DISCUSSION

This work provides a practical application to identifying at-risk groups for participating in impactful population health programs. Clusters create a joint profile of characteristics that reveal the relationship to edentulism, death, and other external variables. Clusters 1, 2, and 3 are the most at-risk clusters. Their numbers of baseline unfavorable risk factors are the highest as well as their links to edentulism and other external variables. From all of the clusters, profiles of individuals are identified using the modes of each risk factor. Clusters naturally consider the interactions of the input variables by grouping similar individuals together. The results reflect a combination of characteristics that affect outcomes, not individual characteristics. Traditional logistic models would need to include multilevel interaction terms to accomplish a similar result.

By ordering the clusters by the rate of edentulism, profiles of individuals with the highest edentulism rates can be compared. Clusters with similar profiles yet differing in their edentulism rates are also seen. The profiles in clusters with the highest rates of edentulism are not necessarily smokers, females, or of low income, as found in prior studies of edentulism using regression techniques. In fact, only 1 of the top 3 clusters for edentulism is smokers, and 2 of the 3 are female or of lower income. Perceived need, whether overall health status, MHS, or any limitations, is consistently high for these clusters.

The clusters are created on the basis of only the baseline risk variables and then labeled on the basis of the rate of edentulism, an external variable. One could easily have labeled the clusters by any of the other external variables. If that is done, Clusters 1–3 for edentulism are seen as the top 3 clusters for death and diabetes, among the top 4 for CVD, and among the top 5 for the Kessler-6 score. Thus, the most at-risk clusters are related across the external variables.

Clusters can be grouped together that are similar to one another in the baseline characteristics but could differ in the outcomes. CVD, diabetes, and high Kessler-6 scores are prevalent across age groups and present in the youngest age groups. In the cases described, the blocks contain younger individuals in clusters who are similar to those in other clusters with individuals who are older and have worse outcomes. These differences can be thought of as a way to forecast the future of these younger individuals. Programs could be designed to provide support for them as they age to help prevent edentulism, such as consistent and affordable preventive care. Dental care could be provided through school-based programs to help lower the onset of edentulism in later years. Adult dental programs could be designed at the state level and include preventive and restorative care. The most productive approach would be to include dental care with medical care. Access to and income support for restorative dental services, such as dentures, could improve the long-term health of those who are edentulous.28 For this study, approximately 30 million people in the U.S. (11.9% of the population) are estimated to be completely edentulous by summing the person-weights of those who are edentulous.

It is concerning that younger age groups are edentulous as well as showing a higher prevalence of CVD and high Kessler-6 scores in certain clusters. In particular, those in Cluster 16 could be targeted for integrated behavioral health programs.

Although the study period includes the coronavirus disease 2019 (COVID-19) years, edentulism is a condition that is years in the making and a marker of poor oral health. By utilizing the longer longitudinal data (4 years rather than 2 years for the MEPS design), a better measure of who is edentulous and the other external variables can be defined. The intent is not to explore any year-by-year changes in health but health over the total interval. An advantage of using self-reported general health status and MHS is to remove the need for a diagnosis from a provider and allow the individual to report how they feel. Thus, lack of access to a provider during the pandemic would not affect this measure. In addition, focusing on the self-reported measures alleviates the need to include variables such as insurance (either medical or dental insurance) and dental treatment itself, which was heavily impacted by the pandemic. The intent was to measure the individual’s health need and not the health care received. The use of the Kessler-6 score at Round 8 (2022) is useful for comparing post–COVID-19 mental health from the beginning of the study.

Clusters are a good tool for exploring the underlying relationships among a number of variables. In this study, the clusters are validated with external variables. However, in contrast to regression methods, there are choices in terms of the type of clustering method used and decisions on how to implement them. Because many of the clustering methods rely on random starts, one could get different results each time the procedure is run. To guard against spurious results, clusters are validated with external variables and other sensitivity analyses to ensure that the results, although not identical, arrive at similar conclusions.

Limitations

There are some limitations to this study. The study is not a causal study. The MEPS data only reflect the non-institutional civilian population; those in nursing homes are excluded from the study if and when they enter an institution. Thus, the results in this study are understated as to the prevalence of edentulism among the total older population. The data are limited in that they do not measure all of the baseline variables and self-reported complete edentulism at the beginning of the study. There is no indication of how long an individual has been living with edentulism, when they became completely edentulous, or their functional dentition.