Handling Missing Data in Veterinary Clinical Research

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

Key Takeaways

• Missing data in veterinary clinical research arise from diverse causes including animal mortality, owner withdrawal, sample degradation (e.g., hemolysis), and equipment failure, necessitating careful consideration of the missingness mechanism (MCAR, MAR, MNAR) before analysis.
• Complete-case analysis, while simple and valid under MCAR, can lead to significant loss of statistical power and introduce bias under MAR or MNAR, especially when the proportion of missing data exceeds 5-10% for primary outcomes.
• Multiple imputation is the recommended default strategy for MAR data, requiring a pre-specified imputation model that includes all analysis variables and relevant auxiliary variables (e.g., baseline weight, age, breed, study site) to improve accuracy.
• For data missing not at random (MNAR), where missingness depends on the unobserved value (e.g., death due to treatment toxicity), routine imputation is insufficient, and advanced methods like pattern-mixture or selection models, often requiring specialist statistical input, are necessary.
• Transparent reporting of missing data, including numbers, reasons for exclusion, and the analytical methods employed, is mandated by guidelines such as ARRIVE 2.0 and is crucial for assessing the robustness of study conclusions.
• Proactive planning at the study design stage, including protocols for managing attrition and sample size calculations that account for anticipated losses, is the most effective strategy to mitigate the impact of missing data on research validity.

Missing data are an unavoidable feature of veterinary clinical research. Animals die, owners withdraw consent, samples hemolyse, and recording instruments fail. How these absent observations are managed determines whether a study's conclusions remain valid or become quietly distorted. This article provides a practical framework for veterinary researchers confronting incomplete datasets, from recognizing the mechanism that generated the missingness to selecting an appropriate analytical strategy. The focus is on clinical studies across species, including companion animal trials, livestock production research, and preclinical models. The methods described are those a researcher can implement with standard statistical software, with emphasis on complete-case analysis, weighting, single imputation, and multiple imputation. Advanced Bayesian approaches are outside the scope of this article.

The central question addressed is deceptively simple: when observations are absent, what can the remaining data still tell you? The answer depends on why the data went missing, how much is missing, and what assumptions the chosen analytical method makes. A researcher who ignores this question risks reporting treatment effects that are artefacts of attrition instead of biology. A researcher who addresses it explicitly can often salvage valid inference from datasets that initially appear compromised.

At a Glance

Missingness Mechanisms and Their Implications

The statistical literature classifies missing data by the relationship between the probability of missingness and the values of the variables in the study. This classification, formalised by Rubin, underpins every practical decision that follows. Three mechanisms are distinguished.

Missing completely at random (MCAR) occurs when the probability of a missing observation is independent of both observed and unobserved values. A laboratory sample lost in transit, a blood tube that breaks in the centrifuge, or a data entry error that corrupts a single cell are examples. Under MCAR, the observed data are a random subsample of the full dataset, and analyzes that discard incomplete records remain unbiased, though they lose precision.

Missing at random (MAR) occurs when the probability of missingness depends on observed variables but not on the unobserved value itself, once the observed variables are accounted for. An animal that becomes too fractious to weigh may be more likely to have missing weight data, but if the probability of fractiousness is fully explained by its previous weights, treatment group, or other recorded covariates, the missingness is MAR. This mechanism is more common than MCAR in clinical research and is the assumption under which multiple imputation and likelihood-based methods produce valid results.

Missing not at random (MNAR) occurs when the probability of missingness depends on the unobserved value itself. An animal that dies because the treatment caused a fatal arrhythmia contributes a missing outcome that is directly informative about its prognosis. Similarly, an owner who withdraws a pet from a pain trial because the animal is deteriorating creates missingness that is MNAR. No amount of observed covariate information can correct for this mechanism, because the very fact of missingness carries information about the missing value.

The distinction between these mechanisms is also academic. It determines whether a simple analysis can be trusted, whether imputation is appropriate, and how the results should be interpreted. The challenge is that the mechanism cannot be definitively identified from the data alone. A researcher can test whether missingness is related to observed covariates, which provides evidence against MCAR, but MAR and MNAR are indistinguishable without external information about the missing values.

Consequences of Ignoring Missing Data

The most common approach to missing data in veterinary research is also the most dangerous: omit the incomplete records and analyze only the animals with complete data. This complete-case analysis is the default in many statistical packages and remains widespread in the veterinary literature. Its validity depends on the missingness mechanism. Under MCAR, complete-case analysis is unbiased but inefficient, discarding potentially useful information from partially observed animals. Under MAR, complete-case analysis can be biased if the probability of missingness depends on covariates that are also related to the outcome. Under MNAR, complete-case analysis is biased and the direction of the bias is unpredictable.

The loss of power from complete-case analysis is not trivial. In a study with 20% missingness on a single covariate, a complete-case analysis may discard far more than 20% of records, because animals missing any variable are excluded entirely. In longitudinal studies, where animals may miss individual time points, the cumulative loss can be severe. The expectation-maximization likelihood ratio test, developed for quantitative trait locus mapping in experimental crosses, illustrates the gains available from methods that use all observed data instead of discarding incomplete records. In simulations, this approach offered greater statistical power than analysis of variance on complete cases, with the advantage becoming more pronounced as missingness increased. The same principle applies across veterinary research settings: methods that use partial information outperform those that throw it away.

Missing Data in Preclinical and Translational Studies

Preclinical studies present distinctive missing-data challenges. Tumor growth inhibition studies, used to prioritize anticancer compounds, rely on serial tumor volume measurements that are lost when animals reach humane endpoints or die. The decision to euthanise an animal is often related to tumor burden, creating missingness that is directly informative about the outcome of interest. This is a textbook example of MNAR, and it cannot be handled by routine imputation methods. The experimental design considerations for preclinical tumor growth inhibition studies explicitly address this problem, noting that the choice of endpoint, the timing of measurements, and the handling of animals that die before study completion all affect the validity of the analysis.

Multicentre preclinical trials, which are increasingly advocated to improve the translational value of animal research, face additional complications. The collective statement from the Preclinical Ischemic Stroke Multicenter Trials group identifies attrition and mortality as central statistical considerations, alongside the need for standardized protocols across sites. When multiple laboratories contribute data, missingness patterns may differ by site, and the reasons for missingness may be entangled with site-specific practices. A site with more rigorous monitoring may detect humane endpoints earlier, creating a different missingness pattern than a site with less frequent assessment. These site effects must be modelled explicitly, or the analysis will conflate site differences with treatment effects.

Reporting Standards and Transparency

The reporting of missing data is as important as the statistical handling. The ARRIVE guidelines, which specify the minimum information required for transparent and reproducible animal research publications, require authors to report the number of animals in each group at each stage of the study, the number excluded, and the reasons for exclusion. The ARRIVE guidelines 2.0 also require that authors state the statistical methods used to handle missing data. This reporting burden is not bureaucratic. It allows readers to assess whether the chosen methods were appropriate and whether the conclusions are robust to alternative assumptions.

The broader reporting landscape, catalogued by the EQUATOR Network, includes study-type specific guidelines such as CONSORT for randomised trials and STROBE for observational studies. Veterinary researchers should consult the relevant guideline for their study design and follow its missing-data reporting requirements. The REFLECT statement, developed specifically for livestock trials, includes similar provisions. A study that reports missing data transparently can be evaluated fairly even if its missingness is substantial. A study that buries its attrition in a footnote cannot.

Planning for Missing Data at the Design Stage

The most effective strategy for handling missing data is to prevent it. Study designs that minimize the opportunity for missing observations, protocols that specify how attrition will be managed, and sample size calculations that account for anticipated losses all reduce the burden on the analysis stage. The MSD Veterinary Manual and the American Veterinary Medical Association practice resources both emphasize the importance of protocol development and standardization in clinical research, and the same principles apply to missing-data planning. A sample size calculation that assumes no attrition will produce an underpowered study when 15% of animals are lost, a common figure in veterinary field trials.

Prevention strategies vary by species and setting. In companion animal trials, owner communication and follow-up protocols reduce withdrawal. In livestock research, the survey of castration methods and associated livestock management practices demonstrates how missing data arise in production settings: of 1,972 invited veterinarians, only 189 responses were analysable after partially completed surveys and missing data were omitted. This response rate, below 10%, is typical of veterinary survey research and illustrates the importance of designing questionnaires that minimize item non-response and planning analyzes that can accommodate the missingness that remains.

The international standards for animal health surveillance and research, published by the World Organization for Animal Health, also have implications for missing-data planning. Studies that inform trade policy or disease control decisions must be able to defend their handling of incomplete data, because the consequences of biased estimates extend beyond the individual study to regulatory decisions affecting animal movements and market access.

Choosing a Handling Strategy: A Decision Framework

The choice of missing data strategy should follow a structured assessment of the missingness mechanism, the proportion of incomplete records, the analysis objective, and the available computational resources. No single method suits all veterinary datasets, and the decision should be made before the primary analysis is run, not after inspecting results.

Step 1: Characterize the Missingness Pattern

Begin by tabulating missingness per variable and per observation. Calculate the proportion of missing values for each endpoint, and examine whether missingness clusters in particular treatment groups, time points, or study sites. Visualize the pattern with a matrix plot or heat map. This descriptive step determines whether the data are missing completely at random (MCAR), missing at random (MAR), or missing not at random (MNAR), as defined in Part 1.

For longitudinal studies, distinguish between intermittent missingness, where an animal misses one visit but returns, and monotone dropout, where an animal is lost permanently. Monotone dropout is more common in chronic disease studies and in production animal trials where animals leave the cohort. The distinction matters because some imputation methods assume monotone patterns and become inefficient or biased when applied to intermittent gaps.

Step 2: Apply the Decision Criteria

The following decision tree summarizes the selection logic. The thresholds are practical guidelines, not universal rules, and should be interpreted with knowledge of the study context.

When missingness is below 5% and appears MCAR, complete-case analysis is usually acceptable, provided the remaining sample retains adequate statistical power. Between 5% and 20%, multiple imputation is the default for MAR data because it uses all observed information and produces valid standard errors when the imputation model is correctly specified. Above 20%, the results become increasingly dependent on the imputation model, and the study conclusions should be interpreted with corresponding caution.

For MNAR data, no imputation method recovers the missing values without strong assumptions. Pattern-mixture models, which stratify the sample by missingness pattern and model each stratum separately, are one option. Selection models, which jointly model the outcome and the probability of missingness, are another. Both require specialist statistical input and should be planned prospectively.

Step 3: Match the Method to the Analysis Type

The primary endpoint dictates the imputation approach. For continuous endpoints measured longitudinally, such as tumor volume in xenograft studies, linear mixed models with maximum likelihood estimation handle missing values under MAR without explicit imputation, because the likelihood uses all available time points. This approach is computationally efficient and avoids the need to generate imputed datasets. However, it assumes the missingness is ignorable, an assumption that should be stated explicitly in the methods section.

For categorical endpoints, such as clinical response or adverse event occurrence, logistic regression or generalized estimating equations are common. Multiple imputation is more straightforward for these models than for mixed models, because the imputation model can be specified at the variable level. For time-to-event endpoints, such as time to tumor recurrence or time to withdrawal from a feeding trial, Cox proportional hazards models accommodate right-censoring naturally, but missing covariate values still require imputation.

In experimental crosses and genetic studies, missing genotype data can be addressed with expectation-maximization algorithms. An expectation-maximization likelihood-ratio test has been shown to offer greater statistical power than analysis of variance on complete cases alone, with the advantage becoming more pronounced as missingness increases. This approach is relevant to veterinary genetic studies using F2 intercross designs or quantitative trait locus mapping.

Multiple Imputation in Practice

Multiple imputation creates several completed datasets, analyzes each one with the intended statistical model, and pools the results using Rubin's rules. The pooled estimates incorporate both within-imputation and between-imputation variance, so the final standard errors reflect the uncertainty introduced by missing values.

Specifying the Imputation Model

The imputation model should include all variables that appear in the analysis model, plus auxiliary variables that predict missingness or correlate with the incomplete variable. Auxiliary variables need not be part of the primary analysis, they exist to improve imputation accuracy. In veterinary studies, useful auxiliary variables include body weight at baseline, age, sex, breed, study site, and any prior measurements of the same endpoint.

For longitudinal data, impute at the variable level using all available time points as predictors. For clustered data, such as animals within litters or pens within farms, include the cluster identifier as a fixed or random effect in the imputation model. Ignoring clustering during imputation attenuates the between-cluster variance and can bias treatment effect estimates.

Number of Imputations and Diagnostics

The number of imputed datasets should be at least 20 for routine analyzes, and higher when the proportion of missing data is large or when the analysis model is complex. Modern computing makes 50 or 100 imputations feasible and reduces Monte Carlo error in the pooled estimates.

After imputation, compare the distribution of imputed values with observed values for each variable. Check that imputed values fall within plausible ranges, that categorical variables receive only valid categories, and that the between-imputation variance is not excessive. Convergence of the imputation algorithm should be assessed with trace plots. If the imputation model fails to converge or produces implausible values, revise the model specification before proceeding.

Complete-Case and Available-Case Analysis

Complete-case analysis restricts the sample to animals with no missing values across all analysis variables. It is simple, transparent, and valid under MCAR. Its limitations are loss of statistical power and potential bias under MAR or MNAR, because the complete cases may differ systematically from the full cohort.

Available-case analysis uses all observed values for each variable, so the sample size varies between analyzes. This approach is common in descriptive studies and in tables of baseline characteriztics. It avoids discarding data but produces inconsistent sample sizes across tables, which can confuse readers. Report the number of observations contributing to each estimate.

In veterinary surveys, complete-case analysis is frequently applied after excluding partially completed responses. A survey of castration practices among bovine veterinarians in the United States omitted partially completed surveys and missing data before analysis, leaving 189 responses from an initial invitation list of nearly 2,000 practitioners. This approach is defensible when the missingness is concentrated in non-essential items, but researchers should report how many responses were excluded and compare the included and excluded groups on available characteriztics.

Species, Setting, and Practical Constraints

The correct handling strategy depends on the study population and the practical realities of data collection. In production animal studies, animals may be removed from the cohort for reasons unrelated to the study endpoint, such as sale, culling, or movement between management groups. These losses are often MNAR if the reason for removal correlates with the outcome. For example, a calf removed because of poor weight gain is informative about the growth endpoint. In such cases, record the reason for removal and consider a pattern-mixture analysis that treats removed animals separately.

In companion animal clinical trials, owner compliance drives missingness. Owners may withdraw animals because of perceived lack of efficacy, adverse events, or financial constraints. Each reason has different implications for the missingness mechanism. Document the reason for every withdrawal and include it as a variable in the imputation model where possible.

In preclinical studies, attrition and mortality are common in ischemic stroke models and tumor growth inhibition studies. Multicentre preclinical trials have identified appropriate handling of attrition and mortality as a key statistical consideration, alongside adaptive sequential designs and covariate adjustment. When animals die before the final endpoint, the missingness is often MNAR because death is the outcome of interest. Analyze mortality as a competing risk instead of imputing the missing endpoint values.

Equipment and software availability also constrain method choice. Multiple imputation is implemented in all major statistical packages, but pattern-mixture models and selection models require specialist routines and statistical expertise. If the study team lacks this expertise, choose a simpler method and acknowledge its limitations in the manuscript.

Reporting the Handling Strategy

Reporting standards for animal research require that the number of animals in each analysis and any exclusions be stated explicitly. The ARRIVE guidelines specify that publications should report the exact number of animals used in each experimental group and the number excluded from analysis, with reasons. The EQUATOR Network maintains a library of reporting guidelines, including CONSORT for randomised trials and STROBE for observational studies, which contain specific items on missing data handling.

The methods section should state the missingness mechanism assumed, the proportion of missing data for each primary endpoint, the handling method used, and the software and parameters for imputation. The results section should report the number of animals contributing to each analysis and the results of any sensitivity analyzes. A sensitivity analysis that repeats the primary analysis under different missingness assumptions is the most informative way to demonstrate that conclusions are robust.

Recognized Complications and Failure Modes

The most common failure in missing data handling is applying an imputation method without first characterizing the missingness mechanism. A complete-case analysis applied to data missing not at random will produce biased estimates that no downstream correction can repair. The second most frequent error is treating imputed values as real observations in the final analysis, which leads to overprecision in confidence intervals and false-positive findings. Multiple imputation addresses this by propagating uncertainty across imputed datasets, but only when the imputation model is correctly specified and the number of imputations is adequate for the proportion of missing data.

A third failure mode is imputing the outcome variable using predictors that include the outcome itself or other post-treatment variables. This creates circularity and artificially strengthens associations. In longitudinal studies, a related error is carrying the last observation forward when the trajectory is nonlinear, such as tumor growth or weight gain in growing animals. The experimental design guidance for tumor growth inhibition studies specifically cautions against this practice because it systematically underestimates the divergence between treatment groups over time.

A fourth complication is the interaction between missing data and attrition due to mortality. In preclinical stroke trials and other studies with high event rates, animals that die before the final measurement are not missing at random, their absence is informative. The PRISM collective statement on multicenter preclinical trials recommends that mortality be analyzed as a competing outcome instead of treated as missing data, and that sensitivity analyzes compare results under different assumptions about the missingness mechanism.

Early detection of these problems requires routine diagnostic checks. Plot missingness against observed values of other variables, compare the distribution of observed outcomes between animals with and without complete follow-up, and examine whether missingness clusters by treatment group, site, or time point. A pattern in which missingness increases with disease severity or with time on study should trigger a re-evaluation of the chosen method.

Common Errors and Corrective Actions

Less experienced analysts often default to complete-case analysis because it is the default in many statistical software packages. The corrective action is to quantify the proportion of incomplete cases and compare the characteriztics of complete and incomplete observations. If more than 5% of cases are incomplete, complete-case analysis should be justified or replaced.

A second common error is using mean imputation, which reduces variance and distorts correlations. The corrective action is to use multiple imputation or, for simple sensitivity checks, available-case analysis. A third error is failing to include auxiliary variables in the imputation model. Variables that predict missingness but are not part of the primary analysis should be included in the imputation model to satisfy the missing-at-random assumption.

A fourth error is imputing baseline covariates that were measured before randomisation. These are almost always observed and should not be imputed unless the measurement itself failed. When baseline data are missing, the imputation model should include treatment assignment and all outcome variables measured at later time points.

Limitations of the Current Evidence

The veterinary literature contains few methodologic studies that directly compare missing data strategies across species and clinical settings. Most guidance is extrapolated from human clinical trials or from statistical theory. The genotyping literature in crop plants demonstrates that imputation methods developed for one data type do not transfer automatically to another, and the same caution applies across veterinary species. What works for a randomised controlled trial in dogs with a fixed follow-up period may not suit an observational study in cattle with variable herd management.

Expert opinion differs on the threshold at which missing data become unmanageable. Some statisticians recommend that if more than 40% of cases have missing outcome data, the study should be considered exploratory regardless of the imputation method. Others argue that the proportion matters less than the mechanism. The expectation-maximization likelihood ratio test literature shows that methods using all observed data outperform complete-case analysis even when missingness is substantial, but the advantage depends on the model being correctly specified.

When to Escalate

Referral to a consulting biostatistician is warranted when the missingness mechanism is uncertain, when the proportion of missing data exceeds 20%, or when the primary analysis changes materially under different missingness assumptions. Laboratory involvement is appropriate when missingness results from assay failure, sample degradation, or detection limits, because the laboratory may be able to re-run samples or provide information about the failure mechanism.

Regulatory reporting obligations arise in specific circumstances. For studies conducted under Good Clinical Practice, the sponsor must document all protocol deviations including missing data and their handling. For pharmacovigilance reports, missing safety data must be reported as such instead of imputed. The ARRIVE guidelines require that the number of animals excluded from analysis and the reasons for exclusion be reported, and the EQUATOR Network reporting standards provide the corresponding checklists for clinical trials and observational studies. When missing data affect the interpretation of safety outcomes in a regulated product, the responsible regulatory authority should be informed through the appropriate reporting pathway.

Frequently Asked Questions

How Many Imputations Do I Need When Computational Resources Are Limited?

The traditional advice of 5 to 10 imputations may suffice for simple analyzes, but modern guidance favours imputing at least 20 to 40 data sets, especially when the fraction of missing information is high. If your computing environment cannot support that volume, consider whether a complete-case analysis with clear reporting of its limitations is more defensible than an underpowered imputation run. Alternatively, reduce the complexity of the imputation model by removing weakly predictive auxiliary variables. Document the number of imputations and the reason for that choice in the methods section, since reviewers will assess whether the computational shortcut materially affected precision. The ARRIVE guidelines require transparent reporting of all analytical decisions, including imputation parameters.

What Should I Do When the Planned Analysis Software Is Unavailable?

Use the software you have, but verify that its missing-data routines implement the method you intend. Many general-purpose statistical packages include multiple imputation procedures, but default settings may assume multivariate normality or use listwise deletion for auxiliary variables. Before committing, run a small simulation or a sensitivity analysis comparing your software's output against a reference implementation if one is available. If your package cannot perform multiple imputation, a well-conducted complete-case analysis with a documented missingness assessment may be preferable to an incorrectly specified imputation. The EQUATOR Network reporting guidelines can help you identify which reporting standard applies to your study design and what analytical details must be disclosed.

How Do I Handle Missing Data in a Multicentre or Multispecies Trial?

Missingness mechanisms often differ across sites or species, so pool the data only after testing whether the missingness pattern and outcome distributions are comparable. Fit the imputation model separately by site or species when sample sizes permit, or include site and species as fixed effects in the imputation model. Attrition due to mortality or euthanasia is a particular concern in animal studies and may represent informative missingness that requires joint modeling of the outcome and the dropout process. Multicentre preclinical trials have identified attrition and mortality as central challenges for data interpretation, and their recommendations emphasize standardized protocols for recording the reasons for missing observations across all participating sites.

What Are the Minimum Records I Must Keep About Missing Data?

Maintain a missing-data log that records, for every missing observation, the animal identifier, the scheduled measurement time point, the reason for missingness, and the date the gap was first noticed. Distinguish between planned missingness, such as samples not collected by protocol design, and unplanned missingness, such as equipment failure or animal death. Keep the original data files untouched and create a separate analytical data set for any imputation work. This audit trail allows reviewers to verify that the missingness assumptions stated in your manuscript match what actually occurred in the study. The ARRIVE guidelines specify that publications must report the number of animals excluded from analysis and the reasons for exclusion.

How Should I Explain Missing Data Handling to a Client or Supervisor?

Frame the explanation around the risk of a wrong conclusion, not the statistical mechanics. State that animals with missing measurements may differ systematically from those with complete data, and that ignoring those gaps can make a treatment appear more effective or safer than it is. Explain that the analysis plan accounts for this by either restricting the analysis to animals with complete data, which is transparent but may lose power, or by using statistical methods that estimate the missing values from the observed data, which preserves sample size but relies on assumptions. Emphasize that the choice was made before the analysis and is reported in the methods so that readers can judge its appropriateness.

Does the Handling Strategy Differ Between Production Animals and Companion Animals?

Yes, in practical ways. Production animal studies often have large sample sizes, which makes complete-case analysis less costly in terms of statistical power, but they also face higher rates of loss due to culling, transport, or management decisions outside the research team's control. Companion animal studies typically have smaller samples, so imputation may be more attractive, but the reasons for missingness, such as owner non-compliance with follow-up visits, may be more strongly associated with the outcome. The WOAH terrestrial animal health standards provide guidance on data quality expectations for studies that inform international animal health policy, and these standards apply regardless of species.

Related Clinical & Scientific Guides

• Conducting Systematic Reviews of Veterinary Diagnostic Test Accuracy
• Bias in Veterinary Research: Types, Sources, and Mitigation
• Cluster Randomized Trials in Veterinary Research: Design and Analysis

References and Further Reading

• Efficient genome-wide genotyping strategies and data integration in crop plants.. 2018.
• Experimental design considerations and statistical analyzes in preclinical tumor growth inhibition studies.. 2025.
• Preclinical Ischemic Stroke Multicenter Trials (PRISM) Collective Statement: Opportunities, Challenges, and Recommendations for a New Era.. 2026.
• A survey of castration methods and associated livestock management practices performed by bovine veterinarians in the United States.. 2010.
• An expectation-maximization-likelihood-ratio test for handling missing data: application in experimental crosses.. 2005.
• Ensemble Machine Learning Approaches for Bathymetry Estimation in Multi-Spectral Images.. 2025.
• ARRIVE Guidelines 2.0 for Reporting Animal Research. PLOS Biology, 2020.
• EQUATOR Network Reporting Guidelines. EQUATOR Network.
• MSD Veterinary Manual, Professional Edition. MSD Veterinary Manual.

Related Articles

• Blinding in Veterinary Clinical Research: Methods and Challenges
• Implementing Adaptive Trial Designs in Veterinary Clinical Research
• Data Management in Veterinary Research: Collection, Storage, and Sharing
• Pragmatic vs Explanatory Trials in Veterinary Clinical Research
• Preparing Data Management Plans for Veterinary 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.