Fisher’s exact test in nursing research helps you analyse an association between categorical variables when a contingency table is sparse. It is especially useful when a standard chi-square approximation is questionable. However, selecting an exact method does not remove the need to check the study design, define the outcome clearly and report uncertainty honestly.
This guide focuses on choosing, running and reporting Fisher’s test for independent groups. It includes a worked nursing example, software checks and results wording. For the broader explanation of categorical association and expected frequencies, read our chi-square guide for nursing research.
Fisher’s exact test in nursing research: quick answer
Use the test when your question concerns an association between two categorical variables and the observations are independent. A common application is a 2 × 2 table comparing two groups on a binary outcome. Before choosing the method, inspect the expected frequencies and confirm that the categories describe counts of people or other independent units.
Report the observed counts, relevant percentages, the exact p-value and the direction of the comparison. Where appropriate, include an effect estimate and confidence interval obtained through a documented method. Interpret these results alongside the design and clinical context, rather than treating a small p-value as proof of an important improvement.
What does the test actually assess?
Imagine comparing whether nurses in two independently recruited groups completed a training module. Each nurse belongs to one group and has one completion status. The resulting table describes the association between group membership and completion. Fisher’s test evaluates the pattern under a null hypothesis of independence.
In a 2 × 2 table, the conditional null can also be expressed as an odds ratio of one. The word “exact” concerns how the reference probabilities are calculated. It does not mean that your sample represents every nursing population or that the estimated effect is free from uncertainty.
The official R documentation explains that the calculation conditions on the table’s marginal totals. It also identifies the alternative hypothesis and the returned estimate. These implementation details matter when comparing output from different analytical packages (R Core Team, n.d.).
Use counts, not percentages alone
The software needs the number of observations in each cell. A table containing 20%, 30%, 40% and 10% without the underlying denominators is insufficient. The same percentages can arise from very different sample sizes, producing different uncertainty and inferential results.
If you enter aggregated counts, check how your software treats the frequency variable. Each row of an aggregated dataset is a category combination, rather than a separate participant. Incorrect weighting can make four table cells appear to be a study of only four people.
When is Fisher’s exact test in nursing research appropriate?
Start with the research question. A binary outcome such as attendance, checklist completion or documented screening can produce a suitable table when compared across two independent groups. The analysis becomes questionable if participants contribute repeated outcomes or if the groups were defined after inspecting the result.
Sparse tables are common in small dissertations and studies involving uncommon outcomes. Kim (2017) discusses exact testing as an alternative to the chi-squared approximation in this setting. Nevertheless, a total sample below a particular number is not a universal decision rule. Examine the distribution of counts and the assumptions of the candidate method.
| Situation | Decision to consider | Main reason |
|---|---|---|
| Independent groups with a sparse binary outcome | Fisher’s exact test | The usual large-sample approximation may be unsuitable |
| Independent groups with adequate expected counts | Pearson chi-square may be suitable | The approximation can be adequate for the table |
| The same nurses measured before and after training | A paired binary method, such as McNemar’s test | Observations within each nurse are related |
| Participants grouped within wards | A method that accounts for clustering | People within a ward may share influences |
| A continuous outcome such as a scale score | A method suited to numerical data | Fisher’s test answers a categorical question |
Discuss the choice before analysing the outcome. A short decision record can state the unit of observation, table dimensions, independence assessment and reason for the selected test. This is more defensible than reporting that the software suggested a method without explaining why.
Check expected counts without changing the data
Expected counts describe the frequencies anticipated under independence. They differ from the observed numbers collected during the study. Software can display both, allowing you to assess whether the Pearson approximation is a reasonable choice.
Do not combine categories simply because an exact test gives an inconvenient result. Grouping must have a clinical or methodological rationale. For example, collapsing several employment roles might obscure a meaningful distinction between registered nurses and support workers.
Likewise, avoid turning a continuous score into “high” and “low” solely to obtain a 2 × 2 table. Such a decision discards information and changes the research question. If a recognised threshold is genuinely relevant, explain its origin and consider what the categorisation means for interpretation.
Preserve your original dataset and retain a clear record of any recoding. A reviewer should be able to trace the displayed counts back to the inclusion criteria and variable definitions. Our nursing data analysis workflow provides the wider preparation steps.
Worked example: training and checklist completion
The following numbers are invented for teaching. They are not findings from a clinical study. Suppose 24 independently observed nurses are divided into two groups. The outcome is whether each nurse completes a simulated checklist on one assessment occasion.
| Group | Completed | Did not complete | Total |
|---|---|---|---|
| Training group | 10 | 2 | 12 |
| Comparison group | 4 | 8 | 12 |
| Total | 14 | 10 | 24 |
Completion is 10/12, or 83.3%, in the training group and 4/12, or 33.3%, in the comparison group. The observed difference is 50 percentage points. These descriptive results should appear before the inferential interpretation, because they show what actually happened in the sample.
The expected counts under independence are seven completions and five non-completions in each group. Therefore, this example is not a case where every expected count is extremely small. Fisher’s test is used here to illustrate an exact calculation that may be specified in a small-sample analysis plan.
For this table, the probability-based two-sided Fisher p-value is approximately 0.0361. The simple cross-product odds ratio is (10 × 8)/(2 × 4), which equals 10. The risk ratio for completion is (10/12)/(4/12), which equals 2.5. These numbers describe different measures and must not be presented as interchangeable.
Interpret the example cautiously
The completion proportions favour the training group in this teaching dataset. However, the test alone does not establish that training caused the difference. That claim depends on allocation, baseline comparability, measurement and other design features that the small table does not describe.
There are only 12 nurses per group. Consequently, the point estimate should not be interpreted as a precise prediction of what would happen in a larger service. A suitable confidence interval would help readers understand the uncertainty, but it must come from an actual calculation rather than a guessed range.
Remember that “ten times the odds” does not mean “ten times the probability”. When an outcome is common, these quantities can differ substantially. In this example, completion is already frequent in the training group, so using odds language as if it described a probability would exaggerate the comparison.
How to run and verify the analysis
In SPSS, begin with a correctly coded dataset and the Crosstabs procedure. Choose the grouping variable and outcome variable, then request observed counts and percentages that answer your question. Inspect the output rather than assuming that every significance column refers to the same test.
Available exact-testing options depend on the installed version, licence and table structure. IBM provides an Exact Tests manual describing the relevant procedures. Check your own installation before promising a particular menu or output table in your methods chapter.
For a standard 2 × 2 result, identify the Fisher’s Exact Test row and the appropriate exact significance column. Do not copy the Pearson asymptotic p-value and label it as Fisher’s result. Save the output alongside the syntax or other record needed to reproduce the analysis.
A reproducible R teaching example
counts <- matrix(c(10, 2, 4, 8), nrow = 2, byrow = TRUE)
dimnames(counts) <- list(
Group = c("Training", "Comparison"),
Completion = c("Yes", "No")
)
fisher.test(counts, alternative = "two.sided")Check the displayed row and column labels before interpreting the estimate. Reversing one category order can reverse an odds ratio. Although the two-sided association test remains equivalent, the narrative direction of the effect changes.
R returns a conditional maximum-likelihood odds ratio estimate, which need not equal the simple cross-product estimate. Label your method accurately if your report includes that estimate. A difference between the two numbers is not automatically evidence that one program has made an error.
Choose one-sided or two-sided testing in advance
A two-sided test allows evidence of an association in either direction. For many dissertation questions, it is the appropriate default unless a clearly justified analysis plan specifies otherwise. The direction should follow the research question and accepted statistical advice.
Do not select a one-sided result after discovering that the two-sided p-value exceeds your threshold. That changes the analytical decision in response to the data. It also hides the possibility that an intervention performs differently from the anticipated direction.
If a one-sided test is justified, state the direction before analysis and ensure that the row and column coding matches it. A software option called “greater” depends on the parameter definition, not on whether a larger value sounds desirable in everyday language.
How to report Fisher’s test in a nursing dissertation
A clear results paragraph combines the table with the test result. Specify which groups were compared, identify the outcome, give the denominators and state whether the p-value was one-sided or two-sided. Include the analytical population if exclusions changed the available sample.
A teaching example could read:
In the simulated dataset, checklist completion was observed in 10 of 12 nurses in the training group (83.3%) and 4 of 12 in the comparison group (33.3%). A two-sided Fisher’s exact test gave p = 0.036. The observed completion proportions differed between groups; the small sample and unspecified allocation process limit causal interpretation.
Adapt the structure to your actual results. If you calculate an odds ratio or another effect measure, add its confidence interval and method. Do not report an interval, adjusted estimate or analysis that you have not performed.
Fisher’s test does not require you to invent a chi-square statistic or degrees of freedom for the results sentence. Some software displays additional statistics for extended exact procedures. Report the method that was actually used, rather than borrowing a template for a different test.
For help connecting tables with prose, see our guide to the nursing SPSS results chapter. Keep the results section descriptive, then use the discussion to consider mechanisms, existing evidence and limitations.
Why the effect estimate and interval still matter
An exact p-value answers a hypothesis-testing question. It does not tell a reader how large the association is or whether its plausible magnitude would matter in practice. A large estimated odds ratio from a sparse table can still be highly uncertain.
Confidence intervals and two-sided exact tests also require methodological care. Fay (2010) explains that different definitions of a two-sided exact test and an exact interval can sometimes produce apparently inconsistent threshold conclusions. Therefore, document the chosen implementation rather than silently replacing one output with another.
If a confidence interval includes the null value while an exact test crosses the significance threshold, investigate whether the procedures are matched. Do not assume that the discrepancy can be resolved by changing the study hypothesis. Your supervisor or a statistician can help identify the relevant definitions.
Clinical interpretation also depends on the absolute pattern. A difference involving an uncommon adverse event may have another practical meaning from the same odds ratio for a frequent training outcome. Explain the numerator, denominator and observation period before discussing implications for nursing practice.
Limitations an exact test cannot repair
Independent observation is fundamental. A table formed from repeated records belonging to the same patient does not become valid merely because Fisher’s test was selected. The relevant method must reflect the dependence in the data.
Selection and measurement can also affect the result. If the comparison group was recruited from a different setting or assessed with another checklist, the apparent association may reflect those differences. An exact probability calculation cannot separate these influences automatically.
Missing outcomes need transparent handling. Explain how many records were excluded and why. If missingness differs between groups, the displayed percentages may represent selected subsets rather than the full recruited sample. Avoid coding a missing answer as “no” unless that coding is genuinely justified.
Multiple testing is another concern. Running many independent tables and highlighting only the smallest p-value creates a misleading account of the evidence. Define the main comparison and distinguish prespecified analyses from exploratory checks.
A non-significant result should not be described as proof that groups are identical. The sample may provide limited information about clinically meaningful differences. Discuss the observed estimates and uncertainty, rather than reducing the entire study to a binary verdict.
Final checks before submitting the analysis
- Confirm that the question concerns categorical counts.
- Identify the independent observational unit.
- Verify eligibility, coding and missing-value decisions.
- Inspect observed and expected counts.
- Record why the selected method fits the design.
- Check the software’s exact significance column.
- Verify the row and column reference categories.
- Report counts, denominators and relevant percentages.
- Label the effect estimate and interval method correctly.
- Explain uncertainty and avoid unsupported causal claims.
Ask a second reader to reconstruct the result from your table. They should be able to identify the comparison, determine which proportion uses which denominator and understand why the test was chosen. If they cannot, improve the reporting before polishing the wording.
Need help checking your categorical analysis?
Use nursing dissertation support to review your test selection, coding, tables and interpretation. Bring your assignment brief, anonymised variable definitions and existing output. Support should help you understand and report your own work, while following your university’s requirements.
Frequently asked questions
Is Fisher’s exact test only for very small samples?
No. Sparse counts and the suitability of the alternative approximation are more useful considerations than a universal total-sample threshold. The study design must still support the analysis.
Can I use it for the same patients before and after treatment?
A standard independent-groups Fisher test does not account for paired observations. Consider a paired binary method and obtain advice appropriate to the design.
Can it analyse a table containing a zero?
A zero cell does not automatically prevent an exact test. However, effect estimates can become unstable or infinite, and an entirely empty row or column needs careful investigation.
Does a significant result show that training worked?
It indicates evidence against the specified null model. A causal conclusion also requires an appropriate design and consideration of alternative explanations.
Should I report only the p-value?
No. Include the counts, relevant percentages and a suitable effect estimate with uncertainty where available. Explain which comparison and analytical sample produced the result.
Choose the test, then explain the evidence
Fisher’s test can support a nursing analysis of sparse categorical data, provided the observations and research question fit the method. A defensible dissertation combines transparent coding, reproducible software choices and clear reporting of the sample pattern. The final interpretation should explain both the evidence and the limits of what the study can establish.
References
- Fay, M.P. (2010) ‘Two-sided exact tests and matching confidence intervals for discrete data’, The R Journal, 2(1), pp. 53–58. Available at: The R Journal article.
- IBM (n.d.) IBM SPSS Exact Tests. Available at: IBM manual (Accessed: 30 September 2026).
- Kim, H.-Y. (2017) ‘Statistical notes for clinical researchers: Chi-squared test and Fisher’s exact test’, Restorative Dentistry & Endodontics, 42(2), pp. 152–155. Available at: Full article.
- R Core Team (n.d.) Fisher’s Exact Test for Count Data. R documentation. Available at: fisher.test documentation (Accessed: 30 September 2026).
