Assumptions and Limitations
You'll learn when the Z-score method works reliably and when it can mislead you.
The Core Idea
The Z-score method is powerful but not universal. It makes a critical assumption: your data follows a normal (Gaussian) distribution. Why? Because Z-scores measure how many standard deviations a point lies from the mean—and in a normal distribution, we know that ~99.7% of data falls within ±3 standard deviations. This mathematical relationship breaks down when your data is skewed, has multiple peaks, or follows other shapes.
The Normality Requirement
Think of Z-scores as a ruler designed for bell-shaped curves. If you try to use that ruler on a U-shaped distribution or a heavily skewed dataset, your measurements become meaningless. A point with Z = 3 might be genuinely rare in normal data, but perfectly common in exponentially-distributed data.
The Small Sample Trap
Z-scores are especially sensitive to outliers in small datasets. Imagine you have 10 data points, and one is extreme. That single outlier pulls the mean and inflates the standard deviation—ironically making itself look less anomalous by its own Z-score! With larger samples (typically 30+), this effect diminishes, but it remains a real risk in small datasets.
When to Be Cautious
- Non-normal distributions: Z-scores will flag false positives or miss real anomalies
- Small samples (n < 30): Outliers distort the very statistics used to detect them
- Multimodal data: Multiple clusters make mean/std deviation misleading
- Heavy-tailed distributions: Extreme values are naturally more common than the normal model predicts
Key Takeaway: The Z-score method only works reliably when data is approximately normally distributed and sample sizes are adequate—always check these conditions before trusting your anomaly flags.