Association Between Duration of Usual Source of Care and Patient-reported Primary Care Quality in Japan: A Nationwide Cross-sectional Study

Target Population and Analytic Sample

Using sex- and age-stratified random sampling, we invited 1112 adults aged 20 to 74 years (the PCPCM validation range) via automated e-mail from the survey company’s secure platform; investigators had no direct contact with potential participants. ⁹ Responses were submitted through anonymized online forms, and the research team received de-identified data only. Respondents were first asked, “Do you have a medical facility that you usually visit when ill or for health advice?” Those answering yes (USC = 1; n = 423) proceeded to the PCPCM and comprised the analytic sample for outcome models. Those answering no (USC = 0; n = 377) did not receive PCPCM items and were excluded from outcome analyses, consistent with recommended gating in non-registration systems.^8,9 Within the analytic sample (USC = 1), there were no missing data on exposure or outcome.

Exposure: Duration-Based Continuity

Continuity was measured as years since first attending the usual institution and years since first consulting the usual physician, reported by USC = 1 respondents. Relationship duration has been used as a patient-centered proxy for interpersonal continuity with supportive associations to preventive care, utilization, and mortality.^29-32 When respondents had more than 1 potential USC, instructions specified the facility used most often for general health problems; ties were resolved by selecting the most recently visited facility. To accommodate 0 years (<1 year) and reduce right-skew, we modeled ln(years+1) (“log1p”) as a continuous predictor, as illustrated in Figure 1.

OPEN IN VIEWER

Figure 1.

Distributions of years since starting to visit: (a) a regular medical institution (“Years with regular institution”), (b) natural log-transformed “years with regular institution,” (c) a regular physician (“Years with regular physician”), and (d) natural log-transformed “years with regular physician.”

Outcome: PX

PX was assessed using the validated Japanese PCPCM (11 domains; each 1-4). ⁹ Consistent with prior PCPCM studies,^13,14 the composite score used in analyses was the mean of the 11 items (range: 1-4), not the sum; thus, values in figures and tables represent average item scores. The eleven items cover accessibility, comprehensiveness, integration, coordination, relationship, continuity, advocacy, family and community context, goal-orientation, and health promotion. The continuity item (“My doctor and I have been through a lot together”) does not include relationship length. The Japanese and original versions of all items are provided in Supplemental Tables S1 and S2.

Covariates

Based on prior associations with continuity and PX, we adjusted for age, sex, education, household income, marital status, occupation, self-rated health, multimorbidity (count of 20 chronic conditions), USC type (clinic vs hospital), and rurality using the Rurality Index for Japan (RIJ; 1-100; higher = rural/remote).^10,34-37 Age was grouped into 5-year bands to align with sampling strata and reduce functional form misspecification. Multimorbidity was defined as the count of 20 self-reported chronic conditions (eg, hypertension, diabetes, dyslipidemia, chronic lung disease, mental illness), and entered as a categorical covariate (0, 1, and ≥2 conditions) in the main models to reflect overall burden while avoiding over-interpretation of specific disease combinations sharing the same count.

Sample Size Planning

From prior studies we anticipated a USC prevalence of ~55% and a PCPCM standard deviation of ~0.6.^9,14,15 Power calculations targeting a small effect for the continuity term (Cohen’s f ²  = 0.02) in multiple regression with ~12 to 14 covariates (α = .05, 1 − β = .80) indicated the need for ≈380 to 420 USC respondents, using the standard sample-size formula for multiple linear regression based on Cohen’s f ² . Assuming ~70% response and 55% USC prevalence, we issued ≈1100 invitations to reach this target while satisfying the ≥10 observations-per-parameter heuristic.

Statistical Analysis

Categorical variables were summarized as counts and percentages; continuous variables as medians and interquartile ranges. We first fitted multivariable linear regression models of PCPCM on ln(years+1) continuity, separately for institution- and physician-level exposures, adjusting for covariates: age, sex, education, household income, marital status, occupation, self-rated health, multimorbidity, USC type, and RIJ. In the primary specification, multimorbidity was included as a categorical variable (0, 1, and ≥2 conditions). As a sensitivity analysis, we re-estimated all models with multimorbidity entered as a continuous count variable (0-20 conditions) to assess the robustness of continuity estimates to this modelling choice. The PCPCM mean score was treated as a continuous outcome. Model assumptions (linearity, homoscedasticity, normality of residuals) were checked using diagnostic plots, and variance inflation factors (VIFs) were examined (threshold 2.5).^38,39 Two-sided P < .05 denoted statistical significance. Analyses used Stata/SE 17.0.

Stratified Analysis

To examine potential effect modification and contextual robustness, we stratified the main models by (a) USC type (clinic vs hospital) and (b) rurality (RIJ 1-20 vs RIJ 21-100, corresponding to urban vs rural/remote). Facility type and geography are known to shape primary care delivery patterns, resource availability, and patient-provider relationships in Japan.^10,11

Sensitivity Analyses

Sensitivity analyses addressed potential construct overlap, alternative exposure specifications, nonlinearity, heteroskedasticity and outliers, missing data, and a negative-control outcome. To address potential construct overlap, we recomputed the PCPCM composite after excluding the continuity item (No. 6) and repeated the primary models. Details of all sensitivity-analysis specifications, including categorical exposures, trimming/winsorization, spline modeling, and missing-data handling, are provided in the Supplemental Material. We also repeated all models with multimorbidity entered as a continuous count variable (0-20 chronic conditions) instead of categories; continuity coefficients were very similar, indicating robustness to this modelling choice. Unless otherwise noted, models used heteroskedasticity-consistent covariance estimator type 3 (HC3) robust standard errors.

Ethical Approval and Consent to Participate

The study was performed in accordance with the Declaration of Helsinki and approved by the Ethics Committee of Yokohama City University (approval number: F220900057). All participants provided electronic informed consent and were allowed to terminate the survey at any time without penalty.

System Context and International Comparison

Regarding the scope and utility of the PCPCM, validation studies in the United States and Canada have established it as a robust measure of high-value primary care functions,^13,15 and cross-national analyses across 35 OECD countries have linked national PCPCM scores to health system performance. ¹⁴ Our findings suggest that the PCPCM also captures the benefits of continuity in Japan, supporting its use beyond the Western cultural and health system contexts in which it was developed.

In registration systems (eg, the UK), list-based assignment and gatekeeping structurally protect interpersonal continuity, which is associated with fewer ambulatory care-sensitive admissions and lower mortality.^20,40 In Canada, administrative data separating physician- from clinic-level continuity show stronger reductions in acute care use for the former. ²³ Our results indicate that even without registration, relationship duration tracks higher PCPCM in Japan, implying that policy tools which approximate registration—such as formal designation of a family physician responsible for first-contact care, preventive services, and longitudinal panel management, together with stable provider assignment in clinics and hospital outpatient departments—could enhance PX under free access. ¹¹

In Japan’s free-access system, patients can self-select both clinics and hospital outpatient departments as their USC. This self-selection may encourage those with positive experiences to maintain long-term relationships and affords patients greater autonomy over whom they regard as their usual doctor or facility. However, self-selection is unlikely to operate uniformly across demographic groups: for example, rural residents and individuals with multimorbidity may place greater value on stable relationships and therefore be more likely to identify a USC than healthy urban adults. ⁴¹ These structural features of the Japanese system may thus concentrate continuity among higher-need groups and should be considered when interpreting continuity-PX gradients and designing continuity-enhancing policies.

Facility Type and Geography

The continuity-PX gradient differed by facility and rurality: clinics in urban areas and hospitals in rural/remote areas showed the steepest associations. This pattern is consistent with rural populations’ higher valuation of interpersonal relationships ⁴¹ and with the reality that regional hospitals often serve as primary contact points in remote areas. In urban settings, clinics typically provide person-centered, longitudinal care with smaller panels and appointment systems, so physician continuity may be easier to maintain and more salient for PX. In contrast, in rural and remote regions of Japan, hospital outpatient departments often function as de facto primary care clinics with limited provider choice; within this context, sustained relationships with hospital-based physicians can play a similar interpersonal role, explaining the strong continuity-PX gradients we observed. These contextual differences argue against one-size-fits-all approaches and suggest facility-specific strategies to bolster continuity.

Interpretation Boundaries and Covariates

Our covariates served to control confounding; we deliberately avoid interpreting their coefficients causally to prevent over-reading adjustment variables as determinants of PX. Estimates for sex, age, income, occupation, and health status should not be over-read. For multimorbidity, we chose a simple categorical specification (0, 1, and ≥2 conditions) because very different disease combinations can share the same count, and the marginal impact of “1 additional condition” is unlikely to be strictly linear. Sensitivity analyses modelling multimorbidity as a continuous count yielded similar continuity estimates, suggesting that our conclusions are not driven by this coding choice.

Conceptual Overlap

The PCPCM includes a domain labeled continuity/relationship, which is conceptually adjacent to our exposure. This raises the possibility of partial construct overlap inflating the association. We therefore caution readers in interpreting the magnitude of effects. However, our sensitivity analysis excluding the continuity item yielded estimates that were substantively similar to the primary results (as detailed in the Results section). These findings suggest that construct overlap does not fully explain the observed gradients, although some inflation of effect sizes remains possible.

Causal Direction and Bidirectionality

Cross-sectional design limits causal inference. Better PX could promote continuity by reducing switching, while continuity could improve PX by enabling accumulated knowledge and coordination. Unmeasured preferences and local supply may shape both. Longitudinal designs with time-varying exposures, or interventional approaches (eg, dedicated attending-physician models or fixed provider assignment), are needed to disentangle directionality.

Implications

Interventions that preserve and extend continuity are expected to support better PX and, according to prior longitudinal and interventional studies, may also contribute to lower utilization and costs.^16-23 Given our cross-sectional design, these possibilities should be interpreted as hypotheses for future research rather than causal conclusions. In Japan, strengthening the Family Physician Function Reporting System and piloting stable provider assignment in hospital outpatient departments—especially in rural regions—could operationalize continuity at scale. For clinics, this could include explicitly assigning each patient a regular physician, documenting this in medical records, and aligning appointment systems and after-hours coverage to maintain that relationship. For hospitals, continuity could be supported by small, stable outpatient teams ⁴² with named lead physicians, structured handovers when staffing changes, and communication to patients about which clinician is responsible for their ongoing care. Telemedicine and panel management may help sustain continuity where workforce is scarce.

Limitations

This study has several limitations. First, as noted above, its cross-sectional design and the potential bidirectionality of the continuity-PX association limit causal inference. Second, because PCPCM items were only administered to respondents who reported a USC, and USC holders differed from non-holders on several sociodemographic and clinical characteristics (Table 1), our analytic sample over-represents individuals who have already established a usual place of care. This likely reflects selective uptake of continuity by people with greater healthcare needs and limits the generalizability of our findings beyond adults with a USC. Third, regarding our sampling method, although we used age- and sex-stratified quotas, our web-based panel is not a probability sample. Relative to face-to-face national surveys, web-based sampling offers rapid recruitment and lower cost but is more vulnerable to selection bias related to internet access and digital literacy. Unlike administrative-data studies, our design captures patient-reported PX but cannot provide population-level estimates. Consequently, we may underrepresent digitally marginalized groups, which could lead to overestimation of continuity and PX among USC holders. Fourth, self-reported relationship duration may be affected by recall error, especially in hospitals with rotating staff. ⁴³ Such error is likely to be nondifferential with respect to PCPCM scores and would therefore tend to attenuate continuity-PX associations rather than create spurious ones; however, some respondents may over-estimate very long relationships, which could modestly inflate effect estimates at the upper end of the duration distribution. Finally, as discussed above, potential conceptual overlap between our continuity exposure and the PCPCM continuity/relationship domain and our use of a simple multimorbidity count (0, 1, and ≥2 conditions) may partially inflate associations, though sensitivity analyses excluding the continuity item and modelling multimorbidity as a continuous count yielded substantively similar estimates. These limitations should be considered when interpreting the magnitude and generalizability of our findings.