When data is not normally distributed, it means your values don’t follow the classic bell-shaped curve where most observations cluster around the average and taper off symmetrically on both sides. This matters because many common statistical tests assume normality, and violating that assumption can, in some cases, affect your results. The good news: the consequences are often less severe than people think, and there are straightforward ways to handle it.
What a Normal Distribution Actually Looks Like
A normal (or Gaussian) distribution has three defining features: it’s symmetric around the mean, its tails drop off at a predictable rate, and the mean, median, and mode are all the same value. About 68% of data falls within one standard deviation of the mean, 95% within two, and 99.7% within three. When your data deviates from any of these properties, it’s non-normal.
Two measurements help quantify exactly how your data departs from normal. Skewness measures symmetry. A normal distribution has a skewness of zero. Positive skewness means the data has a longer tail stretching to the right (think income data, where most people earn modest amounts but a few earn enormous sums). Negative skewness means the tail stretches left. Kurtosis measures whether your data has heavier or lighter tails than a normal distribution. A normal distribution has a kurtosis of three. Values above three mean more extreme observations than expected; values below three mean fewer.
Why Data Ends Up Non-Normal
Non-normality isn’t a mistake. It’s often a genuine feature of what you’re measuring. Some common causes:
- Natural boundaries: If a variable can’t go below zero (like reaction times, hospital stays, or income), the data often piles up near the low end and stretches out to the right, creating a right-skewed distribution.
- Outliers: A handful of extreme values can distort an otherwise normal-looking dataset. These may reflect data entry errors, equipment malfunctions, or genuinely unusual cases. Importantly, what looks like an outlier might actually be a sign that the underlying data simply isn’t normal to begin with.
- Mixed populations: If your dataset combines two distinct groups (say, men and women measured on a variable where they differ substantially), the combined data can appear bimodal or skewed even if each group is normal on its own.
- Count data: When you’re counting events (number of doctor visits, number of errors), the data follows distributions like the Poisson or negative binomial, which are inherently non-normal, especially when counts are low.
Common Non-Normal Distributions
Non-normal data isn’t random chaos. It usually follows a recognizable pattern. The lognormal distribution appears frequently in medicine, social sciences, and economics. It describes data where the logarithm of the values is normally distributed, producing a right-skewed shape. Biological measurements like blood pressure readings, antibody concentrations, and survival times often follow this pattern.
The exponential distribution is common in health and social sciences, typically modeling the time between events (time until equipment failure, time between hospital admissions). Count data, like the number of seizures a patient has per month, often follows a Poisson distribution. When the data has more variability than the Poisson allows, the negative binomial distribution is a better fit. Recognizing which distribution your data actually follows is more useful than simply labeling it “not normal.”
How to Check for Normality
Visual tools are the fastest first step. A histogram gives you an immediate sense of shape. A Q-Q (quantile-quantile) plot is more precise: it plots your data against what you’d expect from a perfectly normal distribution. If the points fall along a straight diagonal line, your data is approximately normal. Points that curve off at the extremes indicate heavy tails, meaning more extreme values than expected. Points that form an arc or curve instead of a line typically indicate skewed data.
For formal testing, the Shapiro-Wilk test is the most popular and powerful option, especially for smaller samples under 50 observations. It directly tests whether your data could plausibly come from a normal distribution. The Kolmogorov-Smirnov test is generally used for larger samples of 50 or more but is less sensitive. You can also calculate z-scores from your skewness and kurtosis values. For small samples under 50, z-values beyond plus or minus 1.96 suggest non-normality. For medium samples (50 to 300), the threshold is plus or minus 3.29. For samples above 300, skewness values beyond 2 or kurtosis values beyond 4 are the practical cutoffs.
One important caveat: with very large samples, formal normality tests will flag even trivial departures from normality as statistically significant. A histogram or Q-Q plot gives you a better sense of whether the departure actually matters in practice.
Does It Actually Affect Your Analysis?
This is where many people overreact. Common parametric tests like the t-test and linear regression are surprisingly robust to non-normality. A large simulation study testing 100 different combinations of non-normal distributions found that, with a sample size of 1,000, false-positive rates stayed between 4.4% and 5.2%, essentially identical to the expected 5%. Even at a sample size of 100, the range was only 3.7% to 5.8%. The power of these tests to detect real effects was also only weakly affected by non-normality, while sample size and effect size mattered far more.
The main risk comes from small samples (around 10) combined with distributions that produce extreme leverage points, like highly skewed data with influential outliers. In those specific scenarios, false-positive rates climbed to around 8% to 11%, which is notably higher than the 5% target but still not catastrophic. One extensive simulation study concluded that the two-sample t-test is so robust against non-normality that it can be recommended in nearly all practical applications.
The Central Limit Theorem explains much of this resilience. It states that the distribution of sample means approaches normality regardless of the underlying data’s shape, as long as the sample is large enough. The conventional threshold is a sample size of 30, at which point the sampling distribution closely approximates a normal distribution.
What to Do About Non-Normal Data
Transform the Data
The most common approach is applying a mathematical transformation to make the data more normal. For right-skewed data, the natural log transformation is the go-to choice. It compresses the long right tail and spreads out the bunched-up lower values. If your data includes zeros, you can add a small constant before logging. The square root transformation is a milder option, traditionally recommended for count data. The reciprocal transformation (dividing 1 by each value) is the most aggressive, useful for strongly right-skewed data.
These are all special cases of what’s called a power transformation, where you raise your data to some exponent between negative 1 and 2. The Box-Cox procedure automates the search for the best exponent, removing the guesswork. After transforming, you run your analysis on the transformed values, then convert the results back to the original scale for interpretation.
Use Non-Parametric Tests
Non-parametric tests don’t assume normality at all. They work by ranking your data from smallest to largest and analyzing the ranks rather than the raw values. The trade-off is that they’re sometimes slightly less powerful when data actually is normal, but they perform well across a wide range of distributions.
The mapping from parametric to non-parametric is straightforward. Instead of an independent-samples t-test comparing two groups, use the Mann-Whitney U test (also called the Wilcoxon rank sum test). Instead of a one-way ANOVA comparing three or more groups, use the Kruskal-Wallis test. Instead of Pearson correlation for measuring linear relationships, use Spearman correlation, which captures any monotonic relationship between variables.
Use a Model That Fits Your Data
The most sophisticated option is to choose a statistical model that matches your data’s actual distribution rather than forcing it into normality. Generalized linear models let you specify the distribution directly. For count data, you’d use Poisson or negative binomial regression. For binary outcomes (yes/no), logistic regression. For right-skewed continuous data, gamma regression. This approach is increasingly standard in fields like epidemiology and ecology, and modern statistical software makes it accessible.
When Non-Normality Matters Most
Non-normality is most consequential when your sample is small (under 30), your data has extreme outliers, or you’re working with strict significance thresholds. It also matters more for prediction intervals and tolerance intervals than for hypothesis tests, because those depend on the shape of the entire distribution rather than just the mean. If you’re building a model to predict individual outcomes rather than estimate averages, getting the distribution right is more important.
For most routine analyses with reasonable sample sizes, non-normality is a nuance rather than a crisis. Check for it, understand the shape of your data, and choose your approach accordingly. But don’t abandon a well-designed analysis just because a normality test returned a low p-value.