Abstract
Tooth loss has been associated with increased mortality risk in older adults. This study examined heterogeneity in the association between tooth loss and mortality among older Japanese adults. Data were obtained from 69,265 individuals aged ≥65 y who participated in the 2013 wave of the Japan Gerontological Evaluation Study. Questionnaire survey responses were linked to mortality records from municipal registries for 9 y. The average treatment effect (ATE) of tooth loss (having <20 natural teeth) on all-cause mortality was estimated using a linear probability model with an inverse probability weighting estimator. Heterogeneity in the association was assessed by estimating conditional average treatment effects (CATEs) using the causal forest machine learning algorithm. A set of 44 covariates, including demographic, economic, social, health, and community-related factors, was incorporated as potential drivers of heterogeneity. During the follow-up period, 26.4% of participants with <20 teeth and 14.4% of those with ≥20 teeth died. Tooth loss was significantly associated with increased mortality risk by 3.2 percentage points after adjustment for covariates (average treatment effect = 0.032; 95% confidence interval, 0.023–0.040). A statistically significant heterogeneity in the estimated effects was observed (median CATE = 0.028; interquartile range = 0.010). Greater effects were identified among subgroups characterized by men, poor health status, and lower socioeconomic conditions. Heart disease, sex, and depression were the most influential contributors to this heterogeneity. Tooth loss was associated with mortality risk, with significant variation in the magnitude across population subgroups. Tooth loss may serve as an informative indicator of elevated mortality risk among older adults, particularly among vulnerable subgroups.
Introduction
The association between tooth loss and the risk of diseases and mortality has been well documented (Koka and Gupta 2018; Seitz et al 2019). A recent systematic review reported a dose–response relationship, showing that complete loss of natural teeth was associated with a 1.5-fold higher risk of mortality (Peng et al 2019). With global population aging, the prevalence of edentulism has doubled over the past 30 y, and the number of incident cases is projected to reach 32 million by 2050 (Bernabe et al 2025; Yan et al 2025).
While causality should be carefully interpreted, several mechanisms have been proposed to explain tooth loss as a potential risk factor for mortality. For example, tooth loss can impair nutrition (Iwasaki et al 2021), which in turn may increase mortality risk (Hiratsuka et al 2020; Kusama et al 2022). Periodontal pathogens may contribute to cardiovascular disease by accelerating atherosclerosis (Kebschull et al 2010). The systemic inflammation pathway has also been proposed, although evidence remains mixed (Peng et al 2019; Hiratsuka et al 2020). Furthermore, tooth loss may adversely affect social relationships (Cooray et al 2023), and poor social relationships are risk factors for mortality (Holt-Lunstad et al 2010). Accordingly, preventing tooth loss may contribute to healthy aging among older adults. Even if the association is not causal, examining the relationship between tooth loss and mortality remains valuable, as tooth loss may serve as a reliable indicator of mortality risk in later life.
Previous studies have primarily examined the average effect of tooth loss on mortality, assuming that the magnitude of the association is constant across individuals. Few studies have directly assessed whether the impact of tooth loss on mortality differs by individual characteristics, although existing evidence suggests possible effect modification of oral health on general health outcomes. For example, the association between tooth loss and cognitive decline was stronger among older adults (Chen et al 2022), and working-age adults with health problems were more vulnerable to the effects of oral conditions on self-rated general health (Brennan and Teusner 2015). In addition, frequent dental visits prevented tooth loss only among individuals with additional risk factors (Giannobile et al 2013), suggesting an interaction between dental care and patient background characteristics. These findings indicate the need for further investigation into effect modifiers of the association between oral health and general health. However, most studies in dentistry have considered only 1 or a limited number of effect modifiers, without examining the combined influence of multiple factors.
Recent advances in causal inference theory and the development of machine learning algorithms have made it possible to incorporate various factors that may derive effect heterogeneity based on high-dimensional data (Athey et al 2019; Shiba and Inoue 2024). A previous study showed that the association between tooth loss and loss of functional capacity was stronger in subgroups characterized by older age, men, unmarried people, those with a lower socioeconomic status, and individuals in poor health (Matsuyama et al 2024). Identifying vulnerable subgroups may facilitate more effective resource allocation and provide insights into the mechanisms linking tooth loss with general health outcomes.
The present study aimed to examine the effect of heterogeneity in the association between tooth loss and mortality among Japanese older adults. A state-of-the-art machine learning–based causal inference framework was employed to account for diverse combinations of background factors and to identify high-dimensional effect heterogeneity.
Methods
Study Participants
This cohort study used data from the Japan Gerontological Evaluation Study (Kondo et al 2018). In 2013, a questionnaire survey was administered to adults aged ≥65 y across 29 municipalities (response rate = 70.7%). Among these, 17 municipalities agreed to provide mortality records. These 17 municipalities had larger populations and higher population density, as well as lower proportions of older residents, higher average incomes, and higher educational attainment than the remaining 12 municipalities. The lengths of follow-up varied slightly across these municipalities, ranging from 9.1 to 9.6 y after the baseline survey. Of the 77,694 respondents residing in the 17 municipalities, 76,472 were successfully linked to mortality records (follow-up rate = 98.4%; median follow-up period = 9.2 y). After excluding participants who were dependent in activities of daily living at baseline (n = 5,218), those who moved out and were lost to follow-up (n = 1,715), and those who died within 6 mo of baseline (n = 274), the final analytic sample consisted of 69,265 individuals.
Mortality
The outcome was all-cause mortality, identified from municipal registry records. The date of death for survey respondents was obtained from the Japanese Long-Term Care Insurance registry, a nationwide system that provides care and support for older adults. The information was linked to the survey data using unique identification numbers. Individuals with potential linkage errors, such as those with an age discrepancy greater than 4 years, were excluded. The Japanese tracking system allowed for a high follow-up rate of 98.4%. For the main analysis, mortality was treated as a dichotomous variable indicating death during follow-up. In addition, follow-up duration was used in survival analysis to confirm the association.
Tooth Loss
The exposure was tooth loss, assessed with the following self-reported question (Petersen et al 2013): “How many natural teeth do you have?” Response options included “no natural teeth,” “1–4 natural teeth,” “5–9 natural teeth,” “10–19 natural teeth,” and “20 or more natural teeth.” For analysis, the variable was dichotomized as having ≥20 teeth versus <20 teeth, following the World Health Organization’s criteria for functional dentition (Petersen et al 2013).
Covariates
Forty-four variables were included as covariates, capturing demographic characteristics (e.g., age, sex), socioeconomic status (e.g., income, education), health status (e.g., comorbid conditions, depressive symptoms [Shin et al 2019], and functional capacity [Koyano et al 1991]), health-related behaviors (e.g., smoking, alcohol drinking, physical activity; Sato et al 2021), social relationships (e.g., marital status, number of friends), and municipality-level factors (e.g., density of dental clinics). These variables were selected from the survey items as potential confounders or effect modifiers, based on prior literature and expert knowledge. The list of variables and coding is provided in Appendix Table 1. For machine learning and regression analyses, categorical variables were coded into smaller categories to maximize information, while broader categories were used for descriptive purposes to improve interpretability (Table 1).
Demographic Characteristics of the Study Participants.
Values are presented as number (%) unless otherwise indicated.
M JPY, million Japanese yen.
Assessed using the 15-item Geriatric Depression Scale (Shin et al 2019); higher values indicate greater depressive symptoms.
Assessed using the 13-item Tokyo Metropolitan Institute of Gerontology Index of Competence Scale (Koyano et al 1991); higher values indicate greater functional capacity.
Calculated based on the self-reported frequency of physical activities and their intensity expressed in metabolic equivalents of task (METs) (Sato et al 2021); higher values indicate greater physical activity.
Statistical Analysis
First, the association between tooth loss and mortality was examined by estimating average treatment effects (ATEs). A linear probability model (LPM) was fitted to estimate ATEs on the risk difference scale. In addition, Cox proportional hazards model was used to confirm the robustness of the association. Both models were estimated using the inverse probability weighting (IPW) estimator, with stabilized weights for the probability of having <20 teeth calculated from a logistic regression including all 44 covariates (Hernán and Robins 2020).
Second, heterogeneity in the association between tooth loss and mortality was examined by estimating conditional average treatment effects (CATEs). We applied the causal forest method from the generalized random forest (GRF) algorithm (Athey et al 2019). Unlike random forests, the prediction target of the causal forest is the effect of an exposure on an outcome rather than the outcome itself. Details of the causal forest are reported elsewhere (Athey et al 2019; Shiba and Inoue 2024). In this study, all tunable parameters were optimized by cross-validation, and 4,000 trees were grown from bootstrapped samples. All 44 covariates were included. Predictions were made using 20-fold cross-fitting. Model performance was evaluated by estimating group average treatment effects (GATEs) for quintile subgroups of estimated CATEs using an augmented inverse propensity weighted (AIPW) estimator. Best linear predictor (BLP) analysis was also performed to formally test for the presence of effect heterogeneity (Chernozhukov et al 2018).
Third, to identify the factors strongly contributing to effect heterogeneity, we applied the Extreme Gradient Boosting (XGBoost) algorithm, a tree-based machine learning method (Chen et al 2025). Estimated CATEs were modeled using the 44 covariates. For hyperparameter tuning, the data were split into training (80%) and validation (20%) subsets. Two hundred parameter combinations were randomly sampled from predefined grids, and the set minimizing the root mean squared error was used for the final prediction on the full dataset (Appendix Table 2). Variables with larger absolute Shapley additive explanations (SHAP) values were considered greater drivers of effect heterogeneity and were used to construct a heatmap to visualize the heterogeneity pattern.
Missing information on variables was imputed by random forest imputation (Stekhoven and Buhlmann 2012). The proportion of missingness ranged from 0% (e.g., age and sex) to 26.1% (equivalent wealth) (Appendix Table 3). The characteristics of the imputed sample more closely resembled those of the baseline sample than those of the complete case sample (Appendix Fig. 1).
Analyses were conducted using Stata MP 19.5 and R 4.5.1. The R packages grf (Tibshirani et al 2024), xgboost (Chen et al 2025), and missRanger (Mayer 2024) were used for GRF, XGBoost, and random forest imputation, respectively. Other analyses were performed using Stata MP 19.5. The study followed the Strengthening the Reporting of Observational Studies in Epidemiology guidelines.
Ethics Approval
This study obtained approval from the ethics committees of Nihon Fukushi University (13-14), Chiba University (No. 2493 and 3442), the National Center for Geriatrics and Gerontology (No. 992 and 1274-2), the Japan Agency for Gerontological Evaluation Study (No. 2019-01), and the Institute of Science Tokyo (No. D2022-040).
Results
Table 1 and Appendix Table 4 describe the demographic characteristics of participants. Overall, 20.1% of participants died during follow-up. The mortality rate was 26.4% among individuals with <20 teeth and 14.4% among those with ≥20 teeth. Participants with <20 teeth were more likely to be older, men, of lower socioeconomic status, and in poorer health (Table 1). They were also more likely to be unmarried, to have poorer social networks and a lack of social support, and to reside in municipalities with slightly lower municipal average income, lower density of dental clinics in the municipality, and lower population density in the municipality (Appendix Table 4).
Table 2 shows the ATEs estimated using IPW. Tooth loss was associated with a 3.2-percentage-point higher risk of mortality (coefficient = 0.032; 95% confidence interval [CI], 0.023–0.040). A similar association was observed using the Cox proportional hazards model (hazard ratio = 1.19; 95% CI, 1.14–1.25).
Association between Tooth Loss and Mortality with Inverse Probability Weighting Estimator.
CI, confidence interval; HR, hazard ratio; LPM, linear probability model; PH, proportional hazard.
Figure 1 illustrates the distribution of estimated CATEs, which was slightly right-skewed, ranging from 0.001 to 0.070. The median, 25th percentile, and 75th percentile were 0.028, 0.024, and 0.034, respectively. These results show that tooth loss was associated with increased mortality across most of the population, with a stronger impact in certain subgroups.
Distribution of estimated CATEs. The right side of the distribution corresponds to individuals for whom tooth loss has a stronger effect on mortality. CATE, conditional average treatment effect.
Model performance tests indicated good calibration. Estimated GATE values increased across subgroups defined by quintiles of CATEs (Appendix Fig. 2). BLP analysis (Appendix Table 5) showed that the forests were well calibrated (coefficient for mean forest prediction = 0.99; P < 0.001) and captured significant effect heterogeneity (coefficient for differential forest prediction = 1.17; P = 0.003).
Table 3 and Appendix Table 6 describe the participant characteristics by quintiles of CATEs. Subgroups with greater CATEs were more likely to include men, individuals with lower socioeconomic status, those in poorer health, and those with unhealthy behaviors (Table 3). They also exhibited poorer social relationships and tended to live in municipalities with a higher population density (Appendix Table 6).
Demographic Characteristics of the Study Participants by Quintile Categories of Estimated Conditional Treatment Effects.
Values are presented as percentages unless otherwise indicated.
CATE, conditional average treatment effect; M JPY, million Japanese yen.
Assessed using the 15-item Geriatric Depression Scale (Shin et al 2019); higher values indicate greater depressive symptoms.
Assessed using the 13-item Tokyo Metropolitan Institute of Gerontology Index of Competence Scale (Koyano et al 1991); higher values indicate greater functional capacity.
Calculated based on the self-reported frequency of physical activities and their intensity expressed in metabolic equivalents of task (METs) (Sato et al 2021); higher values indicate greater physical activity.
Appendix Figure 3 presents the mean absolute SHAP values. The presence of heart disease, sex, and depressive symptoms was considered the 3 strongest contributors to effect heterogeneity. Some variables exhibited relatively smaller SHAP values, suggesting that they contributed only modestly as individual effect modifiers. Although the influence of their combined effects is incorporated into the CATE estimation, the extent of this influence remains uncertain. As shown in Figure 2, the average CATEs differed across subgroups defined by the presence of heart disease, sex, and depressive symptoms, ranging from 0.024 for women without heart disease or depression to 0.054 for men with heart disease and severe depression.
Heatmap of estimated conditional average treatment effects by sex, heart disease, and depression. Values represent the mean (standard deviation) of CATEs in each stratum. Higher values indicate a stronger effect of tooth loss on mortality: for example, tooth loss was associated with a 5.0-percentage-point increase in the risk of mortality for men with heart disease and severe depressive symptoms. Depression was grouped based on the Geriatric Depression Scale score: none (score 0 to 4), moderate (score 5 to 9), and severe (score 10 to 15). CATE, conditional average treatment effect.
Discussion
This study is the first to report high-dimensional effect heterogeneity in the association between tooth loss and all-cause mortality among older adults. Using a machine learning–based causal inference framework, we demonstrated that tooth loss was associated with increased mortality risk over 9 y, with varying magnitudes across subgroups. On average, tooth loss was associated with a 3.2-percentage-point higher mortality risk. The magnitude was stronger in subgroups characterized by men, lower socioeconomic status, and poorer health. In particular, male sex, heart disease, and depressive symptoms were major contributors to the heterogeneity, with estimated effects ranging from a 2.4-percentage-point increase among women without heart disease or depression to a 5.4-percentage-point increase among men with heart disease and severe depression.
Although it is difficult to determine whether the magnitude of heterogeneity observed in this study is large or small—partly due to the scarcity of prior research examining heterogeneous associations between health conditions such as tooth loss and mortality—the magnitudes of both average association and the heterogeneity are likely to be clinically meaningful given that the outcome is mortality. For comparison, an individual participant-level meta-analysis reported that the effects of pharmacological treatments to lower blood pressure on cardiovascular mortality among older adults varied by age, with reductions of up to 1 percentage point (Rahimi et al 2021).
The findings are consistent with previous research reporting associations between tooth loss and mortality (Koka and Gupta 2018; Peng et al 2019). To our knowledge, no previous studies have formally tested effect modification in the association of tooth loss with mortality. Nonetheless, our results align with previous studies reporting that the impact of tooth loss on cognitive impairment was greater among older adults (Chen et al 2022), and the impact of oral conditions on overall health was greater among those with health problems (Brennan and Teusner 2015). The present study extends these findings by applying a machine learning approach to detect high-dimensional heterogeneity in the association between tooth loss and mortality, a hard objective outcome.
The observed heterogeneity may be interpreted in the context of multimorbidity, a risk of excess mortality. Interactions among multiple conditions, as well as greater health care needs, can elevate mortality risk (Nunes et al 2016). Tooth loss has recently been incorporated into the multimorbidity concept (Mirza et al 2024), and its coexistence with other health conditions may increase mortality risk. Heart disease, in particular, was shown as a major contributor to heterogeneity. This result is in line with evidence that tooth loss increases the risk of cardiovascular disease mortality (Aminoshariae et al 2024). Men are more likely to have poorer oral conditions and exhibit a negative attitude toward dental care (Lipsky et al 2021). This gender difference in oral health may result in the greater effect of tooth loss on mortality in men than in women. While the contribution of socioeconomic status appeared modest, poorer health in the vulnerable subgroups may reflect socioeconomic disadvantages, and barriers to health care may also contribute to the mortality impact of tooth loss.
Oral diseases are highly prevalent in older populations (Kassebaum et al 2014). Recognizing this burden, the World Health Organization (WHO 2024) recently set a global target to reduce oral diseases by 10% by 2030. The present findings suggest that differentiated interventions for individuals with conditions such as heart disease and depression may be effective for extending longevity by improving oral health. A high-benefit approach has been proposed in the medical field (Inoue et al 2023). Although its application to dental diseases requires careful consideration—given their high prevalence and significant health inequalities (Peres et al 2019)—a combination of population and high-benefit approaches may maximize the effectiveness of policy and clinical-level interventions in dentistry.
This study has several strengths. The use of a state-of-the-art causal inference framework with machine learning enabled us to incorporate diverse variables and their combinations to detect high-dimensional effect heterogeneity. This approach allowed for a detailed characterization of vulnerable subgroups. In addition, we data-adaptively identified factors that contributed most to the observed heterogeneity, providing insights into its underlying sources.
This study also has limitations. First, although a rich set of covariates was included, residual confounding may still exist. For example, previous studies have additionally adjusted for biomarkers such as lipid profiles and inflammatory markers, family history of heart disease, parental socioeconomic position, and preventive oral care (Peng et al 2019). These unmeasured or unknown confounders may explain the observed association.
Second, tooth loss was measured by self-report; however, previous studies support the validity of self-reported tooth counts (Matsui et al 2017). Third, effect modification does not necessarily indicate causal interaction, and it remains unclear whether interventions on the identified effect modifiers would reduce the impact of tooth loss on mortality. Fourth, tooth loss was dichotomized as having ≥20 or <20 natural teeth to be incorporated into the causal forest algorithm. Although this definition aligns with the WHO criterion for functional dentition (Petersen et al 2013), variation in the number of natural teeth within each category may violate the consistency assumption. Fifth, we did not account for changes in factors during follow-up. Sixth, heterogeneity patterns may vary by the algorithms (Inoue et al 2024). Although meta-learners offer advantages in imbalanced designs (Künzel et al 2019; Inoue et al 2024), the causal forest was well suited for this study, given the balanced sizes of exposed (47.2%) and nonexposed groups (52.8%) and its straightforward implementation. Seventh, generalizability may be limited to older adults in Japan. For instance, Japanese health care systems may have influenced results. Further study in other settings is warranted.
In conclusion, the present study found that tooth loss was associated with all-cause mortality among Japanese older adults. The magnitude of the association varied by participant characteristics, with subgroups characterized by men; poor health conditions, particularly heart disease and depression; and lower socioeconomic status exhibiting a stronger association. Tooth loss may serve as an informative indicator of elevated mortality risk among older adults, particularly among these subgroups.
Author Contributions
Y. Matsuyama, contributed to conception and design, data acquisition, analysis, and interpretation, drafted the manuscript; S. Kiuchi, T. Yamamoto, J. Aida, contributed to data acquisition and interpretation, critically revised the manuscript. All authors gave final approval and agreed to be accountable for all aspects of the work.
Supplemental Material
sj-pdf-1-jdr-10.1177_00220345251414360 – Supplemental material for High-Dimensional Effect Heterogeneity of Tooth Loss on Mortality
Supplemental material, sj-pdf-1-jdr-10.1177_00220345251414360 for High-Dimensional Effect Heterogeneity of Tooth Loss on Mortality by Y. Matsuyama, S. Kiuchi, T. Yamamoto and J. Aida in Journal of Dental Research
Footnotes
Acknowledgements
The authors thank the study participants and municipal officials’ cooperation with the survey. The authors used ChatGPT 5 (
) during the preparation of this work only to correct grammatical errors and improve readability, with minimal adjustments to adhere to academic English standards throughout the manuscript written by the first author. The authors reviewed and edited the content as needed to ensure scientific rigor and accuracy, and they take full responsibility for the publication’s content.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study used data from JAGES (the Japan Gerontological Evaluation Study). This study was supported by Grant-in-Aid for Scientific Research (20H00557, 20K10540, 21H03153, 21H03196, 21K17302, 21K19635, 22H00934, 22H03299, 22K04450, 22K13558, 22K17409, 22K17285, 23H00449, 23H03117, 24K02658) from the Japan Society for the Promotion of Science (JSPS), Health Labour Sciences Research Grants (19FA1012, 19FA2001, 21FA1012, 22FA2001, 22FA1010, 22FG2001, 23FA1022), Research Institute of Science and Technology for Society (JPMJOP1831) from the Japan Science and Technology (JST), a grant from Japan Health Promotion & Fitness Foundation, and the National Research Institute for Earth Science and Disaster Resilience. The views and opinions expressed in this article are those of the authors and do not necessarily reflect the official policy or position of the respective funding organizations.
Data Availability
The use of JAGES data is available from the corresponding author on reasonable request. All enquiries should be addressed to the JAGES data management committee via e-mail:
A supplemental appendix to this article is available online.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
