Choosing the right statistical test becomes much easier when you stop starting with test names.
You do not first decide whether to use a t-test, ANOVA, chi-square test, correlation, or regression. Instead, start with your research question, identify the type of outcome you are analyzing, understand how your observations are structured, and then choose a statistical method that matches that design.
For most Master’s and PhD research, you can narrow down the appropriate statistical test by answering six questions:
- What does your research question ask?
- What type of outcome variable do you have?
- What type of independent or predictor variable do you have?
- How many groups, conditions, or measurements are involved?
- Are the observations independent, paired, repeated, or clustered?
- Does the proposed statistical test meet its assumptions?
This guide walks through each question and shows how to use the answers to choose an appropriate statistical test for your dissertation or thesis.
Quick Statistical Test Selection Table
If you already understand your variables and research design, use this table as a starting point.
| Research question or design | Typical statistical test |
|---|---|
| Compare one sample mean with a known or hypothesized value | One-sample t test |
| Compare a continuous outcome between two independent groups | Independent-samples t test or Welch’s t test |
| Compare a continuous outcome before and after in the same participants | Paired-samples t test |
| Compare a continuous outcome among three or more independent groups | One-way ANOVA or Welch’s ANOVA |
| Compare the same participants at three or more time points | Repeated-measures ANOVA or mixed-effects model |
| Examine the relationship between two continuous variables | Pearson correlation |
| Examine a monotonic relationship between ordinal or non-normal continuous variables | Spearman correlation |
| Examine an association between two categorical variables | Chi-square test of independence |
| Examine an association between categorical variables when expected counts are very small | Fisher’s exact test |
| Predict a continuous outcome from one or more variables | Linear regression |
| Predict a binary outcome such as yes/no | Binary logistic regression |
| Predict an ordinal outcome | Ordinal logistic regression |
| Predict a categorical outcome with three or more unordered categories | Multinomial logistic regression |
| Analyze a count outcome | Poisson or negative binomial regression |
| Analyze time until an event occurs | Survival analysis, such as Kaplan-Meier or Cox regression |
This table should not be used mechanically. Two studies with similar variables can require different methods because their research questions or study designs differ.
The steps below will help you make the final decision.
Step 1: Start With Your Research Question
Your research question is usually the best place to begin.
Most quantitative research questions fall into one of four broad categories.
1. Comparing groups
These questions ask whether groups or conditions differ.
Examples:
- Is job satisfaction different between permanent and contract employees?
- Do three teaching methods produce different examination scores?
- Did anxiety scores change after an intervention?
Common statistical methods include:
- independent-samples t test;
- paired-samples t test;
- ANOVA;
- repeated-measures ANOVA;
- Mann-Whitney U test;
- Wilcoxon signed-rank test; and
- Kruskal-Wallis test.
2. Examining relationships or associations
These questions ask whether variables are related.
Examples:
- Is burnout related to job satisfaction?
- Is gender associated with preferred learning method?
- Is employee engagement associated with intention to leave?
Possible methods include:
- Pearson correlation;
- Spearman correlation;
- chi-square test of independence; and
- Fisher’s exact test.
3. Predicting an outcome
These questions examine whether one or more variables predict an outcome.
Examples:
- Do workload and organizational support predict job satisfaction?
- Does age predict systolic blood pressure?
- Do burnout, salary satisfaction, and leadership support predict whether an employee intends to leave?
Possible methods include:
- linear regression;
- binary logistic regression;
- ordinal logistic regression;
- multinomial logistic regression; and
- other generalized regression models.
4. Examining change over time
These questions involve measurements taken from the same participants more than once.
Examples:
- Did depression scores change from before treatment to after treatment?
- Did academic performance change across three semesters?
- Did blood pressure change at baseline, three months, and six months?
The important point here is that repeated observations from the same person are not independent.
Possible methods include:
- paired-samples t test;
- Wilcoxon signed-rank test;
- repeated-measures ANOVA; and
- mixed-effects models.
Once you know what the research question is asking, you have already eliminated many inappropriate statistical tests.
Step 2: Identify Your Dependent or Outcome Variable
Next, identify the variable you are trying to explain, compare, or predict.
This is usually called the dependent variable, outcome variable, or response variable.
The type of outcome variable has a major influence on the statistical test you should use.
Continuous variables
A continuous variable represents a numerical measurement or score for which differences between values are meaningful.
Examples include:
- age;
- income;
- examination score;
- blood pressure;
- height;
- weight;
- job satisfaction scale score.
Common methods for continuous outcomes include t tests, ANOVA, correlation, and linear regression.
Binary variables
A binary variable has two possible categories.
Examples include:
- yes/no;
- passed/failed;
- employed/unemployed;
- intention to leave/intention to stay;
- disease/no disease.
If you are predicting a binary outcome using several predictors, binary logistic regression is usually more appropriate than linear regression.
Nominal categorical variables
Nominal variables contain categories without a natural ranking.
Examples include:
- marital status;
- department;
- occupation;
- preferred learning method;
- geographic region.
If you are examining whether two categorical variables are associated, a chi-square test of independence may be appropriate.
If the outcome has three or more unordered categories and you want to predict it, multinomial logistic regression may be appropriate.
Ordinal variables
Ordinal variables contain categories that have a meaningful order but do not necessarily have equal distances between categories.
Examples include:
- poor, fair, good, excellent;
- low, medium, high;
- strongly disagree to strongly agree;
- disease stages.
Methods designed specifically for ordinal outcomes, such as ordinal logistic regression, may sometimes be appropriate.
Step 3: Identify Your Independent or Predictor Variable
After identifying the outcome, determine what variable is being used to create groups, explain variation, or predict the outcome.
For example, consider this research question:
Is there a difference in mean job satisfaction between permanent and contract employees?
Here:
- Job satisfaction is the outcome.
- Employment type is the grouping variable.
- Employment type contains two independent groups.
This structure points toward an independent-samples comparison, such as an independent-samples t test or Welch’s t test if the outcome is treated as continuous.
Now consider:
Do workload, burnout, and organizational support predict job satisfaction?
Here:
- Job satisfaction is the continuous outcome.
- Workload, burnout, and organizational support are predictors.
- There are several predictor variables.
Multiple linear regression is therefore a more appropriate starting point than a correlation or t test.
Step 4: Determine How Many Groups or Measurements You Have
The number of groups matters when the research question asks about differences.
One group compared with a known value
Suppose a university knows that the historical average examination score is 70 and wants to determine whether this year’s students have a different mean score.
A one-sample t test can compare the sample mean with the specified value.
Two independent groups
Suppose you want to compare mean job satisfaction between:
- permanent employees; and
- contract employees.
These are two separate groups of people.
An independent-samples t test or Welch’s t test may be appropriate for a continuous outcome.
Three or more independent groups
Suppose you want to compare examination scores among students taught using:
- traditional lectures;
- flipped classrooms; and
- problem-based learning.
There are three independent groups.
A one-way ANOVA may be appropriate.
If the overall ANOVA is significant, a suitable post-hoc procedure can be used to determine which specific groups differ.
Do not simply conduct multiple independent t tests between every pair of groups. Doing many separate tests increases the probability of obtaining a false-positive result.
Two measurements from the same participants
Suppose stress is measured before and after a six-week intervention.
The observations are paired because each participant contributes two scores.
A paired-samples t test may be appropriate.
Three or more measurements from the same participants
If stress is measured at:
- baseline;
- three months; and
- six months,
a repeated-measures method is required.
A repeated-measures ANOVA may be suitable for a relatively simple complete design. More flexible longitudinal designs may require a mixed-effects model.
Step 5: Determine Whether Your Observations Are Independent
This step is easy to overlook, but it can completely change the analysis.
Independent observations
Observations are independent when one participant’s measurement does not form a natural pair with another participant’s measurement.
For example:
- 50 male participants and 50 female participants;
- patients receiving treatment A versus different patients receiving treatment B;
- employees from two unrelated organizations.
Paired observations
Observations are paired when measurements are naturally connected.
Examples include:
- the same participants before and after treatment;
- matched case-control pairs;
- measurements from twins;
- measurements from the same person under two conditions.
An independent-samples t test should not be used when the observations are actually paired.
Repeated observations
If the same individual is measured several times, those observations are correlated.
For example:
Blood pressure is measured monthly for six months.
Treating all six measurements as independent ignores the fact that measurements from the same participant tend to resemble one another.
Repeated-measures or mixed-effects methods are designed to account for this dependency.
Clustered observations
Data can also be grouped in other ways.
For example:
- students are nested within schools;
- patients are nested within hospitals;
- employees are nested within departments;
- repeated observations are nested within participants.
Simple statistical tests assume independent observations and may not be suitable when strong clustering exists.
A multilevel or mixed-effects model may be more appropriate.
Step 6: Decide Whether You Need a Comparison Test or a Regression Model
One common mistake is using several separate tests when the actual research question requires a model.
Suppose your question is:
Does job satisfaction differ between male and female employees?
A two-group comparison may answer the question.
But suppose your question becomes:
Does gender predict job satisfaction after controlling for age, education, salary, and years of employment?
You are no longer conducting a simple two-group comparison.
You need a model that allows several predictors to be considered simultaneously.
For a continuous outcome, that will often be multiple linear regression.
For a binary outcome, it may be multiple logistic regression.
The need to control for confounding variables or include several predictors often moves the analysis from a simple hypothesis test to a regression model.
Which Statistical Test Should I Use for Two Groups?
If you are comparing two groups, first determine whether the groups are independent or paired.
Two independent groups with a continuous outcome
Possible test:
Independent-samples t test or Welch’s t test
Example:
Is mean burnout different between nurses working day shifts and nurses working night shifts?
The two groups contain different nurses.
Two related measurements with a continuous outcome
Possible test:
Paired-samples t test
Example:
Did nurses’ burnout scores change after a workplace intervention?
The same nurses are measured twice.
Two independent groups with an ordinal or strongly non-normal outcome
Depending on the research question and distribution, a Mann-Whitney U test may be considered.
Two paired measurements with ordinal or unsuitable continuous data
A Wilcoxon signed-rank test may be considered.
The important point is that you cannot choose between an independent and paired test simply by counting groups. You must know whether the observations are related.
When Should I Use a t Test or ANOVA?
A t test and ANOVA are both commonly used to compare means.
The main difference in a simple independent-groups setting is the number of groups.
Use an independent-samples t test when comparing two independent groups.
Use a one-way ANOVA when comparing three or more independent groups.
For example:
Compare average salary between men and women.
Two groups → independent-samples t test.
Compare average salary among employees in finance, marketing, operations, and human resources.
Four groups → one-way ANOVA.
However, the number of groups is not the only consideration. The design and assumptions must also be appropriate.
When Should I Use Pearson or Spearman Correlation?
Both Pearson and Spearman correlation examine relationships between variables, but they answer slightly different questions.
Pearson correlation
Pearson correlation is commonly used to measure the strength and direction of a linear relationship between two continuous variables.
Example:
Is organizational support related to job satisfaction?
If both variables are continuous scale scores and the relationship is reasonably linear, Pearson correlation may be appropriate.
Spearman correlation
Spearman correlation is based on ranks and measures the strength and direction of a monotonic relationship.
It is often useful when:
- variables are ordinal;
- the relationship is monotonic but not suitably linear;
- extreme outliers make Pearson correlation inappropriate;
- continuous data do not meet the conditions required for the planned Pearson analysis.
Do not choose Spearman simply because one normality test produced a significant p-value. Examine the variables, relationship, distribution, outliers, sample size, and purpose of the analysis together.
When Should I Use a Chi-Square Test?
A chi-square test of independence is commonly used when both variables are categorical.
For example:
Is preferred learning method associated with student gender?
Suppose the variables are:
Gender
- male;
- female.
Preferred learning method
- online;
- classroom.
Both variables are categorical.
A chi-square test of independence can test whether the distribution of learning preferences differs across gender categories.
If expected cell counts are very small, particularly in a 2 × 2 table, Fisher’s exact test may be more appropriate.
When Should I Use Regression?
Regression is appropriate when the goal goes beyond simply comparing two variables and involves explaining or predicting an outcome.
The type of regression depends largely on the outcome variable.
Linear regression
Use linear regression when the dependent variable is continuous.
Example:
Do workload, salary satisfaction, and organizational support predict job satisfaction?
Outcome:
Job satisfaction score → continuous.
Possible method:
Multiple linear regression
Binary logistic regression
Use binary logistic regression when the outcome has two categories.
Example:
Do burnout, age, salary satisfaction, and workload predict employees’ intention to leave?
Outcome:
- intention to leave;
- no intention to leave.
Possible method:
Binary logistic regression
Ordinal logistic regression
Consider ordinal logistic regression when the dependent variable has ordered categories.
Example:
Which factors predict students’ satisfaction level?
Outcome:
- dissatisfied;
- neutral;
- satisfied;
- very satisfied.
Multinomial logistic regression
Multinomial logistic regression can be used when an outcome has three or more categories that do not have a meaningful order.
What if My Data Are Not Normally Distributed?
Do not automatically assume that a significant normality test means you must abandon every parametric method.
The correct decision depends on factors such as:
- the particular statistical test;
- where the normality assumption applies;
- sample size;
- severity of skewness;
- influential outliers;
- balance between groups;
- variance structure;
- and robustness of the proposed method.
For example, an independent-samples t test does not require every value in the entire dataset to follow one perfect normal distribution.
For regression and many other models, the relevant assumptions concern the model and its residuals rather than requiring every predictor to be normally distributed.
If assumptions are seriously violated, possible alternatives may include:
| Parametric method | Possible alternative |
|---|---|
| Independent-samples t test | Mann-Whitney U test |
| Paired-samples t test | Wilcoxon signed-rank test |
| One-way ANOVA | Kruskal-Wallis test |
| Pearson correlation | Spearman correlation |
These are not automatic replacements. The alternative must still answer the research question appropriately.
What if the Group Variances Are Unequal?
Unequal variances do not automatically mean that you need a nonparametric test.
When comparing two independent means, Welch’s t test is designed to handle unequal variances and unequal sample sizes better than the traditional equal-variance t test.
Similarly, Welch’s ANOVA can be considered when comparing three or more independent group means when the equal-variance assumption is problematic.
The test should match the actual problem with the data rather than switching methods simply because one assumption check failed.
How Do Likert Scale Questions Affect Statistical Test Selection?
Likert data require some thought because a single Likert item and a multi-item scale are not necessarily analyzed in the same way.
A single item such as:
How satisfied are you with your supervisor?
with responses from:
- Strongly disagree
- Disagree
- Neutral
- Agree
- Strongly agree
is ordinal.
However, researchers often combine several related Likert items into a composite scale score after establishing that the items appropriately measure the same construct.
For example, six questionnaire items may be combined into an overall organizational support score.
Depending on the scale, research design, distribution, measurement properties, and disciplinary conventions, the resulting composite score may sometimes be analyzed using methods designed for approximately continuous outcomes.
Do not decide that every variable coded 1 to 5 must automatically be analyzed the same way.
First determine whether you are analyzing:
- one individual Likert item;
- an ordinal outcome;
- or a validated multi-item scale score.
Five Dissertation Examples: From Research Question to Statistical Test
Example 1: Comparing two independent groups
Research question:
Is burnout different between permanent and contract healthcare workers?
Outcome: Burnout score, continuous
Predictor: Employment type, categorical
Groups: Two
Relationship between observations: Independent
Possible test: Independent-samples t test or Welch’s t test, depending on the analysis conditions.
Example 2: Before-and-after study
Research question:
Does anxiety decrease after an eight-week counseling intervention?
Outcome: Anxiety score, continuous
Measurements: Before and after
Participants: Same people measured twice
Possible test: Paired-samples t test if its assumptions are sufficiently satisfied.
Example 3: Association between categorical variables
Research question:
Is employment status associated with intention to leave an organization?
Employment status: Categorical
Intention to leave: Yes/no
Possible test: Chi-square test of independence, provided the contingency-table conditions are appropriate.
Example 4: Predicting a continuous outcome
Research question:
Do workload, leadership support, and salary satisfaction predict employee job satisfaction?
Outcome: Job satisfaction score, continuous
Predictors: Multiple variables
Possible test: Multiple linear regression.
Example 5: Predicting a binary outcome
Research question:
Do burnout, workload, organizational support, and age predict whether healthcare workers intend to leave their jobs?
Outcome: Intention to leave, yes/no
Predictors: Multiple
Possible test: Binary logistic regression.
Notice that the test is selected from the structure of the research question and data. It is not chosen because one statistical test happens to be more familiar.
Common Mistakes When Choosing a Statistical Test
Choosing the test before defining the research question
Do not decide that you are going to use ANOVA or regression simply because you have seen that test in other dissertations.
Start with what you are actually trying to answer.
Ignoring the type of outcome variable
Linear regression is designed for a continuous outcome. Standard binary logistic regression is designed for a two-category outcome.
Changing the outcome type can therefore change the model completely.
Ignoring paired or repeated observations
Before-and-after measurements from the same participants are not independent groups.
Using an independent-samples test for paired observations discards important information about the study design.
Running several t tests instead of ANOVA
When comparing several groups, performing every possible pairwise t test without appropriate multiplicity control increases the probability of false-positive findings.
A suitable ANOVA and planned post-hoc procedure is generally a better starting point for a standard multi-group comparison.
Choosing a test only from a normality p-value
A normality test should not be treated as an automatic statistical-test selector.
Evaluate the research design, variable type, distribution, outliers, sample size, residuals where relevant, and assumptions of the specific method.
Confusing correlation with prediction
Correlation examines association between variables.
Regression models an outcome as a function of one or more predictors.
A significant correlation does not automatically mean that one variable predicts or causes another in a substantive or causal sense.
Ignoring confounding variables
A simple comparison may not answer a research question that explicitly requires adjustment for other characteristics.
For example:
Is organizational support associated with intention to leave after controlling for age, salary, profession, and years of experience?
This requires an adjusted model rather than several disconnected two-variable tests.
Choosing a statistical test because it gives p < .05
The appropriate analysis should be determined by the research question and design, not by which method produces statistical significance.
Trying several tests and reporting only the one that gives a desirable p-value can produce misleading conclusions.
When a Simple Statistical Test Is Not Enough
Not every Master’s or PhD research question fits neatly into a t test, ANOVA, correlation, or chi-square test.
You may need a more advanced method when your research includes:
- several predictors;
- confounding variables;
- interaction effects;
- repeated measurements;
- students nested within schools;
- patients nested within hospitals;
- longitudinal observations;
- multiple dependent variables;
- latent constructs;
- count outcomes;
- time-to-event outcomes;
- mediation or moderation;
- complex survey designs.
For example, a study measuring employees every month for one year may be better suited to a mixed-effects model than twelve separate comparisons.
Similarly, a PhD study involving several latent constructs measured using questionnaire items may require factor analysis or structural equation modeling rather than a series of basic hypothesis tests.
The goal is not to choose the most advanced method. It is to choose the simplest method that correctly answers the research question while respecting the study design and data structure.
A Simple Checklist Before Finalizing Your Statistical Test
Before running your analysis, check the following:
- Research question: What exactly am I trying to determine?
- Outcome: What is my dependent variable?
- Outcome type: Is it continuous, binary, nominal, ordinal, count, or time-to-event?
- Predictor: What is my independent or predictor variable?
- Groups: How many groups or categories are involved?
- Study design: Are observations independent, paired, repeated, or clustered?
- Covariates: Do I need to control for other variables?
- Assumptions: Are the assumptions of the proposed method reasonable?
- Follow-up analysis: Will I need post-hoc comparisons or additional model diagnostics?
- Interpretation: Does the chosen test actually answer my stated research question?
If you cannot answer one of these questions, resolve it before running the statistical test.
Frequently Asked Questions
How do I know which statistical test to use for my dissertation?
Start with your research question. Determine whether you are comparing groups, examining an association, investigating change, or predicting an outcome. Then identify your outcome type, predictor type, number of groups, whether observations are independent or related, and the assumptions of the candidate method.
Should I choose my statistical test before collecting data?
Ideally, yes. Your analysis plan should normally be considered during the research-design stage. This helps ensure that you collect the variables and sample required to answer your research questions. You can still check assumptions and refine the final analysis after seeing the data, but the fundamental analytical approach should follow the study design rather than be selected simply because of the observed results.
What statistical test should I use for two independent groups?
If you are comparing a continuous outcome between two independent groups, an independent-samples t test or Welch’s t test may be appropriate. For ordinal data or situations where the assumptions of the mean-based comparison are unsuitable, a Mann-Whitney U test may be considered.
What test should I use for before-and-after data?
If the same participants are measured before and after an intervention and the outcome is continuous, a paired-samples t test may be appropriate. If the conditions for that analysis are unsuitable, the Wilcoxon signed-rank test may be considered.
Should I use a t test or ANOVA?
For a simple comparison of two independent group means, a t test may be appropriate. For three or more independent groups, one-way ANOVA is commonly used. Repeated measurements require a paired or repeated-measures approach rather than an ordinary independent-groups test.
Should I use Pearson or Spearman correlation?
Pearson correlation is commonly used for a linear relationship between two continuous variables. Spearman correlation is based on ranks and is useful for ordinal variables or monotonic relationships where Pearson correlation is not appropriate.
Which statistical test should I use for Likert scale data?
It depends on whether you are analyzing an individual Likert item or a multi-item scale. A single Likert item is ordinal. A properly constructed composite scale may sometimes be treated as approximately continuous when this is justified. The research question and measurement properties should guide the analysis.
Can SPSS choose the correct statistical test for me?
SPSS can perform many statistical analyses, but you still need to specify an analysis that matches your research question and study design. Statistical software cannot replace understanding your dependent variable, predictors, group structure, repeated measurements, and assumptions.
Final Takeaway
The easiest way to choose the right statistical test is to work from the research question toward the analysis rather than starting with a test name.
Ask:
What am I trying to find out?
Then determine:
What type of outcome do I have?
What variables or groups am I comparing?
Are the observations independent or related?
Do I need to adjust for other variables?
Does the proposed method fit the assumptions and structure of my data?
Once those questions are clear, the number of reasonable statistical methods usually becomes much smaller.
Most importantly, your final choice should be one that you can justify in your dissertation and explain confidently to your supervisor or examination committee.