The FMM Procedure

It is important to note that Roeder’s original analysis proceeds in a different manner than the finite mixture modeling presented here. The technique presented by Roeder first develops a “best” range of scale parameters based on a specific criterion. Roeder then uses fixed scale parameters taken from this range to develop optimal equal-scale Gaussian mixture models.

You can reproduce Roeder’s point estimate for the density by specifying a five-component Gaussian mixture. In addition, use the EQUATE=SCALE option in the MODEL statement and a RESTRICT statement fixing the first component’s scale parameter at 0.9025 (Roeder’s h = 0.95, scale). The combination of these options produces a mixture of five Gaussian components, each with variance 0.9025. The following statements conduct this analysis:

title2 "Five Components, Equal Variances = 0.9025";
ods select DensityPlot;
proc fmm data=galaxies;
   model v = / K=5 equate=scale;
   restrict int 0 (scale 1) = 0.9025;
   ods exclude IterHistory OptInfo ComponentInfo;
run;
ods graphics off;

The output is shown in Figure 37.18 and Figure 37.19.

Figure 37.19: Density Plot for Roeder’s Analysis