This ecological panel study examined whether the home-death rate reflected the actual level of home-based end-of-life care across Japanese municipalities. The unit of analysis was the municipality (shi, cho, or son), and data were collected at three time points (2017, 2020, and 2023), yielding 5,223 observations across 1,741 municipalities. Municipality codes were harmonised across the survey years using the Survey of Municipality Areas from the Geospatial Information Authority of Japan. One municipality was excluded because it reported zero deaths in 2023, resulting in a final analytical sample of 1,740 municipalities.

Data on the medical institutions, home-visit nursing services, population characteristics, and number of home deaths certified by clinics were obtained from the Regional Dataset on Home Medical Care provided by the Ministry of Health, Labour, and Welfare. ¹¹ Because the survey on clinically certified home deaths was conducted every three years, data were available only for 2017, 2020, and 2023. Data on deaths under LTCI-reimbursed home-visit nursing were derived from the Survey of Long-Term Care Service Facilities and Providers. ¹² Geographic boundary data were based on the Survey of Municipality Areas by the Geospatial Information Authority of Japan, ¹³ and prefectural rates of coroner-investigated deaths were obtained from the Statistics on the Handling of Deceased Bodies. ¹⁴ Because visiting-nursing death data were only available at the prefectural level owing to privacy restrictions, uniform values were applied to all municipalities within each prefecture. This approach captures inter-prefectural differences in care structure but may attenuate within-prefecture variability, likely resulting in conservative estimates. All data and analytical codes supporting this study are available in the Open Science Framework. ¹⁵

The dependent variable was the home-death rate, defined as the proportion of deaths occurring at home among all registered deaths in each municipality. This measure is based on physicians’ death certificates, which record the place of death as “home”, “hospital”, “nursing home”, or other categories. ¹

The main explanatory variables captured indicators of home-based end-of-life care obtained from national surveys of clinics and home-visit nursing agencies. All variables were based on the number of home deaths reported in September of each survey year and were multiplied by 12 to approximate the annual totals.

The following four variables were used: 1.

Home deaths certified by home-care support clinics: Deaths at home for which the death certificate was issued by a designated home-care support clinic expressed as the proportion of all deaths in the municipality.

Clinic-related indicators were standardised by the total deaths in each municipality, and nursing-related indicators were standardised by the total deaths in each prefecture to adjust for population size and reporting coverage.

The following control variables were included to consider regional healthcare resources and demographic characteristics: 1.

Number of clinics (log-transformed), representing the availability of primary medical institutions.

The numbers of clinics and home-visit nursing agencies were standardised per 10,000 population to consider population size differences across municipalities.

We employed a multilevel fixed-effects panel regression model using the panel ordinary least squares estimator, incorporating municipality and year fixed effects to control for unobserved time-invariant characteristics and nationwide temporal trends. ¹⁶ Because indicators of home-visit nursing deaths were available only at the prefectural level, the model accounted for a hierarchical data structure, with municipalities nested within prefectures. Cluster-robust standard errors are calculated at the municipal level.

Three nested models were estimated: 1.

Model 1 included only indicators of home deaths certified by clinics (both home-care support and general clinics).

To assess multicollinearity among explanatory variables, variance inflation factors (VIFs) were calculated for Model 3, including all covariates. A VIF exceeding 10 was considered indicative of serious multicollinearity, whereas values below 5 were considered acceptable. Finally, we examined the correlation between residuals from Model 3 (2023 data only) and the prefectural rate of coroner-investigated deaths (per total deaths in 2023) to explore whether the home-death rate reflects deaths without medical involvement, such as solitary deaths.

All analyses were performed in Python (version 3.12.2) using the linearmodels package (version 6.1). Because all data were aggregated and publicly available, ethical approval and informed consent were not required for this study.

The municipal home-death rate partly reflects activities of community-based medical and nursing services, particularly general clinics and medical-insurance-based visiting nursing, which likely capture primary-care contributions to end-of-life care. ^{17, 18} In contrast, the association between home deaths certified by home-care support clinics and the overall home-death rate became nonsignificant post-adjustment. This may stem from the fact that the formal designation of “home-care support clinics” does not necessarily correspond to the actual provision of palliative or home-visit services. As of 2023, Japan had 104,894 general clinics and 14,514 designated home-care support clinics, indicating that the latter accounted for approximately 14% of all clinics. Given their relatively small number and uneven regional distribution, the statistical influence of designated home-care support clinics on municipal home-death rates may be inherently limited compared to that of general clinics. ¹⁹ Furthermore, although home-care support clinics are certified by regulations, their levels of home-visit engagement vary widely. ¹⁸ Conversely, many general clinics, regardless of designation, actively provide home medical care and play a central role in end-of-life support. ¹⁸ Fukui et al. along with Ikezaki and Ikegami demonstrated that the involvement of attending or visiting physicians, rather than institutional status, was a key predictor of home deaths. ^{20, 21} Together, these findings suggest that the home-death rate reflects the functional strength of community-based clinical activities more than formal institutional categories.

The stronger association observed for deaths under medical-insurance-reimbursed visiting nursing than for those under LTCI-reimbursed visiting nursing is consistent with Japan’s dual insurance structure. LTCI mainly covers daily assistance for older adults with chronic support needs; medical insurance applies when continuous medical supervision or terminal-phase care is required. ²² In practice, patients often transition from LTCI to medical insurance once the end-of-life management intensifies. This structural distinction plausibly explains why medical-insurance-reimbursed nursing deaths were more closely aligned with overall home-death rates.

Despite modest effect sizes, the home-death rate only partially represented medically attended deaths. However, the lack of correlation with medicolegal investigation rates suggests that non-medical deaths do not substantially distort the indicator. Thus, the home-death rate retains limited but meaningful interpretability as a proxy for medically attended end-of-life care.

However, conceptually, the place of death alone cannot capture care quality. Dying at home does not always imply better symptom control, communication, or satisfaction. ²³ Rather than serving as a direct quality metric, home-death rate may be better understood as a contextual indicator of community-based end-of-life care capacity that reflects the accessibility and continuity of home medical and nursing services.

This study had several limitations. First, numbers of home deaths by clinics and home-visit nursing agencies were derived from September reports and multiplied by 12 to obtain approximate annual totals. Although this follows national reporting standards, seasonal variations can introduce bias. Second, data on home-visit nursing deaths and medicolegal investigations were only available at the prefectural level, whereas data on other variables were available at the municipal level, potentially reducing within-prefecture variability and attenuating the associations. Third, while fixed-effects models were adjusted for time-invariant heterogeneity and national trends, unobserved confounders, such as regional differences in reporting accuracy or death certification practices, cannot be ruled out; however, these are unlikely to explain the main observed patterns. Despite these limitations, the persistence of key associations, particularly for deaths under medical-insurance-reimbursed visiting nursing and those certified by general clinics, supports the robustness of our main findings.

This study provides empirical evidence that Japan’s home-death rate, derived from physicians’ death certificates, partially reflects medically attended home deaths rather than non-medical or solitary deaths. These findings suggest that this indicator serves as a population-level contextual measure of community-based end-of-life care capacity and captures the extent to which medical and nursing services are integrated into local systems. However, because the place of death alone cannot represent care quality or patient experience, the home-death rate should be interpreted with caution and complemented with other indicators that reflect care involvement, coordination, and symptom management. Future research using individual-level or record-linked data is essential to validate these relationships and to develop more accurate and patient-centred measures for evaluating end-of-life care in the community.