Dissertation statistics help usually starts with one question: which test answers my research question? The answer depends on what you are comparing or predicting, how your variables are measured and whether your data meet the test's assumptions.
This guide gives you a decision table for common tests, explains assumptions, power and effect sizes, shows how to report results in APA style and lists the errors that most often cost marks.
Choosing a Statistical Test in Three Questions
Ask three things. Are you comparing groups or testing a relationship? How is your outcome measured (continuous, ordinal or categorical)? How many groups or predictors do you have, and are the groups independent or the same people measured twice?
Then check the test's assumptions, report an effect size with every test and interpret results in plain language. A p-value alone never answers a research question.
Matching Your Question to a Test
This table covers the tests most often used in master's and doctoral dissertations. It is a starting point; your design may need a more specialized model.
| Research aim | Outcome variable | Parametric test | Non-parametric alternative |
|---|---|---|---|
| Compare two independent groups | Continuous | Independent samples t-test | Mann-Whitney U |
| Compare the same people twice | Continuous | Paired samples t-test | Wilcoxon signed-rank |
| Compare three or more independent groups | Continuous | One-way ANOVA | Kruskal-Wallis |
| Compare the same people three or more times | Continuous | Repeated measures ANOVA | Friedman test |
| Association between two categorical variables | Categorical | Chi-square test of independence | Fisher's exact test (small expected counts) |
| Relationship between two variables | Continuous | Pearson correlation | Spearman's rho |
| Predict an outcome from several variables | Continuous | Multiple linear regression | Robust or bootstrapped regression |
| Predict a yes/no outcome | Binary | Binary logistic regression | Not usually needed |
Fisher's exact test is not strictly a non-parametric version of chi-square, but it is the usual choice when expected cell counts are small. Mediation, moderation, multilevel models and structural equation modeling build on these foundations and need extra planning.
Preparing and Cleaning Your Data
Most analysis problems start before the first test. A clean, well-documented dataset saves hours and protects you when an examiner asks how a figure was produced.
- Keep an untouched copy of the raw data and work on a duplicate.
- Name variables clearly and record value labels, such as 1 = strongly disagree.
- Reverse-score negatively worded items before building scale totals.
- Check for impossible values, duplicates and out-of-range entries.
- Record how much data is missing and decide in advance how to handle it.
- Screen for outliers and decide, with a stated rule, whether to keep them.
Write these decisions down as you make them. They belong in your methodology or results chapter, and they are much harder to reconstruct later.
Checking Assumptions Before You Trust a Result
Every parametric test makes assumptions about your data. Examiners expect you to report how you checked them and what you did if one was violated.
| Assumption | Applies to | How to check | If violated |
|---|---|---|---|
| Normality of residuals or group scores | t-tests, ANOVA, regression | Q-Q plots, histograms; Shapiro-Wilk with care in large samples | Transform, use a non-parametric test or bootstrap |
| Homogeneity of variance | t-tests, ANOVA | Levene's test | Welch's t-test or Welch's ANOVA |
| Independence of observations | Most tests | Design logic | Use repeated measures or multilevel models |
| Linearity | Correlation, regression | Scatterplots, residual plots | Add terms or transform |
| Multicollinearity | Multiple regression | VIF and tolerance | Remove or combine predictors |
| Expected counts | Chi-square | Expected frequencies table | Fisher's exact test or merge categories |
Formal normality tests become oversensitive with large samples and underpowered with small ones, so many statisticians recommend relying mainly on plots. Thresholds such as "VIF above 10" are rules of thumb, and some authors use lower cut-offs.
Sample Size and Power Analysis
A power analysis tells you how many participants you need to detect an effect of a given size. Most dissertations use a significance level of .05 and power of .80, though some fields expect .90.
For example, an independent samples t-test looking for a medium effect (d = 0.5) at alpha .05 and power .80, two-tailed, needs about 64 participants per group. Free tools such as G*Power run these calculations for most common tests.
Base the expected effect size on previous studies or the smallest effect that would matter in practice, not on a small pilot. Report the calculation in your methodology so examiners can follow it.
Effect Sizes and What They Mean
Statistical significance tells you whether an effect is unlikely under the null hypothesis; effect size tells you how large it is. APA style expects both.
| Effect size | Used with | Conventional small / medium / large |
|---|---|---|
| Cohen's d | t-tests | 0.2 / 0.5 / 0.8 |
| Pearson's r | Correlation | .10 / .30 / .50 |
| Eta squared or partial eta squared | ANOVA | .01 / .06 / .14 |
| R squared | Regression | Interpreted in context |
| Odds ratio | Logistic regression | Interpreted in context |
These benchmarks come from Cohen's guidelines and are widely used, but Cohen himself warned they are rough. A "small" effect can matter greatly in health or education, so interpret size in the context of your field.
Reporting Results in APA Style
APA 7th edition sets clear conventions: italicize statistical symbols, give exact p-values to two or three decimals (p < .001 for very small values), omit the leading zero for values that cannot exceed 1, and include an effect size.
Independent t-test. Nurses on self-rostered shifts reported lower burnout (M = 2.41, SD = 0.62) than those on rotating shifts (M = 2.83, SD = 0.71), t(58) = 2.41, p = .019, d = 0.63.
Multiple regression. The model explained 34% of the variance in job satisfaction, R² = .34, F(3, 196) = 33.65, p < .001. Perceived autonomy was the strongest predictor, β = .41, p < .001.
Chi-square. Intention to leave was associated with contract type, χ²(1, N = 180) = 6.12, p = .013, φ = .18.
All figures above are invented for illustration. In your dissertation, every number must come from your own output, and symbols such as t, p, M and SD should be italicized.
SPSS, R, Stata or Something Else?
Most universities accept any recognized package. Choose the one your department supports and your supervisor can help with.
| Software | Strengths | Watch out for |
|---|---|---|
| SPSS | Menus, widely taught in social sciences and health | License cost; advanced models can be awkward |
| R | Free, powerful, reproducible scripts | Steeper learning curve |
| Stata | Strong for econometrics and panel data | License cost |
| jamovi or JASP | Free, point-and-click, APA-style tables | Fewer advanced options |
| Python | Flexible, good for large or messy data | More coding for standard tests |
Whatever you use, cite the software and version in your methodology, and keep your syntax or script. It makes your analysis reproducible and helps if an examiner asks how a figure was produced.
Common Statistical Errors Examiners Notice
| Error | Fix |
|---|---|
| Running many tests without correction | Plan comparisons in advance; use post hoc tests or adjust for multiple comparisons |
| Treating correlation as causation | Use cautious language unless the design is experimental |
| "Approaching significance" for p = .07 | Report the result as non-significant and discuss the effect size |
| Summing Likert items without checking reliability | Report Cronbach's alpha or omega for multi-item scales |
| Pasting raw software output into the chapter | Build clean APA tables and describe them in text |
| Ignoring missing data | Report how much was missing and how it was handled |
How DissertationHorse Helps with Dissertation Statistics
Send your research questions, hypotheses, dataset or software output and your methodology chapter. Your writer can recommend and justify tests, check assumptions, run analysis on the data you supply and write the results chapter in APA or your required style.
We never fabricate, alter or "improve" data or results. Choose a Standard, Master or Elite writer tier, see the price before you pay and get free revisions within the scope of your original request.
Want Your Analysis Planned or Written up Properly?
Send your research questions, dataset or output and any analysis plan. Your writer checks the test choice and assumptions and writes up results accurately in your required style.
Order Your Statistics HelpFree revisions · Full refund if late · Written from scratch for your order
Frequently Asked Questions
It depends on your research question, how the outcome is measured and the number and type of groups or predictors. Use the decision table above, then confirm with your supervisor.
Not always. t-tests and ANOVA are fairly robust with larger, similar-sized groups, and the normality assumption often concerns residuals rather than raw scores. Check plots, then decide; non-parametric tests or bootstrapping are options.
This is debated. Single Likert items are ordinal. Summed or averaged scales of several items are often analyzed with parametric tests, which is common practice, but justify your choice.
Run a power analysis based on your test, expected effect size, alpha and desired power. For a two-group comparison with a medium effect, about 64 per group is needed at power .80.
Report them honestly with effect sizes and confidence intervals, and discuss possible reasons such as power. Non-significant results are still findings.
Two-tailed is the default and safest choice. Use one-tailed only with a strong directional hypothesis stated before the analysis, and justify it.
Usually not. Most universities accept SPSS, R, Stata, jamovi or similar. Check whether your program or supervisor prefers one.
Yes, using data you have collected. A writer runs the agreed analysis and explains it so you can understand and defend every result.