Statistical Inference in Veterinary Epidemiology: Hypothesis Testing and Confidence Intervals

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

Key Takeaways

• Statistical inference in veterinary epidemiology bridges sample data to population-level conclusions, with hypothesis testing and confidence intervals being the primary methods. Hypothesis testing falsifies a null hypothesis (e.g., no treatment effect), while confidence intervals estimate a range of plausible population parameter values, offering greater insight into effect size and precision.
• The unit of analysis is critical; treatments assigned to groups (e.g., pens, herds) necessitate group-level analysis to avoid inflated sample sizes and biased probability statements, a common pitfall in livestock research. Ignoring clustering leads to overly narrow confidence intervals and overly small p-values.
• P-values represent the probability of observing data as extreme as, or more extreme than, the observed data if the null hypothesis were true; they do not indicate the probability that the null hypothesis is true or the magnitude of an effect. Statistical significance (e.g., p < 0.05) does not equate to biological or clinical significance.
• Confidence intervals provide a range of plausible values for a population parameter, with their width reflecting the precision of the estimate. A 95% confidence interval indicates that, in repeated sampling, 95% of such intervals would contain the true population value, offering crucial information for clinical decision-making beyond a simple binary test outcome.
• Census data, representing every member of a defined population (e.g., all racehorse retirements in a specific jurisdiction), do not require statistical inference to a broader population; descriptive statistics are sufficient. Uncertainty in census data relates to measurement error or generalizability, not sampling error.
• Power calculations and sample size determinations must explicitly account for the unit of analysis and the desired precision. Underpowered studies, particularly those with small sample sizes common in veterinary clinical trials, may fail to detect clinically meaningful effects, a limitation best illuminated by the confidence interval.

Statistical inference is the formal process by which veterinary researchers draw conclusions about animal populations from data collected on samples. This article explains the two principal modes of inference, hypothesis testing and confidence interval estimation, as they apply to veterinary data across species and production systems. It is written for veterinary researchers and graduate students who design studies, analyze clinical or field data, and interpret the veterinary literature. The article addresses a specific question: how does one move from observed sample statistics to defensible statements about the target population, and what errors threaten that movement?

The clinical and academic stakes are substantial. A conclusion that a treatment reduces disease incidence, that a diagnostic test discriminates between conditions, or that a management practice alters production outcomes is only as strong as the inferential framework supporting it. Veterinary data present particular challenges, including clustered animals within pens or herds, variable sampling intensity from tracking devices, and censuses that require no inference at all. Each of these contexts modifies how hypothesis tests and confidence intervals should be constructed and read.

At a Glance

The Logic of Hypothesis Testing

Hypothesis testing is a decision procedure built on falsification. The researcher specifies a null hypothesis, typically that no effect exists, and then asks whether the observed data are sufficiently incompatible with that hypothesis to warrant rejection. The alternative hypothesis states the effect the study is designed to detect. The logic is probabilistic: if the null hypothesis were true, how often would data at least as extreme as those observed arise by chance alone? That probability is the p-value.

The p-value must be interpreted with care. It is not the probability that the null hypothesis is true, nor is it the probability that the alternative is false. It is a conditional probability about the data under a specified model. A small p-value indicates that the data are unusual under the null model, which is evidence against the null, but the magnitude of the effect and its biological importance are separate questions. The CDC principles of epidemiology in public health practice emphasize that statistical significance does not equate to public health or clinical significance, a distinction that applies equally to veterinary medicine.

The significance level, alpha, is the pre-specified threshold for rejecting the null hypothesis. Setting alpha at 0.05 means the researcher accepts a 5% chance of a Type I error, concluding an effect exists when it does not. The choice of alpha should reflect the consequences of a false positive. In regulatory toxicology or food safety work, a smaller alpha may be warranted because the cost of wrongly declaring a compound safe is severe. In exploratory studies, a larger alpha may be acceptable to avoid missing candidate effects.

Confidence Intervals and Estimation

Where hypothesis testing answers a yes or no question, confidence intervals answer a quantitative one: what range of values is compatible with the population parameter given the data? A 95% confidence interval is constructed so that, across repeated sampling, 95% of such intervals would contain the true population value. This frequentist interpretation is subtle. For any single interval, the true value either lies within it or does not, the confidence is in the procedure, not the particular interval.

Confidence intervals convey precision directly. A narrow interval around a risk ratio or mean difference indicates precise estimation, while a wide interval signals uncertainty, often from small sample sizes or high variability. Intervals also preserve the connection between statistical and clinical significance. An interval that excludes the null value supports rejection of the null hypothesis at the corresponding alpha level, but an interval that includes the null value does not prove the absence of an effect. It may simply reflect insufficient power.

Units of Analysis and the Pen Problem

Veterinary studies frequently assign treatments to groups of animals instead of to individuals. In pen-based livestock research, for example, a dietary treatment is applied to an entire pen, yet individual animals within the pen are measured. If the analysis treats individual animals as independent observations, the effective sample size is inflated and the test of significance is biased. St-Pierre's review of design and analysis of pen studies in the animal sciences demonstrates that using animals as the error term when treatments are applied to pens produces biased probability statements and invalid causal inference. The pen is the experimental unit for the treatment effect, and the analysis must reflect that structure.

This principle extends beyond pens. Herds, litters, flocks, and even repeated measurements on the same animal all create clustering that violates the independence assumption of standard tests. Ignoring clustering produces p-values that are too small and confidence intervals that are too narrow. The result is overconfidence in effects that may not exist.

Inference in Observational and Tracking Data

Not all veterinary data arise from designed experiments. Observational studies, including those using GPS telemetry, present distinct inferential challenges. The critical review of GPS telemetry data in ecology notes that technological capacity does not guarantee inferential validity. Collar failures, high costs, and reduced sample sizes can weaken study design and compromise statistical inference. Similarly, the guidance on availability sampling in resource selection functions shows that the choice of availability sample influences coefficient estimates and can introduce bias when the spatial extent is mis-specified.

In some observational settings, inference is unnecessary. The descriptive analysis of Thoroughbred racehorse retirements due to tendon injuries explicitly states that because the data constitute a complete census of Hong Kong Jockey Club retirement records, no statistical inference to a population is required. The distinction between census and sample is fundamental. When every member of the target population is measured, descriptive statistics are the final product, and hypothesis testing adds no information.

The Pseudoreplication Debate

The concept of pseudoreplication, the treatment of non-independent observations as independent, has shaped veterinary and ecological study design for decades. Yet the doctrine is not without critics. Schank and Koehnle argue that pseudoreplication is a pseudoproblem, contending that the concept rests on a misunderstanding of statistical independence and the nature of control groups. They advocate judging each study on its own merits instead of applying universal criteria for acceptance or rejection.

This debate matters for veterinary researchers because it affects how manuscripts are reviewed and how study designs are evaluated. The practical resolution is to think carefully about the inferential context. When treatments are applied to groups, the group is the unit of analysis. When the research question concerns individual-level processes and the data support that level of inference, individual-level analysis may be appropriate. The key is that the analysis must match the question and the design.

Choosing Between Hypothesis Testing and Confidence Intervals

The decision to emphasize hypothesis testing or confidence intervals should follow the research question, not habit. Hypothesis testing answers a binary question: does the evidence contradict a specified null value? Confidence intervals answer a quantitative question: within what range of values does the population parameter plausibly lie? For veterinary research, the interval is usually the more informative choice because it preserves the magnitude and direction of an effect, which is what clinical and production decisions require.

A confidence interval that excludes the null value corresponds to a significant test at the same alpha level, but the interval conveys more. A p-value of 0.04 and a p-value of 0.001 can both be labelled significant, yet the second carries stronger evidence. The interval makes that distinction visible. Conversely, a wide interval that includes the null value does not prove the absence of an effect. It indicates that the data are compatible with a range of possibilities, including clinically meaningful ones. This distinction matters in veterinary medicine where sample sizes are often constrained by cost, animal welfare, and the availability of naturally occurring disease.

Setting the Null and Alternative Hypotheses

The null hypothesis must be specified before analysis and should represent a meaningful reference point. In veterinary studies this is often zero difference between treatment groups, but it can also be a threshold such as a minimum acceptable vaccine efficacy or a maximum tolerable prevalence for a herd-level certification program. The alternative hypothesis should reflect the direction of interest when prior evidence supports one. One-sided tests are appropriate only when a difference in the opposite direction is biologically implausible or clinically irrelevant, and they must be justified in the study protocol.

The choice of alpha should be made in advance and justified. The conventional 0.05 is a default, not a law. Studies with small sample sizes, multiple endpoints, or high costs of a false positive may warrant a more stringent threshold. Studies that are exploratory or that screen many candidate risk factors may tolerate a more liberal threshold, provided the results are labelled as hypothesis-generating. The CDC principles of epidemiology in public health practice describe the same logic for human populations, and the reasoning transfers directly to animal health investigations.

Power and Sample Size as Inference Decisions

Statistical inference is only as strong as the study that produces it. A study with low power can miss a real effect, and the resulting non-significant p-value is frequently misread as evidence of no effect. Power depends on the true effect size, the variability in the outcome, the sample size, and the alpha level. Of these, the effect size is the most difficult to specify because it requires a judgment about what difference is biologically or economically important. That judgment should come from clinical experience, published literature, or pilot data, not from the observed data after the study is complete.

Sample size calculations must account for the unit of analysis. When animals are housed in pens and the treatment is applied at the pen level, the pen is the experimental unit and the sample size must be based on the number of pens, not the number of animals. Ignoring this structure inflates the degrees of freedom and produces biased probability statements about treatment effects, as described in design and analysis of pen studies in the animal sciences. The same principle applies to group-housed pigs, floor pens in poultry, and pasture groups in cattle. A study that appears adequately powered on an animal-level calculation may be severely underpowered when the pen-level structure is recognized.

Interpreting Non-Significant Results

A non-significant result has three possible explanations: the null hypothesis is true, the effect is real but smaller than the study could detect, or the study was poorly designed or executed. The confidence interval is the tool that distinguishes among these. If the interval is narrow and excludes clinically meaningful effects, the non-significant result is informative. If the interval is wide and includes both the null value and a clinically important effect, the study is uninformative regardless of the p-value.

This reasoning is particularly relevant in veterinary clinical trials where sample sizes are often small. A trial comparing two analgesic protocols in 20 dogs per group may show no significant difference in pain scores, but the confidence interval may include a difference large enough to matter in practice. Reporting only the p-value in that situation is misleading. The interval should be reported and interpreted against a pre-specified clinically important difference.

Inference in Surveillance and Census Data

Not all veterinary data require statistical inference. When data constitute a complete census of a defined population, the observed values are parameters, not estimates. The descriptive analysis of Thoroughbred racehorse retirements due to tendon injuries at the Hong Kong Jockey Club is an example. Because the study included every retirement in the population over the study period, the authors correctly noted that no statistical inference to a broader population was necessary. The reported cumulative incidence and proportions described the population directly.

Census data are common in veterinary medicine. A herd health record system that captures every animal in a closed herd, a slaughterhouse surveillance program that examines every carcass, and a national notification database that records every confirmed case of a reportable disease all generate census data. In these settings, confidence intervals and p-values describe uncertainty about sampling, but there is no sampling. The uncertainty that remains is about measurement error, missing data, and the generalizability of the findings to other populations or time periods. Those sources of uncertainty are not captured by a confidence interval and must be addressed through study design and external comparison.

Surveillance standards from the World Organization for Animal Health emphasize that the purpose of surveillance is to detect disease and support trade decisions. When a surveillance system is designed to detect a minimum prevalence with a specified confidence, the inference framework is built into the sampling design. The confidence level of the surveillance system is a statement about the probability of detecting at least one positive if the disease is present at or above the design prevalence. This is a different inferential target than estimating prevalence with a confidence interval, and the distinction should be explicit in the reporting.

Documenting Inference Decisions

The study report should state the primary outcome, the null and alternative hypotheses, the alpha level, the power or precision target, and the unit of analysis. Each of these decisions should be made before data collection and documented in the protocol. Post-hoc changes to the analysis plan, such as switching from a two-sided to a one-sided test or changing the primary outcome, weaken the inferential basis of the study and should be disclosed.

The results section should report effect estimates with confidence intervals for all primary outcomes. P-values may be reported alongside the intervals, but they should not replace them. For secondary outcomes, intervals without p-values are often sufficient. The discussion should interpret the intervals against the pre-specified clinically important difference, not against the null value alone. This structure applies across species and production systems, from companion animal clinical trials to feedlot intervention studies to wildlife disease surveillance. The MSD Veterinary Manual and AVMA practice resources both emphasize evidence-based decision making in clinical practice, and the same standards of transparent inference reporting should govern research that informs those decisions.

Recognized Failure Modes in Inference

Statistical inference fails in predictable ways, and most failures are detectable before analysis begins. The most consequential failure is the unit-of-analysis error, where treatments are applied to groups but outcomes are measured on individuals. In pen studies, using individual animals as the error term when treatments are applied to pens produces inflated degrees of freedom and biased probability statements about treatment effects. The discriminating check is simple: identify the level at which treatment was assigned, then confirm that the error term in the model corresponds to that level, not to the level of measurement. St-Pierre's review of pen study design provides worked examples of the bias that results when this correspondence is ignored.

A second failure mode is the conflation of statistical significance with biological or clinical importance. A large study can return a small p-value for a trivial effect, while a small study can miss a clinically meaningful difference. Confidence intervals mitigate this error by displaying the range of plausible effect sizes. When the interval includes values that would change clinical decisions, the evidence is insufficient regardless of the p-value.

A third failure mode is the misuse of availability samples in habitat selection and movement studies. The size and spatial extent of the availability sample influence coefficient estimates, and spatial autocorrelation in covariates exacerbates the resulting bias. Northrup and colleagues' guidance on availability sampling recommends sensitivity analysis: rerun the model under different availability definitions and report whether conclusions shift.

Common Errors and Corrective Actions

Less experienced analysts often default to hypothesis testing when estimation would serve better. A test answers one question, whether the null is plausible. An interval answers the clinical question, how large is the effect and how precisely is it known. When the goal is to inform a treatment recommendation or a culling decision, report the interval and its clinical interpretation.

A related error is the mechanical application of a threshold. Treating p < 0.05 as a binary pass-fail criterion ignores the continuity of evidence. A p-value of 0.051 in a small study and a p-value of 0.049 in a large study do not represent comparable evidence. Report the exact value, the interval, and the study size.

A third error is the neglect of clustering in observational data. Animals from the same herd, pen, or litter are not independent observations. Ignoring this structure produces standard errors that are too small and conclusions that are too confident. The corrective action is to fit a mixed model or use cluster-robust variance estimation, and to report the intracluster correlation coefficient where relevant.

Limitations of the Evidence and Areas of Disagreement

The pseudoreplication debate illustrates genuine disagreement about the universality of design rules. Schank and Koehnle's critique argues that the doctrine rests on a misunderstanding of statistical independence and that no universal criteria can govern acceptance or rejection of research. This does not license careless design. It does mean that the unit-of-analysis problem must be reasoned through for each study instead of applied as a checklist.

Movement ecology presents a different limitation. Hebblewhite and Haydon's review of GPS telemetry notes that collar failures and high cost often force weaker study designs and reduced sample sizes, which directly compromise statistical inference. The field-based understanding of animal ecology that once accompanied data collection is increasingly rare, and the loss affects interpretation even when the mathematics is sound.

Surveillance data carry their own constraints. Complete census data, such as the retirement records of racehorses in a single jurisdiction, require no statistical inference to a broader population, but the descriptive patterns still need careful interpretation. The Hong Kong Jockey Club tendon injury study demonstrates that census data can reveal trends, yet those trends may not generalize to other populations or management systems.

Escalation and Referral

Most inference problems are resolved at the design stage, which means the time to consult a statistician is before data collection, not after. Referral to a statistical collaborator is warranted when the study involves clustered or hierarchical data, repeated measures, missing data, or complex sampling designs. Laboratory involvement is appropriate when diagnostic test performance must be quantified, because sensitivity, specificity, and predictive values carry their own inferential requirements.

Regulatory reporting obligations are distinct from statistical inference. When surveillance or diagnostic findings meet notifiable disease criteria, the reporting duty follows the standards of the World Organization for Animal Health and the WOAH Terrestrial Animal Health Code. These obligations apply regardless of whether the finding is statistically significant, and they vary by jurisdiction and species.

Troubleshooting Table

Frequently Asked Questions

How Should I Handle Statistical Inference When Budget Constraints Limit Sample Size?

Budget constraints are a reality of veterinary research, but they must be addressed transparently instead of ignored. When resources limit sample size, the study's power to detect meaningful effects decreases, and confidence intervals widen. You can partially compensate by using more efficient designs, such as paired or crossover arrangements, or by measuring outcomes with greater precision. However, you cannot manufacture power that the design does not provide. Report the achieved power and the width of confidence intervals honestly, and frame conclusions accordingly. If the funding situation forces a choice between fewer animals with intensive monitoring or more animals with less precise measurement, consider which error type carries greater scientific cost for your specific question.

What Do I Do When Animals Are Grouped in Pens but I Cannot Afford True Pen Replication?

When treatments are applied at the pen level but animals within pens are used as the unit of analysis, probability statements become biased and significance tests are invalid. If true pen replication is impossible, you must acknowledge that causal inference is compromised. One option is to treat the study as observational instead of experimental, analyzing pen-level outcomes descriptively and reporting the limitations explicitly. Another is to use a split-plot framework if the design permits, where pens serve as main plots and animals as subplots. St-Pierre's review of pen studies in animal sciences outlines why ignoring pen-level treatment assignment inflates degrees of freedom and biases treatment effect estimates. When resources preclude proper replication, the honest path is to downgrade the strength of your conclusions.

How Do Inference Principles Differ When Working with Wildlife Tracking Data Versus Herd-Level Clinical Data?

Wildlife tracking data present distinct inference challenges because the sampling process is often dictated by technology instead of study design. GPS collar failures and high costs frequently reduce sample sizes and weaken statistical power, as noted in the critical review of GPS telemetry in ecology. Additionally, the availability sample in habitat selection studies can bias coefficient estimates if the spatial extent is mischaracterised. Herd-level clinical data typically involve planned sampling frames and clearer population definitions, though pen effects and clustering still require attention. For tracking data, you must calibrate measurement error, assess sensitivity to the availability sample, and consider continuous-time movement models that separate the movement process from the sampling schedule. The inferential target in wildlife studies is often a spatial process instead of a treatment effect.

What Should I Document in My Study Records to Make My Inference Decisions Reproducible?

Document the sampling frame, the unit of analysis, and the justification for that choice. Record how the null and alternative hypotheses were specified before data collection, including the effect size considered clinically meaningful. Note the target power, the assumed variability, and the sample size calculation with its inputs. Describe how clustering was handled, whether pens, herds, or litters were treated as random effects, and what denominator degrees of freedom were used. Keep a log of any deviations from the analysis plan, including missing data handling and model selection steps. The WOAH animal health surveillance standards emphasize that surveillance and research activities should be documented sufficiently for independent evaluation. This documentation allows a reviewer to assess whether your inference decisions were defensible.

How Do I Explain a Non-Significant Result to a Producer or Practice Owner Without Misleading Them?

A non-significant result does not mean the treatment had no effect. Explain that the study lacked sufficient power to detect a difference of the size that would matter clinically, or that the confidence interval includes both meaningful benefit and meaningful harm. Use plain language: the data do not allow us to say with confidence that the treatment works, but they also do not prove it fails. If the confidence interval is narrow and excludes clinically important effects, you can state that the treatment is unlikely to have a large effect. If the interval is wide, recommend a larger study or a different outcome measure. Frame the decision in terms of risk and cost instead of statistical significance alone.

When Should I Prefer a Confidence Interval Over a P-Value in Clinical Decision-Making?

Confidence intervals are generally more informative for clinical decisions because they convey the magnitude and precision of an effect, also whether it differs from zero. A p-value tells you whether the observed data are unlikely under the null hypothesis, but it does not tell you how large the effect is or whether it matters clinically. When advising on treatment choices, use the confidence interval to assess whether the plausible range of effects includes values that would change your recommendation. For surveillance and census data, where no sampling error exists, inference to a broader population is unnecessary, as demonstrated in the retrospective analysis of Thoroughbred racehorse tendon injuries. In those settings, descriptive measures may be more appropriate than formal hypothesis tests.

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

• Design and analysis of pen studies in the animal sciences.. 2007.
• Distinguishing technology from biology: a critical review of the use of GPS telemetry data in ecology.. 2010.
• Practical guidance on characterizing availability in resource selection functions under a use-availability design.. 2013.
• Descriptive analysis of retirement of Thoroughbred racehorses due to tendon injuries at the Hong Kong Jockey Club (1992-2004).. 2007.
• Pseudoreplication is a pseudoproblem.. 2009.
• Scale-insensitive estimation of speed and distance traveled from animal tracking data.. 2019.
• WOAH Animal Health Surveillance Standards. WOAH.
• CDC Principles of Epidemiology in Public Health Practice. CDC.
• MSD Veterinary Manual, Professional Edition. MSD Veterinary Manual.

Related Articles

• Using Simulation Models in Veterinary Epidemiology
• Basic Reproductive Ratio (R0) in Veterinary Epidemiology
• Effect Modification and Interaction in Veterinary Epidemiology
• Likelihood Ratios in Veterinary Diagnostic Testing
• Spatio-Temporal Modeling of Animal Diseases

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.