The right statistical test for a final year project depends on the question, the study design, and the kind of data you actually collected. Start with those, not the test your classmate used. Some projects need only descriptive tables, some need an inferential test, and a qualitative project may need no statistical test at all.
This guide gives you a four-step decision path: data type, number of groups, paired or independent, and whether the assumptions of a parametric test hold. Treat it as a discussion aid for your supervisor or departmental statistician, not an automatic prescription.
Your analysis needs a written account that matches the test you actually ran. UniDraft helps organise your confirmed objectives, methods and real results into an editable project document. Start your project write-up, then verify every statistic and interpretation. See the project workflow; it does not choose or execute a test for you.
The short answer
- What type is your outcome? Categories (Yes/No, faculty, level) point towards counts, percentages and chi-square. Scores or measurements (marks, time, a summed scale) open up t-tests, ANOVA, correlation and regression.
- How many groups or variables? Two groups, three or more groups, or a relationship between two measurements.
- Are the observations independent or paired? The same students measured before and after are paired. Two different classes are not automatically independent observations: students can be clustered within classes.
- Do the assumptions hold? If not, review the design and estimand with a methods adviser; a different model, robust approach or rank-based test may be appropriate.
Start with the question shape
| Your question asks | Data and design to check | Possible analysis |
|---|---|---|
| How many or what proportion? | Categories and a defined sample | Counts and percentages |
| What is the typical score? | How the score was constructed and measured | A suitable summary such as median or mean, with spread |
| Are two categories associated? | Independent observations and expected cell counts | Chi-square test of independence, if its assumptions hold |
| Do two independent groups differ? | Outcome scale, distribution, variance, sample design | A suitable independent-groups model or rank-based comparison |
| Did the same people change? | Paired before-and-after measurements | Paired t-test or Wilcoxon signed-rank |
| Do three or more groups differ? | Independent groups and suitable outcome | One-way ANOVA or Kruskal-Wallis |
| Are two measurements related? | Variable scales and relationship shape | Pearson or Spearman correlation |
| Can one or more variables predict an outcome? | Outcome type, design, assumptions, confounders | An appropriate regression model |
The word possible matters. Pearson correlation needs more than two columns of numbers: the relationship should be roughly linear, observations independent, and extreme outliers examined. A single Likert item is ordinal; treating a constructed score as continuous requires justification. A small expected count can make the usual chi-square approximation unsuitable; exact or other methods may be needed. Check your department's methods guidance and seek statistical review before making a claim from a borderline case.
Work backward from your objectives
Suppose your confirmed objective is "describe the proportion of respondents using a campus booking service." A frequency table may answer it. Adding a t-test does not make that objective more rigorous. If the objective is "compare satisfaction scores between users and non-users," you must define the score, the two groups, the sampling method, and the planned comparison before naming a test.
That is why the test belongs to the objectives, not the other way round. Before you collect anything, settle your research questions and objectives, then make sure your questionnaire actually collects the variables you intend to analyse. You cannot recover an unasked question in Chapter 4. If you have hypotheses, write each one so that it names the variables and the direction of comparison, because the wording usually reveals the test.
Step 1: What type of data do you have?
Your variable type does most of the work in narrowing the choice.
- Nominal (categories with no order): faculty, gender, yes/no, state of origin. You describe these with counts and percentages and relate two of them with chi-square.
- Ordinal (ordered categories): a single Likert item ("Strongly disagree" to "Strongly agree"), class of degree, satisfaction rank. You can report medians and frequencies and use rank-based tests.
- Interval or ratio (real measurements): exam scores, time in minutes, temperature, yield per plot, a validated scale total. These can enter t-tests, ANOVA, Pearson correlation and regression, subject to their assumptions.
Many student projects mix these. A questionnaire might give you demographic categories, several ordinal items, and one composite score. Use a different analysis for each objective; you do not need one test for the whole project. If you build composite scores from several items, the validity and reliability of the scale is what makes treating the total as a measurement defensible.
Step 2: How many groups, or what kind of relationship?
Two groups, one outcome. Male and female students, users and non-users, treatment and control. The candidate is the independent-samples t-test (or Mann-Whitney U as the rank-based alternative).
Three or more groups. Compare 100-, 200- and 300-level students, or four fertiliser rates. Running several t-tests inflates your chance of a false positive, so use one-way ANOVA (or Kruskal-Wallis) and follow up with a suitable post hoc comparison if the overall result is significant.
Two categorical variables. "Is use of the service associated with faculty?" A chi-square test of independence compares observed counts with the counts expected if the variables were unrelated.
Two measurements. "Is hours of study related to CGPA?" Use correlation. Pearson measures a linear relationship between interval variables; Spearman works on ranks and suits ordinal data or a clearly monotonic but non-linear pattern.
Prediction. If you want to estimate how far one or several predictors explain an outcome, you are in regression territory. Simple linear regression has one predictor and a continuous outcome; multiple regression adds predictors; logistic regression handles a yes/no outcome. Regression is easy to run in software and easy to misinterpret, so agree the model with your supervisor first.
Step 3: Independent or paired?
This is the step students most often skip, and it changes the test.
- Independent concerns the observation structure, not merely different names. If teaching is assigned by class, students are clustered within classes; an individual-level independent-samples test may be unsuitable.
- Paired means the same people, or matched pairs, measured twice: the same 40 students before and after training.
For paired numerical data, use the paired t-test (or Wilcoxon signed-rank). It analyses the difference within each pair. Using an independent t-test on paired data ignores that link between measurements. For paired categorical data (for example yes/no before and after), a chi-square test of independence does not apply, because its observations must be independent; McNemar's test is the usual choice, according to Kent State University's SPSS guidance.
Step 4: Check the assumptions
Use the following as prompts for statistical review, not a complete diagnostic rule for every model.
| Test | Main use | Assumptions to check |
|---|---|---|
| Independent t-test | Compare means of two independent groups | Interval outcome, roughly normal within groups, independent observations |
| Paired t-test | Compare two related measurements | Interval variables, paired observations, differences distributional conditions for the chosen inference |
| One-way ANOVA | Compare means across three or more groups | Normal outcome within groups, similar variances, independent observations |
| Chi-square test of independence | Association between two categorical variables | Independent observations; expected counts of about 5 or more in most cells |
| Pearson correlation | Linear relationship between two interval variables | Linear pattern, independent cases, no extreme outliers, roughly normal |
| Linear regression | Predict a continuous outcome | Residuals show no pattern, roughly constant spread, roughly normal, independent errors |
Three practical points:
Normality. UCLA's statistics guidance describes the t-test, ANOVA and correlation as applying to normally distributed interval variables, and offers Wilcoxon-Mann-Whitney, Kruskal-Wallis and Wilcoxon signed-rank tests when you do not want to assume that. In practice, plot a histogram or Q-Q plot first. Very small samples make normality hard to judge, which is one reason your sample size matters before you ever reach the analysis.
Equal variances. ANOVA and the standard t-test assume similar spread across groups. NIST's Engineering Statistics Handbook describes the Levene test for equality of variances, and notes that it is less sensitive than the Bartlett test to departures from normality. Most software also offers a t-test version that does not assume equal variances.
Independence. Kent State's guide notes that Pearson correlation and its significance test are not robust when independence is violated. If your respondents filled the questionnaire in groups, copied each other, or the same person is counted twice, the p-values are not trustworthy. Do not describe a convenience sample as random simply because software produced a p-value.
For regression, NIST's handbook stresses graphical residual analysis as the primary way to judge whether a model fits, and warns that a high R² does not guarantee a good fit. Look at the residual plots; do not stop at the R².
Worked examples (illustrative)
These are invented scenarios to show the reasoning, not real data or results.
Example 1: Two groups, independent. Objective: "compare mean practical-exam scores of students taught with lab simulation and those taught by demonstration." Outcome: score (interval). Two different classes. Data roughly bell-shaped. Candidate: independent-samples t-test. Two classes can introduce clustering and confounding with teaching method. The model and unit of analysis need review; do not automatically substitute Mann-Whitney U.
Example 2: Same group before and after. Objective: "determine whether a four-week digital literacy training changed test scores." Same students, two measurements. Candidate: a paired analysis appropriate to the differences and estimand. A signed-rank test needs its own assumptions, including symmetry for the usual location interpretation; a before/after change alone does not prove the training caused it.
Example 3: Category by category. Objective: "examine whether use of the library e-resource portal is associated with level of study." Two categorical variables. Candidate: chi-square test of independence. If several cells have expected counts below 5 (say, few 400-level users), seek statistical advice about an appropriate exact or other analysis. Do not merge categories merely to obtain a favourable p-value.
Example 4: Relationship. Objective: "assess the relationship between weekly study hours and CGPA." Both recorded numerically, with the measurement and scatterplot assessed. Candidate: Pearson correlation; Spearman if CGPA is banded into classes of degree.
Example 5: Descriptive only. Objective: "identify the most common challenges of SIWES students in Lagos." A ranked frequency table is the analysis. No inferential test is needed.
Report the test properly
Plan the test in Chapter 3, where your methodology should name the test, the software, the significance level, and how you will handle missing responses. In Chapter 4, report the actual statistic, degrees of freedom where applicable, sample size, p-value or confidence interval as your department requires, and a plain-language interpretation tied to the original question. The Chapter 4 guide shows how to present verified findings in tables and figures without dumping raw software output into the report.
Significance is not the whole conclusion
If correlation is appropriate and the analysis has been checked, use the correlation-table interpretation worksheet to distinguish the coefficient, p-value and usable N, avoid counting mirrored cells twice and prepare a bounded result paragraph.
"Statistically significant" does not mean "large," "important," or "caused by." Report the size and direction of an observed difference or relationship where your method permits it. A non-significant result is still a real result: do not erase it or keep trying tests until one crosses the 0.05 line. That misrepresents the analysis, regardless of whether a reviewer notices. Discuss what your findings do and do not show, and their limits, in Chapter 5.
Common mistakes supervisors flag
- Test chosen before the objective is clear. The panel asks "what question does this answer?" and there is no answer.
- Independent t-test on paired data, or several t-tests instead of one ANOVA.
- Chi-square with tiny expected counts, or on percentages instead of counts.
- Treating one Likert item as a precise measurement with no justification.
- Correlation described as causation.
- Reporting only the p-value, with no sample size, statistic or effect direction.
- Changing tests after seeing the results without saying so.
- A tool that is not explained: naming SPSS or Excel is fine, but you should be able to explain what the test does.
Pre-analysis checklist
- Each objective is matched to one planned analysis, written in Chapter 3.
- You have identified the type of each variable (nominal, ordinal, interval).
- You know whether groups are independent or paired.
- You have checked the assumptions and agreed how departures will be handled.
- Your sample size and sampling method are stated honestly.
- Your supervisor or departmental statistician has seen the plan.
- Every number in Chapter 4 comes from your real data.
If your department wants hypotheses, how to write hypotheses for a final year project shows how to word them so the test follows naturally.
Where UniDraft fits in
Check the current price and included correction allowance in the document workflow before payment.
Sources
- UCLA OARC: What statistical analysis should I use?: uses and assumptions of t-tests, ANOVA, chi-square, correlation and rank-based alternatives.
- Kent State University Libraries: Chi-square test of independence: independence, expected counts, McNemar's test for paired categories.
- Kent State University Libraries: Pearson correlation: assumptions of Pearson correlation.
- NIST Engineering Statistics Handbook: One-way ANOVA: purpose and normality assumption.
- NIST Engineering Statistics Handbook: Levene test: testing equality of variances.
- NIST Engineering Statistics Handbook: Residual analysis: checking regression assumptions.