Population.

The population of interest will be Equus caballus foals of any breed. The population will be restricted to foals under one year of age, as the equine industry rarely weans foals beyond this age [11,15]. Studies of neonate foals (<one month of age), individuals weaned at or after one year of age, non-Equus caballus species, and those on young or adult horses (≥ 1 year of age) will be excluded. Studies that assess both the target and excluded populations will be included only if data for the target population are reported separately. For this review, it will be assumed that any foal who meets the inclusion criteria will, in principle, be equally likely to be allocated to any eligible intervention.

Interventions.

All studies containing planned artificial weaning or separations of the target population will be included. Weaning in this review will be defined as the commencement of permanent nutritional and physical separation of the foal from the dam [10]. Abrupt weaning will be characterised as the immediate and complete separation of the foal from the dam without prior acclimatisation. This may occur individually (reference intervention), in pairs or in groups of foals, but without the presence of other adults. The reference intervention will be compared to alternative weaning interventions, including:

1. 1) Abrupt weaning with a nanny (adult horse): foal weaned abruptly from the dam but left in the care of another adult horse;
2. 2) Progressive removal of mares: progressive removal of 1–2 dams from an established dam-foal herd through the abrupt weaning of the foal over ≥ 2 days until all mares have been removed;
3. 3) Gradual acclimatisation: foals gradually acclimatised to separations prior to (≥ 2 weeks) the commencement of the weaning process;
4. 4) Gradual partial contact: partial contact maintained for ≥ 2 days prior to complete separation (e.g., fenceline, nutritional separation via covering udder);
5. 5) Gradual over time: increasing separation time ≥ 2 days prior to complete separation;
6. 6) Gradual over distance: increasing the distance separated over ≥ 2 days prior to complete separation; and
7. 7) Other: any other weaning intervention not covered by previous definitions.

Any additional management strategy specifically explored at the study treatment level, while keeping the primary weaning intervention constant, will be classified as a management intervention. The management interventions will be categorised as:

• Environmental (housing, conspecifics, handling)
• Temporal (age at weaning)
• Dietary (pre/post-weaning nutrition)
• Medical (pharmaceutical use).

However, when these factors are observed but not evaluated as a treatment-level comparison, they will be considered study-level characteristics and explored as subgroup or moderating factors. Additionally, experiments reporting temporary separations of the target population from their dam, with a management intervention aimed at reducing stress during the separation will also be included.

Instances where weaning is unplanned, such as due to the mare’s death or emergency health interventions, will be excluded, as will research involving the target population if the study commenced only after the foal was weaned, or the study was not related at all to weaning or separation. Observations of weaning in wild/feral equine populations will also be excluded, as they do not reflect domestic weaning practices.

Comparisons.

All studies assessing weaning will be included if pre- and post-weaning data are reported, regardless of whether a direct comparator group is included. Comparisons may include:

1. 1) Control group (unweaned) versus any form of weaning or separation
2. 2) Abrupt weaning versus any alternative weaning intervention
3. 3) Comparison between two alternative weaning interventions
4. 4) Comparison between management interventions within the same weaning intervention
5. 5) Temporary separation with a management intervention aimed at reducing stress
6. 6) Any intervention with pre- and post-intervention comparisons

Primary outcomes.

Articles will not be excluded based on the outcomes reported [51]. Primary outcomes of interest will include physiological and behavioural indicators measured by independent assessors (Table 1). Outcomes will be extracted as reported (e.g., dichotomous, continuous). Due to potential variation in behavioural definitions, these will also be extracted as reported.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Primary and secondary outcomes.

https://doi.org/10.1371/journal.pone.0352182.t001

Study design.

The review will include randomised controlled trials (RCTs), cluster-RCTs, non-randomised trials (non-RCTs), controlled before-and-after (CBA) trials, prospective and retrospective comparative cohort studies, cross-sectional, case-series, and observational studies in the narrative synthesis. Inclusion of these study designs will ensure a comprehensive assessment of the available evidence where high-quality comparative studies are limited. Narrative synthesis without meta-analysis will only be undertaken if higher-quality evidence is insufficient [52]. Questionnaires, individual case reports, narrative and scoping reviews, will be excluded, as will superseded book or chapter editions, news reports, editorials and opinion pieces. Only RCTs, cluster-RCTs, non-RCTs, and CBA trials will be included in meta-analyses.

Reporting characteristics.

There will be no limitations based on the year of publication, geographical location, publication type, or language. Articles not written in English will be translated where possible, and those that cannot be translated will be excluded from full-text screening.

Exclusion criteria hierarchy.

The reason for exclusion will be reported based on the following hierarchical order:

1. 1) Population – non-Equus caballus species
2. 2) Population – individuals weaned after > 1 year of age
3. 3) Population – neonate foals (≤ 1 month of age)
4. 4) Population – young or adult (including mares) horses (> 1 year of age)
5. 5) Intervention – unplanned weaning (e.g., mare death or emergency health intervention)
6. 6) Intervention – includes target population but data collection commenced only after the foal was weaned
7. 7) Intervention – on target population study not related to weaning or separation
8. 8) Intervention – weaning of foals in wild/feral equine populations
9. 9) Comparator – no included comparator group
10. 10) Study types – narrative and scoping reviews, questionnaires, individual case reports, superseded book or chapter editions, new reports, editorials and opinion pieces
11. 11) English translation unavailable.

Data management.

All search result citations will be imported and managed in Covidence (https://www.covidence.org/), a web-based systematic review management tool that facilitates independence and collaboration among reviewers during study selection, data extraction and risk of bias (RoB) assessment [62]. Articles that cannot be directly imported into Covidence will be imported via EndNote^TM21 (Clarivate Analytics, Philadelphia, PA, USA). Covidence automatically detects and removes duplicate records based on title, year, volume, and author(s). Any duplicate publications identified during screening will be manually checked using the same criteria and reported separately in the PRISMA flow diagram. Multiple reports of the same study will be cross-checked against authorship, start date and duration, population characteristics and sample size, location, ethics approval numbers, and registration numbers, and will be merged, with the most comprehensive report designated as the primary data source for reporting purposes.

Selection process.

The selection process for the systematic review will be documented in a PRISMA flow diagram (Fig 1). Two independent reviewers will be involved in screening, data extraction, RoB, and confidence in evidence assessments. One reviewer (NC) will be involved in all stages of the review, while the second reviewer (a review team member) will vary depending on availability and potential conflicts of interest. Blinding of article selection will be facilitated using the Covidence software. Potentially relevant studies will be identified through title and abstract screening. Reviewers will not be blinded to the article’s bibliographic fields; therefore, bias will be mitigated through the use of two independent reviewers and predefined eligibility criteria. Each record will be assessed against the eligibility criteria within Covidence, and only records with a positive or unclear response to the inclusion criteria will proceed to full-text screening. Any disagreements will be resolved through discussion or arbitrated by a third, independent reviewer (HR). Reasons for exclusion of full-text articles will be recorded in Covidence at the first clear instance of ineligibility and will be documented in a “Characteristics of Excluded Studies” table.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Flow diagram outlining the protocol review stages.

https://doi.org/10.1371/journal.pone.0352182.g001

Data collection process.

Two reviewers will independently extract data from the selected studies using a modified data extraction form within Covidence [63]. The form will be pilot tested on 5 articles to ensure all relevant data is being captured prior to full data extraction. Any changes to the extraction form will be updated and uploaded into OSF before extracting data from the remaining papers. Upon completion of data extraction, any discrepancies will be addressed through discussion and in consultation with a third independent reviewer (HR). While Covidence includes some machine learning that can extract certain types of data, all extracted data will be collected manually due to the complexity and variability in equine weaning research.

Authors of articles with insufficient information for data extraction will be contacted by email to attempt to clarify or obtain the required information. A follow-up email will be sent one month later, and authors will be given three months to provide the requested information. If requested information is not obtained and there is still insufficient information for data extraction, these articles will be excluded from data analysis and documented in the PRISMA flow diagram (Fig 1) and “Characteristics of Excluded Studies” table.

Data extraction.

Data extracted from each article will include, but not be limited to, article identification (e.g., year of publication, country, language, funding, setting, contact author details), study information and methods (e.g., study aim, main objectives, study design, data collection methods, method of randomisation, sample sizes), population (e.g., age, sex, breed), treatment interventions (e.g., intervention name and description, intervention procedure, number of participants, conspecifics, weaning housing, environment, duration), and primary outcomes (e.g., outcome name and definition, unit of measure, scale, time points, statistical method, results).

Subgroups/Moderators.

Study-level characteristics will be analysed in predefined categories as subgroup factors in pairwise meta-analyses and as moderating factors in Bayesian NMA, and will include age (early (<5 months), average (5–7 months), late (> 7 months) [43]), sex (fillies, colts), housing during weaning (stable (<25m2), yard/barn (>25m2 to 0.2 ha), paddock (>0.2ha) [64]), and conspecifics during weaning (individual, pair, > 2 individuals). Additionally, outcome measures will be collected within predefined timeframes (Table 3), to help facilitate exploration of heterogeneity and identification of potential moderating factors.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Pre- and post-weaning outcome timeframes categories.

https://doi.org/10.1371/journal.pone.0352182.t003

Effect measurements.

The primary effect measures for continuous outcomes will be the weighted mean difference (MD, 95% confidence interval (CI)) for outcomes measured on the same scale, and the standardised mean difference (SMD, 95% CI) for outcomes measured on different scales [68]. For dichotomous outcomes, the primary effect measure will be the risk ratio (RR) with 95% CI. Count data will be treated as continuous or dichotomous, depending on the outcome’s frequency. For multi-arm trials, correlations between effect estimates will be accounted for using arm-based models or variance-correction techniques to avoid unit-of-analysis errors.

Outcome measures will be collected for each intervention and study-level characteristic within predefined time frames, allowing assessment of immediate-, short- and longer-term effects [68]. When trials report multiple measurements of the same continuous outcome within a single time window, the effect estimates will be combined using inverse-variance weighting to produce a single summary value per time window, with the corresponding standard error (SE) calculated from the sum of the weights. For dichotomous outcomes, event counts and total observations from multiple measurements within each time window will be aggregated to derive a single study-level RR, 95% CI, and SE.

For pre- versus post-intervention comparisons, continuous outcomes will be calculated as the mean change (post-pre) when measured on the same scale, or as standardised mean change when measured on different scales, using the standard deviation (SD) of the change score [69]. For dichotomous outcomes, event counts and total observations will be used to calculate an RR comparing post- to pre-intervention periods. Where pre-post comparisons are not possible due to missing or incompatible baseline data, post-intervention endpoint values will be used to derive effect estimates (MD, SMD, or RR) for each time window. All effect estimates will be accompanied by their SE, serving as the summary statistics for both pairwise meta-analysis and Bayesian NMA [70].

When means and SD are not reported, the primary authors will be contacted to obtain the original data. If unavailable, means and SDs will be estimated from medians, ranges, or SE using established methods [69]. Graphically presented data will be digitally extracted. All assumptions and computations will be documented, their potential impact assessed through sensitivity analyses, and the resulting biases discussed.

Pairwise meta-analysis.

To assess the overall impact of artificial weaning on foal physiology and behaviour (objective one), pairwise random-effects meta-analysis will be used, as it is better able to account for variability between studies [69]. Analyses will be conducted using Cochrane’s RevMan (version 10.6.0; https://revman.cochrane.org) or R Statistical Software (version 4.6.0; www.r-project.org), or the version available at the time of analysis. Separate analyses will be conducted for each predefined time window. Research reporting only post-weaning scores will be excluded from SMD analyses to avoid mixing sources of variability.

Point estimates and 95% CIs will be presented in a Summary of Findings table, and conclusions will be based on consistency, direction, and magnitude of effects across studies rather than statistical P values alone [71]. Continuous outcome MDs will be reported in natural units, with positive values indicating an increase and negative values indicating a decrease in the outcome post-weaning. The relevance of these changes will depend on the nature and direction of each specific outcome. For outcomes measured using different scales, SMDs will be reported and interpreted according to Cohen’s rules (0.2 = small, 0.5 = moderate, 0.8 = large). For ease of interpretation, SMDs will, where possible, be re-expressed on a familiar scale. When both pre- and post-weaning continuous outcome values are positive and available, the ratio of means will be calculated to provide a relative measure of change. For dichotomous outcomes, both relative and absolute effects will be presented. Precise P values will be reported, but interpretation of their meaningfulness will be made in conjunction with 95% CIs and confidence in evidence ratings.

Statistical between-study heterogeneity will be assessed using the I² statistic and τ² estimates [69]. Data will be considered sufficiently homogeneous when I² is less than 50%, and τ² values are small. If data are not homogeneous, prediction intervals will be presented rather than the average intervention effects to account for the variability [72]. Sensitivity analyses comparing fixed-effect and random-effect estimates will be conducted to assess if small trials are affecting heterogeneity. Additional sources of heterogeneity will be explored using network meta-analysis [73].

Network meta-analysis.

The Bayesian NMA framework is well-suited for managing complex interventions, accounting for between-study heterogeneity, small sample sizes, sparse data, and uncertainty in effect estimates, and was therefore chosen as the method for this review [46]. A standard intervention-level NMA will be used to assess objective two, while a component-based NMA model will explore how moderating factors contribute to weaning outcomes in foals (objective three) [73]. Results from the NMA will inform the assessment and ranking of interventions and moderating factors, indicating which practices will improve foal welfare (objective four) [46].

Feasibility and network structure.

The NMA model will be used when three or more interventions can be linked through a connected evidence network and when outcomes are reported consistently across trials [46]. Studies will be eligible for inclusion if they compare at least two interventions and report sufficient aggregate-level data to calculate relative intervention effects. Connectivity will be visually assessed by mapping all available pairwise comparisons to produce a network diagram and characteristics tables [70]. Disconnected nodes will be excluded from the NMA and analysed separately using pairwise meta-analysis.

Model specification.

A Bayesian NMA will be conducted using a hierarchical random-effects framework using R Statistical Software [70]. The reference intervention was selected to provide a consistent benchmark against which all other interventions will be compared. Abrupt, individual weaning was chosen as it is a standard industry weaning practice [10,11]. Due to the likely small number of comparable studies and the absence of prior data to provide informative prior distributions, weakly informative priors will be used [70,74]. These priors are based on reasonable assumptions regarding plausible effect sizes and heterogeneity and will be used to improve model stability without strongly influencing the posterior estimates. For intervention and moderating factors, a Normal(0, 10) prior will be used, while the between-study standard deviation (τ) will be assigned a Half-Normal(0,1) prior, selected based on the likelihood of small trial sizes and sparse data sets.

Analyses will be performed using Markov Chain Monte Carlo (MCMC) simulations with three chains, each beginning with dispersed starting values to ensure thorough exploration of the posterior distributions [70]. A burn-in period of 10,000 iterations will be used to allow the chains to reach stationarity, followed by 50,000 sampling iterations to obtain posterior inferences. Both point estimates and 95% credible intervals will be reported. Convergence will be assessed using trace plots to visually inspect mixing and stability, and the Gelman-Rubin diagnostic, with values below 1.05 indicating convergence. If convergence is not satisfactory, the model will be re-run, increasing burn-in and sampling iterations by 10% to improve model stationarity until convergence is reached.

Ranking of interventions and moderating factors will be estimated using the Surface Under the Cumulative Ranking Curve (SUCRA) scores and mean ranks [46]. The results will be presented as a league table showing posterior mean effect sizes and 95% credible interval (CrI) for all pairwise comparisons, and rankograms will be generated to visualise the ranking probability of each intervention or study-level characteristic improving foal welfare (objective four).

Model fit.

The posterior mean residual deviance (D‾_res) will be used to assess model fit. The D‾_res computed during MCMC iterations will be compared with the number of data points (the number of studies multiplied by the number of interventions). A D‾_res value close to the number of data points indicates a good fit, values substantially larger suggest a lack of fit, while smaller values indicate model overfit.

Sensitivity analyses.

Sensitivity analyses will be performed to examine the impact of bias and test the robustness of the results. Sensitivity analyses will include investigating the impact of missing data, the choice of effect measures and priors, and the exclusion of outlier, non-RCT, high RoB, and low confidence in evidence studies. Models will be compared with and without subgroups to assess the contribution of study-level characteristics to heterogeneity, and alternative categorisations of study-level definitions will also be tested.

Assumptions.

The validity of this model is based on three key assumptions – heterogeneity, transitivity, and coherence [46]. Heterogeneity between direct head-to-head comparisons will be examined using I² and ‾² as described in the pairwise meta-analysis section. When heterogeneity is present, component-based NMA and sensitivity analyses will help identify important effect modifiers [70].

Transitivity assumes that the population and distribution of effect modifiers are similar across comparisons [46]. The transitivity of the foal population is met through strict population inclusion criteria. Transitivity of key effect modifiers, including intervention characteristics and study conditions, will be assessed by examining whether these effect modifiers are similarly distributed across comparisons. Where substantial differences are identified, subgroup or sensitivity analyses will be conducted to explore the impact of these differences on the network, and comparisons will be excluded if transitivity assumptions cannot be met.

Coherence is the assumption that there is agreement between direct and indirect evidence [46]. Local coherence will be assessed using the node-splitting approach, separating indirect from direct evidence, to compare estimates for each intervention. The difference between these two estimates will be considered coherent when the 95% CrI includes zero. Additionally, if the calculated Bayesian P-values are not significant (>0.05) and the CrI overlaps, local coherence for that intervention arm is further supported. Global coherence across the network will be assessed by comparing the fit of a consistency model with an inconsistency model [70]. Model fit will be evaluated using DIC values. If the inconsistency model has lower DIC values (>3 points), global inconsistency will be inferred [69]. Residual deviance and leverage plots will be used to identify influential studies and those that may cause model misfit.

Synthesis without meta-analysis.

Network meta-analysis will not be used if the network diagram indicates insufficient connectivity, the assumptions of transitivity and consistency have not been met, or outcomes are too heterogeneous [46]. In this instance, where appropriate, results will be synthesised using pairwise random-effects meta-analysis. Where direct evidence is sparse or insufficiently homogeneous, and if pairwise meta-analysis is inappropriate, data will be synthesised using Synthesis Without Meta-analysis (SWiM) methods [75]. Results will be presented in tabulated or graphical form following the SWiM reporting guidelines [52]. The decision not to proceed with any meta-analyses will be documented and justified.