Sample Size Calculation for Veterinary Studies: Methods and Software

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

Key Takeaways

• Sample size calculation is critical for ensuring veterinary studies yield valid conclusions, are ethical, and are publishable, resting on the interplay of effect size, variance, significance level (alpha, conventionally 0.05), and statistical power (conventionally 0.80 or higher).
• Biological variance in veterinary populations (e.g., outbred animals, production species, wildlife) is often higher than in inbred laboratory cohorts, necessitating larger sample sizes and careful estimation using species- and outcome-specific data, or pilot studies.
• Study design elements such as the unit of analysis, clustering (e.g., animals housed in pens or herds), and repeated measures significantly impact the required sample size, demanding explicit accounting for intracluster correlation coefficients to avoid inflated Type I error rates.
• Overestimation of effect size is a common pitfall; effect sizes should be biologically meaningful and derived from systematic reviews or meta-analyses, not optimistic hopes, with sensitivity analyses recommended when evidence is limited.
• Transparent reporting of sample size determination, including primary outcome, effect size source, variance estimate source, alpha, power, attrition allowance, and software used, is mandated by guidelines like ARRIVE 2.0 and is essential for reproducibility and scientific rigor.
• Software options range from general statistical packages (R, SAS) to dedicated tools (G*Power, nQuery), with specialized approaches required for complex designs like microbiome studies or freedom-from-disease surveys that account for compositional data or imperfect test sensitivity/specificity.

Sample size calculation is the process of determining the minimum number of animals or observations required to detect a biologically meaningful effect with acceptable statistical confidence. For veterinary researchers, this calculation is not a formality but a determinant of whether a study can yield valid conclusions, whether it is ethical, and whether it can be published. This article provides a practical framework for performing sample size and power calculations across common veterinary study designs, including experimental trials, observational studies, diagnostic accuracy studies, and freedom-from-disease surveys. It is written for veterinary researchers, graduate students, and clinicians engaged in hypothesis-driven research who need to select an appropriate method, identify the required input parameters, and implement the calculation using available software.

The article covers the statistical logic underlying power analysis, the specific considerations that distinguish veterinary studies from human clinical research, and the practical steps for calculating sample size in each major design category. It also addresses common pitfalls, including failure to account for clustering, overestimation of effect size, and neglect of biological variance in outbred or field populations. Where the evidence base is limited or contested, this is stated explicitly.

At a Glance

The Statistical Basis of Sample Size Determination

Sample size calculation rests on the relationship between four quantities: the effect size, the variance of the outcome, the significance level, and the statistical power. Power is the probability that a study will detect an effect of a given magnitude when that effect truly exists. A study with insufficient power may fail to reject a false null hypothesis, producing a false negative result that is scientifically uninformative and ethically problematic because animals were used without a reasonable chance of answering the question.

The effect size is the magnitude of the difference or association the study aims to detect. It must be specified a priori and should represent the smallest effect that would be clinically or biologically meaningful, not the largest effect the investigator hopes to see. For continuous outcomes, the standardized effect size is often expressed as Cohen's d, the difference between group means divided by the pooled standard deviation. For binary outcomes, the effect size is expressed as a difference in proportions or as an odds ratio. For survival outcomes, the hazard ratio is used.

Variance estimation is frequently the weakest link in veterinary sample size planning. Inbred laboratory rodents housed under controlled conditions show relatively low biological variance, and sample sizes derived from such populations may be inadequate when applied to outbred animals, production species, or free-ranging populations. A systematic review of stem cell therapy in experimental stroke found that fewer than half of the included studies reported randomization or blinded outcome assessment and only three of 117 studies reported a sample size calculation, indicating that underpowered studies with avoidable bias were common in that literature. Similarly, a meta-analysis of the transverse aortic constriction model in mice and rats found that only 2% of studies reported a sample size calculation despite substantial heterogeneity in functional outcome measures, even when experimental design was otherwise homogeneous.

Power Analysis in the Context of Animal Research Reporting

The ARRIVE guidelines for reporting animal research specify that manuscripts must state the sample size, how it was determined, and the statistical methods used to calculate it. These guidelines apply across species and study types and are endorsed by major veterinary and biomedical journals. The EQUATOR Network reporting guidelines provide a searchable library of design-specific reporting standards, including CONSORT for randomized trials, PRISMA for systematic reviews, STROBE for observational studies, and REFLECT for livestock trials, all of which require transparent reporting of sample size determination.

A sample size calculation should be reported with enough detail that a reader can reproduce it. This includes the primary outcome, the expected effect size and its source, the variance estimate and its source, the significance level, the power, the number of groups or arms, the attrition allowance, and the software or formula used. When a study is underpowered because of practical constraints, this should be stated explicitly with an explanation of the implications for interpretation.

Sources of Variance in Veterinary Study Populations

Veterinary studies frequently involve greater biological variance than human clinical trials or laboratory studies using inbred strains. Species differences, breed differences, age, sex, body condition, management systems, and environmental conditions all contribute to outcome variability. In gene expression studies using RNA sequencing, high biological variance is typical of field-based studies and means that larger sample sizes are often needed to achieve the same statistical power as studies based on cell lines or inbred animal models. The same principle applies to many veterinary outcomes, particularly those measured in production animals or wildlife.

Strain and sex differences can be substantial. In a multilaboratory study of distal middle cerebral artery occlusion in mice, important strain differences in infarct volume were reported, and power analysis confirmed that some functional tests required larger sample sizes than others to detect differences reliably. Grip strength and latency to move were useful at long-term time points, but with important differences between strains. These findings illustrate that variance estimates from one strain or population cannot be assumed to transfer to another.

The Unit of Analysis and Clustering

The unit of analysis is the entity to which the treatment is assigned and on which the outcome is measured. In veterinary research, the unit of analysis is not always the individual animal. Animals may be housed in pens, litters, or herds, and treatments may be applied at the group level. When animals within a group are correlated, the effective sample size is smaller than the number of individual animals, and the analysis must account for clustering to avoid inflated type I error rates.

For group-housed animals, the sample size calculation should be based on the number of independent units, not the total number of animals, unless the analysis explicitly models the clustering. Intracluster correlation coefficients are rarely reported in veterinary literature, which complicates planning. A conservative approach is to assume a nonzero intracluster correlation and inflate the sample size accordingly, or to use variance estimates from studies that used similar housing and management conditions.

Effect Size Selection and the Risk of Overestimation

The most common error in sample size calculation is the use of an unrealistically large effect size, which produces a sample size too small to detect smaller but still meaningful effects. Effect sizes should be derived from systematic reviews or meta-analyzes when available, instead of from single studies, which tend to report larger effects than subsequent replication. Publication bias inflates effect estimates in many fields. A systematic review and meta-analysis of stem cell therapy in experimental stroke found evidence of significant publication bias, with nonrandomized studies giving significantly higher estimates of improvement in structural outcome. Meta-analyzes of the transverse aortic constriction model did not detect specific sources of heterogeneity despite substantial variability in outcomes, underscoring the difficulty of predicting effect sizes from aggregated literature.

When no reliable effect size estimate exists, investigators should consider a pilot study, but pilot studies are typically too small to provide stable variance estimates. An alternative is to calculate sample size across a range of plausible effect sizes and variances, then select a sample size that provides adequate power across the plausible range. This sensitivity analysis is more informative than a single calculation and should be reported when the evidence base is weak.

Software Options for Sample Size Calculation

Several software packages are available for sample size and power calculation. General statistical packages such as R, SAS, Stata, and SPSS include power analysis procedures for common designs. Dedicated sample size software, including nQuery, PASS, and G*Power, provides user-friendly interfaces and coverage of a wide range of designs. G*Power is freely available and supports common designs including t tests, ANOVA, regression, and chi-square tests. R offers extensive power analysis packages, including pwr for basic designs and more specialized packages for mixed models, survival analysis, and diagnostic accuracy studies.

For microbiome studies, which involve high-dimensional data with unique distributional features, standard sample size methods do not apply directly. Statistical approaches must account for the compositionality, sparsity, and overdispersion of microbiome data, and specialized R scripts are available for calculating sample sizes when microbiome features are the outcome, the exposure, or the mediator. These methods are relevant to veterinary studies of the gut, skin, and respiratory microbiomes in companion and production animals.

For freedom-from-disease surveys, the FreeCalc program implements a probability formula that accounts for imperfect test sensitivity and specificity and finite population size. This approach is directly relevant to veterinary surveillance and certification programs, where the goal is to demonstrate freedom from disease with a specified level of confidence. The WOAH terrestrial animal health standards provide the international framework for such surveys, and the MSD Veterinary Manual offers species-specific guidance on diagnostic test performance that informs these calculations.

Worked Example: Two-Group Comparison of Infarct Volume in a Murine Stroke Model

Consider a common veterinary research scenario: a preclinical efficacy study comparing infarct volume between treated and control mice after distal middle cerebral artery occlusion. The distal occlusion model is reproducible with low mortality, making it suitable for mid- and long-term functional studies, but strain differences in infarct volume are substantial and directly affect the variance you must enter into the calculation Rosell et al., 2013.

Assume a two-sided independent t test, alpha of 0.05, and power of 0.80. The standardized effect size, Cohen's d, is the difference between group means divided by the pooled standard deviation. For a normally distributed outcome with equal group sizes, the required sample size per group is:

n = 2 × [(z₁₋α/₂ + z₁₋β)²] / d²

where z₁₋α/₂ = 1.96 and z₁₋β = 0.84 for the specified alpha and power. Substituting these values gives:

n = 2 × [(1.96 + 0.84)²] / d² = 2 × 7.84 / d² = 15.68 / d²

For d = 0.8, n = 24.5, so 25 animals per group. For d = 0.5, n = 62.7, so 63 per group. For d = 0.2, n = 392 per group. These figures assume equal variance and no clustering. If you anticipate 10% attrition from perioperative mortality or exclusion for poor occlusion, divide n by 0.90. The distal MCAO model has low mortality, but strain-specific differences in infarct volume mean that variance estimates from the literature should be adjusted upward when your colony differs from the published source Rosell et al., 2013.

Effect Size Benchmarks for Common Veterinary Outcomes

The table below provides reference values for interpreting standardized effect sizes in veterinary research contexts. These are general benchmarks, not field-specific thresholds, and should be calibrated against the published literature for your particular model and outcome.

The medium and large benchmarks are the most defensible starting points for preclinical efficacy studies. Small effects require sample sizes that are rarely feasible in animal research and are better detected through continuous outcomes with repeated measures or through meta-analytic pooling.

Adjusting for Multiple Comparisons and Correlated Outcomes

Most veterinary studies measure more than one endpoint. If you plan to test three behavioral outcomes, each at alpha 0.05, the family-wise error rate rises to approximately 14%. Apply a Bonferroni correction by dividing alpha by the number of comparisons, or use a false discovery rate approach for exploratory endpoints. The correction increases required sample size. For three comparisons at corrected alpha of 0.017, the z term changes from 1.96 to 2.39, and the multiplier in the formula becomes 2 × (2.39 + 0.84)² = 20.9, roughly one-third larger than the uncorrected value.

Correlated outcomes complicate the calculation further. Grip strength and latency to move in the distal MCAO model both detect long-term deficits, but they are not independent measures Rosell et al., 2013. If you analyze them separately with correction, you overestimate the required sample size. A multivariate analysis of variance or a linear mixed model with outcome as a repeated factor accounts for correlation and preserves power, but requires simulation-based sample size estimation instead of a closed-form formula.

Software Selection and Implementation

Free software covers nearly all veterinary sample size requirements. G*Power handles t tests, ANOVA, regression, and chi-squared tests with a graphical interface and is suitable for most two-group and multi-group designs. R packages including pwr, powerSurvEpi, and clusterPower extend this to survival analysis, clustered designs, and simulation-based approaches. For freedom-from-disease surveys, the FreeCalc program implements exact probability formulae that account for imperfect test sensitivity and specificity and finite population size, which is essential when the assumption of a perfect test is invalid Cameron and Baldock, 1998.

The choice of software depends on the study design. G*Power is adequate for simple comparisons. R is necessary when you need to simulate data, model clustering, or incorporate prior information through Bayesian methods. Microbiome studies require specialised approaches because the data are compositional, sparse, and high-dimensional, standard sample size formulae do not apply, and dedicated R scripts are needed to handle hypotheses where the microbiome is the outcome, the exposure, or the mediator Ferdous et al., 2022.

Species and Production System Considerations

The correct sample size calculation changes with species and production context. In companion animal clinical trials, owners may withdraw animals, and the variance of clinical endpoints is often larger than in controlled laboratory settings. In food animal studies, clustering by herd or pen is the norm, and the intracluster correlation coefficient must be estimated from prior work or pilot data. In wildlife studies, capture probability and seasonal availability constrain feasible sample sizes regardless of the statistical ideal.

Laboratory rodent studies benefit from inbred strains and controlled environments, which reduce variance and permit smaller samples. However, the homogeneity of inbred animals can overstate the generalizability of findings. The transverse aortic constriction model shows substantial heterogeneity in functional outcomes despite the preferential use of male C57Bl/6 mice, and only 2% of studies in that literature reported a sample size calculation Bosch et al., 2021. This pattern is not unique to cardiac research. A systematic review of stem cell therapy in experimental stroke found that fewer than half of studies reported randomisation or blinded outcome assessment and only three of 117 publications reported a sample size calculation Lees et al., 2012.

Reporting the Calculation in the Methods Section

The methods section must state the primary outcome, the expected effect size and its source, the estimated variance, the alpha level, the desired power, the software used, and the final sample size after adjustment for attrition. The ARRIVE guidelines 2.0 specify that the sample size calculation should be reported with sufficient detail to allow replication, including the statistical test used and the assumptions entered into the calculation NC3Rs, 2020. If no formal calculation was performed, this must be stated explicitly with justification.

A defensible methods entry reads: "Sample size was calculated in G*Power 3.1 for a two-sided t test with alpha 0.05 and power 0.80. Based on published infarct volume data in C57Bl/6 mice after distal MCAO, we assumed a standardized effect size of 0.8 and allocated 25 animals per group. To account for an anticipated 10% attrition, 28 animals per group were enrolled."

Sensitivity Analysis and Pilot Data

A single sample size calculation based on one effect size estimate is fragile. Run a sensitivity analysis across a plausible range of effect sizes and variance estimates. If the required sample size changes from 20 to 80 per group across that range, the study design is not robust and you should either refine the outcome, reduce variance through stricter inclusion criteria, or reconsider the feasibility of the study.

Pilot data can inform variance estimates, but pilot studies in veterinary research are often too small to provide stable estimates. A pilot of 5 animals per group yields a confidence interval for the standard deviation that is extremely wide. Use pilot data to check feasibility and refine measurement protocols, but base the formal calculation on published values or conservative assumptions. Where the evidence base is limited, as in many wildlife and exotic species, acknowledge the uncertainty and present the calculation as a range instead of a single number.

Documentation and Audit Trail

Record the date of the calculation, the software version, all input parameters, and the source of each assumption. This documentation should be retained with the study protocol and made available to reviewers on request. Journals increasingly require this information at submission, and funding bodies may audit it during grant review. The EQUATOR Network maintains a library of reporting guidelines, including REFLECT for livestock studies and CONSORT for randomised clinical trials, which specify the reporting items expected for sample size and power EQUATOR Network.

Recognized Complications and Early Detection

Sample size calculations fail in predictable ways. The most common complication is silent underestimation of variance, which produces an underpowered study that appears sound on paper. Detect this early by comparing the variance estimate used in the calculation against the variance observed in the first 10 to 20 percent of enrolled animals. If the observed standard deviation exceeds the planned value by more than 20 percent, the study will not reach its target power without an increase in sample size or a reduction in outcome variability.

A second complication is attrition that exceeds the planned dropout rate. Longitudinal veterinary studies, particularly those involving surgical models or production animals, routinely lose subjects to euthanasia, withdrawal, or protocol violation. The ARRIVE guidelines 2.0 for reporting animal research require that attrition be reported transparently, but the calculation itself must anticipate it. Monitor cumulative attrition against the planned rate at each scheduled interim review. If attrition exceeds the allowance, the effective power at study completion will fall below the stated level.

A third complication is effect size drift. The effect size used in the original calculation may become untenable as preliminary data accumulate. This occurs when pilot data show a smaller treatment effect than the literature suggested, or when the outcome measure proves noisier than expected. The systematic review of stem cell therapy in experimental stroke found that fewer than three percent of included studies reported a sample size calculation, and those that did often used effect sizes derived from nonrandomised studies that overestimated benefit. Re-estimate the required sample size whenever the observed effect size falls outside the confidence interval of the planned effect.

Common Errors and Corrective Action

Less experienced researchers frequently confuse the unit of analysis with the unit of assignment. In a pen-based feeding trial, the pen is the experimental unit even when individual animals are weighed. Treating individual animals as independent replicates inflates power and risks a false-positive conclusion. The corrective action is to compute the intracluster correlation coefficient from pilot data and apply a design effect to the sample size.

A second error is using a one-tailed test without justification. Directional hypotheses are rarely defensible in veterinary research because treatments can produce unexpected adverse effects. If the protocol specifies a one-tailed test, the rationale must be explicit and the possibility of a harmful effect must be addressed. The safer default is a two-tailed test with a correspondingly larger sample size.

A third error is failing to account for multiple primary outcomes. Studies that measure infarct volume, grip strength, and latency to move, as in the distal middle cerebral artery occlusion model review, need a correction for multiplicity or a clear statement of which outcome is primary. Without correction, the probability of at least one false positive rises with each additional outcome. The corrective action is to designate one primary outcome for the power calculation and treat all others as secondary.

A fourth error is using a sample size calculator without understanding the underlying assumptions. Software packages differ in their default handling of unequal group sizes, variance estimation, and dropout. The corrective action is to reproduce the calculation manually or in a second software package and confirm that the results agree.

Limitations of the Current Evidence

The evidence base for sample size calculation in veterinary research is uneven. Most methodological work derives from human clinical trials or from laboratory animal models, and the transfer of these methods to veterinary populations is imperfect. Production animals differ from laboratory rodents in genetic heterogeneity, environmental exposure, and management practices, all of which increase variance. The RNA-seq review in ecology and evolution notes that field-based studies typically require larger sample sizes than laboratory studies because of high biological variance, a principle that applies equally to veterinary field research.

Expert opinion differs on the acceptable minimum power. Many veterinary journals expect 80 percent power, but some funding bodies and ethics committees now require 90 percent for studies involving protected species or substantial animal numbers. The transverse aortic constriction model meta-analysis found that only two percent of studies reported a sample size calculation, which suggests that current practice lags behind methodological guidance.

For microbiome studies, the situation is more complex. The microbiome power and sample size review describes methods that account for compositional data, sparsity, and multiple testing, but these methods require specialised software and assumptions that many veterinary researchers are not equipped to evaluate. Where the evidence base is contested, state the uncertainty in the protocol and justify the chosen approach.

Referral, Consultation, and Regulatory Reporting

Consult a statistician before finalising the protocol when the study involves clustered data, longitudinal measurements, survival outcomes, or multiple co-primary endpoints. These designs require methods that exceed the scope of standard software and general guidance. Laboratory involvement is warranted when the outcome measure requires specialised assays, such as microbiome sequencing or histopathological scoring, because the variance of the assay contributes directly to the sample size.

Regulatory reporting applies when the study is conducted under an animal ethics approval that specifies a maximum animal number. If the sample size calculation indicates that more animals are needed than the approved number, the protocol must be amended before enrollment continues. For disease freedom surveys, the WOAH terrestrial animal health standards specify the design prevalence and confidence level that the sample size must satisfy. The formula for freedom from disease surveys accounts for imperfect test sensitivity and finite population size, and the FreeCalc program implements it. If the survey results do not meet the required confidence level, the finding must be reported as inconclusive instead of as evidence of freedom.

Troubleshooting Table

Frequently Asked Questions

How do I justify a smaller sample size when funding or animal welfare limits are strict?

Justify any reduction explicitly in the methods section. Report the target power, the effect size you originally planned to detect, and the power your reduced sample actually provides. If you lower the sample size, you must either accept a larger minimum detectable effect or a higher false-negative risk. Pilot data can help show that the effect size is large enough to justify the reduction. For survival or disease-freedom surveys, the formula-based approach for finite populations with imperfect tests can minimize required numbers while preserving the confidence level. If the reduced design cannot meet your primary objective, state that limitation plainly and consider whether the study should proceed at all.

What sample size approach applies when I cannot use the recommended software or statistical package?

The calculation itself does not depend on a specific program. You can perform the arithmetic manually using standard power formulas for simple two-group comparisons, or use a spreadsheet with built-in statistical functions. For more complex designs, several free online calculators implement the same underlying methods. The EQUATOR Network reporting guidelines list the reporting standards that apply regardless of software, so your methods section must describe the formula, the inputs, and the assumptions even if you used a spreadsheet. If you lack access to specialised software, keep the design simple, use a conservative variance estimate, and have a statistician review the arithmetic before you finalise the protocol.

How does the required sample size change when I move from an inbred rodent model to outbred animals or client-owned pets?

The change is driven by variance, not species. Outbred animals and client-owned populations carry substantially higher biological variance than inbred laboratory strains. Field-based gene expression studies, for example, typically need larger samples than cell-line or inbred-animal work to achieve the same power. Strain differences also affect the outcome itself, as demonstrated by the variation in infarct volume across mouse strains in stroke models. When you move to a more variable population, re-estimate the standard deviation from published data or a pilot study, then recalculate. Do not carry over the sample size from the inbred study. If published variance data are absent, plan a pilot phase and use the upper confidence bound of the variance estimate for the main calculation.

What records should I keep to show that the sample size calculation was performed correctly?

Keep the original protocol with the dated calculation, the software or formula version, all input values, and the name of the person who performed the calculation. Save the pilot data or the literature source used for the variance and effect size estimates. Record any changes made after a blinded interim review or an ethics committee request, with the date and reason for each change. Journals increasingly require this documentation under ARRIVE 2.0 reporting standards, and funders may audit it. If you used a random number generator for allocation, store the seed and the generation method. This audit trail protects you if a reviewer questions the calculation and allows a future meta-analysis to assess the quality of your design.

How do I explain the sample size to an ethics committee or animal welfare body that wants fewer animals?

Frame the calculation as a welfare tool, not an administrative hurdle. An underpowered study wastes every animal it uses because the results cannot support a reliable conclusion. The systematic review of stem cell therapy in stroke models found that fewer than three percent of studies reported a sample size calculation, and this weakness undermines the translational value of the whole field. Present the power curve showing how the false-negative rate rises as the sample shrinks. Offer a sequential design if the committee is concerned about overuse, where you stop early only if the effect is clearly present or clearly absent. This preserves the scientific objective while allowing the smallest number consistent with a defensible answer.

What should I do when the pilot data show much higher variance than the published estimates I used?

Do not proceed with the original calculation. Recalculate using the pilot variance, and if the new sample size is unaffordable, you have three options. First, refine the measurement protocol to reduce technical variance, for example by standardizing the time of day for sampling or using a single blinded assessor. Second, consider a paired or repeated-measures design if each animal can serve as its own control, which removes between-animal variance. Third, revisit the effect size you consider clinically meaningful. The transverse aortic constriction model meta-analysis showed substantial heterogeneity in functional outcomes even under standardized conditions, so your pilot variance is likely more realistic than published values. Document the revised calculation and the reason for the change in the protocol before any further animals are enrolled.

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

• Distal occlusion of the middle cerebral artery in mice: are we ready to assess long-term functional outcome?. 2013.
• Stem cell-based therapy for experimental stroke: a systematic review and meta-analysis.. 2012.
• The transverse aortic constriction heart failure animal model: a systematic review and meta-analysis.. 2021.
• The rise to power of the microbiome: power and sample size calculation for microbiome studies.. 2022.
• A new probability formula for surveys to substantiate freedom from disease.. 1998.
• The power and promise of RNA-seq in ecology and evolution.. 2016.
• 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

• Using Mixed Methods in Veterinary Research
• Conducting Pharmacovigilance Studies in Veterinary Medicine
• How to Write a Research Protocol for Veterinary Studies
• Blinding in Veterinary Clinical Research: Methods and Challenges
• Designing Dose-Response Studies in Veterinary Pharmacology

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.