# Understanding Bias in Veterinary Epidemiological Studies


## Key Takeaways

- Bias is a systematic error in study design, conduct, or analysis that distorts effect estimates, distinct from random error, and cannot be corrected by increasing sample size.
- Selection bias arises when the study population is not representative of the target population, often due to referral bias in veterinary teaching hospitals or owner volunteer bias, leading to distorted exposure-outcome associations.
- Information bias occurs through systematic misclassification of exposure or outcome, such as recall bias where owners of sick animals remember exposures differently than owners of healthy animals, or diagnostic test error where imperfect sensitivity/specificity leads to misclassification.
- Species-specific features like culling practices in production animals (e.g., mastitic cows being culled before study endpoints) and owner-mediated care in companion animals introduce unique bias pathways not present in human epidemiology.
- Design safeguards include explicit eligibility criteria, enrollment before outcome ascertainment, blinded outcome assessment, and validated measurement instruments, while species-specific considerations necessitate adapting general epidemiological principles to animal population realities.
- Reporting standards like WOAH surveillance standards and CDC epidemiology teaching materials provide frameworks for bias-aware study reporting, emphasizing transparent documentation of sampling frames, enrollment completeness, and potential sources of error.

---

Veterinary epidemiological studies generate the evidence base for clinical decisions, herd health programs, and regulatory policy. Their conclusions, however, depend on the validity of the measurements and comparisons that underlie them. Bias, a systematic error in study design, conduct, or analysis, distorts effect estimates away from their true values. This article provides a structured account of the principal bias categories that affect veterinary research, the mechanisms by which they operate, and the practical decisions a researcher faces when designing, executing, and interpreting observational studies in animal populations.

The intended reader is a veterinary researcher or graduate student engaged in study design, manuscript review, or critical appraisal of the literature. The article answers three questions. What forms does bias take in veterinary epidemiology? How do species-specific features of animal populations, such as management systems, diagnostic test availability, and owner-mediated care, create bias pathways that differ from those in human epidemiology? And which design choices reduce or eliminate each bias type? Confounding is treated separately in this reference series and is excluded here, although its effects frequently co-occur with those of selection and information bias.

## At a Glance

| Parameter | Decision or fact |
| --- | --- |
| Bias definition | Systematic error in design, conduct, or analysis that distorts an effect estimate, distinct from random error |
| Selection bias | Arises when the study population is not representative of the target population, or when participation depends on both exposure and outcome |
| Information bias | Arises from systematic misclassification of exposure, outcome, or covariates |
| Differential misclassification | Error that differs between comparison groups, biases effect estimates in unpredictable directions |
| Non-differential misclassification | Error that is independent of group membership, typically biases estimates toward the null |
| Key design safeguards | Explicit eligibility criteria, enrollment before outcome ascertainment, blinded outcome assessment, validated measurement instruments |
| Species-specific considerations | Management systems, culling practices, diagnostic test performance, and owner reporting behavior modify bias risk |
| Reporting standard | WOAH surveillance standards and CDC epidemiology teaching materials provide frameworks for bias-aware study reporting |

## The Logic of Systematic Error

Bias differs from random error in a fundamental way. Random error, driven by sampling variation, diminishes as sample size increases and can be quantified with confidence intervals. Bias does not shrink with larger samples. A large study that is biased produces a precise estimate of the wrong quantity. The researcher's task is therefore to anticipate bias before data collection begins, because many bias mechanisms cannot be repaired after the fact.

The target population is the group to which the researcher wishes to generalize. The study population is the group actually sampled. The difference between these two groups, when it is systematic instead of random, is the raw material of selection bias. Similarly, the true exposure and outcome status of each animal is the ground truth that the measurement instruments attempt to capture. The gap between measured and true status, when it is systematic, constitutes information bias.

Veterinary research presents bias risks that human epidemiology does not face. Animals cannot consent to participation, so enrollment decisions are made by owners, herd managers, or attending clinicians. Culling and slaughter remove animals from observation at rates that may depend on the very exposures under study. Diagnostic tests validated in one species or production system may perform differently in another, as illustrated by the species-specific epidemiology of mastitis pathogens and the molecular typing methods required to distinguish them [Zadoks et al., molecular epidemiology of mastitis pathogens of dairy cattle and comparative relevance to humans](https://pubmed.ncbi.nlm.nih.gov/21968538/). These features demand that the veterinary researcher adapt general epidemiological principles to the realities of animal populations.

## Selection Bias

Selection bias occurs when the probability of being included in the study is related to both the exposure and the outcome of interest. The resulting study population no longer reflects the target population, and the observed association between exposure and outcome diverges from the true association.

### Sources of Selection Bias in Animal Studies

Referral bias is common in veterinary clinical research conducted at teaching hospitals and specialty referral centers. A primary care population of dogs with a given condition differs systematically from the referral population, because referral decisions depend on case severity, owner resources, and clinician judgment. Studies based on referral populations therefore overestimate the severity spectrum of disease and may distort exposure-outcome associations.

Volunteer bias operates when owners decide whether to enrol their animals. Owners who participate in research tend to differ from non-participants in education, income, and health-related behaviors, including preventive care and vaccination compliance. These owner characteriztics are frequently associated with the animal's exposure status, creating a selection pathway that the researcher cannot fully observe.

Loss to follow-up is a form of selection bias in longitudinal studies. Animals may be withdrawn because they die, are euthanised, are sold, or move to a different management system. Each of these events can be related to the exposure under study. In production animal research, culling decisions are often driven by productivity or health, so animals that leave the cohort are not a random subset. The systematic review of echinococcosis epidemiology in domestic and wild animals illustrates the related problem of study heterogeneity, where differences in animal management and sampling across studies complicate pooled inference [Otero-Abad and Torgerson, a systematic review of the epidemiology of echinococcosis in domestic and wild animals](https://pubmed.ncbi.nlm.nih.gov/23755310/).

### Design Strategies

Incidence density sampling, in which controls are selected from the population at risk at the time each case occurs, reduces selection bias in case-control studies nested within cohorts. Enrollment before disease onset, where feasible, eliminates the possibility that outcome status influences participation. Explicit, written eligibility criteria applied uniformly to all candidate animals reduce the discretion that invites bias. For production animal studies, the researcher should document the source population, the herd selection process, and the reasons for animal exclusion at each stage, following the surveillance reporting frameworks published by the [World Organization for Animal Health](https://www.woah.org/en/what-we-do/animal-health-and-welfare/disease-data-collection/).

## Information Bias

Information bias arises when the measurement of exposure, outcome, or covariates is systematically inaccurate. The direction and magnitude of the resulting distortion depend on whether the misclassification is differential or non-differential.

### Misclassification of Exposure and Outcome

Non-differential misclassification occurs when measurement error is independent of the other variable. In a cohort study, if exposure is measured with error that does not depend on outcome status, the effect estimate is generally attenuated toward the null. The same logic applies to outcome misclassification that is independent of exposure. This attenuation can obscure genuine associations, particularly when the exposure is common or the measurement error is large.

Differential misclassification is more serious because it can bias estimates in either direction, sometimes producing associations that are entirely spurious. Recall bias is the classic example. Owners of animals with a diagnosed disease may search their memory more thoroughly for potential exposures than owners of healthy animals, producing differential reporting of historical exposures. Interviewer bias operates similarly when data collectors who know the outcome status probe more diligently for exposures in cases than in controls.

### Diagnostic Test Error as Information Bias

Veterinary studies depend on diagnostic tests that are rarely perfect. Sensitivity and specificity determine the direction of bias when test results are used to classify outcome status. If a test has imperfect sensitivity, some diseased animals are classified as non-diseased. When this misclassification is non-differential, the effect estimate is biased toward the null. When test performance differs between exposure groups, for example because the test was validated in one species but applied to another, the bias becomes differential and unpredictable.

The molecular epidemiology of bovine mastitis demonstrates the importance of test choice. Different typing methods, from electrophoretic banding patterns to whole genome sequencing, have different discriminatory power, and the choice of method affects the apparent distribution of strains within and between herds [Zadoks et al., molecular epidemiology of mastitis pathogens of dairy cattle and comparative relevance to humans](https://pubmed.ncbi.nlm.nih.gov/21968538/). A researcher who uses a low-resolution typing method will systematically misclassify distinct strains as identical, attenuating apparent transmission rates.

### Measurement Instruments and Blinding

Blinding of outcome assessors to exposure status, and of exposure assessors to outcome status, is the primary defense against differential information bias. Where blinding is impossible, the researcher should document the reasons and quantify the potential for bias through sensitivity analyzes. Standardized data collection instruments, applied identically to all study animals, reduce non-differential error. Validation of instruments against a reference standard in the target species, instead of in a related species, is essential before the instrument is used in the main study. The [CDC principles of epidemiology in public health practice](https://www.cdc.gov/csels/dsepd/ss1978/index.html) provide a structured approach to measurement error that veterinary researchers can adapt to animal populations.

## Bias in Study Design and Conduct

### Bias in Cross-Sectional and Prevalence Studies

Cross-sectional studies in veterinary medicine commonly estimate disease prevalence, yet their validity depends on the representativeness of the sampled population. Prevalence estimates are distorted when the sampling frame excludes subpopulations with different exposure or disease status. For example, surveys of echinococcosis in dogs that recruit animals presented at veterinary clinics will overestimate prevalence in the general canine population, because owners who seek veterinary care are more likely to have observed clinical signs or to practice higher standards of husbandry. A systematic review of echinococcosis epidemiology identified feeding with raw viscera and lack of anthelmintic treatment as consistent risk factors across studies, but the review authors noted that studies relying on convenience samples of dogs produced risk estimates that differed from those based on random village-level sampling [Otero-Abad and Torgerson, systematic review of echinococcosis epidemiology](https://pubmed.ncbi.nlm.nih.gov/23755310/).

The timing of sampling introduces a further layer of error. Prevalence studies conducted during a single season capture only the point prevalence, which may differ markedly from the annual average for diseases with seasonal transmission. This matters for vector-borne infections, for periparturient diseases, and for parasitic infections whose egg output varies with host immunity and weather. When the exposure of interest is itself seasonal, such as pasture turnout or housing, the measured association between exposure and disease can be reversed if the study period does not span the full transmission cycle.

### Bias in Longitudinal and Cohort Studies

Cohort studies in veterinary settings are vulnerable to attrition bias, which arises when loss to follow-up is related to both exposure and outcome. Animals that die or are culled before the study endpoint are not missing at random. In dairy herds, cows with clinical mastitis are more likely to be culled than herdmates, so a cohort study of mastitis outcomes that ignores culling will underestimate the true incidence of severe disease [Zadoks et al., molecular epidemiology of mastitis pathogens](https://pubmed.ncbi.nlm.nih.gov/21968538/). The direction and magnitude of attrition bias depend on which animals leave and why. Differential loss of exposed animals with poor outcomes biases the risk ratio toward the null, whereas differential loss of unexposed animals with poor outcomes biases it away from the null.

Surveillance bias operates when the intensity of outcome detection differs between exposure groups. In production animal medicine, herds enrolled in health schemes receive more frequent veterinary visits and more diagnostic testing than non-enrolled herds. A comparison of disease incidence between enrolled and non-enrolled herds will overestimate the effect of any risk factor that is more common in enrolled herds, simply because more cases are detected. This bias is particularly problematic in studies of diseases with a wide subclinical spectrum, such as mastitis, lameness, and parasitism.

### Bias in Diagnostic Accuracy Studies

Diagnostic accuracy studies compare a candidate test against a reference standard. Verification bias occurs when only a subset of animals receives the reference standard, and the decision to verify is influenced by the candidate test result. In veterinary studies, this commonly arises when the reference standard is invasive, expensive, or requires necropsy. If animals with negative candidate test results are less likely to undergo the reference procedure, sensitivity will be overestimated and specificity underestimated. The magnitude of this bias can be quantified using the verification ratio, the proportion of test-positive and test-negative animals that receive the reference standard.

Spectrum bias describes the variation in test performance across different disease stages, severities, and host subpopulations. A test that performs well in animals with advanced clinical disease may perform poorly in subclinical or early infections. For example, molecular typing methods for mastitis pathogens have different discriminatory power depending on the bacterial species and the within-herd strain diversity [Zadoks et al., molecular epidemiology of mastitis pathogens](https://pubmed.ncbi.nlm.nih.gov/21968538/). Diagnostic accuracy studies that enrol only severely affected animals will overestimate sensitivity in the target population. The same principle applies to serological tests evaluated in vaccinated versus naturally infected animals, where antibody kinetics differ substantially.

### Bias in Surveillance and Outbreak Investigations

Animal health surveillance systems are subject to reporting bias, which arises when the probability that a case is reported depends on factors other than the true disease status. Passive surveillance relies on clinicians and producers to submit reports, and submission rates vary with disease awareness, economic incentives, and the perceived consequences of reporting. The World Organization for Animal Health surveillance standards emphasize that the sensitivity of a surveillance system must be evaluated, not assumed, and that the choice of surveillance method should be matched to the disease, the production system, and the purpose of the data [WOAH animal health surveillance standards](https://www.woah.org/en/what-we-do/animal-health-and-welfare/disease-data-collection/). A system that detects clinical cases well may miss subclinical infections entirely, leading to systematic underestimation of disease burden.

Outbreak investigations face a distinct set of biases. Recall bias is prominent when exposure histories are collected after the outbreak is recognized, because affected herd owners search their memory more intensely than unaffected controls. Interviewer bias arises when the investigator, knowing the outbreak source hypothesis, probes exposed individuals more thoroughly. The case definition itself can introduce bias if it is applied differently across herds or species. Standardized case definitions and structured questionnaires, applied before the outbreak source is revealed to the field team, reduce these errors.

### Bias in Molecular Epidemiological Studies

Molecular epidemiology adds a layer of complexity to bias assessment. Strain typing results depend on the sampling strategy, the number of isolates collected per animal and per herd, and the typing method used. Studies that collect a single isolate per animal will miss within-host strain diversity, which is documented for several mastitis pathogens where multiple strains can coexist in the same mammary gland [Zadoks et al., molecular epidemiology of mastitis pathogens](https://pubmed.ncbi.nlm.nih.gov/21968538/). The choice of typing method determines the discriminatory power and therefore the conclusions about transmission chains. Methods with low discriminatory power will falsely suggest that distinct strains are identical, overestimating transmission between animals.

Selection bias in molecular studies often operates at the isolate level instead of the animal level. If isolates are chosen for typing based on phenotypic characteriztics, such as antimicrobial resistance or colony morphology, the resulting strain distribution will not reflect the population from which the isolates were drawn. This is a form of verification bias applied to microbial specimens. The sampling frame for isolates should be defined independently of the genetic characteriztics under investigation.

### A Classification Framework for Veterinary Bias

The following table organizes the principal biases by study phase and provides veterinary examples. The classification follows the framework used in the CDC principles of epidemiology, adapted to animal populations [CDC principles of epidemiology in public health practice](https://www.cdc.gov/csels/dsepd/ss1978/index.html).

| Bias | Study phase | Mechanism | Veterinary example | Direction of effect |
|---|---|---|---|---|
| Selection bias | Sampling | Sampling frame excludes part of the target population | Clinic-based prevalence surveys overrepresent animals with access to veterinary care | Prevalence overestimated |
| Attrition bias | Follow-up | Loss to follow-up related to exposure and outcome | Culling of mastitic cows before study end | Risk ratio biased toward null |
| Surveillance bias | Outcome detection | Detection intensity differs by exposure group | Herds in health schemes tested more frequently | Spurious association |
| Verification bias | Diagnostic accuracy | Reference standard applied selectively | Necropsy only in test-positive animals | Sensitivity overestimated |
| Spectrum bias | Diagnostic accuracy | Test performance varies by disease stage | Serology evaluated only in clinical cases | Sensitivity overestimated |
| Recall bias | Data collection | Exposure recall differs by outcome status | Herd owners with outbreaks recall exposures more thoroughly | Odds ratio biased away from null |
| Reporting bias | Surveillance | Case reporting probability varies | Passive surveillance misses subclinical cases | Incidence underestimated |
| Isolate selection bias | Molecular typing | Isolates chosen by phenotype | Resistant isolates typed preferentially | Transmission overestimated |

The correct mitigation strategy depends on the bias, the study design, and the resources available. For selection bias, the remedy is a defined sampling frame and random or systematic selection within it. For information bias, the remedy is blinding, standardized measurement, and validation of instruments against an appropriate reference. For attrition bias, the remedy is active follow-up, analysis of dropouts, and sensitivity analyzes that model the missing data under different assumptions. No single design eliminates all bias, but explicit identification of the most influential biases in a given study allows the investigator to prioritize mitigation efforts and to state the residual limitations honestly.

## Recognized Failure Modes and Early Detection

Bias in veterinary epidemiological studies rarely announces itself. It accumulates through decisions made at each stage of design, execution, and analysis, and its effects become visible only when results are compared across settings or when replication fails. The most common failure modes are consistent across species and study types.

**Restriction of the source population.** When enrollment criteria exclude animals that are difficult to sample, such as free-ranging wildlife or herds without electronic records, the study population no longer represents the target population. Detection requires comparing the demographic and clinical profile of enrolled subjects against the full eligible population, using whatever denominator data exist. In production animal studies, herd-level participation rates below 70% warrant scrutiny, as non-participating herds often differ systematically in management intensity and disease burden.

**Left truncation and late entry.** Cohort studies that enrol animals after the exposure period begins miss early events and misclassify person-time at risk. This is particularly problematic in long-latency conditions such as neoplasia or chronic parasitism. Early detection involves plotting the distribution of entry times relative to exposure onset and testing whether outcome rates differ between early and late entrants.

**Survivor bias in cross-sectional sampling.** Prevalence studies that sample only animals present at a single visit exclude those that died or were culled before sampling. In mastitis research, for example, cows with severe acute infections may be culled before bacteriological sampling occurs, producing prevalence estimates that understate the true burden of virulent strains. The discriminating check is to compare culling rates between exposure groups and to model the outcome using survival methods when attrition exceeds 10%.

**Verification bias in diagnostic accuracy studies.** When only animals with positive screening tests undergo the reference standard, sensitivity is overestimated and specificity underestimated. This is common in studies of novel diagnostic tests for chronic infections. Detection requires auditing the proportion of screen-negative animals that received the reference standard and, when feasible, applying the reference standard to a random subset of screen-negatives.

**Recall and reporting asymmetry.** In questionnaire-based studies, owners of affected animals recall exposures with greater completeness than owners of unaffected animals. This is especially problematic in companion animal oncology and toxicology studies. Early detection uses validation substudies comparing questionnaire responses against veterinary records or pharmacy dispensing data.

The following table summarizes common failure modes and their discriminating checks.

| Observation | Likely cause | Discriminating check |
|---|---|---|
| Effect estimate changes when sample restricted | Selection bias from enrollment criteria | Compare characteriztics of included versus excluded animals |
| Prevalence lower than expected for known endemic disease | Survivor bias or left truncation | Examine culling or mortality rates before sampling |
| Sensitivity and specificity sum exceeds expected range | Verification bias | Audit reference standard application in screen-negatives |
| Exposure-outcome association present only in one subgroup | Recall bias or differential misclassification | Validate exposures against independent records |
| Results differ markedly between study sites | Unmeasured selection pressure or effect modification | Stratify by site and compare enrollment completeness |

## Common Errors and Corrective Actions

Less experienced investigators frequently conflate bias with confounding or with random error. Bias is a systematic distortion produced by the study process itself, whereas confounding arises from the natural association of exposures and outcomes with third variables. The distinction matters because the remedies differ: confounding can be addressed through design or adjustment, while bias cannot be corrected after data collection.

A second recurring error is treating non-response as a nuisance instead of a source of selection bias. When 30% of invited herds decline participation, the study is no longer a random sample of the target population. The corrective action is to document the reasons for non-participation, compare participants with non-participants on available variables, and conduct sensitivity analyzes that re-weight results under different assumptions about the non-participants.

A third error involves the use of convenience samples for prevalence estimation. Samples of convenience, such as diagnostic laboratory submissions or shelter populations, reflect the referral and submission patterns of the source facilities, not the underlying population. Corrective action is to restrict inference to the population that the sampling frame actually represents and to state that restriction explicitly in the methods and limitations.

A fourth error is the failure to blind outcome assessors in studies where the exposure is visible. In field trials of surgical techniques or therapeutic protocols, the surgeon or clinician who knows the treatment allocation may unconsciously influence outcome measurement. The corrective action is to use independent outcome assessors who are masked to allocation, and to document the masking procedure in the study protocol.

## Limitations of Current Evidence and Areas of Expert Disagreement

The veterinary literature contains relatively few formal bias assessment tools adapted from human epidemiology. Most veterinary systematic reviews rely on checklists developed for human studies, which may not capture species-specific sources of bias such as culling decisions, production economics, or owner compliance. The [systematic review of echinococcosis epidemiology in domestic and wild animals](https://pubmed.ncbi.nlm.nih.gov/23755310/) illustrates this gap, as the included studies varied widely in sampling strategy and diagnostic methods, complicating direct comparison of risk factor estimates.

Expert opinion differs on the magnitude of bias introduced by diagnostic test imperfection in field studies. Some investigators advocate routine correction of prevalence estimates using known test sensitivity and specificity, while others argue that such corrections introduce new assumptions that may be less defensible than the original bias. The [CDC principles of epidemiology in public health practice](https://www.cdc.gov/csels/dsepd/ss1978/index.html) recommend transparent reporting of test characteriztics and explicit discussion of their influence on estimates, a position that most veterinary epidemiologists now accept.

A further area of disagreement concerns the role of molecular typing in reducing misclassification. While [molecular epidemiological studies of mastitis pathogens](https://pubmed.ncbi.nlm.nih.gov/21968538/) have clarified transmission routes and strain distributions, the choice of typing method can itself introduce bias if the method lacks discriminatory power for the target organizm. Experts differ on whether whole genome sequencing should be the default reference method or whether less expensive methods suffice for routine surveillance.

## Referral, Consultation, and Reporting Triggers

Veterinary researchers should seek specialist consultation when study design decisions have irreversible consequences for bias. This includes the selection of sampling frames for wildlife or free-ranging populations, the design of diagnostic accuracy studies, and the analysis of data from passive surveillance systems. Consultation with a veterinary epidemiologist or biostatistician is warranted before data collection begins, not after problems emerge.

Laboratory involvement is indicated when diagnostic test performance is uncertain or when sample handling may introduce measurement error. Reference laboratories can provide validation data for the specific test and matrix used in the study, and can advise on storage, transport, and processing conditions that affect test accuracy.

Regulatory reporting obligations arise when bias in surveillance or diagnostic testing could affect animal health decisions with trade or public health consequences. The [WOAH terrestrial animal health standards](https://www.woah.org/en/what-we-do/standards/codes-and-manuals/terrestrial-code-online-access/) and [WOAH animal health surveillance standards](https://www.woah.org/en/what-we-do/animal-health-and-welfare/disease-data-collection/) specify reporting requirements for notifiable diseases and define the surveillance quality expected of member countries. Investigators whose studies detect notifiable disease should confirm reporting obligations with their national veterinary authority before publication, as delayed reporting can compromise disease control efforts.

When bias threatens the validity of a study that informs regulatory decisions, the appropriate action is to disclose the limitation to the relevant authority and, where feasible, to conduct supplementary analyzes that bound the effect of the bias. Withholding such information, even when publication is not yet complete, undermines the evidence base for animal health policy.

## Frequently Asked Questions

**How much bias can I tolerate in a resource-limited field study?**

There is no universal threshold. Bias tolerance depends on the study's purpose and the decisions it will inform. For exploratory work or hypothesis generation, moderate selection bias may be acceptable if acknowledged. For studies guiding regulatory action, trade decisions, or treatment protocols, the standard is much stricter. Quantify the likely direction and magnitude of bias before starting. If you cannot measure a key confounder or achieve adequate follow-up, consider whether a smaller, more rigorous study answers the question better than a large, biased one. Document all compromises explicitly in the methods and limitations sections. Surveillance standards from the [World Organization for Animal Health](https://www.woah.org/en/what-we-do/animal-health-and-welfare/disease-data-collection/) can help frame what level of evidence is fit for purpose.

**What do I do when blinding is impossible in a field setting?**

Blinding is often impractical in livestock studies where the investigator must handle animals or read ear tags. When true blinding fails, use objective outcome definitions that leave little room for interpretation. For example, use laboratory thresholds instead of subjective clinical scores. Automate measurement where possible, such as electronic scales or automated milk recording. If the person assessing outcome must know the exposure, separate the outcome assessment from the exposure measurement and have a second person, unaware of the study question, collect outcome data. Document the blinding status of each study phase. The [CDC principles of epidemiology](https://www.cdc.gov/csels/dsepd/ss1978/index.html) describe standard approaches to reducing observer error that apply directly to veterinary field work.

**How do I explain bias to a producer or practice owner who wants quick answers?**

Use concrete language. Explain that a study can be wrong in a consistent direction, like a scale that always reads two kilograms heavy. Give an example from their species. For a dairy client, describe how testing only cows with high somatic cell counts overestimates the herd's mastitis problem. Emphasize that bias is not about dishonesty, it is about how animals are chosen and measured. Connect the explanation to their decision. If they are deciding whether to change a vaccination protocol, the cost of acting on a biased result may exceed the cost of the study itself. The [MSD Veterinary Manual](https://www.msdvetmanual.com/) offers species-specific examples of how clinical signs and test results can mislead when interpreted without accounting for population context.

**How does bias differ between companion animal and production animal studies?**

The dominant bias sources differ by species and setting. Companion animal studies rely heavily on hospital populations, so referral bias and owner self-selection dominate. Owners who seek specialty care differ systematically from the general pet population in income, attachment, and willingness to pursue treatment. Production animal studies face different pressures: culling decisions, herd-level clustering, and selective reporting of affected groups. In both settings, loss to follow-up is common but for different reasons, owners move or elect euthanasia in companion animals, animals are sold or die in production systems. Molecular typing has shown how strain distribution varies across host species and management systems, which affects generalization of findings between settings. The [molecular epidemiology of mastitis pathogens](https://pubmed.ncbi.nlm.nih.gov/21968538/) illustrates how conclusions from one herd type may not transfer to another.

**What records should I keep to allow later bias assessment?**

Keep a study log separate from the data file. Record the number of eligible animals, how many were enrolled, and reasons for non-enrollment. Note any animals that crossed over between exposure groups or received concurrent treatments. Document changes to measurement protocols, personnel, or equipment during the study. Keep versions of questionnaires and data forms with dates. Record the timing and method of outcome assessment for each animal. For surveillance data, note changes in reporting intensity or diagnostic capacity over time. These records allow you to reconstruct the study population and assess selection and information bias after the fact. The [World Organization for Animal Health terrestrial code](https://www.woah.org/en/what-we-do/standards/codes-and-manuals/terrestrial-code-online-access/) provides guidance on documentation standards for disease surveillance that apply equally to research data.

**When should I consult a veterinary epidemiologist or biostatistician?**

Consult before you finalise the study design, not after data collection. A brief consultation at the design stage can prevent bias that no analysis can fix. Seek help when you are planning a study with multiple herds or flocks, when the outcome is rare, when you cannot measure all suspected confounders, or when you plan to combine data from different sources. Also consult if a reviewer or supervisor questions your sampling strategy. Many institutions offer free statistical consulting. Bring a one-page summary of your research question, target population, and available resources. The [AVMA practice resources](https://www.avma.org/resources-tools) list professional networks that can help identify consultants with veterinary epidemiological expertise.

## Related Clinical & Scientific Guides

* [Evaluating Veterinary Surveillance System Attributes](/knowledge/veterinary-medicine/veterinary-epidemiology/evaluating-veterinary-surveillance-system-attributes)
* [Network Analysis for Infectious Disease Spread in Animal Populations](/knowledge/veterinary-medicine/veterinary-epidemiology/network-analysis-infectious-disease-spread-animal-populations)
* [Randomized Controlled Trials in Veterinary Field Settings](/knowledge/veterinary-medicine/veterinary-epidemiology/randomized-controlled-trials-veterinary-field-settings)


## References and Further Reading

- [A systematic review of the epidemiology of echinococcosis in domestic and wild animals.](https://pubmed.ncbi.nlm.nih.gov/23755310/). 2013.
- [Human monkeypox - After 40 years, an unintended consequence of smallpox eradication.](https://pubmed.ncbi.nlm.nih.gov/32417140/). 2020.
- [Concerns over use of glyphosate-based herbicides and risks associated with exposures: a consensus statement.](https://pubmed.ncbi.nlm.nih.gov/26883814/). 2016.
- [Prevention of atrial fibrillation: report from a national heart, lung, and blood institute workshop.](https://pubmed.ncbi.nlm.nih.gov/19188521/). 2009.
- [Molecular epidemiology of mastitis pathogens of dairy cattle and comparative relevance to humans.](https://pubmed.ncbi.nlm.nih.gov/21968538/). 2011.
- [Human infections due to Streptococcus dysgalactiae subspecies equisimilis.](https://pubmed.ncbi.nlm.nih.gov/19635028/). 2009.
- [WOAH Animal Health Surveillance Standards](https://www.woah.org/en/what-we-do/animal-health-and-welfare/disease-data-collection/). WOAH.
- [CDC Principles of Epidemiology in Public Health Practice](https://www.cdc.gov/csels/dsepd/ss1978/index.html). CDC.
- [MSD Veterinary Manual, Professional Edition](https://www.msdvetmanual.com/). MSD Veterinary Manual.

## Related Articles

- [Understanding Ecological Studies in Veterinary Epidemiology](/knowledge/veterinary-medicine/veterinary-epidemiology/understanding-ecological-studies-veterinary-epidemiology)
- [Sample Size Calculation for Veterinary Epidemiological Studies](/knowledge/veterinary-medicine/veterinary-epidemiology/sample-size-calculation-veterinary-epidemiological-studies)
- [Case-Control Studies in Veterinary Epidemiology: Selection and Analysis](/knowledge/veterinary-medicine/veterinary-epidemiology/case-control-studies-veterinary-epidemiology-selection-analysis)
- [Using Simulation Models in Veterinary Epidemiology](/knowledge/veterinary-medicine/veterinary-epidemiology/using-simulation-models-veterinary-epidemiology)
- [Cluster Sampling in Veterinary Field Studies](/knowledge/veterinary-medicine/veterinary-epidemiology/cluster-sampling-veterinary-field-studies)

> 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.