Evaluating Diagnostic Tests in the Absence of a Gold Standard

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

Key Takeaways

• Evaluating diagnostic tests without a gold standard necessitates latent class analysis (LCA), which models true disease status as an unobserved variable, allowing estimation of sensitivity, specificity, and prevalence for each test.
• LCA requires at least two conditionally independent tests or three tests with a known dependence structure, applied to populations with varying disease prevalence to ensure parameter identifiability.
• Bayesian LCA, utilizing Markov chain Monte Carlo estimation, is preferred for its ability to incorporate prior knowledge, handle complex dependence structures (e.g., between bacteriology and PCR for bovine tuberculosis), and provide posterior distributions for parameter uncertainty.
• Conditional independence between tests, given true disease status, is a critical assumption; violations, such as using two serological tests detecting antibodies to the same antigen, can lead to biased sensitivity and specificity estimates.
• Study design is paramount, requiring sampling from multiple populations with differing prevalences and adequate sample sizes, often determined through simulation, to achieve reliable parameter estimation and narrow credible intervals.
• Reporting of LCA results should include posterior medians with 95% credible intervals, clearly distinguishing them from frequentist confidence intervals, and address potential failure modes like non-identifiability or unmodeled conditional dependence.

Veterinary researchers frequently confront a fundamental problem in diagnostic test evaluation: no perfect reference standard exists for the condition under study. Ante-mortem diagnosis of bovine respiratory disease, confirmatory testing for bovine tuberculosis, and serological detection of chronic infections such as toxoplasmosis all share this limitation. When a gold standard is unavailable, conventional estimates of sensitivity and specificity become biased because the imperfect reference test misclassifies animals, and the magnitude and direction of that bias are generally unknown.

This article provides a procedural framework for evaluating diagnostic test accuracy when no gold standard exists. It serves veterinary researchers designing validation studies, epidemiologists interpreting published test performance data, and clinicians who need to understand why published sensitivity and specificity estimates vary across populations. The methods described include latent class analysis, Bayesian latent class modeling, and the assumptions that govern their valid application. Simple sensitivity and specificity estimation against a known reference standard is outside the scope of this article.

The central question addressed is practical: given two or more imperfect tests applied to a population of unknown disease status, how can the true accuracy of each test be estimated? The answer requires abandoning the fiction of a perfect reference test and instead modeling disease status as an unobserved, or latent, variable.

At a Glance

The Problem of the Imperfect Reference Standard

Diagnostic test evaluation traditionally proceeds by comparing a candidate test against a reference standard. When that reference standard is itself imperfect, the resulting sensitivity and specificity estimates are distorted. The direction of bias depends on whether the reference test errors are differential or non-differential with respect to the candidate test. Non-differential misclassification of the reference test generally biases sensitivity and specificity estimates toward their null values, while differential misclassification can bias estimates in either direction.

Consider the evaluation of clinical illness detection for bovine respiratory disease in feedlot cattle. Lung lesions at slaughter have been used as a reference test, yet clinical illness and lung lesions are biologically distinct phenomena. A systematic review and hierarchical Bayesian latent-class meta-analysis of this question found that clinical illness detection had a pooled sensitivity of 0.27 and specificity of 0.92, while lung lesions had substantially different accuracy, confirming that neither test could serve as a gold standard for the other Diagnostic accuracy of clinical illness for BRD diagnosis in beef cattle placed in feedlots.

The same problem arises in bovine tuberculosis confirmatory testing. Bacteriology, histopathology, and PCR are all used for confirmatory diagnosis, but none is perfect. Bacteriology suffers from slow growth and contamination, histopathology requires skilled interpretation, and PCR detects DNA from both viable and non-viable organizms. A latent class analysis of 5,211 animals in France estimated PCR sensitivity at 87.7% and bacteriology sensitivity at 78.1%, with both tests exceeding 97% specificity Estimation of sensitivity and specificity of bacteriology, histopathology and PCR for bovine tuberculosis. These estimates would have been distorted had either test been treated as a gold standard.

Latent Class Analysis: Conceptual Foundation

Latent class analysis treats the true disease status of each animal as an unobserved categorical variable. The observed test results are modeled as functions of this latent status plus test-specific error parameters. The model estimates three sets of parameters simultaneously: the sensitivity and specificity of each test, and the prevalence of disease in each sampled population.

The logic is analogous to solving a system of equations. With two tests and one population, the model has more unknown parameters than observable data points, and the system is underdetermined. Adding a second population with different disease prevalence, or a third test, provides additional constraints that allow unique estimation. This identifiability requirement is the most common practical obstacle in study design.

Bayesian estimation is the standard implementation. Prior distributions are specified for all unknown parameters, and Markov chain Monte Carlo methods generate posterior distributions. The posterior distributions for sensitivity and specificity directly express the uncertainty remaining after combining prior knowledge with observed data. This framework accommodates the hierarchical structure of multi-study data, as demonstrated in the bovine respiratory disease meta-analysis where within-study and between-study variability in clinical illness accuracy were modeled jointly Diagnostic accuracy of clinical illness for BRD diagnosis in beef cattle placed in feedlots.

Conditional Dependence and Independence

The foundational assumption of standard latent class analysis is conditional independence: given the true disease status, the results of different tests are statistically independent. This assumption holds when tests measure different biological phenomena. Clinical illness detection and lung lesion evaluation for bovine respiratory disease operate on different biological principles, supporting the conditional independence assumption in that analysis Diagnostic accuracy of clinical illness for BRD diagnosis in beef cattle placed in feedlots.

When tests share a biological basis, conditional independence fails. Two serological tests detecting antibodies to the same antigen, or bacteriology and PCR both detecting the same organizm, are likely to be correlated within disease status categories. Unmodeled conditional dependence inflates sensitivity estimates and deflates specificity estimates. Modern Bayesian latent class models allow dependence to be modeled explicitly, typically by adding covariance parameters between test pairs. The bovine tuberculosis analyzes cited above modeled dependence between bacteriology and PCR, and between the single intradermal comparative tuberculin test and the gamma-interferon test, while assuming independence from histopathology and the multiplex immunoassay respectively Estimation of sensitivity and specificity of bacteriology, histopathology and PCR for bovine tuberculosis, Using latent class analysis to estimate the test characteriztics of the gamma-interferon test, the single intradermal comparative tuberculin test and a multiplex immunoassay under Irish conditions.

Study Design Requirements

The validity of latent class analysis depends on study design decisions made before data collection. Sampling from at least two populations with expected differences in disease prevalence improves identifiability and precision. The Irish bovine tuberculosis study sampled herds with no known tuberculosis problems and herds experiencing confirmed breakdowns, producing prevalence differences that supported stable estimation Using latent class analysis to estimate the test characteriztics of the gamma-interferon test, the single intradermal comparative tuberculin test and a multiplex immunoassay under Irish conditions. Similarly, a porcine toxoplasmosis study sampled indoor and outdoor farms, modeling within-farm clustering with random effects to account for the non-independence of animals from the same farm Assessment of diagnostic accuracy of a commercial ELISA for the detection of Toxoplasma gondii infection in pigs.

Sample size requirements for latent class analysis exceed those for conventional test evaluation. Simulation studies are recommended during the design phase to verify that the planned sample sizes and test combinations will yield identifiable parameters with acceptably narrow posterior intervals. The number of tests, the number of populations, the expected prevalence values, and the anticipated test accuracy all influence the required sample size.

Model Selection: A Decision Framework

Choosing among latent class approaches depends on the data structure, the number of tests available, and the biological relationship between tests. The first decision concerns whether a frequentist or Bayesian framework is appropriate. Frequentist latent class analysis, typically implemented via maximum likelihood estimation, works well with large samples and simple model structures. Bayesian methods are preferred when prior information exists, when sample sizes are modest, or when the model must accommodate conditional dependence between tests, as demonstrated in evaluations of bovine tuberculosis diagnostics where bacteriology and PCR share biological basis Courcoul et al., estimation of bacteriology, histopathology and PCR accuracy using latent class analysis.

The second decision concerns the number of populations sampled. A single-population design with two tests is unidentified without strong priors. Three tests in a single population, or two tests across two populations with different prevalences, provide identifiable models under conditional independence. When conditional dependence is expected, at least two populations and three tests, or two tests with a third test known to be conditionally independent, are required.

The third decision concerns how to handle conditional dependence. Tests based on similar biological principles, such as two serological assays detecting antibodies to the same pathogen, will likely be correlated conditional on true infection status. The bovine tuberculosis literature illustrates this pattern repeatedly: interferon-gamma release assays and tuberculin skin tests are often modelled as conditionally dependent, while tests based on different biological principles, such as culture and histopathology, can be treated as independent Clegg et al., latent class analysis of tuberculin and interferon-gamma tests under Irish conditions.

Clustering deserves explicit attention. Animals within a herd are not independent observations. Herd-level factors, including management practices and pathogen strain, influence both true infection status and test results. Ignoring this clustering can produce artificially narrow credible intervals and biased accuracy estimates. The evaluation of a commercial ELISA for Toxoplasma gondii in pigs incorporated random effects to adjust for within-farm clustering, recognizing that indoor and outdoor production systems represent distinct epidemiological contexts Basso et al., Bayesian latent class assessment of a commercial ELISA for Toxoplasma gondii in pigs. When study populations are drawn from multiple herds or production systems, the model should include a herd-level random effect.

Worked Example: Three Tests, Two Populations

Consider a study evaluating two ante-mortem tests for bovine tuberculosis, the single intradermal comparative cervical tuberculin test and an interferon-gamma release assay, against post-mortem meat inspection. The data come from two populations: herds with a confirmed breakdown and herds with no known tuberculosis history. This design mirrors published work in Northern Ireland, where retrospective Bayesian latent class analysis was applied to 71,185 cattle from 806 chronically infected herds Lahuerta-Marin et al., Bayesian latent class estimation of bovine tuberculosis test accuracy in Northern Ireland.

The model structure proceeds as follows. Each of the three tests has a sensitivity and specificity parameter. Each population has a prevalence parameter. The tuberculin test and interferon-gamma assay are modelled as conditionally dependent, because both measure cell-mediated immune responses to mycobacterial antigens. Meat inspection is modelled as conditionally independent of the two ante-mortem tests, because it detects gross lesions instead of immune responses.

Prior distributions require specification. For sensitivity and specificity, beta distributions centerd on published estimates with wide variance are appropriate. For prevalence, a uniform prior between zero and one works when no external information exists, though informative priors based on regional surveillance data can improve precision. The analysis should include a sensitivity analysis comparing results under alternative priors to confirm that conclusions are not driven by prior specification.

The output provides posterior medians and 95% credible intervals for each parameter. In the Northern Ireland example, the standard interpretation of the tuberculin test had estimated sensitivity of 40.5% to 57.7% and specificity of 96.3% to 99.7%, while the interferon-gamma assay had higher sensitivity but lower specificity Lahuerta-Marin et al., Bayesian latent class estimation of bovine tuberculosis test accuracy in Northern Ireland. These estimates differ meaningfully from those derived in populations with known infection status, illustrating why latent class methods matter when no reference standard exists.

Stratified analyzes can reveal heterogeneity. The same study found lower sensitivity estimates in dairy cattle than beef cattle, a difference attributed to variations in disease presentation and management. When such heterogeneity is anticipated, the model can include covariates or the analysis can be run separately within strata.

Interpreting and Reporting Results

Latent class analysis outputs require careful interpretation. Posterior credible intervals reflect uncertainty from both sampling variability and model structure. Wide intervals, as seen in the pooled sensitivity estimate of 0.27 with a credible interval from 0.12 to 0.65 for clinical illness detection of bovine respiratory disease, indicate substantial uncertainty that should temper clinical recommendations Timsit et al., hierarchical Bayesian latent-class meta-analysis of bovine respiratory disease diagnosis.

Convergence diagnostics matter. Multiple chains with dispersed starting values should be run, and trace plots, Gelman-Rubin statistics, and effective sample sizes should be examined. Poor convergence indicates model misspecification or identifiability problems. Common remedies include simplifying the dependence structure, adding informative priors, or collecting additional data.

Reporting should follow established epidemiological standards. The CDC principles of epidemiology in public health practice provide a framework for describing study design, population selection, and analytic methods. International reporting for notifiable diseases should align with WOAH animal health surveillance standards, which specify requirements for test validation and surveillance design.

Species and Production System Considerations

The correct analytical approach varies with species and production context. In food animals, where individual animal value is lower and slaughter confirmation is feasible, post-mortem tests can serve as one of the three required tests. In companion animals, where post-mortem sampling is rarely acceptable, the analyst must rely on multiple ante-mortem tests, often with smaller sample sizes and consequently wider credible intervals.

Production system affects both disease prevalence and test performance. Indoor and outdoor pig production systems present different Toxoplasma gondii exposure patterns, and pooling data across systems without adjustment can bias estimates Basso et al., Bayesian latent class assessment of a commercial ELISA for Toxoplasma gondii in pigs. Similarly, dairy and beef cattle differ in tuberculosis test performance, as demonstrated in the Northern Ireland analysis Lahuerta-Marin et al., Bayesian latent class estimation of bovine tuberculosis test accuracy in Northern Ireland.

Resource constraints also shape design choices. When only two tests are available and a third cannot be added, the analyst must rely on multiple populations and informative priors. When funding limits sample size, a Bayesian approach with informative priors derived from previous studies may be the only feasible option. The MSD Veterinary Manual provides species-specific background on disease pathogenesis and test interpretation that can inform prior specification.

Common Failure Modes

Model non-identifiability is the most common failure. Symptoms include non-convergence, extreme posterior estimates, or estimates that track the priors. The remedy is structural: add a test, add a population, or impose stronger priors.

Conditional dependence misspecification produces biased estimates. If two tests are modelled as independent when they are in fact dependent, sensitivity estimates tend to be inflated. The direction of bias depends on the correlation structure, but the practical consequence is overconfidence in test accuracy.

Population heterogeneity can masquerade as test inaccuracy. If prevalence varies within a sampled population, the latent class model may absorb this variation into test parameter estimates. Stratified analysis or random effects models address this problem.

Finally, the assumption that latent class membership corresponds to a single underlying disease state deserves scrutiny. Conditions with a spectrum of severity, such as allergic bronchopulmonary aspergillosis, present particular challenges because the boundary between affected and unaffected is itself uncertain Saxena et al., latent class analysis of diagnostic criteria for allergic bronchopulmonary aspergillosis. In such settings, the latent class may represent a composite state, and the estimated test accuracy applies to that composite instead of to a precisely defined pathological entity.

Recognized Complications and Early Detection

Latent class analyzes fail in characteriztic ways, and most failures become visible only after the model has run. The most common complication is non-identifiability, where the data cannot support the number of parameters being estimated. A model with three tests in one population has seven estimable parameters from seven data cells, leaving no degrees of freedom for prior information to be updated. The model will converge, but posterior estimates will mirror the priors. Detect this early by comparing posterior distributions to prior distributions. If the 95% credible intervals have not narrowed appreciably, the data are not informing the estimates.

A second complication is poor Markov chain Monte Carlo convergence. Trace plots that show drifting chains, high autocorrelation, or multiple distinct modes indicate that the sampler has not explored the full posterior surface. Run at least three chains from dispersed starting values and examine the Gelman-Rubin statistic. Values above 1.1 warrant longer chains, reparameterisation, or reconsideration of the model structure.

Conditional dependence between tests is the third major failure mode. When two tests rely on the same biological principle, such as bacteriology and PCR both detecting mycobacterial DNA or viable organizms, their errors are correlated. Ignoring this dependence biases sensitivity estimates upward and specificity estimates downward. The discriminating check is to fit both an independent and a dependent model and compare them using the deviance information criterion or posterior predictive checks. The bovine tuberculosis literature provides repeated examples of this necessity, with models explicitly allowing dependence between bacteriology and PCR while assuming independence from histopathology Courcoul et al., estimation of sensitivity and specificity of bacteriology, histopathology and PCR for bovine tuberculosis using latent class analysis.

Common Errors and Corrective Action

Less experienced analysts frequently select tests that are too similar. Two serological tests measuring the same immunoglobulin class against overlapping antigens will violate conditional independence assumptions, yet the model will appear to run successfully. The corrective action is to select tests based on distinct biological principles, as done in the Irish bovine tuberculosis work where the cell-mediated and humoral arms of the immune response were deliberately paired Clegg et al., latent class analysis of the gamma-interferon test, single intradermal comparative tuberculin test and multiplex immunoassay under Irish conditions.

A second common error is pooling populations with different disease prevalences into a single analysis. Prevalence heterogeneity is an assumption of the two-population latent class model, not a nuisance to be eliminated. Stratify by herd type, production system, or region before analysis. The Northern Ireland bovine tuberculosis study found sensitivity estimates were lower in dairy cattle than in beef cattle, a difference that would have been obscured by pooling Lahuerta-Marin et al., Bayesian latent class estimation of sensitivity and specificity for bovine tuberculosis in chronically infected herds in Northern Ireland.

A third error is misinterpreting posterior credible intervals as frequentist confidence intervals. The credible interval describes the posterior probability that the parameter lies within the stated bounds, conditional on the model and priors. It does not describe the long-run frequency of capturing the true value. Report intervals with their Bayesian interpretation explicitly.

Limitations of Current Evidence

The evidence base for latent class analyzes in veterinary medicine is concentrated in a small number of disease systems. Bovine tuberculosis dominates the literature, with multiple independent studies across Ireland, France, and Northern Ireland. Other diseases are comparatively neglected. Toxoplasma gondii serology in pigs has been evaluated with latent class methods, but the study populations were small, with fewer than 300 samples across two production systems Basso et al., assessment of diagnostic accuracy of a commercial ELISA for Toxoplasma gondii in pigs using a Bayesian latent class approach. The bovine respiratory disease complex has been addressed through hierarchical meta-analysis, but the pooled sensitivity estimates carried wide credible intervals, reflecting substantial between-study heterogeneity Timsit et al., diagnostic accuracy of clinical illness for bovine respiratory disease in feedlot cattle, hierarchical Bayesian latent-class meta-analysis.

Expert opinion still differs on the appropriate prior specifications for specificity. Some analysts argue for informative priors based on experimental challenge data, while others prefer vague priors to let the field data speak. The choice materially affects posterior estimates when data are sparse. There is no consensus resolution, and the sensitivity of conclusions to prior choice should be reported.

Referral and Escalation

Referral to a veterinary epidemiologist or biostatistician is warranted when the study design involves more than three tests, when conditional dependence is suspected but cannot be modelled confidently, or when regulatory decisions depend on the results. Laboratories should be consulted early in the design phase to confirm that test protocols and cut-offs will be applied consistently across all study sites. Cut-off drift between batches or laboratories invalidates the assumption that a single sensitivity and specificity describe the test.

Regulatory reporting obligations apply when the tests under evaluation are used for trade, surveillance, or disease control programs. The World Organization for Animal Health maintains standards for test validation and surveillance that should be consulted before results are used to support official claims about freedom from infection or test performance WOAH animal health surveillance standards. Where results will inform international trade, the terrestrial animal health code provisions on diagnostic testing apply WOAH terrestrial animal health code.

Frequently Asked Questions

How Many Animals and Tests Do I Need for a Latent Class Analysis?

There is no universal minimum, but the model must be identifiable. At minimum, you need at least three tests applied to a single population, or two tests applied to at least two populations with different expected prevalences. More informative designs use three tests across two or more populations. Sample size depends on the expected sensitivity and specificity of each test, the anticipated prevalence, and the strength of conditional dependence. As a practical guide, studies with fewer than 200 animals per population often produce wide credible intervals that cannot discriminate between plausible test accuracies. When resources are limited, prioritize a two-population, two-test design over a single-population, three-test design because prevalence differences help separate test accuracy from disease frequency. Consult a biostatistician before data collection to confirm identifiability for your specific design.

What Can I Do When Only Two Imperfect Tests Are Available?

A two-test, single-population design is not identifiable without strong assumptions. You can proceed if you are willing to fix one parameter, such as assuming conditional independence between tests or fixing the specificity of one test from prior evidence. Alternatively, recruit a second population with a distinctly different expected prevalence, which permits estimation of all parameters. If neither option is feasible, consider a latent class model that incorporates continuous test results instead of dichotomised outcomes, as this can improve identifiability. Another option is to use a composite reference standard, but this reintroduces bias if the composite is imperfect. When the evidence base is thin, report results as ranges instead of point estimates and state clearly which assumptions drove the analysis. The WOAH terrestrial animal health standards emphasize transparent reporting of test validation limitations.

How Do I Choose Between a Frequentist and a Bayesian Latent Class Approach?

Bayesian methods dominate veterinary latent class applications because they handle small samples, incorporate prior information, and naturally produce credible intervals. Frequentist approaches, such as maximum likelihood estimation, are useful when you have large datasets and wish to avoid specifying priors. Choose Bayesian analysis when prior information exists from earlier studies, when data are sparse, or when you need to model conditional dependence flexibly. Choose frequentist methods when you have thousands of observations and prefer estimates driven entirely by the data. In practice, Bayesian models with weakly informative priors often behave similarly to frequentist approaches in large samples. The choice matters most in small studies where priors can materially shift estimates. For regulatory submissions, check whether the relevant authority expects a particular framework. The WOAH terrestrial code online access provides guidance on acceptable validation approaches for trade-related testing.

How Should I Explain Latent Class Results to a Client or Supervisor?

Focus on the practical consequence instead of the statistical machinery. State that no available test is perfect, and the analysis estimates how often each test correctly identifies infected and uninfected animals in your population. Give the sensitivity and specificity with their uncertainty ranges, and explain what those ranges mean for interpreting individual results. For example, a test with 90% sensitivity will miss roughly one in ten infected animals. If the supervisor asks why you did not use a gold standard, explain that the reference test itself has error, and latent class analysis accounts for that error. Provide a one-page summary with the test characteriztics, the population studied, and the main limitations. The CDC principles of epidemiology offer useful language for explaining diagnostic test properties to non-specialists.

Can I Apply Results From One Species or Production System to Another?

Transferability is limited. Test accuracy depends on disease prevalence, infection stage, sample quality, laboratory protocols, and host factors. For example, sensitivity estimates for bovine tuberculosis tests differ between dairy and beef cattle, as shown in a Bayesian latent class analysis of chronically infected herds in Northern Ireland, where sensitivity was lower in dairy cattle than in beef cattle. Similarly, a commercial ELISA for Toxoplasma gondii in pigs performed differently in indoor versus outdoor production systems. If you must apply results from another setting, treat the published values as priors and validate locally with a smaller study. Report the population characteriztics whenever you publish or present estimates so others can judge transferability. When local validation is impossible, state the uncertainty explicitly and recommend conservative interpretation of test results.

What Records Do I Need to Keep for a Latent Class Study?

Document the sampling frame, inclusion and exclusion criteria, and the date and location of sampling for every animal. Record the laboratory that performed each test, the specific test version or kit lot, the technician, and the exact interpretation threshold used. Note any deviations from the standard protocol, including sample storage times, transport conditions, and repeat testing. Keep the raw test results, also the final positive or negative classification, because continuous data allow more flexible modeling. Record the prevalence context, such as known disease status of the herd or region, because this informs prior specification. Preserve the analysis code and model specification, including priors and convergence diagnostics. The AVMA practice resources provide general guidance on medical record keeping that applies to research data as well as clinical records.

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

• Diagnostic accuracy of clinical illness for bovine respiratory disease (BRD) diagnosis in beef cattle placed in feedlots: A systematic literature review and hierarchical Bayesian latent-class meta-analysis.. 2016.
• Estimation of sensitivity and specificity of bacteriology, histopathology and PCR for the confirmatory diagnosis of bovine tuberculosis using latent class analysis.. 2014.
• Using latent class analysis to estimate the test characteriztics of the γ-interferon test, the single intradermal comparative tuberculin test and a multiplex immunoassay under Irish conditions.. 2011.
• Assessment of diagnostic accuracy of a commercial ELISA for the detection of Toxoplasma gondii infection in pigs compared with IFAT, TgSAG1-ELISA and Western blot, using a Bayesian latent class approach.. 2013.
• Bayesian latent class estimation of sensitivity and specificity parameters of diagnostic tests for bovine tuberculosis in chronically infected herds in Northern Ireland.. 2018.
• Which Are the Optimal Criteria for the Diagnosis of Allergic Bronchopulmonary Aspergillosis? A Latent Class Analysis.. 2021.
• 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

• Bayesian Methods for Diagnostic Test Evaluation in Animals
• Evaluating Veterinary Surveillance System Attributes
• Sensitivity Analysis in Veterinary Disease Models
• Likelihood Ratios in Veterinary Diagnostic Testing
• Risk Factor Analysis for Disease in Animal Populations

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.