A data analysis plan says, for each research question, which variables answer it, how they are measured, which test you will run, and why that test fits. It also says how you will prepare the data, which assumptions you will check, and what you will do if one is violated. In a five-chapter dissertation it is the Data Analysis section of Chapter 3, and your committee approves it at the proposal defense, before you collect anything.
Committees are not grading the sophistication of the statistics. They are checking alignment: does each analysis answer the question it is attached to, with the data you will have? A simple test that fits the question is easier to defend than a complex one that doesn't.
The plan also protects you later. If the analysis in Chapter 4 matches what the committee approved in Chapter 3, there is much less to argue about at the final defense.
What goes in the plan
Check your program's Chapter 3 template first. If it doesn't set the order, this one works for a quantitative study.
- Research questions and hypotheses. Restate them exactly as they appear in Chapter 1. Everything after this is organized around them.
- Variables. For each one, give its name, how it is measured, and its level of measurement: categorical, ordinal, or continuous. Say which is the outcome and which are the predictors.
- Data preparation. How you will screen the data, handle missing values, score any scales, and check their reliability.
- Descriptive statistics. What you will report to describe the sample and every variable before any test is run. If you are using secondary data and already have access to the data, sometimes committees will ask that you do preliminary descriptives.
- The analysis for each research question. Name the test and say why it fits the question and the variables. One short paragraph per question.
- Assumptions. Which assumptions each test makes, how you will check them, and what you will do if one is violated.
- Decision rules. Your alpha level, the effect size you will report, and how you will handle multiple comparisons if you run several tests on the same outcome.
- Sample size. The power analysis for the test that needs the most participants. See the sample size guide.
The alignment table
Before you write a paragraph, build one table. Each research question gets a row. If you can't fill in a row, the problem is in the question or the measure, and it is much cheaper to find that now than after data collection.
This example is a made-up survey study of hospital nurses.
| Research question | Outcome | Predictor or grouping variable | Test | What you will report |
|---|---|---|---|---|
| 1. Does burnout differ between night-shift and day-shift nurses? | Burnout score (continuous) | Shift (categorical, 2 groups) | Independent-samples t-test | Group means and SDs, t, df, p, Cohen's d |
| 2. Is burnout related to years of experience? | Burnout score (continuous) | Years of experience (continuous) | Pearson correlation | r, df, p |
| 3. Do shift, years of experience, and mentoring predict intent to stay? | Intent to stay score (continuous) | Shift, years of experience, mentoring (3 predictors) | Multiple regression | R², F, then b, SE, β, and p for each predictor |
Three things this table makes visible:
- Every question has a test, and every test has a question. Nothing is run just because you want a fancy test done.
- The test follows from how the variables are measured. Change the outcome in question 3 to a yes or no (stayed or left) and the test changes to logistic regression.
- The sample size follows from the table. With a medium effect, these three tests need 128, 84, and 77 participants. You recruit for the largest, which is the t-test.
Ask your chair whether this table should appear in Chapter 3 itself.
Matching the test to the question
The test depends on two things: what kind of variable the outcome is, and what you are comparing it against. This table covers the tests most dissertations use.
| Your question | Outcome | Usual test | If assumptions fail |
|---|---|---|---|
| Do two separate groups differ? | Continuous | Independent-samples t-test | Welch's t-test or Mann-Whitney U |
| Do the same people differ at two time points? | Continuous | Paired-samples t-test | Wilcoxon signed-rank |
| Do three or more groups differ? | Continuous | One-way ANOVA with post hoc tests | Welch's ANOVA or Kruskal-Wallis |
| Are two continuous variables related? | Continuous | Pearson correlation | Spearman correlation |
| Do several variables predict an outcome? | Continuous | Multiple regression | Robust standard errors or a transformation |
| Do several variables predict a yes or no outcome? | Binary | Logistic regression | |
| Are two categorical variables related? | Categorical | Chi-square test of independence | Fisher's exact test |
For a longer version of this table, and how to write the justification for each choice, see How to Choose a Statistical Test for Your Dissertation.
If your design is not here (repeated measures across several time points, nested data such as nurses within hospitals, or latent variables), the logic is the same but the test is not. That is a good point to get advice before the proposal.
Assumptions, missing data, and what you will do about them
Committees expect you to show you know what each test assumes. That does not mean an exhaustive list of diagnostics. It means naming the assumptions, saying how you will check them, and saying in advance what you will do if one fails. Silence on this raises more concern than an imperfect result.
| Test | Assumptions to name | How you will check | If it fails |
|---|---|---|---|
| t-test, ANOVA | Independent observations, roughly normal outcome in each group, similar variances | Histograms or Q-Q plots, Levene's test | Welch's version, or a nonparametric test |
| Pearson correlation | Linear relationship, no extreme outliers | Scatterplot | Spearman correlation |
| Multiple regression | Linearity, independent errors, constant error variance, roughly normal residuals, predictors not too highly correlated | Residual plots, Q-Q plot, variance inflation factors | Robust standard errors, a transformation, or dropping or combining predictors |
| Chi-square | Independent observations, expected counts large enough in each cell | Expected counts table | Fisher's exact test, or combining categories |
Three more decisions belong in the plan, because making them after you see the data looks like you chose the answer you wanted:
- Missing data. Say how you will report how much is missing, and what you will do about it: analyze complete cases, or use a method such as multiple imputation. Say what rule you will use to score a scale when a participant skips items.
- Multiple comparisons. If you run several tests on the same outcome, say whether you will adjust alpha and how. For an ANOVA, name the post hoc test.
- Outliers. Say how you will identify them and what you will do with them. "Removed because they changed the result" is not a rule.
Sample paragraphs
These are models to adapt to your own study, not text to copy. The study is made up for illustration.
Weak: "Data will be entered into SPSS and analyzed using descriptive and inferential statistics. T-tests and regression will be used to test the hypotheses. Results will be considered significant at p < .05."
Stronger: "Research question 1 asks whether burnout differs between night-shift and day-shift nurses. Burnout is a continuous score on the [instrument], and shift is a categorical variable with two independent groups, so an independent-samples t-test will be used. Normality will be examined with histograms and Q-Q plots, and equality of variances with Levene's test. If variances are unequal, Welch's t-test will be reported. Group means and standard deviations, the test statistic, degrees of freedom, exact p value, and Cohen's d will be reported, with alpha set at .05."
The weak version could describe any study. It never connects a test to a question or a variable. The stronger version names the question, says how each variable is measured, gives the reason the test fits, and says what happens if an assumption fails. Write one paragraph like it for each research question.
Citing a method is not the same as justifying it. "Multiple regression was used, following [Author, year]" tells the committee someone else did it. "Multiple regression will be used because the outcome is continuous and the question asks how three predictors relate to it together" tells them why it fits your study.
Qualitative and mixed methods plans
Qualitative
A qualitative analysis plan answers the same question in a different form: how will you get from raw data to findings, and why should a reader trust the route? It should say:
- The approach, by name, such as thematic analysis, and why it fits your research questions.
- The steps. How transcripts will be prepared, how coding will proceed, how codes become themes, and how themes will be checked against the data.
- Who does the coding, and whether anyone else reviews it.
- How you will establish trustworthiness. The specific strategies you will use, not a list of every term in the textbook.
- How the themes connect to the research questions.
- The software, if you are using any, in one sentence.
The same alignment table works here. Each research question gets a row, with the data source that answers it and the analytic step that produces the answer.
Mixed methods
Write a plan for each strand by its own rules, then add the part that makes the study mixed: where the two strands meet and how. Say whether one strand informs the design of the other or whether the results are brought together at interpretation, and how you will display them side by side. Integration is a design decision. A plan that leaves it for the results chapter usually ends with two studies stapled together.
What committees push back on
Most concerns about a quantitative plan are not about wrong statistics. They are about alignment and explanation.
- A test that doesn't match the research question. Usually because the method was chosen for familiarity, because a prior study used it, or because the software made it easy.
- A test that doesn't match how the variable is measured. A t-test planned for an outcome that is really a yes or no.
- Variables that are never defined. "Engagement" appears in the research question and the plan never says which instrument or items measure it.
- Software in place of reasoning. Steps and menus described in detail, with no explanation of why the test fits.
- No mention of assumptions, or a list of assumptions with no plan for checking them.
- A method cited but not justified. A reference to a prior study where an explanation should be.
- More complexity than the question needs. An advanced model where a simpler one answers the question. Committees value coherence over sophistication.
- A power analysis for a different test than the one in the plan.
- Nothing on missing data.
- Analyses that appear with no research question attached.
- Trying to use the analysis to apply causality. A type of analysis alone doesn't establish causality. Your sampling procedure plays as much (if not more) of a role.
Checklist before it goes to your chair
- The plan follows your program's Chapter 3 template, if it has one.
- Every research question has an analysis, and every analysis has a research question.
- Every variable is defined, with its measure and level of measurement.
- Each test fits the question and the way the outcome is measured, and the plan says why.
- Data preparation is described: screening, missing data, scale scoring, reliability.
- Assumptions are named for each test, with how you will check them and what you will do if one fails.
- Alpha, effect sizes, and any adjustment for multiple comparisons are stated.
- The power analysis is for the test that needs the largest sample.
- Software is named in a sentence, not described in a page.
- For a mixed methods study, the plan says where and how the strands are integrated.
Get the analysis plan worksheet
A free two-page Word file: a blank alignment table, a table for your assumptions and what you will do if they fail, a fill-in version of the analysis paragraph, and the checklist.
I'll email you once or twice about dissertation help. No newsletter. Reply "stop" any time.
Related guides
- How to Justify Your Sample Size in a Dissertation Proposal
- How to Write the Results Chapter of a Dissertation
- Choosing the Right Statistical Software
- How to Conduct Thematic Analysis That Committees Find Credible
Work with Matt
If you want the plan built with you before your proposal defense, that's the Analysis Plan ($595): research questions mapped to variables and tests, power analysis, assumptions, and the reasoning behind each choice, with one revision included. Stuck on one question, such as which test fits? The Methods Consult ($195) is a 90-minute working session with a written summary you can bring to your chair.