One Way ANOVA Calculator – Free Online F Test Tool

A one way ANOVA tests whether three or more group means differ, by dividing the variation between groups by the variation inside them to give an F ratio.

One factor, one continuous outcome, independent groups. A large F means the group means are spread further apart than the noise inside each group can explain. The p value says how often that would happen by chance if every population mean were identical. The test tells you that at least one mean differs, never which one, so a post hoc test comes next.

Rule of thumb: reject the null when p is below 0.05. Worked example on this page: three teaching methods, 8 students each, F(2, 21) = 19.06, p < 0.001, η² = 0.645, so group explains 64.5% of the variance, a large effect.

A one way ANOVA, short for one way analysis of variance, extends the independent samples t test to three or more groups. It tests the null hypothesis that all group population means are equal, by partitioning the total variance in the outcome into a between groups component and a within groups component. One categorical factor, one continuous dependent variable, independent observations.

What it actually measures. Imagine every observation sitting on a number line. Some of the spread comes from the fact that people are in different groups, and some comes from the fact that people inside the same group still differ. ANOVA measures both, then divides one by the other. If group membership matters, the between groups part is large relative to the within groups part and F is well above 1. If group membership means nothing, both parts estimate the same error variance and F sits near 1.

Why not just run several t tests? Because each test carries its own 5% false positive risk. With three groups you would run three t tests and the chance of at least one false positive rises to about 14%. With five groups you would run ten tests and the risk passes 40%. ANOVA asks the question once, at one alpha level, and keeps the error rate where you set it.

A worked one-liner. Three teaching methods, 8 students each. Means 53.00, 65.00 and 56.63. SS_between = 606.08 over 2 degrees of freedom gives MS_between = 303.04. SS_within = 333.88 over 21 degrees of freedom gives MS_within = 15.90. So F = 303.04 / 15.90 = 19.06, p < 0.001.

Who uses it. Psychologists comparing three experimental conditions. Agronomists comparing fertiliser treatments across field plots. Clinical researchers comparing a placebo with two drug doses. Educators comparing teaching methods. Manufacturing engineers comparing machines or suppliers. Ecologists comparing species richness across habitat types. Anywhere one grouping variable splits a continuous measurement into three or more independent samples.

1.1 One way ANOVA compared with its nearest relatives

The most common mistake is using a one way ANOVA when the same people were measured more than once. That is a repeated measures design, and treating it as independent groups throws away the pairing and usually loses power.

Nothing is computed until you press the button. Changing any input clears the results so you never read stale numbers.

4.1 What the F ratio is, in words

F is a ratio of two variances. The numerator, the between groups mean square, asks how far apart the group means are. The denominator, the within groups mean square, asks how much people differ from their own group mean. Both are estimates of the same error variance when the null hypothesis is true, so F sits near 1. When the groups really differ, only the numerator grows, and F climbs above 1.

That is why F can never be negative and why the test is always one tailed on the right. There is no such thing as a significantly small F in a normal ANOVA. A very small F, well below 1, sometimes signals a different problem: data that are too neat, such as fabricated numbers or a design where the groups were matched too closely.

4.2 Reading the p value correctly

The p value is the probability of getting an F ratio at least as large as yours if every population mean were truly identical. It is not the probability that the null hypothesis is true, and 1 minus p is not the probability that your alternative is true. A p of 0.03 does not mean there is a 97% chance the groups differ.

Compare p with the alpha you set before seeing the data, usually 0.05. Below alpha, you reject the null and say at least one mean differs. At or above alpha, you retain the null and say this study could not detect a difference. Retaining the null is never proof that the groups are the same, only that the evidence was not strong enough to separate them.

4.3 Degrees of freedom, and why they must be reported

A one way ANOVA has two degrees of freedom values and both belong in the write up. The first, df₁ = k − 1, counts the number of independent comparisons between group means. The second, df₂ = N − k, counts the observations left over once each group mean has been estimated.

Written as F(2, 21) = 19.06, the reader can immediately reconstruct your design: 3 groups and 24 observations. That is why journals insist on the bracketed pair. If you see F(2, 21) but the paper reports 30 participants, something is wrong, usually undeclared missing data.

4.4 Effect size: eta squared, omega squared and Cohen's f

The p value tells you whether an effect is detectable. The effect size tells you how big it is, and only the second one is comparable across studies. Eta squared is the share of the total sum of squares that belongs to the between groups term, so it is read directly as a percentage of variance explained. Cohen's benchmarks are 0.01 small, 0.06 medium and 0.14 large, and they are conventions rather than laws.

Eta squared is biased upward, because sampling noise always inflates the between groups sum of squares a little. Omega squared subtracts an estimate of that noise and is the better statistic when group sizes are small, which is why this calculator reports both. With 30 or more per group the two are usually close. With 5 per group they can differ noticeably, and the honest choice is omega squared.

Cohen's f is a third scale, defined as the square root of eta squared over one minus eta squared. It exists because G*Power and most power calculators ask for f rather than eta squared. Benchmarks are 0.10 small, 0.25 medium and 0.40 large.

4.5 The omnibus trap: F is significant, so what now?

A significant F is not the finding. It is permission to go looking for the finding. It says at least one of the k means differs from at least one other, which with four groups leaves six possible pairwise stories and no way to choose between them from F alone.

Post hoc tests answer the follow-up question while keeping the family-wise error rate under control. Tukey HSD is the standard choice for equal variances and roughly equal group sizes. Games-Howell is the right choice when variances differ, because it uses each pair's own variances and its own degrees of freedom. Bonferroni works anywhere but is conservative, and it loses power quickly as the number of comparisons grows.

4.6 What to do when Levene's test fails

Levene's test asks whether group variances are equal. When its p value falls below 0.05, the classic F test's denominator is no longer a fair estimate of a single common error variance, and the reported p value can be badly wrong, especially with unequal group sizes.

The direction of the error is worth knowing. If the larger group also has the larger variance, the F test becomes conservative and you lose power. If the smaller group has the larger variance, F becomes liberal, meaning your true false positive rate is above the 5% you think you are running. That second case is the dangerous one. Welch's one way ANOVA fixes both, and modern statistical advice increasingly says to use Welch by default rather than testing first and switching.

4.7 What to do when normality fails

ANOVA assumes the residuals are normal, not the raw data, and not each group's raw data. With equal group sizes the F test is remarkably robust: even quite skewed residuals leave the false positive rate close to 5% once each group has around 15 observations, thanks to the central limit theorem.

The failures that matter are heavy outliers and strong skew in small samples. Inspect the Q-Q plot rather than relying on the Shapiro-Wilk p value alone, because with large N the test flags trivial departures and with small N it misses real ones. If the Q-Q plot bends badly, options are a transformation such as the log or square root, a Kruskal-Wallis test on ranks, or a bootstrap or permutation version of the F test.

4.8 Sample size, power and the non-significant result

A non-significant ANOVA has two possible causes, and they are very different. Either there is genuinely no difference, or there is one and this study was too small to see it. The p value cannot tell you which.

The way to tell them apart is a power calculation using the Cohen's f the calculator reports. As a rough guide, detecting a medium effect (f = 0.25) across three groups at 80% power needs about 53 observations per group. Detecting a large effect (f = 0.40) needs about 22 per group. If your study had 10 per group, a null result is nearly uninformative, and saying so is better science than claiming the groups are the same.

4.9 Confidence intervals do more work than the p value

Every group mean in the results table carries a 95% confidence interval built on the pooled error term, and every post hoc comparison carries an interval on the difference. Those intervals carry more information than the p value, because they show both the direction and the plausible size of the effect.

The reading rule is simple. If a difference interval excludes zero, that pair differs at your alpha. If it includes zero, it does not. But an interval running from −0.2 to 0.4 says something quite different from one running from −18 to 22, even though both include zero and both give p above 0.05. The first rules out a large effect; the second rules out nothing.

4.10 Common mistakes this calculator prevents

• Running three t tests instead of one ANOVA. The calculator does the omnibus test first and only then breaks it down, with the correction applied.
• Reporting a p value with no effect size. Four effect sizes are computed automatically, and the APA sentence includes one.
• Ignoring unequal variances. Levene's test runs on every analysis, and Welch's F is computed alongside the classic F whether you asked for it or not.
• Quoting F without its degrees of freedom. Every output block writes them as F(df1, df2).
• Reading post hoc tests after a non-significant F. The tool prints a warning when you do.
• Treating repeated measurements as independent groups. The assumption panel names this explicitly as something only you can confirm.
• Rounding p to 0.000. Small p values are shown as less than 0.0001 rather than as zero, which is never a correct p value.

4.11 How to decide what to report

4.12 A note on multiple outcomes

Running a separate one way ANOVA on each of ten outcome variables reintroduces exactly the multiplicity problem ANOVA was designed to solve, one level up. If the outcomes are conceptually one construct, MANOVA is the right test. If they are separate questions, say so in advance, pre-register them, and correct across the family with Holm or a false discovery rate procedure. Silence on this point is the single most common reason a reviewer asks for a revision.

▶ Run the analysis above to auto-fill all five templates with your own numbers.

When reporting one way ANOVA results, four things must appear together: the F ratio with both degrees of freedom, the exact p value, an effect size, and the descriptive statistics for every group. A sentence missing any one of those cannot be interpreted or reproduced by a reader.

5.1 The Methods sentence

Say which test you ran, why, at what alpha, in which software, and what you planned to do if an assumption failed. Write it in the past tense and in the passive or first person plural, whichever your field prefers. State the post hoc method here, not in the Results, because choosing it after seeing the p values is a form of data dredging.

5.2 The Results sentence

5.3 Table and figure conventions

Give the ANOVA table its own numbered table with the four standard columns: source, SS, df, MS, F and p. Put descriptive statistics in a separate table, one row per group, with n, mean, SD and the confidence interval. Never repeat in the text every number that already appears in a table; quote only the F statistic and the effect size.

For figures, a bar chart of means with 95% confidence interval whiskers is the standard, and the caption must say what the whiskers represent. An unlabelled error bar is grounds for revision at most journals, because standard deviation, standard error and confidence interval whiskers look identical and mean quite different things.

5.4 The Discussion sentence

The Discussion turns statistics into meaning. Name the direction of the difference, its size in the units your readers care about, and the practical consequence. Then name the limits: the design, the sample, the assumptions that were shaky, and the comparisons that were exploratory rather than planned. Resist the urge to explain a non-significant result as though it were significant.

5.5 Reporting checklist

• F ratio with both degrees of freedom, written as F(df1, df2)
• Exact p value, or p < .001 when smaller
• At least one effect size, with the benchmark you are reading it against
• n, mean and SD for every group
• The assumption checks with their own statistics, not just a claim
• The post hoc method, named and justified, with confidence intervals on each comparison
• The alpha level, and whether it was corrected across multiple outcomes
• The software and version used

Seven formulas do all the work in a one way ANOVA. They run from the raw sums of squares up to the effect size and the confidence interval, and every symbol in every equation is defined underneath it.

1. Enter your data as comma separated numbers. The default tab takes values exactly as the placeholder shows: 52, 48, 55, 61, 47, .... Newlines, tabs, semicolons and spaces all work too, so a column pasted straight from Excel is fine.
2. Name each group. The group name field above every box is editable. Type something meaningful such as Flipped classroom and it flows into the results table, all four charts, the post hoc comparisons and the exported report.
3. Add or remove groups. Press Add cluster for a fourth or fifth condition. Each box has its own Clear and Remove button, and Remove stops at two groups because a one way ANOVA cannot run on fewer.
4. Or build the table by hand. The Manual Entry Table tab gives you one editable column per group. Rename any column heading, press Add column or Add row to grow it, use the × buttons to remove one, then press Use this table.
5. Or upload a file. On the Upload tab choose a CSV or Excel file, then click the columns that should each become a cluster. Every column you select is loaded as a separate group. Non-numeric columns are greyed out.
6. Try a sample dataset. Ten built-in datasets cover the cases that matter: a textbook classic, a clear effect, a borderline p value, a null result, small samples, badly skewed data and a set with unequal variances where Welch is needed.
7. Set your alpha. 0.05 is the default. Choose it before you look at the data, not after.
8. Choose a post hoc method. Leave it on Automatic and the tool picks Tukey HSD when variances are equal and Games-Howell when they are not, using Levene's test to decide.
9. Press Run One Way ANOVA. Nothing is computed until you do, and changing any input clears the results so you never read stale numbers.
10. Read the assumption panel before the p value. If Levene fails, the Welch row in the robust alternatives table is your real headline. If Shapiro-Wilk fails, look at the Q-Q plot in Chart 4 and check the Kruskal-Wallis row.
11. Copy the reporting sentence and export. Section 5 fills five reporting styles with your numbers. The buttons in Section 3 export a text report, a CSV or a print ready PDF, and each file ends with the StatsUnlock credit line.

Excel can run a one way ANOVA two ways: through the Analysis ToolPak, which gives you the whole table in one click, or with worksheet functions, which show you every step. This section does both on the same worked dataset used in Sections 9, 10 and 12, so the three sets of numbers can be checked against each other. The one function that makes the manual route easy is DEVSQ, which returns a sum of squared deviations from the mean directly.

8.1 The functions you need

8.2 Step by step

8.3 The awkward cases Excel makes hard

No effect size at all. Excel has no eta squared or omega squared function. Build them from the sums of squares the ToolPak already prints: =SSB/SST for eta squared, and =(SSB-2*MSW)/(SST+MSW) for omega squared. On this dataset that is 0.6448 and 0.6008.

No post hoc tests. There is no Tukey HSD in Excel. You can approximate it by computing the studentized range statistic yourself, =ABS(mean1-mean2)/SQRT(MSW/2*(1/n1+1/n2)), and comparing it with a critical q value from Table 11.4 in Section 11. Excel cannot give you the exact Tukey p value, because it has no studentized range distribution function.

No Welch's ANOVA. The ToolPak only does the classic equal variance test. If Levene's test fails, Excel cannot help and you should move to R, Python or this calculator.

No Levene's test. Build it by hand: compute the absolute deviation of every value from its own group median with =ABS(A2-MEDIAN($A$2:$A$9)), fill that down for all three columns, then run a second Anova: Single Factor on those deviations. The F it prints is Levene's statistic.

Long format data. If your data have one score column and one group column, the ToolPak cannot read them. Either pivot to wide with a PivotTable, or use =IF($B2="Flipped",$A2,"") style helper columns to split them out.

8.4 Why your Excel answer might differ from R or Python

8.5 Charting the result in Excel

1. Put the group names in one column and their means in the next, with the standard errors in a third.
2. Insert → Charts → Clustered Column, selecting only the names and the means.
3. Click the chart, press the + button, tick Error Bars, then More Options → Custom → Specify Value and point both boxes at your standard error column.
4. Say in the caption whether the bars are standard deviations, standard errors or 95% confidence intervals. They look identical and mean different things, and an unlabelled error bar is grounds for revision at most journals.
5. For a figure that shows the raw data too, add a scatter series of the individual points over the columns using a combo chart.

Chart 1 in Section 3 of this page draws the group means with proper 95% confidence intervals automatically, and Chart 2 shows every raw observation, which is usually faster than rebuilding both in Excel.

8.6 Excel errors you will probably hit

R gives you a one way ANOVA in one line, summary(aov(score ~ group, data = dat)), but that line alone is not what you should report. It has no effect size, no assumption checks and no post hoc comparisons. The script below is complete and runnable, uses base R only so there is nothing to install, prints a plain English verdict, and saves one publication ready figure at 300 dpi.

All three ship with R itself. aov, TukeyHSD, shapiro.test, bartlett.test, oneway.test and kruskal.test all live in stats, which is exactly what reviewers expect to see cited.

9.1 The complete script

# ------------------------------------------------------------------
# One-way ANOVA in R, base R only, nothing to install.
# Change ONLY the data block below to use your own numbers.
# Saves one 300 dpi figure: anova_plot.png
# ------------------------------------------------------------------

set.seed(42)   # keeps the jitter in the figure reproducible

# ---- 1. YOUR DATA -------------------------------------------------
traditional <- c(52, 48, 55, 61, 47, 50, 58, 53)
flipped     <- c(63, 67, 59, 71, 65, 62, 69, 64)
online      <- c(55, 58, 52, 60, 57, 54, 61, 56)

score <- c(traditional, flipped, online)
group <- factor(rep(c("Traditional", "Flipped", "Online"),
                    times = c(length(traditional), length(flipped), length(online))),
                levels = c("Traditional", "Flipped", "Online"))
dat <- data.frame(score, group)

# From a CSV instead? Use these two lines and delete the block above.
# dat <- read.csv("my_data.csv")                    # columns: score, group
# dat <- dat[complete.cases(dat$score, dat$group), ]   # drop missing values
# dat$group <- factor(dat$group)

# ---- 2. DESCRIPTIVES ----------------------------------------------
desc <- aggregate(score ~ group, data = dat,
                  FUN = function(x) c(n = length(x), mean = mean(x), sd = sd(x)))
desc <- do.call(data.frame, desc)
names(desc) <- c("group", "n", "mean", "sd")
print(desc, row.names = FALSE)

# ---- 3. THE ONE-WAY ANOVA -----------------------------------------
fit <- aov(score ~ group, data = dat)
tab <- summary(fit)[[1]]
Fval <- tab[1, "F value"]
pval <- tab[1, "Pr(>F)"]
df1  <- tab[1, "Df"]
df2  <- tab[2, "Df"]
SSb  <- tab[1, "Sum Sq"]
SSw  <- tab[2, "Sum Sq"]
MSw  <- tab[2, "Mean Sq"]
print(tab)

# ---- 4. EFFECT SIZES ----------------------------------------------
eta2   <- SSb / (SSb + SSw)
omega2 <- (SSb - df1 * MSw) / (SSb + SSw + MSw)
cohenf <- sqrt(eta2 / (1 - eta2))

# ---- 5. ASSUMPTION CHECKS -----------------------------------------
sw  <- shapiro.test(residuals(fit))                 # normality of residuals
bar <- bartlett.test(score ~ group, data = dat)     # equal variances
# Levene (Brown-Forsythe, median centred) without loading the car package:
absdev <- abs(dat$score - ave(dat$score, dat$group, FUN = median))
lev <- summary(aov(absdev ~ dat$group))[[1]]

# ---- 6. ROBUST ALTERNATIVES ---------------------------------------
welch <- oneway.test(score ~ group, data = dat, var.equal = FALSE)
kw    <- kruskal.test(score ~ group, data = dat)

# ---- 7. POST HOC --------------------------------------------------
tk <- TukeyHSD(fit, conf.level = 0.95)
print(tk)

# ---- 8. PLAIN-ENGLISH VERDICT -------------------------------------
size <- if (eta2 >= 0.14) "large" else if (eta2 >= 0.06) "medium" else "small"
cat("\n================ ONE-WAY ANOVA VERDICT ================\n")
cat(sprintf("F(%d, %d) = %.4f, p = %.6f\n", df1, df2, Fval, pval))
cat(sprintf("Eta squared = %.4f (%s), omega squared = %.4f, Cohen's f = %.4f\n",
            eta2, size, omega2, cohenf))
cat(sprintf("Shapiro-Wilk on residuals: W = %.4f, p = %.4f -> %s\n",
            sw$statistic, sw$p.value,
            ifelse(sw$p.value > 0.05, "normality OK", "NORMALITY FAILED")))
cat(sprintf("Levene (median centred): F = %.4f, p = %.4f -> %s\n",
            lev[1, "F value"], lev[1, "Pr(>F)"],
            ifelse(lev[1, "Pr(>F)"] > 0.05, "equal variances OK", "VARIANCES DIFFER")))
cat(sprintf("Welch's ANOVA: F(%.0f, %.2f) = %.4f, p = %.6f\n",
            welch$parameter[1], welch$parameter[2], welch$statistic, welch$p.value))
cat(sprintf("Kruskal-Wallis: H = %.4f, df = %d, p = %.6f\n",
            kw$statistic, kw$parameter, kw$p.value))
if (pval < 0.05) {
  cat("=> At least one group mean differs. Read the Tukey table above.\n")
} else {
  cat("=> No detectable difference between the group means.\n")
}
cat("=======================================================\n")

# ---- 9. ONE PUBLICATION-READY FIGURE, 300 dpi ---------------------
png("anova_plot.png", width = 2400, height = 1000, res = 300)
par(mfrow = c(1, 2), mar = c(4.2, 4.2, 2.6, 1))
cols <- c("#2563eb", "#16a34a", "#ea580c")
boxplot(score ~ group, data = dat, col = adjustcolor(cols, 0.35),
        border = cols, ylab = "Score", xlab = "", main = "Scores by group")
stripchart(score ~ group, data = dat, vertical = TRUE, method = "jitter",
           jitter = 0.12, pch = 19, col = cols, add = TRUE)
points(1:3, desc$mean, pch = 18, cex = 1.6, col = "black")
qqnorm(residuals(fit), pch = 19, col = "#9333ea", main = "Q-Q plot of residuals")
qqline(residuals(fit), col = "#ea580c", lwd = 2, lty = 2)
dev.off()
cat("Figure saved as anova_plot.png (300 dpi)\n")

9.2 What each part does

9.3 What the figure shows

The left panel is a box plot of the three groups with every raw observation jittered over it and each group mean marked as a black diamond. It answers two questions at once: are the groups separated, and is any group being driven by one extreme value. On the worked data the Flipped group sits clearly above the other two, and the Traditional and Online boxes overlap heavily, which is exactly the pattern the Tukey table then confirms.

The right panel is a normal Q-Q plot of the residuals. Points hugging the dashed line mean the normality assumption is safe. A pronounced S shape means skew, and points peeling away at both ends mean heavy tails, which usually comes from outliers rather than from the distribution itself.

9.4 Common R problems and fixes

9.5 Useful one-liners

# 1. The whole test in one line
summary(aov(score ~ group, data = dat))

# 2. Welch's ANOVA, safest when variances differ
oneway.test(score ~ group, data = dat, var.equal = FALSE)

# 3. Eta squared straight from the table
with(summary(aov(score ~ group, dat))[[1]], `Sum Sq`[1] / sum(`Sum Sq`))

# 4. Group means, SDs and counts in one call
aggregate(score ~ group, dat, function(x) c(n = length(x), m = mean(x), s = sd(x)))

# 5. Tukey HSD as a tidy data frame
as.data.frame(TukeyHSD(aov(score ~ group, dat))$group)

# 6. Non-parametric alternative plus its post hoc
kruskal.test(score ~ group, data = dat)
pairwise.wilcox.test(dat$score, dat$group, p.adjust.method = "holm")

# 7. Bootstrap confidence interval for eta squared, 2000 resamples
set.seed(1); replicate(2000, {i <- sample(nrow(dat), replace = TRUE)
  t <- summary(aov(score ~ group, dat[i, ]))[[1]]; t$`Sum Sq`[1] / sum(t$`Sum Sq`)}) |> quantile(c(.025, .975))

# 8. Write every result to CSV
write.csv(as.data.frame(summary(aov(score ~ group, dat))[[1]]), "anova_table.csv")

In Python the test itself is scipy.stats.f_oneway(*groups), which returns the F ratio and the p value and nothing else. As in R, that is not enough to report. The script below has the same shape as the R script in Section 9, so a reader who scrolled past it finds everything in the same order: data, descriptives, the ANOVA table, effect sizes, assumption checks, robust alternatives, Tukey HSD, a plain English verdict and one 300 dpi figure.

Install with pip install numpy scipy matplotlib. NumPy holds the arrays and the arithmetic, scipy.stats holds every test used here, and matplotlib draws the figure. pandas is only needed for the commented read_csv alternative. scipy.stats.tukey_hsd requires SciPy 1.11 or newer.

10.1 The complete script

# ------------------------------------------------------------------
# One-way ANOVA in Python. numpy + scipy + matplotlib.
# pip install numpy scipy matplotlib
# Change ONLY the data block below to use your own numbers.
# Saves one 300 dpi figure: anova_plot.png
# ------------------------------------------------------------------

import numpy as np
import scipy.stats as st
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt

rng = np.random.default_rng(42)   # seeded, so the jitter is reproducible

# ---- 1. YOUR DATA -------------------------------------------------
traditional = np.array([52, 48, 55, 61, 47, 50, 58, 53], dtype=float)
flipped     = np.array([63, 67, 59, 71, 65, 62, 69, 64], dtype=float)
online      = np.array([55, 58, 52, 60, 57, 54, 61, 56], dtype=float)

groups = [traditional, flipped, online]
names  = ["Traditional", "Flipped", "Online"]

# From a CSV instead? Use these three lines and delete the block above.
# import pandas as pd
# df = pd.read_csv("my_data.csv").dropna(subset=["score", "group"])
# names = list(df["group"].unique()); groups = [df.loc[df["group"] == g, "score"].to_numpy(float) for g in names]

# ---- 2. DESCRIPTIVES ----------------------------------------------
print(f"{'Group':<14}{'n':>4}{'Mean':>10}{'SD':>10}")
for nm, g in zip(names, groups):
    print(f"{nm:<14}{g.size:>4}{g.mean():>10.4f}{g.std(ddof=1):>10.4f}")

# ---- 3. THE ONE-WAY ANOVA -----------------------------------------
allv = np.concatenate(groups)
N, k = allv.size, len(groups)
grand = allv.mean()
SSb = sum(g.size * (g.mean() - grand) ** 2 for g in groups)
SSw = sum(((g - g.mean()) ** 2).sum() for g in groups)
SSt = SSb + SSw
df1, df2 = k - 1, N - k
MSb, MSw = SSb / df1, SSw / df2
F, p = st.f_oneway(*groups)          # same F and p as the hand calculation

print("\nSource          SS         df        MS          F         p")
print(f"Between  {SSb:10.4f} {df1:9d} {MSb:10.4f} {F:10.4f} {p:10.6f}")
print(f"Within   {SSw:10.4f} {df2:9d} {MSw:10.4f}")
print(f"Total    {SSt:10.4f} {N-1:9d}")

# ---- 4. EFFECT SIZES ----------------------------------------------
eta2   = SSb / SSt
omega2 = (SSb - df1 * MSw) / (SSt + MSw)
cohenf = np.sqrt(eta2 / (1 - eta2))

# ---- 5. ASSUMPTION CHECKS -----------------------------------------
resid = np.concatenate([g - g.mean() for g in groups])
sw = st.shapiro(resid)                                  # normality of residuals
lev = st.levene(*groups, center="median")               # Brown-Forsythe
bart = st.bartlett(*groups)                             # assumes normality

# ---- 6. ROBUST ALTERNATIVES ---------------------------------------
w = np.array([g.size / g.var(ddof=1) for g in groups])
mw = (w * np.array([g.mean() for g in groups])).sum() / w.sum()
lam = sum((1 - wi / w.sum()) ** 2 / (g.size - 1) for wi, g in zip(w, groups))
A = (w * (np.array([g.mean() for g in groups]) - mw) ** 2).sum() / df1
Fw = A / (1 + 2 * (k - 2) * lam / (k * k - 1))
dfw = (k * k - 1) / (3 * lam)
pw = st.f.sf(Fw, df1, dfw)
H, pkw = st.kruskal(*groups)

# ---- 7. POST HOC, TUKEY HSD ---------------------------------------
tk = st.tukey_hsd(*groups)
print("\nTukey HSD")
for i in range(k):
    for j in range(i + 1, k):
        lo, hi = tk.confidence_interval().low[i, j], tk.confidence_interval().high[i, j]
        print(f"{names[i]:<12} vs {names[j]:<12} diff={tk.statistic[i, j]:8.4f} "
              f"p={tk.pvalue[i, j]:9.6f} 95% CI [{lo:.4f}, {hi:.4f}]")

# ---- 8. PLAIN-ENGLISH VERDICT -------------------------------------
size = "large" if eta2 >= 0.14 else "medium" if eta2 >= 0.06 else "small"
print("\n================ ONE-WAY ANOVA VERDICT ================")
print(f"F({df1}, {df2}) = {F:.4f}, p = {p:.6f}")
print(f"Eta squared = {eta2:.4f} ({size}), omega squared = {omega2:.4f}, Cohen's f = {cohenf:.4f}")
print(f"Shapiro-Wilk on residuals: W = {sw.statistic:.4f}, p = {sw.pvalue:.4f} -> "
      f"{'normality OK' if sw.pvalue > 0.05 else 'NORMALITY FAILED'}")
print(f"Levene (median centred): W = {lev.statistic:.4f}, p = {lev.pvalue:.4f} -> "
      f"{'equal variances OK' if lev.pvalue > 0.05 else 'VARIANCES DIFFER'}")
print(f"Bartlett: chi2 = {bart.statistic:.4f}, p = {bart.pvalue:.4f}")
print(f"Welch's ANOVA: F({df1}, {dfw:.2f}) = {Fw:.4f}, p = {pw:.6f}")
print(f"Kruskal-Wallis: H = {H:.4f}, df = {k-1}, p = {pkw:.6f}")
print("=> At least one group mean differs." if p < 0.05 else "=> No detectable difference.")
print("=======================================================")

# ---- 9. ONE PUBLICATION-READY FIGURE, 300 dpi ---------------------
fig, ax = plt.subplots(1, 2, figsize=(11, 4.2))
cols = ["#2563eb", "#16a34a", "#ea580c"]
bp = ax[0].boxplot(groups, patch_artist=True, widths=0.55)
ax[0].set_xticks(range(1, k + 1)); ax[0].set_xticklabels(names)
for patch, c in zip(bp["boxes"], cols):
    patch.set_facecolor(c); patch.set_alpha(0.30); patch.set_edgecolor(c)
for i, (g, c) in enumerate(zip(groups, cols), start=1):
    x = rng.normal(i, 0.045, g.size)
    ax[0].plot(x, g, "o", color=c, alpha=0.85, markersize=5)
    ax[0].plot(i, g.mean(), "D", color="black", markersize=7)
ax[0].set_ylabel("Score"); ax[0].set_title("Scores by group")
st.probplot(resid, dist="norm", plot=ax[1])
ax[1].set_title("Q-Q plot of residuals")
ax[1].get_lines()[0].set_color("#9333ea"); ax[1].get_lines()[0].set_markersize(5)
ax[1].get_lines()[1].set_color("#ea580c"); ax[1].get_lines()[1].set_linestyle("--")
fig.suptitle(f"One-way ANOVA: F({df1}, {df2}) = {F:.2f}, p = {p:.4g}, eta squared = {eta2:.3f}")
fig.tight_layout()
fig.savefig("anova_plot.png", dpi=300, bbox_inches="tight")
print("Figure saved as anova_plot.png (300 dpi)")

10.2 What each part does

10.3 What the figure shows

The left panel is a box plot with every observation drawn over it and each group mean marked as a black diamond, so a reader sees the spread and the centre together. The right panel is the Q-Q plot of the residuals against a normal reference. The title strip carries F, p and eta squared, which makes the file usable as a figure in a slide deck without any further editing.

10.4 Common Python problems and fixes

10.5 Useful one-liners

# 1. The whole test in one line
F, p = st.f_oneway(traditional, flipped, online)

# 2. Eta squared in one expression
eta2 = SSb / (SSb + SSw)

# 3. Group summary straight out of pandas
df.groupby("group")["score"].agg(["count", "mean", "std"])

# 4. Same ANOVA through statsmodels, with a proper table
import statsmodels.api as sm; from statsmodels.formula.api import ols
sm.stats.anova_lm(ols("score ~ C(group)", data=df).fit(), typ=2)

# 5. Tukey HSD through statsmodels, works on older SciPy
from statsmodels.stats.multicomp import pairwise_tukeyhsd
print(pairwise_tukeyhsd(df["score"], df["group"], alpha=0.05))

# 6. Non-parametric alternative
st.kruskal(traditional, flipped, online)

# 7. Permutation test when you trust no distribution at all
st.permutation_test((traditional, flipped, online), lambda *a: st.f_oneway(*a).statistic,
                    permutation_type="independent", alternative="greater", n_resamples=9999, random_state=42)

# 8. Required sample size per group for a given effect
from statsmodels.stats.power import FTestAnovaPower
FTestAnovaPower().solve_power(effect_size=0.25, k_groups=3, alpha=0.05, power=0.8)

To run any of this from a terminal, save it as anova.py and type python anova.py. In Jupyter, remove the matplotlib.use("Agg") line and add plt.show() at the end so the figure appears inline. Sections 8, 9 and 10 all use the same worked dataset and agree to four decimal places: F = 19.0607, p = 0.0000191, eta squared = 0.6448.

Eight lookup tables for the one way ANOVA. Every value here was computed from the exact distribution functions, not copied from a textbook scan, so they agree with R, SciPy and Excel to four decimal places.

11.1 Critical F values at alpha = 0.05

What to conclude: if your F is larger than the value in the cell for your degrees of freedom, reject the null hypothesis at alpha = 0.05. With 3 groups and 24 observations, df₁ = 2 and df₂ = 21, so the critical value is 3.4668. Source: the F distribution, matching F.INV.RT in Excel and qf(0.95, df1, df2) in R.

11.2 Critical F values at alpha = 0.01

What to conclude: use this table only if you set alpha to 0.01 before collecting data. Moving the threshold after seeing the p value is not a stricter standard, it is a different experiment. Notice how much larger every value is: the price of a lower false positive rate is a higher false negative rate.

11.3 Effect size interpretation bands

What to conclude: these are conventions, not laws, and they were proposed for psychology. In fields with tightly controlled measurement, such as analytical chemistry, an eta squared of 0.14 would be unremarkable. In epidemiology it would be enormous. Always report the raw difference in real units beside the standardised effect size.

11.4 Critical studentized range q at alpha = 0.05, for Tukey HSD

What to conclude: a pair differs at alpha = 0.05 when its q statistic exceeds the value in the cell for your number of groups k and your error degrees of freedom. With k = 3 and df₂ = 21 the critical q is 3.565. Multiply it by the standard error of the difference to get the honestly significant difference itself.

11.5 Sample size per group for 80% and 90% power at alpha = 0.05

What to conclude: read the column for your design and the row for the smallest effect you would not want to miss. Detecting a medium effect across three groups needs about 53 people per group. If your study has 10 per group, only a very large effect is detectable and a null result tells you almost nothing. Computed with the non-central F distribution, matching G*Power.

11.6 Assumption check thresholds

What to conclude: the two rows that decide which test you report are Levene and Shapiro-Wilk. Everything else is context. Section 4.11 turns these rows into a single decision table.

11.7 Which test to use when a one way ANOVA does not fit

What to conclude: a one way ANOVA is the right answer only when there is one factor, three or more independent groups and one continuous outcome. Any other shape has a better test.

11.8 Turning a p value into a sentence

What to conclude: significance is a yes or no decision against a threshold you set in advance. The size of p is not the size of the effect, which is why the effect size column exists at all. Report both in the same sentence, every time.

Eight worked one way ANOVA examples, spanning the situations you will actually meet: a textbook case, a huge effect, a borderline p value, a null result, a small sample, a significant but trivial effect, an assumption failure and a four group field trial. Every figure and every number here is static, so they are readable without running the calculator.

The design a one way ANOVA assumes. One categorical factor with three or more levels, each unit measured once, units assigned to levels independently, and one continuous outcome measured the same way in every group. Anything else needs a different test, and no amount of care during analysis can repair a design that does not match.

13.1 Twelve steps to data a one way ANOVA can accept

1. Write the hypothesis before collecting anything. The null is that all k population means are equal. State the alternative, the alpha level and the post hoc method now, not after you have seen the numbers.
2. Define the factor and its levels precisely. "Teaching method" is a factor; traditional, flipped and online are its three levels. Levels must be mutually exclusive, so no unit can belong to two.
3. Define the outcome and its unit. It must be continuous or at least interval, measured on the same instrument, in the same units, with the same precision in every group.
4. Run a power calculation for the one way ANOVA before recruiting. Use Table 11.5. Decide the smallest effect worth detecting and recruit for it. A study that cannot detect the effect it is looking for wastes everyone's time.
5. Randomise assignment where you can. Random allocation is what turns a difference into a causal claim. Without it you have an observational comparison, and confounding is your main enemy.
6. Aim for equal group sizes. Balanced designs are the most robust to unequal variances and give the most power for a fixed total N.
7. Measure each unit once. If the same person appears in two groups, or is measured twice, independence is broken and a one way ANOVA is the wrong test.
8. Blind the measurement where possible. The person recording the outcome should not know which group the unit is in.
9. Record the raw value, never a category. Do not bin a continuous outcome into high, medium and low before analysis. Binning throws away information and lowers power.
10. Record missing data explicitly. Leave the cell blank with a reason recorded elsewhere. Never type a zero for a value that was not measured.
11. Keep one row per observation. Two columns: the outcome value and the group label. This long format works in R, Python, SPSS and JASP without reshaping.
12. Log everything as you go. Dates, instrument, operator, calibration, exclusions. Reviewers ask, and a year later you will not remember.

13.2 The data layout this one way ANOVA calculator expects

Two layouts work. On the Paste tab, one box per group, values comma separated. On the Upload tab, one column per group, with the group name in the header row.

That long format is the one to store. To use the Upload tab, pivot it so each group becomes its own column, then click the columns you want loaded as clusters.

13.3 Minimum sample size for a one way ANOVA

Two observations per group is the arithmetic minimum, because a group of one contributes no variance. That is not a useful minimum. Aim for at least 15 per group, which is where the central limit theorem starts to protect you against non-normal residuals, and use Table 11.5 to check whether your target effect is reachable at that size. To detect a medium effect (Cohen's f = 0.25) across three groups at 80% power you need about 53 per group, 159 in total.

13.4 Pre-registration line

Copy this into your pre-registration and edit the bracketed parts: "Group differences in [outcome] will be tested with a one way between-subjects ANOVA at alpha = 0.05. If Levene's test is significant at p < .05, Welch's ANOVA will be reported instead. If the residuals fail Shapiro-Wilk and the design is unbalanced, a Kruskal-Wallis test will be reported. A significant omnibus test will be followed by [Tukey HSD / Games-Howell] pairwise comparisons. The target sample of [n] per group gives 80% power to detect f = [effect]."

13.5 Common mistakes that make data unusable

• Measuring the same people in every condition, then analysing them as independent groups. This is the single most common fatal error.
• Different measurement instruments across groups, which turns an instrument difference into a group difference.
• Collecting until the p value drops below 0.05. Optional stopping inflates the false positive rate far above 5%.
• Dropping a group after seeing the results because it "did not work".
• Typing zeros for missing values, which silently drags a group mean down and shrinks its variance.
• Rounding during data entry, which loses precision that cannot be recovered.
• Mixing units, such as recording some weights in kilograms and some in pounds.
• Losing the group labels when data are pasted between spreadsheets. Always keep the id column.

This free one way ANOVA calculator is designed for anyone comparing the means of three or more independent groups on one continuous outcome: students checking coursework, researchers checking software output, teachers preparing worked examples, and analysts who need an answer before opening R.

14.1 Use it when

• You have one grouping factor with three or more levels.
• Each observation comes from a different, unrelated unit.
• The outcome is continuous: a score, a weight, a time, a concentration, a yield.
• You want the F ratio, the p value, an effect size and post hoc comparisons together.
• You need to check whether your equal variance and normality assumptions hold.
• You need an APA formatted results sentence you can paste into a paper.
• You want to verify a number produced by SPSS, R, Python, JASP, Minitab or Excel.

14.2 Do not use it when

• You have exactly two groups. Use an independent samples t test. ANOVA gives the same answer, since F equals t squared, but the t test reports the direction.
• The same subjects appear in every condition. Use a repeated measures ANOVA or a linear mixed model.
• You have two grouping factors. Use a two way ANOVA so you can see the interaction.
• Your outcome is a count or a proportion. Use Poisson or logistic regression, or a chi-square test.
• Your data are clustered, such as pupils inside schools or leaves on the same tree. Independence fails and F will be too large.
• You want to control for a continuous covariate. Use ANCOVA.
• You have several correlated outcomes. Use MANOVA rather than one ANOVA per outcome.

14.3 Four real situations

14.4 Decision tree

16.1 Assumptions

1. Independence of observations. Requires: every value comes from a different unit, and no unit influences another. Check: by design, not by statistic. If it fails: F is inflated, often badly, and the false positive rate can rise from 5% to well over 30% with moderate clustering. Use a repeated measures ANOVA or a mixed model instead.
2. Normality of residuals. Requires: the residuals, not the raw scores, follow a normal distribution. Check: Shapiro-Wilk on the residuals plus the Q-Q plot in Chart 4. If it fails: with balanced groups of 15 or more the effect on the false positive rate is small. With small unbalanced groups and strong skew, F becomes liberal, meaning too many significant results. Use Kruskal-Wallis or a transformation.
3. Homogeneity of variance. Requires: the population variances are equal across groups. Check: Levene's test, median centred, plus the largest to smallest variance ratio. If it fails: the direction of the bias depends on the design. If the smaller group has the larger variance, F is liberal and your real alpha is above 5%. If the larger group has the larger variance, F is conservative and you lose power. Use Welch's ANOVA.
4. Continuous outcome measured on an interval scale. Requires: the differences between values mean the same thing everywhere on the scale. Check: by inspection of the measurement instrument. If it fails: means become hard to interpret, as with Likert items. Use Kruskal-Wallis on ranks.
5. No extreme outliers. Requires: no single value dominates a group. Check: Chart 2 and Tukey fences at 1.5 times the interquartile range. If it fails: the outlier inflates that group's variance, which inflates MS within and pushes F down, so you lose power. It can also drag the mean and create a false difference.
6. Random sampling and, for causal claims, random assignment. Requires: units representative of the population, allocated to groups by chance. Check: the study protocol. If it fails: the p value is still computable but no longer generalises, and no causal claim is available.

16.2 Limitations

1. It is an omnibus test. A significant F never says which groups differ. Use Tukey HSD or Games-Howell to find out.
2. It only handles one factor. Two grouping variables need a two way ANOVA, which also tests their interaction.
3. It cannot handle a covariate. Use ANCOVA when a continuous nuisance variable needs controlling.
4. It assumes one measurement per unit. Repeated measurements need a repeated measures ANOVA or a mixed model.
5. It is not robust to dependence. Of the three main assumptions, independence is the one that cannot be patched afterwards.
6. A non-significant result is not evidence of equivalence. To claim two groups are the same you need an equivalence test such as TOST, with a pre-specified equivalence margin.
7. Effect size conventions are field specific. Cohen's bands come from psychology and may be far too lenient or far too strict elsewhere.
8. Post hoc power is not informative. Computing power from your observed effect size after the fact just re-expresses the p value. Power calculations belong before data collection.
9. Multiple outcomes reintroduce multiplicity. Ten separate ANOVAs on ten outcomes need their own correction across the family.
10. This calculator runs in your browser and is for checking, teaching and exploration. For formal submission, reproduce the numbers in R, Python, SPSS or SAS and cite that software.

17.1 The one way ANOVA in one paragraph

A one way ANOVA answers a single question: are these three or more group means further apart than chance alone would put them? It answers it by splitting the total variation into the part explained by group membership and the part left over, turning each into a mean square, and dividing one by the other. The result is the F ratio, and the p value is simply how far into the right tail of the F distribution that ratio lands.

17.2 What good practice looks like

Set alpha before you look at the data. Check the assumptions before you read the p value. Report the F ratio with both degrees of freedom, the exact p value, an effect size and the descriptive statistics for every group. Follow a significant omnibus test with a corrected post hoc procedure chosen in advance, and give confidence intervals on every pairwise difference. Say plainly what the result does not show. That sequence, followed every time, is what separates a defensible analysis from a lucky one.

17.3 Where to go next

If your groups are the same people measured repeatedly, move to a repeated measures ANOVA. If a second factor matters, move to a two way ANOVA. If a continuous covariate needs controlling, move to ANCOVA. If the residuals will not behave, move to Kruskal-Wallis. And whichever you choose, reproduce the final numbers in R or Python before you submit, using the scripts in Sections 9 and 10 of this one way ANOVA calculator as a starting point.

Cite the software you actually ran the final analysis in. If you reproduced these numbers in R or Python, cite R or SciPy as well, using the references in Section 22 as a starting point.

Twenty peer-reviewed journal articles behind every method this one way ANOVA calculator uses, in APA 7th edition, each with a working DOI link.

1. Bartlett, M. S. (1937). Properties of sufficiency and statistical tests. Proceedings of the Royal Society of London. Series A, 160(901), 268–282. https://doi.org/10.1098/rspa.1937.0109
2. Blanca, M. J., Alarcón, R., Arnau, J., Bono, R., & Bendayan, R. (2017). Non-normal data: Is ANOVA still a valid option? Psicothema, 29(4), 552–557. https://doi.org/10.7334/psicothema2016.383
3. Box, G. E. P. (1954). Some theorems on quadratic forms applied in the study of analysis of variance problems, I. Effect of inequality of variance in the one-way classification. The Annals of Mathematical Statistics, 25(2), 290–302. https://doi.org/10.1214/aoms/1177728786
4. Brown, M. B., & Forsythe, A. B. (1974). Robust tests for the equality of variances. Journal of the American Statistical Association, 69(346), 364–367. https://doi.org/10.1080/01621459.1974.10482955
5. Cohen, J. (1992). A power primer. Psychological Bulletin, 112(1), 155–159. https://doi.org/10.1037/0033-2909.112.1.155
6. Cumming, G. (2014). The new statistics: Why and how. Psychological Science, 25(1), 7–29. https://doi.org/10.1177/0956797613504966
7. Delacre, M., Leys, C., Mora, Y. L., & Lakens, D. (2019). Taking parametric assumptions seriously: Arguments for the use of Welch's F-test instead of the classical F-test in one-way ANOVA. International Review of Social Psychology, 32(1), 13. https://doi.org/10.5334/irsp.198
8. Faul, F., Erdfelder, E., Lang, A.-G., & Buchner, A. (2007). G*Power 3: A flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behavior Research Methods, 39(2), 175–191. https://doi.org/10.3758/BF03193146
9. Games, P. A., & Howell, J. F. (1976). Pairwise multiple comparison procedures with unequal n's and/or variances: A Monte Carlo study. Journal of Educational Statistics, 1(2), 113–125. https://doi.org/10.3102/10769986001002113
10. Glass, G. V., Peckham, P. D., & Sanders, J. R. (1972). Consequences of failure to meet assumptions underlying the fixed effects analyses of variance and covariance. Review of Educational Research, 42(3), 237–288. https://doi.org/10.3102/00346543042003237
11. Keselman, H. J., Huberty, C. J., Lix, L. M., Olejnik, S., Cribbie, R. A., Donahue, B., Kowalchuk, R. K., Lowman, L. L., Petoskey, M. D., Keselman, J. C., & Levin, J. R. (1998). Statistical practices of educational researchers: An analysis of their ANOVA, MANOVA, and ANCOVA analyses. Review of Educational Research, 68(3), 350–386. https://doi.org/10.3102/00346543068003350
12. Kruskal, W. H., & Wallis, W. A. (1952). Use of ranks in one-criterion variance analysis. Journal of the American Statistical Association, 47(260), 583–621. https://doi.org/10.1080/01621459.1952.10483441
13. Lakens, D. (2013). Calculating and reporting effect sizes to facilitate cumulative science: A practical primer for t-tests and ANOVAs. Frontiers in Psychology, 4, 863. https://doi.org/10.3389/fpsyg.2013.00863
14. Olejnik, S., & Algina, J. (2003). Generalized eta and omega squared statistics: Measures of effect size for some common research designs. Psychological Methods, 8(4), 434–447. https://doi.org/10.1037/1082-989X.8.4.434
15. Royston, P. (1995). Remark AS R94: A remark on Algorithm AS 181: The W-test for normality. Journal of the Royal Statistical Society. Series C (Applied Statistics), 44(4), 547–551. https://doi.org/10.2307/2986146
16. Schmider, E., Ziegler, M., Danay, E., Beyer, L., & Bühner, M. (2010). Is it really robust? Reinvestigating the robustness of ANOVA against violations of the normal distribution assumption. Methodology, 6(4), 147–151. https://doi.org/10.1027/1614-2241/a000016
17. Shapiro, S. S., & Wilk, M. B. (1965). An analysis of variance test for normality (complete samples). Biometrika, 52(3/4), 591–611. https://doi.org/10.2307/2333709
18. Tukey, J. W. (1949). Comparing individual means in the analysis of variance. Biometrics, 5(2), 99–114. https://doi.org/10.2307/3001913
19. Wasserstein, R. L., & Lazar, N. A. (2016). The ASA statement on p-values: Context, process, and purpose. The American Statistician, 70(2), 129–133. https://doi.org/10.1080/00031305.2016.1154108
20. Welch, B. L. (1951). On the comparison of several mean values: An alternative approach. Biometrika, 38(3/4), 330–336. https://doi.org/10.2307/2332579