Section 3 of 5
RESULTS
Marjorie A. Rosenberg · about 6 minutes
Initially, the components of the baseline risk of the clusters and their association with edentulism were described. Then the clusters were validated with the external variables to show the viability of the approach.
Overall, 11.9% of the cohort were edentulous, with 50.3% of Cluster 1 being edentulous and Clusters 2–4 having rates above 30%. Figure 1 helps to visualize the profiles of the clusters, with 1 showing the details. The clusters are labeled on the basis of the highest edentulism rate (Cluster 1) to the lowest (Cluster 17). Because all of the baseline variables are defined as binary, only 1 category per variable is needed. The percentage of each cluster is shown for smokers, those with good/fair/poor PHS or MHS, those with any limitation, those with poor/near poor/low income, females, the unemployed, and those aged ≥50 years. If the category is 50% or greater, it is considered the cluster mode. These collective categories are viewed as unfavorable risk factors.
Figure 1: Summary of baseline risk, with clusters sorted from highest to lowest edentulous percentage. G/F/P, good/fair/poor; MHS, mental health status; PHS, perceived health status.
Cluster 1 is formed by those who are smokers; those with good, fair, or poor PHS and MHS; those with limitations; those with poor, near poor, or low income; those unemployed for the first round in the study; and male aged >50 years. Those in the 17th cluster are described as nonsmokers, those with either excellent or very good PHS and MHS, those with no limitations, those with higher income, those employed, and female aged <50 years. Clusters 8, 15, and 17 are the largest clusters, with 13.5%, 13.9%, and 13.4% of the population, respectively. Cluster 16 is the smallest, with 0.7% of the population.
Looking across the clusters in Figure 1, the first 3 clusters are seen to have a greater number of unfavorable risk factors. Interestingly, the cluster with the highest edentulism rate is male dominated, and the second and third highest clusters for edentulism are nonsmoker dominated. Cluster 16 also has a higher number of unfavorable risk factors but has a low edentulism rate.
These differences can be better visualized with a dendrogram output from the hierarchical clustering procedure shown in Figure 2. The height of the dendrogram reflects the distance between the clusters.19 Three blocks of clusters (A, B, and C) are defined as shown on the dendrogram.
Figure 2: Dendrogram showing similarity among the 17 clusters.
This dendrogram can be better examined through a comparison within and across the blocks. Clusters 10 and 14 and Clusters 15 and 17 in Block B are the most similar pairs of clusters, because their paired distances are the smallest among all pairs of clusters. Clusters 2 and 3 in Block A are similar, as are Clusters 6 and 16, which are joined with Cluster 1. Clusters 4, 5, and 7 in Block C are most similar to one another. Block B has the least number of unfavorable risk factors, whereas Block A has the most. These similarities are more easily visualized by sorting the cluster output in the order of the dendrogram, as shown in Figure 3.
Figure 3: Summary of cluster percentages baseline risk, with clusters sorted by dendrogram output. G/F/P, good/fair/poor; MHS, mental health status; PHS, perceived health status.
The differences in external variables shown in Figure 4 are explored for clusters that are similar in their risk profile. 1 shows the details. First, differences between Clusters 1, 2, and 3 and Clusters 6 and 16 in the highest at-risk Block A are identified. Then, the differences between Clusters 9 and 11 and Clusters 4, 5, and 7 in Block C are also examined. Finally, the most similar pairs, 10 and 14, as well as 15 and 17, are in Block B of those least at risk.
Figure 4: Summary of external variables by cluster, sorted by dendrogram output. CVD, cardiovascular disease.
Clusters 1, 2, and 3 have the highest edentulism rates as well as a high prevalence of other unfavorable risk factors, together with the worst outcomes. Overall, the 4-year death rate is 1.8%, yet Clusters 1 and 3 have death rates of 11.0% and 13.6%, respectively, the highest across all the clusters. Clusters 1–3 have the highest cluster percentages (over 28%) of those with diabetes. Percentages of those with CVD are among the highest in Clusters 1–4. Clusters 1–3 represent 8.4% of the population, accounting for 25.2% of the edentulous population and 50.0% of the deaths over the 4-year period measured for those who started in the study. These 3 clusters represent 21.7% of those with diabetes, 16.3% of those with CVD, and 33.7% of those with a high Kessler-6 score. Clusters 6 and 16 have a younger age distribution than Clusters 1, 2, and 3. Clusters 6 and 16’s external variables do not show the same level of poor outcomes as those in Clusters 1–3; however, the percentages of those with CVD are high in both clusters, and the high Kessler-6 score in Cluster 16 is the highest percentage across all the clusters.
Block C Clusters 9, 11, 4, 5, and 7 have fewer unfavorable risk factors than those in Block A. Clusters 9 and 11 are nonsmokers with a high percentage of individuals with risk factors for self-reported general health status and MHS, whereas Clusters 4, 5, and 7 are smokers but are mixed in the prevalence of those with high PHS and MHS. Clusters 5 and 7 are similar in the external variables, but Cluster 4 has a very high CVD cluster percentage of 78.6% as well as an older age distribution. The external variables for Clusters 9 and 11 are more similar to those for Clusters 5 and 7.
Clusters 10 and 14 in Block B mainly differ in the input variables, with Cluster 14 being entirely female, whereas Cluster 10 is entirely male. Their 4-year death rates and high Kessler-6 scores are low, but there is a greater percentage of those with diabetes and a much greater percentage of those with CVD in Cluster 10. The results for the external variables are very similar for Clusters 15 and 17, yet Cluster 15 is male, and Cluster 17 is female. Both of these clusters are skewed toward the younger age groups, in contrast to Clusters 10 and 14 (shown in Figure 5). Thus, although these individuals generally reported their health as excellent or very good, they have a greater prevalence of diagnosed conditions.
Figure 5: Summary of external variables by cluster and age group, sorted by dendrogram output. CVD, cardiovascular disease.
The discussion above indicates that age may be one reason why clusters with similar profiles may have different outcomes. Although the clusters are constructed with just a binary variable, Figure 5 shows detailed output by age category, sorted in the order of the dendrogram, with support in 1. The first panel is the distribution of age overall by cluster, with the columns summing to 1. Clusters 1–3 are skewed toward the higher age groups, whereas Clusters 6 and 16 are skewed toward the youngest age categories. Clusters 10 and 14 have distributions opposite to those of Clusters 15 and 17. Cluster 4 has the highest percentage of the oldest ages relative to Clusters 9, 11, 5, and 7.
The remaining 5 panels detail age by cluster by external variable (edentulism, death, CVD, diabetes, and high Kessler-6 scores). Each of the rows sums to the prevalence of the outcome by age category, the sums of the columns show the exposure by external variable, and the sum of all of the cells is 100%.
Although the presence of edentulism is associated with older ages, 34% of those aged <50 years are edentulous. The prevalence of edentulism at younger ages is noticeable. Similarly, although most deaths occur in the highest age groups, deaths in the 18–29 age group are present for Clusters 6, 17, and 7. CVD, diabetes, and high Kessler-6 scores are also present at many of the younger ages.