Nursing theses in the Philippines are built from a small set of designs — a knowledge, attitude and practice survey, a before-and-after health teaching intervention, a comparison of two wards or two groups of patients, a compliance audit — and each of those designs has one right statistical test and several tempting wrong ones. BSN groups pick the wrong one for a specific reason: nursing data is often paired, often ordinal and often small, and the tests taught for large independent survey samples do not fit it.
The table below puts every test a BSN thesis is likely to need against the sub-problem it answers, the data it requires and the nursing study it actually fits. The ranked shortlist follows, with where each test falls short, then one recommendation you can defend in front of a panel that includes a statistician.
The comparison table
| Test | Sub-problem it answers | Data it needs | Typical BSN use | Non-parametric or exact alternative |
|---|---|---|---|---|
| Frequency and percentage | What is the profile of the respondents? | Nominal or ordinal | Age group, sex, ward assignment, years in service, unit type | Not applicable |
| Mean and standard deviation, or median and IQR | What is the level of knowledge, attitude or practice? | Interval, or ordinal for the median | KAP scores, Likert-scaled attitude subscales, compliance checklists | Median and interquartile range for skewed scores |
| Paired samples t-test | Did the same respondents change between pre-test and post-test? | Two interval measures on the same people, differences roughly normal | Knowledge score before and after a health teaching session | Wilcoxon signed-rank test |
| Independent samples t-test | Do two separate groups differ? | Interval outcome, two groups, roughly normal | Anxiety scores of patients in Ward A versus Ward B; staff nurses versus nursing students | Mann-Whitney U test |
| One-way ANOVA | Do three or more groups differ? | Interval outcome, three or more groups | Compliance across three shifts; stress across year levels | Kruskal-Wallis H test |
| Chi-square test of independence | Are two categorical variables related? | Two nominal variables, expected count of at least 5 in most cells | Vaccination status against sex; hand hygiene compliance (yes or no) against unit | Fisher’s exact test for small cells |
| McNemar test | Did a yes-or-no outcome change in the same people? | Paired nominal data | Correct versus incorrect inhaler technique before and after demonstration | Exact McNemar for small samples |
| Pearson r | Is there a relationship between two interval variables? | Both interval, linear, roughly normal | Knowledge score against practice score | Spearman rho |
| Spearman rho | Is there a relationship involving an ordinal variable? | Ordinal or ranked | Pain rating against satisfaction rating; years in service against attitude rank | Not applicable; it is the alternative |
How to read the table against your own sub-problems
Three questions decide the choice, and nursing studies answer the first one differently from every other course.
Are the two sets of numbers from the same people? A pre-test and post-test on the same thirty patients is paired data, and it takes a paired test — the paired t-test, or Wilcoxon if the differences are not normal. Treating a before-and-after design as two independent groups is the single most common statistical error in BSN theses, and it usually hides a real effect. Two different wards, two different hospitals, staff versus students: those are independent groups.
What did your instrument actually produce? A knowledge test scored right or wrong gives an interval total. A single Likert item is ordinal. A checklist observed as done or not done is nominal. A pain scale from 0 to 10 is usually treated as ordinal. The test has to match the level of measurement, and a panel statistician will ask you what level your outcome variable is before anything else.
How many respondents do you really have? Nursing samples are small because wards are small. Thirty patients, twenty-five staff nurses, one batch of clinical instructors. Small samples fail normality tests, empty the cells of a chi-square table and make the exact and non-parametric alternatives the honest choice rather than the fallback. Plan for them from the start.
The ranked shortlist for BSN studies
1. Paired samples t-test, with Wilcoxon signed-rank as the planned alternative
The health teaching intervention is the signature BSN design: a knowledge or practice score measured before a demonstration, lecture or leaflet, and again after. Every one of those studies is a paired comparison. Run the Shapiro-Wilk test on the difference scores, not on the pre-test and post-test separately; if the differences are roughly normal, report the paired t-test with the mean difference and its confidence interval, and if they are not, report Wilcoxon with the median difference.
Suits: pre-test and post-test on the same respondents, whether patients, mothers, caregivers or student nurses. Falls short: it cannot tell you whether the change was caused by your intervention rather than by time, retesting or the attention itself, because there is no control group. Say so in your limitations; a panel respects the honesty and punishes the overclaim.
2. Chi-square test of independence, with Fisher’s exact test ready
Nursing outcomes are often binary: complied or did not, immunised or not, correct technique or not, readmitted or not. Chi-square answers whether a binary outcome is related to a categorical grouping such as unit, shift or sex. The expected count rule matters more here than in any other course, because a 2-by-2 table from thirty respondents routinely has an expected cell below five. When it does, report Fisher’s exact test instead. Both are in SPSS under Crosstabs, and the menu path is in the walkthrough on analysing thesis data in SPSS step by step.
Suits: category against category. Falls short: it discards ordering, so do not collapse a knowledge score into pass or fail just to run a chi-square; use the score.
3. Independent samples t-test, with Mann-Whitney U as the planned alternative
Comparing two genuinely separate groups — medical versus surgical ward patients, day versus night shift nurses, rural versus urban health centres — is a two-group difference question. Check normality within each group and the homogeneity of variance, then report the t-test, or Mann-Whitney U if the assumptions fail. The decision rule and the wording for the switch are set out in the guide on what to do when thesis data is not normally distributed.
Suits: two independent groups on one interval outcome. Falls short: it is misused on paired designs, and it cannot handle three groups; three groups is ANOVA or Kruskal-Wallis, never three t-tests.
4. Spearman rho for the KAP relationship questions
Almost every knowledge, attitude and practice study ends with a sub-problem asking whether knowledge is related to practice, or attitude to practice. Knowledge is usually an interval total; attitude and practice are often Likert composites that behave as ordinal, and practice is sometimes a self-report checklist that is heavily skewed. Spearman rho handles all of those without apology. Use Pearson r only when both variables are interval and roughly normal, and report the coefficient with its direction, strength and p-value.
Suits: relationship questions in KAP designs. Falls short: a correlation is not a cause, and a panel will ask you to say so in Chapter 5.
5. McNemar test — the test BSN groups do not know they need
When the outcome is yes-or-no and the design is before-and-after on the same people — correct handwashing technique observed before and after a demonstration, presence or absence of a symptom before and after an intervention — neither chi-square nor the paired t-test is right. The McNemar test is the paired equivalent of chi-square, and it is the test that turns a common BSN design from statistically wrong to statistically clean. SPSS runs it under Crosstabs with the McNemar box ticked, or under Nonparametric Tests for related samples.
Suits: paired binary outcomes. Falls short: it needs a reasonable number of respondents who changed in either direction; if almost nobody changed, the test has nothing to work with, and the descriptive count is your finding.
The recommendation
For the typical BSN thesis — a KAP survey with a before-and-after teaching component, thirty to sixty respondents, one clinical setting — use frequency and percentage for the profile; mean and standard deviation, or median and interquartile range, for each KAP level; a paired t-test or Wilcoxon signed-rank test for the pre-test and post-test comparison; and Spearman rho for the knowledge-attitude-practice relationships. Run Shapiro-Wilk first on every score that feeds a parametric test, and write the alternative into your Chapter 3 before you collect a single questionnaire, so that switching later is a planned branch and not a last-minute apology.
Reporting the tests the way a nursing panel expects
Nursing panels read Chapter 4 for clinical meaning, not just significance. For a paired comparison, report the pre-test mean, the post-test mean, the mean difference, the statistic, the p-value and one sentence on the size of the change; a three-point rise on a twenty-item test is significant and modest, and the sentence that says so is the one the panel waits for.
For every test, name the sub-problem it answers, state the level of significance you set, and report the actual p-value rather than only the words significant or not significant. The anatomy of the Chapter 3 section that lists these decisions is set out in the explainer on what statistical treatment of data means in a thesis, and the table conventions for presenting them are in the guide to writing Chapter 4 table by table.
Three nursing-specific traps
Treating the pain scale as interval. A 0-to-10 numeric rating scale is widely analysed with means in clinical papers, but a strict panel statistician will call it ordinal. Report the median and use Wilcoxon or Mann-Whitney unless your adviser explicitly approves treating it as interval, and say which you did.
Comparing percentages without a test. Saying that compliance rose from 60 per cent to 80 per cent is a description. Whether that rise is more than chance in thirty respondents is a McNemar question, and the answer is frequently no. Do not let a striking percentage stand in for a test.
Running a t-test on a single Likert item. One item rated from 1 to 5 is ordinal, however many respondents answered it. Sum the subscale into a composite before you treat it as interval, and if the sub-problem is about one item, use Mann-Whitney or Wilcoxon and report the median.
Which software to use
SPSS runs every test in the table, including McNemar and Fisher’s exact, and is what most Philippine nursing research offices expect to see named. JASP is free, open source and produces APA-style tables, and jamovi has a free cloud tier that suits a group sharing one laptop. Excel with the Analysis ToolPak covers the t-tests, ANOVA, Pearson r and descriptives, but it does not document the non-parametric tests, McNemar or Fisher’s exact — the trade-offs are laid out in the head-to-head on SPSS versus Excel for thesis statistics.
Get the statistical treatment section written tonight
Choosing the test is judgement. Writing the Chapter 3 paragraph that names each test, ties it to a numbered sub-problem, states the decision rule and the planned alternative, in the wording a nursing panel expects, is drafting — and it is where BSN groups lose the evening before submission. Tesify drafts that section from your own sub-problems, keeps the test names identical between Chapter 3 and Chapter 4, and formats your references in APA 7th as you go. Over 9,000 students have used it to write more than 15,000 chapters, and every word is still written by you.
Start your nursing thesis on Tesify · Check your manuscript before the research office does
Frequently asked questions
What statistical treatment is used in a nursing thesis?
Most BSN theses use frequency and percentage for the profile, mean and standard deviation or median and interquartile range for knowledge, attitude and practice levels, a paired t-test or Wilcoxon signed-rank test for pre-test and post-test comparisons, and Spearman rho or Pearson r for relationships. Chi-square, Fisher’s exact and McNemar cover categorical outcomes.
Paired t-test or independent t-test for a pre-test and post-test?
Paired. The same respondents were measured twice, so the two sets of scores are related and a paired test is required. Using an independent t-test on a before-and-after design is a common error that usually hides a real effect.
When do I use Wilcoxon instead of the paired t-test?
When the Shapiro-Wilk test on the difference scores rejects normality, when the outcome is ordinal such as a single Likert item or a pain rating, or when the sample is very small. Report the median difference rather than the mean difference when you switch.
What is the McNemar test and when does a nursing study need it?
McNemar is the paired version of chi-square. Use it when a yes-or-no outcome is observed in the same people before and after an intervention, such as correct technique before and after a demonstration. Neither chi-square nor the paired t-test is appropriate for that design.
My chi-square table has an expected count below five. What do I do?
Report Fisher’s exact test instead. Small nursing samples routinely produce a two-by-two table with a low expected count, and Fisher’s exact test is the correct replacement, available in SPSS under Crosstabs.
Should I use Pearson r or Spearman rho for knowledge and practice scores?
Use Pearson r only when both scores are interval and roughly normal. Knowledge totals often are; attitude and practice composites often are not. When in doubt, Spearman rho is defensible for both and is the safer choice for skewed self-report practice scores.
How many respondents does a BSN thesis need for these tests?
Enough to satisfy the sub-problem, not a fixed number. Thirty paired observations support a paired t-test or Wilcoxon reasonably well; a chi-square needs expected cell counts of at least five; regression needs roughly ten to fifteen respondents per predictor, which most ward-based samples cannot provide.