Statistics Glossary → z-score (standardization)
z-score (standardization)concept
A unit-free score showing how many standard deviations a value lies from the mean: z = (x - mean) / standard deviation.
Subtracting the variable's mean from each value and dividing the result by the standard deviation gives the z-scores. The transformed variable has a mean of 0 and a standard deviation of 1. A z of 2 means the value is two standard deviations above the mean; a z of -0.5 means half a standard deviation below it. Variables measured in different units (kilograms, years, lira) can thus be compared on a common scale.
Standardization does not change the shape of the distribution: a skewed variable is still skewed after conversion to z-scores, so this is not 'normalization'. Turning z-scores into normal probabilities (for example, only 5% of values have |z| > 1.96) is valid only if the variable is roughly normal. In distance-based methods (clustering, k-nearest neighbours, multidimensional scaling) and in penalized regression the scale of the variables directly affects the result, so standardization is often essential. Using n - 1 or n in the denominator makes only small differences; min-max scaling, which squeezes values into the 0-1 range, is a different method.
An everyday example
Ayşe scored 72 in maths (class mean 60, standard deviation 8) and 85 in Turkish (mean 80, standard deviation 10). The raw scores make Turkish look better, but her z-scores are (72 - 60) / 8 = 1.5 in maths and (85 - 80) / 10 = 0.5 in Turkish. Relative to her class, she is much further ahead in maths.
In MerQur
It appears as 'Variable transformation' (z-score and others) in K-Means and Hierarchical Clustering, 'z-standardize' in DBSCAN / HDBSCAN, 'Standardize indicators (z)' in Latent Profile Analysis, 'z-standardization' in Multidimensional Scaling and 'Feature scaling' in k-Nearest Neighbors. In Anomaly Detection, the 'Z threshold' flags observations whose |z| exceeds it. t-SNE z-standardizes its inputs, and the Neural Network z-standardizes the target in regression and back-transforms the predictions.
Analyses where this concept appears
Neural Network (MLP)K-Means ClusteringHierarchical ClusteringDBSCAN / HDBSCANLatent Profile Analysis (LPA)t-SNE (t-Distributed Stochastic Neighbor Embedding)Multidimensional Scaling (MDS)UMAP (Uniform Manifold Approximation and Projection)Regularized Regression (LASSO / Ridge / ElasticNet)Anomaly DetectionLOESS (Local Regression)k-Nearest Neighbors (KNN)Getis-Ord Gi* Hotspot
Related concepts
Median absolute deviation (MAD) and modified z-scoreDistance measure (Euclidean, Manhattan, Gower, great-circle)