Using Propensity Scores in Veterinary Observational Research

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

Key Takeaways

• Propensity scores are essential for mitigating confounding in veterinary observational research by estimating the probability of treatment receipt based on observed baseline covariates, enabling the construction of comparable treatment and control groups.
• Key assumptions for valid inference include no unmeasured confounding (strong ignorability), positivity (non-zero probability of receiving either treatment for all subjects), and the stable unit treatment value assumption (no interference between subjects).
• Analytic strategies include matching (pairs treated with untreated subjects), inverse probability of treatment weighting (assigns weights to create a pseudo-population), stratification (divides subjects into strata based on score), and covariate adjustment (uses score as a covariate in outcome regression).
• Balance assessment is critical, typically using standardized mean differences below 0.10-0.20 for all covariates, and requires checking for overlap in propensity score distributions between groups to ensure valid comparisons.
• Common failure modes include inadequate overlap in propensity score distributions, residual covariate imbalance after adjustment (indicating model misspecification), and unmeasured confounding, necessitating sensitivity analyses to quantify their potential impact.
• Reporting standards like STROBE-Vet or REFLECT are crucial for transparency, detailing propensity score model specification, covariate selection (excluding mediators), matching/weighting algorithms, and comprehensive balance diagnostics.

Veterinary clinical practice generates abundant observational data: electronic medical records, hospital referral populations, insurance claims, and national surveillance databases. These data capture treatments as they are actually delivered, across diverse species, breeds, ages, and clinical contexts. Yet observational comparisons of treatment effects are vulnerable to confounding, because the decision to administer a treatment is rarely random. Animals that receive a drug, undergo surgery, or receive nutritional support differ systematically from those that do not, and those differences, not the treatment itself, may drive observed outcome differences.

Propensity score methods offer a structured approach to reducing this confounding in veterinary observational research. A propensity score is the conditional probability of receiving a treatment given a set of observed baseline covariates. By balancing treated and untreated groups on this single summary score, researchers can construct comparison groups that approximate the exchangeability achieved by randomisation, at least with respect to measured confounders. This article explains the conceptual basis of propensity scores, describes the principal analytic strategies, and provides practical guidance for designing, executing, and reporting propensity score analyzes in veterinary studies. It is written for veterinary researchers and graduate students who design observational studies or critically appraise the veterinary and comparative literature.

The methods described here apply across species and clinical settings, from companion animal oncology to production animal medicine. The statistical principles are species-independent, although the practical challenges of data quality, sample size, and covariate measurement differ by context. The article does not cover randomised trial methodology, but the reporting standards discussed are relevant to any study that uses observational data to estimate causal effects.

At a Glance

The Confounding Problem in Veterinary Observational Data

Observational studies answer questions that trials cannot, often because randomisation is unethical, impractical, or prohibitively expensive. A veterinary clinician cannot randomise critically ill neonates to different feeding protocols, nor assign cancer patients to receive or withhold a drug on the basis of a coin flip. The published veterinary and comparative literature therefore relies heavily on observational designs. For example, population-based cohort studies have examined whether beta-blocker exposure is associated with cancer incidence and whether statin use after cancer diagnosis is associated with survival, both using administrative and registry data that carry substantial confounding risk.

The central threat in such designs is that treatment groups differ in prognosis before treatment begins. Consider a study of postoperative analgesia and cancer recurrence. Animals receiving epidural analgesia may be healthier, undergo different surgical procedures, or receive different adjuvant therapy than animals receiving systemic analgesia alone. Any observed association between epidural use and recurrence could reflect these baseline differences instead of a true analgesic effect. The same logic applies in reverse: sicker animals may be preferentially given a treatment, producing confounding by indication that biases the treatment effect toward harm.

Randomisation solves this problem by making treatment assignment independent of all covariates, measured and unmeasured. Propensity score methods cannot achieve that guarantee. They balance only the covariates that the researcher measures and includes in the score. This limitation is fundamental and should govern the interpretation of every propensity score analysis. The methods reduce confounding by observed covariates, and the validity of the resulting estimate depends on the assumption that no important unmeasured confounders remain.

The Propensity Score and Its Theoretical Basis

The propensity score, defined by Rosenbaum and Rubin, is the probability that a subject receives the treatment of interest, conditional on observed covariates. In a veterinary study, this might be the probability that a dog with lymphoma receives a particular chemotherapy protocol given its age, breed, stage, histologic subtype, and comorbidity burden. The score collapses a potentially high-dimensional covariate vector into a single scalar between zero and one.

Three theoretical properties make the score useful. First, treated and untreated subjects with the same propensity score have, on average, the same distribution of the included covariates. This property, called balancing, means that conditioning on the score is equivalent to conditioning on all covariates simultaneously. Second, if treatment assignment is strongly ignorable, meaning that it is independent of potential outcomes given the observed covariates, then it is also independent of potential outcomes given the propensity score alone. Third, the score can be used in several analytic strategies, all of which exploit the balancing property to construct valid comparisons.

The propensity score is almost always unknown and must be estimated, typically with logistic regression of treatment assignment on covariates. The estimated score is then used in matching, weighting, stratification, or covariate adjustment. The quality of the estimated score depends on correct specification of the treatment model, including relevant interactions and nonlinear terms. Misspecification of this model can leave residual imbalance even when the score appears to balance the sample.

Assumptions Required for Valid Inference

Propensity score methods rest on three assumptions that should be stated explicitly in any veterinary manuscript. The first is no unmeasured confounding, also called strong ignorability. This assumption holds that all variables influencing both treatment assignment and outcome have been measured and included in the score. It cannot be tested directly, although sensitivity analyzes can assess how strongly an unmeasured confounder would need to be associated with treatment and outcome to overturn the findings.

The second assumption is positivity, meaning that every subject has a nonzero probability of receiving either treatment. In veterinary data, this assumption fails when a treatment is never given to certain subgroups, for example, if a drug is contraindicated in animals with renal disease and no such animal in the database ever received it. Positivity violations produce propensity scores of exactly zero or one, and the affected subjects cannot contribute to the comparison.

The third assumption is the stable unit treatment value assumption, which requires that one subject's treatment does not affect another subject's outcome. This assumption is violated in herd-level studies where treatments have indirect effects, such as vaccination programs that confer herd immunity. In such settings, the unit of analysis and the definition of treatment must be chosen carefully.

Analytic Strategies Using the Propensity Score

Four principal strategies use the estimated propensity score, and the choice among them depends on the research question, sample size, and the target population of inference.

Propensity score matching pairs each treated subject with one or more untreated subjects who have a similar score. The matched sample is then analyzed as if it came from a paired or block-randomised design. Matching is intuitive and produces a sample with directly comparable covariate distributions, but it discards unmatched subjects and can reduce the sample size substantially. The effect estimated in a matched sample is the average treatment effect on the treated, because untreated subjects are selected to resemble the treated group.

Inverse probability of treatment weighting assigns each subject a weight equal to the inverse of the probability of receiving the treatment actually received. Weighted analyzes create a pseudo-population in which treatment is independent of the measured covariates. This approach retains all subjects and can estimate either the average treatment effect or the average treatment effect on the treated, depending on the weighting scheme. Extreme weights, arising from propensity scores near zero or one, can inflate variance and bias estimates, and weight truncation is often necessary.

Stratification divides subjects into strata based on the propensity score, typically quintiles, and pools within-stratum treatment effects. This method is simple but coarse, and residual confounding can remain within strata. Covariate adjustment uses the propensity score as a single covariate in an outcome regression model. This approach is statistically efficient but relies heavily on correct model specification and can be sensitive to misspecification of the outcome model.

The choice of strategy should be made before analysis and justified in the study protocol. Sensitivity analyzes using more than one strategy can demonstrate robustness, as illustrated in comparative studies that report both multivariable adjustment and propensity score analyzes with similar conclusions.

Building the Propensity Score Model

The propensity score is the conditional probability of receiving the treatment given observed covariates. In practice, you estimate this probability with a logistic regression model in which treatment assignment is the outcome and the covariates are predictors. The model need not predict treatment perfectly. It must balance the covariate distributions between groups.

Begin by listing every baseline variable that plausibly influences both treatment assignment and outcome. In veterinary data, these often include age, breed, body weight, sex, comorbidity count, disease severity scores, clinic or hospital site, season, and prior medication use. Include variables that are known confounders from the literature even if they do not reach statistical significance in your sample. A variable that predicts outcome but not treatment still belongs in the model because it reduces residual outcome variation and improves precision.

Do not include variables that lie on the causal pathway between treatment and outcome. These are mediators, not confounders. Adjusting for them removes part of the treatment effect you intend to estimate. Likewise, do not include post-treatment variables unless you are explicitly estimating a controlled direct effect, a scenario that requires more advanced methods than this article covers.

Model specification deserves deliberate attention. Continuous covariates such as age or body weight rarely have linear log-odds relationships with treatment. Consider flexible specifications using restricted cubic splines or fractional polynomials. Include clinically plausible interaction terms, for example between age and comorbidity burden. The goal is a model that adequately describes the treatment assignment mechanism, not a parsimonious model. Overfitting is a lesser concern than residual imbalance.

After fitting the model, examine the distribution of estimated propensity scores in each treatment group. Substantial overlap is required for valid inference. If large regions of the score distribution contain only treated or only untreated subjects, the data cannot support causal comparisons in those regions. This condition, called the positivity assumption, fails when some patients have near-zero or near-one probabilities of receiving either treatment. In veterinary data, this occurs when a procedure is nearly universal in one referral population and nearly absent in another.

Matching, Weighting, and Stratification

Three analytic strategies use the estimated propensity score. Each has distinct properties that suit different data structures.

Propensity Score Matching

Matching pairs each treated subject with one or more untreated subjects who have similar propensity scores. The most common implementation is nearest-neighbour matching without replacement, often with a caliper width of 0.2 of the pooled standard deviation of the logit of the propensity score. This caliper value balances bias reduction against sample retention in most settings. Greedy matching processes subjects sequentially, while optimal matching minimizes the total distance across all pairs. Optimal matching is computationally more demanding but produces better balance in small samples.

After matching, you must assess covariate balance in the matched sample. The standardized mean difference, calculated as the difference in means divided by the pooled standard deviation, should be below 0.1 for every covariate. This threshold is a diagnostic convention, not a guarantee of validity. Examine the full covariate distribution, also means. Quantile-quantile plots and empirical cumulative distribution functions can reveal residual imbalance that means alone obscure.

Matching discards unmatched subjects. In veterinary datasets with modest sample sizes, this can reduce power substantially. The observational study of antifungal prophylaxis in critically ill patients illustrates the trade-off: with 132 patients and 75 treated, matching would retain only a fraction of the sample, which is why the authors used propensity score adjustment instead.

Inverse Probability of Treatment Weighting

Weighting assigns each subject a weight equal to the inverse of the probability of receiving the treatment they actually received. Treated subjects receive weight 1 divided by their propensity score. Untreated subjects receive weight 1 divided by 1 minus their propensity score. The weighted sample approximates a population in which treatment assignment is independent of the measured covariates.

Stabilized weights, which multiply the inverse probability weights by the marginal probability of treatment, reduce variance and are generally preferred. Trimming extreme weights, for example discarding subjects with weights above the 99th percentile, can improve precision at the cost of changing the estimand. The population-based cohort study of propranolol and cancer risk used propensity score matching in a large administrative database, where the ample sample size made matching feasible without excessive loss.

Weighting uses all subjects and preserves the full sample. It is particularly useful when the treatment groups are imbalanced in size. The cost is sensitivity to misspecification: subjects with very small propensity scores receive very large weights and can dominate the analysis. Check the distribution of weights before proceeding.

Stratification

Stratification divides subjects into strata, commonly quintiles of the propensity score, and pools treatment effect estimates across strata. This approach removes approximately 90 percent of the bias due to measured confounders when the score is well specified. It is simple and transparent but leaves residual within-stratum imbalance. Stratification is a reasonable choice when the sample is too small for matching and the propensity score distribution is not too extreme.

Diagnostic Checks After Estimation

Whatever strategy you choose, the same diagnostic sequence applies. First, verify that the propensity score model converged and that no covariate perfectly predicts treatment. Perfect prediction produces separation in logistic regression and unstable estimates. Second, examine the overlap of propensity score distributions between groups. A histogram or density plot overlaid by treatment group is the standard visual check. Third, compute standardized mean differences for every covariate before and after adjustment. The sensitivity analysis using propensity score methods in the breast cancer statin cohort demonstrates how these diagnostics support the primary analysis and reassure readers about robustness.

Fourth, assess balance in higher moments and interactions. A covariate may have similar means but different variances across groups. Include clinically important two-way interactions in the balance assessment even if they were not in the propensity score model. Fifth, re-estimate the treatment effect using an alternative propensity score strategy. If matching, weighting, and adjustment give materially different answers, the propensity score model is probably misspecified or the positivity assumption is violated.

Worked Example: Enteral Feeding and Necrotising Enterocolitis

The national surveillance study of necrotising enterocolitis in English neonatal units provides a useful template. The investigators used propensity score analysis to estimate the effect of feeding with own mother's milk, with or without donor milk, on the risk of severe necrotising enterocolitis in infants born before 32 weeks of gestation.

The treatment definition was exposure to bovine-origin formula or fortifier avoidance in the first 14 postnatal days. The propensity score model included gestational age, birth weight, sex, mode of delivery, and network-level factors. The outcome was severe necrotising enterocolitis confirmed by laparotomy or leading to death.

The analysis proceeded in stages. First, the propensity score model was fitted and overlap was confirmed. Second, the treatment effect was estimated in the propensity score-adjusted sample. Third, sensitivity analyzes examined the robustness of the result to unmeasured confounding. The study illustrates a key practical point: the propensity score does not eliminate confounding by unmeasured variables. It balances only the covariates you measured and modelled.

Reporting Standards

Observational studies using propensity scores must report their methods with sufficient detail for replication. The ARRIVE guidelines for animal research specify the minimum information required for transparent reporting, including sample size determination, randomisation or allocation procedures, and statistical methods. The EQUATOR Network reporting guidelines catalogue the relevant checklists, including STROBE for observational studies and REFLECT for livestock studies.

Your methods section should state the propensity score model specification, including all covariates and their functional forms. Report the matching algorithm, caliper width, and whether matching was with or without replacement. For weighting, report the weight type and any trimming. Report the balance diagnostics with numeric standardized mean differences before and after adjustment. State how many subjects were excluded at each stage and why.

Species differences matter in reporting. In companion animal studies, breed and body weight are often strong confounders because they influence both treatment choice and outcome. In production animal studies, herd-level factors such as management system and biosecurity status may dominate. In wildlife studies, capture method and season may drive both treatment and outcome. The MSD Veterinary Manual provides species-specific clinical context that helps identify plausible confounders, while AVMA practice resources and WOAH terrestrial animal health standards frame the regulatory and professional context in which the research will be interpreted.

Common Failure Modes

Propensity score analyzes fail in predictable ways. The most common is inadequate overlap. If the treated and untreated groups occupy nearly disjoint regions of covariate space, no adjustment method can produce credible estimates. The analysis should stop and the question should be reframed.

The second failure mode is residual imbalance after adjustment. This usually indicates a misspecified propensity score model. Add nonlinear terms, interactions, or additional covariates. If balance remains poor, consider whether the treatment groups are fundamentally incomparable.

The third failure mode is unmeasured confounding. Propensity scores address only measured confounders. A sensitivity analysis, such as the E-value or a Rosenbaum sensitivity analysis, quantifies how strong an unmeasured confounder would need to be to overturn the results. The follow-up study of neonatal morphine effects illustrates this concern: the apparent difference in intelligence quotient disappeared after propensity score adjustment, suggesting that confounding by clinical severity, not the morphine itself, drove the unadjusted finding.

The fourth failure mode is treating the propensity score as a magic solution. It is a balancing tool, not a causal inference machine. The validity of the estimate still rests on the assumptions of no unmeasured confounding, positivity, and correct model specification.

Recognized Complications and Early Detection

Propensity score methods fail in characteriztic ways, and most failures are detectable before outcome analysis begins. The first and most consequential failure is non-overlap of propensity score distributions between treatment groups. When treated and untreated animals occupy different regions of the score distribution, every analytic strategy relies on extrapolation into regions where no comparable animals exist. Early detection requires plotting the estimated score distributions for both groups and quantifying the region of common support. A caliper width of 0.2 times the pooled standard deviation of the logit of the propensity score is a common threshold, but the visual inspection of the overlap plot remains the primary diagnostic.

The second failure mode is residual imbalance after adjustment. Propensity score methods balance measured covariates only in expectation, and small samples frequently leave meaningful imbalance. The standardized mean difference, calculated for every covariate before and after adjustment, should fall below 0.1 for all variables. Values between 0.1 and 0.2 warrant scrutiny, and values above 0.2 indicate that the propensity score model omitted an important covariate or that the matching procedure was poorly executed. The veterinary literature contains examples where propensity score adjustment changed conclusions, such as the association between epidural analgesia and cancer recurrence after colorectal surgery, where the primary analysis and the propensity score sensitivity analysis produced similar results Gottschalk et al., epidural analgesia and cancer recurrence after colorectal cancer surgery. That concordance is the expected outcome when the propensity score model is adequate.

The third failure is covariate imbalance that arises after weighting. Inverse probability weights can create extreme weights when a small number of animals have very low probability of receiving the treatment they actually received. These extreme weights dominate the analysis and inflate variance. The discriminating check is the distribution of weights themselves. Weight values exceeding 10, or a weight distribution where the largest weight exceeds 10 percent of the total weight mass, indicate a fragile analysis. Trimming or truncating weights at a prespecified percentile is a recognized remedy, but it changes the estimand and must be reported.

Common Errors and Corrective Action

Less experienced analysts frequently build the propensity score model using every measured variable without regard to causal structure. Including variables that are affected by treatment, mediators on the causal pathway, induces collider bias and distorts the estimated effect. The corrective action is to specify the causal model before estimation and to exclude post-treatment variables from the propensity score model. A related error is including an instrumental variable, a variable that predicts treatment but not outcome, which reduces balance without reducing bias and can increase variance.

A second common error is treating the propensity score as a sufficient statistic without checking covariate balance within the matched or weighted sample. The propensity score balances covariates in expectation, but in any finite sample, chance imbalance persists. The corrective action is routine balance assessment after every adjustment procedure, also before it.

A third error is interpreting the propensity score analysis as equivalent to a randomised trial. Propensity score methods address measured confounding only. Unmeasured confounders, such as owner compliance, subclinical disease severity, or farm-level management practices, can still bias the estimate. The study of propranolol and cancer risk in a large administrative database illustrates this limitation: the propensity score matched cohorts on measured claims data, but lifestyle factors and over-the-counter medication use were unavailable Chang et al., propranolol and cancer risk population-based cohort study. Sensitivity analyzes for unmeasured confounding should accompany the primary analysis.

Limitations of Current Evidence

The veterinary evidence base for propensity score methods is thin relative to human medicine. Most veterinary applications appear in companion animal oncology and production animal medicine, and few studies report the full set of diagnostics recommended here. The neonatal necrotising enterocolitis surveillance study in England used propensity score analysis on a large national database, demonstrating feasibility in a registry setting, but such registries are rare in veterinary medicine Battersby et al., incidence and enteral feed antecedents of severe neonatal necrotising enterocolitis. Veterinary electronic medical records are fragmented across practice management systems, and data on important confounders such as body condition, client compliance, and environmental exposures are often missing or recorded inconsistently.

Expert opinion still differs on whether propensity score matching or inverse probability weighting should be preferred in veterinary settings. Matching is intuitive and produces a sample that resembles a randomised trial, but it discards animals and reduces precision. Weighting retains the full sample but is sensitive to extreme weights and model misspecification. Doubly robust methods, which combine outcome regression with propensity score weighting, are increasingly recommended but are rarely used in veterinary publications. The reporting standards from the EQUATOR Network library, including STROBE for observational studies, provide a structured framework for what should be disclosed, and the ARRIVE guidelines specify the minimum information required for transparent animal research reporting EQUATOR Network reporting guidelines and ARRIVE guidelines for reporting animal research.

Escalation and Consultation

Referral to a biostatistician is warranted when the propensity score model fails to achieve balance after iterative revision, when the study involves time-varying treatments or competing risks, or when the analysis will inform regulatory decisions. Laboratory involvement is indicated when outcome ascertainment depends on diagnostic tests with imperfect sensitivity or specificity, because misclassification of outcomes can bias effect estimates in either direction. Regulatory reporting obligations arise when the study involves notifiable diseases, and the World Organization for Animal Health terrestrial animal health standards define the surveillance and reporting framework that applies across member countries WOAH terrestrial animal health code. Species-specific clinical reference materials should be consulted when interpreting the biological plausibility of findings, as the MSD Veterinary Manual provides peer-reviewed guidance across species MSD Veterinary Manual professional edition.

Frequently Asked Questions

How many events per variable do I need for a propensity score model in a veterinary dataset?

The traditional 10 events per variable rule applies to the outcome model, not the propensity score model. For the score model itself, simulation work suggests that 5 to 10 outcome events per covariate is adequate when using the score for matching or weighting. Sparse veterinary datasets, such as those from referral hospital records, often contain fewer events. In these situations, reduce the number of covariates in the score model, combine related covariates, or use a high-dimensional propensity score approach with variable selection. Check covariate balance after estimation regardless of event counts. If balance remains poor, consider inverse probability weighting with stabilized weights instead of matching, because weighting uses the full sample and may perform better with limited data.

Can I use propensity scores when my study includes multiple species or breeds?

Yes, but the score model must account for species and breed as covariates or through stratification. A single score estimated across species assumes the treatment assignment mechanism is comparable, which is rarely true. For example, the decision to administer a particular analgesic may depend on species-specific metabolic pathways and licensed indications. Fit the propensity model separately within species or include species-by-covariate interactions. After matching or weighting, verify balance within each species subgroup, also in the pooled sample. If sample sizes within species are too small for separate models, consider restricting the study to the most common species and reporting this as a limitation.

What should I do when the treatment and control groups have no region of common support?

Common support failure means some treated animals have no comparable control animals. Examine the propensity score distributions graphically before analysis. When overlap is poor, trim the sample to the region of overlap and report how many animals were excluded. Alternatively, use weighting with truncation at specified percentiles, such as the 1st and 99th, to reduce the influence of extreme weights. Avoid matching without replacement in this setting because it forces matches of poor quality. If the non-overlap is extensive, the treatment effect is not estimable in the available data without strong assumptions. Consider whether the treatment is reserved for a specific clinical subgroup and reframe the research question accordingly.

How do I handle missing data in covariates used to estimate the propensity score?

Complete-case analysis is acceptable only when missingness is minimal and unrelated to treatment or outcome. Multiple imputation is preferred when more than a small fraction of observations have missing covariates. Impute the covariates, estimate the propensity score in each imputed dataset, then combine the estimated scores across imputations for matching or weighting. For matching, a practical approach is to estimate the score in each imputed dataset and match within each dataset, then pool the treatment effect estimates using Rubin's rules. Do not impute the outcome solely to improve the propensity model. Sensitivity analyzes comparing complete-case and imputed results help assess whether missing data influenced conclusions.

How should I report propensity score analyzes in a veterinary journal?

Follow the STROBE guidelines for observational studies and consult the EQUATOR Network reporting guidelines for the relevant checklist. Report how the propensity score was estimated, which covariates were included and why, and the balance diagnostics before and after adjustment. State the matching algorithm, caliper width, and whether matching was with or without replacement. Report the number of unmatched animals and the effective sample size after weighting. Provide both unadjusted and adjusted estimates. For animal research, also follow the ARRIVE guidelines for reporting animal research to ensure the study is reproducible.

Do propensity score methods apply to studies of production animals or wildlife?

They apply whenever treatment assignment is non-random and confounders are measurable. In production medicine, herd-level treatments such as vaccination protocols or metaphylaxis create clustering that must be addressed, often with multilevel propensity score models or by estimating scores within herds. Wildlife studies face additional challenges: sparse data, incomplete covariate measurement, and non-independence due to social structure or spatial proximity. Propensity scores do not remove confounding by unmeasured variables in any setting. In wildlife and production contexts, consider sensitivity analyzes for unmeasured confounding and consult species-specific reference materials from sources such as the MSD Veterinary Manual when defining clinically meaningful covariates.

Related Clinical & Scientific Guides

• Conducting Systematic Reviews of Veterinary Diagnostic Test Accuracy
• Bias in Veterinary Research: Types, Sources, and Mitigation
• Cluster Randomized Trials in Veterinary Research: Design and Analysis

References and Further Reading

• Antifungal prophylaxis for prevention of COVID-19-associated pulmonary aspergillosis in critically ill patients: an observational study.. 2021.
• Association between epidural analgesia and cancer recurrence after colorectal cancer surgery.. 2010.
• Long-term effects of routine morphine infusion in mechanically ventilated neonates on children's functioning: five-year follow-up of a randomized controlled trial.. 2011.
• Incidence and enteral feed antecedents of severe neonatal necrotising enterocolitis across neonatal networks in England, 2012-13: a whole-population surveillance study.. 2017.
• Propranolol Reduces Cancer Risk: A Population-Based Cohort Study.. 2015.
• Statin use after diagnosis of breast cancer and survival: a population-based cohort study.. 2015.
• ARRIVE Guidelines 2.0 for Reporting Animal Research. PLOS Biology, 2020.
• EQUATOR Network Reporting Guidelines. EQUATOR Network.
• MSD Veterinary Manual, Professional Edition. MSD Veterinary Manual.

Related Articles

• Using Mixed Methods in Veterinary Research
• Conducting Qualitative Research in Veterinary Settings
• How to Write a Research Protocol for Veterinary Studies
• Applying Competing Risks Analysis in Veterinary Research
• The ARRIVE Guidelines for Animal Research: Implementation and Compliance

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.