Which Statistical Test Should You Use? An SPSS Guide for PhD Scholars

Start with the question, not the software

Every test answers one of four question types: is there a difference between groups, is there a relationship between variables, does one variable predict another, or does a structure underlie a set of items. Naming your question type first eliminates most of the options.

Then check your variables

Write down each variable and its measurement level: nominal, ordinal, interval or ratio. How many groups are you comparing? Are the groups independent, or the same people measured twice?

The decision table

Check assumptions before you trust the output

• Normality: Shapiro-Wilk for samples under 50, otherwise skewness and kurtosis between -2 and +2, plus a histogram.
• Homogeneity of variance: Levene's test, reported with ANOVA and the t-test.
• Linearity: scatterplots before any correlation or regression.
• Multicollinearity: VIF under 5 (some argue under 10) in multiple regression.
• Independence of residuals: Durbin-Watson between roughly 1.5 and 2.5.
• Reliability: Cronbach's alpha of 0.70 or above before you compute scale scores.

If an assumption fails, you have three honest options: transform the variable, switch to the non-parametric equivalent, or use a robust method. Ignoring the failure is not one of them.

Reporting results the way examiners expect

Report the test, the statistic, degrees of freedom, the exact p value and an effect size. For example: t(148) = 3.42, p = .001, d = 0.56, followed by one sentence in plain language explaining what it means for your research question. An effect size with no interpretation, or a p value with no effect size, will be queried in the viva.

The mistakes that cost marks

• Running t-tests repeatedly instead of one ANOVA, which inflates the error rate.
• Reading significance without checking assumptions first.
• Treating a correlation as proof of causation.
• Reporting "p = .000", which should be written as p < .001.
• Deleting outliers with no stated rule.

What examiners ask about your analysis

In a viva, the statistics questions are predictable. Prepare answers for these six:

1. Why this test rather than the obvious alternative?
2. How did you check the assumptions, and what did you find?
3. What is the practical meaning of this effect size?
4. How did you handle missing data, and why that way?
5. What did you do about outliers, and what rule did you use?
6. What would change if the assumption you relaxed had held?

Missing data and outliers, handled defensibly

• Report how much data is missing and whether it looks random. Under 5% at random is usually manageable.
• Prefer multiple imputation or maximum likelihood over deleting cases, and say which you used.
• Define an outlier rule before you look at the data — for example, standardised residuals beyond ±3.29 — and report cases removed with the reason.

When SPSS is not the right tool

SPSS handles most survey-based doctoral analysis well. Move to AMOS for structural equation modelling with latent variables, to R or Python for large datasets, custom models, machine learning or reproducible pipelines, and to NVivo or similar for qualitative coding. If your analysis involves prediction rather than explanation, see our guide on AI and machine learning research, where the evaluation rules are different.

Frequently asked questions

Which SPSS test should a PhD scholar use for Likert-scale data?

If a PhD scholar analyses individual Likert items, they are treated as ordinal and use non-parametric tests. If you sum several items into a validated scale, the total is usually treated as continuous, which allows t-tests, ANOVA and regression.

What sample size does a PhD researcher need for regression in SPSS?

For doctoral research, a common rule is at least 10 to 15 cases per predictor, with 100 or more cases preferred. For a defensible number, run a power analysis in G*Power and report it.

What should a research scholar do if the data is not normally distributed?

Use the non-parametric equivalent of your test, transform the variable, or use bootstrapping. State clearly in your methods chapter which route you took and why.

Do I need to report effect sizes in a PhD thesis?

Yes. A p value tells you whether an effect is unlikely under the null hypothesis; it does not tell you how large or important the effect is. Examiners and journals now expect both.

How should a PhD scholar handle missing data in SPSS?

Report the amount and pattern of missingness, then use multiple imputation or maximum likelihood estimation where appropriate. Listwise deletion is acceptable only when missingness is minimal and random, and you should say so explicitly.

Is SPSS still accepted for PhD research in 2026?

Yes. SPSS remains standard for survey-based research in management, social sciences, education and health. R and Python are preferred where reproducibility, large data or custom modelling matter.