Social determinants and machine learning prediction of complete tooth loss in older adults across United States counties

INTRODUCTION

Complete tooth loss, the loss of all natural teeth, is the end stage of untreated caries and periodontal disease and remains common among older adults despite decades of improvement in high-income countries. Worldwide, edentulism affects a substantial share of adults aged 60 years and over, with prevalence that varies more than fourfold across regions and that persists as a marker of accumulated oral disadvantage.(1) Estimates from the Global Burden of Disease study document a large and only slowly declining global burden, concentrated among women and older age groups, and national analyses in China and elsewhere confirm that edentulism remains a durable public health problem even as overall oral health improves.(2,3) In the United States, the prevalence of edentulism declined steadily through the twentieth century, but the decline has been uneven, leaving large geographic and social gaps between the groups that retained their teeth and those that did not.(4-6) Complete tooth loss carries functional and systemic consequences, including impaired chewing, dietary restriction, and associations with cognitive decline, frailty, and cardiovascular and all-cause mortality, so its distribution is a tractable target for population surveillance.(7-10) With the rapid aging of populations worldwide, the absolute number of older adults who retain no teeth will keep rising for some time even where prevalence is falling, which keeps the geographic monitoring of edentulism relevant to planning for prosthodontic and supportive care.

The social gradient of tooth loss is among the strongest and most replicated findings in oral epidemiology. Individual-level studies consistently find that income, education, race and ethnicity, and rural residence predict both the number of teeth retained and the odds of becoming edentulous.(11,12) Because complete tooth loss is the accumulated endpoint of a lifetime of disease and treatment decisions, it is a particularly informative outcome for studying how social position shapes oral health, in contrast to point-in-time measures such as untreated caries. These associations are not only compositional; area-level deprivation predicts edentulism above and beyond individual characteristics, as shown in county and census-tract analyses in the United States.(13,14) Comparable geographic gradients in socioeconomic inequality and oral outcomes have been reported across European, Latin American, and Asian settings, indicating that the relationship between place-based disadvantage and tooth loss is not specific to a single health system.(15-17) Decomposition studies further suggest that socioeconomic position, rather than behavioral differences alone, accounts for a large share of the racial and ethnic inequities in oral health.(18) Because public health and social services are planned and funded at the county and subcounty level, understanding the county-level correlates of edentulism is directly actionable for targeting programs where the need is greatest.

Recent advances in small-area estimation and machine learning now make it possible to model the geographic distribution of a condition directly and to interrogate which determinants matter most for prediction. The CDC PLACES project publishes model-based county estimates of chronic disease, health behaviors, and health-related social needs that support local planning without new survey collection.(19,20) Machine learning methods, particularly tree ensembles, have been applied to public health and geriatric problems to capture non-linear and interacting relationships, and explainability tools such as permutation importance and partial dependence make these models auditable for decision makers.(21-23) The appeal of these tools for oral health is that they can be fitted to existing routine data and updated as new small-area estimates are released, rather than requiring new and costly surveys. What remains unclear is whether such models trained on small-area data generalize across space, since random cross-validation can inflate apparent performance when units are spatially autocorrelated.( 24-26)

This study addresses that gap for complete tooth loss among older adults. Prior county-level models of edentulism have not, to our knowledge, combined model-agnostic explainability with spatially grouped validation, so the transferability of such models beyond the counties on which they were trained remains unestablished. The objective was to identify the social determinants that best predict county-level prevalence of complete tooth loss among adults aged 65 years and over and to assess the stability of machine learning predictions under spatial cross-validation. Three hypotheses were tested: H1, county poverty and low educational attainment are positively associated with complete tooth loss prevalence independent of age structure and dental care access; H2, gradient boosting predicts complete tooth loss prevalence more accurately than a linear ridge model; and H3, the gradient-boosting advantage and the ranking of social determinants persist under spatial cross-validation.

METHODS

RESULTS

Hypothesis 3: stability under spatial cross-validation

Under spatial block cross-validation, performance declined for both models, but the decline was markedly larger for the linear model. The ridge baseline fell to an R-squared of 0.34, whereas the gradient-boosted model retained an R-squared of 0.53 with a root mean square error of 2.94 percentage points (Table 2; Figure 1). The scatter of out-of-fold predictions against observed values showed reasonable agreement across most of the prevalence range, with the largest scatter at the upper extreme, where few counties lie (Figure 1). The leave-one-state-out sensitivity analysis was consistent, with a gradient-boosting R-squared of 0.68. In permutation importance, self-rated health (0.170), poverty (0.162), and current smoking (0.117) were the three leading predictors, and poverty was the highest-ranked social determinant (Table 3). The partial dependence functions confirmed a monotonic increase in predicted complete tooth loss across the observed range of poverty and low educational attainment (Figure 2). The gradient-boosting advantage and the prominence of socioeconomic deprivation therefore persisted under spatial validation, supporting H3.

Predictor	Importance	SD
GHLTH	0.170	0.006
POV150	0.162	0.006
CSMOKING	0.117	0.004
DENTAL	0.045	0.003
ACCESS2	0.030	0.002
NOHSDP	0.026	0.002
REMNRTY	0.024	0.001
AGE65	0.018	0.001
DIABETES	0.017	0.001
HCOST	0.014	0.001
SNGPNT	0.011	0.001
CROWD	0.011	0.000
BROAD	0.010	0.001
UNEMP	0.010	0.001

Note: Importance is the mean decrease in the coefficient of determination after permuting each predictor across ten repeats. SD = standard deviation. GHLTH = fair or poor self-rated health; POV150 = below 150% of poverty; CSMOKING = current smoking; DENTAL = dental visit in the past year; ACCESS2 = lack of health insurance; NOHSDP = no high school diploma; REMNRTY = racial or ethnic minority; AGE65 = aged 65 years or older. Source: Own elaboration from the analysis of CDC PLACES and ACS data.

DISCUSSION

The main finding is that county socioeconomic deprivation independently predicts complete tooth loss in older adults, and that a gradient-boosted model transfers across space markedly better than a linear model even though the two are indistinguishable under random validation. Poverty and low educational attainment each showed a robust, adjusted association with complete tooth loss, and poverty was the leading social determinant in the explainability analysis. These results bear directly on the growing use of small-area models to target oral health resources, because they show that the apparent accuracy of such models can be misleading unless validation respects geographic structure.

Hypothesis 1 was supported. The positive association between deprivation and edentulism is consistent with a long line of work. In nationally representative United States data, tooth loss tracks income, education, and racial or ethnic status, and social stratification explains a widening of disparities between advantaged and disadvantaged groups over time.(6,12) County and census-tract analyses likewise found that area-level poverty and other measures of disadvantage predict edentulism among seniors.(13,14) The present estimates are of comparable direction and magnitude, with poverty showing the strongest standardized effect, and they extend earlier work by showing that the association survives adjustment for dental care use, age structure, and several chronic-disease covariates. In absolute terms, moving a county from the 25th to the 75th percentile of poverty corresponds to on the order of two additional percentage points of complete tooth loss, a difference that is meaningful when aggregated across the older population of a county. Cross-national studies of socioeconomic inequality in oral health reached similar conclusions, so the gradient is unlikely to be an artifact of the United States health system.(15,16) A study of older adults in Sichuan Province similarly found socioeconomic status to be the dominant correlate of oral disease, and analyses that decompose racial and ethnic oral health gaps into their components point to socioeconomic position as the principal driver.( 17,18)

Material hardship beyond income shows the same pattern. Food insecurity has been associated with dental caries and periodontitis among middle-aged and older adults in the United States and Korea, and the PLACES platform itself publishes county measures of food and housing insecurity that capture the material dimension of disadvantage.(32) The present model did not include these measures because they are not available for every county at the same resolution as the ACS indicators, but their omission probably means the predictive value of deprivation is underestimated rather than overstated.

Not every study finds a strong area-level effect. Research using individual-level data in some settings reports that neighborhood or county disadvantage contributes comparatively little to tooth loss once individual income and education are accounted for, whereas other studies find that contextual deprivation matters independently. These divergences probably reflect differences in the geographic scale studied, the covariates available for adjustment, and the extent of residential socioeconomic segregation in each setting.(14,16) The present county-level estimates, which lack individual-level adjustment, should be read as capturing the combined individual and contextual dimensions of disadvantage rather than isolating one from the other.

Some individual-level studies emphasize clinical and behavioral mediators such as smoking, diabetes, and dental attendance rather than income itself.(8,12) The permutation ranking here, in which smoking and fair or poor self-rated health also rank highly alongside poverty, is compatible with a pathway in which deprivation operates partly through these intermediate factors. Social support has likewise been linked to severe tooth loss among older adults, suggesting that the relevant exposures extend beyond material resources.(33) That low educational attainment showed a smaller but still substantial independent effect is consistent with education operating through knowledge, health literacy, and lifetime patterns of care seeking rather than through current income alone.(6) The near-zero bivariate correlation of the proportion aged 65 years or older with tooth loss is reassuring, because it indicates that the geographic pattern is not simply an artifact of counties that differ in their age composition.

Hypothesis 2 was not supported. When folds were assigned at random, a well-specified linear model captured most of the predictive signal in these variables, and gradient boosting added no accuracy. This echoes a broader pattern in which simpler models match or approach tree ensembles for many tabular health prediction tasks, particularly when the underlying relationships are close to linear.(22,34) The result is a useful caution against assuming that model complexity confers predictive value in ecological health data, and it also carries a positive implication for practice: where the goal is within-sample ranking and the relationships are linear, a transparent regression whose coefficients can be read directly is preferable to an ensemble precisely because it is easier to communicate and audit. The near-equivalence under random validation also implies that the value of the more complex model lies entirely in its robustness to spatial transfer, a property that would be invisible in a conventional single-validation analysis. The non-linear model earns its place only where it is actually required, which the spatial results below indicate is the case for out-of-area prediction.

Hypothesis 3 was supported, and this is the central methodological contribution. The divergence between models appeared only under spatial validation, where the linear model's R-squared fell to roughly half of its random value while the gradient-boosted model retained more of its performance. The leave-one-state-out result fell between the random and spatial-block estimates, as expected, because state boundaries group counties less tightly than coordinate-based blocks, so they leak less information across folds than a random split but more than a purely spatial partition. This finding aligns with evidence that random cross-validation inflates reported performance for spatially structured data and that spatial validation reveals poor transferability of large-scale mapping models.(24,25) It also matches methodological work showing that the choice of cross-validation strategy changes both the estimated performance and the interpretation of which predictors matter.(26,29) Studies in species distribution modeling and agricultural prediction have reported the same pattern, in which models that appeared accurate under conventional validation degraded sharply once spatial blocking was imposed.(30,31)

Several mechanisms may explain why the machine learning model generalizes better across space. Gradient boosting captures non-linearities and interactions, such as the sharper increase in tooth loss at the extreme of the poverty distribution visible in the partial dependence curve, and its subsampling regularization may be more robust to the spatial clustering of unmeasured regional factors than a global linear fit. The interpretability tools applied here show that the model's spatial generalization does not come at the cost of opacity: poverty and low educational attainment have monotonic, plausible relationships with the outcome, and the ranking of determinants is consistent with the regression analysis. Substantively, the poverty-tooth loss association likely operates through several interlocking channels, including the cost of dental care relative to income, competing demands on household budgets, higher smoking prevalence and diabetes burden, and differential access to preventive services across the life course.(11,12) The inverse association of dental visit prevalence with tooth loss, and the prominence of smoking and self-rated health in the importance ranking, are consistent with these intermediate pathways. The downstream health consequences of edentulism, including frailty, cognitive impairment, and cardiovascular risk, give these predictions clinical weight beyond the oral cavity.( 7,10,35)

Beyond the substantive findings, the study illustrates a methodological point of general relevance. The two models were indistinguishable under random validation and yet differed by roughly twenty percentage points of explained variance under spatial validation. Any small-area model intended to inform decisions outside its training area should therefore report spatial as well as random performance, because the two can lead to opposite conclusions about which model to deploy. This concern extends to county and subnational health models beyond oral health, wherever neighboring units share unmeasured characteristics, and it reinforces the argument that validation strategy is a scientific choice rather than a technical detail.(26,31) Reporting guidance for prediction models now emphasizes validation in settings that differ from the development data, and the present comparison suggests that geographic structure is one of the most consequential dimensions along which a development and an application setting can differ for area-level health models.

Taken together, the three hypotheses point to a practical conclusion for population oral health. If the goal is to rank counties by predicted tooth loss for resource allocation within the area used for training, a linear model is adequate. If the goal is to predict previously unseen counties or regions, the machine learning model is the more defensible choice, and its feature ranking remains interpretable. The difference between the two validation schemes, rather than any property of the algorithm itself, is what separates a model that appears accurate from one that is accurate where it will actually be used. For researchers building small-area oral health models, a simple rule follows: report both random and spatial performance, and choose the model class according to how the model will actually be deployed.

The strengths of the analysis include its national coverage of nearly three thousand counties, the use of open and reproducible data, the direct comparison of a transparent linear baseline against a flexible ensemble, and the inclusion of both random and spatial validation, which allowed the two to be contrasted. The study also has limitations. First, it is ecological, so the associations describe counties rather than individuals and cannot rule out confounding by unmeasured county characteristics; an individual-level multilevel design would be needed to separate compositional and contextual effects. Second, the outcome and several covariates are model-based small-area estimates rather than direct observations, so measurement error and the borrowing of information across neighboring areas may attenuate or, for the outcome, partially induce spatial structure; direct survey or administrative counts would strengthen inference. Third, the analysis is cross-sectional and identifies associations, not causal effects. Fourth, spatial blocks were defined by k-means clustering rather than administrative or ecological boundaries, so results may differ under alternative spatial partitions, although the leave-one-state-out sensitivity was concordant. Fifth, complete-case analysis excluded 197 counties with missing data, most of them small, and their omission could influence the estimated associations; multiple imputation or inclusion of county-equivalent units would address this. Sixth, the set of social determinants was limited to those available from a single ACS-derived source, and it omits measures of area-level wealth, employment quality, and racial residential segregation that may capture additional predictive signal; future models could incorporate these to test whether the ranking of poverty is robust.

Future work could extend this analysis in several directions. Adding county-level measures of dental workforce supply, insurance coverage policy, and water fluoridation would test whether supply-side factors modify the deprivation gradient. Multilevel designs linking individual survey respondents to their counties would separate the compositional and contextual contributions that the ecological design conflates. Applying the same spatially validated framework to other oral outcomes, such as untreated caries or periodontal disease, would indicate whether the spatial transferability advantage of non-linear models is specific to tooth loss or a general feature of small-area oral health modeling.

For practice and policy, these results suggest that county-level models of edentulism can support the geographic targeting of oral health programs for older adults, but only if their performance is reported under spatial validation. The finding that deprivation, and poverty in particular, is the leading social determinant indicates that interventions addressing affordability and access are likely to yield the largest reductions in geographic inequality in tooth loss. For example, a program that subsidizes dentures or expands coverage for older adults could use the predicted county ranking to prioritize outreach to the counties where the expected burden, and therefore the return on each dollar, is highest. Because the predictors are drawn from routine open data, the approach can be reproduced in other countries that publish small-area health and census indicators, which would test whether the social gradient generalizes across health systems.

CONCLUSIONS

County socioeconomic deprivation, and poverty above all, independently predicted complete tooth loss among older adults across United States counties, supporting H1. Gradient boosting did not exceed a linear model under random cross-validation, so H2 was not supported. The machine learning model transferred better across space and retained the ranking of social determinants under spatial validation, supporting H3 and confirming the need for spatially aware evaluation of small-area predictive models.

References

1. 1 Borg-Bartolo R, Roccuzzo A, Molinero-Mourelle P, Schimmel M, Gambetta-Tessini K, Chaurasia A, et al. Global prevalence of edentulism and dental caries in middle-aged and elderly persons: A systematic review and meta-analysis. Journal of Dentistry. 2022;127:104335. doi: 10.1016/j.jdent.2022.104335
2. 2 Chen H, Jing Y, Song Y, Jia Z, He X, Zhang S, et al. Edentulism Trends in Chinese Women vs US and Global Women: GBD 2021 Outlook. IJWH. 2025;Volume 17:4661-4670. doi: 10.2147/ijwh.s558621
3. 3 Qin X, He J, He H, Yuan X, Su X, Zeng X. Long-term trends in the burden of edentulism in China over three decades: A Joinpoint regression and age-period-cohort analysis based on the global burden of disease study 2019. Front Public Health. 2023;11:1099194. doi: 10.3389/fpubh.2023.1099194
4. 4 Wu B, Liang J, Plassman BL, Remle C, Luo X. Edentulism trends among middle‐aged and older adults in the United States: comparison of five racial/ethnic groups. Comm Dent Oral Epid. 2012;40(2):145-153. doi: 10.1111/j.1600-0528.2011.00640.x
5. 5 Wu B, Liang J, Plassman BL, Remle RC, Bai L. Oral health among white, black, and Mexican-American elders: an examination of edentulism and dental caries. Journal of Public Health Dentistry. 2011;71(4):308-317. doi: 10.1111/j.1752-7325.2011.00273.x
6. 6 Wu B, Hybels C, Liang J, Landerman L, Plassman B. Social stratification and tooth loss among middle‐aged and older Americans from 1988 to 2004. Comm Dent Oral Epid. 2014;42(6):495-502. doi: 10.1111/cdoe.12116
7. 7 Matsuyama Y, Jürges H, Listl S. Causal Effect of Tooth Loss on Cardiovascular Diseases. J Dent Res. 2023;102(1):37-44. doi: 10.1177/00220345221120164
8. 8 Tung H, Ford R. Incident edentulism and number of comorbidities among middle‐aged and older Americans. Gerodontology. 2023;40(4):484-490. doi: 10.1111/ger.12675
9. 9 Mira R, Newton JT, Sabbah W. The Longitudinal Relationship between Edentulism and the Progress of Multimorbidity. Nutrients. 2024;16(14):2234. doi: 10.3390/nu16142234
10. 10 Dibello V, Lobbezoo F, Lozupone M, Sardone R, Ballini A, Berardino G, et al. Oral frailty indicators to target major adverse health-related outcomes in older age: a systematic review. GeroScience. 2023;45(2):663-706. doi: 10.1007/s11357-022-00663-8
11. 11 Gilbert GH, Paul duncan R, Shelton BJ. Social Determinants of Tooth Loss. Health Services Research. 2003;38(6p2):1843-1862. doi: 10.1111/j.1475-6773.2003.00205.x
12. 12 Lee H, Kim D, Jung A, Chae W. Ethnicity, Social, and Clinical Risk Factors to Tooth Loss among Older Adults in the U.S., NHANES 2011–2018. IJERPH. 2022;19(4):2382. doi: 10.3390/ijerph19042382
13. 13 Mitchell J, Bennett K, Brock‐Martin A. Edentulism in High Poverty Rural Counties. The Journal of Rural Health. 2013;29(1):30-38. doi: 10.1111/j.1748-0361.2012.00440.x
14. 14 Chong GTF, Gansky SA. Census tract geospatial analysis comparing social determinants of health with tooth loss in California seniors: An ecologic study. Comm Dent Oral Epid. 2024;52(6):889-899. doi: 10.1111/cdoe.12995
15. 15 Pitchika V, Jordan RA, Norderyd O, Rolander B, Welk A, Völzke H, et al. Factors influencing tooth loss in European populations. J Clinic Periodontology. 2022;49(7):642-653. doi: 10.1111/jcpe.13642
16. 16 Kailembo A, Preet R, Stewart Williams J. Socioeconomic inequality in self-reported unmet need for oral health services in adults aged 50 years and over in China, Ghana, and India. Int J Equity Health. 2018;17(1):99. doi: 10.1186/s12939-018-0812-2
17. 17 Wang L, Cheng L, Yuan B, Hong X, Hu T. Association between socio-economic status and dental caries in elderly people in Sichuan Province, China: a cross-sectional study. BMJ Open. 2017;7(9):e016557. doi: 10.1136/bmjopen-2017-016557
18. 18 Jamieson L, Wu B, Chrisopoulos S, Luzzi L, Mejia G, Ju X. Oral health inequities over time among older adults of different racial/ethnic backgrounds: a comparative decomposition analysis across Australia and the United States. BMC Geriatr. 2024;24(1):936. doi: 10.1186/s12877-024-05525-w
19. 19 Greenlund KJ, Lu H, Wang Y, Matthews KA, LeClercq JM, Lee B, et al. PLACES: Local Data for Better Health. Prev Chronic Dis. 2022;19:210459. doi: 10.5888/pcd19.210459
20. 20 Wang Y, Holt JB, Zhang X, Lu H, Shah SN, Dooley DP, et al. Comparison of Methods for Estimating Prevalence of Chronic Diseases and Health Behaviors for Small Geographic Areas: Boston Validation Study, 2013. Prev Chronic Dis. 2017;14:170281. doi: 10.5888/pcd14.170281
21. 21 Rasheed K, Qayyum A, Ghaly M, Al-Fuqaha A, Razi A, Qadir J. Explainable, trustworthy, and ethical machine learning for healthcare: A survey. Computers in Biology and Medicine. 2022;149:106043. doi: 10.1016/j.compbiomed.2022.106043
22. 22 Woodman RJ, Mangoni AA. A comprehensive review of machine learning algorithms and their application in geriatric medicine: present and future. Aging Clin Exp Res. 2023;35(11):2363-2397. doi: 10.1007/s40520-023-02552-2
23. 23 van Schaik P, Peng Y, Ojelabi A, Ling J. Explainable statistical learning in public health for policy development: the case of real-world suicide data. BMC Med Res Methodol. 2019;19(1):152. doi: 10.1186/s12874-019-0796-7
24. 24 Ploton P, Mortier F, Réjou-Méchain M, Barbier N, Picard N, Rossi V, et al. Spatial validation reveals poor predictive performance of large-scale ecological mapping models. Nat Commun. 2020;11(1):4540. doi: 10.1038/s41467-020-18321-y
25. 25 Kattenborn T, Schiefer F, Frey J, Feilhauer H, Mahecha MD, Dormann CF. Spatially autocorrelated training and validation samples inflate performance assessment of convolutional neural networks. ISPRS Open Journal of Photogrammetry and Remote Sensing. 2022;5:100018. doi: 10.1016/j.ophoto.2022.100018
26. 26 Sweet LB, Müller C, Anand M, Zscheischler J. Cross-Validation Strategy Impacts the Performance and Interpretation of Machine Learning Models. Artificial Intelligence for the Earth Systems. 2023;2(4). doi: 10.1175/aies-d-23-0026.1
27. 27 Centers for Disease Control and Prevention. PLACES: Local Data for Better Health, County Data, 2025 release [Internet]. 2025. Available from: https://data.cdc.gov/Public-Health-Surveillance/PLACES-Local-Data-for-Better-Health-County-Data-2/swc5-untb
28. 28 Centers for Disease Control and Prevention. Non-Medical Factor Measures for County, American Community Survey 2017-2021 [Internet]. 2023. Available from: https://data.cdc.gov/500-Cities-Places/Non-Medical-Factor-Measures-for-County-ACS-2017-/i6u4-y3g4
29. 29 Deppner J, Cajias M. Accounting for Spatial Autocorrelation in Algorithm-Driven Hedonic Models: A Spatial Cross-Validation Approach. J Real Estate Finan Econ. 2024;68(2):235-273. doi: 10.1007/s11146-022-09915-y
30. 30 Milà C, Mateu J, Pebesma E, Meyer H. Nearest neighbour distance matching Leave‐One‐Out Cross‐Validation for map validation. Methods Ecol Evol. 2022;13(6):1304-1316. doi: 10.1111/2041-210x.13851
31. 31 Yates LA, Aandahl Z, Richards SA, Brook BW. Cross validation for model selection: A review with examples from ecology. Ecological Monographs. 2023;93(1):e1557. doi: 10.1002/ecm.1557
32. 32 Lee JJ, Kim Y, Scott JM, Hill CM, Chi DL. The association of food insecurity with tooth decay and periodontitis among middle‐aged and older adults in the United States and Korea. Gerodontology. 2023;40(2):251-262. doi: 10.1111/ger.12651
33. 33 Laniado N, Cloidt M, Shah P. Social support and oral health among working‐age and older adults in the United States. J Public Health Dent. 2023;83(3):247-253. doi: 10.1111/jphd.12573
34. 34 Bellinger C, Mohomed Jabbar MS, Zaïane O, Osornio-Vargas A. A systematic review of data mining and machine learning for air pollution epidemiology. BMC Public Health. 2017;17(1):907. doi: 10.1186/s12889-017-4914-3
35. 35 Choi SE, Mo E, Palmer N, Fox K, Da Silva JD, Nagai S, et al. Cognitive impairment and edentulism among older adults: an observational study using claims data. BMC Geriatr. 2022;22(1):278. doi: 10.1186/s12877-022-02985-w