How to Choose a Statistical Test for a Dissertation or Research Project
Choose an analysis from the research question, outcome type, study design and data conditions. Software cannot make that decision for you, and a single decision tree cannot cover every design.
Write the question in analysis terms
First ask whether you want to compare groups, examine an association or estimate a predictor’s relationship with an outcome. Then identify the outcome variable, the explanatory variable or groups, and whether observations are independent or repeated on the same people. The UCLA OARC SPSS guide works through these distinctions with examples and assumptions.
“Do employees with different training histories report different satisfaction scores?” is not yet enough. Define the training groups, how satisfaction was measured, whether respondents appear once or repeatedly, and how missing responses will be handled. The analysis should follow the approved protocol or supervisor’s requirements, especially for a preregistered study.
Common questions and candidate methods
- Two independent groups, numeric outcome: an independent-samples t-test may fit when its conditions are met; alternatives or different models may be needed when they are not.
- Same participants measured twice: a paired analysis is relevant because the observations are linked. An independent-groups test ignores that design.
- More than two independent groups, numeric outcome: ANOVA or an appropriate regression model may address group differences, depending on design and assumptions.
- Two categorical variables: a contingency-table approach such as chi-square may fit; sparse expected counts can change the choice.
- Association or prediction: select correlation or regression from the variable types and the question. A relationship does not by itself establish causation.
These are candidate methods, not a substitute for a statistical analysis plan. UCLA’s SPSS learning module demonstrates how common procedures behave. More complex designs, clustered samples, multiple time points or missing data often require specialist input.
Why “Likert data” is not one automatic test choice
A single ordered response item and a multi-item score constructed from several items are different measurements. How a scale is scored, its reliability, distribution and intended interpretation all matter. It is unsafe to assert that every Likert response requires a non-parametric test, or that a normality p-value alone makes a parametric test valid. Explain your scoring and method in the study’s context.
Check the observed data: missingness, coding errors, outliers, group sizes and distribution. Examine assumptions relevant to the candidate method, not a generic checklist. If a method is unsuitable, choose a defensible alternative and explain it before looking for whichever test produces the smallest p-value.
A small example of a justified decision
Suppose one survey asks whether job-satisfaction scores differ between two separately recruited departments. The outcome is a pre-specified multi-item score; the groups are independent. An independent-samples comparison is a candidate. Before running it, check the score’s construction, group size, distribution, unusual values and relevant variance conditions. If the same workers answered before and after training, the observations are paired and the original test choice no longer matches the design.
Write down that logic with the research question and planned analysis. In the results, report estimates, uncertainty and limitations as well as the test statistic. Our SPSS output and APA reporting guide covers how to read and present output after the choice is justified; our SPSS workflow guide covers data checks and interpretation.
Need help matching the method to your study?
Labeeb can review your research question, variable definitions and analysis plan with you, then explain defensible options. The student or researcher retains responsibility for their data, decisions and submitted work.