Statistical Methods
Choose a Linkage Rule Before Building a Dendrogram
Define the task
A dendrogram depends on the distance definition, variable scaling and linkage rule. State whether input is an observation matrix or a condensed distance vector. A square matrix of distances may be interpreted as observations by an API expecting a two-dimensional feature matrix, producing an unintended analysis. Retain object labels in order and verify at least one distance pair. Choose the rule to match the purpose of clustering rather than selecting whichever picture appears most convincing.
Carry out the check
Single linkage uses the smallest cross-cluster distance, complete linkage the largest, and average linkage the mean over cross-cluster pairs. These choices can produce different shapes and sensitivity to bridges or outliers. Ward, centroid and median linkage require appropriate Euclidean geometry; supplying arbitrary precomputed dissimilarities does not establish that requirement. Record preprocessing and software versions. Clusters are algorithmic groupings, not validated biological subtypes. Evaluate stability and external relevance separately and avoid naming groups as diagnoses without supporting evidence.
Worked hypothetical example
For hypothetical one-dimensional points 0,2,5, the first merge combines 0 and 2 at distance 2. Their distances to 5 are 5 and 3, so the next merge height is 3 under single linkage, 5 under complete linkage and 4 under average linkage. These differences occur with identical observations. Independently reconstruct this small example before interpreting a larger dendrogram, and state which rule determines its vertical scale.
Method checked 16 November 2026 against SciPy documentation: linkage.