Sample Size Calculation for Veterinary Epidemiological Studies

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

Key Takeaways

• Veterinary epidemiological studies must account for the hierarchical structure of animal populations (e.g., herds, flocks, litters), where individuals within a group share common exposures and genetics, leading to non-independent outcomes. This necessitates the use of a "design effect" to inflate sample sizes calculated for independent sampling, preventing underpowered studies.
• Sample size calculations for prevalence surveys are driven by the expected prevalence and desired precision (half-width of the confidence interval), while comparative studies depend on the significance level (α), power (1-β), and the minimum clinically meaningful effect size. For instance, estimating a rare disease (e.g., 1% prevalence) with high precision requires a substantially larger sample than initially apparent due to the variance of proportions.
• The intracluster correlation coefficient (ρ) is a critical parameter for calculating the design effect, quantifying within-group similarity. Reliable ρ estimates are often scarce, necessitating sensitivity analyses across a plausible range (e.g., 0.01 to 0.10) to assess the impact on the required sample size and determine study feasibility.
• Common errors include ignoring clustering, using inappropriate expected prevalence from different populations, confusing sampling units with analysis units (e.g., animals vs. herds), and failing to adjust for losses to follow-up (e.g., culling, death). For example, a study designed for independent sampling of 100 animals from 10 herds might require over 300 animals if the ICC is 0.10 and cluster size is 20.
• When resources are limited, recalculate the study objectives around a reduced scope, such as a single primary outcome or a larger minimum effect size, rather than accepting a smaller sample size without justification. Documenting all input parameters, their sources, and conducting sensitivity analyses is crucial for grant applications and peer review.
• Consultation with a veterinary epidemiologist or biostatistician is recommended for complex designs involving multiple levels of clustering, rare outcomes, or when planning logistic regression with numerous predictors, as standard formulas may be insufficient.

Sample size calculation is a mandatory component of study planning in veterinary epidemiology. Funding agencies and ethical review boards require justification of the number of animals enrolled, and the assumptions underlying that number must be stated explicitly. This article provides practical guidance for veterinary researchers calculating sample sizes for prevalence surveys, comparative studies, and observational designs common in animal populations. It explains the statistical logic, the parameters that must be specified, and the adjustments required when animals are clustered within herds, flocks, or litters.

The central problem is balancing two failure modes. A study with too few subjects risks failing to detect a real difference or an acceptable prevalence threshold, a type II error that wastes the entire investigation. A study with too many subjects wastes time, resources, and animal lives. The researcher's task is to identify the smallest number that meets the study's objectives with acceptable statistical certainty. This article answers the practical question: which formula do I use, what values do I plug in, and how do I interpret the result for a veterinary population?

At a Glance

Why Veterinary Sample Sizes Differ from Human Studies

Individual animals in a veterinary setting are almost always aggregated into hierarchical groups. Animals within a herd share housing, management, genetics, and pathogen exposure, so their outcomes are not independent. Sample size formulae that assume independence, the standard approach in human epidemiology, systematically underestimate the required number of animals when applied to clustered veterinary populations. Stevenson's overview of sample size estimation in veterinary epidemiologic research identifies this lack of data independence as the central complication distinguishing veterinary from human sample size work.

The consequence is that a crude sample size calculated for a simple random sample must be inflated by a design effect. The design effect depends on the average cluster size and the intracluster correlation coefficient, which measures how similar outcomes are within a group. Ignoring this adjustment produces studies that are underpowered, sometimes severely, even though the arithmetic appears correct.

The Statistical Logic of Sample Size

The Hypothesis Testing Framework

Most comparative studies test a null hypothesis of no difference between groups. The sample size must be large enough that, if a true difference of a specified magnitude exists, the test will reject the null hypothesis with probability equal to the chosen power. Four quantities are mathematically linked: significance level, power, effect size, and sample size. Fix any three and the fourth is determined. The researcher's judgment enters in choosing the effect size, which must represent the smallest difference that would change clinical or policy decisions.

Precision-Based Estimation for Prevalence Surveys

Prevalence surveys do not test a hypothesis. They estimate a proportion with a specified confidence interval. The required sample size depends on the expected prevalence and the desired precision, defined as the half-width of the confidence interval. A survey aiming to detect a prevalence of 5% with precision of ±2% requires a different sample than one aiming to detect 50% with the same precision. The variance of a proportion is largest at 50%, so surveys targeting rare conditions can often use smaller samples, provided the expected prevalence is accurately specified.

The Role of the Design Effect

When animals are sampled in clusters, the effective sample size is smaller than the number of animals sampled. The design effect quantifies this loss. For a given cluster size and intracluster correlation, the design effect multiplies the sample size that would be required under simple random sampling. WOAH animal health surveillance standards explicitly recognize that surveillance designs must account for population structure, and the same principle applies to research studies.

Specifying Input Parameters

Expected Prevalence and Effect Size

The expected prevalence should come from prior studies in comparable populations, pilot data, or structured expert opinion. When the literature is sparse, a conservative approach is to use 50% prevalence, which maximizes the required sample size for a given precision. For comparative studies, the effect size should be the smallest difference that would alter a management decision, not the largest difference the researcher hopes to find. Using an unrealistically large effect size produces a small sample that cannot detect smaller, still meaningful differences.

Significance Level and Power

Convention sets α at 0.05 and power at 0.80, but these values are defaults, not requirements. Studies with serious consequences for a false negative, such as failure to detect a notifiable disease, may justify power of 0.90 or higher. Studies with multiple primary outcomes require adjustment to the significance level to control the family-wise error rate. The researcher should state these choices explicitly and justify them in the study protocol.

Population Size and the Finite Population Correction

When the sampling fraction exceeds roughly 5% of the population, the finite population correction reduces the required sample size. For large populations, the correction is negligible and can be ignored. For small herds or flocks, the correction matters and should be applied. The correction factor is (N, n)/(N, 1), where N is the population size and n is the uncorrected sample size.

Common Study Designs and Their Formulas

Simple Random Sampling for Prevalence

For a simple random sample estimating a proportion, the required sample size is calculated from the expected prevalence, the desired precision, and the z-value corresponding to the chosen confidence level. This formula assumes independence and is appropriate only when animals are sampled individually from a population with no clustering structure.

Comparative Studies with Two Groups

For comparing two proportions or two means, the sample size depends on the expected values in each group, the variability of the outcome, and the chosen power and significance level. Continuous outcomes require an estimate of the standard deviation, which often comes from prior studies. Binary outcomes require the expected proportion in each group. The formula yields the number per group, and the researcher must decide whether groups are equal in size or whether one group is expected to be larger, as in cohort studies with unequal exposure groups.

Logistic Regression and Multivariable Models

When the primary analysis uses logistic regression, sample size calculation becomes more complex. Broll, Glaser, and Kreienbrock describe a simplified method for calculating sample size bounds in multiple logistic regression that extends earlier work by Hsieh. This approach requires specifying the prevalence of the outcome, the prevalence of each exposure, and the expected odds ratio for each predictor. The method demands additional information that is often difficult to obtain, but it is preferable to calculating sample size for a univariable test and then applying the result to a multivariable analysis, which systematically underpower the model.

Hierarchical Data and the Design Effect in Practice

Estimating the Intracluster Correlation Coefficient

The intracluster correlation coefficient is rarely known with precision before a study begins. Published values exist for common outcomes in production animal species, but the researcher should treat these as estimates with uncertainty. A sensitivity analysis, calculating sample size across a plausible range of ρ values, is more informative than a single point estimate. If the study is the first in a population, a conservative ρ of 0.1 to 0.2 is often used, but this choice should be justified.

Calculating the Design Effect

The design effect is calculated as 1 + (m, 1)ρ, where m is the average cluster size and ρ is the intracluster correlation. The inflated sample size is the crude sample size multiplied by the design effect. This inflation applies to the total number of animals, not the number of clusters. The number of clusters is then determined by dividing the inflated sample by the average cluster size. Studies with few clusters and large cluster sizes require the largest inflation, and the number of clusters itself may become the limiting factor.

When the Number of Clusters Is Small

Statistical inference from a small number of clusters is unreliable regardless of the total number of animals sampled. The effective degrees of freedom are determined by the number of clusters, not the number of individuals. A study with 10 clusters of 50 animals each has less statistical information than a study with 50 clusters of 10 animals each, even though both sample 500 animals. Researchers planning studies with fewer than 20 to 30 clusters should consider whether the design can support the planned analysis, and whether cluster-level random effects or generalized estimating equations will be needed in the analysis stage.

Worked Examples for Prevalence Surveys

The formulas in the previous section become practical when applied to a concrete scenario. Consider a cross-sectional survey designed to estimate the prevalence of bovine viral diarrhea virus (BVDV) persistently infected cattle in a region where the expected prevalence is 1%. The researcher chooses a 95% confidence level, a desired absolute precision of 0.5%, and a population of 10,000 cattle.

The crude sample size for simple random sampling is calculated as:

n = (Z² × p × (1 - p)) / d²

where Z = 1.96, p = 0.01, and d = 0.005. This yields:

n = (1.96² × 0.01 × 0.99) / 0.005² = 1,521

Because the population is finite, apply the finite population correction:

n_adj = (n × N) / (n + N) = (1,521 × 10,000) / (1,521 + 10,000) = 1,320

If cattle are sampled from herds instead of individually, the design effect must be applied. With an estimated intracluster correlation coefficient of 0.05 and an average of 50 cattle sampled per herd, the design effect is:

DE = 1 + (m - 1) × ICC = 1 + (50 - 1) × 0.05 = 3.45

The required sample size becomes 1,320 × 3.45 = 4,554 cattle. This figure exceeds the population of many small herds, which signals that the sampling strategy itself may need revision. Reducing the number sampled per herd to 20 lowers the design effect to 1.95 and the required sample to 2,574 cattle. The researcher must weigh the logistical cost of visiting more herds against the statistical cost of sampling more animals within each herd. Stevenson's overview of sample size estimation in veterinary epidemiologic research emphasizes that hierarchical data structures are the norm in veterinary settings and that ignoring them produces sample sizes that are too small.

Worked Examples for Comparative Studies

For a two-group comparative study, the sample size per group depends on the expected difference in outcomes and the variability of the outcome measure. Suppose a researcher plans a randomised trial comparing two treatment protocols for canine otitis externa, with bacterial cure at 14 days as the primary outcome. The standard protocol is expected to achieve a 70% cure rate, and the new protocol is considered clinically superior if it achieves 85%. With α = 0.05 and power = 0.80, the sample size per group is:

n = [(Z_α/2 + Z_β)² × (p₁(1 - p₁) + p₂(1 - p₂))] / (p₁ - p₂)²

where p₁ = 0.70, p₂ = 0.85, Z_α/2 = 1.96, and Z_β = 0.84. This gives:

n = [(1.96 + 0.84)² × (0.21 + 0.1275)] / (0.15)² = 7.84 × 0.3375 / 0.0225 = 117.6

Each group requires 118 dogs. If the researcher instead chose a 10 percentage point difference as the minimum clinically important effect, the required sample size rises to 199 dogs per group. The choice of the minimum clinically important difference is therefore the single most influential decision in the calculation. The simplified method for sample size in logistic regression models extends this logic to multivariable settings, where additional covariates absorb some of the outcome variance and the required sample size depends on the prevalence of each exposure and the correlation between exposures.

For continuous outcomes, the formula uses the standard deviation of the outcome instead of proportions. A study comparing weight gain in lambs on two feeding regimens, with a standard deviation of 1.5 kg and a target difference of 1.0 kg, requires:

n = [2 × (Z_α/2 + Z_β)² × σ²] / δ² = [2 × 7.84 × 2.25] / 1.0 = 35.3

Each group requires 36 lambs. The standard deviation must come from published literature or pilot data, and an underestimate produces an underpowered study. When the standard deviation is uncertain, inflate it by 20% to 30% as a conservative measure.

Software and Computational Tools

Several software packages implement the formulas described above. Open-source options include R packages such as epiR and pwr, which handle prevalence surveys, comparative studies, and design effect adjustments. Commercial packages include PASS and nQuery, which offer graphical interfaces and extensive documentation. Spreadsheet implementations are feasible for simple random sampling and two-group comparisons, but they become error-prone when design effects or multivariable adjustments are introduced.

The choice of software matters less than the clarity of the assumptions entered. A calculation performed in any package is only as valid as the inputs. Document every input parameter and its source before running the calculation, and record the version of the software and the date. This documentation supports the grant application and allows reviewers to reproduce the calculation. Funding agencies now require justification of the number of subjects enrolled and details of the assumptions and methodologies used to derive sample size estimates, so the audit trail is part of the scientific record.

Sensitivity Analysis and Reporting

A single sample size calculation conveys false certainty. The inputs, particularly the expected prevalence, the minimum clinically important difference, and the intracluster correlation coefficient, are estimates with their own uncertainty. A sensitivity analysis repeats the calculation across a plausible range of each input and reports the range of sample sizes that result.

The table illustrates a general principle: when the sample size range spans more than a factor of two across plausible inputs, the study design itself needs reconsideration instead of a single point estimate. A prevalence survey that requires 6,500 animals under pessimistic assumptions may be logistically impossible, and the researcher should consider a different sampling strategy, a two-stage design, or a broader case definition. The stochastic modeling approach used to examine canine leishmaniasis seroprevalence demonstrates how population structure and sampling strategy interact to affect prevalence estimates, and the same logic applies to sample size planning.

Report the sample size calculation in the methods section with the following elements: the primary outcome and its expected value in each group, the minimum clinically important difference, the significance level, the power, the design effect and its components, the software used, and the results of the sensitivity analysis. This level of detail allows a reviewer to identify the weakest assumption and to judge whether the study is adequately powered under realistic conditions.

Common Errors and How to Avoid Them

The most frequent error in veterinary sample size calculations is ignoring clustering. Animals within a herd or kennel are more similar to each other than to animals in other groups, and treating them as independent inflates the effective sample size. The design effect correction is mandatory whenever the sampling unit differs from the analysis unit. Veterinary study subjects are almost always aggregated into hierarchical groups, and sample size estimates calculated using formulae that assume data independence are not appropriate.

The second most common error is using the expected prevalence from a different population or time period. Prevalence estimates vary with geography, production system, breed, and season. A prevalence of 1% in one region may be 5% in another, and the sample size changes by a factor of nearly five. Use local surveillance data where available, and consult WOAH animal health surveillance standards for guidance on data quality and representativeness.

The third error is confusing the unit of analysis with the unit of sampling. In a herd-level study, the herd is the unit of analysis and the number of herds determines the power. Sampling more animals within each herd improves the precision of the herd-level estimate but does not increase the number of independent observations. The distinction is critical in production animal research, where the number of herds is often small and the number of animals per herd is large.

The fourth error is failing to account for losses to follow-up. Veterinary studies lose subjects to death, culling, withdrawal, and owner non-compliance. A study that requires 118 dogs per group with 10% expected loss needs 132 dogs per group at enrollment. The loss rate should be estimated from previous studies in the same population and stated explicitly in the methods.

The fifth error is using a one-sided test when a two-sided test is appropriate. One-sided tests are only justified when a difference in the opposite direction is clinically impossible or irrelevant. In most veterinary comparisons, the new treatment could plausibly be worse than the standard, and a two-sided test is required. The choice affects the Z value and therefore the sample size, and it must be justified in the protocol.

Recognized Complications and Failure Modes

The most consequential failure in sample size work is not mathematical. It is the mismatch between the calculated sample and the population actually sampled. A prevalence survey designed for a simple random sample will produce biased estimates if field constraints force convenience sampling. The calculated precision then describes a study that was not performed. Detect this early by comparing the sampling frame used in the calculation with the recruitment log after the first week of data collection. If the animals enrolled differ systematically from the target population by age, breed, or production class, the sample size calculation must be repeated with the realised sampling design.

A second failure mode is the silent loss of statistical power through attrition. Sample size formulae assume all enrolled subjects contribute complete data. In longitudinal veterinary studies, losses from culling, death, or owner withdrawal are common. Monitor the cumulative loss rate against the anticipated rate used in the calculation. If losses exceed 10 percent of the enrolled sample, the effective power has dropped below the planned level. The corrective action is either to enrol additional subjects prospectively or to acknowledge the reduced power in the final report.

A third failure mode concerns the design effect. Researchers frequently borrow an intracluster correlation coefficient from published work without verifying that the cluster size and outcome prevalence resemble their own setting. The design effect scales linearly with cluster size, so a small error in the ICC produces a large error in the final sample size. Check the sensitivity of your result to the ICC by recalculating with a range of plausible values before finalising the protocol.

Common Errors and Corrective Actions

Less experienced researchers often confuse the unit of analysis with the unit of sampling. In a herd-level study, the herd is the sampling unit and the animal is the observational unit. Calculating sample size at the animal level without a design effect produces a sample that is far too small, because animals within a herd are correlated. The corrective action is to identify the primary unit of inference first and then apply the appropriate formula or inflation factor.

A second frequent error is the misuse of one-tailed tests. Veterinary researchers sometimes choose a one-tailed test to reduce the required sample size. This is only defensible when a difference in the opposite direction is scientifically impossible or clinically irrelevant. For most comparative studies, a two-tailed test is the conservative and expected choice. Reviewers and funding agencies will question a one-tailed justification that rests on convenience instead of biological reasoning.

A third error is the failure to adjust for multiple outcomes. Studies that measure several primary outcomes require either a larger sample size to preserve overall power or a stated hierarchy of outcomes with a single primary endpoint. Without this adjustment, the probability of at least one false positive finding rises steeply. The corrective action is to designate one primary outcome before the study begins and treat all others as secondary.

Limitations of the Current Evidence

The evidence base for sample size methods in veterinary epidemiology is thinner than in human epidemiology. Most published formulae derive from human clinical research and are adapted to animal populations with varying success. The institutional review of sample size estimation in veterinary epidemiologic research notes that veterinary study subjects are almost always aggregated into hierarchical groups, and formulae assuming data independence are therefore inappropriate. The same review acknowledges that reliable ICC estimates for many species and production systems are scarce.

Expert opinion still differs on the minimum number of clusters required for valid inference. Some authors accept as few as 10 clusters when the ICC is small, while others recommend 30 or more. There is no universal threshold, and the appropriate minimum depends on the balance between cluster size, ICC, and the expected effect size. Researchers should state their assumption explicitly and justify it from the literature or from pilot data.

Bayesian approaches offer an alternative when sample sizes are necessarily small. Bayesian accelerated failure time methods can be applied at any sample size and allow formal incorporation of prior information. However, the choice of prior distribution materially affects the results, and this subjectivity deters some researchers. The simplified method for logistic regression sample sizes is useful but demands additional information that is often difficult to obtain, particularly the prevalence of the outcome and the correlation among predictors.

Escalation and Consultation

Most sample size problems can be resolved with careful reading of the CDC principles of epidemiology and the WOAH animal health surveillance standards. Refer to a veterinary epidemiologist or biostatistician when the study involves hierarchical data with more than two levels, when the outcome is rare, or when the sampling design departs from simple random sampling. Laboratory involvement is warranted when diagnostic test sensitivity and specificity materially affect the required sample size, since imperfect tests inflate the number of animals needed to achieve a given precision.

Regulatory reporting obligations arise when the study is part of a notifiable disease surveillance program. The WOAH terrestrial animal health code specifies surveillance requirements for listed diseases, and sample sizes for these programs must satisfy international standards instead of local convenience. Consult the relevant veterinary authority before finalising the protocol.

Frequently Asked Questions

How Do I Adjust My Sample Size When Funding or Time Is Limited?

When resources constrain the ideal sample size, recalculate the study around a reduced objective instead of silently accepting a smaller number. First, consider whether a narrower research question, such as a single primary outcome instead of several, preserves the core aim. Second, increase the expected effect size or prevalence difference only if the literature supports that choice, since inflating these inputs to justify a smaller sample introduces bias. Third, accept a lower power, for example 0.80 instead of 0.90, and state this explicitly in the methods. Stevenson's overview of sample size estimation in veterinary research notes that underpowered studies risk failing to detect genuine differences, so document every compromise and report the achieved power in the final manuscript.

What Should I Do If I Cannot Obtain a Reliable Estimate of the Intracluster Correlation Coefficient?

Use published ICC values from similar species, production systems, and outcomes, and then run a sensitivity analysis across a plausible range. For example, if the literature suggests an ICC between 0.05 and 0.15, calculate the design effect at both extremes and report the resulting sample size range. If no published value exists, collect pilot data from two or three clusters and estimate the ICC directly. Stevenson's review of veterinary sample size methods emphasizes that hierarchical data require design effect adjustment, and the cost of guessing too low is a study that is genuinely underpowered. When clusters are few, consider increasing the number of clusters instead of the cluster size, since the design effect responds more to cluster count.

How Does the Required Sample Size Change When I Move from Cattle to Dogs or Cats?

The statistical formula does not change across species, but the input parameters often do. Companion animal populations are typically less structured than production herds, so the design effect may be smaller or unnecessary if animals are sampled independently. However, household clustering matters in multi-pet households, and shelter or colony housing introduces clustering comparable to a herd. Expected prevalence varies by species, region, and management system, so use species-specific estimates from regional surveillance data. The World Organization for Animal Health surveillance standards describe how population structure and sampling frames differ across production systems, and those differences should guide your assumptions about clustering and prevalence instead of the species label itself.

What Records Must I Keep to Support My Sample Size Justification During Peer Review?

Retain the protocol document that lists every input parameter, the source for each value, and the formula or software used. Record the version of the software and the date of calculation. Keep the sensitivity analysis outputs, since reviewers often ask how the sample size would change under alternative assumptions. Document any deviations from the original plan, including the date, the reason, and who approved the change. Funding agencies increasingly require sample size justification as a mandatory component of grant applications, as noted in Stevenson's review of veterinary epidemiologic research methods. A spreadsheet with named cells for each parameter is easier to audit than a calculation embedded in prose.

How Do I Explain the Sample Size to a Producer or Practice Owner Who Wants Results Quickly?

Frame the sample size as a guarantee of interpretability instead of an administrative burden. Explain that too few animals produce results that could miss a real disease problem or falsely suggest one exists, and that either outcome wastes the money already spent. Use a concrete example from their operation, such as the number of animals needed to detect a 10 percent prevalence increase with confidence. Emphasize that the calculation protects their investment in testing and treatment decisions. The CDC principles of epidemiology in public health practice describe how sample size and study design determine whether surveillance findings can support action, and that logic translates directly when discussing expected disease frequency with an owner.

When Should I Consult a Statistician instead of Proceeding with a Standard Formula?

Consult a statistician when your data have more than two levels of clustering, when you plan a logistic regression with many predictors, or when the number of clusters is fewer than roughly 15. Broll and colleagues describe a simplified method for sample size in multiple logistic regression that requires additional information about covariate distributions, and that information is often difficult to obtain without expert input. Also seek help when you expect sparse outcomes, such as a rare disease, or when you plan a Bayesian analysis, since Bayesian sample size approaches differ fundamentally from frequentist calculations. A statistician can also help when your sampling frame is incomplete or when you must combine data from multiple sources with different structures.

Related Clinical & Scientific Guides

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

References and Further Reading

• Sample Size Estimation in Veterinary Epidemiologic Research.. 2020.
• Calculating sample size bounds for logistic regression.. 2002.
• Bayesian accelerated failure time analysis with application to veterinary epidemiology.. 2000.
• Variations in seroprevalences of canine leishmaniasis: Could it be a consequence of the population structure?. 2016.
• WOAH Animal Health Surveillance Standards. WOAH.
• CDC Principles of Epidemiology in Public Health Practice. CDC.
• MSD Veterinary Manual, Professional Edition. MSD Veterinary Manual.
• American Veterinary Medical Association Practice Resources. American Veterinary Medical Association.
• WOAH Terrestrial Animal Health Code. WOAH.

Related Articles

• Sample Size Calculations for Veterinary Surveys
• Understanding Bias in Veterinary Epidemiological Studies
• Cluster Sampling in Veterinary Field Studies
• Confounding in Veterinary Studies: Identification and Control
• Understanding Ecological Studies in Veterinary Epidemiology

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.