Grubbs' Test Calculator

📊 Grubbs' test detects a single outlier in a univariate dataset that is assumed to follow an approximately normal distribution. It tests whether the most extreme data point (the observation furthest from the sample mean) is statistically significantly different from the rest of the data.

📊 The null hypothesis is that no outliers exist; the alternative is that exactly one outlier is present. If the test statistic G exceeds the critical value, the extreme point is declared an outlier. Grubbs' test does not identify which direction the outlier lies in — it simply flags the most extreme point regardless of whether it is the maximum or minimum value.

⚠️ A two-tailed Grubbs' test tests whether the most extreme value in either direction (the maximum or the minimum) is an outlier. This is the default and most common form, because typically you do not know in advance whether the outlier will be unusually high or unusually low. A one-tailed Grubbs' test tests whether only the maximum value is an outlier (upper-tailed) or only the minimum value is an outlier (lower-tailed).

📐 Use the one-tailed version when you have a priori knowledge about the direction of the potential outlier — for example, if you are specifically concerned about contamination inflating a measurement upward, you would use the upper-tailed test. The one-tailed test is more powerful for detecting outliers in the specified direction because it uses α/n rather than α/(2n) in the critical value calculation, but it will completely miss an outlier in the opposite direction.

📊 Grubbs' test derives its critical values from the assumption that the data (excluding the outlier) follows a normal distribution. The critical value depends on the t-distribution, which itself assumes normally distributed data. If the underlying data is heavily skewed or has heavy tails (e.

📊 g., from a log-normal, exponential, or Cauchy distribution), the test may produce too many false positives — flagging legitimate extreme values from the tails of the non-normal distribution as outliers. Conversely, with light-tailed distributions, the test may fail to detect genuine outliers.

📊 Always verify normality before applying Grubbs' test using a histogram, Q-Q plot, or formal normality test. If normality is violated, consider using Dixon's Q-test (for small samples), the median absolute deviation (MAD) method, or a robust outlier detection approach that does not assume normality.

🔍 The mathematical minimum is n = 3, but Grubbs' test is unreliable with very small samples because the critical values become extremely high — with n = 3 at α = 0.05, the critical value is approximately 1.155, meaning an observation must deviate by more than 1.155 standard deviations to be flagged, which is barely distinguishable from normal variation.

🔍 With n = 6, G_crit ≈ 1.887, and with n = 10, G_crit ≈ 2.286. For practical reliability, use at least n ≥ 8 observations. With larger samples (n > 30), the critical value approaches a stable limit near 3.0 for α = 0.05, and the test becomes more powerful at detecting moderate outliers. However, very large samples can also make the test overly sensitive — detecting deviations that are statistically significant but practically unimportant.

⚠️ Yes, Grubbs' test can be applied iteratively: run the test, remove the identified outlier if one is found, then re-run the test on the remaining n − 1 data points, and repeat until no more outliers are detected. However, this procedure has an important limitation: each iteration inflates the overall Type I error rate.

❌ If you run the test k times at α = 0.05, the probability of falsely declaring at least one non-outlier point as an outlier across all iterations can be substantially higher than 0.05. To control this, apply a Bonferroni correction by using α/k for the k-th iteration, or use the Generalized Extreme Studentized Deviate (ESD) test (also called Rosner's test), which is specifically designed to detect multiple outliers while controlling the overall error rate.

💻 Our calculator performs a single Grubbs' test; for iterative testing, you would need to manually remove the outlier and re-run the test.

📊 Both Grubbs' test and Dixon's Q-test detect single outliers in univariate data, but they differ in their approach and applicable sample sizes. Grubbs' test uses the mean and standard deviation of the entire dataset, making it sensitive to the outlier itself (because the outlier inflates both statistics).

🔍 Dixon's Q-test uses only the gap between the suspected outlier and its nearest neighbor, divided by the overall range, making it less influenced by the outlier on the denominator. Dixon's Q-test is preferred for very small samples (n = 3 to 7) because its critical values are well-established for these sizes and the test is more powerful when n is tiny.

📊 Grubbs' test is preferred for larger samples (n ≥ 8) because it makes fuller use of the data and has better statistical power. Both tests assume normality. For datasets with n > 25, Grubbs' test is generally the better choice.

📊 When you remove an outlier identified by Grubbs' test, the sample mean and standard deviation change — typically, the mean shifts away from the outlier and the standard deviation decreases, because the extreme value that was inflating the variance has been removed.

⚠️ This means the remaining data will appear more tightly clustered, which can sometimes reveal a second outlier that was previously masked by the first. However, removing data points should never be done lightly — always investigate the cause of the outlier before deleting it. If the outlier arose from a measurement error, equipment malfunction, or data-entry mistake, removal is justified.

📊 If the outlier represents a genuine rare event, removing it may obscure important information about the population. Document every removal with a justification in your analysis report.

📊 The Grubbs test statistic G = max|xᵢ − x̄| / s is the ratio of the largest absolute deviation from the sample mean to the sample standard deviation. Conceptually, it tells you how many standard deviations the most extreme observation lies from the center of the data — similar to a z-score, but computed from sample statistics rather than known population parameters.

📊 In a normal distribution without outliers, G typically ranges from about 1.5 to 2.5 depending on sample size. Values significantly above the critical value indicate that the most extreme point is implausibly far from the rest of the data, given the assumed normal distribution. The larger G is relative to the critical value, the stronger the evidence that the extreme point is an outlier.

📐 The critical value G_crit is derived from the t-distribution and depends on two parameters: the sample size n and the significance level α. For a two-tailed test, the formula is G_crit = (n − 1) / √n × √(t²_{α/(2n), n−2} / (n − 2 + t²_{α/(2n), n−2})), where t_{α/(2n), n−2} is the (1 − α/(2n)) quantile of the t-distribution with n − 2 degrees of freedom.

🔍 The α/(2n) term reflects a Bonferroni adjustment for testing both tails simultaneously. As n increases, the critical value increases because larger samples are more likely to contain extreme observations by chance alone, so the threshold for declaring an outlier must be higher. As α decreases (e.g., from 0.05 to 0.01), the critical value also increases, making the test more conservative.

📊 Grubbs' test is not recommended for non-normal data because its critical values and p-values are derived under the assumption of normality. If your data is skewed (e.g., income data, reaction times), the test may flag legitimate tail values as outliers, producing false positives. If your data has heavy tails (e.g., financial returns), the test may also over-detect.

📊 For non-normal data, consider these alternatives: (1) Transform the data (e.g., log transformation for right-skewed data) to achieve approximate normality, then apply Grubbs' test on the transformed values. (2) Use the median absolute deviation (MAD) method, which uses the median and MAD instead of the mean and standard deviation and is robust to non-normality. (3) Use a boxplot rule (1.5 × IQR), which does not assume normality. (4) Use robust z-scores based on the median and MAD.

❌ Grubbs' test detects exactly one outlier per application. The Generalized Extreme Studentized Deviate (ESD) test, also known as Rosner's test, is an extension designed to detect up to k outliers simultaneously while controlling the overall Type I error rate.

📊 It works by running Grubbs' test k times iteratively, each time removing the most extreme point, and then comparing each test statistic to a progressively adjusted critical value. The key advantage over simple iterative Grubbs' testing is that the Generalized ESD test uses pre-specified critical values that account for the multiple testing, so the overall error rate remains at α.

🔍 If you suspect multiple outliers, the Generalized ESD test is statistically more rigorous than running Grubbs' test repeatedly with the same α level.

📊 The choice of significance level α balances two risks: α = 0.05 is the most common choice and provides a 5% chance of falsely declaring a legitimate data point as an outlier (Type I error). α = 0.01 is more conservative — only a 1% chance of a false positive, but a higher chance of missing a real outlier (Type II error).

📊 Use α = 0.05 when the cost of missing an outlier is high (e.g., undetected contamination in pharmaceutical manufacturing could endanger patients). Use α = 0.01 when the cost of incorrectly removing a data point is high (e.g., in clinical trials where removing a patient's data could bias the results). Some practitioners recommend α = 0.10 for exploratory analysis to cast a wider net, then investigating flagged points carefully before deciding on removal.

📊 The standard Grubbs' test is designed for univariate data only — a single variable measured on n observations. It cannot be directly applied to paired (X, Y) data or multivariate datasets. For paired data, you can apply Grubbs' test separately to each variable, but this misses outliers that are unusual in the joint distribution (e.g., a point with individually normal X and Y values but an extreme combination).

📊 For multivariate outlier detection, use Mahalanobis distance, which measures how far each observation is from the multivariate mean accounting for correlations between variables. Outliers are identified by comparing Mahalanobis distances to a chi-squared distribution. For residuals from a regression analysis, apply Grubbs' test to the residual values, which are univariate, to detect outlier residuals that may indicate influential data points.

📊 Yes, applying Grubbs' test to regression residuals is a valid way to detect a single outlier residual — an observation whose vertical distance from the regression line is unusually large. This can identify influential data points that may be distorting the regression model. However, be aware that the residuals from OLS regression are not fully independent (they sum to zero and have n − p − 1 degrees of freedom), so the nominal p-value from Grubbs' test is only approximate.

📊 For simple linear regression with one predictor, the effective degrees of freedom are n − 2 rather than n − 1, which slightly affects the critical value. For practical purposes, this approximation is usually adequate, especially with moderate sample sizes (n > 10). You can use our Regression Assumptions Checker to complement Grubbs' test with a full assessment of regression assumptions.

📊 The maximum normed residual test is another name for Grubbs' test. The term "normed" refers to the fact that the absolute deviation is divided by (normalized by) the standard deviation s, producing a dimensionless ratio G.

📊 The term "maximum" indicates that the test focuses on the single largest such ratio — the observation with the greatest standardized distance from the mean. This alternative name is commonly used in engineering standards, particularly in ASTM E178 — the American Society for Testing and Materials standard for dealing with outlying observations. The mathematical formulation and interpretation are identical regardless of which name is used.