Choose Language Hide Translation Bar
Hierarchical clustering

In my previous JMP Blog post, I talked about two algorithms used for clustering: k-means and Normal Mixtures (using Expectation Maximization). Here, I look at a hierarchical clustering example.DrKm.pngWhile k-means generally gives hard cluster assignments – each data point belongs to only one cluster – Normal Mixtures gives soft, or probabilistic, cluster assignments. In Normal Mixtures clustering, each data point – whether it’s a person or a cell culture – belongs to every cluster with a certain probability. Similarly, hierarchical clustering allows us to view points as belonging to multiple clusters. But in hierarchical clustering, the clusters are – perhaps obviously – hierarchical in nature.

When to Use Hierarchical Clustering

Hierarchical clustering is most useful when there’s a hierarchical structure in the data. Hierarchical data has groups that are themselves made up of subgroups that are made of subgroups ... so on and so forth forever. For example, when we look at blood, we can have red blood cells, white blood cells, plasma, and platelets.Chelsea shared a drawing with you 38.png

Within white blood cells, you can have neutrophils, basophils, lymphocytes ... and many others.Chelsea shared a drawing with you 39.png

Within lymphocytes, you can have T-cells, B-cells, and natural killer cells.Chelsea shared a drawing with you 40.pngWithin T-cells, you can have Helper T-cells, or Memory T-cells, Suppressor T-cells, or Killer T-cells. I could go on, but this is a statistics blog, not an immunology blog.Chelsea shared a drawing with you 41.png


While looking at very specific groups of cells is often useful, there are many situations where we need to generalize and look at a higher-level group, like when doctors look at a white blood cell count from a blood test.Chelsea shared a drawing with you 42.png

When grouping customers based on their purchasing habits, you may sometimes have enough funding to create 50 different marketing emails to cater to shoppers’ specific habits. But other times, you may only have a budget to create five. Hierarchical clustering allows you to select any number of clusters, while providing clusters that are as cohesive as possible.

TL;DR: Use hierarchical clustering when you think your data might have a hierarchical structure. It really can be that simple.

Agglomerative Hierarchical Clustering

JMP uses Hierarchical Agglomerative Clustering (HAC for short), which models hierarchical relationships in the data by successively merging similar clusters together until all the data are in one cluster. At first, each data point is its own singleton cluster. But, that’s awfully lonely, so we then merge the two closest clusters together to form a new cluster. Then we merge the next two closest clusters together, and then we do it again. And again. And again. Until all the data are in one cluster. We often refer to HAC as bottom-up clustering because we started at the bottom, and now we’re here, with one cluster containing all our data points. You could say that the whole team is here.Chelsea shared a drawing with you 38 copy 4.png

Distance Metrics

Deciding which clusters are close to each other is not always straightforward; there are many linkage criteria we can use to determine how far apart two clusters are.

Single linkage defines the distance between two clusters (A and B) as the minimum distance between any point in A and any point in B.Chelsea shared a drawing with you 38 copy.pngAverage linkage defines the distance between to clusters (A and B) as the average distance between points in A and points in B.Chelsea shared a drawing with you 41 copy.png

Complete linkage defines the distance between to clusters (A and B) as the maximum distance between any point in A and any point in B.

Chelsea shared a drawing with you 39 copy.pngThe Centroid method defines the distance between to clusters (A and B) as the distance between their respective mean vectors (centroids).Chelsea shared a drawing with you 40 copy.png

Ward’s method is slightly different from the above methods. Instead of using a pure measure of distance, a univariate ANOVA is fit using cluster assignments as groups. Two clusters (A and B) are combined if they result in the smallest increase in error sums of squares when another ANOVA is run with A and B combined into one group. From our ANOVA article, we know that error is the within group variation. So Ward’s method asks which two groups we can combine and still have low within group variation (i.e., which two groups are the most similar).Chelsea shared a drawing with you 38 copy 2.png


How to Read a Dendrogram

Once we’ve determined the hierarchical structure for our data, we can examine the different clusters at any level. We often want to make observations about the structure. The easiest way to do that is by looking at a dendrogram.

Dendrograms show how clusters were combined. In JMP, the Y axis represents data points/clusters, and the X axis represents distance. For example if two clusters join lower on the dendrogram (to the left), then they are more similar than two other clusters that join higher up on the dendrogram (more to the right).Screen Shot 2019-03-19 at 2.11.24 PM.png


Dendrograms that are dense at the bottom (left) usually have clusters that are more separable, whereas dendrograms that are not dense at the bottom (left) tend to have less separable, more spread-out clusters.

Clustering is more of an art than a science, so the interpretation of your clusters likely requires domain expertise to produce something meaningful. Using your dendrogram, you can create any number of clusters from 1 to N (where N is the number of data points that you have). “Cutting” the dendrogram at any level will tell you the membership of each cluster.Chelsea shared a drawing with you 38 copy 3.png



Let’s take a look at a data set that records the average grams of protein, fat, and carbs from 150 people. There are many reasons we might want to cluster these different diets. Perhaps we want to look at whether different diets have different rates of heart disease. Or maybe we want to look at the cholesterol or iron levels of each cluster.

In JMP, performing hierarchical clustering is pretty simple. Just go to Analyze > Clustering > Hierarchical Cluster and select the columns you want to cluster on. If each data point has a name or label, select that and add it as a Label. You can also select your linkage criteria here. Click OK, and JMP will create your clusters and a dendrogram for you.GIF1.gif


Once inside the Hierarchical platform, you can color your dendrogram by clusters, which can help you pick out different patterns in the clusters.GIF2.gif


You can also grab hard cluster assignments for any number of clusters from 1 to N (where N is your sample size). Just go to the red triangle menu and select Number of Clusters and type in your desired number. You can also drag the small black Diamond on the dendrogram to select a number of clusters. Then go back to the red triangle menu and select Save Clusters. When you select a k (number of clusters), JMP will “cut” the dendrogram at the place where there are exactly k clusters.

You can also drag the diamond on the side of the plot to the point on the dendrogram at which you want to cut.GIF3.gif


Saving our clusters as a column allows us to better visualize the clusters, and use the cluster assignment as a group variable in further analyses (like an ANOVA to see whether cluster membership has a significant effect on rates of cardiac disease).

When looking at just two clusters, we can see that there is a relatively high protein, low-carb, low-fat cluster, and a relatively low protein, high fat and carbs cluster.Screen Shot 2019-04-09 at 3.47.59 PM.png


We could also look at the three-cluster solution, which allows us to tease apart the high protein, low-carb, low-fat group into an extreme and more moderate group. The hierarchical structure helps us understand how these groups relate together, while still allowing us to treat them as separate when needed.Screen Shot 2019-04-09 at 3.48.33 PM.png


JMP will recommend a k for you based on the distance graph shown below the dendrogram. The x axis of the graph represents each time two clusters are merged. The y axis represents the distance between the last two clusters that were joined. When the slope dramatically changes, that means that the last two clusters merged were relatively far apart. This suggests that the remaining clusters are relatively separate, and that the current configuration is a good choice for k.

The dendrogram for this data set is most dense at the bottom (left), meaning that there’s likely a few separable clusters, which we can confirm looking at this plots above.

Try It

Hierarchical clustering (and clustering in general) is not a panacea for all your unlabeled data woes. But it is extremely good at modeling the hierarchical structure in data that actually has a hierarchical structure. It gives us information above and beyond non-hierarchical clustering because it tells us how different clusters relate to each other. This information can be invaluable if we need to combine or split up clusters. Hierarchical clustering gives you some flexibility in how you group your data, and in the end, we all need to be a little more flexible. Try it yourself with the data table attached to this post.

Level II

Hi Chelsea,

Thank you for this article. it is very clear and explanatory.

I have a question about clustering and I was wondering if you could help.

I am trying hierarchical (and K-means) clustering with a relatively small set of data (17 rows and +/-90 columns).

When i do the clustering using all columns, some of the rows get excluded from the clustering. So I may have 1-2-3-etc clusters depending on the chosen number of clusters but the number of rows (observations) in those clusters is smaller than 17 (i.e. some rows do not get assigned any cluster at all).

Is there any reason why this should happen?

Any help is welcome. 



Hi Asier,

I can't think of a math/stats reason why this would happen. It might be
that one or more of your rows is missing an entry. I don't know how JMP
exactly employs the Hierarchical Clustering algorithm, but you usually need
complete data (No missing values). I hope that helps!

Level I

Hello Chelsea, I have a query related to HAC. Please find the picture. What is the significance of this graph.lkjk.jpg

Community Manager

@ASHU4JMP, the easiest way to learn parts of a JMP report is use the question mark tool in the Tools menu and click on the part of the report you have a question about.


In your case, clicking on that graph would take you to the Dendrogram Report help page where you'd learn that that is a Distance Graph. 


Distance Graph

The Distance Graph is the plot that appears beneath the dendrogram. This graph has a point for each step where two clusters are joined into a single cluster. The horizontal coordinates represent the numbers of clusters, which decrease from left to right. The vertical coordinate of the point is the distance between the clusters that were joined at the given step.


You can click and drag either diamond-shaped handle in the dendrogram to control the chosen number of clusters. When you click and drag the diamond, a vertical line appears in the plot that moves to correspond to the number of clusters. Often there is a point where the slope of the distance graph levels off. Such a point suggests a natural break and helps you determine the number of clusters.


Level I

Where is the similarity axis for the HCA? How do we turn it on?

I looked through google and through the manual briefly and could find nothing. It appears in previous versions of JMP the similarity was depicted on the X axis. Was this removed?