Meta-analysis and evidence synthesis for veterinary research

Evidence synthesis in veterinary research

Veterinary research serves several groups: owners of companion animals, farmers, veterinarians and public health authorities. Synthesis questions include which treatments work for a disease in dogs or cattle, how vaccines affect herd outcomes, what the prevalence of an infection is in a region, how housing or feeding affects welfare and productivity, and how antimicrobial use relates to resistance. Systematic reviews of veterinary interventions use methods adapted from human medicine, and there are guidelines for reporting trials in livestock (REFLECT) and for reviews.

The evidence base has particular features. Animals are clustered in herds, pens and farms, and treatment is often applied to the group. Trials are smaller and fewer than in human medicine, and reports often miss details about randomization, blinding and sample size. The diversity of species and breeds limits comparability. Our methods follow systematic review and meta-analysis practice, adapted to these features. This page builds on the general guidance for agriculture and food sciences. It describes research methods only and gives no clinical advice about any animal.

Clustering of animals

When treatment is allocated to groups but analyzed as if animals were independent, standard errors are too small. Cluster-randomized trials need an adjustment for the intracluster correlation, which is often unreported. A review should extract the unit of allocation, the number of clusters and the correlation if given, and adjust effective sample sizes using plausible values, with a sensitivity analysis. In pooled analyses of herd-level outcomes, the number of herds, not the number of animals, determines precision. Multilevel models for individual-level outcomes can include herd as a random effect.

Clinical trials in animals

Trials in companion animals are often small and conducted in referral clinics, and those in livestock are done on farms with varying management. Many lack randomization or blinding. Outcomes can be objective (death, milk yield) or subjective (owner-reported pain, clinician-scored lameness), and subjective outcomes are more prone to bias when the assessor knows the treatment. A review should assess risk of bias with a tool appropriate to the design, record blinding of owners and assessors, and compare effects by blinding. Because few trials exist for many conditions, network meta-analysis can compare several treatments, with the assumption that the trials are similar enough, which needs checking.

Prevalence and surveillance

Prevalence of infection in animal populations is pooled with the methods in the prevalence meta-analysis page. The results depend on the diagnostic test and its sensitivity and specificity, the sampling frame (random sample versus animals presented for illness), and whether prevalence is measured at animal level or herd level. A review should record all of these, and adjust for test accuracy where information is available. Different tests have different performance and may classify the same animal differently, so pooled prevalence from studies using different tests is a mix of true prevalence and test properties.

Antimicrobial use and resistance

Antimicrobial use in animals is a public health issue. Studies relate use to resistance in animal bacteria and sometimes in humans, using quantities such as treatment days, defined daily doses or amounts sold. Measures differ across countries and species, and use is correlated with other factors, such as herd size and production system. A review should record how use was measured, the level of aggregation and the laboratory method for resistance, and should not interpret an association as proving that reducing use will lower resistance by a given amount. Interventions that reduce use are better studied in trials and before-after designs, and their effects on production and health are examined together.

Welfare and behavior

Welfare research uses behavioral observations, physiological markers and indicators of health. Different welfare frameworks emphasize different aspects, and measures are not fully interchangeable. A review should describe the indicator set and avoid combining a behavioral measure of preference with a physiological measure of stress as if they were the same quantity. Studies of housing, enrichment or handling are often small and use groups of animals as the unit, so cluster adjustment matters. Ethical review and licensing of animal studies are standard, and reviews note whether the primary studies reported approval and whether animals were euthanized or followed after the study, since this affects what long-term outcomes can be known.

Production outcomes and economics

Outcomes such as milk yield, weight gain, feed conversion and fertility are continuous and depend on breed, age, stage and management. Effects are reported on the original scale or as percentages, and units differ across species and systems. Economic evaluations of interventions combine biological effects with prices, which change. A review should extract the biological effect separately from the economic assumptions and, if economics are covered, state the price basis and the year.

Publication bias and reporting

Many veterinary trials are sponsored by companies that make the product tested. Industry sponsorship has been associated with more favorable results in some comparisons, and reviews should record funding and test whether effects differ. Trial registration is not common, which limits comparison of reported with planned outcomes. We use funnel plots, searches of conference proceedings and theses, and contact authors for missing data.

One Health, zoonoses and wildlife

Many questions cross the boundary between animal, human and environmental health: zoonotic infections, antimicrobial resistance, food-borne disease and the role of wildlife as reservoirs. Studies come from veterinary, medical and ecological research, each with its own methods and reporting. A synthesis in this area should keep the evidence streams separate (animal infection, human infection, environmental sampling) and be careful about linking them. A pooled prevalence in animals tells nothing by itself about human risk, and an association between farm exposure and human infection in a case-control study does not show the animal source without typing of strains.

Wildlife studies depend on opportunistic sampling, such as animals found dead or trapped for other reasons, and sample sizes are often small. Prevalence estimates from such samples are biased, and the review should note the sampling design and avoid projecting them to whole populations. Results of serological surveys depend on cut-offs that vary between laboratories and species.

Companion animal studies and owner-reported outcomes

Companion animal research relies heavily on owner questionnaires and on clinic populations. Owners who seek referral care are not typical, and owner-reported outcomes such as quality of life or pain are subjective and open to expectation effects. Validated instruments exist for some conditions, such as canine osteoarthritis pain scales, and a review should record which were used and whether they were validated. Breed-related disease studies also face selection and the lack of population denominators. Reviews should state clearly when estimates come from clinic-based samples, and avoid presenting them as breed-wide rates. Veterinary insurance claims and electronic practice records are growing data sources, but they reflect which animals receive care and how conditions are coded, and a review should describe how they were used and what population they represent.

Common pitfalls we look for

• Ignoring clustering when treatment is allocated to pens or herds.
• Using owner- or clinician-rated outcomes without blinding.
• Pooling prevalence from different tests without accounting for accuracy.
• Treating use-resistance associations as causal effects.
• Combining species and breeds without a moderator.
• Not recording sponsorship.

Planning a veterinary synthesis

We help define the species and population, the intervention or exposure, the comparator and the outcomes, plan searches in CAB Abstracts, MEDLINE, Embase, Web of Science and Vet-specific databases, and set up coding of species, unit of allocation, cluster information, test characteristics, blinding and sponsorship. Reporting follows PRISMA and, for trials, the REFLECT and CONSORT-based items. See the systematic review service for scope and process.

An invented example of cluster adjustment

Suppose an invented trial gives a vaccine to 20 herds with 100 animals each and compares with 20 herds that are not vaccinated, and the animal-level analysis reports 2,000 versus 2,000 animals. If the intracluster correlation is 0.05, the design effect is 1 plus (100 minus 1) times 0.05, which is 5.95. The effective sample size per arm is about 2,000 divided by 5.95, around 336 animals, rather than 2,000. Pooling the trial with its nominal numbers would give it nearly six times too much weight in a meta-analysis of precision-weighted effects. This is why the review records the unit of allocation, the number of clusters and any reported correlation, and adjusts or tests sensitivity.

Coding and transparency

Coding frames record species, breed, age and production stage, setting, unit of allocation, number of clusters and animals, intracluster correlation, test used for diagnosis, outcome definition, blinding, sponsorship and country. Two coders extract data independently on a sample and the coded data and code are shared with the final report.

How we support research projects in this area

Boundaries of this service

A veterinary research synthesis describes average findings across published studies. It does not diagnose or treat any animal, advise on drug dosing, provide herd health plans or replace consultation with a licensed veterinarian. Evidence for many conditions is limited, and results may not apply to other species, breeds or production systems.

Frequently asked questions