Cluster Analysis
Also known as: clustering, unsupervised classification, data clustering, numerical taxonomy
Cluster analysis is a family of unsupervised multivariate techniques that partition a set of objects or observations into internally homogeneous, mutually distinct groups — clusters — based on measured characteristics, without any prior knowledge of group membership. It is widely used in market segmentation, bioinformatics, psychology, and social science to reveal natural groupings in data.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
+5 more
When to use it
Use cluster analysis when the goal is to discover natural groups in data without a predefined outcome variable — for example to identify market segments, patient subgroups, species types, or respondent profiles. Prerequisites: observations should be exchangeable (no meaningful time order) and variables should be measured consistently. Standardise continuous variables to equalise scale effects. Avoid cluster analysis when groups are already known and the goal is classification (use discriminant analysis) or when the data are too sparse or high-dimensional for distance measures to be meaningful without dimensionality reduction first. Require at least 20 to 30 observations per anticipated cluster for stable results, and always validate the cluster solution on new data or with a split-sample approach before drawing substantive conclusions.
Strengths & limitations
- Entirely data-driven and unsupervised — reveals structure without imposing a theory-driven grouping.
- Applicable to diverse data types (continuous, ordinal, binary, mixed) with appropriate distance measures.
- Hierarchical methods provide a full picture of nested structure through the dendrogram without committing to a fixed number of clusters.
- Scalable: k-means and related algorithms handle large datasets efficiently.
- Complementary to factor analysis and PCA — can cluster on factor scores or reduced dimensions.
- The number of clusters must be chosen by the analyst; no method for determining k is universally agreed upon.
- Results depend heavily on the choice of distance measure, linkage criterion, and starting values for iterative algorithms.
- All methods will find clusters even in random data — validation is essential.
- Sensitive to outliers, which can distort centroids and linkage in non-robust implementations.
- Clusters are hard partitions; observations at cluster boundaries are forced into one group even when membership is ambiguous.
Frequently asked
How do I decide how many clusters to use?
No single rule is definitive. Common approaches include the elbow method (plot within-cluster sum of squares against k and look for a clear bend), the silhouette width (higher average silhouette indicates better-separated clusters), and the Calinski-Harabasz index. For hierarchical solutions, inspect the dendrogram and the agglomeration schedule for large jumps in merge distances. Always verify that the chosen number of clusters makes substantive sense.
What is the difference between hierarchical and k-means clustering?
Hierarchical clustering builds a tree of nested groupings by successively merging (agglomerative) or splitting (divisive) observations; the analyst chooses where to cut the tree. K-means requires specifying the number of clusters k in advance and then iteratively assigns each observation to the nearest centroid. Hierarchical methods are better for exploring the full range of groupings and for small samples; k-means is more efficient for large datasets when k is approximately known.
Should I standardise my variables before clustering?
Yes, if variables are measured on different scales. Without standardisation, variables with larger numeric ranges will dominate Euclidean distances and small-scale variables will contribute little to the grouping. Z-score standardisation (subtract mean, divide by standard deviation) is the default; range standardisation to [0, 1] is an alternative if the original scale is meaningful.
Can cluster analysis handle categorical variables?
Standard Euclidean distance is not appropriate for categorical variables. Use Gower's distance for mixed data types, or Jaccard / simple matching coefficients for binary data. Alternatively, convert categorical variables to dummy codes, but interpret such clusters cautiously. Latent class analysis is a model-based alternative specifically designed for categorical indicators.
How do I validate a cluster solution?
Internal validation compares within-cluster cohesion and between-cluster separation using indices such as silhouette width. External validation checks whether clusters correspond to meaningful external criteria (e.g., known group membership in a hold-out sample). Stability is assessed by repeating the analysis on bootstrap resamples or random splits of the data and measuring how consistently observations are assigned together.
Sources
- Everitt, B. S., Landau, S., Leese, M. & Stahl, D. (2011). Cluster Analysis (5th ed.). Wiley. ISBN: 978-0470749913
- Hair, J. F., Black, W. C., Babin, B. J. & Anderson, R. E. (2019). Multivariate Data Analysis (8th ed.). Cengage Learning. ISBN: 978-1473756540
How to cite this page
ScholarGate. (2026, June 3). Cluster Analysis. ScholarGate. https://scholargate.app/en/statistics/cluster-analysis
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Discriminant AnalysisStatistics↔ compare
- EFAStatistics↔ compare
- Latent Class AnalysisStatistics↔ compare
- Mixture ModelingStatistics↔ compare
- Multidimensional ScalingStatistics↔ compare