Network Meta-Analysis for Veterinary Treatment Comparisons

By Dr. Zubair Khalid, DVM, MS, PhD · 2026-08-11

Key Takeaways

• Network meta-analysis (NMA) extends pairwise meta-analysis by integrating direct and indirect evidence to compare three or more veterinary treatments simultaneously, filling evidence gaps where head-to-head trials are absent.
• The core assumption of transitivity requires that effect modifiers (e.g., species, breed, disease severity, production system) are similarly distributed across all treatment comparisons within the network for indirect estimates to be valid.
• Consistency, a statistical property, is assessed by checking if direct and indirect estimates for the same comparison agree within statistical uncertainty; inconsistency signals potential violations of transitivity or systematic study differences, invalidating pooled indirect estimates.
• Network structure is critical; nodes represent treatments, edges represent direct comparisons, and a connected network is necessary for indirect estimation, with network geometry (density and study support per edge) influencing the precision and reliability of treatment rankings.
• Key output metrics include relative effect estimates with credible intervals and treatment rankings, such as the Surface Under the Cumulative Ranking Curve (SUCRA), which quantifies the probability of a treatment being the best, but should be interpreted alongside effect estimates and their uncertainty.
• Veterinary NMAs, exemplified by ranking 8 antimicrobials for bovine respiratory disease using 37 trials, require careful definition of populations (e.g., feedlot calves vs. dairy calves) and harmonized outcome definitions (e.g., morbidity incidence windows) to maintain transitivity and ensure valid comparisons.

Veterinary clinicians and researchers increasingly face a common dilemma: multiple treatments exist for the same condition, yet few trials compare all relevant options directly. A network meta-analysis (NMA), also called a mixed treatment comparison, extends conventional pairwise meta-analysis by combining direct and indirect evidence across a network of trials. This article explains the conceptual foundation, core assumptions, and interpretation of NMA for veterinary applications, with worked examples drawn from published analyzes in bovine respiratory disease, diabetic mouse models, and other settings. It is written for veterinary researchers who design, appraise, or apply evidence syntheses and who need to judge when an NMA is valid and what its results actually mean.

The central question an NMA answers is comparative: among several interventions, which is most likely to produce the best outcome, and with what degree of certainty? Conventional meta-analysis can only compare two treatments at a time. When a clinician must choose among five antimicrobials for metaphylaxis, or among three anticoagulant strategies, pairwise evidence alone leaves gaps. An NMA fills those gaps by using a common comparator, such as placebo or standard care, to create an indirect estimate between treatments never directly compared. The method has been applied to veterinary questions including antimicrobial metaphylaxis for bovine respiratory disease, where a mixed treatment comparison combined 37 trials and 8 antimicrobials to rank morbidity and mortality outcomes, and to preclinical animal models, where a network meta-analysis of short-chain fatty acids in diabetic mice compared five interventions for glycemic control.

This article proceeds from the logic of indirect comparison through the assumptions that make NMA valid, the structure of a network, and the interpretation of output metrics such as surface under the cumulative ranking curve (SUCRA). It closes with practical guidance for appraising published NMAs and for planning new ones. Software-specific coding is excluded, the focus is on concepts, decisions, and reporting standards that apply across platforms.

At a Glance

The Logic of Indirect Comparison

A pairwise meta-analysis pools studies that compare the same two treatments, say A and B. An NMA extends this by including studies comparing A to C and B to C, then estimates the A versus B effect indirectly as the difference between A versus C and B versus C. This indirect estimate gains precision because it uses all available data, and it permits ranking of all treatments on a common scale.

The gain in information is real but conditional. Indirect estimates inherit the validity of the direct comparisons that generate them. If the A versus C trials are biased, the indirect A versus B estimate is biased. If the patient populations differ systematically across the A versus C and B versus C trials, the indirect estimate may reflect population differences instead of treatment differences. The assumption that these populations are exchangeable is called transitivity, and it is the conceptual foundation on which all NMA rests.

Transitivity requires that the relative effect of A versus C would be the same in a population of patients eligible for a B versus C trial as it is in the population actually studied. Effect modifiers, such as disease severity, species, breed, or production system, must be similarly distributed across comparisons. In the bovine respiratory disease network, for example, transitivity would be violated if tulathromycin trials enrolled high-risk calves with severe disease while gamithromycin trials enrolled low-risk calves, because the indirect comparison would then conflate drug effect with disease severity. The published mixed treatment comparison addressed this by restricting inclusion to randomized controlled trials with defined arrival windows and standardized outcome definitions, but the assumption remains a judgment call supported by clinical reasoning and baseline data inspection.

Network Structure and Evidence Geometry

A network is a graph. Nodes represent treatments, and edges represent direct comparisons supported by at least one study. A connected network, in which every node can be reached from every other node through some path, is required for full indirect estimation. Disconnected components can sometimes be linked by a common comparator, but if no such link exists, no NMA is possible.

The geometry of the network matters for interpretation. A network with many studies on one edge and a single small study on another edge will produce precise estimates for the well-supported comparison and imprecise, heavily indirect estimates elsewhere. The bovine respiratory disease network, with 37 trials and 8 antimicrobials, had sufficient density to support ranking of morbidity outcomes, but the authors reported wider credible intervals for mortality and retreatment outcomes, reflecting sparser evidence on those edges. Readers should inspect the network graph before trusting any ranking, because a treatment can appear best simply because it is connected through a single, small, or biased study.

Assumptions: Transitivity and Consistency

Transitivity is a design assumption about the studies that enter the network. Consistency is a statistical property of the data: direct and indirect estimates for the same comparison should agree within sampling error. Inconsistency, also called incoherence, signals that transitivity has failed or that some studies are systematically different from others.

Consistency can be assessed globally with a chi-square type test comparing the fit of a consistent model to an inconsistent one, or locally by comparing direct and indirect estimates for specific comparisons. The cancer-associated thrombosis network meta-analysis provides a cautionary example: the authors observed inconsistency between pairwise and network estimates for major bleeding, with direct comparisons showing a higher bleeding rate for direct oral anticoagulants than the network estimate suggested. They reported this discrepancy transparently instead of forcing a single pooled estimate.

For veterinary readers, the practical implication is that an NMA is not a single number but a set of estimates whose validity depends on assumptions that must be checked and reported. The ARRIVE guidelines for reporting animal research and the reporting standards indexed by the EQUATOR Network both emphasize transparent reporting of methods and assumptions, and the PRISMA extension for NMAs specifies what a published analysis should include.

Planning a Network Meta-Analysis in Veterinary Settings

The first decision is whether a network meta-analysis is warranted at all. A conventional pairwise meta-analysis remains the correct choice when the clinical question involves only two interventions and direct trials are plentiful. A network meta-analysis earns its additional complexity when three or more treatments compete for the same indication and the evidence base contains no head-to-head trials for at least one comparison of interest. The bovine respiratory disease literature illustrates this situation well: a mixed treatment comparison of eight parenteral metaphylactic antimicrobials was able to rank treatments for morbidity and mortality outcomes using a network that combined direct and indirect evidence across 37 trials Abell et al., mixed treatment comparison of metaphylaxis treatments for bovine respiratory disease.

Defining the Clinical Question and Eligibility Criteria

The population, intervention, comparator, and outcome structure must be specified before literature searching begins. In veterinary networks, the population definition carries particular weight because species, breed, age class, production system, and disease pathogenesis all modify treatment response. A network built for feedlot calves within 48 hours of arrival cannot be extended to dairy calves with established respiratory disease without violating transitivity. The bovine respiratory disease network restricted inclusion to randomized controlled trials of parenteral metaphylaxis administered to incoming feeder or stocker calves, a narrow population definition that supported credible indirect comparisons Abell et al., mixed treatment comparison of metaphylaxis treatments for bovine respiratory disease.

Outcome definitions must be harmonised across trials before any quantitative synthesis. The bovine respiratory disease example categorised outcomes as morbidity cumulative incidence from day 1 to day 60 or less, morbidity to closeout, mortality to closeout, and retreatment incidence to closeout Abell et al., mixed treatment comparison of metaphylaxis treatments for bovine respiratory disease. Each outcome required its own network because the treatment rankings differed by outcome. A treatment that ranked highly for early morbidity did not necessarily rank highly for mortality.

Eligibility criteria should specify the minimum follow-up duration, the permitted concurrent interventions, and the definition of the treatment arm. Trials that permit rescue therapy as an outcome event measure something different from trials that exclude rescue therapy. The distinction matters for interpretation and should be recorded during data extraction.

Constructing the Evidence Base

Systematic searching follows the same principles as for any systematic review. Multiple databases should be searched, and the search strategy should be reported in full so that replication is possible. The reporting standards for animal research, including the ARRIVE guidelines, provide the framework for assessing whether individual trials contain the information needed for network construction ARRIVE guidelines for reporting animal research. The broader reporting guideline libraries maintained by the EQUATOR Network list the relevant checklists for randomised trials and systematic reviews EQUATOR Network reporting guidelines.

Data extraction for a network meta-analysis requires additional fields beyond those needed for pairwise synthesis. For each trial, record the exact interventions compared, the dose and route for each arm, the timing of outcome assessment, and the number of events and total animals per arm. When trials report multiple doses of the same drug, decide in advance whether doses will be pooled or treated as separate nodes. Pooling doses of the same drug is common but can obscure dose-response relationships. Treating each dose as a separate node preserves information but fragments the network and may prevent estimation.

Assessing Risk of Bias and Reporting Quality

Risk of bias assessment should use a tool appropriate to the study design. For randomised trials, the Cochrane Risk of Bias tool is the standard. The assessment should be conducted independently by two reviewers with disagreements resolved by discussion. Trials at high risk of bias should not automatically be excluded, but sensitivity analyzes should examine whether the network conclusions change when high-risk trials are removed.

Reporting quality deserves separate attention in veterinary networks. Incomplete reporting of randomisation methods, allocation concealment, and blinding is common in animal research. The ARRIVE guidelines specify the minimum information required for transparent animal research reporting, and their use as a reporting benchmark during screening can identify trials that lack essential methodological detail ARRIVE guidelines for reporting animal research. Trials that cannot provide basic methodological information may contribute unreliable effect estimates to the network.

Worked Example: Comparing Preoperative Skin Antiseptics

A network meta-analysis of preoperative skin antiseptics for surgical site infection prevention illustrates the practical steps of network construction and interpretation. The evidence base included randomised controlled trials comparing chlorhexidine, iodine, and olanexidine in alcohol-based and aqueous solutions across adult surgical patients Jalalzadeh et al., network meta-analysis of preoperative skin antiseptics. Although this example comes from human surgery, the methodological structure transfers directly to veterinary surgical site infection prevention.

The network contained multiple nodes because each antiseptic was evaluated in more than one formulation. Chlorhexidine in alcohol, chlorhexidine in aqueous solution, iodine in alcohol, and iodine in aqueous solution were treated as separate nodes. This decision preserved clinically meaningful distinctions because the alcohol vehicle affects both antimicrobial activity and drying time. The network meta-analysis used a frequentist random effects model, a choice that reflects the availability of direct comparisons between several node pairs.

The key interpretive step was examining whether the network ranking matched clinical expectations. Alcohol-based chlorhexidine ranked highest for preventing surgical site infections, a finding consistent with the mechanism of action and with direct trial evidence. The network also allowed comparison between antiseptics that had never been tested head-to-head, which is the primary utility of the method.

Checklist for Assessing Transitivity and Consistency

The following checklist provides a structured approach to evaluating whether a veterinary network meta-analysis can be trusted. Each item should be addressed explicitly in the methods and results sections of the report.

Consistency assessment requires formal statistical testing. The node-splitting approach compares direct and indirect evidence for each pairwise comparison and reports a Bayesian p-value for disagreement. The design-by-treatment interaction test evaluates consistency across the entire network simultaneously. Both tests should be reported, and the results should be interpreted in the context of the network geometry. A network with few closed loops has limited power to detect inconsistency, so failure to reject consistency does not prove that the assumption holds.

When inconsistency is detected, the first response is to examine trial characteriztics instead of to abandon the analysis. Differences in species, production system, disease chronicity, or concurrent medication may explain the disagreement. The bovine respiratory disease network, for example, would require separate consistency assessment for each outcome because the treatment rankings differed by outcome Abell et al., mixed treatment comparison of metaphylaxis treatments for bovine respiratory disease. A network that is consistent for mortality may be inconsistent for retreatment incidence, and the reasons for that difference may be clinically informative.

Species and Production System Considerations

The correct choice of network structure depends on the species and production system under study. Companion animal networks can often include trials from multiple countries because the patient population and standard of care are relatively similar across regions. Production animal networks require more careful population definition because management systems, disease epidemiology, and regulatory constraints on drug use differ substantially between regions. The WOAH terrestrial animal health standards document the international framework for animal health and disease control, and these standards can inform decisions about which populations are comparable for synthesis WOAH terrestrial animal health code.

Regulatory approval status also affects network construction. A drug that is licensed for a given indication in one country but not another may have a different evidence base, and trials conducted under different regulatory frameworks may use different outcome definitions. The MSD Veterinary Manual provides species-specific information on drug approvals and clinical use that can help identify whether trials from different regions are likely to be comparable MSD Veterinary Manual professional edition. Professional practice resources from organizations such as the AVMA can also indicate regional differences in standard of care that might violate transitivity AVMA professional practice resources.

The evidence base for veterinary network meta-analyzes remains limited compared with human medicine. Many veterinary networks will contain fewer than ten trials, and some comparisons will rest on a single trial. In these circumstances, the network meta-analysis should be presented as an exploratory synthesis instead of a definitive ranking. The confidence in the estimates should be graded accordingly, and the limitations should be stated plainly in the discussion.

Recognized Complications and Failure Modes

Network meta-analyzes fail in characteriztic ways, and most failures are detectable before the analysis is finalised. The most consequential failure is inconsistency, where direct and indirect estimates for the same comparison disagree beyond sampling error. A global inconsistency test, such as the design-by-treatment interaction model, provides a single summary statistic, but it can miss inconsistency confined to one loop. Loop-specific tests, which compare direct and indirect estimates within a closed triangle of treatments, localize the problem. When inconsistency is detected, the first step is to verify that the transitivity assumption holds for the studies forming the loop. If study populations differ systematically across comparisons, the indirect estimate is biased and should not be pooled with the direct estimate.

A second failure mode is sparse network geometry. When some treatments are connected by a single study, or when a comparator node has few events, the network meta-analysis produces wide credibility intervals that are easily overinterpreted. The discriminating check is the network plot itself: nodes with few connecting edges and comparisons supported by one small trial should be flagged in the results as low confidence. A related problem is the inclusion of studies with zero events in one arm. Standard continuity corrections can distort estimates when event rates are low, and a sensitivity analysis using alternative corrections, or an exact method, is advisable.

Publication bias remains a concern in veterinary networks. Funnel plot asymmetry is harder to interpret in a network than in a pairwise meta-analysis, because each comparison has its own set of studies. Comparison-adjusted funnel plots can be used, but their power is limited when the network is small. The practical response is to search grey literature and conference proceedings during the evidence construction phase, and to report the search dates and databases explicitly.

Common Errors and Corrective Actions

Less experienced analysts often confuse the consistency assumption with the transitivity assumption. Transitivity is a clinical judgment made before analysis: it asks whether the studies comparing treatments A and B are sufficiently similar to the studies comparing B and C that an indirect A versus C estimate is meaningful. Consistency is a statistical property checked after analysis. A network can be transitive in design and still show inconsistency by chance, and a statistically consistent network can rest on clinically intransitive evidence. The corrective action is to document the transitivity assessment in the protocol, with the same rigour applied to eligibility criteria.

A second common error is ranking treatments by surface under the cumulative ranking curve (SUCRA) values without reporting the underlying effect estimates and their uncertainty. SUCRA compresses a distribution into a single number, and two treatments with nearly identical SUCRA values may differ clinically. Report the mean effect with its credibility interval for every pairwise comparison, and treat ranking as a secondary summary. A third error is the inclusion of studies with different follow-up durations in the same network without checking whether the outcome definition is time-consistent. For example, bovine respiratory disease morbidity can be reported at day 60 or at closeout, and pooling these as a single outcome requires justification Abell and colleagues, mixed treatment comparison of metaphylaxis treatments for bovine respiratory disease.

A fourth error is the failure to pre-specify the analysis plan. Post hoc decisions about which studies to include, which effect measure to use, or which inconsistency test to run inflate the risk of false positive findings. A protocol registered before data extraction, following the reporting standards catalogued by the EQUATOR Network, protects against this.

Limitations of the Current Evidence

The veterinary evidence base for network meta-analysis is thinner than the human counterpart. Many veterinary networks are built from a handful of small trials, and the trials themselves often have short follow-up and surrogate outcomes. The bovine respiratory disease network included 29 studies and 37 trials, which is large by veterinary standards, yet the authors still reported wide credibility intervals for several comparisons Abell and colleagues, mixed treatment comparison of metaphylaxis treatments for bovine respiratory disease. Networks in companion animal medicine are smaller still.

Expert opinion differs on how to handle studies with different breeds, production systems, or housing conditions within one network. Some analysts advocate restricting the network to a single production system, while others prefer a broader network with covariates. The correct choice depends on the clinical question, and the decision should be made explicit. Reporting standards for animal research, such as the ARRIVE guidelines, improve the transparency of the primary studies, but they do not resolve the underlying scarcity of randomised trials in many veterinary fields.

Escalation and Consultation

Referral to a statistician with network meta-analysis experience is warranted when inconsistency is detected and cannot be explained, when the network contains multi-arm trials that require correlated effect modeling, or when the analysis plan involves complex covariates. Laboratory involvement is indicated when outcome definitions vary across studies and require harmonisation, for example when diagnostic thresholds differ between sites. Regulatory reporting may be required when the network meta-analysis informs a label claim or a regulatory submission, in those settings, the analysis should follow the standards expected by the relevant authority, and the WOAH terrestrial animal health standards provide a reference for international expectations.

When the network meta-analysis is intended to guide clinical practice, the results should be interpreted alongside species-specific guidance from sources such as the MSD Veterinary Manual and professional resources from the American Veterinary Medical Association, because statistical ranking does not replace clinical judgment about drug availability, cost, or regulatory constraints in a given region.

Frequently Asked Questions

How Many Studies Are Needed Before a Network Meta-Analysis Becomes Worthwhile?

There is no fixed minimum, but the network must contain at least one closed loop to gain an advantage over pairwise meta-analysis. A connected network with three or more treatments and at least one direct comparison between two of them permits indirect estimation. Sparse networks produce wide credibility intervals and fragile ranking estimates. The bovine respiratory disease network included 37 trials across 8 antimicrobials, which supported stable estimates for morbidity and mortality outcomes. If your network has fewer than 10 studies, examine whether the indirect estimates are precise enough to inform clinical decisions. When the evidence base is thin, a narrative synthesis with pairwise meta-analysis may serve your question better than a network that cannot be validated.

What Should I Do When Direct and Indirect Evidence Disagree?

Disagreement between direct and indirect evidence signals inconsistency, a violation of the consistency assumption. First, verify that transitivity holds by comparing the distribution of effect modifiers across comparisons. If patient or study characteriztics differ systematically, the indirect estimate is biased. Second, inspect the network geometry for a single influential study bridging two treatment arms. The anticoagulant therapy network for cancer-associated thrombosis showed inconsistency between pairwise and network estimates for major bleeding, which prompted caution in interpreting that outcome. When inconsistency persists, report both direct and indirect estimates separately, present the network estimate with a clear caveat, and consider whether the conflicting comparisons should be pooled at all.

Can I Apply Results From a Human or Laboratory Animal Network to My Clinical Population?

Only when the biological mechanism and outcome definitions are sufficiently similar. Networks built on human trials, such as the preoperative skin antiseptic comparison, may inform veterinary surgical practice where no veterinary trials exist, but species differences in skin flora, hair density, and wound healing limit direct extrapolation. Conversely, networks from mouse models of diabetes provide mechanistic insight but do not establish efficacy in companion animals. The MSD Veterinary Manual offers species-specific guidance that should be weighed alongside network estimates. When applying cross-species evidence, state the assumption of transitivity explicitly and downgrade the certainty of recommendations. If a veterinary network exists for your question, prefer it over extrapolated human data.

How Do I Account for Different Outcome Definitions Across Studies?

Outcome heterogeneity is a common reason networks fail. The bovine respiratory disease network addressed this by categorising outcomes into distinct cumulative incidence windows, such as morbidity from day 1 to day 60 versus day 1 to closeout. Before pooling, map each study's outcome definition to a standardized taxonomy. If studies report outcomes at incompatible time points, consider a multi-outcome network or restrict the analysis to the most consistently reported definition. Sensitivity analyzes can test whether results change when borderline studies are excluded. Report the outcome definitions used in your protocol and justify any harmonisation decisions. When definitions cannot be reconciled, the studies are not comparable and should not share a network node.

What Resources and Expertise Are Required to Conduct a Network Meta-Analysis?

A network meta-analysis demands more than statistical software. You need a librarian or information specialist for the systematic search, two reviewers for screening and data extraction, and a statistician familiar with Bayesian or frequentist network models. The EQUATOR Network reporting guidelines library lists the PRISMA extension for network meta-analyzes, which should be followed from protocol stage. Time commitment typically ranges from several months to a year. If institutional statisticians lack network meta-analysis experience, consider collaboration with a methodologist at another institution. Budget for access to full-text databases and, if using Bayesian methods, sufficient computing power for Markov chain Monte Carlo estimation. A feasible alternative is a frequentist approach, which requires less computational intensity.

How Should I Present Network Meta-Analysis Results to Clinicians or Regulators?

Lead with the clinical question and the network diagram, then present effect estimates for the most clinically relevant comparisons instead of the full matrix. Ranking probabilities, such as surface under the cumulative ranking curve values, are intuitive but should be accompanied by the underlying effect sizes and uncertainty intervals. The short-chain fatty acid network ranked butyrate as probably most effective, but the confidence intervals overlapped with other interventions, so ranking alone would overstate certainty. For regulators, emphasize the consistency assessment and risk of bias evaluation. For clinicians, frame results as the probability of a meaningful benefit over the current standard. The AVMA practice resources can guide translation of evidence into practice recommendations. Always state the certainty of evidence and the populations to which results apply.

Related Clinical & Scientific Guides

• Conducting Systematic Reviews of Veterinary Diagnostic Test Accuracy
• Bias in Veterinary Research: Types, Sources, and Mitigation
• Cluster Randomized Trials in Veterinary Research: Design and Analysis

References and Further Reading

• A mixed treatment comparison meta-analysis of metaphylaxis treatments for bovine respiratory disease in beef cattle.. 2017.
• Anticoagulant therapy for acute venous thrombo-embolism in cancer patients: A systematic review and network meta-analysis.. 2019.
• Effects of short-chain fatty acids on blood glucose and lipid levels in mouse models of diabetes mellitus: A systematic review and network meta-analysis.. 2024.
• Do bone grafts or barrier membranes provide additional treatment effects for infrabony lesions treated with enamel matrix derivatives? A network meta-analysis of randomized-controlled trials.. 2010.
• Salivary biomarkers for early detection of oral squamous cell carcinoma (OSCC) and head/neck squamous cell carcinoma (HNSCC): A systematic review and network meta-analysis.. 2024.
• Efficacy of different preoperative skin antiseptics on the incidence of surgical site infections: a systematic review, GRADE assessment, and network meta-analysis.. 2022.
• ARRIVE Guidelines 2.0 for Reporting Animal Research. PLOS Biology, 2020.
• EQUATOR Network Reporting Guidelines. EQUATOR Network.
• MSD Veterinary Manual, Professional Edition. MSD Veterinary Manual.

Related Articles

• Meta-Analysis of Veterinary Diagnostic Test Accuracy
• Applying Competing Risks Analysis in Veterinary Research
• Cluster Randomized Trials in Veterinary Research: Design and Analysis
• Statistical Analysis of Veterinary Clinical Trials: Common Methods and Misconceptions
• Appraising Diagnostic Accuracy Studies in Veterinary Medicine

This article is educational professional reference material for veterinary audiences. It is not a substitute for veterinary diagnosis, individual clinical judgment, current product labeling, or applicable regulatory requirements.