Scholars often treat statistical test selection as a memory problem — a hundred test names to somehow keep straight. It isn't. It's a decision problem with three questions, and once you answer them honestly for your study, the test almost always chooses itself. This guide walks the same decision path a statistician uses (and that our free Statistical Test Selector automates).
Question 1 — What are you actually testing?
- A difference between groups — do groups differ on an outcome? (t-tests, ANOVA and their cousins)
- A relationship between variables — do two variables move together? (correlation, chi-square)
- A prediction — can one or more variables predict an outcome? (regression families)
- A comparison against a known value — does your sample differ from a benchmark or scale midpoint? (one-sample tests)
Question 2 — How many groups, and are they independent?
For difference questions, count the groups (two vs three or more) and ask whether the measurements are independent (different people in each group) or related (the same people measured repeatedly, or matched pairs). Mixing these up is the single most common test-selection error — a before/after comparison of the same respondents needs a paired test, not an independent one.
Question 3 — What type is your outcome variable?
Continuous and roughly normal (scale scores, age, income) unlocks the parametric tests. Ordinal or non-normal data route you to rank-based alternatives. Categorical outcomes (yes/no, categories) call for chi-square-family tests. A single Likert item is ordinal; a multi-item scale score with acceptable reliability is conventionally treated as continuous — see our Likert analysis guide.
The decision map
Comparing two groups
- Independent groups, continuous outcome → Independent-samples t-test (non-parametric: Mann-Whitney U)
- Same participants twice, continuous outcome → Paired-samples t-test (non-parametric: Wilcoxon signed-rank)
- Categorical outcome → Chi-square test of independence (small 2×2 samples: Fisher's exact; paired binary: McNemar)
Comparing three or more groups
- Independent groups, continuous outcome → One-way ANOVA with Tukey post-hoc (non-parametric: Kruskal-Wallis H)
- Repeated measurements, continuous outcome → Repeated-measures ANOVA (non-parametric: Friedman)
Relationships and prediction
- Two continuous variables → Pearson correlation (ordinal/non-normal: Spearman's rho)
- Two categorical variables → Chi-square + Cramér's V
- Predicting a continuous outcome → Multiple linear regression
- Predicting a binary outcome → Binary logistic regression
- Multiple latent constructs and paths → SEM in AMOS or SmartPLS
Check assumptions before you trust the output
Parametric tests assume approximate normality (check skewness/kurtosis, Shapiro-Wilk for small samples), homogeneity of variance (Levene's test) and independent observations. Violations aren't failures — they're routing instructions to the non-parametric alternative. And always report an [effect size](/knowledge-hub/effect-size) alongside the p-value: significance says a difference exists; effect size says whether it matters.
The free Statistical Test Selector walks you through these exact questions and returns the test, its assumptions, the SPSS menu path and the effect size to report — in under a minute.
If your analysis chapter involves several of these tests and you want an experienced statistician checking the logic, our SPSS mentoring and statistics services exist for exactly that.
Frequently asked
What is the most commonly used statistical test in theses?+
For survey-based management and social-science theses: reliability (Cronbach's alpha), descriptives, then t-tests/ANOVA for group differences, Pearson correlation for relationships, and multiple regression for prediction — often capped with SEM when the model has latent constructs.
When should I use a non-parametric test?+
When the outcome is ordinal, the sample is small, or normality is clearly violated (marked skew, outliers). Rank-based tests like Mann-Whitney and Kruskal-Wallis trade a little power for robustness.
Can I run multiple t-tests instead of ANOVA?+
No — running several t-tests across three or more groups inflates the Type I error rate. Run one ANOVA, and only if it is significant, use post-hoc tests (e.g. Tukey HSD) to locate which groups differ.
phdguide's mentors are senior academics, former supervisors, statisticians and publication specialists with 25+ years' average experience guiding MBA, MPhil and PhD scholars from topic to viva.
Ethical, compliant guidance: We provide academic support, mentoring, analysis, editing and structuring — not authorship. Your work stays compliant with university policies.