Planning Stepped-Wedge Cluster Trials in Veterinary Settings

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

Key Takeaways

• Stepped-wedge cluster trials are indicated when simultaneous intervention rollout is infeasible or unethical, and the intervention is applied at a cluster level (e.g., herd, practice, region) rather than individual animal.
• The design requires explicit modeling of time effects to distinguish intervention impacts from secular trends, such as seasonal disease patterns or market shifts, which can otherwise bias estimates.
• Adequate statistical power is critically dependent on the number of clusters; trials with fewer than ten clusters have limited ability to estimate cluster-level variance, leading to wide confidence intervals.
• Transition periods between control and intervention phases must be pre-specified and handled carefully in analysis, as they represent a ramp-up phase where the intervention is not yet fully operational.
• Common failure modes include secular trend confounding, contamination between clusters, implementation failure (poor fidelity), and attrition, all of which can lead to biased estimates or underpowered comparisons.
• Reporting should adhere to extensions of CONSORT for cluster trials and specific guidelines like REFLECT for livestock, ensuring transparency regarding randomization, blinding (if applicable), and statistical analysis methods.

Stepped-wedge cluster trials offer veterinary researchers a practical route to evaluating interventions when withholding a promising intervention from entire herds, practices, or regions is ethically or logistically untenable. In this design, all clusters eventually receive the intervention, but the timing of rollout is randomized. This article explains the design logic, the conditions under which the design is preferable, and the planning decisions that determine whether the resulting evidence will be credible. It is written for veterinary researchers designing implementation studies, antimicrobial stewardship programs, or practice-level quality improvement initiatives across production animal, companion animal, and One Health settings.

The central question this article answers is practical: when should you choose a stepped-wedge design, how do you structure the rollout, and what must you specify before data collection begins? The design has been used to evaluate veterinary antimicrobial stewardship programs in weaned pig production, where an eight-month intervention supporting adherence to a clinical practice guideline achieved a 25% reduction in farm-level antimicrobial use at the first measurement and 49% at the second, compared with baseline. The same design logic appears in human health implementation research, including protocols for postoperative pain management and smoking cessation in maternity wards, and in One Health rabies elimination programs in India. These examples illustrate the design's versatility, but they also expose its complexity. A stepped-wedge trial that is poorly planned will produce biased estimates, underpowered comparisons, or both.

At a Glance

Design Logic and Indications

The stepped-wedge design is a variant of the cluster randomized trial. Clusters are randomly allocated to one of several rollout sequences. At baseline, no cluster receives the intervention. At the first step, a randomly selected subset of clusters crosses over to the intervention condition. At each subsequent step, another subset crosses over, until all clusters are exposed. Measurements are taken throughout, so each cluster contributes both control-period and intervention-period observations.

The design is indicated when three conditions hold simultaneously. First, the intervention is delivered at the cluster level, such as a practice-level antimicrobial stewardship program, a biosecurity protocol applied to a herd, or a surveillance system implemented across administrative blocks. Second, there is a reason to believe the intervention is beneficial, making it difficult to justify a parallel-arm design in which some clusters never receive it. Third, practical constraints prevent simultaneous implementation across all clusters. These constraints may include limited training capacity, the need to phase in equipment or software, or the requirement to coordinate with existing disease control campaigns.

The design is not a default choice. It requires more measurements and more complex analysis than a parallel cluster trial, and it is more vulnerable to time-related confounding. If the intervention can be rolled out to all clusters simultaneously and a control group is acceptable, a parallel cluster trial is simpler and often more efficient. The stepped-wedge design earns its complexity only when the ethical or logistical case for staggered rollout is genuine.

Cluster Definition and Randomization

The cluster is the unit of randomization and the unit of intervention delivery. In veterinary settings, clusters may be farms, veterinary practices, practice networks, markets, or geographic regions. The choice of cluster definition determines the analysis structure, the number of independent units available, and the generalizability of the findings.

Randomization should be constrained where necessary to ensure balance on cluster-level characteriztics that are strongly associated with the outcome. The Mayo Clinic NOHARM trial, evaluating a non-pharmacological pain management bundle across 22 surgical practice clusters, used constrained randomization to assign clusters to one of five steps. This approach preserves the benefits of randomization while reducing the risk of severe imbalance on prognostic factors. In veterinary trials, relevant constraints might include herd size, production type, baseline disease prevalence, or practice caseload.

The number of clusters is the single most important determinant of statistical power. A stepped-wedge trial with fewer than ten clusters will have limited ability to estimate the cluster-level random effect variance, and the confidence interval for the intervention effect will be wide regardless of the number of animals measured within each cluster. Researchers should plan for the maximum feasible number of clusters, recognizing that adding clusters is almost always more valuable than increasing cluster size beyond a modest threshold.

Time Effects and the Transition Period

The stepped-wedge design assumes that the intervention effect can be distinguished from secular trends. This assumption is testable only if the model includes a time effect, and the design's validity depends on specifying that effect correctly. Outcomes in veterinary practice often trend upward or downward for reasons unrelated to the intervention, including seasonal disease patterns, market conditions, or concurrent policy changes. A model that omits time effects will attribute these trends to the intervention.

The transition period between control and intervention conditions requires explicit planning. Some interventions take effect immediately, while others require a ramp-up period during which staff are trained, protocols are embedded, or behavior change accumulates. The 5A-QUIT-N trial in maternity wards planned a transition period between control and intervention phases to deploy the intervention, recognizing that immediate measurement after rollout would capture implementation noise instead of the intervention's true effect. In veterinary settings, a transition period might be needed for veterinarians to complete training, for software to become operational, or for herd-level protocols to be fully adopted. The duration of this period should be specified in the protocol, and measurements taken during transition should be handled according to a pre-specified rule, typically either excluded from analysis or assigned to the control condition.

Analysis Principles

The primary analysis of a stepped-wedge trial uses a mixed-effects model with fixed effects for time period and intervention status, and random effects for cluster. The time effect accounts for secular trends, and the cluster random effect accounts for the correlation of outcomes within clusters. The intervention effect is estimated as the difference between intervention-period and control-period outcomes, adjusted for time.

The analysis must account for the correlation structure induced by repeated measurements within clusters. Ignoring this correlation will produce standard errors that are too small and p-values that are too small, increasing the risk of false-positive conclusions. The intracluster correlation coefficient, which quantifies the proportion of total variance attributable to between-cluster differences, is a key input to both sample size calculation and analysis. Researchers should report the estimated intracluster correlation coefficient and the method used to estimate it.

Reporting should follow established standards. The EQUATOR Network maintains a library of reporting guidelines, including the CONSORT extension for cluster randomized trials and the REFLECT statement for livestock trials. These guidelines specify the minimum information needed for readers to assess the risk of bias and the generalizability of the findings. The ARRIVE guidelines, developed by the NC3Rs, provide complementary standards for reporting animal research, including the number of animals, the experimental unit, and the statistical methods used.

Sample Size Planning

Sample size calculation for a stepped-wedge cluster trial must account for the same clustering effect that governs parallel cluster trials, plus additional variance components introduced by the staggered rollout. The intracluster correlation coefficient (ICC) describes the proportion of total outcome variance attributable to between-cluster differences. In veterinary field settings, ICC values commonly range from 0.01 to 0.20 depending on the outcome, the species, and the production system. Herd-level outcomes such as antimicrobial use or mortality tend to show higher ICCs than individual-animal outcomes measured within the same herds.

The stepped-wedge design is often less statistically efficient than a parallel cluster trial with the same number of clusters and measurement periods, because the intervention effect is estimated partly from within-cluster comparisons over time. However, the design can require fewer total clusters when logistical constraints prevent simultaneous rollout. The number of clusters, the number of steps, the cluster size, and the number of measurement periods per step all influence power. Increasing the number of steps generally improves precision, but each additional step extends trial duration and increases the risk of secular drift confounding the intervention effect.

A key planning decision is whether to treat the intervention effect as immediate and sustained or as gradual and potentially decaying. The analysis model must reflect the expected time course. For an intervention that requires behavior change, such as a veterinary antimicrobial stewardship program, the effect may accumulate over months. The stepped-wedge trial of a Streptococcus suis guideline adherence program in weaned pigs measured outcomes at baseline, during the intervention, and after the intervention, allowing the analysis to distinguish early from later effects reduced antimicrobial use in weaned pigs through a guideline adherence intervention. A sample size calculation that assumes an immediate full effect will underpower a trial where the true effect emerges gradually.

Worked Example

Consider a trial evaluating a herd-level biosecurity intervention across 12 veterinary practices, with each practice serving as a cluster. The primary outcome is the proportion of herds testing positive for a target pathogen at the end of each 3-month period. Six steps are planned, with two clusters crossing from control to intervention at each step. Each cluster contributes one measurement per period, and cluster size is fixed at 20 herds per period.

Assume an expected reduction in the proportion of positive herds from 0.40 to 0.25, an ICC of 0.05, and a coefficient of variation in cluster size of 0.3. A standard formula for stepped-wedge trials with binary outcomes, adjusted for the design effect of clustering and the number of steps, yields an estimated requirement of approximately 14 clusters for 80% power at alpha 0.05. With only 12 clusters available, the trial would be underpowered unless the expected effect size is increased, the ICC is lower than assumed, or additional measurement periods are added.

The calculation should be repeated under pessimistic and optimiztic assumptions. If the ICC is actually 0.10, the required number of clusters rises substantially. If the intervention effect is delayed by one full period, the effective number of post-intervention measurements decreases, and power falls accordingly. Sensitivity analysis across these parameters should be reported in the protocol, as recommended by reporting standards for cluster trials indexed in the EQUATOR Network reporting guidelines library.

Implementation Logistics

The transition period between control and intervention conditions requires explicit planning. During this period, the intervention is being introduced but is not yet fully operational, and data from this period are typically excluded from the primary analysis. The duration of the transition period should reflect the time needed to train personnel, distribute materials, and resolve equipment failures. In the 5A-QUIT-N trial of an organizational intervention for smoking cessation support in maternity wards, a dedicated transition period was planned between control and intervention phases to deploy the intervention protocol for the 5A-QUIT-N stepped-wedge cluster randomised trial. Veterinary equivalents include the time needed to train practice staff on a new diagnostic algorithm, to install and validate point-of-care testing equipment, or to implement a new prescribing protocol across multiple farm visits.

Cluster-level buy-in must be secured before randomisation. A practice that withdraws after randomisation but before its scheduled crossover creates an imbalance in the number of clusters per step and complicates the analysis. Contamination between clusters is less of a concern in stepped-wedge designs than in parallel designs, because all clusters eventually receive the intervention. However, contamination can still occur if veterinarians or producers in control-period clusters learn about the intervention from colleagues in intervention-period clusters and begin to adopt components early. This is a particular risk when the same veterinary practice serves multiple clusters or when regional meetings bring practitioners together.

Data Collection and Monitoring

Each cluster should contribute data at every measurement period, including periods before its crossover. Missing data at the cluster level is a serious threat to validity, because the analysis relies on comparing each cluster's pre-intervention trajectory with its post-intervention trajectory. Strategies to minimize missing data include automated data extraction from practice management software, scheduled reminder systems, and pre-specified procedures for handling herds that leave the study population.

The choice of outcome measure determines the monitoring schedule. For antimicrobial use outcomes, the metric must be defined precisely: defined daily doses per animal per year, treatment incidence per 100 animal-days, or total mass of active ingredient. These metrics are not interchangeable, and the sample size calculation must use the metric that will be analyzed. The Streptococcus suis guideline trial used farm-level antimicrobial use and veterinarian-level prescribing indicators as co-primary outcomes, requiring different analytical approaches for each level reduced antimicrobial use in weaned pigs through a guideline adherence intervention.

Documentation and Reporting

The protocol must specify the randomisation procedure, including whether constrained randomisation was used to balance cluster-level covariates across steps. The Mayo Clinic NOHARM trial used constrained randomisation to assign 22 practice clusters to five steps, balancing surgical specialty and geographic location across steps NOHARM stepped-wedge cluster-randomized pragmatic trial protocol. Veterinary trials should similarly balance species, production type, herd size, and baseline outcome levels across steps.

Reporting should follow the CONSORT extension for cluster randomised trials, with additional items specific to stepped-wedge designs. The ARRIVE guidelines for reporting animal research apply to the experimental components, while the EQUATOR Network library provides the relevant reporting checklists for the trial design itself EQUATOR Network reporting guidelines. The completed checklist should be submitted with the manuscript.

Planning Checklist

The following checklist consolidates the decisions required before a stepped-wedge cluster trial in a veterinary setting can begin.

1. Define the intervention and its expected time course. Specify whether the effect is immediate, gradual, or transient.
2. Define clusters and justify the cluster level. State the expected ICC and its source.
3. Determine the number of steps and the number of clusters per step. Confirm that the total number of clusters meets the sample size requirement under pessimistic assumptions.
4. Specify the transition period duration and the criteria for declaring a cluster fully operational.
5. Select primary and secondary outcomes. Define each outcome precisely, including the denominator and the measurement instrument.
6. Choose the analysis model. State whether the model includes random effects for cluster and time, and how secular trends are handled.
7. Plan for missing data. Specify the expected rate and the imputation or weighting approach.
8. Document the randomisation procedure, including any constraints.
9. Register the trial and publish the protocol before the first cluster crosses over.
10. Prepare reporting checklists and data-sharing statements before data collection begins.

Species and production system modify several of these decisions. In companion animal practice, clusters may be individual clinics with relatively small patient volumes per period, requiring longer measurement periods or more clusters. In production animal settings, clusters may be herds or practices, and the outcome may be measured at the group level instead of the individual level. In wildlife or free-ranging populations, cluster boundaries are harder to define, and the ICC may be higher due to spatial autocorrelation. The WOAH terrestrial animal health standards provide relevant guidance for surveillance-related outcomes in animal populations. Each setting requires the sample size calculation to be revisited with parameters appropriate to that context.

Recognized Complications and Failure Modes

Stepped-wedge designs fail in predictable ways. The most consequential failure is secular trend confounding, where an external change, such as a new antimicrobial restriction policy or a disease outbreak, coincides with the intervention rollout. Because all clusters eventually receive the intervention, the design cannot distinguish the intervention effect from a background temporal trend. Detection relies on the pre-intervention baseline period: if outcome trends are non-linear or unstable across the baseline steps, the parallel-trends assumption is threatened. Plot cluster-specific outcome trajectories by step and inspect for drift before the intervention begins.

Contamination across clusters is a second major failure mode. Veterinary practices within the same geographic region or corporate group may share information, referral networks, or locum staff, so control-period clusters may inadvertently adopt elements of the intervention. The discriminating check is process measurement: record intervention adherence indicators in all clusters during all periods, also in active clusters. If adherence rises in clusters still designated as control, contamination is present and the estimated effect will be biased toward the null.

Implementation failure, where the intervention is delivered late, incompletely, or with poor fidelity, is common in pragmatic veterinary trials. The stepped-wedge design assumes the intervention is fully active from the designated step onward. Early detection requires a run-in period with pilot testing and a formal fidelity monitoring plan. Track delivery dates against the randomisation schedule and record the proportion of intended components actually delivered in each cluster-period.

Attrition and cluster dropout create missing data that are particularly damaging in stepped-wedge designs because each cluster contributes both control and intervention observations. Differential dropout, where clusters leave during their control period but remain for the intervention period, can induce selection bias. Monitor cluster retention at each step and compare characteriztics of completing versus lost clusters.

Common Errors and Corrective Actions

A frequent analytical error is ignoring the cluster-period structure and analyzing the data as if observations were independent. This produces artificially narrow confidence intervals and inflated significance. The corrective action is to fit a mixed-effects model with random intercepts for cluster and, where appropriate, cluster-period terms, as described in the analysis principles section.

A second error is treating the transition period as either fully control or fully intervention. The transition period, during which the intervention is being deployed, should be excluded from both designations or modelled separately. Including transition-period data in the control arm dilutes the intervention effect, including it in the intervention arm underestimates the effect during early implementation.

A third error is failing to account for the correlation between repeated measurements from the same animal or farm within a cluster-period. The intracluster correlation coefficient must be estimated and incorporated at both the cluster and the within-cluster-period level. Using only a single-level correlation structure when two levels exist leads to overprecision.

A fourth error is selecting the number of steps without considering the minimum detectable effect size. Fewer steps with more clusters per step improve power but lengthen the trial. The sample size planning section should be revisited iteratively, not treated as a one-time calculation.

Limitations of Current Evidence

The veterinary stepped-wedge literature is sparse. The swine antimicrobial stewardship trial by Wayop and colleagues provides one of the few completed veterinary applications, demonstrating a 25% reduction in farm-level antimicrobial use at the first measurement and 49% at the second, but it also illustrates the challenges of participant attrition, with only 33 of 49 enrolled veterinarians contributing complete data Reduced antimicrobial use in weaned pigs through an intervention program supporting veterinarians' adherence to the clinical practice Streptococcus suis guideline. Human healthcare protocols, such as the NOHARM trial in surgical pain management and the 5A-QUIT-N trial in maternity care, offer methodological templates for complex organizational interventions, but their transferability to veterinary settings is untested Non-pharmacological Options in Postoperative Hospital-Based and Rehabilitation Pain Management (NOHARM): Protocol for a Stepped-Wedge Cluster-Randomized Pragmatic Clinical Trial.

Expert opinion still differs on whether the stepped-wedge design should be preferred over parallel cluster designs when the intervention is believed to be beneficial. Some argue that the ethical advantage of eventual exposure to all clusters justifies the design even with reduced statistical efficiency. Others contend that the design's vulnerability to time confounding makes it inferior unless the secular trend is known to be minimal. The One Health rabies implementation protocol from Kerala illustrates the design's appeal for public health interventions where withholding the intervention from control areas is ethically problematic Integrated Bite Case Management within a One Health Framework for Rabies Elimination at the Primary Care Level in Kerala, India: An Implementation Research Protocol Using a Stepped-Wedge Cluster Randomized Design.

Escalation and Referral

Statistical consultation is warranted when the intracluster correlation coefficient is unknown, when the number of clusters is small, or when the analysis plan requires adjustment for time-varying covariates. A veterinary epidemiologist or biostatistician should be engaged before randomisation, not after data collection begins.

Laboratory involvement is required when outcome measurement depends on diagnostic testing, such as bacteriological culture or antimicrobial susceptibility testing. The swine trial's performance indicators included the use of bacteriological examination, and standardization of laboratory protocols across clusters is essential for valid comparison Reduced antimicrobial use in weaned pigs through an intervention program supporting veterinarians' adherence to the clinical practice Streptococcus suis guideline.

Regulatory reporting obligations arise when the intervention involves changes to antimicrobial use, vaccine administration, or other regulated products. Reporting standards for animal research, including the ARRIVE guidelines, should be followed at publication ARRIVE Guidelines 2.0 for Reporting Animal Research. Where the intervention affects food-producing animals, national residue monitoring requirements and international standards from the World Organization for Animal Health may apply WOAH terrestrial animal health standards.

Troubleshooting Table

Frequently Asked Questions

How do I estimate the budget when the intervention requires repeated cluster-level training or equipment purchases?

Budget for three cost categories: fixed setup costs per cluster, per-step delivery costs, and ongoing maintenance costs. Fixed costs include training materials, equipment, and certification of personnel. Per-step costs recur each time a new cluster crosses over, such as travel, trainer time, and consumables. Maintenance costs cover refresher training, software licences, and replacement supplies. Pilot one cluster to measure actual resource use before finalising the budget. The stepped-wedge design spreads delivery costs across steps, which can ease cash flow but extends the total study duration. Include contingency for delayed steps, as protocol drift in delivery timing threatens the validity of the time-effect adjustment.

What can I do when the planned number of clusters is unaffordable or logistically impossible?

Reduce the number of clusters only after considering alternatives. Increasing the number of measurement periods per cluster can partially compensate for fewer clusters, because the design gains precision from repeated observations. Alternatively, shorten the step interval to fit more steps into the same calendar window. If clusters are heterogeneous, constrained randomisation can balance prognostic factors across steps and improve efficiency. When the cluster count is fixed and small, consider a cohort design with repeated measures on the same animals instead of cross-sectional sampling. Report the achieved power honestly. A pragmatic trial with fewer clusters than ideal may still produce useful effect estimates if the analysis accounts for the reduced precision.

How does the design change when the unit of intervention is a herd but the unit of measurement is an individual animal?

The cluster is the herd, and individual animals within a herd are correlated. Sample size calculations must inflate for the intracluster correlation coefficient, and the analysis must use mixed-effects models with herd as a random effect. If animals are sampled repeatedly from the same herd across periods, model the repeated measures on individual animals when animals can be identified. When animals are not identifiable, treat each period's sample as independent within the herd. The reporting standards for animal research require clear specification of the cluster and measurement units. In production settings, consider that herd-level outcomes such as antimicrobial use per animal may be more robust than individual animal outcomes.

What record keeping is essential during the transition period when a cluster is preparing to cross over?

Document the start and end dates of the transition period for each cluster, the training completion date, and the date the intervention became fully operational. Record any components delivered early or late, and note staff turnover during training. Keep logs of intervention fidelity, such as checklists completed, materials distributed, or software activated. These records allow the analysis to model the intervention as a continuous variable instead of a binary switch, which is valuable when implementation is gradual. The EQUATOR Network reporting guidelines list the items needed for transparent reporting of implementation trials. Incomplete transition records are a common cause of unanalysable stepped-wedge data.

How should I explain the stepped-wedge design to a referring veterinarian or practice owner who is asked to participate?

Explain that every participating practice eventually receives the intervention, which is often more acceptable than randomisation to a control group that never receives it. The order in which practices start is determined by chance, and the staggered start allows comparison of outcomes between practices that have started and those that have not. Emphasize that the intervention is delivered as part of normal practice, so the burden is limited to data collection and training. Clarify that the practice's own patients benefit from the intervention during the study period. For practice owners concerned about staff time, refer to the AVMA practice resources for guidance on integrating research activities into clinical workflows.

When is a stepped-wedge design inappropriate even when it seems logistically attractive?

Avoid the design when the intervention effect is expected to change over time within a cluster, because the design cannot separate the intervention effect from time effects without strong assumptions. Avoid it when the outcome is measured once per cluster at the end of the study, as the design requires repeated measurements. Avoid it when contamination between clusters is likely, because the staggered rollout increases the opportunity for practices to share information. Also avoid it when the intervention has a rapid, large effect that makes the delayed rollout ethically questionable. In these situations, a parallel cluster trial or an interrupted time series may be more appropriate. Consult the MSD Veterinary Manual for guidance on outcome measurement in veterinary clinical settings.

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

• Reduced antimicrobial use in weaned pigs through an intervention program supporting veterinarians' adherence to the clinical practice Streptococcus suis guideline.. 2024.
• Non-pharmacological Options in Postoperative Hospital-Based and Rehabilitation Pain Management (NOHARM): Protocol for a Stepped-Wedge Cluster-Randomized Pragmatic Clinical Trial.. 2022.
• Integrated Bite Case Management within a One Health Framework for Rabies Elimination at the Primary Care Level in Kerala, India: An Implementation Research Protocol Using a Stepped-Wedge Cluster Randomized Design. 2026.
• Personal, social, and environmental correlates of physical activity in adults living in rural south-west England: a cross-sectional analysis.. 2013.
• Effectiveness evaluation of an organizational intervention, targeting pregnancy and addiction care professionals, among women who have just given birth in maternity wards and smoked tobacco during pregnancy (5A-QUIT-N): study protocol for a stepped-wedge cluster randomised trial.. 2024.
• 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.
• American Veterinary Medical Association Practice Resources. American Veterinary Medical Association.

Related Articles

• Designing Crossover Trials for Veterinary Therapeutics
• Cluster Randomized Trials in Veterinary Research: Design and Analysis
• Designing Equivalence Trials for Veterinary Generic Drugs
• Equivalence and Non-Inferiority Trials in Veterinary Medicine
• Pragmatic vs Explanatory Trials in Veterinary Clinical Research

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.