Statistics Glossary → Normality assumption
Normality assumptionconcept
The assumption, made by many parametric tests, that the data or the model residuals come from a normal (Gaussian) distribution.
Methods such as the t-tests, ANOVA, linear regression and Pearson correlation compute their p values and confidence intervals under a normal distribution. The assumption usually concerns not the raw data themselves but the quantity that matters for the method: the residuals in regression and ANOVA, the differences in a paired t-test, and each group in an independent t-test. Normality is examined with tests such as Shapiro-Wilk, Kolmogorov-Smirnov (Lilliefors) and Anderson-Darling, and with Q-Q plots, histograms, skewness and kurtosis.
Thanks to the central limit theorem, the distribution of means approaches normality as the sample grows, so in large samples the t-test and ANOVA tolerate moderate departures from normality. On the other hand, normality tests flag even trivial departures in large samples and can miss serious ones in small samples, so their results should be read alongside plots. When the assumption is seriously violated, transformations, robust methods, the bootstrap or nonparametric tests (Mann-Whitney, Kruskal-Wallis) are the usual alternatives.
An everyday example
The weights of loaves from a bakery usually cluster around 500 grams; a few loaves come out slightly light or slightly heavy, and very extreme weights are rare. This bell-shaped picture fits a normal distribution. Household incomes in a city, by contrast, have a long tail to the right: most households earn a middle income while a few earn a great deal. For data like these the normality assumption does not hold.
In MerQur
It is tested with the Shapiro-Wilk, K-S (Lilliefors), Anderson-Darling, D'Agostino K², Jarque-Bera, Cramér-von Mises and Shapiro-Francia tests in Normality Tests, with Shapiro-Wilk in the t-tests and with the 'Assumption checks: Shapiro' options in ANOVA and ANCOVA. Correlation Analysis offers 'Assumption checks', Bland-Altman analysis normality of differences, Likert Scale Analysis item normality, the Nested Mixed Model 'Residual normality and homogeneity of variance', Tobit Regression the conditional moment normality test; Confirmatory Factor Analysis, Path Analysis and Canonical Correlation give Mardia's multivariate normality test, and ARIMA gives 'Residual diagnostics (ACF + normality)'.
Analyses where this concept appears
Descriptive StatisticsNormality TestsOne-Sample t-TestIndependent-Samples t-TestPaired t-TestOne-Way ANOVATwo-Way ANOVAANCOVAPermutation TestMann-Whitney UWilcoxon Signed-RankCorrelation AnalysisBland-Altman Agreement AnalysisCanonical CorrelationNested Mixed Model (Nested LMM, Multi-Level Hierarchical Mixed Model)Likert Scale AnalysisConfirmatory Factor Analysis (CFA)Tobit Regression (Censored Regression)Path AnalysisARIMA ForecastVAR (Vector Autoregression)Moran's I (Spatial Autocorrelation)
Related concepts
Shapiro-Wilk testQ-Q and P-P plotsKolmogorov-Smirnov test and Lilliefors correction