Which Statistical Test Should I Use? A Decision Guide for Dissertations

You have collected your data, the spreadsheet is full, and you are staring at it with no idea which statistical test to run. This is one of the most common places a quantitative dissertation stalls, and choosing wrong here is costly, because a mismatched test produces results that do not answer your research question and get your chapter sent back. The good news is that test selection is not guesswork, and it is not a matter of memorizing a hundred tests. It is a structured decision driven by a small number of features of your study: what your research question is asking, how your variables are measured, how many groups you have, and whether a few statistical assumptions hold. Once you can read those features, the correct test is almost always determined for you. This guide walks through that decision the way a statistician does, so you can identify the right test for your own data and defend the choice to your committee. When you want the analysis run and written to examiner standard, our dissertation data analysis service does exactly that.

Quick Answer:

The correct statistical test is determined by six things: whether your research question is about differences between groups or relationships between variables; the level of measurement of each variable (nominal, ordinal, interval, or ratio); the number of groups or conditions; whether your samples are independent or related (paired or repeated); the number of dependent and independent variables; and whether the assumptions of parametric tests are met. Work through those in order, and the test selects itself. For example, a difference between two independent groups on a continuous, normally distributed outcome calls for an independent-samples t-test; if the outcome is ordinal or the normality assumption fails, its non-parametric equivalent, the Mann-Whitney U test, applies instead.

Start With the Question, Not the Test

The single most common mistake is to start by asking "which test is popular in my field?" and work backward. Start instead with what your research question is actually asking, because that determines the entire branch of statistics you are in. Almost every quantitative research question is one of two types.

The first type asks about differences between groups: does one group score higher than another, did scores change after an intervention, or do three conditions differ? These questions lead to tests like t-tests and ANOVA. The second type asks about relationships or prediction: are two variables associated, does one variable predict another, can we model an outcome from several predictors? These questions lead to correlation and regression. A smaller third category asks about associations between categorical variables, which leads to chi-square tests. Naming which of these your question belongs to is the first fork in the decision tree, and it rules out most of the options immediately. If you are still deciding whether your study is even quantitative, our guide on whether your dissertation should be qualitative or quantitative is the place to settle that first.

The Six Factors That Determine Your Test

Test selection is a six-factor decision, not a lookup in a table. Each factor narrows the field, and together they almost always point to a single correct test. Work through them in this order.

One, the research question type. Difference or relationship, as described above. This sets the branch.

Two, the level of measurement of each variable. This is the factor students most often overlook, and it is decisive. Nominal variables are unordered categories, such as treatment group or sex. Ordinal variables are ordered but without equal intervals, such as a Likert response or a pain rating. Interval and ratio variables are continuous, with equal intervals, such as blood pressure, test scores, or reaction time. Continuous outcomes open the door to parametric tests; ordinal or nominal outcomes generally require non-parametric or categorical methods. You cannot choose a test without knowing the measurement level of your dependent variable in particular.

Three, the number of groups or conditions. Comparing two groups leads to a different test than comparing three or more. A two-group difference on a continuous outcome is a t-test; three or more groups is an ANOVA, because running multiple t-tests inflates your error rate.

Four, independent or related samples. Independent samples are different people in each group, such as a treatment group and a separate control group. Related samples are the same people measured more than once, such as a pre-test and a post-test on one group, or matched pairs. This distinction changes the test entirely: an independent-samples t-test versus a paired-samples t-test, a one-way ANOVA versus a repeated-measures ANOVA. Getting it wrong is a serious error because related-samples designs violate the independence assumption of the independent tests.

Five, the number of dependent and independent variables. One outcome or several? One predictor or many? A single continuous outcome predicted by several variables is multiple regression; several outcomes analyzed together are MANOVA. The count of variables on each side of your question shapes the test.

Six, whether parametric assumptions hold. Parametric tests assume, among other things, that your continuous outcome is approximately normally distributed and, for group comparisons, that variances are roughly equal. When those assumptions hold, parametric tests are the default because they are more powerful. When they are meaningfully violated, you move to a non-parametric equivalent or a robust method. This factor is checked after you have provisionally chosen a test, and it can redirect you.

Table 1: The Six Factors That Determine Your Statistical Test

Tests for Differences Between Groups

Once your question is about differences and you know your measurement level, group count, and sample structure, the test is essentially determined.

For a continuous outcome compared between two independent groups, use the independent-samples t-test. For the same comparison across three or more independent groups, use a one-way ANOVA, which tests whether any group differs and is followed by post-hoc tests to locate which pairs differ. When the same participants are measured twice, use the paired-samples t-test; when they are measured three or more times, use the repeated-measures ANOVA. If you have more than one grouping factor, a factorial ANOVA tests each factor and, importantly, their interaction. If you need to compare groups while adjusting for a continuous covariate, ANCOVA does that. And when you have several dependent variables that are conceptually related and want to analyze them together, MANOVA is the multivariate extension.

Each of these is a parametric test, and each has a non-parametric counterpart for when the assumptions do not hold, which is where many students go wrong by forcing a parametric test onto data that cannot support it.

Tests for Relationships and Prediction

When your question is about association or prediction rather than group differences, you are in the correlation and regression family.

To measure the strength and direction of the linear relationship between two continuous variables, use Pearson's correlation coefficient. If either variable is ordinal, or the relationship is monotonic but not linear, or normality fails, use Spearman's rho instead. To predict a continuous outcome from a single predictor, use simple linear regression; from several predictors, multiple linear regression. When you want to enter predictors in theoretically motivated blocks and see what each block adds, hierarchical regression does that. And when your outcome is not continuous but binary, such as improved versus not improved, you use logistic regression, which reports odds ratios rather than raw coefficients. Logistic regression extends to multinomial and ordinal outcomes as well.

Tests for Categorical Variables

A distinct branch handles questions about associations between categorical variables. To test whether two categorical variables are associated, such as a treatment group and a yes or no outcome, use the chi-square test of independence. To test whether the distribution of a single categorical variable matches an expected distribution, use the chi-square goodness-of-fit test. When your expected cell counts are small, conventionally below five in any cell, the chi-square approximation becomes unreliable, and you use Fisher's exact test instead. For paired categorical data, such as a yes or no measured before and after on the same people, McNemar's test is the correct choice, and it is one that dissertation and DNP candidates frequently miss.

Parametric and Non-Parametric: Making the Right Switch

The choice between a parametric test and its non-parametric equivalent is where sophistication shows, and where a lot of chapters are either over-corrected or under-corrected. Parametric tests, the t-tests, ANOVAs, and Pearson correlation, assume your continuous outcome is approximately normally distributed and, for comparisons, that group variances are similar. They are the first choice when those conditions are met, because they have more statistical power. Non-parametric tests make fewer distributional assumptions and are the right choice when the outcome is ordinal, or when the normality assumption is meaningfully violated, particularly in small samples.

The mapping is clean. The independent-samples t-test maps to the Mann-Whitney U test. The paired-samples t-test maps to the Wilcoxon signed-rank test. The one-way ANOVA maps to the Kruskal-Wallis test. The repeated-measures ANOVA maps to the Friedman test. Pearson's correlation maps to Spearman's rho. Knowing these pairs means that when an assumption fails, you already know where to go.

Table 2: Parametric Tests and Their Non-Parametric Equivalents

One important nuance, and stating it correctly signals a real understanding of the method: mild departures from normality often do not matter. Under the central limit theorem, the sampling distribution of the mean approaches normality as sample size grows, so with samples larger than roughly 30 per group, parametric tests are generally robust to modest non-normality. This means you should not abandon a t-test the moment a Shapiro-Wilk test returns a significant result, because in large samples, that test flags trivial deviations. Judgment about whether a violation is severe enough to matter is part of the analysis, not a mechanical rule, and it is one of the things examiners probe.

The Assumptions Your Examiner Will Check

Choosing the right test is only half the task; you also have to verify that its assumptions hold and document what you did when they did not. This is where a large share of dissertations are sent back, so it is worth doing properly. Normality is checked with the Shapiro-Wilk test alongside Q-Q plots and histograms, not any single one of them alone. Homogeneity of variance, the assumption that groups have similar spread, is checked with Levene's test, and when it fails, you can use a Welch correction rather than abandoning the analysis. In regression, you check linearity with scatterplots, multicollinearity with the variance inflation factor, independence of residuals with the Durbin-Watson statistic, and homoscedasticity with a residual-versus-fitted plot. Repeated-measures ANOVA adds the sphericity assumption, checked with Mauchly's test, with a Greenhouse-Geisser correction applied when it is violated.

When an assumption is seriously violated, you have three standard routes: transform the variable, for instance, with a log or square-root transformation; switch to a robust method such as Welch's test or bootstrapping; or move to the matched non-parametric test. What you cannot do is ignore the violation and report the parametric result as if nothing happened, because that is exactly what a careful examiner looks for. The full set of checks, and what to do about each, is involved enough that we treat it in a dedicated guide, and it is a core part of what our dissertation data analysis service documents in every deliverable.

Do Not Stop at the P-Value: Effect Sizes and Confidence Intervals

Selecting and running the right test is not the end. Under APA 7th edition and its reporting standards, reporting a p-value alone is one of the most frequently flagged errors. It is empirically common: studies of graduate theses find effect sizes reported for only a minority of tests. A p-value tells you whether an effect is likely to exist; it says nothing about how large or how meaningful that effect is. That is what an effect size conveys, and examiners now expect one beside every inferential result, reported with its confidence interval.

The common effect sizes track the tests. For t-tests, report Cohen's d. For ANOVA, report eta-squared or partial eta-squared. For correlation, the r value is itself the effect size. For regression, report R-squared. For chi-square, report phi or Cramer's V. Cohen offered rough benchmarks, around 0.2, 0.5, and 0.8 for small, medium, and large d, and 0.1, 0.3, and 0.5 for r, but it is worth being clear that these are conventions, not laws. Cohen himself described them as last-resort values to be used only when field-specific norms are unavailable, and in some fields, the typical effect is smaller or larger than his benchmarks suggest. Report the number, interpret it against your field, and avoid presenting the thresholds as if they were fixed scientific constants. Reporting your results this way is precisely what the results chapter requires: the statistic, the degrees of freedom, the exact p-value, the effect size, and the confidence interval. Getting the whole chapter right is covered in our related guidance on Chapter 4 reporting.

Sample Size and Power: The Decision You Should Have Made Earlier

One decision belongs before data collection, but is worth naming here, because a missing power analysis is a routine committee and review board expectation that trips candidates up. An a priori power analysis determines the minimum sample size you need to detect an effect of a given size, before you collect anything. The standard free tool is G*Power, into which you enter your expected effect size, your alpha level (usually .05), and your desired power (usually .80 or .90), and it returns the minimum N. If you are still at the proposal stage, do this now; if you are past it, you should still report the power analysis that justified your sample. Committees increasingly ask for it, and an underpowered study cannot be rescued by any test choice afterward. This is one of several points where the analysis plan should have been set in the methodology chapter, a linkage our overview of the ten most common dissertation mistakes returns to.

A Note for Nursing and DNP Candidates

Statistical analysis in a DNP scholarly project is often distinct from a traditional PhD dissertation, and the difference matters for test selection. A DNP project is practice-focused and translational: it evaluates the outcome of a quality improvement intervention rather than generating generalizable new knowledge. That changes the analytic toolkit. Much DNP work centers on run charts and statistical process control charts, which plot a measure over time and distinguish ordinary common-cause variation from special-cause signals that indicate real change. These often tell you more than a single pre-versus-post comparison, because they show whether an improvement held.

Where a pre-versus-post comparison is used, the tests follow the same logic as the rest of this guide. Use a paired-samples t-test for a continuous outcome measured before and after on the same patients, the Wilcoxon signed-rank test when its assumptions fail, and McNemar's test for a paired categorical outcome. The clinical question is usually framed with PICOT, which drives the outcome measure and therefore the analysis, and our guide on writing a PICOT question covers that framing. DNP projects are also governed by their own reporting standard, SQUIRE 2.0, and their own competency framework in the AACN Essentials, both of which shape how the analysis is presented. This QI-specific analysis, run charts, control charts, and pre-post testing, is a distinct specialism, and it is central to our DNP capstone support.

Defending Your Choice

Whatever test you land on, you must be able to defend the choice, because "the software suggested it" or "my field uses it" will not survive a viva. A defensible justification names the research question type, the measurement level of the variables, the group structure, and the assumption checks that led to this test rather than its alternatives. If you switched to a non-parametric test, say which assumption failed and how you confirmed it. If you applied a correction, name it. The candidates who handle the statistics section of the viva well are the ones who can walk the examiner through the decision as a chain of reasoning. That is the same chain this guide has followed, and our dissertation defense preparation guide rehearses the walkthrough. When you would rather have the analysis run, documented, and made defensible for you, our dissertation support covers the whole results stage.

Frequently Asked Questions

How do I decide which statistical test to use?

Work through six factors in order. They are: whether your research question is about differences or relationships, the level of measurement of each variable, the number of groups, whether samples are independent or related, the number of dependent and independent variables, and whether parametric assumptions hold. Each factor narrows the options, and together they almost always point to a single correct test. Starting from the research question rather than from a familiar test is the key.

What is the difference between a parametric and a non-parametric test?

Parametric tests, such as t-tests, ANOVA, and Pearson correlation, assume your continuous outcome is approximately normally distributed and that group variances are similar. They are the first choice when those assumptions hold, because they are more powerful. Non-parametric tests, such as Mann-Whitney U and Kruskal-Wallis, make fewer distributional assumptions and are used for ordinal data or when normality is meaningfully violated, especially in small samples.

Which test do I use to compare two groups?

It depends on the measurement level and whether the groups are independent. For a continuous, roughly normal outcome in two independent groups, use the independent-samples t-test; if normality fails or the outcome is ordinal, use the Mann-Whitney U test. If the same people are measured twice, use the paired-samples t-test, or the Wilcoxon signed-rank test when its assumptions do not hold. For a categorical outcome, use a chi-square test, or McNemar's test if the data are paired.

When should I use a non-parametric test instead of a t-test?

Switch when your outcome is ordinal rather than continuous, or when the normality assumption is meaningfully violated, particularly in a small sample. Note that mild non-normality often does not matter in samples larger than about 30 per group, because of the central limit theorem. Hence, a single significant Shapiro-Wilk result is not automatically a reason to abandon a t-test. Judgment about the severity of the violation is part of the analysis.

Do I need to report effect sizes as well as p-values?

Yes. Under APA 7th edition and its reporting standards, a p-value alone is one of the most commonly flagged errors. The p-value indicates whether an effect exists; the effect size indicates how large and meaningful it is, and it should be reported with a confidence interval. Use Cohen's d for t-tests, eta-squared for ANOVA, r for correlation, and R-squared for regression, and interpret the value against your field rather than treating Cohen's benchmarks as fixed rules.

How do I choose a statistical test for a DNP project?

DNP projects are practice-focused quality improvement work, so the analysis often centers on run charts and statistical process control charts, which track a measure over time and distinguish common-cause from special-cause variation. Where a pre-versus-post comparison is used, apply the same logic as any study: a paired-samples t-test for a continuous outcome, the Wilcoxon signed-rank test when assumptions fail, or McNemar's test for a paired categorical outcome. DNP reporting follows SQUIRE 2.0.

What statistical software should I use for my dissertation?

Use what your program accepts and what you can defend. SPSS is the menu-driven standard many committees recognize, and R is free, reproducible, and the most flexible. Stata and SAS are common in economics, epidemiology, and clinical trials. Jamovi and JASP are free modern interfaces that produce clean APA tables. Name the software and version in your methods section, and confirm your supervisor accepts it, since some programs require SPSS output specifically.

Letting the Test Select Itself

The right statistical test is not something you memorize; it is something you derive. Name whether your question is about differences or relationships, identify the measurement level of your variables, count your groups, decide whether your samples are independent or related, count your dependent and independent variables, and check whether the parametric assumptions hold. Follow that chain and the correct test presents itself, and you will be able to explain to any examiner exactly why it was the right one. Then report it with an effect size and a confidence interval, document the assumptions you checked, and your results chapter will hold.

If you would rather a PhD statistician run the analysis, verify the assumptions, and hand you a defensible results chapter with reproducible code, send us your data and research questions. You will have an itemized quote within 2 to 4 business hours, with no obligation.