Patients clustered within wards may share staff, routines, resources and organisational conditions. Their outcomes can therefore be related. A nursing dissertation that treats every patient as an entirely independent observation may misunderstand the study’s design and the precision of its findings.
The central question is not simply how many patients appear in the dataset. You also need to know how they are grouped, where an intervention operates and how the analysis handles that grouping. A large patient count cannot automatically compensate for very few wards.
This guide focuses on patients grouped within wards or similar settings. For a different issue, such as repeated outcomes from the same person, use our repeated measures and mixed ANOVA guide. The two forms of dependence can coexist, but they are not interchangeable.
Why patients clustered within wards matter
Clustering means observations belong to groups whose members may resemble one another. A ward-level process can influence several patients together. Therefore, the information supplied by 100 patients in one ward need not match the information supplied by 100 independent patients across unrelated settings.
The CONSORT extension for cluster randomised trials explains why allocation, recruitment, analysis and reporting must recognise the cluster design. Use it to appraise the study rather than assuming that a conventional patient-level test is adequate.
In your writing, identify the actual cluster. It might be a ward, hospital, practice or community team. Avoid using the word cluster as a vague synonym for any group in a table.
Separate the unit of allocation, measurement and analysis
The unit of allocation is what receives the assigned condition in a randomised study. The unit of measurement is where an outcome is recorded. The analytical approach must respect the design and the dependence in those measurements.
For example, a study may allocate wards to different handover approaches but record outcomes for individual patients. The presence of patient-level outcomes does not turn ward allocation into individual randomisation.
Write these units explicitly in your appraisal notes. Many misunderstandings become visible once you place allocation, intervention delivery and outcome measurement side by side. If the paper leaves a unit unclear, report that as a limitation rather than guessing.
| Design feature | Illustrative answer | Why it matters |
|---|---|---|
| Unit allocated | Ward | The assigned condition operates at group level |
| Outcome recorded | Individual patient | Measurements remain nested within wards |
| Shared conditions | Staffing and local workflow | Patient outcomes may be related |
| Analysis question | How was within-ward dependence handled? | Precision should reflect the design |
Use a fictional example to expose the problem
Imagine a fictional study with eight wards and 40 patients in each ward. Four wards use one process and four use another. The dataset contains 320 patient records, but the intervention was assigned only eight times.
A weak appraisal might praise the sample size of 320 without considering the ward structure. A stronger appraisal would examine the number of wards, the similarity of patients within each ward and the method used to estimate uncertainty.
The example is entirely invented. It illustrates a design question and does not establish that eight wards are sufficient or insufficient for a particular study. Adequacy depends on the aim, assumptions, effect of interest and analytical design.
Explain the intracluster correlation coefficient
The intracluster correlation coefficient, or ICC, describes similarity within clusters relative to the overall variation under a specified model. Its interpretation depends on the outcome and model. It is not a universal property of a ward.
In an appraisal, record which outcome the reported ICC concerns and how the authors obtained it. A value from another population or measure may help planning only if the transfer is justified.
Do not write that a small ICC makes clustering irrelevant. The effect on precision can also depend on how many people belong to each cluster. The size of the cluster and the strength of similarity need consideration together.
Understand a simple design-effect calculation
For a simplified equal-sized cluster setting, a commonly used design effect is 1 + (m − 1) × ICC, where m is the cluster size. This illustrates how within-cluster similarity can increase the information requirement compared with an independent-observation assumption.
Using fictional values of m = 40 and ICC = 0.05 gives 1 + 39 × 0.05 = 2.95. Dividing 320 by 2.95 gives roughly 108 as a rough independent-observation equivalent for this illustration.
Do not present that number as a complete sample-size calculation or a replacement for a cluster-trial analysis. The formula’s assumptions matter, and real designs may involve unequal cluster sizes, small numbers of clusters or additional levels of dependence.
Do not confuse design effect with a corrected p-value
A planning illustration and an analysis of observed data are different tasks. Calculating a design effect after reading a paper does not automatically repair an analysis that ignored the design.
Likewise, dividing the sample size by a design effect does not tell you the correct treatment estimate, confidence interval or degrees of freedom for every outcome. Your appraisal should ask how the actual analysis handled dependence.
If the method is unclear, describe the concern precisely. You can say that the paper does not provide enough information to assess the handling of within-ward correlation. Avoid inventing a corrected result and treating it as the study’s finding.
Recognise analysis approaches without prescribing one universally
Authors may use methods such as cluster-level analysis, multilevel models or generalised estimating equations. These approaches address dependence in different ways and require assumptions appropriate to the outcome and design.
A nursing appraisal does not need to claim that one named method is always best. Instead, explain whether the selected approach recognises the ward structure and whether the paper reports enough detail to assess its use.
For example, identify the grouping level, outcome distribution, covariates and relevant uncertainty method where reported. A sentence merely stating that advanced statistics were used does not show that the model matches the study.
Consider the number of wards, not only patients
When an intervention operates at ward level, adding more patients to the same wards is different from adding more independent wards. The distinction affects how much the study can learn about variation across settings.
In the fictional example, 320 records do not create 320 independent allocation units. A conclusion based on eight wards should make that design visible. Do not describe the study as a large individually randomised trial.
Also consider whether one unusual ward could strongly influence the result. Your appraisal can discuss that possibility without claiming it occurred. Look for the authors’ reporting of cluster characteristics, analysis choices and sensitivity checks.
Inspect unequal cluster sizes
Real wards may contribute different numbers of patients. A simple equal-size illustration may therefore be inadequate for interpreting the actual design. Record the distribution where the paper provides it.
Ask whether a few large wards dominate the dataset and how the analysis accounts for the structure. The answer depends on the model and estimand, so avoid assuming that equal weighting or patient weighting is always correct.
In your dissertation, state the reporting gap if cluster sizes are absent. This limits your ability to assess the study’s precision and applicability. It does not justify inventing equal sizes to complete an appraisal table.
Keep clustering distinct from confounding
Accounting for correlated outcomes and adjusting for confounding are related analytical concerns, but they are not the same task. A model can include patient characteristics while still needing to recognise ward-level dependence.
Similarly, acknowledging clustering does not remove every difference between wards. Staffing, case mix or service organisation may still matter to interpretation, depending on the design and allocation.
A clear appraisal identifies each issue separately. Describe whether the study addresses dependence, baseline differences and relevant covariates rather than using the single phrase adjusted analysis as a complete quality judgement.
Examine recruitment after cluster allocation
In a cluster trial, the timing of participant recruitment can matter. If people enter the study after the ward’s condition is known, examine what the paper reports about selection and recruitment procedures.
The relevant question is whether the process could create differences between groups. Do not assume bias merely because recruitment follows allocation, but do not ignore the timing either.
Record the evidence: who identified eligible patients, whether the process differed between conditions and how the authors addressed the concern. This creates a defensible appraisal rather than a generic statement that all trials have selection bias.
Distinguish ward clustering from repeated measurements
A study can contain patients within wards and several observations within each patient. Those are separate levels of dependence. A repeated-measures approach that handles observations within a person does not necessarily address differences between wards.
Before discussing the analysis, sketch the data hierarchy in words: hospital, ward, patient and time point, as applicable. Then identify which levels the reported model recognises.
Do not assume that every level requires an identical solution. The design, outcome and analytical purpose determine the appropriate method. Your assignment should explain the hierarchy and appraise the reported approach rather than choosing a model by name alone.
Interpret confidence intervals in light of the design
A confidence interval communicates uncertainty under the analysis’s assumptions. If dependence is ignored, the apparent precision may not reflect the design appropriately. Therefore, examine how the interval was estimated before praising a narrow range.
Also keep the effect estimate and its precision distinct. A model may change the uncertainty even when the central estimate looks similar. Do not judge the importance of design adjustment solely by whether the headline effect changes.
For the underlying interpretation, our confidence intervals guide explains uncertainty more broadly. This article adds the specific question of whether the model recognises patients grouped within wards.
Write the limitation precisely in a literature review
A useful critique identifies the design feature, the reporting or analytical concern and its consequence. It should not simply state that the study used a small sample.
For example: although the paper reports many patient records, the intervention was allocated across a limited number of wards. The appraisal therefore needs to consider cluster-level variation and whether the analysis handled within-ward dependence when estimating uncertainty.
If the paper did use an appropriate model, acknowledge that strength. Critical appraisal does not require finding a fault in every study. Explain both the issue and the evidence that the authors addressed it.
Use a structured extraction table
Add columns for cluster definition, number of clusters, patients per cluster, allocation level, outcome level, analysis method and reported ICC. Include a column for information not reported.
Next, separate extraction from judgement. Record the paper’s actual method before evaluating it. This prevents a mistaken summary from becoming the basis for a confident critique.
Finally, compare like with like across studies. A trial with ward-level allocation and a survey of patients from one hospital have different designs. Both can involve grouped data, but the same appraisal sentence may not fit both.
Check the claim made in the discussion
The study’s design affects more than its p-value. Consider the population and settings represented by the clusters. A finding from a limited set of wards does not automatically describe every hospital or service configuration.
Explain what supports transferability and what remains uncertain. Relevant features may include the intervention’s delivery, staff roles and local workflow. Avoid treating the word multicentre as sufficient evidence that all settings are represented.
Also keep the intervention level visible in recommendations. A ward-level change may require organisational support. Do not translate it into an individual nurse’s responsibility without considering the system in which it operates.
Common mistakes to avoid
- Counting patient records as independent allocation units.
- Ignoring a small ICC because it looks numerically minor.
- Using a simple design effect as a universal analysis correction.
- Confusing repeated measurements with ward-level grouping.
- Assuming covariate adjustment automatically handles clustering.
- Inventing cluster sizes or ICC values absent from the paper.
- Equating a named statistical model with a fully appropriate analysis.
- Criticising a study without checking whether the authors addressed the issue.
Explain what to do when the ICC is not reported
Absence of a reported ICC does not automatically prove that the analysis ignored clustering. Read the methods and uncertainty estimation before drawing that conclusion. Authors may describe an appropriate model without presenting every planning parameter.
However, missing information can still limit appraisal. State exactly what you cannot assess, such as the assumptions behind a sample-size calculation or the extent of within-ward similarity for a particular outcome. This is more precise than declaring the entire study invalid.
Next, distinguish planning values from estimates obtained during the study. A planned ICC may come from earlier evidence, while a reported analysis may estimate a related quantity from the observed data. Record which value the paper gives and do not describe them as necessarily identical.
For a dissertation extraction table, use not reported where appropriate. Do not insert zero, an average from unrelated studies or the value used in this article’s fictional example. If you discuss a plausible range to explain uncertainty, label it as your illustrative sensitivity discussion rather than the authors’ result.
Finally, explain the consequence for synthesis. You may retain a study while qualifying confidence in its precision or reporting. The appraisal should show how the concern affects your use of the finding, rather than adding a criticism that has no influence on the conclusion.
Frequently asked questions
Does every hospital study require a cluster trial analysis?
No. First identify the design and dependence structure. A hospital setting alone does not determine the allocation unit or the appropriate model.
Can a small ICC still matter?
Yes. Its consequence depends partly on cluster size and design. Explain both factors rather than judging the ICC in isolation.
Is the design effect the same as the ICC?
No. The ICC concerns within-cluster similarity under a model. A design effect describes a consequence for information or variance under specified assumptions.
Can I fix an ignored cluster structure by changing the sample size?
Not as a general solution. The actual analysis needs an appropriate method. A rough calculation can illustrate the issue but does not create a valid corrected result for every design.
Do I need to run multilevel software for a literature review?
Usually the task is to appraise the published methods, not reanalyse unavailable patient data. Follow the brief and describe any limits in the information the paper provides.
Get feedback on your research appraisal
Our nursing data-analysis and results support can help you identify the design question and interpret methods using legitimate data and appropriate academic guidance.
Get My Nursing Assignment Quote. Include your research question, data hierarchy or appraisal draft so feedback addresses the actual problem.
Conclusion
Patients clustered within wards create a question about dependence, not merely a bigger headcount. Identify the allocation and measurement units, examine the cluster structure and check how uncertainty was estimated. A precise appraisal explains what the analysis recognises and what the available reporting leaves unresolved.
Reference
Campbell, M.K. et al. (2012) ‘CONSORT 2010 statement: extension to cluster randomised trials’, BMJ, 345, e5661. Publication record and journal article.
