# Constructing Clinical Prediction Rules for Veterinary Medicine


## Key Takeaways

- Clinical prediction rules (CPRs) are structured instruments that combine history, physical examination, and diagnostic test results to estimate the probability of a specific diagnosis, outcome, or treatment response, requiring prospective cohort study designs for derivation and rigorous internal and external validation.
- Candidate predictor selection must be prespecified based on biological plausibility and clinical availability, with a minimum of 10-20 outcome events per predictor variable to mitigate overfitting, a common failure mode leading to poor transportability.
- Model performance is assessed by discrimination (e.g., c-statistic/AUC) and calibration (observed vs. predicted agreement), with internal validation techniques like bootstrapping crucial for correcting optimism and estimating true performance.
- Transportability of CPRs is checked through temporal and external geographic validation, as performance can drift due to differences in case-mix, measurement protocols (e.g., assay variability for lactate or creatinine), or evolving clinical practice.
- Reporting standards such as TRIPOD for prediction models and REFLECT for livestock studies are essential for transparency and reproducibility, ensuring that inclusion criteria, predictor definitions, and handling of missing data are clearly documented.
- CPRs serve as decision aids, not replacements for clinical judgment; clinicians must assess threshold instability and potential predictor drift, and regulatory reporting obligations (e.g., WOAH terrestrial animal health standards) remain paramount regardless of rule output.

---

Clinical prediction rules are structured instruments that combine history, physical examination findings, and diagnostic test results to estimate the probability of a specific diagnosis, outcome, or treatment response. In veterinary medicine, these rules address questions such as whether a dog with acute diarrhea requires hospitalization, which cats with hypertrophic cardiomyopathy will develop congestive heart failure, or whether a bovine respiratory disease case will respond to first-line antimicrobial therapy. This article explains the methodological pathway from candidate predictor identification through derivation, internal validation, external validation, and impact analysis. It serves veterinary researchers designing prognostic or diagnostic instruments and clinicians appraising published rules before adopting them in practice.

The procedural focus distinguishes a clinical prediction rule from a statistical model reported solely for association. A rule must be actionable, meaning it guides a clinical decision that changes patient management. It must also be transportable, performing acceptably across populations, settings, and time periods beyond the derivation cohort. The article covers study design selection, sample size planning, predictor coding, model building, discrimination and calibration assessment, validation strategies, and reporting standards. Species-specific examples illustrate principles, but the methodological framework applies across companion animals, livestock, and exotic species.

## At a Glance

| Parameter | Decision or Fact |
|---|---|
| Primary design | Prospective cohort with consecutive or random enrollment for derivation |
| Minimum events per predictor | 10 events per candidate predictor variable, with 20 preferred |
| Candidate predictor selection | Based on literature review, clinical plausibility, and availability in target settings |
| Model performance metrics | Discrimination (AUC or c-statistic) and calibration (observed versus predicted agreement) |
| Validation hierarchy | Internal validation, temporal validation, external geographic validation, impact analysis |
| Reporting standard | TRIPOD for prediction models, REFLECT for livestock and food animal studies |
| Common failure mode | Overfitting from too many predictors relative to outcome events |
| Transportability check | Performance drift when applied to different breeds, ages, or practice types |

## Defining the Clinical Question and Target Population

The research question determines every subsequent design choice. A prediction rule for a binary outcome, such as 30-day mortality after surgery, requires different analytical methods than a rule for a time-to-event outcome, such as survival time after lymphoma diagnosis. The question must specify the target population precisely, including species, breed predispositions, age range, disease stage, and clinical setting. A rule derived in a tertiary referral hospital may not transfer to primary care practice because the prevalence of severe disease and the spectrum of presenting signs differ substantially.

The intended use of the rule shapes its design. A rule intended to rule out a condition must maximize sensitivity, accepting lower specificity to avoid missed diagnoses. A rule intended to rule in a condition requiring expensive or invasive treatment must maximize specificity. The clinical consequence of misclassification determines the acceptable trade-off. For example, a rule identifying dogs at low risk of portosystemic shunt may safely avoid advanced imaging, whereas a rule predicting postoperative hemorrhage must not miss high-risk patients.

## Candidate Predictor Selection

Candidate predictors should be selected before data collection begins, based on biological plausibility, published evidence, and clinical availability. The [ARRIVE guidelines for reporting animal research](https://arriveguidelines.org/) emphasize prespecification of study objectives and outcomes, a principle that applies equally to predictor selection. Post hoc selection of predictors after examining associations in the data inflates the risk of spurious findings and produces rules that fail on external validation.

Each candidate predictor must be defined with explicit operational criteria. Body condition score requires a named scoring system and a defined measurement protocol. Owner-reported variables require standardized question wording. Laboratory variables require specification of the assay, sample handling, and reference intervals. Predictors that are expensive, invasive, or unavailable in the intended clinical setting should be excluded even if they show strong association, because a rule that cannot be applied in practice has no clinical utility.

The number of candidate predictors must be constrained relative to the number of outcome events. A commonly cited minimum is 10 events per predictor variable, though simulation studies suggest that 20 events per predictor provides more stable regression coefficients. With 100 outcome events, no more than 5 to 10 candidate predictors should enter the modeling process. Strategies to reduce predictor count include combining correlated variables, collapsing ordinal categories, and excluding variables with excessive missing data.

## Outcome Definition and Measurement

The outcome must be clinically meaningful, objectively measurable, and ascertained without knowledge of the predictor values. Blinded outcome assessment prevents differential verification bias, where the outcome determination is influenced by the predictor values under study. For survival outcomes, the follow-up period must be long enough to capture the event of interest, and loss to follow-up must be minimized and reported.

Outcome definitions should follow established criteria where they exist. The [WOAH terrestrial animal health standards](https://www.woah.org/en/what-we-do/standards/codes-and-manuals/terrestrial-code-online-access/) provide case definitions for notifiable diseases that can anchor outcome ascertainment in livestock studies. For other conditions, the research team must develop explicit criteria and pilot-test them for inter-observer reliability. A prediction rule built on an unreliable outcome cannot achieve reproducible performance regardless of the modeling technique.

## Study Design Selection

Prospective cohort designs are the reference standard for prediction rule derivation because they allow standardized predictor measurement, complete data collection, and blinded outcome ascertainment. Retrospective designs using medical records are faster and less expensive but introduce missing data, inconsistent measurement, and selection bias from the referral population. The [EQUATOR Network reporting guidelines](https://www.equator-network.org/) catalog the design-specific reporting standards that apply, including STROBE for observational studies and TRIPOD for prediction models.

Nested case-control designs can be efficient when the outcome is rare, but they cannot estimate absolute risk without external prevalence data. Case-control designs also amplify the risk of spectrum bias because the case and control groups may differ in ways unrelated to the prediction question. For most veterinary prediction rules, a prospective cohort with consecutive enrollment remains the preferred approach.

## Sample Size, Events Per Variable, and Model Overfitting

The number of outcome events, not the total number of animals, governs the maximum model complexity. A common heuristic requires at least 10 events per predictor variable considered for inclusion, though simulation work suggests that 10 events per variable can still produce regression coefficients with meaningful bias and unstable calibration. For binary outcomes, the events count is the smaller of the two outcome categories. A study with 400 dogs but only 40 deaths supports a model with no more than four candidate predictors if the 10 events per variable rule is applied strictly.

When the candidate predictor list exceeds what the event count supports, use penalised regression methods such as ridge, lasso, or elastic net. These methods shrink coefficients toward zero and trade a small amount of bias for a substantial reduction in variance. Lasso performs predictor selection by driving some coefficients exactly to zero, which suits situations where the goal is a sparse, deployable rule. Ridge retains all predictors with shrunken coefficients, which often gives better calibration when predictors are correlated. Elastic net sits between the two and handles groups of correlated predictors more gracefully than lasso alone.

Overfitting manifests in three detectable ways. First, apparent performance in the derivation sample is optimiztic. Second, calibration slopes fall below 1 when the model is applied to new data, meaning predictions are too extreme. Third, discrimination indices such as the concordance statistic (c-statistic) drop on external evaluation. Internal validation techniques quantify this optimizm and allow correction.

## Internal Validation Techniques

Split-sample validation, where the data are divided into derivation and validation portions, is inefficient in veterinary datasets that are often small. The validation portion is wasted for model fitting, and the resulting performance estimate is unstable because it depends heavily on the arbitrary split. Bootstrapping is preferred. Draw a bootstrap sample with replacement from the original data, fit the model in that sample, and evaluate it on both the bootstrap sample and the original data. The difference in performance between these two evaluations estimates the optimizm. Repeat this process 200 to 500 times and average the optimizm estimates, then subtract the average from the apparent performance. This procedure uses the full dataset for model fitting and produces a corrected performance estimate with lower variance than split-sample approaches.

Cross-validation is a reasonable alternative when the dataset is large enough that withholding a fold does not materially reduce the fitting sample. Ten-fold cross-validation is common. For time-to-event outcomes, use cross-validation schemes that respect the temporal structure of the data, fitting on earlier events and validating on later ones, because random folds can leak future information into the training set.

The optimizm correction applies to measures of discrimination and calibration. Report both the apparent and optimizm-corrected values. A model whose corrected c-statistic is substantially lower than its apparent value has likely overfit.

## Model Building Strategies

Full model fitting, where every candidate predictor is entered simultaneously, is appropriate when the candidate list is short and grounded in strong prior evidence. Stepwise selection procedures, whether forward, backward, or bidirectional, are discouraged because they produce coefficients that are biased away from zero, exaggerate statistical significance, and yield models that perform poorly in external samples. The standard errors from stepwise procedures do not account for the selection process, so confidence intervals are too narrow.

A more defensible strategy is hierarchical or blockwise entry. Group predictors into blocks based on clinical plausibility, for example signalment, history, physical examination findings, clinicopathologic data, and imaging findings. Enter blocks in a prespecified order and retain predictors within a block based on a threshold such as a likelihood ratio test. This approach preserves clinical interpretability and reduces the number of models examined.

For continuous predictors, avoid arbitrary dichotomisation at round numbers. Dichotomisation discards information and can create spurious effects depending on the cut point chosen. Use restricted cubic splines to allow nonlinear relationships, or prespecify clinically meaningful thresholds based on published reference intervals or prior studies. If a predictor has a known nonlinear relationship with the outcome, such as body condition score and survival in some diseases, model it accordingly instead of forcing linearity.

## Calibration and Discrimination Assessment

Discrimination describes whether predicted probabilities are higher for animals that experience the outcome than for those that do not. The c-statistic, equivalent to the area under the receiver operating characteriztic curve for binary outcomes, is the most common measure. A c-statistic of 0.5 indicates no discrimination, 0.7 to 0.8 is considered acceptable in most veterinary contexts, and values above 0.9 often signal overfitting or an overly narrow outcome definition.

Calibration describes whether predicted probabilities match observed frequencies. A calibration plot with predicted probability on the x-axis and observed proportion on the y-axis is the primary tool. The calibration intercept and slope summarize the plot numerically. An intercept near 0 with a slope near 1 indicates good calibration. An intercept greater than 0 with a slope less than 1 indicates that predictions are too extreme, a pattern typical of overfit models. The Hosmer-Lemeshow test is widely reported but has low power in small samples and is sensitive to the number of groups chosen, it should not be the sole calibration assessment.

For time-to-event outcomes, use the concordance index for discrimination and compare predicted survival curves against Kaplan-Meier estimates within risk strata for calibration.

## External Validation and Temporal Transportability

Internal validation corrects optimizm but does not establish generalizability. External validation applies the final model to data collected independently from the derivation sample. The strongest external validation uses a different institution, region, or time period. Temporal validation, where the model is applied to animals treated after the derivation period at the same institution, is weaker than geographic validation but still informative because it tests whether the model remains valid as clinical practice evolves.

When external validation reveals poor performance, the causes include case-mix differences, measurement variation between centers, and changes in the outcome definition or treatment protocols. A model developed on referral hospital data may not perform in primary care populations because the spectrum of disease severity differs. Before abandoning a model, examine whether recalibration, adjusting the intercept and slope, restores acceptable performance. If recalibration fails, the model structure itself may be inappropriate for the new setting.

The [ARRIVE guidelines for reporting animal research](https://arriveguidelines.org/) specify the minimum information needed to reproduce a study, and adherence improves the likelihood that another group can apply the same predictor definitions and outcome ascertainment in an external validation. Reporting standards for prediction models are catalogued in the [EQUATOR Network reporting guideline library](https://www.equator-network.org/), and authors should consult the relevant checklist before submitting.

## Practical Worked Example: Postoperative Peritonitis in Dogs

Consider a derivation cohort of 600 dogs undergoing gastrointestinal surgery, with 60 cases of postoperative peritonitis. The 10 events per variable rule supports six candidate predictors. The clinical team preselects six: preoperative serum albumin, duration of surgery, intraoperative contamination grade, heart rate at recovery, blood lactate at 12 hours postoperatively, and whether a surgeon with more than five years of experience performed the procedure. These are entered as a single block because each has prior evidence of association.

The full model yields a c-statistic of 0.82 with apparent performance. Bootstrap validation with 300 resamples produces an optimizm-corrected c-statistic of 0.78 and a calibration slope of 0.88, indicating modest overfitting. The team simplifies the model by removing heart rate at recovery, which has a coefficient near zero and a wide confidence interval. The reduced model has similar corrected discrimination and better calibration.

The final rule assigns points based on the remaining five predictors. A dog with a total score above a prespecified threshold is classified as high risk and receives intensive monitoring and early surgical re-exploration criteria. The threshold is chosen to achieve a sensitivity of at least 90% in the derivation sample, accepting lower specificity because the consequence of missing peritonitis is severe.

The model is then applied to a temporally distinct cohort of 200 dogs treated at the same hospital in the subsequent year. The c-statistic in this cohort is 0.74 and the calibration slope is 0.92. The team reports both the derivation and validation results, including the optimizm-corrected estimates, and publishes the full scoring algorithm so that other institutions can test it.

Species differences matter. A rule developed in dogs cannot be assumed to transfer to cats, where the same surgical procedures carry different complication rates and where reference intervals for predictors such as lactate differ. Production animals present additional constraints: the cost of monitoring, the availability of point-of-care testing, and the economic value of the individual animal all change the risk threshold that justifies intervention. The [MSD Veterinary Manual](https://www.msdvetmanual.com/) provides species-specific reference intervals and clinical norms that should inform predictor coding, and [AVMA practice resources](https://www.avma.org/resources-tools) offer guidance on professional standards for diagnostic decision-making. For food animals, the [WOAH terrestrial animal health standards](https://www.woah.org/en/what-we-do/standards/codes-and-manuals/terrestrial-code-online-access/) may define outcome ascertainment for conditions with regulatory or trade implications, and these definitions should be followed when they apply.

## Recognized Failure Modes and Early Detection

Clinical prediction rules fail in characteriztic patterns. The most common is deterioration of performance when applied outside the derivation setting. A rule built on intensive care unit dogs may lose discrimination when applied to first-opinion consultations, because disease prevalence, case severity, and referral filters differ. Detect this early by tracking the rule's calibration slope and discrimination index in a prospective pilot cohort before full deployment.

A second failure mode is predictor drift. Variables such as lactate, creatinine, or band neutrophil count may be measured by different analyzers, at different times relative to intervention, or under different fasting states across sites. The rule silently degrades as measurement protocols diverge. Early detection requires a data audit that compares predictor distributions and measurement timing across participating sites.

A third mode is automation bias in users. Clinicians who trust the rule may stop collecting data the rule does not use, which erodes the very variables needed for future recalibration. Monitor documentation completeness for candidate predictors that were excluded from the final model.

A fourth mode is threshold instability. The optimal probability cut-point shifts with case mix and clinical setting. A rule that flags high risk at 40% probability in a referral hospital may need a 25% threshold in primary care to achieve equivalent sensitivity. Re-estimate thresholds on local data instead of assuming portability.

## Common Errors in Rule Construction

Less experienced analysts frequently dichotomise continuous predictors at data-derived cut-points before modeling. This discards information and inflates apparent effects. The corrective action is to model continuous predictors with flexible terms, such as restricted cubic splines, and only dichotomise after examining the shape of the association.

A second error is selecting predictors by univariable screening with a liberal p-value threshold, then entering survivors into multivariable regression. This procedure distorts coefficient estimates and overstates predictive performance. Use penalised regression, bootstrap selection, or a prespecified model based on clinical reasoning instead.

A third error is evaluating model performance on the same data used to build it. Reported area under the receiver operating characteriztic curve from derivation data is optimiztically biased. Use internal validation techniques such as bootstrapping or cross-validation, and report optimizm-corrected estimates.

A fourth error is ignoring clustering. When multiple records come from the same animal, such as repeated visits, or from the same clinic, observations are not independent. Failure to account for clustering produces confidence intervals that are too narrow. Use mixed-effects models or cluster-robust standard errors.

A fifth error is confusing association with causation when interpreting predictor coefficients. A rule can be perfectly predictive without any predictor being causal. Do not use the rule's coefficients to infer biological mechanisms or to guide therapeutic decisions that assume a causal link.

## Limitations of Current Evidence and Contested Expert Opinion

The veterinary prediction literature is sparse relative to human medicine. Most published rules are single-center, small-sample, and internally validated only. External validation studies are uncommon, and head-to-head comparisons of competing rules are rare. The [EQUATOR Network reporting guidelines](https://www.equator-network.org/) catalogue the reporting standards that should govern such studies, but adherence in veterinary journals remains inconsistent.

Expert opinion diverges on several points. One is whether rules should be species-specific or can be adapted across closely related species. Some argue that cardiopulmonary and hemodynamic predictors transfer between dogs and cats with adjustment, while others contend that species differences in physiology and disease expression invalidate cross-species borrowing. A second contested area is the role of subjective clinical judgment. Some authorities advocate combining rule output with clinician gestalt, while others argue that structured rules outperform unstructured judgment and should replace it. A third area is the handling of missing predictor values at the point of care. Imputation methods developed for research datasets may not be feasible in a busy clinic, and point-of-care rules often require a fallback algorithm that uses only the predictors actually measured.

The [ARRIVE guidelines](https://arriveguidelines.org/) emphasize transparent reporting of animal characteriztics, housing, and experimental procedures. For prediction rules, analogous transparency requires reporting of inclusion criteria, predictor definitions, and handling of missing data. Where this information is absent, the rule cannot be evaluated or reproduced.

## Escalation, Referral, and Reporting Triggers

Prediction rules are decision aids, not substitutes for clinical judgment. Escalate when the rule's output conflicts with a clinician's assessment of immediate physiological instability. A rule that predicts low risk does not override evidence of progressive deterioration. Referral to a specialist or intensive care facility is warranted when the rule identifies high risk and the current setting cannot provide the monitoring or intervention the risk implies.

Laboratory involvement is indicated when predictor values fall outside the range used in model derivation. A rule built on analyzer-specific reference intervals may not accommodate values from a different laboratory platform. Consult the laboratory about assay comparability before applying the rule to a new patient population.

Regulatory reporting obligations are independent of prediction rule outputs. Where a rule identifies a pattern of unexpected deaths or adverse events, the clinician must still follow statutory reporting requirements for pharmacovigilance or notifiable disease. The [WOAH terrestrial animal health standards](https://www.woah.org/en/what-we-do/standards/codes-and-manuals/terrestrial-code-online-access/) define which diseases require reporting to veterinary authorities, and the [AVMA practice resources](https://www.avma.org/resources-tools) provide guidance on professional obligations in the United States. A prediction rule never supersedes these duties.

## Troubleshooting Table

| Observation | Likely Cause | Discriminating Check |
| --- | --- | --- |
| Rule performs well in derivation, poorly in new site | Case mix or measurement protocol differs | Compare predictor distributions and outcome prevalence between sites |
| Calibration slope consistently below 1 | Model overfit in derivation | Re-fit with penalised regression, re-validate externally |
| Discrimination declines over time | Predictor drift or population shift | Re-measure predictor distributions quarterly, monitor outcome incidence |
| Clinicians stop recording unused predictors | Automation bias | Audit medical records for completeness of candidate predictor fields |
| Rule output conflicts with clinical impression | Threshold mismatch or unmeasured confounder | Re-estimate local threshold, assess for variables absent from the model |
| Confidence intervals implausibly narrow | Clustering ignored | Check for repeated measures per animal or per clinic, fit mixed-effects model |

## Frequently Asked Questions

### How Do I Build a Prediction Rule When My Caseload Cannot Support a Large Derivation Cohort?

Small clinical populations are a common constraint in veterinary research. When a single institution cannot supply the required events per variable, consider multi-center collaboration or extending the accrual window. If neither is feasible, simplify the model by reducing candidate predictors before modeling, using prior literature and clinical reasoning instead of data-driven selection. A rule derived from fewer than 10 events per variable risks severe overfitting and will likely fail external validation. Report the events per variable achieved and frame the result as preliminary. The ARRIVE guidelines provide a framework for transparent reporting of sample size constraints and their implications for interpretation.

### What Is the Minimum Equipment and Laboratory Infrastructure Needed to Apply a Rule in Practice?

The required infrastructure depends entirely on the predictors the rule contains. A rule built around physical examination findings and historical data needs no laboratory equipment. A rule requiring advanced imaging or specialised assays is unusable in settings without those resources. Before adopting a published rule, audit your practice against its predictor list. If a predictor is unavailable, you cannot substitute a correlated variable without re-validating the rule. Consider whether the rule's developers reported a reduced model using only routinely available data. The MSD Veterinary Manual offers species-specific guidance on which diagnostic tests are considered standard in general practice.

### Does a Rule Developed in Dogs Transfer to Cats or Other Species?

No. Species differ in physiology, disease expression, and reference intervals, so a rule validated in one species cannot be assumed to work in another. Even breeds within a species can shift predictor distributions. Transferring a rule across species requires full re-derivation or at minimum external validation in the target species with recalibration of intercept and coefficients. Some predictors may lose relevance entirely, while new species-specific predictors may need to be added. Report any cross-species application as exploratory. The WOAH terrestrial animal health standards emphasize that diagnostic and prognostic tools must be fit for purpose within the specific animal population where they will be applied.

### How Should I Document Rule Application in the Medical Record?

Record the rule's name, version, the date of application, and the raw predictor values, also the final risk score. Note who calculated the score and whether any predictors were missing or estimated. Document the clinical decision that followed the score and any deviation from the rule's recommendation, with justification. This creates an audit trail that supports later evaluation of the rule's performance in your population. If the rule is part of a research study, the ARRIVE guidelines specify the minimum information needed for transparent reporting of how clinical assessments were conducted and recorded.

### How Do I Explain a Prediction Rule to an Owner Who Wants a Definite Answer?

Explain that the rule provides a probability, not a certainty. Use natural frequencies instead of percentages where possible, for example "about 3 in 10 dogs with these findings will develop the complication." Clarify that the rule aggregates experience from many similar patients and that the individual animal may still follow an unexpected course. Distinguish between what the rule suggests and what you recommend, since treatment decisions incorporate owner resources, patient temperament, and concurrent disease. Offer the reasoning behind the recommendation in plain language and invite questions. The AVMA practice resources include communication guidance for discussing risk and uncertainty with clients.

### What Should I Do When a Rule's Recommendation Conflicts With My Clinical Judgment?

First, check whether the rule was applied correctly. Verify each predictor value and confirm the rule is appropriate for the patient's species, age, and clinical context. If the rule was applied correctly and still conflicts with your judgment, document both positions in the record and explain your reasoning for overriding the rule. A rule is a decision aid, not a replacement for clinical reasoning. Consider whether the conflict arises from a predictor distribution outside the rule's derivation range, which is a recognized limitation. Report such cases to the rule's developers if possible, since they may indicate a need for recalibration or extension of the rule's scope.

## Related Clinical & Scientific Guides

* [Conducting Systematic Reviews of Veterinary Diagnostic Test Accuracy](/knowledge/veterinary-medicine/veterinary-research-methods/conducting-systematic-reviews-veterinary-diagnostic-test-accuracy)
* [Bias in Veterinary Research: Types, Sources, and Mitigation](/knowledge/veterinary-medicine/veterinary-research-methods/bias-veterinary-research-types-sources-mitigation)
* [Cluster Randomized Trials in Veterinary Research: Design and Analysis](/knowledge/veterinary-medicine/veterinary-research-methods/cluster-randomized-trials-veterinary-research-design-analysis)


## References and Further Reading

- [ARRIVE Guidelines 2.0 for Reporting Animal Research](https://arriveguidelines.org/). PLOS Biology, 2020.
- [EQUATOR Network Reporting Guidelines](https://www.equator-network.org/). EQUATOR Network.
- [MSD Veterinary Manual, Professional Edition](https://www.msdvetmanual.com/). MSD Veterinary Manual.
- [American Veterinary Medical Association Practice Resources](https://www.avma.org/resources-tools). American Veterinary Medical Association.
- [WOAH Terrestrial Animal Health Code](https://www.woah.org/en/what-we-do/standards/codes-and-manuals/terrestrial-code-online-access/). WOAH.

## Related Articles

- [Designing Adaptive Clinical Trials for Veterinary Medicine](/knowledge/veterinary-medicine/veterinary-research-methods/designing-adaptive-clinical-trials-veterinary-medicine)
- [Formulating Answerable Clinical Questions in Veterinary Medicine](/knowledge/veterinary-medicine/veterinary-research-methods/formulating-answerable-clinical-questions-veterinary-medicine)
- [Conducting Pharmacovigilance Studies in Veterinary Medicine](/knowledge/veterinary-medicine/veterinary-research-methods/conducting-pharmacovigilance-studies-veterinary-medicine)
- [Blinding in Veterinary Clinical Research: Methods and Challenges](/knowledge/veterinary-medicine/veterinary-research-methods/blinding-veterinary-clinical-research-methods-challenges)
- [Handling Missing Data in Veterinary Clinical Research](/knowledge/veterinary-medicine/veterinary-research-methods/handling-missing-data-veterinary-clinical-research)

> 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.