Psychosocial determinants of antibiotic stewardship decisions among Chinese clinicians across hospital tiers: a longitudinal study Authors: Xiao Liu, Le Han, Peng Xian, Yue Chang, Ying Liu, Lijun Ma, Jie Ren, Yunshuang Liu, Xiaosong Wang, Yujuan Long, Li Ren, Hui Tang, Qi Liu, Ling Liang, Fang Jin, Guyan Wang Published: 2026 ABSTRACT AbstractInappropriate antibiotic use contributes to antimicrobial resistance, yet much stewardship evidence is cross-sectional and intention-based. We conducted a two-wave longitudinal survey of clinicians in Chinese hospitals approximately 11 months apart (baseline September–October 2024; follow-up August–September 2025) to examine whether baseline psychosocial factors predicted follow-up stewardship-aligned responses to four clinical vignettes and whether associations varied by hospital level. Among 1,420 clinicians with linked responses, higher follow-up vignette scores were independently associated with stronger reasoned-action beliefs, fewer perceived barriers, greater self-efficacy, stronger knowledge and skills, greater social support, and higher perceived patient-centred benefits in both structural equation modelling and robust regression analyses. TPB-related beliefs were not independently associated after mutual adjustment. Hospital level moderated the association between reasoned-action beliefs and follow-up vignette responses, with stronger associations in tertiary than primary hospitals. Stewardship interventions may therefore need to address capability, confidence, supportive environments, and institutional context in addition to attitudes.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-66491-4.Keywords: Antimicrobial stewardship, Antibiotic prescribing, Longitudinal study, Clinicians, Psychosocial determinants, Clinical vignettesSubject terms: Diseases, Health care, Medical research, Psychology, Psychology Introduction Inappropriate antibiotic use remains a major driver of antimicrobial resistance (AMR) and preventable patient harm1,2. Although antimicrobial stewardship has been widely promoted through international guidance and national programmes, achieving consistent improvement in routine prescribing remains challenging3,4. Chinese studies have also shown variation in stewardship structures and antibiotic appropriateness across institutions and care settings4–6. Our previous cross-sectional work suggested that psychosocial factors may help explain antimicrobial prescribing decisions. In one study, social support influenced prescribing through self-efficacy, knowledge and skills, and health beliefs7. In another, theory-informed psychological and behavioural measures strongly predicted clinicians’ intention to use antibiotics appropriately8. Together, these findings suggest that attitudes, capabilities, and contextual support may shape stewardship behaviour, but longitudinal evidence remains limited. Behavioural science provides a useful framework for identifying candidate determinants without prespecifying a single mechanism. Planned-behaviour models emphasise attitudes, norms, and perceived behavioural control9, the Health Belief Model highlights perceived barriers10, and self-efficacy theory emphasises confidence in performing a target behaviour11. Because intention does not automatically translate into action12,13, vignette-based assessments can provide a more behaviour-proximal outcome than intention alone14. Diagnostic infrastructure and microbiology support may also influence whether stewardship-consistent actions are feasible in practice15,16. Prior work has also shown that theory-based psychosocial models can explain antibiotic prescribing decisions in China and other settings17,18. China’s hierarchical hospital system provides a particularly relevant setting for this question because staffing, diagnostic capacity, case complexity, and organisational stewardship support differ across primary, secondary, and tertiary hospitals4,6,19. These tier-related differences may moderate psychosocial–behavioural associations because access to diagnostics, specialist support, and formal stewardship systems may influence the extent to which clinicians’ beliefs, confidence, and skills are reflected in stewardship-aligned decisions. We therefore conducted a two-wave longitudinal study to examine whether baseline psychosocial constructs predicted follow-up vignette-based stewardship responses among Chinese clinicians and whether these associations varied by hospital level. Methods Study design and reporting We conducted a two-wave longitudinal survey with an approximately 11-month interval between measurements. Wave 1 data were collected from September to October 2024 and Wave 2 data from August to September 2025. The study focused on clinicians working in Chinese hospital settings. The primary outcome was a follow-up vignette-based score reflecting stewardship-aligned responses to four common antibiotic-prescribing scenarios (BI_W2). Participants and eligibility Clinicians were eligible if they were practising in clinical settings, completed the baseline questionnaire, and could be linked to the follow-up questionnaire using a unique non-identifiable code. Participants without matched responses at both waves were excluded from paired longitudinal analyses. The final paired analytic sample included 1,420 clinicians. Data collection procedures and confidentiality Data were collected using a self-administered online questionnaire. No personally identifiable information (for example, name, phone number, national identification number, or detailed workplace identifiers) was collected. Each participant generated or was assigned a unique anonymous code to link Wave 1 and Wave 2 responses while maintaining anonymity. Data were stored in password-protected files accessible only to the research team. Theoretical framework We used an integrated psychosocial framework drawing on the Theory of Planned Behaviour, the Health Belief Model, self-efficacy theory, and social support theory. To remain consistent with the questionnaire used in this research programme, one reasoned-action domain is labelled RAT throughout the manuscript. We also included knowledge and skills (KS) and perceived patient-centred benefits (PB) of rational antibiotic use. Hospital level (primary, secondary, tertiary) was treated as a contextual moderator because resource availability and stewardship infrastructure differ across tiers. Study hypotheses Based on the framework and prior evidence, we tested the following hypotheses: H1: Baseline psychosocial constructs (RAT, TPB-related beliefs, barrier-focused health beliefs, self-efficacy, knowledge and skills, social support, and perceived patient-centred benefits) would be independently associated with BI_W2. H2: The associations between baseline psychosocial constructs and BI_W2 would vary by hospital level. Measures Baseline demographic variables included age, gender, education level, years of practice, professional position, and hospital level. Prespecified model covariates were baseline vignette-based stewardship responses (BI_W1), age group, gender, highest educational attainment, years of clinical practice, and hospital level. Hospital level was categorised as primary/community, secondary, or tertiary. Psychosocial constructs (Wave 1) At Wave 1, psychosocial constructs were measured using multi-item 5-point Likert scales (1 = strongly disagree to 5 = strongly agree). The domains were reasoned-action beliefs (RAT), TPB-related beliefs, barrier-focused health beliefs (reverse-scored so higher values indicated fewer perceived barriers), self-efficacy (SET), knowledge and skills (KS), social support (SS), and perceived patient-centred benefits (PB). Behavioural outcome (Wave 2) The primary outcome, vignette-based stewardship responses at Wave 2 (BI_W2), was assessed using four short clinical vignettes reflecting common antibiotic-prescribing scenarios. Participants rated the likelihood of taking stewardship-aligned actions on the same 5-point scale, and BI_W2 was computed as the mean score across vignette items, with higher scores indicating more appropriate responses. The same four vignettes were administered at Wave 1, and their mean score was denoted BI_W1 and included as a baseline covariate. Measurement model evaluation Internal consistency was evaluated using Cronbach’s alpha20. Convergent validity was assessed using composite reliability (CR) and average variance extracted (AVE), using conventional benchmarks of CR ≥ 0.70 and AVE ≥ 0.5021,22. Longitudinal structural equation modelling (SEM) To test H1, we estimated a latent-variable SEM in which Wave 1 psychosocial constructs predicted BI_W2. The SEM adjusted for BI_W1, age group, gender, highest educational attainment, years of clinical practice, and hospital level. Model fit was evaluated using the comparative fit index (CFI), Tucker–Lewis index (TLI), root mean square error of approximation (RMSEA), and standardised root mean square residual (SRMR)23. Statistical analysis As a complementary analysis, we fitted robust linear regression models with HC3 standard errors. The seven psychosocial domains were included simultaneously, and models adjusted for BI_W1, age group, gender, highest educational attainment, years of clinical practice, and hospital level. HC3 heteroscedasticity-consistent standard errors were used to reduce sensitivity to non-constant residual variance and high-leverage observations. In the main-effects models, hospital level followed the questionnaire coding (1 = primary/community, 2 = secondary, and 3 = tertiary). For moderation analyses, hospital level was represented by indicator variables with primary/community hospitals as the reference group. Separate construct-specific interaction models were fitted; within each model, the secondary-hospital × construct and tertiary-hospital × construct terms were entered simultaneously and evaluated using a two-degree-of-freedom joint Wald test. Analyses were implemented in Python using semopy for SEM and statsmodels for robust OLS. Results Baseline demographic characteristics A total of 1,500 clinicians completed the baseline survey, and 1,420 provided linked responses at follow-up. Most participants were aged 26–35 years (36.5%) or 36–45 years (27.3%), and 62.9% were female. Educational attainment was high: 38.4% held a master’s degree and 18.9% held a doctorate, whereas 27.4% held a bachelor’s degree and 14.7% reported a college diploma. More than half (50.7%) had fewer than five years of practice. Participants worked in primary/community hospitals (39.5%), secondary hospitals (36.7%), and tertiary hospitals (23.8%), providing substantial variation in practice context (Table 1). Compared with retained participants, those lost to follow-up appeared broadly similar on most baseline variables, although hospital level showed a noticeable imbalance, with a higher proportion of tertiary-hospital clinicians among non-retained participants (Supplementary Table S1). Table 1Baseline demographic characteristics of the study participants (N = 1,500).Characteristicn (%)Age group (years) 18–25375 (25.0) 26–35548 (36.5) 36–45409 (27.3) 45–55155 (10.3) > 5513 (0.9)Gender Male556 (37.1) Female944 (62.9)Educational attainment College diploma221 (14.7) Bachelor’s degree411 (27.4) Master’s degree576 (38.4) Doctoral degree284 (18.9) Postdoctoral training8 (0.5)Years of clinical practice < 5 years761 (50.7) 5–15 years347 (23.1) 16–25 years260 (17.3) > 25 years132 (8.8)Hospital level Primary/community hospital593 (39.5) Secondary hospital550 (36.7) Tertiary hospital357 (23.8) Baseline demographic characteristics of the study participants (N = 1,500). Psychometric properties and SEM results Table 2 summarises the measurement properties of the study constructs. Wave 1 psychosocial scales showed acceptable internal consistency, with Cronbach’s α ranging from 0.788 to 0.82020. The follow-up vignette-based outcome (BI_W2) had α = 0.698, which is acceptable for a short four-item scenario measure. Standardised factor loadings were 0.680–0.753 for Wave 1 constructs and 0.592–0.624 for BI_W2. Composite reliability met conventional criteria across constructs (CR = 0.698–0.822)21. AVE exceeded 0.50 for TPB, HBM, SET, KS, SS, and PB, whereas RAT (0.475) and BI_W2 (0.366) were lower, suggesting more modest shared variance for those indicators. Table 2Measurement properties of study constructs (Wave 1) and outcome (Wave 2).ConstructItems (n)Cronbach’s αStd. loading range (min–max)CRAVERAT40.7880.687–0.6920.7840.475TPB40.8170.715–0.7340.8190.530HBM (barriers)40.8160.707–0.7390.8150.524SET40.8200.707–0.7530.8220.537KS40.8020.680–0.7280.8020.503SS40.8070.710–0.7260.8080.513PB (perceived benefit)40.8120.696–0.7510.8110.517BI_W2 (scenario-based)40.6980.592–0.6240.6980.366RAT = reasoned-action beliefs (questionnaire construct label); TPB = TPB-related beliefs; HBM (barriers) = barrier-focused health beliefs (reverse-scored so higher values indicate fewer perceived barriers); SET = self-efficacy; KS = knowledge and skills; SS = social support; PB = perceived patient-centred benefits of rational antibiotic use; BI_W2 = follow-up vignette-based stewardship responses; CR = composite reliability; AVE = average variance extracted. Measurement properties of study constructs (Wave 1) and outcome (Wave 2). RAT = reasoned-action beliefs (questionnaire construct label); TPB = TPB-related beliefs; HBM (barriers) = barrier-focused health beliefs (reverse-scored so higher values indicate fewer perceived barriers); SET = self-efficacy; KS = knowledge and skills; SS = social support; PB = perceived patient-centred benefits of rational antibiotic use; BI_W2 = follow-up vignette-based stewardship responses; CR = composite reliability; AVE = average variance extracted. The SEM showed excellent fit (CFI = 0.999, TLI = 0.999, RMSEA = 0.003, SRMR = 0.010)23. In the mutually adjusted model, higher BI_W2 was associated with stronger reasoned-action beliefs (β = 0.135, p = 0.001), fewer perceived barriers within the HBM domain (β = 0.133, p = 0.002), greater self-efficacy (β = 0.116, p = 0.022), stronger knowledge and skills (β = 0.148, p < 0.001), stronger social support (β = 0.081, p = 0.048), and higher perceived patient-centred benefits (β = 0.148, p < 0.001) (Table 3; Fig. 1). TPB-related beliefs were not independently associated with BI_W2 after mutual adjustment (β = 0.040, p = 0.374). Table 3Longitudinal SEM paths predicting follow-up vignette-based stewardship responses (BI_W2).PathStandardized coefficient (β)p-valueRAT → BI_W20.1350.001TPB → BI_W20.0400.374HBM (barriers) → BI_W20.1330.002SET → BI_W20.1160.022KS → BI_W20.148< 0.001SS → BI_W20.0810.048PB (perceived benefit) → BI_W20.148< 0.001BI_W2 denotes follow-up vignette-based stewardship responses, computed as the mean of four clinical vignette items (higher scores indicate more appropriate stewardship-aligned responses). Standardized coefficients are shown from the fully specified SEM, which adjusted for BI_W1, age group, gender, highest educational attainment, years of clinical practice, and hospital level. Longitudinal SEM paths predicting follow-up vignette-based stewardship responses (BI_W2). BI_W2 denotes follow-up vignette-based stewardship responses, computed as the mean of four clinical vignette items (higher scores indicate more appropriate stewardship-aligned responses). Standardized coefficients are shown from the fully specified SEM, which adjusted for BI_W1, age group, gender, highest educational attainment, years of clinical practice, and hospital level. Structural equation model predicting follow-up vignette-based stewardship responses (BI_W2). Structural equation model predicting follow-up vignette-based stewardship responses (BI_W2). Standardized path coefficients (β) are shown. Solid lines indicate statistically significant paths (p < 0.05); dashed lines indicate non-significant paths. The model adjusted for BI_W1, age group, gender, highest educational attainment, years of clinical practice, and hospital level; covariate paths are omitted from the figure for clarity. Model fit indices: CFI = 0.999; TLI = 0.999; RMSEA = 0.003; SRMR = 0.010. Overall, the SEM results point to capability, self-efficacy, supportive context, perceived patient-centred benefit, and lower perceived barriers as the principal prospective correlates of stewardship-aligned vignette responses. Robust OLS regression results for psychosocial determinants of stewardship responses (Table 4) Table 4Robust OLS (HC3) results for baseline predictors of follow-up vignette-based stewardship responses (BI_W2).PredictorβSEz95% CI p RAT0.1300.0324.0100.066 to 0.193< 0.001TPB0.0490.0261.920−0.001 to 0.0990.055HBM (barriers)0.1130.0303.8000.055 to 0.172< 0.001SET0.0820.0263.1500.031 to 0.1330.002KS0.1110.0284.0300.057 to 0.166< 0.001SS0.0750.0302.4800.016 to 0.1340.013PB (perceived benefit)0.1250.0294.3300.068 to 0.181< 0.001Age group−0.0180.025−0.736−0.066 to 0.0300.462Gender0.0200.0230.836−0.026 to 0.0660.403Education−0.0060.015−0.434−0.036 to 0.0230.665Years of practice0.0050.0230.211−0.041 to 0.0500.833Hospital level−0.0060.019−0.337−0.043 to 0.0310.736All seven psychosocial domains were included simultaneously. The model adjusted for BI_W1, age group, gender, highest educational attainment, years of clinical practice, and hospital level. Hospital level followed questionnaire coding (1 = primary/community, 2 = secondary, 3 = tertiary) in the main-effects model. BI_W1 was included as a covariate but is not displayed. Robust OLS (HC3) results for baseline predictors of follow-up vignette-based stewardship responses (BI_W2). All seven psychosocial domains were included simultaneously. The model adjusted for BI_W1, age group, gender, highest educational attainment, years of clinical practice, and hospital level. Hospital level followed questionnaire coding (1 = primary/community, 2 = secondary, 3 = tertiary) in the main-effects model. BI_W1 was included as a covariate but is not displayed. In the fully adjusted robust OLS model, several baseline psychosocial constructs showed independent positive associations with BI_W2 (Table 4). Reasoned-action beliefs were associated with higher BI_W2 (β = 0.130, 95% CI 0.066–0.193, p < 0.001), as were fewer perceived barriers (β = 0.113, 95% CI 0.055–0.172, p < 0.001), self-efficacy (β = 0.082, 95% CI 0.031–0.133, p = 0.002), knowledge and skills (β = 0.111, 95% CI 0.057–0.166, p < 0.001), social support (β = 0.075, 95% CI 0.016–0.134, p = 0.013), and perceived patient-centred benefits (β = 0.125, 95% CI 0.068–0.181, p < 0.001). TPB-related beliefs showed only borderline evidence of association (β = 0.049, 95% CI − 0.001 to 0.099, p = 0.055). Demographic covariates were not statistically significant in the fully adjusted model. Effect modification by hospital level Hospital level significantly moderated the association between reasoned-action beliefs (RAT) and BI_W2 (joint p = 0.007; Supplementary Table S2). The RAT–BI_W2 association was stronger in tertiary hospitals than in primary/community hospitals (interaction b = 0.108, p = 0.002), whereas the difference between secondary and primary/community hospitals showed a weaker, non-significant trend (interaction b = 0.048, p = 0.090). For other psychosocial constructs, there was limited evidence of effect modification by hospital level, including TPB-related beliefs (joint p = 0.887). Discussion This two-wave study provides prospective evidence that stewardship-aligned vignette responses among Chinese clinicians are associated more strongly with capability- and enactment-related factors than with general TPB-related beliefs. The pattern was consistent across SEM and robust regression analyses, which strengthens confidence that the main findings do not depend on a single modelling strategy. The results also extend our earlier cross-sectional work by showing that social support, knowledge and skills, and related psychosocial domains are not only correlates of intention but predictors of later behaviour-proximal responses7,8. Knowledge and skills emerged as one of the strongest predictors of BI_W2. This is consistent with evidence that inappropriate prescribing remains common when diagnostic uncertainty, limited training, or weak stewardship systems constrain clinical decision-making5,19,24. The association of self-efficacy with BI_W2 suggests that clinicians need not only technical knowledge but also confidence to apply stewardship principles under routine pressure11. Perceived patient-centred benefits, social support, and lower perceived barriers were also independently associated with better follow-up responses. Together, these findings suggest that stewardship-aligned responses are associated with both individual capability and the surrounding practice environment. This interpretation accords with behavioural and qualitative literature showing that antimicrobial prescribing is associated with social norms, workflow, access to diagnostic support, and the perceived consequences of delaying or narrowing treatment15–17,25. Hospital level moderated the association between reasoned-action beliefs and BI_W2, with the strongest relationship observed in tertiary hospitals. One possible explanation is that tertiary hospitals have stronger infrastructure, diagnostics, and stewardship governance, which may provide conditions in which reasoned-action beliefs are more closely associated with stewardship-aligned vignette responses4,6,19. This explanation remains hypothetical because the study did not directly test the institutional mechanisms underlying the interaction. In lower-tier settings, structural constraints may likewise weaken the observed association. These findings have practical implications. Interventions that focus only on attitudes are unlikely to be sufficient. Programmes may need to combine education and skills development with social and organisational supports, barrier reduction, and improved diagnostic stewardship3,15,26–29. Tailoring intervention packages by hospital tier may be more effective than assuming a single stewardship model fits all institutions. Conclusion In this longitudinal cohort of Chinese clinicians, follow-up stewardship-aligned vignette responses were most consistently associated with knowledge and skills, self-efficacy, social support, perceived patient-centred benefits, fewer perceived barriers, and reasoned-action beliefs. TPB-related beliefs were not independently predictive after mutual adjustment. The moderation by hospital level indicates that institutional context matters and that stewardship interventions should be designed with hospital tier in mind. Limitations Several limitations should be considered. First, BI_W2 was based on self-reported clinical vignettes rather than observed prescribing, so responses may overestimate real-world stewardship under time pressure or patient demand. Second, although the study was longitudinal and adjusted for baseline vignette score, the observational design does not establish causality. Unmeasured factors may influence both baseline psychosocial constructs and follow-up responses, and reverse causality cannot be ruled out; for example, prior stewardship experience may have influenced baseline psychosocial beliefs. Third, some attrition occurred between waves, and retention differed somewhat by hospital level. Finally, the sample was hospital-based and drawn from Chinese settings, which may limit generalisability to other health systems or primary-care contexts outside the study frame. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Author contributions Xiao Liu, Le Han, and Guyan Wang conceptualized the study. Xiao Liu and Le Han performed the statistical analyses and drafted the manuscript. Peng Xian, Yue Chang, Ying Liu, Lijun Ma, Jie Ren, Yunshuang Liu, Xiaosong Wang, Yujuan Long, Li Ren, Hui Tang, Qi Liu, Ling Liang, and Fang Jin contributed to data collection, interpretation of findings, and critical revision of the manuscript. Guyan Wang supervised the study. All authors approved the final manuscript. Funding This study was supported by the Beijing High-level Public Health Technical Talent Training Plan (Lingjunrencai-03–01). Data availability All data supporting the findings of this study are available within the paper and its Supplementary Information. Declarations Ethics approval and consent to participate The study protocol was reviewed and approved by the Ethics Committee of Beijing Tongren Hospital (Approval No. TREC2025-KYS230). All participants provided informed consent electronically before completing the questionnaire. The study was conducted in accordance with the Declaration of Helsinki. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement All data supporting the findings of this study are available within the paper and its Supplementary Information.