Statistical Methods
Interpret a Shapiro–Wilk Check in the Analysis Context
Define the task
Before testing normality, identify which distribution the analysis assumes. A regression model may concern errors conditional on predictors rather than the raw pooled outcome. Mixing distinct groups can make the pooled distribution non-normal even when a model is useful. Record the observation unit, dependence, sample size and handling of missing values. A Shapiro–Wilk result is one diagnostic, not an automatic instruction to select a different statistical test.
Carry out the check
Inspect a quantile plot, outliers, rounding, boundaries and the consequences of departures for the intended estimator. The documented routine requires at least three observations and cautions that for more than 5,000 observations its W statistic is accurate but its p-value may not be. Small samples can miss important departures; large samples can flag departures that have limited practical consequences. A high p-value does not prove normality. Report the procedure and software version and avoid repeatedly transforming data until a diagnostic crosses a desired threshold.
Worked hypothetical example
Suppose a fictional study has two equally sized groups centered at 0 and 10 with the same within-group standard deviation of 1. The equally weighted mixture has population variance 1+25=26, despite each group's variance being 1. Testing the pooled outcomes targets that mixture, not the within-group errors. This arithmetic example illustrates why the diagnostic target matters; it does not supply a Shapiro–Wilk p-value. Review model residuals and the study design before drawing an analysis decision.
Method checked 18 November 2026 against Primary method documentation: shapiro.