Small sample size is one of the most stressful problems in dissertation research. A student may design a study expecting 150 participants, collect data for months, and end up with 42 usable responses. Another may have enough participants overall, but too few cases in one subgroup to run the originally planned analysis. A third may have a survey dataset where missing data reduces the final analytic sample far more than expected.
When this happens, the question is often urgent:
What analyses are still defensible with a small sample size?
The answer depends on the research questions, the measurement level of the variables, the number of predictors, the distribution of the data, and the expectations of the dissertation committee. A small sample does not automatically make analysis impossible. But it does limit what can be estimated, how much statistical power the study has, and how strongly the results can be interpreted.
This post explains how to think about small sample size in a dissertation and what types of statistical analyses may still be appropriate.
Small Sample Size Is a Design and Interpretation Problem
Small sample size is often discussed as if it is only a power problem. Power matters, but it is not the only issue.
A small sample can affect a dissertation in several ways:
- Lower statistical power: The study may have difficulty detecting real effects.
- Less precise estimates: Confidence intervals may be wide.
- Unstable models: Regression coefficients may change substantially if a few cases are added or removed.
- Difficulty checking assumptions: Normality, outliers, and model fit can be harder to evaluate.
- Sparse categories: Some groups may have too few cases for chi-square tests, logistic regression, or subgroup comparisons.
- Reduced generalizability: The final sample may not represent the target population well.
These problems do not mean the study must be abandoned. They mean the analysis needs to be carefully matched to what the data can support.
What Counts as a Small Sample?
There is no single number that defines a small sample. A sample of 40 may be small for a multiple regression with ten predictors, but adequate for a focused descriptive analysis. A sample of 100 may be large enough for some comparisons but too small for a complex structural equation model.
Rather than asking only, “Is my sample size small?” it is better to ask:
- How many usable cases are available for each research question?
- How many groups are being compared?
- How many predictors are in the model?
- Is the outcome continuous, binary, ordinal, categorical, or a count?
- Are there rare categories or sparse cells?
- How large are the expected effects?
- Are the results intended to be confirmatory or exploratory?
Sample size has to be judged relative to the analysis. A simple analysis may be defensible with a modest sample. A complex model may not be.
Start With the Dissertation Research Questions
When the sample is smaller than expected, the first step is to return to the research questions and hypotheses.
Ask whether each research question can still be answered with the available data. Some questions may remain fully answerable. Others may need to be treated as exploratory, simplified, or removed with committee approval.
For example, a dissertation might originally include these research questions:
- Do three groups differ in burnout?
- Do age, workload, trauma exposure, training, and support predict burnout?
- Does support moderate the relationship between workload and burnout?
If the final sample is small, the first question might still be addressed with caution, depending on group sizes. The second question may require reducing the number of predictors. The third question may be difficult to defend if the study is underpowered for interaction effects.
The goal is not to force every planned analysis into a dataset that cannot support it. The goal is to produce the most defensible analysis possible.
Descriptive Statistics Are Often Still Valuable
With a small sample, descriptive statistics become especially important.
Descriptive analysis may include:
- Sample demographics
- Means and standard deviations for continuous variables
- Medians and interquartile ranges for skewed variables
- Frequencies and percentages for categorical variables
- Reliability estimates for multi-item scales
- Missing-data summaries
- Graphs or tables showing distributions
In some dissertations, the descriptive results may be among the most useful findings. This is especially true for exploratory studies, pilot studies, needs assessments, program evaluations, rare populations, or specialized professional groups.
Descriptive statistics do not solve the power problem, but they help readers understand the data and the context for any inferential tests that follow.
Group Comparisons With Small Samples
Many dissertation studies compare groups. Examples include treatment vs. control, men vs. women, students in different programs, or participants with and without a particular experience.
With small samples, group comparisons require extra caution.
| Question | Possible analysis | Small-sample caution |
|---|---|---|
| Two independent groups, continuous outcome | Independent-samples t-test or Mann-Whitney U test | Check group sizes, outliers, and distributional assumptions |
| Two paired measurements | Paired-samples t-test or Wilcoxon signed-rank test | Focus on the distribution of change scores |
| Three or more groups, continuous outcome | ANOVA or Kruskal-Wallis test | Very small group sizes can make results unstable |
| Two categorical variables | Chi-square test or Fisher’s exact test | Fisher’s exact test may be better when expected cell counts are small |
Nonparametric tests are sometimes used with small samples, especially when distributions are skewed or outliers are severe. However, nonparametric tests are not a universal fix. They still require enough data to produce meaningful results, and they may test slightly different questions than their parametric counterparts.
Regression With a Small Sample
Regression is common in dissertation research, but it can become risky when the sample is small.
The main issue is the ratio of usable cases to predictors. A model with one or two predictors may be defensible with a modest sample. A model with ten predictors, interaction terms, categorical predictors with multiple levels, and control variables may not be.
With small samples, regression models can produce:
- Large standard errors
- Wide confidence intervals
- Unstable coefficients
- Overfitting
- Difficulty detecting assumption violations
- Results driven by a small number of influential cases
This does not mean regression is impossible. It means the model should be simple, theoretically justified, and carefully reported.
Strategies for Small-Sample Regression
Possible strategies include:
- Limit the number of predictors.
- Prioritize predictors based on theory or the dissertation’s primary research questions.
- Avoid unnecessary control variables.
- Avoid testing many interaction terms unless the sample can support them.
- Check for influential cases.
- Report confidence intervals, not just p-values.
- Interpret nonsignificant findings cautiously.
- Consider sensitivity analyses when appropriate.
In dissertation writing, it is usually better to present a simpler defensible model than a complex model that the data cannot support.
Logistic Regression and Sparse Data
Logistic regression can be especially difficult with small samples. The issue is not just the total number of cases, but the number of cases in each outcome category.
For example, suppose a study predicts whether participants have a particular diagnosis. If 45 participants have the diagnosis and only 3 do not, the binary outcome has very limited variation. A logistic regression model may be unstable or impossible to estimate reliably.
Problems can also occur when a predictor perfectly or nearly perfectly separates the outcome. For example, if every participant with a certain characteristic has the outcome, the model may produce extremely large coefficients, warnings, or unreliable estimates.
When logistic regression is not defensible, alternatives may include:
- Descriptive reporting
- Cross-tabulations
- Fisher’s exact test for simple categorical comparisons
- Combining categories, if theoretically justified
- Reducing the number of predictors
- Treating the analysis as exploratory
The right choice depends on the research question and the data structure.
Correlation With Small Samples
Correlation can be used with small samples, but interpretation must be cautious. A correlation based on a small sample can be strongly affected by outliers and may have a wide confidence interval.
For continuous variables, Pearson correlation may be appropriate when the relationship is approximately linear and the data are not dominated by outliers. Spearman correlation may be more appropriate for ordinal variables, skewed variables, or monotonic relationships that are not well described by a straight line.
When reporting correlations from a small sample, it is useful to include the sample size and avoid overinterpreting p-values. A nonsignificant correlation in a small sample does not necessarily prove that no relationship exists. It may simply mean the study lacked power.
When to Use Nonparametric Tests
Nonparametric tests are often considered when sample size is small or assumptions are questionable. Common examples include:
- Mann-Whitney U test instead of an independent-samples t-test
- Wilcoxon signed-rank test instead of a paired-samples t-test
- Kruskal-Wallis test instead of one-way ANOVA
- Spearman correlation instead of Pearson correlation
- Fisher’s exact test instead of chi-square test
These tests can be useful, but they should not be chosen automatically. The test still needs to match the research question and the measurement level of the variables.
For example, the Mann-Whitney U test does not simply compare means in the same way an independent-samples t-test does. Depending on the distributions, it may be better understood as comparing the relative ranks or distributions of the two groups. That may be appropriate, but it should be interpreted correctly.
Effect Sizes Matter More With Small Samples
In small-sample dissertation research, effect sizes are especially important.
A p-value answers a narrow question: assuming the null hypothesis is true and the model assumptions hold, how surprising are the data? It does not tell you whether an effect is large, important, or practically meaningful.
With a small sample, a potentially meaningful effect may not reach statistical significance. Conversely, a statistically significant finding in a small sample may be unstable if it depends heavily on a few cases.
Useful effect-size measures may include:
- Cohen’s d for mean differences
- Eta squared or partial eta squared for ANOVA-type analyses
- Correlation coefficients for associations
- Odds ratios for logistic regression
- Regression coefficients and standardized coefficients
- Confidence intervals around key estimates
A defensible dissertation results section should discuss both statistical significance and substantive meaning.
Confidence Intervals Are Often More Informative Than p-Values Alone
Confidence intervals are useful because they show the range of values that are reasonably compatible with the data and model.
In a small sample, confidence intervals are often wide. That is not a reporting problem; it is useful information. A wide confidence interval tells the reader that the estimate is imprecise.
For example, suppose a treatment group has lower anxiety than a comparison group, but the confidence interval for the mean difference is very wide. The result may suggest a potentially useful effect, but the study may not estimate that effect precisely. That is a more nuanced and honest interpretation than simply saying the result was significant or not significant.
Can You Still Answer the Research Questions?
Sometimes the correct answer is yes, with caution. Sometimes it is partially. Sometimes it is no.
A small sample may still support:
- Descriptive analysis
- Simple group comparisons
- Simple correlations
- Simple regression models
- Exploratory analysis
- Preliminary or pilot-study conclusions
A small sample may not support:
- Complex regression models with many predictors
- Multiple interaction effects
- Subgroup analyses with very small groups
- Structural equation modeling with many parameters
- Hierarchical models with too few clusters
- Logistic regression with sparse outcome categories
- Strong claims of no effect based only on nonsignificant p-values
If a dissertation’s original analysis plan is no longer defensible, the student may need to consult the committee about revising the analysis plan. This should be done carefully and transparently.
How to Write About Small Sample Size as a Limitation
Small sample size should be discussed clearly, but not in a way that unnecessarily undermines the entire dissertation.
A weak limitation statement would be:
The study had a small sample size.
A stronger statement would be:
The final analytic sample was smaller than originally planned, which reduced statistical power and limited the complexity of the models that could be estimated. As a result, nonsignificant findings should be interpreted cautiously, and the results should be viewed as preliminary rather than definitive.
If the small sample affected specific analyses, the limitation should say so:
Because some subgroups contained relatively few participants, subgroup comparisons were interpreted cautiously. The small cell sizes limited the ability to detect group differences and may have produced unstable estimates.
A good limitation statement explains what the limitation was, how it affected the analysis, and how it should shape interpretation.
How to Strengthen a Small-Sample Dissertation
Even when the sample is small, there are ways to make the dissertation stronger.
- Be transparent about the final sample size.
- Report missing data and exclusions clearly.
- Use analyses that match the available data.
- Keep models simple and theory-driven.
- Report effect sizes and confidence intervals.
- Avoid overstating nonsignificant findings.
- Use sensitivity analyses when useful.
- Connect findings back to the research questions.
- Discuss limitations honestly.
- Make practical recommendations that are proportional to the evidence.
The committee may be more receptive to a modest but defensible analysis than to an overly complex analysis that is not supported by the data.
Final Thoughts
A small sample size is not ideal, but it is also not unusual in dissertation research. Many students work with hard-to-reach populations, limited recruitment windows, low response rates, incomplete surveys, or specialized professional samples.
The key is to choose analyses that the data can reasonably support. Descriptive statistics, simple comparisons, correlations, and carefully limited regression models may still be defensible. Complex models, excessive predictors, sparse logistic regression, and overinterpreted subgroup analyses may not be.
The best dissertation analysis is not the most advanced analysis. It is the analysis that best answers the research questions while respecting the limits of the data.
If you are working with a small dissertation sample and are unsure which analyses are still defensible, dissertation statistical analysis support can help you review your research questions, dataset, sample size, missing data, and proposed analysis plan before you commit to an approach.