Z-Score Applications and Limitations
You'll learn: How z-scores are used in real data science tasks and when they might mislead you.
Now that you understand z-scores thoroughly, let's see where they shine—and where they stumble.
Where Z-Scores Excel
Outlier Detection: If a data point has a z-score beyond ±3, it's unusual enough to investigate. In quality control, a product measurement with z = 4.2 might signal a defective batch.
Hypothesis Testing: Z-scores help us ask "Is this result surprising?" If you observe a sample mean with z = 2.5, you know it's in the top ~1% of possibilities under your null hypothesis—strong evidence something's changed.
Standardization for Comparison: Comparing SAT scores (mean 1050, SD 200) to ACT scores (mean 21, SD 5) directly is meaningless. Converting both to z-scores puts them on the same scale, making comparisons fair.
Machine Learning Feature Scaling: Many algorithms (like k-nearest neighbors or neural networks) perform better when features are standardized to similar ranges.
Critical Limitations
Assumes Normal Distribution: Z-scores interpret best when data follows a normal distribution. For skewed or multi-modal data, a z-score of 2 might not mean what you think.
Sensitive to Mean and SD: A few extreme outliers distort the mean and standard deviation, making z-scores unreliable. In income data with billionaires, most people might have "negative" z-scores!
Not Robust: For non-normal data, percentiles or IQR-based methods often work better than z-scores.
Key Takeaway: Z-scores are powerful for outlier detection, hypothesis testing, and fair comparisons when data is approximately normal, but they can mislead with skewed distributions or when outliers corrupt the mean and standard deviation.