# Regression Analysis in Veterinary Epidemiology: Logistic and Poisson Models


## Key Takeaways

- Logistic regression is essential for binary veterinary outcomes (e.g., infected/not infected, presence/absence of lesions) and models the log odds, yielding odds ratios. This is particularly relevant for case-control studies and cross-sectional surveys where direct risk estimation is challenging.
- Poisson regression is the model of choice for count data in veterinary epidemiology (e.g., number of cases per herd per month, mortality rates) and models the log of the expected count, producing incidence rate ratios when an offset (like animal-time at risk) is included.
- Overdispersion, where the variance exceeds the mean in count data, is a common failure mode in veterinary epidemiology (e.g., clustered disease events within herds) and necessitates using negative binomial or quasi-Poisson models to avoid underestimated standard errors.
- Clustering and repeated measures in veterinary data (e.g., multiple animals from the same herd, serial diagnostic test results from an individual) violate independence assumptions, requiring mixed-effects models or generalized estimating equations to produce valid inference.
- Model selection must align with the outcome variable's structure (binary vs. count) and study design, with careful consideration of potential confounders, interactions (e.g., synergistic effects of virulence genes in *E. coli*), and model diagnostics like Hosmer-Lemeshow tests or deviance-to-degrees-of-freedom ratios.
- Common analytical errors include treating continuous predictors as categorical without justification, misinterpreting odds ratios as risk ratios for frequent outcomes, and ignoring clustering, all of which can lead to biased estimates and erroneous conclusions regarding risk factors for diseases like bovine tuberculosis or avian influenza.

---

Veterinary epidemiological data frequently present outcomes that are binary, such as infected or not infected, or counted, such as cases per herd per month. Linear regression, which assumes a continuous, normally distributed outcome, is inappropriate for these data structures. Logistic and Poisson regression models extend the generalized linear model framework to accommodate binary and count outcomes, respectively, while allowing adjustment for multiple predictors and estimation of effect sizes that are directly interpretable in clinical and population contexts.

This article serves veterinary researchers and graduate students who design observational studies, analyze surveillance data, or appraise the epidemiological literature. It explains when each model is appropriate, how to specify and interpret the models, and how to recognize common analytical errors. The focus is on model selection, parameter interpretation, and reporting standards. Survival analysis, which addresses time-to-event outcomes, is outside the scope of this article.

The analytical logic presented here follows the principles taught in standard epidemiological training, including the framework used by the [CDC Principles of Epidemiology in Public Health Practice](https://www.cdc.gov/csels/dsepd/ss1978/index.html). Those principles emphasize that the choice of statistical model must follow the structure of the outcome variable and the study design, not the convenience of the analyst.

## At a Glance

| Parameter | Logistic Regression | Poisson Regression |
| --- | --- | --- |
| Outcome type | Binary (event or no event) | Count (events per unit time or space) |
| Link function | Logit | Log |
| Measure of association | Odds ratio | Incidence rate ratio |
| Underlying distribution | Binomial | Poisson |
| Key assumption | Linearity in the logit | Mean equals variance |
| Common veterinary application | Case-control studies, cross-sectional surveys | Outbreak counts, mortality rates, incidence density |
| Major failure mode | Separation, overfitting, ignored clustering | Overdispersion, zero inflation |
| Alternative when assumptions fail | Exact logistic regression, Firth correction | Negative binomial, zero-inflated Poisson |

## The Generalized Linear Model Framework

Logistic and Poisson regression are both members of the generalized linear model family. A generalized linear model has three components: a random component that specifies the probability distribution of the outcome, a systematic component that is a linear combination of predictors, and a link function that connects the two. This structure allows the analyst to model outcomes whose distributions are not normal while retaining the familiar machinery of linear predictors and maximum likelihood estimation.

The choice between logistic and Poisson regression is driven by the outcome variable. If the outcome is binary, each observation falls into one of two categories and the binomial distribution applies. If the outcome is a count, the number of events in a fixed interval or region follows a Poisson distribution under specified conditions. Applying linear regression to either type of outcome can produce predicted values outside the permissible range, such as probabilities below zero or negative counts, and violates the distributional assumptions that underpin hypothesis testing.

## Logistic Regression for Binary Outcomes

Logistic regression models the log odds of an event as a linear function of predictors. The odds ratio is the exponentiated coefficient and represents the multiplicative change in odds for a one-unit increase in the predictor, holding other variables constant. This measure is particularly useful in case-control studies, where the odds ratio approximates the relative risk when the outcome is rare.

The model requires that observations be independent conditional on the predictors. This assumption is frequently violated in veterinary data when multiple animals are sampled from the same herd or when repeated measurements are taken from the same individual. Pooling observations without accounting for this clustering can produce misleading coefficient estimates and standard errors. A review of resource selection studies demonstrated that categorical and continuous variables changed in statistical significance and even in coefficient sign depending on whether a random effect for individual animals was included [Gillies et al., random effects in resource selection studies](https://pubmed.ncbi.nlm.nih.gov/17009752/). Veterinary researchers analyzing herd-level data should therefore consider mixed-effects logistic regression or generalized estimating equations when clustering is present.

### Model Building and Variable Selection

Multivariable logistic regression allows simultaneous adjustment for multiple risk factors. A study of Shiga toxin-producing *Escherichia coli* isolates illustrates the value of this approach. Univariate analysis showed significant associations between the enterohemorrhagic *E. coli* hemolysin gene and severe disease, but this association disappeared in the multivariate model because of a strong association with the intimin gene [Boerlin et al., virulence factor associations in Shiga toxin-producing E. coli](https://pubmed.ncbi.nlm.nih.gov/9986802/). The multivariate analysis also revealed a statistical interaction between the intimin and Shiga toxin 2 genes, supporting a biological hypothesis of synergism. This example demonstrates that univariate screening alone can misidentify risk factors and that interaction terms should be examined when biological plausibility supports them.

### Interpretation and Reporting

The odds ratio from a logistic regression model is conditional on all other variables in the model. This conditional interpretation differs from the crude odds ratio obtained in univariate analysis. A study of allergic sensitization in Austrian children found that the odds ratio for farming and allergic sensitization changed from 0.48 to 0.75 after adjusting for regular contact with livestock and poultry, indicating that the protective effect of farming was largely mediated by animal contact [Riedler et al., farming and allergic sensitization in children](https://pubmed.ncbi.nlm.nih.gov/10651771/). Reporting both crude and adjusted estimates, with confidence intervals, allows readers to assess the impact of confounding.

## Poisson Regression for Count Outcomes

Poisson regression models the log of the expected count as a linear function of predictors. When an offset is included, typically the log of person-time or animal-time at risk, the model estimates incidence rate ratios. This makes Poisson regression the natural choice for analyzing mortality rates, disease incidence, or any count of events over a defined exposure period.

The Poisson distribution assumes that the mean equals the variance. Veterinary count data frequently violate this assumption because events cluster within herds or because individual animals differ in susceptibility. When overdispersion is present, standard errors are underestimated and p-values are too small. The analyst should check the ratio of deviance to degrees of freedom and consider a negative binomial model or a quasi-Poisson model when overdispersion is detected.

### Offsets and Rate Estimation

An offset is a predictor whose coefficient is fixed at one instead of estimated. Including the log of the time at risk as an offset converts a count model into a rate model. This is essential when observation periods differ across herds or individuals. Failure to include an offset when exposure time varies will bias the estimated effects of all predictors.

## Model Diagnostics and Validation

Both logistic and Poisson models require assessment of fit and influence. For logistic regression, the Hosmer-Lemeshow test and area under the receiver operating characteriztic curve provide complementary information about calibration and discrimination. For Poisson models, the deviance and Pearson chi-square statistics indicate whether the model adequately describes the dispersion in the data.

Separation occurs in logistic regression when a predictor perfectly predicts the outcome, producing infinite coefficient estimates. This is common in small veterinary datasets with rare events. Firth's penalised likelihood or exact logistic regression provides stable estimates in these situations. For count data, zero inflation occurs when the number of zero observations exceeds what the Poisson distribution predicts, requiring a zero-inflated model that distinguishes structural zeros from sampling zeros.

## Applied Workflow: From Research Question to Regression Model

The practical value of logistic and Poisson regression emerges when the analytical framework is matched to the study design and the data structure at hand. A structured workflow reduces the risk of model misspecification and improves the interpretability of results for veterinary audiences.

### Step 1: Define the Outcome and Select the Model Family

The first decision is whether the outcome is binary or a count. Binary outcomes include disease present or absent, survival to discharge, or detection of a pathogen in a sample. Count outcomes include the number of clinical episodes per animal-year, the number of lesions per carcass, or the number of new cases per herd per month.

Logistic regression models the log odds of the outcome as a linear function of predictors. The exponentiated coefficients are odds ratios, which approximate relative risks only when the outcome is rare. Poisson regression models the log of the expected count as a linear function of predictors. The exponentiated coefficients are rate ratios when an offset is included, or incidence density ratios when person-time or animal-time at risk is accounted for.

Table 1 summarizes the decision points that distinguish the two approaches.

| Feature | Logistic Regression | Poisson Regression |
|---|---|---|
| Outcome type | Binary (0/1) | Count (non-negative integer) |
| Link function | Logit | Log |
| Parameter estimate | Odds ratio | Rate ratio or incidence density ratio |
| Baseline risk assumption | No assumption about outcome frequency | Assumes variance equals mean (may need adjustment) |
| Common veterinary applications | Case-control studies, cross-sectional surveys, clinical trials with binary endpoints | Longitudinal incidence studies, disease surveillance counts, parasite egg counts |
| Key diagnostic concern | Separation, sparse data bias | Overdispersion, excess zeros |

### Step 2: Account for Clustering and Repeated Measures

Veterinary data frequently arise from clustered structures. Animals are grouped within herds, herds within regions, and repeated measurements are taken from the same individual over time. Standard logistic and Poisson models assume independence of observations. When this assumption is violated, standard errors are underestimated and p-values are artificially small.

Random effects models extend the regression framework to accommodate clustering. In resource selection studies, for example, individual animals are monitored and their habitat use is compared with available habitat. Pooling observations across individuals without accounting for individual variation can change both the statistical significance and the direction of coefficient estimates, as demonstrated in a review of random-effects applications to resource selection by animals [Gillies et al., random effects in resource selection studies](https://pubmed.ncbi.nlm.nih.gov/17009752/). The same principle applies to repeated disease measurements from the same herd or animal.

When clustering is present, the analyst should fit a mixed-effects logistic or Poisson model with a random intercept for the cluster unit. The number of clusters matters. With fewer than approximately 10 to 20 clusters, random effects estimates can be unstable, and fixed effects for clusters or generalized estimating equations may be preferable. Species and production system influence this choice. In dairy herd health studies with hundreds of herds, random intercepts are standard. In wildlife studies with a small number of collared individuals, the analyst must be more cautious.

### Step 3: Build the Model With a Prespecified Strategy

Variable selection should follow a plan established before analysis begins. The plan should be based on causal reasoning, not purely on statistical criteria. Directed acyclic graphs help identify confounders, mediators, and colliders. Adjusting for a mediator removes part of the total effect. Adjusting for a collider introduces selection bias.

A common approach is to include variables with a priori biological plausibility, then use purposeful selection to retain variables that change the exposure coefficient by more than 10% or that meet a prespecified significance threshold. Automated stepwise procedures are discouraged because they produce coefficients that are biased upward and standard errors that are too small.

The number of events per variable is a critical constraint. For logistic regression, a minimum of 10 events per predictor variable is a widely cited rule, although simulation studies suggest that 5 to 9 events per variable may be acceptable in some settings. For Poisson regression, the equivalent constraint applies to the number of events relative to the number of parameters estimated. Sparse data produce unstable estimates and wide confidence intervals. When events are few, the analyst should consider penalised regression or exact methods.

### Step 4: Check for Interaction and Confounding

Interaction occurs when the effect of one predictor depends on the level of another. In a study of Shiga toxin-producing *Escherichia coli* virulence factors, multivariate logistic regression revealed an interaction between the intimin gene (*eae*) and Shiga toxin 2 (*stx2*), supporting a synergistic relationship between the adhesin and the toxin [Boerlin et al., virulence factor associations in STEC](https://pubmed.ncbi.nlm.nih.gov/9986802/). Without testing for interaction, the independent effects of each gene would have been misinterpreted.

Confounding is present when a third variable is associated with both the exposure and the outcome and is not on the causal pathway. In a cross-sectional study of farming and allergic sensitization in children, the crude odds ratio for farming was 0.48. After adjusting for genetic background, parental education, and housing conditions, the odds ratio changed only slightly. Including regular contact with livestock and poultry in the model changed the odds ratio to 0.75, indicating that animal contact mediated much of the protective association [Riedler et al., farming and allergic sensitization in children](https://pubmed.ncbi.nlm.nih.gov/10651771/). This example illustrates why both confounding and mediation must be considered explicitly.

### Step 5: Assess Model Fit and Calibration

For logistic regression, the Hosmer-Lemeshow test compares observed and predicted event rates across deciles of predicted risk. A non-significant result suggests adequate calibration, but the test has low power in small samples. The area under the receiver operating characteriztic curve summarizes discrimination. Values above 0.8 indicate good discrimination, although the threshold depends on the clinical context.

For Poisson regression, the ratio of the deviance to its degrees of freedom provides a simple check for overdispersion. A ratio substantially greater than 1 indicates that the variance exceeds the mean. Overdispersion can arise from unmeasured heterogeneity, clustering, or an incorrect link function. Options include using a quasi-Poisson model, a negative binomial model, or adding a random effect. Zero-inflated models are appropriate when the data contain more zeros than a Poisson distribution would predict, such as parasite egg counts from animals that are resistant to infection.

### Case Study: Logistic Regression in a Cross-Sectional Survey

Consider a cross-sectional survey of 2,283 schoolchildren aged 8 to 10 years in rural Austria, designed to test whether living on a farm protects against allergic sensitization [Riedler et al., farming and allergic sensitization in children](https://pubmed.ncbi.nlm.nih.gov/10651771/). The outcome was a positive skin prick test to at least one of seven local allergens. The exposure was living on a farm.

The crude prevalence of sensitization was 18.8% in farm children and 32.7% in non-farm children. A univariate logistic regression gave a crude odds ratio of approximately 0.48. The analysis then proceeded through a sequence of models. Adjusting for genetic background, parental education, living conditions, and dietary factors did not materially change the odds ratio. Adding regular contact with livestock and poultry to the model changed the odds ratio to 0.75, with a confidence interval that included 1. This finding indicated that the apparent protective effect of farming was largely explained by animal contact.

The case study demonstrates the reporting standards expected of logistic regression in veterinary and One Health research. The authors reported crude and adjusted odds ratios, confidence intervals, and the specific variables included in each model. They also reported the change in the odds ratio when the mediating variable was added, which allowed readers to assess the causal interpretation.

### Case Study: Logistic Regression in a Clinical Cohort

A second example comes from a cohort of 216 consecutive patients with Parkinson's disease, in which the prevalence and risk factors of hallucinations were investigated [Fénelon et al., hallucinations in Parkinson's disease](https://pubmed.ncbi.nlm.nih.gov/10734005/). Hallucinations were present in 39.8% of patients in the previous three months. Logistic regression identified three factors independently associated with hallucinations. The analysis used non-parametric tests to compare groups first, then entered significant variables into the regression model.

This two-stage approach is common in clinical veterinary research. It reduces the number of candidate predictors before multivariable modeling, which is useful when the number of events is modest. The limitation is that variables excluded in univariate screening cannot be examined for interaction or confounding in the multivariable model. Variables with a univariate p-value above 0.20 are sometimes retained in the multivariable stage to avoid missing confounders, a practice that is acceptable when the number of predictors is small relative to the number of events.

### Species and Production System Considerations

The correct model choice depends on the species and the production system. In companion animal medicine, data are often collected from individual patients in a single clinic, and clustering by clinic or by veterinarian may be relevant. In production animal medicine, data are naturally hierarchical, with animals nested within herds and herds within regions. Ignoring this hierarchy in a Poisson model of disease incidence will produce confidence intervals that are too narrow and may lead to false declarations of significance.

In wildlife epidemiology, telemetry data from collared animals are autocorrelated in space and time. Logistic regression models of resource selection must account for individual variation in availability and selection, as random-effects approaches do [Gillies et al., random effects in resource selection studies](https://pubmed.ncbi.nlm.nih.gov/17009752/). The number of locations per animal and the number of animals both influence the stability of the estimates.

In laboratory animal research, experimental designs are typically balanced and clustering is less of a concern. However, repeated measurements from the same animal over time still require mixed-effects models or generalized estimating equations.

### Documentation and Reporting Standards

The final model should be reported with sufficient detail for replication. This includes the number of observations, the number of events, the number of predictors, the method of variable selection, the handling of missing data, and the model diagnostics performed. Odds ratios and rate ratios should be reported with confidence intervals, not p-values alone. The units of the offset in a Poisson model must be stated explicitly, for example animal-years at risk or herd-months.

International surveillance standards emphasize the importance of transparent and reproducible analytical methods in animal health reporting [WOAH animal health surveillance standards](https://www.woah.org/en/what-we-do/animal-health-and-welfare/disease-data-collection/). The same standards apply to research publications. A model that cannot be reproduced from the methods section is of limited value to the veterinary community.

## Common Failure Modes and Early Detection

Regression models in veterinary epidemiology fail in predictable ways. The most consequential failure is overfitting, where the model captures noise instead of signal. Overfitting is detected by comparing the number of events per predictor variable. For logistic regression, a minimum of 10 events per variable is a widely used threshold, though this guidance originates from human epidemiology and should be applied with caution in veterinary datasets where clustering is common. When the ratio falls below this threshold, the model will produce unstable coefficient estimates and optimiztic apparent performance.

Separation occurs when a predictor perfectly distinguishes outcomes, producing absurdly large odds ratios with wide confidence intervals. This is common in small veterinary datasets with rare outcomes. The discriminating check is inspection of coefficient magnitudes and standard errors. A coefficient exceeding 10 with a standard error above 3 warrants suspicion.

Convergence failure appears as warning messages from the fitting algorithm. It often results from collinearity between predictors, sparse categories, or an excessively complex model relative to sample size. Early detection requires routine examination of convergence diagnostics and correlation matrices before interpreting any output.

## Common Errors and Corrective Actions

Less experienced analysts frequently treat continuous predictors as categorical without justification. This discards information and produces models that are difficult to interpret across studies. The corrective action is to fit continuous predictors as linear terms first, then test for nonlinearity using restricted cubic splines or fractional polynomials.

Another recurring error is the inclusion of variables selected by univariate screening at an arbitrary threshold such as P < 0.20. This practice can exclude important confounders and include noise variables. A prespecified model-building strategy based on causal reasoning is preferable. The [CDC principles of epidemiology](https://www.cdc.gov/csels/dsepd/ss1978/index.html) emphasize that variable selection should follow the study design and causal framework, not purely statistical criteria.

Misinterpretation of odds ratios as risk ratios is common when outcomes are frequent. In cohort studies with outcome prevalence above 10 percent, odds ratios overstate the association. Poisson regression with robust variance or log-binomial models should be considered for such data.

Ignoring clustering produces standard errors that are too small and P values that are too significant. The [random effects approach to resource selection](https://pubmed.ncbi.nlm.nih.gov/17009752/) demonstrates how pooled analyzes that ignore individual variation can change both the sign and significance of coefficients. When animals are sampled repeatedly, mixed-effects models or generalized estimating equations are required.

## Troubleshooting Table

| Observation | Likely Cause | Discriminating Check |
|---|---|---|
| Huge odds ratio with wide CI | Separation or sparse data | Examine cross-tabulation of predictor and outcome |
| Model fails to converge | Collinearity or sparse categories | Check variance inflation factors and category counts |
| Coefficients change sign between univariate and multivariable models | Confounding or suppression | Compare models with and without the suspected confounder |
| Apparent good fit but poor validation | Overfitting | Use bootstrap or cross-validation to compare optimizm |
| Standard errors too small | Clustering ignored | Fit random intercept model and compare inference |

## Limitations of Current Evidence

The veterinary literature on regression modeling is uneven. Much of the methodological guidance derives from human epidemiology and wildlife ecology. The [WOAH animal health surveillance standards](https://www.woah.org/en/what-we-do/animal-health-and-welfare/disease-data-collection/) provide frameworks for data quality and reporting, but they do not prescribe specific analytical methods. Expert opinion still differs on the handling of herd-level clustering when herd sizes vary widely, on the choice between fixed and random effects for small numbers of clusters, and on the minimum sample size for reliable estimation in Poisson models.

The [relationship between chest compression rate and outcome](https://pubmed.ncbi.nlm.nih.gov/22623717/) illustrates a further limitation: many veterinary datasets are observational and subject to residual confounding that no amount of adjustment can remove. Causal claims require either experimental designs or explicit causal modeling frameworks.

## Referral and Escalation

Most regression analyzes can be completed by a veterinarian with graduate training in epidemiology. Referral to a statistical consultant is warranted when the data structure is complex, such as hierarchical sampling with multiple levels of clustering, when outcomes are rare and separation occurs, or when the analysis informs regulatory submissions or publication in high-impact journals.

Laboratory involvement is indicated when diagnostic test results are used as predictors or outcomes and test sensitivity or specificity varies across subgroups. Misclassification bias can distort regression coefficients, and laboratory scientists can provide the test performance data needed for correction.

Regulatory reporting obligations arise when regression analyzes identify notifiable disease signals. The [WOAH terrestrial animal health code](https://www.woah.org/en/what-we-do/standards/codes-and-manuals/terrestrial-code-online-access/) specifies reporting requirements for listed diseases, and national authorities may have additional obligations. When a model reveals an unexpected cluster of cases or an elevated risk in a defined population, the responsible action is to contact the relevant animal health authority before publication or further dissemination.

## Frequently Asked Questions

**How many events per predictor variable do I need for a reliable logistic regression model?**

A common rule of thumb is at least 10 events per predictor variable, where events are the less frequent outcome category. This threshold helps prevent overfitting and unstable coefficient estimates. In veterinary datasets with clustered sampling, such as multiple animals from one herd, the effective sample size is smaller than the raw count. The random-effects literature demonstrates that ignoring clustering can change both coefficient sign and statistical significance, so account for grouping before finalising your sample size. When events are scarce, consider penalised likelihood approaches or reduce the predictor set using prespecified hypotheses instead of data-driven selection.

**My count data show more zeros than a Poisson model predicts. What are my options?**

Excess zeros commonly arise when measuring parasite egg counts, lesion counts, or clinical episodes across a population where many animals are never at risk. Zero-inflated Poisson or negative binomial models treat zeros as arising from two processes: a structural zero group and a count process. Alternatively, hurdle models model the probability of any event separately from the count given that at least one event occurred. Compare these models using AIC or likelihood ratio tests where nested. The choice depends on whether your zeros represent animals that cannot experience the event or animals that simply did not during the observation window. Consult the [CDC principles of epidemiology](https://www.cdc.gov/csels/dsepd/ss1978/index.html) for guidance on defining at-risk populations.

**When should I use a negative binomial model instead of Poisson regression?**

Poisson regression assumes the variance equals the mean. Veterinary count data frequently show overdispersion, where the variance exceeds the mean due to unmeasured heterogeneity between animals or herds. Fit a negative binomial model when the Pearson chi-square divided by degrees of freedom exceeds 1.5 or when a likelihood ratio test comparing the models is significant. Overdispersion that remains after adding relevant predictors suggests missing covariates or clustering. The negative binomial adds a dispersion parameter that accommodates this extra variation. In production medicine, somatic cell counts, lameness episodes, and disease recurrence data routinely require this extension. Report the dispersion parameter estimate so readers can assess whether the model adequately captured the heterogeneity.

**How do I handle data from multiple herds or farms in a single regression model?**

Observations from the same herd are correlated, violating the independence assumption of standard logistic and Poisson models. Options include generalized estimating equations for population-averaged effects, random intercept models for herd-specific effects, or including herd as a fixed effect when the number of herds is small. Random-effects models allow inference to the broader population of herds, whereas fixed effects restrict inference to the herds studied. The resource selection literature shows that ignoring random effects can reverse the direction of associations, so this is not a minor technicality. Use likelihood ratio tests to determine whether the random effect significantly improves fit. For small numbers of herds, consider whether clustering is better addressed through robust standard errors.

**What should I report in the methods section so my regression analysis is reproducible?**

State the model family, link function, and software with version. List all candidate predictors and the prespecified selection strategy, including how continuous variables were handled, whether transformations were considered, and the threshold for retaining variables. Report how clustering was addressed and justify the choice of random or fixed effects. Describe the handling of missing data, including the proportion missing per variable and the imputation method if used. Provide the final model equation with coefficient estimates, standard errors, odds ratios or rate ratios with confidence intervals, and the dispersion parameter for count models. The [WOAH surveillance standards](https://www.woah.org/en/what-we-do/animal-health-and-welfare/disease-data-collection/) emphasize transparent reporting for surveillance data, and the same principle applies to regression analyzes intended for clinical or policy use.

**How do I explain regression results to a producer or practice owner who lacks statistical training?**

Translate odds ratios and rate ratios into absolute terms. An odds ratio of 2.5 for lameness with increasing body condition score is more meaningful when expressed as the predicted probability of lameness at each score level. Present a simple table showing predicted outcomes across the range of the key predictor, holding other variables at their mean. Use plain language: animals with condition score X were about twice as likely to be lame as animals with score Y. Avoid causal language unless the study design supports it. Explain that the model accounts for other factors simultaneously, which is why the result may differ from a simple comparison. The [MSD Veterinary Manual](https://www.msdvetmanual.com/) provides models for communicating clinical findings to lay audiences, and the same principles apply to quantitative results.

## 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

- [Application of random effects to the study of resource selection by animals.](https://pubmed.ncbi.nlm.nih.gov/17009752/). 2006.
- [Relationship between chest compression rates and outcomes from cardiac arrest.](https://pubmed.ncbi.nlm.nih.gov/22623717/). 2012.
- [Serum FGF21 levels are increased in obesity and are independently associated with the metabolic syndrome in humans.](https://pubmed.ncbi.nlm.nih.gov/18252893/). 2008.
- [Associations between virulence factors of Shiga toxin-producing Escherichia coli and disease in humans.](https://pubmed.ncbi.nlm.nih.gov/9986802/). 1999.
- [Austrian children living on a farm have less hay fever, asthma and allergic sensitization.](https://pubmed.ncbi.nlm.nih.gov/10651771/). 2000.
- [Hallucinations in Parkinson's disease: prevalence, phenomenology and risk factors.](https://pubmed.ncbi.nlm.nih.gov/10734005/). 2000.
- [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.

## Related Articles

- [Using Simulation Models in Veterinary Epidemiology](/knowledge/veterinary-medicine/veterinary-epidemiology/using-simulation-models-veterinary-epidemiology)
- [Sensitivity Analysis in Veterinary Disease Models](/knowledge/veterinary-medicine/veterinary-epidemiology/sensitivity-analysis-veterinary-disease-models)
- [Case-Control Studies in Veterinary Epidemiology: Selection and Analysis](/knowledge/veterinary-medicine/veterinary-epidemiology/case-control-studies-veterinary-epidemiology-selection-analysis)
- [Survival Analysis in Veterinary Medicine: Kaplan-Meier and Cox Regression](/knowledge/veterinary-medicine/veterinary-epidemiology/survival-analysis-veterinary-medicine-kaplan-meier-cox-regression)
- [Compartmental Models in Veterinary Disease Dynamics](/knowledge/veterinary-medicine/veterinary-epidemiology/compartmental-models-veterinary-disease-dynamics)

> 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.