The CLUSTER procedure hierarchically clusters the observations in a SAS data set by using one of 11 methods. The data can be coordinates or distances. If the data are coordinates, PROC CLUSTER computes (possibly squared) Euclidean distances. If you want non-Euclidean distances, use the DISTANCE procedure (see Chapter 33) to compute an appropriate distance data set that can then be used as input to PROC CLUSTER.

The clustering methods are: average linkage, the centroid method, complete linkage, density linkage (including Wong’s hybrid and *k*th-nearest-neighbor methods), maximum likelihood for mixtures of spherical multivariate normal distributions with equal variances but possibly unequal mixing proportions, the flexible-beta method, McQuitty’s similarity analysis, the median method, single linkage, two-stage density linkage, and Ward’s minimum-variance method. Each method is described in the section Clustering Methods.

All methods are based on the usual agglomerative hierarchical clustering procedure. Each observation begins in a cluster by itself. The two closest clusters are merged to form a new cluster that replaces the two old clusters. Merging of the two closest clusters is repeated until only one cluster is left. The various clustering methods differ in how the distance between two clusters is computed.

The CLUSTER procedure is not practical for very large data sets because the CPU time is roughly proportional to the square or cube of the number of observations. The FASTCLUS procedure (see Chapter 35) requires time proportional to the number of observations and thus can be used with much larger data sets than PROC CLUSTER. If you want to cluster a very large data set hierarchically, use PROC FASTCLUS for a preliminary cluster analysis to produce a large number of clusters. Then use PROC CLUSTER to cluster the preliminary clusters hierarchically. This method is illustrated in Example 30.3.

PROC CLUSTER displays a history of the clustering process, showing statistics useful for estimating the number of clusters in the population from which the data are sampled. It creates a dendrogram when ODS Graphics is enabled. PROC CLUSTER also creates an output data set that can be used by the TREE procedure to output the cluster membership at any desired level. For example, to obtain the six-cluster solution, you could first use PROC CLUSTER with the OUTTREE= option, and then use this output data set as the input data set to the TREE procedure. With PROC TREE, specify the NCLUSTERS=6 and the OUT= options to obtain the six-cluster solution. For an example, see Example 94.1 in Chapter 94, The TREE Procedure.

For coordinate data, Euclidean distances are computed from differences between coordinate values. The use of differences has several important consequences:

For differences to be valid, the variables must have an interval or stronger scale of measurement. Ordinal or ranked data are generally not appropriate for cluster analysis.

For Euclidean distances to be comparable, equal differences should have equal practical importance. You might need to transform the variables linearly or nonlinearly to satisfy this condition. For example, if one variable is measured in dollars and one in euros, you might need to convert to the same currency. Or, if ratios are more meaningful than differences, take logarithms.

Variables with large variances tend to have more effect on the resulting clusters than variables with small variances. If you consider all variables to be equally important, you can use the STD option in PROC CLUSTER to standardize the variables to mean 0 and standard deviation 1. However, standardization is not always appropriate. See Milligan and Cooper (1987) for a Monte Carlo study on various methods of variable standardization. You should remove outliers before using PROC CLUSTER with the STD option unless you specify the TRIM= option. The STDIZE procedure (see Chapter 84) provides additional methods for standardizing variables and imputing missing values.

The ACECLUS procedure (see Chapter 23) is useful for linear transformations of the variables if any of the following conditions hold:

You have no idea how the variables should be scaled.

You want to detect natural clusters regardless of whether some variables have more influence than others.

You want to use a clustering method designed for finding compact clusters, but you want to be able to detect elongated clusters.

Agglomerative hierarchical clustering is discussed in all standard references on cluster analysis, such as Anderberg (1973), Sneath and Sokal (1973), Hartigan (1975), Everitt (1980), and Spath (1980). An especially good introduction is given by Massart and Kaufman (1983). Anyone considering doing a hierarchical cluster analysis should study the Monte Carlo results of Milligan (1980), Milligan and Cooper (1985), and Cooper and Milligan (1988).

Other essential, though more advanced, references on hierarchical clustering include Hartigan (1977, pp. 60–68; 1981), Wong (1982), Wong and Schaack (1982), and Wong and Lane (1983). See Blashfield and Aldenderfer (1978) for a discussion of the confusing terminology in hierarchical cluster analysis.