Why Dendrograms Are Easy to Misread
A dendrogram compresses a sequence of hierarchical clustering decisions into a picture. Its compact form can invite several errors: interpreting vertical position as a measure of membership, reading the left-to-right leaf order as meaningful, or treating software-assigned cluster numbers as natural group labels. These readings confuse the geometry of the drawing with the partition and merge history represented by the tree.
A reliable reading starts with the branch structure. Leaves represent individual observations. Each internal node represents a merge, and its height records the dissimilarity at that merge under the chosen linkage rule and distance convention. A cut through the tree can define a partition, but the interpretation of a height cut needs particular care when merge heights are not monotone.
What Branch Heights Do—and Do Not—Say
For a given merge, the height is the dissimilarity value assigned to joining the two clusters. Its meaning therefore depends on the distance measure, feature scaling, and linkage rule used to construct the hierarchy. A height of 4 is not intrinsically “large” or “small”; it is interpreted relative to the dissimilarities in that analysis and their units or convention.
Height is not a cluster’s size, its average feature value, its probability of belonging to a group, or a direct measure of how substantively important the group is. Nor does a merge height alone tell us whether an observation is a good or poor member of a cluster. It records a particular joining dissimilarity, not a full description of the resulting group.
When merge heights are monotone, each parent merge occurs at least as high as the merges below it. A horizontal cut at height \(h\) then has a straightforward interpretation: subtrees joined at or below \(h\) are intact clusters, while merges above \(h\) have not yet joined their descendant groups. This works because no parent merge is placed below one of its descendant merges.
The justification is structural. Begin at the leaves and move upward through the tree. Every merge node at or below \(h\) joins its two descendant groups before the cut is reached. A merge above \(h\) joins groups only beyond the cut, so those groups remain separate. Monotonicity ensures that a merge below the cut cannot have a descendant merge above it. Without that condition, this simple “include every merge at or below \(h\)” description can fail.
Worked Example: Reading a Monotone Height Cut
A hierarchical analysis of four hypothetical insurance portfolios produces the following merges: \(A\) and \(B\) merge at height 2; \(C\) and \(D\) merge at height 3; and \(\{A,B\}\) merges with \(\{C,D\}\) at height 8. What partition does a horizontal cut at \(h=4\) produce? What does the height 8 merge say?
Method. Compare each merge height with the cut. The merges at 2 and 3 are below 4, while the merge at 8 is above 4.
Membership. The first merge joins \(A\) with \(B\), so they form one cluster. The second joins \(C\) with \(D\), so they form another. The final merge is above the cut, so these two clusters have not joined at the cut. The partition is therefore \(\{\{A,B\},\{C,D\}\}\), with two clusters.
Interpretation. Height 8 is the dissimilarity assigned to joining the two existing clusters under the selected linkage and distance convention. It does not say that either cluster has eight members or that the portfolios have a substantive value of eight.
Check. The four leaves appear in exactly one of the two groups, and the two groups are disjoint. The common mistake is to interpret a branch’s vertical position as a measurement of cluster size or meaning.
Leaf Order Is Not a Ranking
The left-to-right order of leaves is usually not a substantive ranking. A dendrogram’s branches can often be rotated at an internal node: swapping the two child subtrees changes the displayed order while preserving the same merges and cluster memberships. Thus, if a plot displays leaves \(A,B,C,D\), another valid display may show \(B,A,D,C\) without changing the clustering.
Leaf order can be useful for drawing the tree clearly, and particular software may choose an order using a plotting convention. But unless the analysis explicitly imposes an ordering with a separate meaning, horizontal position is not a distance, a similarity score, a time sequence, or an ordered risk category. The branch connections—not the row of labels—determine which observations belong together.
Worked Example: Separating Leaf Order From Membership
A dendrogram has leaves displayed as \(M,N,P,Q,R\). Its merges are \(M\) with \(N\) at height 1, \(P\) with \(Q\) at height 2, \(\{M,N\}\) with \(R\) at height 5, and \(\{M,N,R\}\) with \(\{P,Q\}\) at height 9. Identify the partition after two merges and decide whether the displayed order ranks observations by similarity.
Step 1: Count the completed merges. There are five leaves initially. Each binary merge reduces the number of current clusters by one, so after two merges there are \(5-2=3\) clusters.
Step 2: Follow those merges. The first merge makes \(\{M,N\}\). The second makes \(\{P,Q\}\). Observation \(R\) has not merged with either group yet. The partition is \(\{\{M,N\},\{P,Q\},\{R\}\}\).
Step 3: Assess the leaf order. The display \(M,N,P,Q,R\) does not rank the observations. For example, the \(P,Q\) subtree could be flipped to display \(Q,P\) without changing any merge. The only membership conclusions come from the connections in the tree.
Check. The partition contains all five observations exactly once and has three clusters, as required after two merges. The common mistake is to infer that adjacent leaves must be more similar than leaves farther apart on the page.
Cluster Labels Are Identifiers, Not Explanations
A clustering program may assign labels such as 1, 2, and 3 to the groups from a cut. These are identifiers, often assigned according to software-specific ordering rules. The label “1” is not inherently the earliest, largest, safest, or most important cluster. Relabeling the groups while keeping their memberships unchanged does not change the result.
A defensible description names the membership or describes a pattern observed in the features after the clusters have been formed. For instance, “cluster 2 contains observations \(P\) and \(Q\)” reports membership. “Cluster 2 consists of high-frequency, low-severity observations” is a substantive interpretation that requires checking the underlying data; the cluster number alone does not justify it.
Worked Example: Interpreting Software Cluster Numbers
A five-observation clustering is cut after two merges. The software reports labels \(R=1,\ S=1,\ P=2,\ Q=2,\ T=3\). The merge history shows \(P\) with \(Q\) first and \(R\) with \(S\) second. State the groups, and evaluate the claim that cluster 1 is “the first and most important group.”
Step 1: Collect observations by label. Label 1 contains \(\{R,S\}\), label 2 contains \(\{P,Q\}\), and label 3 contains \(\{T\}\). Thus the partition is \(\{\{R,S\},\{P,Q\},\{T\}\}\), with three groups.
Step 2: Compare with the merge history. The two completed merges account for \(\{P,Q\}\) and \(\{R,S\}\); \(T\) remains alone. The reported memberships agree with that history.
Step 3: Assess the claim. The membership data do not support calling cluster 1 “most important.” The label is an identifier, not a measure of importance, and it does not by itself establish which group formed first. Importance would require a separate, explicitly defined substantive criterion.
Check. The labels define three disjoint groups covering all five observations. The common mistake is to attach an ordering or value judgment to arbitrary numeric identifiers.
Horizontal Cuts When Heights Are Inverted
Some linkage methods can produce inversions: a later merge has a lower recorded height than an earlier merge. Centroid linkage is a familiar example. The merge sequence still records a valid agglomerative history—each step joins two clusters that exist at that point—but the heights are no longer monotone along every path from leaves to root.
In that situation, do not apply the monotone-height shortcut without checking the tree. A geometric horizontal cut still separates branches where the cut crosses them, but it need not correspond to the claim “include all chronological merges whose heights are at most \(h\).” That claim can be impossible: a later parent merge may have a smaller height than its earlier descendant merges. If software offers a height-cut operation, check its definition and any monotonicity requirements. When the question specifies a fixed number of clusters, following the merge sequence and counting completed merges is often the clearest approach.
Worked Example: Diagnosing an Inverted Dendrogram
A valid four-object agglomerative sequence is recorded as follows: \(A\) and \(B\) merge at height 8; \(C\) and \(D\) merge at height 9; then \(\{A,B\}\) merges with \(\{C,D\}\) at height 7. Explain why this is an inversion, give the partition after two merges, and identify the error in interpreting a cut at \(h=7.5\) by merge heights alone.
Step 1: Check that the sequence is possible. Initially there are four singleton clusters. The first merge leaves three: \(\{A,B\},\{C\},\{D\}\). The second leaves two: \(\{A,B\},\{C,D\}\). The third joins those two, leaving one cluster. Each merge uses clusters that exist at that stage, and there are exactly \(4-1=3\) binary merges.
Step 2: Identify the inversion. The third merge is the root merge, but its height 7 is below both descendant merge heights, 8 and 9. The heights therefore decrease along the path from the earlier merges to the final merge, violating monotonicity.
Step 3: Use the sequence for a fixed merge count. After two merges the partition is \(\{\{A,B\},\{C,D\}\}\), containing two clusters. This conclusion follows from counting the first two chronological merges, not from sorting the heights.
Step 4: Diagnose the threshold trap. The statement “include every merge at height at most 7.5” would include the final merge at 7 but exclude its prerequisite merges at 8 and 9. It cannot be used as a chronological agglomerative history. A geometric horizontal cut at 7.5 can be read from the drawn branches, but it should not be described as the result of simply including all merges with heights at most 7.5. The correct interpretation depends on the actual tree geometry and the software’s cut convention.
Check. The valid sequence has three merges for four objects, and the two-merge partition has two groups. The common mistake is to list a later merge with a cluster that has already been absorbed, or to add an extra merge after all objects have joined.
A Careful Reading Procedure
Check the distance measure, feature scaling, linkage rule, and any special convention for the vertical axis. Interpret heights only in that context.
For a stated number of clusters, count merges in sequence. For a height cut, check whether the heights are monotone and then read the branches at the cut.
Do not infer ranking from leaf position or substantive meaning from numeric cluster labels.
For \(n\) leaves, a binary agglomerative hierarchy has \(n-1\) merges in total. Each merge must combine two clusters that exist immediately before that step.
Common Exam Traps
- Reading height as membership. A high merge indicates a larger recorded merge dissimilarity under the chosen method; it does not itself determine how many observations are in a group.
- Treating leaf adjacency as similarity. Leaves can be reordered by rotating branches. Follow their common branch structure instead.
- Giving cluster numbers a meaning. Report membership first. Interpret group characteristics only by examining the relevant features.
- Applying a height threshold to an inversion as if heights were monotone. A later merge below an earlier merge breaks the simple threshold-as-sequence interpretation. Check the geometry and the stated software convention.
- Writing an impossible merge history. Once an observation belongs to a merged cluster, it cannot be merged separately at a later step. After \(n-1\) binary merges, one cluster remains.
Check Your Understanding
For each question, use the branch structure or merge history rather than the visual order or numeric label alone.
- A monotone dendrogram has merges \(U,V\) at height 1, \(W,X\) at height 4, and \(\{U,V\}\) with \(\{W,X\}\) at height 6. What partition does a cut at height 5 give? (Answer: \(\{\{U,V\},\{W,X\}\}\), two clusters.)
- A six-observation hierarchy has completed three binary merges. How many clusters remain, regardless of the leaf order? (Answer: 3 clusters.)
- A program reports labels \(A=4,B=4,C=2,D=2,E=1\). What are the groups, and does label 4 imply greater importance than label 2? (Answer: \(\{A,B\},\{C,D\},\{E\}\); no, the labels are identifiers.)
- Leaves \(J,K,L,M\) are displayed in that order, but the branches join \(J\) with \(L\) and \(K\) with \(M\) at the first two merges. What are the two groups after those merges, and does displayed adjacency determine them? (Answer: \(\{J,L\}\) and \(\{K,M\}\); no, branch connections determine membership.)
- For four objects, the merge sequence is \(A,B\) at 5; \(C,D\) at 6; then \(\{A,B\},\{C,D\}\) at 3. How many clusters remain after two merges, and what feature makes the heights inverted? (Answer: 2 clusters; the third, parent merge has height 3, below its descendant merge heights 5 and 6.)