.27 on the log odds scale) for journal random intercepts; however, the limited number of studies within each journal (2–28, with four journals having fewer than 5 studies) means that this estimate should be treated with caution. A greater number of studies per journal would have allowed us to better detect any systematic differences. The variation between individual studies (SD = 0.31 on the log odds scale) was comparable to or slightly larger than the variation between journals, suggesting that study-level factors may be as important as journal-level factors in determining adherence.

Model fit was assessed by comparing predicted and observed study-level R-AMSTAR adherence rates (Fig 3D). Studies are ordered by raw percentage of items met from lowest to highest. The hierarchical structure induced shrinkage toward the population mean for extreme values. The model’s tendency to predict values slightly closer to the population average than observed extreme values demonstrates the regularising effect of the hierarchical structure, which borrows information across studies and publications to improve overall prediction quality. The 95% credible intervals appropriately captured the observed variation, indicating well-calibrated model predictions.

Exploratory analysis

In the design of our preregistered model, we attempted to build in robustness to the influence of outliers. Specifically, the hierarchical structure of the model provides inherent protection against the undue influence of individual reviews or journals; partial pooling across studies and journals shrinks estimates toward the overall mean which attenuates the leverage of outliers. Nevertheless, to further interrogate the robustness of the observed temporal trend, we conducted sensitivity analyses examining: (1) whether improvement was driven by the sparse early reviews and (2) whether a non-linear temporal pattern better fit the data.

First, we re-fitted the hierarchical logistic regression model excluding reviews published before 2010 (k = 5 studies removed, leaving k = 59 post-2010 reviews). This tested whether the observed improvement was primarily driven by particularly low adherence in early reviews rather than reflecting consistent progress across the entire study period. Results: The positive temporal trend persisted when restricting to post-2010 data (average marginal effect = 0.7 (95% CI: [−0.32, 1.67]) percentage points per year. So, while the magnitude of the effect was somewhat attenuated, the direction of the effect remained consistent with the full-sample analysis, indicating that improvement was not solely attributable to early outliers.

Second, we fitted a Bayesian hierarchical Generalized Additive Model (GAM) with a smooth term for year (thin plate regression splines, k = 5 basis functions) to allow the data to reveal non-linear patterns. This addressed whether improvement was accelerating, decelerating, or otherwise non-linear over time. The average marginal effect from the GAM was 0.93 (95% CI: [0.05, 1.64]) percentage points increase per year, which was broadly consistent with the findings of the original logistic hierarchical model estimate. Model comparison using leave-one-out cross-validation (LOO-IC) revealed no meaningful difference between the linear and GAM specifications (ΔLOO = −0.2, SE = 0.5), indicating that the linear model adequately captured the temporal trend without requiring additional flexibility. These sensitivity analyses collectively demonstrate that: (1) the temporal improvement in R-AMSTAR adherence is present across the full range of years included in the study; (2) the relationship between year and adherence is genuinely linear rather than constrained by our modelling choices.

We conducted a further sensitivity analysis excluding 12 reviews that focused on non-intervention or basic research topics, retaining only those studies that evaluated interventions or procedures related to socially significant outcomes (k = 52). We maintained the same random effects structure and priors as in the primary analysis. The results were consistent with the primary analysis: the average marginal effect of publication year on R-AMSTAR adherence was 1.15 percentage points per year (95% CI [0.69, 1.62]). This suggests that the inclusion of non-intervention reviews did not substantially influence our conclusions regarding temporal trends in methodological quality.

Finally, we examined whether methodological quality differed between reviews that included a meta-analysis and those that did not. Specifically, meta-analytic status was included as a binary fixed effect in the multilevel Bernoulli model alongside publication year, retaining the same random effects structure and priors as the primary analysis. A weakly informative normal prior (mean = 0, SD = 0.5) was specified for the meta-analytic status coefficient, which served to regularise estimates toward zero while permitting credible effects of plausible magnitude. The predicted probability of meeting any given criterion was 38.9% (95% CI: 31.2–45.5%) for non-meta-analytic reviews and 46.8% (95% CI: 38.1–55.1%) for meta-analytic reviews, a difference of 8.0 percentage points (95% CI: 1.6–14.2%). This suggests that reviews incorporating a meta-analysis tend to demonstrate modestly higher methodological quality than those that do not.