The question I am asked most often is some version of “which test should I run?” It usually arrives after the data is already collected, which is the wrong end of the process, but it is a very common place to be. Here is the order I work through it in.
Start from the question, not the software
A statistical test is an answer to a question you have already asked. If the research question is vague, no amount of SPSS, R or AMOS will rescue it. Write the question as a sentence with the variables named in it: “Does perceived usefulness predict continued use of the platform among staff who were trained in 2024?” Once that sentence exists, the test is usually obvious.
1. Are you comparing groups, or looking at a relationship?
Almost every analysis in social science and health research falls into one of these two families, and the families use different tools.
Comparing groups. Two independent groups and a continuous outcome: independent samples t-test. The same people measured twice: paired samples t-test. Three or more groups: one-way ANOVA, followed by a post-hoc test to find out which pairs differ. If the outcome is badly skewed or ordinal, the non-parametric versions are Mann-Whitney U, Wilcoxon signed-rank and Kruskal-Wallis in that same order.
Looking at a relationship. Two continuous variables, roughly linear: Pearson correlation. Ordinal data, or a relationship that is monotonic but not linear: Spearman. Two categorical variables: chi-square test of independence.
2. Are you predicting, or just describing?
Correlation says two things move together. Regression says how much of one you expect given the other, and lets you hold other variables constant.
Continuous outcome: linear regression. Binary outcome such as adopted or did not adopt: logistic regression. Count outcome such as number of visits: Poisson or negative binomial. The moment you have more than one predictor, regression is almost always the better choice, because it tells you what each variable adds once the others are accounted for.
3. Are your variables measured directly, or through a scale?
This is the fork that sends people to structural equation modelling. If your construct is something like satisfaction, trust or perceived ease of use, you did not measure it directly. You measured several items that stand in for it.
Averaging those items into one score and running a regression is common, and it quietly assumes the items are all equally good indicators and measured without error. SEM does not assume that. It estimates the measurement model and the structural model together, so the paths between constructs are corrected for how well each construct was measured. If you have several constructs and several relationships between them, and the constructs are latent, SEM is the honest choice.
4. Check the assumptions before you report anything
Every test above carries conditions. Independence of observations. Residuals that are approximately normal, not raw scores. Homogeneity of variance across groups. Linearity where linearity is assumed. Enough cases for the number of parameters you are estimating.
Checking these is not an administrative step you do at the end to satisfy a reviewer. A violated assumption can reverse a conclusion, and it is far cheaper to find out before the results are in your discussion chapter.
The step most people skip
Decide the analysis before you collect the data. Write the test next to each hypothesis in your proposal. It takes an afternoon, and it tells you whether your instrument will actually produce data the test can use. The alternative is discovering, after 300 questionnaires, that your key variable was collected on a scale that will not support the model you had in mind.
If you are at that point already, the analysis is usually still salvageable, but the route there depends on what you have. That is what the free call is for.