Nearly every SPSS problem we are sent falls into one of three categories: the wrong test was chosen for the design, an assumption was never checked, or the output was pasted into the document raw. None of them are difficult to avoid once you know what to look for.
This guide walks through the decisions in the order you actually meet them.
Start with your variables, not with the test
The test is determined by three things: the measurement level of your outcome variable, how many groups or predictors you have, and whether your observations are independent or repeated. Decide those first and the choice usually makes itself.
In SPSS, measurement level is set in Variable View. It is worth being accurate there, because SPSS uses it to decide which procedures it will offer you.
| What you want to know | Outcome type | Usual test |
|---|---|---|
| Difference between two independent groups | Continuous | Independent-samples t-test |
| Difference between two related measures | Continuous | Paired-samples t-test |
| Difference across three or more groups | Continuous | One-way ANOVA |
| Difference across groups, controlling a covariate | Continuous | ANCOVA |
| Relationship between two continuous variables | Continuous | Pearson correlation |
| Predicting a continuous outcome | Continuous | Multiple linear regression |
| Predicting a yes/no outcome | Binary | Binary logistic regression |
| Association between two categorical variables | Categorical | Chi-square test of independence |
Check the assumptions that matter for your test
Assumption checking has a reputation for being a box-ticking ritual. It is not — a violated assumption can change your conclusion entirely, and a marker who spots an unchecked one will doubt everything after it.
What matters is that you check the assumptions of the test you actually ran, report what you found, and say what you did about any violation.
- Normality — for t-tests and ANOVA this applies to the residuals, not the raw data. Shapiro-Wilk is over-sensitive in large samples; look at a Q-Q plot alongside it.
- Homogeneity of variance — Levene's test. If it is significant, use Welch's correction, which SPSS reports for you.
- Linearity — for regression and correlation, plot it. A strong curvilinear relationship can produce a correlation near zero.
- Multicollinearity — VIF above about 5 (some use 10) means your predictors are explaining the same variance and the coefficients become unstable.
- Independence — this one comes from your design, not from a test. Repeated measures or clustered data need a model that accounts for the structure.
A violated assumption is not a failed project. It is a finding you report and respond to — with a robust alternative, a transformation, bootstrapping, or a non-parametric equivalent.
Reading the output without over-reading it
A p-value below .05 tells you the result is unlikely under the null hypothesis. It does not tell you the effect is large, or important, or causal.
This is why effect sizes are not optional. APA requires them, and they are what makes the result meaningful. A statistically significant correlation of r = .09 in a sample of 5,000 is real and almost certainly irrelevant.
- Report the test statistic, degrees of freedom, exact p-value and effect size every time.
- Use Cohen's d for t-tests, partial η² for ANOVA, and R² with standardised betas for regression.
- Give confidence intervals where you can — they show precision in a way a p-value never does.
- Write p = .032, not p < .05. Report p < .001 only when it genuinely is.
Turning SPSS output into APA tables
SPSS output pasted straight into a dissertation is one of the most common presentation failures. SPSS tables carry columns you do not need, formatting that is not APA, and decimal places that imply precision you do not have.
Build the table yourself in Word or Excel with only the columns the reader needs.
- Include only relevant columns — drop the ones your reader will not use.
- Round to two decimal places; three for p-values below .01.
- Remove leading zeros from values that cannot exceed 1 (write .05, not 0.05).
- Use APA table rules: no vertical lines, horizontal lines only above and below the header and at the foot.
- Give every table a number and a descriptive title, and refer to it in the text.
Always save the syntax
Every SPSS dialog has a Paste button. Use it. It writes the command into a syntax window instead of running it immediately.
This matters for two reasons. It makes your analysis reproducible, which is what "rerun everything with the two outliers removed" actually requires. And at a viva it means you can show exactly what you did rather than trying to recall which menu options you clicked eight months ago.
Keep three files together: the raw data exactly as collected, the syntax that cleans and analyses it, and the output. That set is reproducible; a single edited .sav file is not.
The mistakes that cost the most marks
- Running a parametric test on ordinal data because it is the familiar procedure.
- Deleting outliers without justification, or keeping them without acknowledgement.
- Reporting significance without effect size.
- Interpreting a correlation or regression coefficient as causal in a cross-sectional design.
- Running dozens of tests and reporting only the significant ones without correcting for multiplicity.
- Letting SPSS delete cases listwise without reporting how many were lost.
Questions this raises
First check whether the test actually assumes normality of the raw data — for t-tests and ANOVA it is the residuals that matter, and with reasonably large groups those tests are robust to moderate departures. If the violation is genuine, use a non-parametric equivalent, a transformation, or bootstrapped confidence intervals, and report what you did.
Run a power analysis before collecting data, using your test, your alpha, the power you want (usually .80) and the smallest effect size worth detecting. G*Power is free and does this well.
If your department teaches and expects it, yes — using it removes a whole category of friction. For reproducible or unusual analysis, R and Python are better tools.
Still stuck after reading this? That is usually the point at which it is worth asking someone. Describe your project or ask on WhatsApp.
