Still prevalent in 2026: the idea that you must first “prove” the normality of the data, then automatically switch to non-parametric tests as a backup.

A better approach consists of choosing the statistical data analysis method so that it answers the research question, taking into account the available sample size and the sensitivity/robustness of the method to deviations from normality.

This bad practice is misleading for several reasons

  • Normality tests are often misused because they are not used to definitively “validate” that the data are normal, but to detect deviations from a normal distribution. There are nearly 100 normality tests, each capable of investigating different deviations. Consequently, it is not uncommon for them to contradict each other.
  • With large samples (n), they become very sensitive and often reject normality for minimal deviations that parametric methods would tolerate.
  • With small n, these parametric tests sometimes lack the power to detect real effects.
  • Consequently, “pass or fail” a normality test does not provide a reliable binary answer for deciding on the subsequent statistical data analysis.
  • I won’t name any software, but several are to blame for perpetuating this myth.

Why nonparametric tests are not just a backup plan

  • Nonparametric tests are not “weak” versions of parametric tests. They rely on different logic, often based on ranks rather than raw values.
  • 💣 They sometimes modify your initial hypothesis.
  • They are useful when the data is ordinal, highly skewed, or when the assumptions of parametric tests simply don’t hold. But they are not automatically better: they can be less powerful than parametric tests when the latter are appropriate (see alternatives below).

Sound Practical Logic

  • Instead of asking “Are my data normal?”, it is better to ask “Is my statistical methodology robust to deviations from normality, and how much data do I have?”.
  • Parametric tests often remain valid if deviations from normality are moderate, especially with reasonable sample sizes and well-maintained assumptions of independence.
  • In practice, a Q-Q plot and a look at skewness, extreme values, and the scientific context are, in my experience, much more informative than a normality test taken in isolation—a p-value, in fact.

Consider Robust Alternatives

  • Permutation and bootstrap tests
  • Generalized Linear Models
  • Generalized Estimating Equations (GEE)
  • Etc.
  • Consulting a seasoned statistician is invariably an investment.