Dunnett's Test: When and How to Use It

What Is Dunnett's Test?

After running a one-way ANOVA and finding a significant F-statistic, you know at least one group mean differs — but not which one. Post hoc tests answer that follow-up question. The choice of test depends entirely on what comparisons the study was designed to make.

When you need every group compared to every other group, Tukey's HSD or the Bonferroni correction are appropriate. When one group is a natural control and the only question is whether each treatment differs from it, Dunnett's Test is the right tool — and it delivers greater power for each individual comparison.

Why Dunnett's Test Exists: The Multiple Comparisons Problem

Every statistical test at α = 0.05 carries a 5% false-positive risk. Run multiple tests on the same dataset and that risk compounds. With k treatment groups and one control, you face k−1 pairwise comparisons. If each uses α = 0.05 independently, the family-wise error rate reaches 1 − (0.95)^(k−1). With four treatments that figure climbs to roughly 14% — a 14% chance of at least one spurious result in the set.

Dunnett's Test adjusts its critical values so the probability of any false rejection across all k−1 comparisons stays exactly at α. It does this more efficiently than the Bonferroni correction because it uses the exact joint distribution of the test statistics, which are correlated through the shared control group and shared MSE. That correlation means the actual joint error rate under Bonferroni is below α — making Bonferroni needlessly conservative. Dunnett's Test exploits that correlation to set tighter (less conservative) critical values and gain power.

The Dunnett's Test Formula

The test statistic for comparing treatment group i to the control group (group 0) is:

The statistic dᵢ follows a multivariate t-distribution. Its critical value d* is obtained from Dunnett's table at significance level α, with k (the number of treatment groups, not counting the control) and ν = N − k − 1 degrees of freedom where N is the total sample size across all groups.

Decision Rule

For each treatment group i, compare the absolute value |dᵢ| to the two-tailed critical value d*(α, k, ν):

• If |dᵢ| ≥ d* → Reject H₀. Statistically significant evidence that treatment i differs from the control.
• If |dᵢ| < d* → Fail to reject H₀. Insufficient evidence of a difference at the chosen α level.

For a one-tailed test — for example, testing whether a drug reduces blood pressure rather than merely changes it — compare the signed dᵢ to the one-tailed critical value and reject only when the sign matches the predicted direction.

Simultaneous Confidence Intervals

Dunnett's procedure also produces simultaneous 95% confidence intervals for each difference μᵢ − μ₀. All k−1 intervals jointly cover the true differences with probability 1 − α:

An interval that excludes zero indicates a statistically significant difference between that treatment and the control. Report these intervals alongside p-values — they communicate the direction and plausible magnitude of each effect, not just whether it clears a significance threshold.

Assumptions of Dunnett's Test

Dunnett's Test shares the assumptions of one-way ANOVA. Violating them can produce inaccurate p-values and incorrect conclusions.

When to Use Dunnett's Test

Real-World Applications

Worked Example 1: Drug Efficacy Study

This is the most common application context. A pharmaceutical company tests three drug doses against a placebo control, measuring systolic blood pressure (mmHg) after four weeks of treatment.

Worked Example 2: Agricultural Fertilizer Trial

An agronomist tests four fertilizer treatments against an untreated control, recording wheat yield (kg per plot) across a growing season. The control-versus-treatment design is standard in field trials and maps directly onto Dunnett's Test.

Dunnett's Test vs. Tukey HSD vs. Bonferroni vs. Scheffé

Choosing among post hoc tests is one of the most common sources of confusion in applied statistics. The comparison below covers the four main options across every dimension that matters for a real analysis decision.

Step-by-Step Procedure

Dunnett's Critical Values Reference Table (α = 0.05, Two-Tailed)

This table gives approximate two-tailed critical values d* at α = 0.05. Rows are error degrees of freedom (ν) from the ANOVA; columns are the number of treatment groups k, not counting the control. For the complete table including α = 0.01 and one-tailed values, see Dunnett's Table.

Running Dunnett's Test in Software

R

The DescTools package provides the most direct implementation. The multcomp package offers additional flexibility for contrast plots and confidence intervals.

# install.packages("DescTools")
library(DescTools)

# Data: Drug efficacy study
bp <- c(
  131,129,133,128,132,130,127,131,133,126,  # Control
  125,123,126,122,125,124,121,126,127,121,  # Drug A
  118,117,120,116,119,118,115,120,119,117,  # Drug B
  112,111,114,110,113,112,110,113,111,114   # Drug C
)
grp <- factor(rep(c("Control","DrugA","DrugB","DrugC"), each=10))

# Step 1: One-way ANOVA
model <- aov(bp ~ grp)
summary(model)

# Step 2: Dunnett's Test — specify the control group
DunnettTest(x=bp, g=grp, control="Control", conf.level=0.95)

library(multcomp)
model <- aov(bp ~ grp)

# Dunnett contrasts vs. control (alphabetically first)
dc <- glht(model, linfct=mcp(grp="Dunnett"))
summary(dc)      # adjusted p-values
confint(dc)     # simultaneous confidence intervals
plot(confint(dc))  # CI plot

Python

# pip install scikit-posthocs scipy
import numpy as np
import pandas as pd
from scipy import stats
import scikit_posthocs as sp

bp_data = {
    "Control": [131,129,133,128,132,130,127,131,133,126],
    "DrugA":   [125,123,126,122,125,124,121,126,127,121],
    "DrugB":   [118,117,120,116,119,118,115,120,119,117],
    "DrugC":   [112,111,114,110,113,112,110,113,111,114]
}

values = np.concatenate(list(bp_data.values()))
labels = np.repeat(list(bp_data.keys()), [len(v) for v in bp_data.values()])

# ANOVA first
f, p = stats.f_oneway(*bp_data.values())
print(ff"F = {f:.3f}, p = {p:.4f}")

# Dunnett's Test vs. control
df = pd.DataFrame({"val": values, "group": labels})
result = sp.posthoc_dunnett(df, val_col="val",
                             group_col="group",
                             control="Control")
print(result)

SPSS

1. Go to Analyze → Compare Means → One-Way ANOVA
2. Move the dependent variable to the Dependent List and the grouping variable to Factor
3. Click Post Hoc
4. Under "Tests that assume equal variance," check Dunnett
5. In the "Control Category" field, specify whether control is the first or last category, or enter the value directly
6. Set α = 0.05 and click Continue → OK

Worked Example 3: Manufacturing Quality Control

Effect Size: Cohen's d for Each Comparison

Statistical significance tells you whether a difference is real given your sample size. Effect size tells you how large that difference is in practical terms. For each Dunnett comparison, report Cohen's d:

For Drug C in Example 1: Cohen's d = (112.1 − 130.2) / √64 = −18.1 / 8 = −2.26. This is a very large effect — the blood pressure reduction is more than twice the pooled standard deviation. A statistically significant result with a tiny effect size may have limited clinical or practical relevance; always report both. See the full effect size guide for more on interpreting Cohen's d.

APA Reporting Format

The APA Publication Manual (7th edition) does not specify a unique template for Dunnett's Test, so follow the general guidance for post hoc comparisons. The key elements are the ANOVA F-statistic, the name of the post hoc test, group means and SDs, mean differences, d statistics, adjusted p-values, and confidence intervals.

Interactive Dunnett's Test Calculator

Enter group means, sample size per group (balanced design), and the ANOVA MSE to compute all Dunnett d statistics and compare each to the approximate critical value. For exact critical values at any ν and k, use R or the full Dunnett's table.

Common Mistakes and How to Avoid Them

Frequently Asked Questions

Summary