What Builds Patient Trust in Doctors? A Japanese Survey Finds Few Easy Answers

Trust between a patient and a physician is often described as the invisible foundation of medicine. It shapes whether patients follow prescriptions, return for follow-up visits, disclose sensitive symptoms, and accept difficult diagnoses. Yet despite decades of research, the question of what actually produces that trust remains surprisingly open. A new study from Kyoto University researchers, published in BMC Health Services Research, set out to test whether some of the most commonly cited candidates—waiting times, the use of local dialect, and a patient’s attitudes toward the physician’s medical school—can genuinely explain how much a patient trusts their doctor. The answer, in short, is that these measurable factors explain very little, a finding that may reshape how health services researchers think about the origins of trust in clinical care.

The research team, led by Masashi Ikuno and colleagues at Kyoto University’s Center for Medical Education and Internationalization, conducted a cross-sectional web panel survey of adults across Japan who reported having a usual physician. The survey used a non-probability commercial web panel with nationwide geographic coverage, and 2,173 respondents completed the questionnaire. The outcome measure was Trust100, a score ranging from 0 to 100 derived from five items based on the Japanese version of the Abbreviated Wake Forest Interpersonal Trust in Physician Scale, a validated instrument widely used in international research on physician–patient relationships. Across the full sample, the mean Trust100 score was 66.8, suggesting a moderately high baseline level of trust among Japanese patients with an established regular physician.

The study’s design was notable for its deliberate focus on three conceptually distinct domains of potential trust determinants. The first was the care-process domain, represented by waiting time before consultations. The second was a linguistic domain, capturing whether the physician and respondent shared use of a local dialect—an intriguingly Japanese hypothesis, given the country’s rich regional linguistic variation and the possibility that dialect concordance signals shared identity and warmth. The third domain concerned attitudes toward and awareness of the physician’s medical school, testing whether the perceived prestige or familiarity of a doctor’s training institution colors patient confidence. Each domain was analyzed with adjusted linear regression models using heteroskedasticity-robust covariance estimates, and the three global domain-level tests were corrected for multiple comparisons using the Holm procedure, a conservative approach that guards against false-positive findings when several hypotheses are tested simultaneously.

The results were striking in their restraint. Waiting time showed a weak negative association with trust in the expected direction—longer waits, lower trust—but after Holm correction the adjusted association did not remain statistically significant, with an adjusted p-value of 0.073 and a partial R-squared of just 0.0043. In practical terms, waiting time accounted for less than half of one percent of the variation in trust scores. The dialect analysis was even more definitive: in a pooled analysis adjusted for region, no clear association emerged between physician–respondent dialect-use combinations and trust (p = 0.783; partial R-squared = 0.0005), and the researchers found no statistical evidence that the relationship varied across Japan’s regions. Whatever intuitive appeal the dialect hypothesis holds, the data did not support it.

The medical-school domain fared somewhat better but still modestly. The global test for attitudes toward and awareness of the physician’s medical school remained statistically significant after Holm correction, making it the only domain to survive multiplicity adjustment. Even so, its explanatory contribution was limited, with a partial R-squared of 0.022—roughly two percent of the variance in trust. Moreover, the pattern of association was non-monotonic, meaning that trust did not rise or fall in a simple, orderly way with increasing favorability of attitudes toward the medical school, and the researchers observed no interaction effects. This suggests that while perceptions of a physician’s educational background may carry some signal, the relationship is complicated and cannot be reduced to a simple prestige effect.

Beyond the domain-specific analyses, the team undertook an exploratory prediction exercise, asking a more ambitious question: how well can patient trust be predicted at all from questionnaire data? They compared four modeling approaches—mean prediction as a baseline, ordinary least-squares regression, ridge regression, and Extra Trees, an ensemble machine-learning method—using repeated nested cross-validation, a rigorous framework that separates model tuning from performance estimation and reduces the risk of optimistic overfitting. Permutation importance, a technique that measures how much a model’s accuracy degrades when a variable’s values are randomly shuffled, identified consultation time as the variable with the highest mean importance, prompting a post hoc examination of that factor.

That post hoc comparison produced one of the study’s most eye-catching descriptive findings. Respondents who reported consultation times of at least three minutes had a mean Trust100 score 7.69 points higher than those reporting less than three minutes, with a 95 percent confidence interval of 6.10 to 9.27. In a field where effects are often small and noisy, a nearly eight-point gap on a 100-point trust scale is substantial. The finding aligns with a long line of international research linking longer, less rushed consultations to stronger therapeutic relationships, and it resonates with ongoing debates in Japan and elsewhere about extreme time pressures in primary care. The authors are careful, however, to frame this as a post hoc descriptive comparison rather than a pre-specified causal test, and self-reported consultation time is subject to recall and rounding error.

The study’s overall conclusion is candid: patient trust was only weakly explained and predicted by the measured questionnaire variables. Even the best-performing predictive models could not extract much signal from the available data. The researchers offer two complementary interpretations. First, trust may be driven largely by unmeasured relational or contextual factors—accumulated experiences of being listened to, moments of competence or compassion, continuity of care over years, community reputation, or the subtle nonverbal choreography of a consultation that no questionnaire item captures. Second, measurement limitations may play a role; a five-item abbreviated scale and single-shot survey items may simply be too coarse to detect the associations that exist. Both explanations carry implications: the first suggests trust research should look deeper into relational dynamics, while the second calls for richer, more sensitive instruments.

Methodologically, the paper is a case study in statistical discipline. The use of heteroskedasticity-robust standard errors protects inference from unequal variance across subgroups, and the Holm correction across the three domain-level tests prevents the all-too-common scenario in survey research where one of many tested associations is trumpeted as significant by chance. The exploratory prediction component, with its nested cross-validation design, avoids the leakage problems that plague much of the machine-learning literature in health services research. The authors also disclose that the study was not prospectively registered, although a dated study protocol and questionnaire had been prepared before data collection—a transparency note that contextualizes the post hoc elements of the analysis. The survey was implemented with support from INTAGE Research Inc., and the work was funded by Japan Society for the Promotion of Science KAKENHI grants 24K13315 and 25K13411, with the funder having no role in the design, analysis, or writing. The study was approved by the Ethics Committee for Medical Research at Kyoto University Graduate School and Faculty of Medicine, and all participants provided electronic informed consent.

For clinicians and health policymakers, the takeaway is both humbling and clarifying. Trust in physicians appears to be a deep, relational phenomenon that resists reduction to waiting-room logistics, linguistic mirroring, or institutional pedigree. If anything can move the needle, the consultation-time signal hints that giving patients genuine time with their doctor may matter more than any of the structural variables studied here. As health systems worldwide grapple with physician shortages and compressed appointments, this Japanese nationwide survey offers a data-driven reminder that the minutes spent face to face may be among the most trust-producing resources medicine has—while cautioning that the true architecture of patient trust still lies largely beyond the reach of the questionnaires we currently know how to ask.

Subject of Research: Patient trust in physicians and its measurable determinants in a Japanese nationwide web panel survey

Article Title: Factors associated with patient trust in physicians: domain-specific association analyses and exploratory prediction in a cross-sectional web panel survey in Japan

Article References: Ikuno, M., Wakabayashi, T., Fukuhisa, A., Onizuka, D., Tokinobu, A., Miyoshi, T., Watari, T., & Kataoka, H. (2026). Factors associated with patient trust in physicians: domain-specific association analyses and exploratory prediction in a cross-sectional web panel survey in Japan. BMC Health Services Research. https://doi.org/10.1186/s12913-026-15698-2

Image Credits: AI Generated

DOI: 10.1186/s12913-026-15698-2

Keywords: patient trust, physician-patient relationship, Japan, web panel survey, waiting time, consultation time, local dialect, medical school, machine learning, cross-validation, health services research, Kyoto University