Bayesian Methods for Diagnostic Test Evaluation in Animals

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

Key Takeaways

• Bayesian Latent Class Analysis (BLCA) estimates diagnostic test sensitivity, specificity, and true prevalence without a perfect gold standard, crucial for animal pathogens where definitive classification is challenging (e.g., bovine tuberculosis).
• The foundational Hui-Walter model requires at least two conditionally independent tests applied to at least two populations with differing true prevalences to achieve model identifiability.
• Conditional independence, the assumption that test errors are independent given true disease status, is critical; violations (e.g., two ELISAs detecting antibodies to the same antigen) bias estimates, necessitating extensions to model test correlations.
• Informative priors, derived from published literature or expert opinion, are often essential for model identifiability, particularly with low prevalence or poor test accuracy, but require sensitivity analysis to assess their influence.
• Results are presented as posterior probability distributions, summarized by medians and 95% credible intervals, offering a direct probabilistic interpretation of parameter uncertainty.
• Common study designs include two tests in two populations or three tests in one population, with model selection driven by test availability and population characteristics, such as expected prevalence differences.

Diagnostic test evaluation in veterinary medicine conventionally assumes a reference standard that perfectly classifies infection status. That assumption fails for many animal pathogens, where no gold standard exists, where reference tests are imperfect, or where sampling live animals constrains the use of definitive post-mortem methods. Bayesian latent class analysis (BLCA) offers a coherent statistical framework for estimating test sensitivity and specificity, and true prevalence, when the true infection status of each animal is unknown. This article provides veterinary researchers with the conceptual foundation, model structures, and practical decision criteria for applying Bayesian methods to diagnostic test evaluation across species.

The intended reader is a veterinary researcher or graduate student familiar with diagnostic test terminology and basic epidemiological measures, who needs to design or interpret a test evaluation study without a gold standard. The article covers when BLCA is appropriate, how the Hui-Walter model and its extensions work, how to specify priors, how to assess model fit and convergence, and how to report results. Frequentist approaches, including maximum likelihood estimation of latent class models, are excluded except where direct contrast clarifies a Bayesian concept.

At a Glance

The Problem of the Imperfect Reference Standard

Traditional test evaluation compares a candidate test against a reference standard and reports sensitivity and specificity as fixed values. When the reference standard is imperfect, the resulting estimates are biased: apparent sensitivity is underestimated if the reference misses true cases, and apparent specificity is underestimated if the reference falsely classifies uninfected animals. This problem is not hypothetical. For bovine tuberculosis, the single intradermal comparative cervical tuberculin test has a median sensitivity of approximately 0.50 with wide credible intervals when evaluated through Bayesian meta-analysis, yet it has historically served as a reference for other tests. The WOAH terrestrial animal health standards recognize that no single test provides perfect classification for many notifiable diseases, which is why surveillance definitions often combine multiple tests or require confirmatory testing.

Latent class analysis solves this problem by treating true infection status as an unobserved, or latent, categorical variable. The observed test results are modelled as probabilistic functions of this latent status. The model simultaneously estimates the sensitivity and specificity of each test and the prevalence of infection in each sampled population. The approach does not require any animal to have a known true status, which makes it applicable to field samples, archived sera, and situations where post-mortem confirmation is impractical.

The Hui-Walter Model and Its Assumptions

The foundational structure for veterinary applications is the Hui-Walter model, which requires at least two conditionally independent tests applied to at least two populations with different true prevalences. The model estimates one sensitivity and one specificity parameter per test, plus one prevalence parameter per population. With two tests in two populations, the model has six unknown parameters and six degrees of freedom from the observed test result combinations, making it just identifiable.

The critical assumption is conditional independence: given an animal's true infection status, the results of one test provide no additional information about the results of another test. This assumption is violated when tests share a biological mechanism, such as two ELISAs detecting antibodies to the same antigen, or when test performance varies with disease stage in ways that affect both tests simultaneously. The review of Bayesian latent class analysis methods emphasizes that the suitability of BLCA for a given pathogen, the availability of appropriate samples, and the number and structure of diagnostic tests must be considered before model construction.

When conditional independence fails, sensitivity and specificity estimates are biased upward. Extensions of the Hui-Walter model allow for correlated test errors by adding covariance parameters, but these require additional data structure, typically a third test or additional populations, to remain identifiable. The Bayesian modeling framework for diagnostic test evaluation describes models for two correlated tests in two or more populations and for three tests where two are correlated but jointly independent of the third.

Bayesian Estimation and Posterior Inference

Bayesian analysis combines prior distributions with the likelihood of the observed data to produce posterior distributions for all unknown parameters. For diagnostic test evaluation, the parameters of interest are the sensitivity and specificity of each test and the prevalence in each population. The posterior distribution summarizes all information about these parameters after accounting for both the data and the prior.

Computation uses Markov chain Monte Carlo (MCMC) methods, most commonly implemented in WinBUGS, JAGS, or Stan. The published WinBUGS code for standard test evaluation scenarios provides a practical starting point that can be adapted to different data structures. Convergence of the MCMC chains must be assessed using diagnostics such as the Gelman-Rubin statistic and visual inspection of trace plots. Posterior summaries are typically reported as medians with 95% credible intervals, which have a direct probabilistic interpretation: there is a 95% probability that the true value lies within the interval, given the model, the priors, and the data.

Prior Specification and Identifiability

Priors encode existing knowledge about test performance before the current data are collected. Informative priors are often necessary because the latent class model may be weakly identified from data alone, particularly when prevalence is low or when test accuracy is poor. Prior information can come from published evaluations, meta-analyzes, or expert elicitation. For example, the evaluation of brucellosis tests in Ethiopian dairy herds used prior information generated from published data on bovine brucellosis test performance, then estimated posterior sensitivity of 96.8% for an indirect ELISA and 89.6% for the Rose Bengal test.

Prior choice affects posterior estimates, so sensitivity analysis is mandatory. Researchers should rerun the model with different prior specifications, including vague or non-informative priors, and report whether conclusions change. When priors dominate the posterior, the data contribute little information, and the study design should be reconsidered. The application of latent class models to foot-and-mouth disease ELISA evaluation demonstrates that even with informative priors, a test with genuinely poor performance, such as the commercial CHEKIT kit with very low sensitivity, will be correctly identified as inadequate.

Model Structures for Common Study Designs

The simplest identifiable structure is two tests in two or more populations, assuming conditional independence. This design is common in cross-sectional studies where samples are collected from herds or regions with expected differences in prevalence. The trypanosome PCR evaluation in Western Kenya used this structure with a T. brucei specific PCR and an ITS-PCR applied to blood samples from cattle, pigs, sheep, and goats, estimating sensitivities of 76.0% and 64.0% respectively with both specificities above 99%.

Three tests in one population is an alternative structure that avoids the need for multiple populations. This design assumes that at least one pair of tests is conditionally independent, or that the correlation structure is explicitly modelled. The brucellosis test evaluation in Ethiopia applied a three test, one population model to Rose Bengal, complement fixation, and indirect ELISA results from 278 sera, with the assumption that the dairy herds shared a similar management system and unknown disease status.

When more than two populations are available, the model gains degrees of freedom that can be used to relax assumptions, such as allowing prevalence to vary freely across populations while holding test accuracy constant. The bovine tuberculosis meta-analysis extended this logic across studies, using Bayesian logistic regression with random effects to account for unexplained heterogeneity between settings, an approach that recognizes that sensitivity and specificity are not fixed properties of a test but vary with population, disease stage, and sampling conditions.

Practical Workflow for Bayesian Test Evaluation

The applied use of Bayesian latent class analysis follows a structured sequence that begins before any samples are collected. The sequence is: define the target condition and study population, select candidate tests, specify the model structure, elicit priors, collect and cross-classify data, run the analysis, and assess convergence and fit. Each step contains decision points that materially affect the validity of the resulting estimates.

Defining the Target Condition and Study Population

The target condition must be defined independently of the tests under evaluation. For infectious agents, this means specifying the infection stage of interest, such as acute infection, chronic carriage, or past exposure. The choice matters because a test may perform differently across stages. For example, molecular tests for Trypanosoma brucei detect current parasitaemia, while serological tests detect exposure, and the two measure different biological states de Clare Bronsvoort et al., 2010.

The study population must be drawn from the setting where the test will be used. Prevalence, infection intensity, and cross-reacting organizms vary by region, production system, and species. A test evaluated in dairy cattle in a high-income country may not perform identically in beef cattle in a tropical setting. The population should be defined by explicit inclusion and exclusion criteria, and the sampling frame should be described with enough detail that another researcher could replicate it.

Selecting Tests and Structuring the Model

The number of tests and populations determines which model structures are available. The Hui-Walter model requires at least two tests and two populations with different prevalences, under the assumption of conditional independence between tests given disease status. When this assumption is violated, the model can be extended to allow for correlation between tests, but this requires additional information, typically from a third test or from informative priors Branscum, Gardner, and Johnson, 2005.

Three tests in a single population is a common design when only one population is accessible. This structure permits estimation of sensitivity and specificity for all three tests without a reference standard, provided the conditional independence assumptions are met. The three test-one population model was used to evaluate Rose Bengal, complement fixation, and indirect ELISA tests for bovine brucellosis in Ethiopian dairy herds, where the authors assumed similar management across herds and unknown disease status Getachew et al., 2016.

The choice between a two test-two population and a three test-one population design depends on practical constraints. If two populations with genuinely different prevalences are available, the two test design is simpler and requires fewer tests per sample. If only one population is accessible, three tests are required. The trade-off is that the three test design demands more careful specification of the correlation structure between tests.

Eliciting Informative Priors

Priors should be derived from published literature, previous studies in similar populations, or expert opinion, and they should be documented transparently. For bovine brucellosis, prior information on sensitivity and specificity was taken from published data and incorporated into the Bayesian model Getachew et al., 2016. The choice of prior distribution, typically beta distributions for sensitivity and specificity, should reflect the strength of existing evidence.

When prior information is weak, vague priors can be used, but this may lead to identifiability problems. When prior information is strong, informative priors can stabilize estimation, but they also carry the risk of biasing results if the prior is wrong. Sensitivity analysis, in which the analysis is repeated with different priors, is a standard check on the influence of prior assumptions.

Data Collection and Cross-Classification

Each sample must be tested with every test in the model. The results are cross-classified into a contingency table that records the combination of test outcomes for each sample. For two tests, this produces a two by two table of positive and negative results. For three tests, the table expands to a two by two by two structure.

Sample size requirements for latent class analysis are larger than for simple test evaluation against a reference standard. The required sample size depends on the expected sensitivity and specificity of the tests, the prevalence in each population, and the number of parameters to be estimated. Simulation studies are often used to determine adequate sample sizes before data collection begins.

Running the Analysis and Assessing Convergence

Markov chain Monte Carlo methods are used to sample from the posterior distribution. Software implementations include WinBUGS, OpenBUGS, JAGS, and Stan, each with its own syntax and convergence diagnostics. The analysis should use multiple chains with dispersed starting values, and convergence should be assessed using the Gelman-Rubin statistic and visual inspection of trace plots.

Posterior summaries, including medians and credible intervals, are reported for sensitivity, specificity, and prevalence. The credible interval is the Bayesian analogue of the confidence interval and has a direct probabilistic interpretation: given the data and the model, the true value lies within the interval with the stated probability.

Reporting and Documentation

Results should be reported with the full model specification, including the prior distributions used, the number of chains and iterations, convergence diagnostics, and the results of sensitivity analyzes. This transparency allows readers to assess the robustness of the findings and to replicate the analysis. The World Organization for Animal Health surveillance standards emphasize the importance of documenting diagnostic test performance in the context of surveillance systems, and the same principle applies to research reporting.

Conceptual Diagram of the Bayesian Latent Class Approach

The following diagram illustrates the flow of information in a Bayesian latent class analysis for test evaluation.

┌─────────────────┐
│  Prior          │
│  Distributions  │
│  for Se, Sp,    │
│  Prevalence     │
└────────┬────────┘
         │
         ▼
┌─────────────────┐     ┌─────────────────┐
│  Observed       │────▶│  Latent Class   │
│  Test Results   │     │  Model          │
│  (Cross-        │     │  (Hui-Walter    │
│  classified)    │     │  or extension)  │
└─────────────────┘     └────────┬────────┘
                                 │
                                 ▼
                    ┌─────────────────────┐
                    │  Posterior          │
                    │  Distributions      │
                    │  for Se, Sp,        │
                    │  Prevalence         │
                    └─────────────────────┘

The observed test results are the data. The latent class model links these data to the unobserved true disease status. Priors supply external information. The posterior distributions combine the likelihood of the data with the priors to produce updated estimates of test performance.

Case Study: Evaluating Three FMD Non-Structural Protein ELISAs

A field study in Cameroon evaluated three enzyme-linked immunosorbent assays for foot-and-mouth disease non-structural protein antibodies using sera from cattle in an endemic setting Bronsvoort et al., 2006. The study used a Bayesian formulation of the Hui-Walter latent class model to estimate sensitivity and specificity in the absence of a gold standard.

The three tests were a Danish competitive ELISA, the World Organization for Animal Health recommended South American indirect ELISA, and a commercial kit. The analysis found high sensitivity and specificity for the Danish and South American tests, while the commercial kit had high specificity but very low sensitivity. This result had direct practical consequences: the commercial kit would produce many false negatives if used for surveillance in this population, and its use would underestimate the true prevalence of foot-and-mouth disease exposure.

The case illustrates several points. First, the latent class approach allowed evaluation of all three tests simultaneously without assuming any one was perfect. Second, the results were specific to the population studied, the commercial kit might perform differently in other settings. Third, the study demonstrated that latent class models are a useful alternative to traditional evaluation against a reference standard, particularly when the reference standard itself is imperfect Bronsvoort et al., 2006.

Model Selection Criteria

The choice of design is constrained by the number of tests available and the number of populations that can be sampled. When a new test is being evaluated alongside an established test, the two test-two population design is often the most efficient. When no established test exists, three tests are needed to achieve identifiability in a single population Cheung et al., 2021.

Species and Production System Considerations

The correct choice of model and interpretation of results depends on the species and production system. In wildlife populations, sampling may be opportunistic and prevalence may vary widely across regions, which affects the choice of populations for a Hui-Walter design. In production animals, management practices such as vaccination, biosecurity, and test-and-cull programs alter the prevalence and the performance of diagnostic tests.

For diseases with regional eradication programs, such as bovine tuberculosis in the UK and Ireland, diagnostic test performance estimates feed directly into surveillance and control decisions. Meta-analyzes of test sensitivity and specificity for bovine tuberculosis have used Bayesian logistic regression models to adjust for confounding factors and account for heterogeneity across studies Nuñez-Garcia et al., 2018. The resulting estimates, such as a median sensitivity of 0.50 for the single intradermal comparative cervical tuberculin test under standard interpretation, have substantial implications for the design of surveillance programs.

The MSD Veterinary Manual provides species-specific guidance on the clinical use of diagnostic tests, and the WOAH terrestrial animal health code sets international standards for test validation and surveillance. Both should be consulted when designing a test evaluation study, because the requirements for test validation may differ between species and between trade-related and research purposes.

Limitations and Common Pitfalls

The most common pitfall in Bayesian latent class analysis is the violation of conditional independence. When two tests are correlated beyond their shared association with disease status, the model will produce biased estimates unless the correlation is explicitly modelled. This is a particular concern when two serological tests detect antibodies to the same pathogen, because cross-reactivity and shared immune responses create correlation.

A second pitfall is the use of priors that are too informative relative to the data. If the prior dominates the likelihood, the posterior will reflect the prior more than the data, and the results will be misleading. Sensitivity analysis is essential to determine how much the conclusions depend on the prior specification.

A third pitfall is the misinterpretation of credible intervals. A 95% credible interval does not mean that 95% of the data fall within the interval. It means that, conditional on the model and the data, there is a 95% probability that the true value lies within the interval. This distinction is important when communicating results to stakeholders.

Finally, the results of a latent

Recognized Complications and Failure Modes

Bayesian latent class analyzes fail in characteriztic patterns. The most common is non-identifiability, where the posterior distribution does not converge to a unique solution because the model contains more parameters than the data can inform. The Hui-Walter model requires at least two tests and two populations with different prevalences, and each test must have constant sensitivity and specificity across populations. When these conditions are unmet, the analysis produces estimates that depend heavily on prior specifications instead of on the observed data Bronsvoort et al., evaluation of three 3ABC ELISAs using latent class analysis.

A second failure mode is conditional dependence between tests. Tests that measure the same biological phenomenon, such as two antibody ELISAs targeting the same immunoglobulin class, violate the conditional independence assumption. The result is biased sensitivity and specificity estimates, often with artificially narrow credible intervals. The review by Cheung et al. on Bayesian latent class analysis when the reference test is imperfect emphasizes that model structure must reflect known test correlations Cheung et al., Bayesian latent class analysis when the reference test is imperfect.

A third pattern is prior dominance. When informative priors are misspecified, the posterior estimates may simply reproduce the priors. This occurs most often in small samples or when prevalence is extreme. The posterior distribution should be compared with the prior distribution, substantial overlap that does not shift toward the data suggests the likelihood is contributing little information.

Common Errors and Corrective Actions

Less experienced analysts frequently choose the number of populations based on convenience instead of on expected prevalence differences. Two populations with nearly identical disease prevalence provide little cross-classification information, and the model struggles to separate test accuracy from prevalence. The corrective action is to sample populations with expected prevalence differences, or to add a third test to improve identifiability Branscum et al., estimation of diagnostic-test sensitivity and specificity through Bayesian modeling.

Another frequent error is treating all priors as interchangeable. Sensitivity and specificity priors should be elicited from published literature, expert opinion, or pilot data, and the analysis should include a sensitivity analysis to prior choices. The Ethiopian brucellosis study used published prior information for the three tests and reported posterior estimates that shifted meaningfully from those priors, which is the expected behavior when data are informative Getachew et al., Bayesian estimation of sensitivity and specificity of Rose Bengal, complement fixation, and indirect ELISA tests for bovine brucellosis in Ethiopia.

A third error is ignoring the target condition definition. If the latent class represents infection status but the tests detect exposure or antibody, the model estimates will be difficult to interpret. The target condition must be defined before model construction, and each test must be understood in relation to that definition.

Troubleshooting Table

Limitations of Current Evidence

The evidence base for Bayesian test evaluation in veterinary medicine is uneven. Most published applications address cattle diseases, including foot-and-mouth disease, brucellosis, and bovine tuberculosis, while companion animal and poultry applications are less common. The bovine tuberculosis meta-analysis illustrates the challenge of synthesising results across studies with different designs, populations, and test protocols, where unexplained heterogeneity produces wide credible intervals Nuñez-Garcia et al., meta-analyzes of sensitivity and specificity of ante-mortem and post-mortem diagnostic tests for bovine tuberculosis.

Expert opinion still differs on the acceptable level of prior informativeness. Some analysts advocate minimally informative priors to let data dominate, while others argue that well-elicited informative priors are essential for identifiability in small samples. Both positions have merit, and the choice should be justified explicitly in the analysis report.

Referral and Reporting Considerations

Veterinary researchers who lack formal training in Bayesian statistics should consult a biostatistician or epidemiologist before conducting a latent class analysis. The computational implementation, convergence assessment, and prior elicitation all require specialised expertise. Laboratory involvement is warranted when test results are ambiguous or when the target condition definition depends on laboratory interpretation.

Regulatory reporting obligations vary by jurisdiction and disease. For diseases subject to official control programs, such as foot-and-mouth disease or bovine tuberculosis, test evaluation results may inform national surveillance strategies and trade decisions. The World Organization for Animal Health provides international standards for surveillance and diagnostic test validation, and researchers should consult these standards when their work informs official disease status WOAH animal health surveillance standards and WOAH terrestrial animal health code. When a test evaluation reveals poor performance in a test used for regulatory purposes, the relevant veterinary authority should be notified so that surveillance interpretations can be adjusted accordingly.

Frequently Asked Questions

How many animals and test results do I need for a Bayesian latent class analysis?

Sample size requirements depend on the number of tests, the number of populations, and the strength of your priors. As a general guide, the three test, one population design used in the Ethiopian brucellosis evaluation was based on 278 serum samples and produced reasonably narrow posterior intervals for sensitivity and specificity Bayesian estimation of brucellosis test performance in Ethiopia. With two tests in two populations, you typically need at least 100 animals per population, and often more when prevalence is low or test accuracy is poor. When resources are constrained, stronger informative priors can compensate for smaller sample sizes, but this trades precision for increased reliance on prior assumptions. Run a simulation study before data collection to confirm that your planned sample size yields posterior intervals narrow enough for your intended use.

What can I do when I can only afford two tests instead of three?

Two conditionally independent tests in two or more populations can be identified if you have at least two populations with different prevalences. This design requires the assumption of conditional independence between tests within each population, which is often difficult to justify when tests detect the same biological phenomenon. The review of latent class methods for veterinary test evaluation notes that the two test, two population model is identifiable but relies heavily on the independence assumption Bayesian latent class analysis when the reference test is imperfect. If you cannot meet this assumption, consider adding a third test even if it is imperfect, or use a single test in multiple populations with strong priors on specificity. The single test design is identifiable only when informative priors are available for both sensitivity and specificity.

How do I handle Bayesian test evaluation in wildlife or exotic species where prior data are scarce?

When published estimates are unavailable for the target species, borrow priors from closely related domestic species and widen the prior variance to reflect added uncertainty. The meta-analysis of bovine tuberculosis tests demonstrates that test performance varies with host species, disease stage, and test protocol, so priors transferred across species should be treated as weakly informative Meta-analyzes of diagnostic tests for bovine tuberculosis. For wildlife, sampling design often dominates the analysis: non-random sampling, pooling, and imperfect sample storage can violate model assumptions more severely than prior misspecification. Consider using a model that accommodates correlated tests, and report sensitivity analyzes that show how posterior estimates change under different prior specifications. If the species is protected or difficult to sample, consult the relevant WOAH surveillance standards for sampling guidance WOAH animal health surveillance standards.

What records do I need to keep for a defensible Bayesian test evaluation?

Document the target condition definition, the study population and sampling frame, the exact test protocols and interpretative criteria, and the raw cross-classified results for every animal. Record the prior distributions, their sources, and the rationale for each choice before running the analysis. Keep versioned copies of the model code, the data files, and the Markov chain Monte Carlo settings, including burn-in, thinning, and chain length. The practical guidance on Bayesian latent class analysis recommends reporting convergence diagnostics and posterior predictive checks alongside the parameter estimates Bayesian latent class analysis when the reference test is imperfect. These records allow another analyst to reproduce your results and support regulatory or publication review. Store the data in a format that preserves the link between laboratory identifiers and test outcomes.

How should I explain Bayesian test evaluation to a referring veterinarian or herd owner?

Describe the problem in practical terms: no test is perfect, and comparing a new test against an imperfect reference test can mislead. Explain that Bayesian methods use the pattern of agreement and disagreement among multiple tests to estimate how often each test is right, without assuming any single test is the truth. Use a concrete example, such as the foot-and-mouth disease ELISA evaluation, where latent class analysis showed that a commercial kit had much lower sensitivity than the other two ELISAs despite being widely used Evaluation of three 3ABC ELISAs using latent class analysis. Emphasize that the result is a probability range, not a single number, and that the analysis incorporates prior knowledge from published studies. Avoid technical terms like posterior distribution unless the client asks.

When should I seek statistical collaboration instead of running the analysis myself?

Seek collaboration when your study design departs from the standard structures, when you have correlated tests, when prevalence is very low or very high, or when you need to adjust for covariates such as age, breed, or vaccination status. The review of Bayesian latent class methods emphasizes that model misspecification, especially incorrect assumptions about conditional independence, can produce seriously biased estimates Bayesian latent class analysis when the reference test is imperfect. Collaboration is also advisable when the results will inform regulatory decisions, trade certification, or disease control policy, where the WOAH terrestrial code requires defensible test performance data WOAH terrestrial animal health code. A statistician with experience in latent class models can help with prior elicitation, convergence assessment, and sensitivity analysis. The cost of collaboration is usually small relative to the cost of an analysis that fails peer review.

Related Clinical & Scientific Guides

• Evaluating Veterinary Surveillance System Attributes
• Network Analysis for Infectious Disease Spread in Animal Populations
• Randomized Controlled Trials in Veterinary Field Settings

References and Further Reading

• Evaluation of three 3ABC ELISAs for foot-and-mouth disease non-structural antibodies using latent class analysis.. 2006.
• Bayesian latent class analysis when the reference test is imperfect.. 2021.
• Bayesian Estimation of Sensitivity and Specificity of Rose Bengal, Complement Fixation, and Indirect ELISA Tests for the Diagnosis of Bovine Brucellosis in Ethiopia.. 2016.
• No gold standard estimation of the sensitivity and specificity of two molecular diagnostic protocols for Trypanosoma brucei spp. in Western Kenya.. 2010.
• Estimation of diagnostic-test sensitivity and specificity through Bayesian modeling.. 2005.
• Meta-analyzes of the sensitivity and specificity of ante-mortem and post-mortem diagnostic tests for bovine tuberculosis in the UK and Ireland.. 2018.
• WOAH Animal Health Surveillance Standards. WOAH.
• CDC Principles of Epidemiology in Public Health Practice. CDC.
• MSD Veterinary Manual, Professional Edition. MSD Veterinary Manual.

Related Articles

• Diagnostic Test Evaluation: Sensitivity and Specificity in Veterinary Medicine
• Evaluating Veterinary Surveillance System Attributes
• Interpreting Diagnostic Test Accuracy: ROC Curves in Veterinary Medicine
• Evaluating Diagnostic Tests in the Absence of a Gold Standard
• Likelihood Ratios in Veterinary Diagnostic Testing

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.