Which Statistical Test Should a Medical Dissertation Use? The 2026 Guide to Clinical Data, Diagnostic Accuracy and Survival

Medical dissertations fail their statistics section for a narrower set of reasons than students expect. It is rarely because the wrong exotic model was fitted. It is because a paired design was analysed as if it were unpaired, because a time-to-event outcome was reduced to a proportion, because an odds ratio was described as a risk, or because 30 significance tests were run on one dataset without a word about multiplicity.

This guide maps the clinical study designs a medical student actually runs — retrospective case note review, a cross-sectional survey of patients or clinicians, a diagnostic accuracy study, an audit-turned-project, a small prospective cohort — onto the analysis each one requires, and sets out what the reporting guidelines expect alongside the p value.

Key points first

• The outcome variable dictates the model family: continuous outcomes go to t tests, ANOVA or linear regression; binary outcomes to chi-square and logistic regression; time-to-event outcomes to Kaplan-Meier and Cox regression; count outcomes such as admissions per year to Poisson or negative binomial regression.
• Whether the observations are paired or independent changes the test entirely, and paired designs are far more common in medicine than students realise — pre and post, left and right, matched controls, repeat measures on the same patient.
• Diagnostic accuracy studies use sensitivity, specificity, predictive values and the area under the ROC curve, not a t test comparing means.
• Effect sizes with confidence intervals matter more than p values in every modern reporting guideline. A mean difference of 2 mmHg with a 95% CI of −1 to 5 is a finding; “p = 0.31” alone is not.
• Case-control designs cannot produce risk ratios or incidence. They produce odds ratios. Cross-sectional studies produce prevalence, not incidence.

Start with three questions, not with SPSS

Before opening any software, answer these in writing:

1. What is my outcome and what type of variable is it? Continuous, binary, ordinal, count, or time-to-event.
2. How many groups am I comparing, and are the observations independent? Two independent groups, two paired measurements, three or more groups, or a continuous exposure.
3. Am I comparing, predicting, or measuring agreement? These are three different statistical questions and only the first is a hypothesis test.

Those three answers determine the analysis almost completely. If you want the general version of this logic across all disciplines, our decision guide to choosing a statistical test works through the same tree with worked examples.

Matching the design to the test

Comparing a continuous outcome between two independent groups

Examples: mean HbA1c in patients on two different regimens; length of stay in two hospitals; BMI in cases versus controls.

Use an independent-samples t test if the outcome is approximately normally distributed within each group. Report the mean difference with a 95% confidence interval, not just the p value, and state whether you used Welch’s correction for unequal variances — with unequal group sizes, which is the norm in retrospective clinical data, Welch’s version is the safer default. Our step-by-step walkthrough of running an independent-samples t test in SPSS covers Levene’s test, Cohen’s d and the write-up.

If the outcome is markedly skewed — length of stay, waiting times, biomarker concentrations and hospital costs almost always are — the Mann-Whitney U test compares distributions instead, and you report medians with interquartile ranges. Our overview of non-parametric tests explains what Mann-Whitney actually tests and how to phrase the conclusion correctly.

Comparing a continuous outcome measured twice in the same patients

Examples: blood pressure before and after an intervention; pain score at admission and discharge; measurements taken on both eyes.

This is a paired design and requires a paired-samples t test or, if the differences are skewed, the Wilcoxon signed-rank test. Analysing paired data as if it were independent is one of the most common and most heavily penalised errors in medical student projects; it throws away the very thing that makes the design efficient. Our guide to running a paired-samples t test in SPSS shows the data layout that prevents the mistake, since the two designs require different spreadsheet structures.

Comparing three or more groups

Examples: mean recovery time across three surgical approaches; symptom score across four disease stages.

One-way ANOVA with a post-hoc correction (Tukey when group sizes are similar, Games-Howell when variances differ), or the Kruskal-Wallis test with Dunn’s post-hoc for skewed outcomes. If patients were measured at several time points, you need a repeated-measures approach instead, and sphericity becomes an issue — our walkthrough of repeated-measures ANOVA, Mauchly’s test and Greenhouse-Geisser covers the corrections examiners expect to see named. If you are unsure which member of the family your design needs, our comparison of one-way, two-way, repeated-measures, mixed and ANCOVA designs lays them out side by side. ANCOVA in particular is the right choice when you want to compare outcomes while adjusting for baseline severity.

Comparing proportions

Examples: complication rate in two groups; proportion achieving target INR; readmission within 30 days.

Pearson’s chi-square test of independence, with Fisher’s exact test when expected cell counts fall below five — which happens constantly in small clinical samples. For paired proportions, such as a diagnosis before and after a new protocol in the same patients, use McNemar’s test. Always report the absolute difference in proportions with a confidence interval alongside the test, because a relative reduction of 50% means something very different at a baseline risk of 40% than at 0.4%.

Predicting a binary outcome from several variables

Examples: predictors of 30-day mortality; factors associated with treatment non-adherence; determinants of a positive culture.

Binary logistic regression, reporting adjusted odds ratios with 95% confidence intervals. Two disciplines matter here. First, respect the events-per-variable guidance: with fewer than roughly 10 outcome events per predictor, your model is overfitted and the confidence intervals are unreliable. A retrospective review with 43 patients and 9 deaths supports one predictor, not six. Second, do not interpret an odds ratio as a risk ratio when the outcome is common; at 30% baseline risk an odds ratio of 2.0 corresponds to a risk ratio closer to 1.5. Our comparison of linear versus logistic regression covers how to choose between them, and if your outcome has ordered categories such as mild, moderate and severe, see ordinal versus multinomial logistic regression.

Time-to-event outcomes

Examples: survival after diagnosis; time to relapse; time to discharge; graft failure.

This is where medical dissertations diverge most sharply from other disciplines. If some patients have not yet had the event when the data ends, or were lost to follow-up, the observation is censored — and converting the outcome to “died: yes/no at two years” discards both the timing and the censoring, which is a substantive analytical error rather than a simplification. Use Kaplan-Meier curves with the log-rank test to compare groups, and Cox proportional hazards regression to adjust for covariates, reporting hazard ratios with confidence intervals and stating how you checked the proportional hazards assumption. Our full walkthrough of Kaplan-Meier, the log-rank test and Cox regression covers the numbers-at-risk table that reviewers expect beneath every survival plot.

Diagnostic and screening accuracy

Examples: does a point-of-care test agree with the laboratory reference standard? Does a risk score identify sepsis?

Build a 2×2 table against the reference standard and report sensitivity, specificity, positive and negative predictive values, and likelihood ratios, each with a confidence interval. For a continuous index test, plot the ROC curve and report the area under it with a confidence interval, then justify any threshold you propose — Youden’s index is the usual criterion. Remember that predictive values depend on prevalence, so a PPV calculated in a highly selected hospital sample does not transfer to a screening population. Report against the STARD checklist, and include the flow of participants.

Agreement and reliability

Examples: two clinicians scoring the same radiographs; a new device versus an established one.

Agreement is not correlation. Two instruments can correlate at 0.99 while one reads consistently 10 units higher. For categorical ratings use Cohen’s kappa (weighted kappa for ordered categories); for continuous measurements use the intraclass correlation coefficient and a Bland-Altman plot showing bias and limits of agreement. Presenting a Pearson correlation as evidence of agreement is a classic error that examiners spot immediately.

Counts and rates

Examples: number of exacerbations per patient-year; attendances per month.

Poisson regression, or negative binomial regression when the variance exceeds the mean, which it usually does in clinical count data. Include an offset for exposure time when follow-up varies between patients.

Data that is not independent

Much clinical data is clustered: several measurements per patient, several patients per consultant, several consultants per hospital. Standard tests assume independent observations, and violating that assumption produces confidence intervals that are too narrow and p values that are too small. If your design has nesting of this kind, a mixed-effects model is the correct answer, and our introduction to multilevel models for nested data explains when the added complexity is genuinely necessary and when a simpler summary-measures approach will do.

Reporting: what the guidelines require

Name the relevant reporting guideline in your methods and follow it. STROBE governs observational studies (cohort, case-control, cross-sectional), CONSORT randomised trials, STARD diagnostic accuracy studies, and PRISMA systematic reviews. Whichever applies, your statistics section should state:

• The software and version used.
• How continuous variables were summarised (mean and SD, or median and IQR) and why.
• Every test used, with the reason it matched the data.
• How missing data were handled — complete-case analysis is acceptable if declared and justified, silent deletion is not.
• The significance threshold and any adjustment for multiple comparisons.
• Effect estimates with 95% confidence intervals throughout.

Report exact p values to three decimal places (p = 0.032, not p < 0.05), and write p < 0.001 only below that threshold. Reporting errors of this kind are surprisingly widespread even in published work, as our analysis of statistical reporting errors in the literature documents — which is exactly why examiners check yours.

Three traps specific to medical dissertations

Turning an audit into research after the fact. The classification of your project as research, service evaluation or clinical audit determines the ethics route, and retrofitting hypotheses to audit data creates governance problems as well as statistical ones. Our explainer on the research, service evaluation and audit distinction sets out where the line falls.

Testing everything in the dataset. Running 40 comparisons at α = 0.05 will produce two significant results by chance alone. Pre-specify your primary outcome, label everything else exploratory, and consider a Bonferroni or false-discovery-rate correction where you have a genuine family of tests.

Treating a non-significant result as proof of no effect. In a sample of 60 patients, absence of evidence is overwhelmingly likely. Report the confidence interval, discuss what effect sizes your study could and could not have detected, and interpret accordingly — our guide on what to do when results are not significant shows how to write that up as a credible finding rather than an apology.

Frequently asked questions