Choosing the wrong statistical test is one of the most common errors in published life science research, and it happens not because researchers are careless but because the landscape of available tests is genuinely confusing. A t-test, a Mann-Whitney U test, and a one-way ANOVA all involve comparing groups, but they are not interchangeable.
This guide gives you a practical decision framework for choosing the right test for the most common experimental designs in life science research. It covers the key assumptions, the most common violations, and the specific tests to reach for in each situation.
The Core Question Before Any Test: What Are You Actually Comparing?
Before worrying about which test to use, get precise about what you are trying to know. Most statistical comparisons in biology fall into one of four categories:
- Comparing two groups (treatment vs. control)
- Comparing three or more groups (multiple treatments, multiple time points)
- Assessing relationship between two variables (correlation, regression)
- Comparing proportions or frequencies (categorical outcomes, survival data)
Within each category, the choice of test depends on two additional questions: Are the data normally distributed? Are the groups independent or paired (repeated measurements from the same subjects)?
Getting these questions right before you run a test will get you to the correct analysis in most cases.
Parametric vs. Non-Parametric Tests: What Actually Matters
The distinction between parametric and non-parametric tests is central to choosing correctly, and it is often misunderstood.
Parametric tests (t-test, ANOVA, Pearson correlation) assume that your data come from a population with a specific distribution, usually a normal (Gaussian) distribution. When this assumption holds, parametric tests are more powerful, meaning they can detect a real difference more reliably with smaller sample sizes.
Non-parametric tests (Mann-Whitney U, Kruskal-Wallis, Spearman correlation) make no distributional assumption. They work by ranking values rather than using the raw data. They are less powerful than their parametric equivalents when the data actually are normally distributed, but they are the correct choice when the normality assumption is violated.
The important nuance: for reasonably large samples (n > 30 per group in most cases), parametric tests are quite robust to departures from normality because of the central limit theorem. The normality assumption matters most with small samples, which is exactly what most biology experiments have.
How to Test for Normality
The most practical approach for small biological samples is to use the Shapiro-Wilk test, which is more sensitive to departures from normality in small samples than the Kolmogorov-Smirnov test. In R, this is shapiro.test(). In GraphPad Prism, it is available under the column statistics analysis.
Practical caveat: with very small samples (n = 3 to 5), which are common in cell biology and animal experiments, the Shapiro-Wilk test has low statistical power to detect non-normality. In these cases, visual inspection of the data and knowledge of the measurement type matter more than a formal test result. Many biological measurements (cell counts, expression ratios, enzyme activity) are naturally right-skewed or log-normal.
Comparing Two Groups
This is the most common scenario in biology: treated vs. untreated, knockout vs. wild-type, drug vs. vehicle.
Independent Two-Group Comparisons
When to use Student’s t-test: Two independent groups, data approximately normally distributed, variances roughly equal. This is appropriate for well-powered experiments with continuous measurements that you expect to follow a roughly normal distribution (body weight, many biochemical measurements, behavioral scores in animal studies).
When to use Welch’s t-test: Two independent groups, data approximately normal, but the variances are unequal between groups. Welch’s t-test does not assume equal variances. In most statistical software (R, Prism), Welch’s is now the default t-test, which is correct practice. If you are using software that still defaults to Student’s t-test, switch to Welch’s or explicitly test variance equality first.
When to use Mann-Whitney U (Wilcoxon rank-sum test): Two independent groups, data are not normally distributed, or you have small samples where normality cannot be confirmed. This is appropriate for ordinal data, for skewed continuous data (many expression fold-changes, viability assays), and for small n experiments (n = 3 to 8 per group) where the normality assumption is uncertain.
The Mann-Whitney U test is not a test of means. It tests whether one group tends to have larger values than the other (technically, whether one distribution is stochastically greater than the other). This distinction matters for how you report results.
Paired Two-Group Comparisons
Paired measurements occur when you measure the same subject twice: before and after treatment, matched patient samples, or left vs. right side comparisons. Ignoring the pairing and using an independent test is a mistake that discards information and reduces power.
When to use paired t-test: Same subjects measured twice, data approximately normally distributed. Examples: body weight before and after treatment in the same animals, gene expression in matched tumor-normal pairs.
When to use Wilcoxon signed-rank test: Same subjects measured twice, data not normally distributed or small sample size. The paired non-parametric equivalent of the paired t-test.
Comparing Three or More Groups
Adding a third group changes everything. Running multiple pairwise t-tests is incorrect because each test carries a Type I error rate (false positive rate), and performing multiple tests inflates the overall error rate. With three groups and three pairwise comparisons, the chance of at least one false positive at alpha = 0.05 is already close to 14%. With five groups and ten comparisons, it exceeds 40%.
The correct approach is to use a single omnibus test first, then apply appropriate post-hoc correction if the omnibus test is significant.
Independent Multi-Group Comparisons
When to use one-way ANOVA: Three or more independent groups, data approximately normally distributed, roughly equal variance across groups. ANOVA tests whether any of the group means differ from each other. A significant result tells you that at least one group is different, not which specific pairs differ. You then apply a post-hoc test (Tukey HSD is most common for all pairwise comparisons; Dunnett’s test when comparing multiple groups to a single control).
When to use Kruskal-Wallis test: Three or more independent groups, data not normally distributed or ordinal. The non-parametric equivalent of one-way ANOVA. After a significant Kruskal-Wallis result, use Dunn’s test for post-hoc pairwise comparisons with appropriate multiple comparison correction.
Two-Factor Experiments
Many biological experiments involve two independent variables simultaneously: for example, drug treatment (treated vs. untreated) crossed with genotype (wild-type vs. knockout). Looking at each factor separately with individual ANOVAs is incorrect because it misses the interaction between factors.
When to use two-way ANOVA: Two independent variables, data approximately normally distributed. Two-way ANOVA tests the effect of each factor and their interaction. The interaction term is often biologically the most interesting result: does the drug work differently in mutant vs. wild-type animals?
Post-hoc corrections after two-way ANOVA: Sidak correction is appropriate for pre-planned comparisons; Tukey HSD for exploratory pairwise comparisons.
Repeated-Measures Designs
When the same subjects are measured at multiple time points or under multiple conditions, use repeated-measures ANOVA (or its non-parametric equivalent, Friedman’s test). This accounts for the correlation between measurements from the same individual and substantially increases statistical power.
Correlation and Regression
Pearson correlation: Measures the linear relationship between two continuous variables. Assumes both variables are normally distributed. Appropriate for assessing linear relationships in well-powered continuous data.
Spearman rank correlation: Measures the monotonic relationship between two variables without assuming normality. Appropriate for skewed data, ordinal data, or when one or both variables contain outliers. In genomics and many biological datasets, Spearman is often the more appropriate default.
Linear regression: Goes beyond correlation to model the actual relationship between variables and generate predictions. Use when you want to quantify how much one variable changes per unit change in another, or when controlling for covariates.
A critical point: correlation does not equal causation, and this is not just a methodological cliche. Many spurious correlations exist in biological datasets because of batch effects, confounding variables, and multiple testing. Always think carefully about what a significant correlation could mean causally before drawing conclusions.
Survival Analysis
For time-to-event outcomes (time to death, time to tumor formation, time to disease progression), survival analysis methods are required. Using a t-test on survival times is incorrect because it ignores censoring, which occurs when subjects are still alive at the end of the study or are lost to follow-up.
Kaplan-Meier curves are the standard way to visualize survival data. The log-rank test is the standard statistical test for comparing survival between two groups. For more than two groups or when controlling for covariates, the Cox proportional hazards model is appropriate.
Common Mistakes in Life Science Statistics
Running multiple t-tests instead of ANOVA. This inflates the false positive rate. Use ANOVA plus post-hoc correction.
Ignoring the difference between independent and paired designs. If your measurements are paired, use a paired test. You will gain statistical power and get a more accurate result.
Using parametric tests on very small samples without justification. With n = 3 per group, you have essentially no ability to verify normality. Consider whether non-parametric tests or exact permutation tests are more appropriate.
Testing for normality on the raw data rather than the residuals. For ANOVA, what matters is whether the residuals are normally distributed, not whether the raw data are. This is a subtle but real distinction.
Applying Bonferroni correction when tests are correlated. Bonferroni correction is highly conservative and assumes all tests are independent. For correlated tests (e.g., multiple measurements on the same patients), it over-corrects. Benjamini-Hochberg false discovery rate (FDR) correction is more appropriate in most high-throughput biological settings.
Reporting only p-values without effect sizes. A statistically significant result with a tiny effect size may be biologically meaningless. Always report effect sizes (Cohen’s d, fold change, odds ratio) alongside p-values. Sullivan and Feinn (2012) in the Journal of Graduate Medical Education provide a clear discussion of why effect sizes are essential for interpreting statistical results in clinical and biological research.
Not pre-registering or pre-specifying tests. Running multiple tests and reporting only the significant ones (p-hacking) is a significant source of false findings in the biological literature. Where possible, specify your primary statistical test before data collection.
A Quick Decision Guide
Use this as a starting point, not a rigid rule:
| Scenario | Normal data | Non-normal or small n |
|---|---|---|
| 2 independent groups | Welch’s t-test | Mann-Whitney U |
| 2 paired groups | Paired t-test | Wilcoxon signed-rank |
| 3+ independent groups | One-way ANOVA + post-hoc | Kruskal-Wallis + Dunn’s |
| 3+ conditions, same subjects | Repeated-measures ANOVA | Friedman’s test |
| 2 continuous variables | Pearson correlation | Spearman correlation |
| Time-to-event data | Log-rank test / Cox model | Log-rank test / Cox model |
| Two factors (factorial design) | Two-way ANOVA | Aligned rank transform + two-way ANOVA |
Software Tools
GraphPad Prism remains the most commonly used statistical software in wet lab biology. It guides you through test selection with built-in analysis wizards and handles the most common biological designs well. If you are not yet using a statistical software package and want something approachable, Prism is the practical starting point for most wet lab researchers.
R with packages such as rstatix, coin, and emmeans gives you the full range of statistical methods with complete control. It has a steeper learning curve but is the standard in bioinformatics and increasingly in experimental biology.
JASP is a free, open-source alternative with a point-and-click interface that covers most common tests and adds Bayesian analysis options.
Next Steps
The decision framework here covers the most common scenarios in life science research, but statistical design is genuinely deep. For a thorough treatment of experimental design and analysis for biologists, Zar’s Biostatistical Analysis remains the definitive reference text and is worth having on your shelf if you regularly design and analyze experiments. For a more approachable introduction focused specifically on biologists, Whitlock and Schluter’s The Analysis of Biological Data is widely used in graduate courses and explains the reasoning behind each method clearly.
The most important rule is to decide on your primary statistical test before you collect data, not after you have looked at the results. Post-hoc test selection based on the direction of the data is a subtle form of p-hacking, even when done unconsciously.