# Propensity Score Methods in Veterinary Observational Studies


## Key Takeaways

- Propensity score methods are essential for reducing confounding in veterinary observational studies by balancing measured covariates between exposed and unexposed groups, thereby approximating randomization. The core assumption is that all confounders are measured and correctly modeled, a critical limitation given the potential for unmeasured factors like subclinical disease or owner compliance.
- The propensity score, defined as the probability of receiving an exposure conditional on observed covariates, is typically estimated using logistic regression. Key methods for bias reduction include matching, inverse probability of treatment weighting (IPTW), and stratification, each with distinct strengths and limitations regarding data retention and generalizability.
- Balance diagnostics, primarily using standardized mean differences with a common threshold of 0.10, are crucial for verifying that propensity score adjustment has effectively reduced covariate imbalances between groups. Poor overlap in propensity score distributions between exposure groups is a common failure mode, necessitating trimming or alternative approaches that alter the target estimand.
- Covariate selection for the propensity score model must be guided by causal reasoning and a pre-constructed causal diagram, including variables that influence both exposure and outcome, while excluding mediators and instrumental variables to prevent bias. The outcome itself must never be used in the score model construction.
- Reporting transparency is paramount, requiring detailed descriptions of the propensity score model, included covariates, balance diagnostics before and after adjustment, and the number of subjects excluded due to lack of overlap. Sensitivity analyses, such as comparing results from matching versus weighting or incorporating doubly robust estimators, are vital for assessing the robustness of causal inferences.

---

Veterinary clinical research frequently relies on observational data because randomised controlled trials are costly, ethically constrained, or impractical in many clinical and production settings. When treatment assignment is not randomised, exposed and unexposed animals typically differ in baseline characteriztics that also influence the outcome. These imbalances, known as confounding, can distort or even reverse the apparent effect of an exposure. Propensity score methods offer a structured approach to reducing this bias by balancing measured covariates between comparison groups before estimating the treatment effect.

This article explains how to design, execute, and report propensity score analyzes in veterinary observational studies. It serves veterinary researchers who need a procedural reference for matching, weighting, and stratification, and it assumes familiarity with regression modeling and epidemiologic study design. The methods described apply across species, from companion animal clinical records to livestock production data, and the worked examples draw on published veterinary and comparative research. The article does not cover traditional multivariable outcome adjustment, which conditions on covariates inside the outcome model instead of balancing them in the design phase.

The central question this article answers is practical: given a veterinary dataset with measured confounders, how does a researcher construct a propensity score, verify that it performs adequately, and use it to estimate a causal effect with defensible uncertainty? The answer requires attention to covariate selection, model specification, balance diagnostics, and the choice among matching, weighting, and stratification. Each decision carries trade-offs in precision, bias reduction, and the population to which the estimated effect generalizes.

## At a Glance

| Parameter | Decision or fact |
|---|---|
| Purpose | Reduce confounding from measured covariates in observational comparisons |
| Core assumption | All confounders are measured and modelled correctly (no unmeasured confounding) |
| Score definition | Probability of receiving the exposure, conditional on observed covariates |
| Primary methods | Matching, inverse probability weighting, stratification |
| Balance metric | Standardized mean difference, commonly with a threshold of 0.10 |
| Common failure mode | Poor overlap in score distributions between exposure groups |
| Reporting requirement | Covariate balance table before and after adjustment |
| Software | Standard statistical packages with propensity score or weighting libraries |
| Complementary approach | Doubly robust estimation combining propensity scores with outcome regression |

## The Logic of the Propensity Score

The propensity score is the conditional probability of receiving the exposure given a set of observed covariates, typically estimated with logistic regression. Rosenbaum and Rubin demonstrated that if this score is correctly specified, then conditioning on the score alone balances the distribution of all included covariates between exposure groups. This property is what makes the score useful: instead of matching on many covariates simultaneously, which becomes impractical as the covariate set grows, the researcher matches on a single scalar summary.

The score does not remove confounding by itself. It reorganises the comparison so that exposed and unexposed animals with similar scores are exchangeable with respect to the measured covariates. Any remaining difference in outcomes can then be attributed to the exposure, provided the assumptions hold. The first assumption is that all confounders are measured and included in the score model. The second is that every animal has a nonzero probability of receiving either exposure, an assumption known as positivity or overlap. The third is that the outcome of one animal does not depend on the treatment assignment of another, the stable unit treatment value assumption.

These assumptions are strong. In veterinary datasets, unmeasured confounders such as subclinical disease status, owner compliance, or farm management practices can persist after propensity score adjustment. Sensitivity analyzes that probe how large an unmeasured confounder would need to be to overturn the findings are therefore advisable. The [CDC principles of epidemiology in public health practice](https://www.cdc.gov/csels/dsepd/ss1978/index.html) provide a useful framework for thinking about the sources of confounding and the assumptions embedded in observational comparisons.

## Constructing the Propensity Score Model

Covariate selection for the score model should be guided by the causal question, not by statistical significance. Variables that are known or suspected confounders, meaning they influence both the exposure and the outcome, belong in the model. Variables that are affected by the exposure, such as intermediate outcomes on the causal pathway, should be excluded because conditioning on them introduces collider bias. Instrumental variables, which predict the exposure but not the outcome except through the exposure, are also best excluded because they can increase variance without reducing bias.

The outcome itself must not be used when building the score model. This is a common error in retrospective analyzes where outcome data are already available. Peeking at the outcome when selecting covariates or specifying the model can inflate the apparent performance of the score and produce overfitted balance statistics. The score model should be built and evaluated with covariate data only.

Model specification matters. Logistic regression with main effects for all covariates is the default, but interactions and nonlinear terms may be needed to achieve balance. The researcher should check whether the model fits adequately and whether the predicted scores separate the exposure groups in a way that supports the chosen analytic method. In a study of PRRSV outbreaks in sow farms, for example, the investigators constructed a causal diagram before modeling to visualize the relationships among exposure, outcome, and potential confounders such as season, average parity at farrow, and sow farm Mycoplasma hyopneumoniae status, and then applied propensity score matching alongside regression and doubly robust methods to estimate the effect of epidemic status on nursery mortality. The causal diagram step is worth copying because it forces explicit decisions about which variables enter the model and why.

## Assessing Overlap and Balance

Before any matching or weighting, the researcher must examine the distribution of propensity scores in the exposed and unexposed groups. Poor overlap means that some animals have scores so extreme that no comparable counterpart exists in the other group. These animals contribute little information and can distort weighted estimates. Trimming the sample to the region of common support is one remedy, but it changes the estimand to the effect in the trimmed population, which should be reported clearly.

Balance is assessed by comparing covariate distributions between exposure groups after adjustment. The standardized mean difference, calculated as the difference in means divided by the pooled standard deviation, is the preferred metric because it does not depend on sample size. A threshold of 0.10 is commonly used to indicate acceptable balance, although this is a convention instead of a statistical law. Balance should be checked for every covariate in the score model, and for interactions and squared terms if they were included. In the canine hepatocellular carcinoma study, the investigators used propensity score matching to compare clinical characteriztics between affected dogs and a reference hospital population, allowing them to examine features such as thrombocyte counts while accounting for breed, age, and sex differences that would otherwise have confounded the comparison.

## Matching, Weighting, and Stratification: Choosing an Approach

The three principal uses of the propensity score each answer a slightly different causal question, and the choice among them should follow the structure of the data and the research objective.

**Propensity score matching** pairs each treated subject with one or more untreated subjects who have a similar estimated probability of treatment. Matching directly constructs a comparison sample with covariate distributions that approximate those of a randomised trial, which makes the analytic plan straightforward and the result easy to communicate. The principal cost is loss of data: unmatched subjects are discarded, and the effective sample size can shrink substantially. Matching is most appropriate when the treated group is small relative to the available pool of untreated subjects and when the region of common support is adequate. In the swine industry study of PRRSV outbreaks, propensity score matching allowed the authors to compare nursery mortality between epidemic and non-epidemic lots while balancing season, average parity at farrow, and sow farm *Mycoplasma hyopneumoniae* status ([PRRSV outbreak impact assessed with causal inference methods](https://pubmed.ncbi.nlm.nih.gov/40370830/)).

**Inverse probability of treatment weighting** (IPTW) assigns each subject a weight equal to the inverse of the probability of receiving the treatment actually received. The weighted sample has a pseudo-population in which treatment assignment is independent of the measured covariates. IPTW retains the full sample, which preserves statistical power, but it is sensitive to extreme weights when propensity scores approach zero or one. Weighting is often preferred when the exposure is common, when the sample is small, or when the goal is to estimate an average treatment effect in the total population instead of in a matched subset.

**Stratification** divides subjects into strata based on the propensity score, typically quintiles, and pools within-stratum treatment effects. This approach is simple and robust, but it controls confounding only coarsely. Residual confounding within strata is possible, particularly when the propensity score is strongly associated with the outcome. Stratification is best used as a sensitivity analysis or when the sample is too small for matching or weighting to be stable.

A doubly robust approach combines propensity score weighting with outcome regression. The estimate remains consistent if either the propensity score model or the outcome model is correctly specified, which provides protection against model misspecification. The PRRSV study used this strategy alongside matching, and the consistency of findings across approaches strengthened the causal interpretation ([PRRSV outbreak impact assessed with causal inference methods](https://pubmed.ncbi.nlm.nih.gov/40370830/)).

| Approach | Data retained | Primary strength | Primary limitation | Preferred setting |
|---|---|---|---|---|
| Matching | Treated and matched controls only | Direct construction of a balanced comparison group | Discards unmatched subjects | Small treated group, large untreated pool |
| IPTW | All subjects | Preserves power, estimates population average effect | Sensitive to extreme weights | Common exposure, full-sample inference desired |
| Stratification | All subjects | Simple, transparent | Coarse confounding control | Small samples, sensitivity analyzes |
| Doubly robust | All subjects | Consistent if either model is correct | Requires two correctly structured models | When outcome model is well understood |

## Step-by-Step Protocol for Propensity Score Analysis

**Step 1: Define the causal question and draw the causal diagram.** Specify the exposure, the outcome, and the hypothesised causal structure linking covariates to both. The diagram determines which variables enter the propensity score model. Include variables that influence treatment assignment and outcome, exclude instrumental variables that affect treatment but not outcome, and exclude colliders. The CDC epidemiology self-study course provides the foundational logic for distinguishing confounders from other covariate types ([CDC principles of epidemiology in public health practice](https://www.cdc.gov/csels/dsepd/ss1978/index.html)).

**Step 2: Select the analytic approach before examining the data.** Pre-specify whether matching, weighting, or stratification will be the primary analysis. This decision should rest on the sample size, the expected overlap, and the target estimand, not on which approach yields the most favourable result.

**Step 3: Estimate the propensity score.** Fit a logistic regression model with treatment as the outcome and the selected covariates as predictors. Consider including interaction terms or non-linear terms if the data support them. Assess model discrimination with the c-statistic, but do not use it as the sole criterion for model adequacy. A c-statistic near 0.5 can still accompany a valid propensity score if the covariates are weak predictors of treatment.

**Step 4: Assess overlap and balance.** Examine the distribution of propensity scores in the treated and untreated groups. Subjects with scores outside the common support region should be identified and their influence evaluated. After matching or weighting, compute standardized differences for every covariate. A standardized difference below 0.10 is commonly used as a threshold for acceptable balance, though this is a guideline instead of a statistical test.

**Step 5: Estimate the treatment effect.** In the matched or weighted sample, estimate the outcome difference between treatment groups. For time-to-event outcomes, use a Cox model in the matched sample or a weighted Cox model with robust standard errors. For binary outcomes, use a weighted logistic model or a conditional model appropriate to the matching structure.

**Step 6: Conduct sensitivity analyzes.** Repeat the analysis using a different propensity score implementation, such as switching from matching to weighting, or adding a doubly robust estimator. If the conclusions change materially across approaches, report this and interpret the findings cautiously. The consistency of results across methods in the PRRSV study, where regression, matching, and doubly robust approaches all indicated higher nursery mortality in epidemic lots, is the pattern that supports a robust causal claim ([PRRSV outbreak impact assessed with causal inference methods](https://pubmed.ncbi.nlm.nih.gov/40370830/)).

**Step 7: Report transparently.** Describe the propensity score model, the covariates included, the balance diagnostics, the number of subjects discarded, and the distribution of weights. This reporting standard allows readers to judge the validity of the analysis and to reproduce it.

## Diagnostics and Common Failure Modes

**Extreme weights** arise when treated subjects have very low propensity scores or untreated subjects have very high scores. These subjects exert disproportionate influence on the weighted estimate. Inspect the distribution of weights and consider truncating weights at a pre-specified percentile, commonly the 1st and 99th percentiles, or using stabilized weights.

**Poor overlap** occurs when the propensity score distributions of the two groups barely intersect. In this situation, the treated and untreated populations are so different that no analytic method can credibly estimate a treatment effect. Restricting the analysis to the region of common support is possible, but the estimand then changes to the effect in that restricted population.

**Residual imbalance** after matching or weighting indicates that the propensity score model was misspecified or that important confounders were omitted. Re-examine the model, consider adding interactions or non-linear terms, and verify that balance holds across all covariates, also those with the largest standardized differences.

**Unmeasured confounding** remains a threat in any observational study. Propensity score methods balance only measured covariates. The veterinary literature includes examples where propensity score matching revealed associations that were not apparent in conventional analyzes, such as the identification of breed predispositions and concurrent disorders in canine hepatocellular carcinoma ([epidemiology of massive hepatocellular carcinoma in dogs](https://pubmed.ncbi.nlm.nih.gov/31113567/)). However, the method cannot address confounders that were never recorded.

## Species and Setting Considerations

The choice of covariates for the propensity score model depends on the species and production system. In food animal research, herd-level factors such as season, parity distribution, and co-infection status are often the dominant confounders, as illustrated by the PRRSV study's inclusion of these variables ([PRRSV outbreak impact assessed with causal inference methods](https://pubmed.ncbi.nlm.nih.gov/40370830/)). In companion animal studies, breed, age, and concurrent disease are typically central, as demonstrated in the canine HCC analysis ([epidemiology of massive hepatocellular carcinoma in dogs](https://pubmed.ncbi.nlm.nih.gov/31113567/)).

Electronic medical record systems in veterinary teaching hospitals and large production systems increasingly support the covariate extraction needed for propensity score analysis. When such data are unavailable, the feasibility of the approach diminishes. Studies relying on manually abstracted records may have too few subjects or too many missing covariates to support matching or weighting. In those settings, stratification on a small number of key confounders may be the only practical option.

The target population also matters. A propensity score estimated in one production system or geographic region may not transport to another. The score is a function of the covariate distribution in the study sample, and the estimated treatment effect is conditional on that distribution. Generalizing findings beyond the study population requires explicit assumptions about transportability, which the [WOAH terrestrial animal health standards](https://www.woah.org/en/what-we-do/standards/codes-and-manuals/terrestrial-code-online-access/) implicitly recognize in their emphasis on surveillance and reporting frameworks that account for local disease ecology.

## Recognized Complications and Early Detection

Propensity score methods fail in characteriztic patterns, and each pattern produces a detectable signature before it corrupts the final estimate.

**Unmeasured confounding** remains the most consequential limitation. The propensity score balances only covariates that were measured and included in the score model. A confounder that was never recorded, such as farm-level biosecurity practices or clinician prescribing preference, remains unbalanced regardless of how precisely the score is estimated. Early detection relies on sensitivity analysis: repeat the primary analysis with a plausible unmeasured confounder added to the model and observe whether the treatment effect estimate changes materially. If it does, the original result cannot be interpreted causally.

**Extreme or near-zero propensity scores** indicate poor overlap between treatment groups. When a treated subject has a score of 0.98 and an untreated subject has a score of 0.02, matching forces comparisons between clinically incomparable animals. The standard diagnostic is a histogram or density plot of the score distribution by treatment group. When the distributions do not overlap substantially, trimming the sample to the region of common support is appropriate, but the generalizability of the resulting estimate narrows accordingly.

**Post-matching imbalance** can persist even when the score model appears adequate. The correct check is standardized mean difference for each covariate, not the p-value from a paired test. A standardized difference below 0.10 for every covariate is the conventional threshold, though some authors accept 0.20. When imbalance persists, the score model needs additional terms, including interactions or nonlinear specifications.

**Weighting instability** arises in inverse probability weighting when a small number of subjects receive very large weights. The diagnostic is the distribution of weights themselves: a weight exceeding 10 or a few weights dominating the analysis signals instability. Truncating weights at a prespecified percentile, commonly the 1st and 99th, stabilizes the estimate at the cost of introducing mild bias.

| Observation | Likely cause | Discriminating check |
|---|---|---|
| Large treatment effect appears after matching but not in crude analysis | Unmeasured confounding or misspecified score model | Sensitivity analysis for unmeasured confounding, inspect covariate balance |
| Few matched pairs available despite large sample | Poor overlap, extreme scores | Plot score distributions, examine common support region |
| Weighted analysis differs sharply from matched analysis | Weight instability or model misspecification | Compare weight distribution, check for extreme weights |
| Balance improves for some covariates but worsens for others | Score model omitted interactions or nonlinear terms | Re-specify model, re-check standardized mean differences |
| Effect estimate changes when one covariate is removed | Collinearity or overfitting in score model | Examine variance inflation, simplify model |

## Common Errors and Corrective Action

Less experienced analysts frequently include the outcome in the propensity score model. This is a structural error: the score must be estimated from covariates only, and including outcome information induces bias by conditioning on a post-treatment variable. The corrective action is to verify that the outcome variable is absent from the score equation before estimation.

Another frequent error is treating the propensity score as a substitute for careful covariate selection. The score model should include all measured confounders and predictors of the outcome, but indiscriminate inclusion of dozens of covariates, particularly those measured after treatment initiation, can introduce bias. The corrective action is to construct a causal diagram before analysis and include only variables that are confounders or outcome predictors, excluding mediators and colliders.

A third error is failing to assess balance after matching, relying instead on the assumption that a good score model produces balance automatically. Balance is an empirical property of the matched sample, not a mathematical consequence of the score. The corrective action is routine post-matching balance assessment using standardized mean differences.

Finally, analysts sometimes interpret the propensity score as a measure of treatment probability in the target population. The score is an artefact of the study sample and the covariates chosen. It does not estimate the true probability of treatment assignment in any external population.

## Limitations of Current Evidence

The veterinary literature applying propensity score methods remains sparse, and most published examples come from companion animal oncology or production medicine. The study of canine hepatocellular carcinoma used propensity score matching to compare clinical characteriztics between affected dogs and a hospital reference population, demonstrating the feasibility of the approach in a single teaching hospital setting [canine hepatocellular carcinoma epidemiology study](https://pubmed.ncbi.nlm.nih.gov/31113567/). In swine production research, propensity score matching and doubly robust methods have been applied to estimate the effect of PRRSV outbreaks on nursery mortality, with the doubly robust approach combining outcome regression with propensity weighting to provide protection against misspecification of either component [PRRSV outbreak causal inference study](https://pubmed.ncbi.nlm.nih.gov/40370830/).

Expert opinion still differs on several points. Whether matching or weighting should be preferred in small veterinary samples remains contested. Matching discards unmatched subjects and reduces sample size, while weighting retains all subjects but can produce unstable estimates. Some authors argue that weighting is more efficient, others that matching is more robust to model misspecification. No consensus has emerged, and the choice should be guided by the specific data structure instead of habit.

The evidence base is also limited by the absence of veterinary-specific simulation studies. Most methodological guidance derives from human epidemiology, where sample sizes are larger and covariate structures differ. The performance of propensity score methods in veterinary datasets with small treatment groups, clustered data from multiple farms or hospitals, and hierarchical structures has not been systematically evaluated.

## Referral, Consultation, and Reporting

When the analysis produces unstable weights, persistent imbalance, or results that change substantially under sensitivity analysis, consultation with a veterinary epidemiologist or biostatistician is warranted before the findings are reported. This is particularly important when the study informs clinical recommendations or regulatory decisions.

Regulatory reporting obligations arise when the study involves notifiable diseases. The World Organization for Animal Health maintains surveillance standards and reporting frameworks that define which diseases must be reported to national authorities [WOAH animal health surveillance standards](https://www.woah.org/en/what-we-do/animal-health-and-welfare/disease-data-collection/). Investigators should confirm their reporting obligations under the relevant national legislation before commencing a study that involves potentially notifiable pathogens. The WOAH Terrestrial Animal Health Code provides the international reference for disease notification and trade-related health measures [WOAH terrestrial animal health code](https://www.woah.org/en/what-we-do/standards/codes-and-manuals/terrestrial-code-online-access/).

Laboratory involvement is indicated when diagnostic misclassification could bias the propensity score or the outcome. If the exposure or outcome relies on laboratory-confirmed diagnoses, the laboratory should be engaged early to confirm test performance characteriztics and to ensure that diagnostic methods are consistent across study sites and time periods.

## Frequently Asked Questions

### How many animals do I need for propensity score matching in a veterinary study?

There is no universal minimum, but the limiting factor is covariate overlap, not total sample size. Matching requires a sufficient pool of untreated animals to serve as controls for treated animals across the propensity score distribution. A common practical benchmark is at least 5 to 10 untreated animals per treated animal before matching, though fewer may suffice when the score distributions overlap well. Examine the propensity score histograms before matching. If the treated group has scores in regions where no untreated animals exist, those animals cannot be matched and are discarded, which reduces effective sample size and generalizability. Report the number of matched pairs and the proportion of treated animals retained after matching.

### Can I use propensity score methods when my dataset has missing covariate values?

Missing data must be handled before propensity score estimation, because the score model requires complete covariate vectors. Multiple imputation is the preferred approach for veterinary datasets with moderate missingness, provided the missing-at-random assumption is plausible. Complete-case analysis is acceptable when missingness is low, typically under 5%, and when missingness is unrelated to both exposure and outcome. Avoid dropping covariates with missing values from the score model, as this reintroduces confounding. After imputation, estimate the propensity score within each imputed dataset and combine the final effect estimates using Rubin's rules. Sensitivity analyzes comparing complete-case and imputed results help assess whether missingness influenced the conclusions.

### What should I do when the treatment and control groups have very different sizes?

Severe imbalance in group sizes is common in veterinary studies, particularly when the exposure is rare. Propensity score weighting handles unequal group sizes more efficiently than matching, because matching discards unmatched controls. Inverse probability of treatment weighting retains all animals and reweights each observation to create a pseudo-population with balanced covariate distributions. When the treated group is small, consider caliper matching with replacement, which allows a single control to serve multiple treated animals. This reduces bias but requires adjustment for the repeated use of controls in variance estimation. Alternatively, use weighting with stabilized weights and trim extreme weights to reduce variance inflation from near-zero propensity scores.

### How do I report propensity score analyzes in a veterinary journal manuscript?

Follow reporting guidelines for observational studies and describe the propensity score model explicitly. State which covariates entered the score model and how they were measured. Report the matching ratio, caliper width, and whether matching was performed with or without replacement. Present balance diagnostics before and after adjustment, including standardized mean differences for every covariate. Report the number of animals excluded due to lack of overlap. Describe the outcome model used after propensity score adjustment, and present both unadjusted and adjusted effect estimates. Acknowledge that propensity scores balance measured covariates only, and note that unmeasured confounding remains possible. The [CDC principles of epidemiology](https://www.cdc.gov/csels/dsepd/ss1978/index.html) provide useful background on confounding and study design considerations that strengthen the methods section.

### How do I explain propensity score results to a referring veterinarian or practice owner?

Frame the explanation around the clinical question, not the statistical machinery. State that the analysis compared animals that were similar in age, breed, disease severity, and other relevant factors, so the observed difference in outcomes is more likely attributable to the treatment or exposure itself. Use absolute measures when possible, such as the difference in mortality percentage or the number needed to treat, instead of hazard ratios alone. Acknowledge the limitations honestly, including that the study was observational and cannot prove causation. For practice-level decisions, connect the findings to the specific patient population the practice serves. The [AVMA practice resources](https://www.avma.org/resources-tools) offer guidance on translating clinical evidence into practice decisions.

### When should I consult a biostatistician instead of proceeding alone?

Consult a biostatistician before data collection if you anticipate complex confounding, clustered data structures, or time-varying exposures. Propensity score methods assume correct specification of the score model, and misspecification can introduce bias that is difficult to detect after the fact. Seek expert input when your exposure is rare, when you have many covariates relative to the number of events, or when you plan to use doubly robust methods. A biostatistician can also help with complex variance estimation after matching with replacement or weighting. For straightforward comparisons with a modest covariate set and adequate sample sizes, standard software implementations are accessible, but peer review will scrutinise methodological choices. The [MSD Veterinary Manual](https://www.msdvetmanual.com/) and [WOAH animal health surveillance standards](https://www.woah.org/en/what-we-do/animal-health-and-welfare/disease-data-collection/) provide context on study design expectations in clinical and regulatory settings.

## Related Clinical & Scientific Guides

* [Evaluating Veterinary Surveillance System Attributes](/knowledge/veterinary-medicine/veterinary-epidemiology/evaluating-veterinary-surveillance-system-attributes)
* [Network Analysis for Infectious Disease Spread in Animal Populations](/knowledge/veterinary-medicine/veterinary-epidemiology/network-analysis-infectious-disease-spread-animal-populations)
* [Randomized Controlled Trials in Veterinary Field Settings](/knowledge/veterinary-medicine/veterinary-epidemiology/randomized-controlled-trials-veterinary-field-settings)


## References and Further Reading

- [Measuring the impact of sow farm outbreaks with PRRS virus on the downstream mortality using causal inference methods.](https://pubmed.ncbi.nlm.nih.gov/40370830/). 2025.
- [Cross-sectional analysis of dyslipidemia risk in coal mine workers: from epidemiology to animal models.](https://pubmed.ncbi.nlm.nih.gov/39505893/). 2024.
- [Epidemiology of massive hepatocellular carcinoma in dogs: A 4-year retrospective study.](https://pubmed.ncbi.nlm.nih.gov/31113567/). 2019.
- [Effectiveness and safety of rivaroxaban compared with low-molecular-weight heparin in cancer-associated thromboembolism.](https://pubmed.ncbi.nlm.nih.gov/32853381/). 2020.
- [WOAH Animal Health Surveillance Standards](https://www.woah.org/en/what-we-do/animal-health-and-welfare/disease-data-collection/). WOAH.
- [CDC Principles of Epidemiology in Public Health Practice](https://www.cdc.gov/csels/dsepd/ss1978/index.html). CDC.
- [MSD Veterinary Manual, Professional Edition](https://www.msdvetmanual.com/). MSD Veterinary Manual.
- [American Veterinary Medical Association Practice Resources](https://www.avma.org/resources-tools). American Veterinary Medical Association.
- [WOAH Terrestrial Animal Health Code](https://www.woah.org/en/what-we-do/standards/codes-and-manuals/terrestrial-code-online-access/). WOAH.

## Related Articles

- [Cluster Sampling in Veterinary Field Studies](/knowledge/veterinary-medicine/veterinary-epidemiology/cluster-sampling-veterinary-field-studies)
- [Confounding in Veterinary Studies: Identification and Control](/knowledge/veterinary-medicine/veterinary-epidemiology/confounding-veterinary-studies-identification-control)
- [Understanding Bias in Veterinary Epidemiological Studies](/knowledge/veterinary-medicine/veterinary-epidemiology/understanding-bias-veterinary-epidemiological-studies)
- [Understanding Ecological Studies in Veterinary Epidemiology](/knowledge/veterinary-medicine/veterinary-epidemiology/understanding-ecological-studies-veterinary-epidemiology)
- [Bayesian Methods for Diagnostic Test Evaluation in Animals](/knowledge/veterinary-medicine/veterinary-epidemiology/bayesian-methods-diagnostic-test-evaluation-animals)

> 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.