Understanding COVID-19 vaccine hesitancy among older adults in post-zero-COVID China: a predictive modeling study

Abstract

Background

The inadequate vaccination rates observed in China’s older population underscore the urgent need for immediate action to accelerate the immunization campaign and mitigate the consequences of adjusting the zero COVID-19 strategy.

Objective

This study’s objective was to identify key predictors of COVID-19 vaccine hesitancy among older adults in China during the post-zero-COVID period and to develop interpretable models to inform the development of targeted intervention strategies.

Methods

We conducted a cross-sectional study between January and March 2023, sampling 647 older persons across fifteen Chinese provinces. Predictors included sociodemographic characteristics, health status, psychological antecedents of vaccination, perceptions related to the COVID-19, and mental health. Group LASSO regression was employed for feature selection, followed by binary multivariable logistic regression and Random Forest modeling. SHapley Additive Explanations (SHAP) was used to illuminate the variable significance.

Results

Among the participants, the mean age was 68.36 ± 6.26 years, and 49.9% were male. The prevalence of vaccine hesitancy was 53.3% (95% CI, 49.5%–57.2%). Significant predictors of reduced vaccine hesitancy identified in the logistic regression model included elevated confidence (β = -0.852, P < 0.001), increased fear of COVID-19 (β = -0.060, P = 0.002), and residence in the Midlands (β = -0.840, P = 0.007) or Western regions (β = -0.899, P = 0.004) relative to the Eastern region. Heightened perceived constraint (β = 0.390, P < 0.001) significantly predicted higher vaccine hesitation. The logistic regression model (AUC = 0.827) and Random Forest model (AUC = 0.808) demonstrated high predictive performance. SHAP analysis confirmed the importance of confidence, constraints, region, COVID-19 vaccination status, education level, and collective responsibility in shaping individual-level predictions.

Conclusion

COVID-19 vaccine hesitancy among older adults in China is shaped by psychological attitudes, structural barriers to access, and changing risk perceptions following the post-zero-COVID strategy era. Our study demonstrates that interpretable modeling effectively identifies these key drivers, providing a clear evidence base for developing targeted public health strategies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12877-025-06758-z.

Introduction

Expanding COVID-19 vaccination coverage among older adults remains a pressing global public health priority [1]. Although early national surveys in China indicated that over 90% of adults expressed willingness to receive a COVID-19 vaccine [2], actual vaccination rates among those aged 60 years and older have consistently lagged behind younger age groups [3]. As of March 2023, only 73.15% of older adults in China had received their first booster dose [4]—substantially lower than in countries such as the United Kingdom (88.25%) and Australia (94.58%) [4]. The sudden end of China’s zero-COVID policy in December 2022 [5, 6] triggered a sharp rise in infections among older individuals [7], prompting the government to intensify efforts through targeted vaccination campaigns. However, vaccine hesitancy continues to pose a major barrier to achieving adequate coverage in this vulnerable population [8].

Acceptance of vaccination is an outcome behavior resulting from a complex decision-making process [9]. Vaccine hesitancy refers to a delay in acceptance or refusal of vaccination despite the availability of vaccination services [9]. Understanding the predictors driving vaccine hesitancy in older adults is essential for informing effective interventions. Previous studies have identified demographic characteristics [10], health-related factors [11], psychological antecedents (e.g., confidence, complacency) [12], and mental health conditions [13] as important contributors to vaccine hesitancy. However, despite these insights, few studies have employed comprehensive analytical frameworks that integrate psychological, epidemic-related, and sociodemographic predictors within older populations [14, 15]. In addition, changes in public opinion following the zero-COVID strategy and the psychological and behavioral effects of extended containment programs may have influenced vaccine decisions in ways not fully explored by earlier surveys [16–18].

Recent advancements in machine learning (ML) offer valuable tools for addressing these gaps, particularly by modeling complex and potentially non-linear relationships among multiple predictors. Group Least Absolute Shrinkage and Selection Operator (Group LASSO) allows for systematic feature selection while preserving grouped categorical structures [19]. Unlike traditional methods, this approach minimizes the risk of overfitting and improves interpretability by ensuring that entire categorical variables are either included or excluded as a group. Following feature selection, multivariable logistic regression quantifies linear associations, whereas Random Forest (RF) models capture non-linear relationships and interaction effects [20, 21]. To further enhance model transparency, SHapley Additive Explanations (SHAP) provide insights into both global and individual-level predictor contributions [22]. Despite these advances, integrated machine learning approaches that combine feature selection, linear modeling, and non-linear interpretation have not yet been extensively applied to understand vaccine hesitancy among older adults in China during the post-zero-COVID period, where risk perceptions and behavioral patterns may differ from those in previous waves of the pandemic.

In this study, we aim to identify key predictors of vaccine hesitancy in this population by modelling epidemiological, demographic, psychological determinants using interpretable machine learning tools. Specifically, we employ Group LASSO for variable selection, followed by multivariable logistic regression and RF modeling to characterize both linear and non-linear associations. SHAP values are used to interpret the results at both global and individual levels. This study aims to provide new insights into the changing factors influencing vaccine hesitancy and to guide strategies for enhancing vaccine uptake among older adults in China.

Methods

Study Population, Sampling, and recruitment

We conducted a cross-sectional study between January and March 2023. The study population consisted of older adults recruited from 22 community health centers (CHCs) across 15 provinces in China, including 6 provinces in the eastern region (7 CHCs), 6 provinces in the central region (9 CHCs), and 3 provinces in the western region (6 CHCs).

We conducted recruitment using a consecutive convenience sampling method, whereby all eligible individuals attending the health centers during the study periods were invited to participate. Inclusion criteria for this study were: (1) being aged 60 years or older; (2) attending one of the selected community health centers during the study period; and (3) being able to understand the study’s purpose and provide informed consent. Participants were excluded if they were unable to complete the questionnaire due to severe hearing, visual, or mobility impairments. This was assessed on-site by trained research assistants, who conducted a brief screening via conversation and observation to ensure participants could reliably understand and respond to the questionnaire items before enrollment. An entire recruitment and selection process is summarized in a flowchart (Fig. 1).

Fig. 1.

Flowchart of study participant selection

The final sample included 647 participants. A post-hoc power analysis confirmed that this sample size provided over 99% power to detect a small effect size (Odds Ratio = 1.5, α = 0.05) in the logistic regression analysis, indicating the sample was sufficient for the study’s objectives.

Data collection and ethical considerations

We collected data using a pre-designed, pre-tested, and semi-structured questionnaire. Participants completed either a paper-based questionnaire or an electronic questionnaire via QR code scanning on their smartphones. To encourage participation, respondents received a small incentive (e.g., medical masks or disinfecting alcohol). Ethics approval was obtained from the Ethics Review Committee of the School of Public Health and Nursing at Shanghai Jiao Tong University (approved on 20 March 2020; Approval No. SJUPN-202018).All participants provided informed consent, and anonymity was guaranteed throughout the study.

Assessments

General information

Baseline data were categorized into two groups: (1) Sociodemographic variables, including sex, age, region, marital status, religious beliefs, and education level; and (2) Health-related factors, such as chronic disease status, medical background, perceived health changes before and after COVID-19, SARS-CoV-2 infection history, COVID-19 vaccination status, and frequency of influenza vaccination over the past three years.

Vaccine hesitancy

In accordance with the WHO criteria and relevant research [9, 23], vaccine hesitancy is assessed by allowing participants the option of choosing whether or not to be vaccinated based on their actual situation. The propensity to choose was divided into six options, ranging from complete unwillingness to complete willingness. Participants who selected the first five options—complete refusal, inclined to refuse but uncertain, wait and see/hold back/delay, partially inclined to be willing, or willing to be keen but uncertain—were classified as having vaccine hesitancy.

Psychological antecedents of vaccination

Psychological antecedents were measured using the validated 5 C scale developed by a German research team [24]. The scale comprises five constructs—confidence, complacency, constraints, calculation, and collective responsibility—each rated on a seven-point Likert scale (1 = strongly disagree, 7 = strongly agree). These constructs are confidence (belief in vaccine effectiveness and safety), complacency (low perceived infection risk, belief vaccination is unnecessary), constraints (practical/psychological barriers like access or scheduling), calculation (extensive information seeking and risk-benefit analysis), and collective responsibility (willingness to vaccinate to protect others). The average score of each construct was calculated. The Chinese version of the scale was adapted through translation, cultural adjustment, and psychometric testing (details provided in Supplementary Material 1).

COVID-19 Epidemic-Related perceptions

(1) Benefit-Finding from Epidemic Scale (BFES): A 10-item scale developed in China [25] that assesses perceptions of benefits derived from the COVID-19 pandemic, including health awareness, familial values, meaning in life, and social connection. Responses were rated on a seven-point Likert scale, with higher scores indicating greater benefit perception. The Cronbach’s α coefficient in this study was 0.87, indicating strong internal consistency.

(2) Chinese version of the 2019 Coronavirus Disease Fear Scale (FCV-19-C): A seven-item scale used to measure the level of concern surrounding the COVID-19 pandemic, with scores ranging from 7 to 35, where higher scores indicate greater fear [26]. The Chinese version demonstrated high internal consistency [27], with Cronbach’s α of 0.927 in this study.

Mental health status

(1) Generalized anxiety disorder 7-items questionnaire (GAD-7): A seven-item scale that assessed anxiety symptoms based on a four-point scale (0–3) indicating symptom frequency over the past two weeks [28]. The total score ranges from 0 to 21, with higher scores indicating greater anxiety severity. The Chinese version of the GAD-7 has demonstrated high reliability [29], with Cronbach’s α of 0.940. In this study, the GAD-7 total score was used as a continuous variable in the primary analyses.

(2) Depression patient health questionnaire-9 (PHQ-9): A nine-item scale assessing depressive symptoms on a four-point scale (0–3) [30]. The total score ranges from 0 to 27, with higher scores indicating greater depression severity. The Chinese version is highly reliable [31], with Cronbach’s α of 0.945 in this study. In this study, the PHQ-9 total score was used as a continuous variable in the primary analyses.

Statistical analyses

All statistical analyses were performed using R (version 4.4.1) and Python (version 3.8.16). To minimize the influence of confounding, our analytical strategy involved systematic feature selection followed by multivariable modeling. Descriptive statistics were employed to summarize sociodemographic characteristics, health-related information, psychological antecedents, COVID-19 epidemic-related perceptions, and mental health status. A comparison of these characteristics between vaccine-hesitant and non-hesitant groups is detailed in Table 1. The normality of all continuous variables was assessed using the Shapiro-Wilk test. Although continuous variables were described using mean and standard deviation (mean ± SD); however, due to their non-normal distribution, group differences were assessed using the Wilcoxon rank sum test (also known as the Mann-Whitney U test). Categorical variables were presented as frequencies (n) and percentages (%) and evaluated using chi-square (χ²) tests for group differences. Fisher’s exact test was applied to categorical variables with anticipated low cell counts. To account for multiple comparisons in Table 1, all p-values from the initial bivariate group comparisons (i.e., Wilcoxon, chi-square, and Fisher’s tests) were adjusted using the Benjamini-Hochberg procedure to control the false discovery rate.

Table 1.

Comparison of participant characteristics between hesitant and Non-Hesitant groups (N = 647)

Variables	Total (m = 647)	Hesitant (n = 345)	Non-Hesitant (n = 302)	χ²/W	P value	P value adjusted
Age (years)	68.36 ± 6.26	68.29 ± 6.23	68.42 ± 6.35	51,764	0.889	0.983
Gender				0.56	0.456	0.552
Male	323 (49.9%)	167 (48.4%)	156 (51.7%)
Female	324 (50.1%)	178 (51.6%)	146 (48.3%)
Region				38.09	< 0.001	< 0.001
Eastern	255 (39.4%)	173 (50.1%)	82 (27.2%)
Midlands	156 (24.1%)	76 (22.1%)	80 (26.4%)
Western	236 (36.5%)	96 (27.8%)	140 (46.4%)
Marital status				-	0.951	0.983
Unmarried	10 (1.6%)	5 (1.4%)	5 (1.7%)
Married	568 (87.8%)	302 (87.5%)	266 (88.1%)
Divorced/widowed	69 (10.6%)	38 (11.1%)	31 (10.2%)
Religious beliefs				5.55	0.018	0.029
No	567 (87.6%)	292 (84.6%)	275 (91.1%)
Yes	80 (12.4%)	53 (15.4%)	27 (8.9%)
Education level				28.72	< 0.001	< 0.001
Less than high school	400 (61.8%)	181 (52.5%)	219 (72.5%)
High school diploma	144 (22.3%)	100 (29.0%)	44 (14.6%)
College degree or more	103 (15.9%)	64 (185%)	39 (12.9%)
Chronic disease				0	0.983	0.983
No	327 (50.5%)	175 (50.7%)	152 (50.3%)
Yes	320 (49.5%)	170 (49.3%)	150 (49.7%)
Medical background				6.39	0.041	0.052
None	478 (73.9%)	241 (69.9%)	237 (78.5%)
I had medical background	26 (4.0%)	17 (4.9%)	9 (3.0%)
My family had medical background	143 (22.1%)	87 (25.2%)	56 (18.5%)
Health status				8.2	0.017	0.029
Remain unchanged	331 (51.2%)	175 (50.7%)	156 (51.7%)
Better	94 (14.5%)	39 (11.3%)	55 (18.2%)
Worse	222 (34.3%)	131 (38.0%)	91 (30.1%)
SARS-CoV-2 infection				7.63	0.022	0.031
Nobody got infected in my family	63 (9.7%)	25 (7.2%)	38 (12.6%)
At least one of my family got infected	339 (52.4%)	195 (56.5%)	144 (47.7%)
I got infected	245 (37.9%)	125 (36.3%)	120 (39.7%)
COVID-19 vaccination				42.5	< 0.001	< 0.001
Unvaccinated	71 (12.0%)	61 (17.6%)	10 (3.3%)
Primary series	271 (41.9%)	151 (43.8%)	120 (39.7%)
Booster doses	305 (47.1%)	133 (38.6%)	172 (57.0%)
Frequency of influenza vaccination in the past 3 years				14.78	< 0.001	0.001
Every year	175 (27.1%)	75 (21.7%)	100 (33.1%)
Once/twice vaccinated	231 (35.6%)	121 (35.1%)	110 (36.4%)
Never	241 (37.3%)	149 (43.2%)	92 (30.5%)
5 C scale dimensions
Confidence	5.48 ± 1.50	4.74 ± 1.48	6.33 ± 1.01	18,226	< 0.001	< 0.001
Complacency	3.21 ± 1.57	3.54 ± 1.43	2.84 ± 1.63	67,116	< 0.001	< 0.001
Constraint	2.55 ± 1.56	3.01 ± 1.56	2.02 ± 1.39	72,846	< 0.001	< 0.001
Calculation	4.82 ± 2.09	4.91 ± 1.88	4.72 ± 2.31	51,955	0.952	0.983
Collective responsibility	5.32 ± 1.64	4.87 ± 1.61	5.84 ± 1.52	32,242	< 0.001	< 0.001
Benefit-Finding from epidemic	56.81 ± 11.26	54.82 ± 11.94	59.1 ± 9.96	41,092	< 0.001	< 0.001
Fear of COVID-19	17.48 ± 8.65	18.49 ± 8.33	16.37 ± 8.9	60,926	< 0.001	< 0.001
Anxiety	3.53 ± 4.62	4.22 ± 5.04	2.73 ± 3.96	62,028	< 0.001	< 0.001
Depression	3.79 ± 5.38	4.86 ± 5.97	2.56 ± 4.30	65,066	< 0.001	< 0.001

Data are presented as n (%) for categorical variables and mean ± SD (standard deviation) for continuous variables. The ‘χ²/W’ column displays the test statistic: W-values from the Wilcoxon rank-sum test were used for continuous variables, while χ²-values from Pearson’s χ² test were used for categorical variables. For Marital Status, Fisher’s exact test was used due to low expected cell counts; a ‘-’ is indicated in the column. The ‘P value’ column displays the unadjusted results from these initial bivariate tests. These p-values were then adjusted using the Benjamini-Hochberg procedure to control the false discovery rate, with the final results presented in the ‘P value adjusted’ column

Data preprocessing

Data preprocessing was conducted before model development. For the Group LASSO and logistic regression models, categorical predictors were transformed into dummy variables using one-hot encoding; dummy variables derived from the same original categorical predictor were subsequently treated as a unified group during the penalized regression. For the Group LASSO feature selection phase, continuous predictors were standardized by centering and scaling. The Random Forest model utilized continuous predictors on their original scale and categorical variables as factors. The design matrix for Group LASSO comprised these standardized continuous variables and the one-hot encoded categorical variables from the training set.

Feature selection with group LASSO

To identify the most relevant predictors of COVID-19 vaccine hesitancy while preserving the inherent grouping of categorical variables, we used Group LASSO regression method. The dataset was partitioned into a training set (80%) and a testing set (20%). The Group LASSO was applied to the training data (with one-hot encoded categorical predictors and standardized continuous predictors). This method facilitates systematic variable selection by retaining or removing entire groups of dummy variables corresponding to a single categorical predictor, thereby enhancing model interpretability and mitigating overfitting. Ten-fold cross-validation on the training set determined the optimal regularization parameter (λ). The λ value that minimized the cross-validated misclassification error (λ_min) was selected. This cross-validation process serves as the performance evaluation for the feature selection step, ensuring the chosen predictors are optimized for predictive accuracy. Features with non-zero coefficients in the resulting Group LASSO model were considered selected for subsequent analyses.

Interpretive model development

The original predictor variables corresponding to the features selected by the Group LASSO model were used to fit a standard multivariable logistic regression model to the training data. This model aimed to determine the association between these selected predictors (with continuous variables on their original scales) and vaccine hesitancy, which was the binary dependent variable (Yes/No). Coefficients (β), standard errors (SE), z-values, and p-values were estimated to evaluate the magnitude, direction, and statistical significance of each predictor’s association with vaccine hesitancy.

Predictive model development and performance evaluation

The goodness-of-fit of the final logistic regression model on the training data was assessed using the Hosmer-Lemeshow test, which indicated a good model fit (χ² = 7.56, df = 8, P = 0.478). The predictive performance of this final logistic regression model and a RF model was evaluated on the independent test set. For the RF model, categorical predictors were treated as factors, and continuous predictors were used on their original numeric scales. The RF model was constructed with 500 trees. The performance of both models on the test set was assessed using accuracy, sensitivity (correctly identifying hesitant individuals), and specificity (correctly identifying non-hesitant individuals) derived from their respective confusion matrices. The Area Under the Receiver Operating Characteristic curve (AUC) was also calculated to evaluate each model’s overall discriminative power.

Model interpretation using SHAP

To enhance the interpretability of the RF model, SHAP values were estimated for predictions on the test set, quantifying each predictor’s contribution to the model output. A SHAP summary plot was used for global interpretability, displaying the overall importance and directionality of each feature based on its SHAP values across all instances in the test set. For individualized interpretation, SHAP waterfall and force plots were generated for representative cases, illustrating how specific feature values for a particular participant influenced the model’s predicted probability of vaccine hesitancy.

Results

Participant characteristics

A total of 647 participants aged 60 years and older were included, of whom 345 (53.3%, 95% CI: 49.5%–57.2%) demonstrating vaccine hesitancy (Table 1). The overall mean age was 68.36 ± 6.26 years (range: 60–95 years); no significant age difference was observed between hesitant and non-hesitant groups (Adjusted P = 0.983). Gender (49.9% male) and marital status (87.8% married) were not significantly associated with vaccine hesitancy (Adjusted P = 0.552 and P = 0.983, respectively). Significant associations with vaccine hesitancy (Adjusted P < 0.05) were observed for region of residence, religious beliefs, education level, health status, and SARS-CoV-2 infection history. Chronic disease status was not a significant factor (Adjusted P = 0.983). COVID-19 vaccination status and frequency of past influenza vaccination were strongly associated to vaccine hesitancy (Adjusted P < 0.001 and P = 0.010, respectively).

All 5 C psychological antecedents (confidence, complacency, constraint, collective responsibility), except for calculation (Adjusted P = 0.983), showed significant differences between groups (Adjusted P < 0.001). Hesitant individuals reported lower mean confidence (4.74 ± 1.48 vs. 6.33 ± 1.01) and collective responsibility (4.87 ± 1.61 vs. 5.84 ± 1.52), and higher complacency (3.54 ± 1.43 vs. 2.84 ± 1.63) and constraint (3.01 ± 1.56 vs. 2.02 ± 1.39). Epidemic-related perceptions (benefit-finding, fear of COVID-19) and mental health scores (anxiety, depression) also differed significantly between groups (all Adjusted P < 0.001).

Variable selection via group LASSO

The Group LASSO regression with ten-fold cross-validation identified an optimal λ_min value of 0.0114. Figure 2A shows the coefficient paths for all predictor groups as the penalty parameter λ changes, demonstrating how coefficients are progressively shrunk toward zero and eventually excluded from the model. The cross-validation curve (Fig. 2B) shows the misclassification error plotted against log(λ). The vertical dashed blue line shows the chosen λ_min. Using this λ_min with the Group LASSO model trained on the training data picked predictors whose coefficients stayed non-zero. Age, region, SARS-CoV-2 infection history, confidence, complacency, constraint, calculation, collective responsibility, fear of COVID-19, and depression were the most important variables chosen by this method.

Fig. 2.

Group LASSO Coefficient Paths and Cross-ValidationA Complete coefficient paths for 21 predictor groups across log(λ). Each colored line shows how a group’s standardized coefficient evolves; coefficients shrink to zero as λ increases. Labels mark each predictor. B Ten-fold CV misclassification error (red dots) ± 1 SE (gray bars) plotted against log(λ). Blue dashed line = λ_min (min CV error); red dashed line = λ_1se (largest λ within one SE)

Predictors of vaccine hesitancy from logistic regression analysis

Following feature selection, a multivariable logistic regression model was fitted on the training data using the selected predictors to examine their associations with vaccine hesitancy, with detailed results presented in Table 2. Significant predictors of decreased vaccine hesitancy included higher confidence (β = −0.852, P < 0.001) and greater fear of COVID-19 (β = −0.060, P = 0.002). Living in certain regions compared to the reference region (Eastern) also significantly decreased hesitancy (e.g., Midlands β = −0.840, P = 0.007; Western: β = −0.899, P = 0.004). A significant predictor of increased vaccine hesitancy was greater perceived Constraint (β = 0.390, P < 0.001). Other variables included in the model, such as age, SARS-CoV-2 infection history, complacency, calculation, collective responsibility, and depression, did not show a statistically significant association with vaccine hesitancy in this multivariable model (all P > 0.05).

Table 2.

Predictors of COVID-19 vaccine hesitancy from binary multivariable logistic regression analysis

Predictor	Estimate (β)	Std. Error	z value	P-value
Intercept	4.203	1.543	2.724	0.006
Age	0.012	0.019	0.612	0.540
Region (Midlands vs. Eastern)	−0.840	0.311	−2.696	0.007
Region (Western vs. Eastern)	−0.899	0.316	−2.843	0.004
SARS-CoV-2 Infection (At least one family member infected vs. I got infected)	0.620	0.388	1.601	0.109
SARS-CoV-2 Infection (Nobody infected vs. I got infected)	0.173	0.403	0.430	0.667
Confidence	−0.852	0.109	−7.945	< 0.001
Complacency	0.088	0.084	1.057	0.291
Constraint	0.390	0.102	3.843	< 0.001
Calculation	0.059	0.065	0.906	0.365
Collective Responsibility	−0.086	0.078	−1.101	0.271
Fear of COVID-19	−0.060	0.020	−3.080	0.002
Depression	0.016	0.031	0.511	0.609

β beta coefficient from logistic regression, Std. Error Standard Error. The model’s goodness-of-fit was confirmed with the Hosmer-Lemeshow test (P = 0.478). The model was fitted on the training data using variables selected by Group LASSO. For this logistic regression model, continuous predictor variables were used on their original scales

Model performance comparison

The independent test set was used to evaluate the predictive performance of both the logistic regression and RF models, with “vaccine hesitancy” defined as the positive class. As shown in Fig. 3, Panels A and C present the confusion matrix and ROC curve for the logistic regression model, while Panels B and D present those for the RF model. Both models achieved the same overall accuracy (74.4%). The logistic regression model showed a slightly higher AUC (0.827 vs. 0.808), indicating marginally better discriminative ability. It also demonstrated higher specificity in identifying non-hesitant individuals (75.0% vs. 70.0%), whereas the RF model achieved higher sensitivity in detecting hesitant individuals (78.3% vs. 73.9%). In general, both models demonstrated strong predictive capabilities.

Fig. 3.

Model Performance ComparisonA Confusion matrix for the logistic regression model. B Confusion matrix for the Random Forest model. C ROC curve for the logistic regression model. D ROC curve for the Random Forest model

Importance of features related to vaccine hesitancy interpreted by SHAP values

The SHAP summary plot (Fig. 4A) highlighted the global impact of features on predicted vaccine hesitancy in the RF model. Higher vaccine confidence and stronger collective responsibility were key factors in reducing predicted hesitancy (negative SHAP values). Conversely, higher perceived constraints and increased complacency resulted in higher predicted hesitancy. Regional differences indicated that residing in Western China was associated with lower predicted hesitancy, while residents in the Eastern region showed higher predicted hesitancy. Receiving a COVID-19 booster dose corresponded to lower predicted hesitancy, whereas being unvaccinated was associated with higher hesitancy. Lower educational attainment generally increased predicted hesitancy, while higher calculation scores modestly reduced it. Higher depression scores and older age were associated with slightly increased predicted hesitancy.

Fig. 4.

SHAP values for predictors of COVID-19 vaccine hesitancy in the Random Forest model.A SHAP summary plot. B SHAP waterfall plot. C. SHAP force plot

For a representative participant (Fig. 4B and C), SHAP waterfall and force plots illustrated how individual feature values contributed to their specific prediction. For this individual, whose features culminated in a final predicted vaccine hesitancy of approximately 0.740 (from a baseline average of 0.546), higher confidence and collective responsibility unexpectedly increased their predicted hesitancy. Other factors increasing their predicted hesitancy included residing in the Midlands region and poorer health status. Conversely, very low perceived constraints and lower education attainment significantly reduced this individual’s predicted hesitancy.

Discussion

This study examined the prevalence and predictors of COVID-19 vaccine hesitancy in Chinese older adults after adjusting for a zero-covid strategy. Our findings showed that 53.3% of participants exhibited hesitancy, highlighting a significant public health concern for this high-risk population. The main predictors were low self-confidence, decreased collective responsibility, high complacency, and perceived constraints. The logistic regression model had an AUC of 0.827 with 74.4% accuracy, and the random forest model had an AUC of 0.808 with 74.4% accuracy. These results validate the robustness of findings regarding these influences and suggest that both linear (logistic regression) and nonlinear (random forest) relationships were explored in predicting vaccine hesitancy.

The primary methodological advantage of this study is its integrated analytical process. We first used Group LASSO to select features based on data, which reduces researcher bias and the risk of overfitting. Subsequently, the complementary use of multivariable logistic regression and a Random Forest model provided a more comprehensive understanding than a single model could. Although logistic regression accurately measured essential linear relationships, SHAP-interpreted random forests illuminated the subtle significance of predictors like complacency and collective responsibility, which were not deemed significant in the final linear model. The “transparent box” methodology suggests that these variables are essential determinants of hesitation, presumably influencing through complex or nonlinear mechanisms.

Both the multivariable logistic regression analysis, which yielded β coefficients for statistically significant associations, and the SHAP importance scores from the Random Forest model, which illuminated the global influence of features, informed the interpretation of predictors in this study. Vaccine confidence demonstrated the strongest negative association with hesitancy (β = − 0.852) and ranked highest in global SHAP importance, underscoring its central role across modeling approaches. Confidence may reflect trust in vaccine safety, efficacy, and the public health system [32]. Among older adults, skepticism arising from rapid vaccine development and concerns about long-term side effects may undermine this trust [33]. Perceived constraints were positively associated with vaccine hesitancy (β = +0.390) and ranked second in SHAP importance, highlighting their strong impact across models. Despite efforts by the Chinese government, such as free vaccination, community-based delivery, and in-home services for seniors (47), barriers remained. Long wait times due to limited health personnel [34], along with low digital literacy and unfamiliarity with appointment systems among older adults [35], contributed to procedural difficulties. Beyond these structural issues, a form of vaccine fatigue may have emerged after prolonged exposure to pandemic-related messaging [36]. Some individuals may have experienced informational overload and psychological weariness, leading to passive disengagement rather than active refusal [37]. This behavioral inertia represents a subtler but important dimension of perceived constraints in the post-pandemic context. In addition, greater fear of COVID-19 was a significant predictor of decreased vaccine hesitancy in the logistic regression model (β = −0.060), suggesting that individuals with heightened perceived risk [38] were more motivated to vaccinate.

Although not statistically significant in the final multivariable logistic regression model, SHAP analysis for the Random Forest model identified stronger collective responsibility as a key factor reducing predicted hesitancy. That suggest that prosocial motivation—such as protecting grandchildren or supporting community immunity—may activate internalized social norms that encourage vaccination [39]. These results emphasize the relevance of value-based messaging, especially in collectivist cultural settings [40]. Similarly, while complacency and calculation were not a statistically significant predictors in the final logistic regression model, SHAP analysis indicated that elevated complacency and calculation increased predicted hesitancy. Complacency reflects a perceived lack of vulnerability to COVID-19 and may reduce the perceived necessity of vaccination, particularly among individuals who view themselves as healthy [41, 42]. Meanwhile, high levels of calculation may indicate over-reliance on personal deliberation or information seeking, which, in the presence of conflicting narratives or misinformation, could reinforce indecision and doubt [24, 43].

Both logistic regression and SHAP analyses revealed that vaccine hesitancy was higher among individuals residing in the eastern part of China, specifically in more economically developed areas. Regarding education, while SHAP results indicated that lower educational attainment independently increased predicted hesitancy, descriptive analysis conversely revealed an unexpected pattern: older adults with higher education levels reported greater vaccine hesitancy. This latter pattern of higher hesitancy among more educated older adults—is also observed in other settings [44–46]. Notably, this does not suggest that higher education is inherently “worse”; instead, it may foster greater autonomy and demand for transparent, evidence-based justifications [47]. Indeed, prior research has shown that educated individuals may experience information overload or develop greater sensitivity to inconsistencies in risk communication, leading to increased skepticism of vaccines [48]. Reinforcing this, previous nationwide research in China yielded comparable results, indicating that vaccine hesitancy was more prevalent among older people with higher levels of education residing in developed regions [44]. Similar trends have been observed outside of China, where vaccination rates against influenza are sometimes lower among those with higher socioeconomic status [46, 49]. A potential explanation is that people who are financially secure and well-educated have more access to a wide range of information sources. More exposure may lead to information overload and an increased risk of receiving misleading information, which could make people less likely to get vaccinated [50]. Therefore, precise and personalized communication strategies that strike a balance between scientific complexity and simplicity are necessary.

SHAP force and waterfall plots illustrated how individual-level factors contributed to predicted hesitancy. For example, a representative individual with high predicted hesitancy (approximately 0.740) exhibited an unexpected pattern: higher confidence and stronger collective responsibility seemed to increase risk (Fig. 4B). In certain circumstances, people might require additional information and convincing evidence before getting vaccinated, which could explain this paradox. People who reported they were in poorer health were even much more likely to be hesitant, probably because they were more worried about the safety or effectiveness of the vaccine [51]. Additionally, this individual’s predicted hesitancy increased with residence in the Midlands, while very low perceived constraints and a lower education level reduced it. These SHAP insights at the individual level underscore that constructs like “confidence” are context-dependent and may have different implications across subpopulations. This challenges the effectiveness of one-size-fits-all intervention strategies. Systemic issues, like inconsistent vaccination practices where providers may refuse to vaccinate people with comorbidities because they are worried about risking legal action, can make people less trusting and more hesitant [52]. Resolving these issues requires implementing consistent national policies and establishing precise, reliable communication from healthcare providers.

Implications

The public health implications of these findings are multifaceted. First, interventions should focus on rebuilding vaccine confidence by ensuring transparent and consistent communication, preferably conveyed by trusted local figures like primary care providers or community health workers [53]. Second, it is essential to tackle structural barriers by introducing walk-in services, streamlining offline booking processes, and developing home-based vaccination programs specifically designed for older adults [54].

Additionally, health communication strategies ought to be tailored to individual psychological profiles. For example, individuals exhibiting high complacency may find messaging that highlights vulnerability to be beneficial [47], whereas those demonstrating low collective responsibility may respond more effectively to appeals focused on community protection. Community-based platforms, including neighborhood committees and senior centers, can function as effective channels for targeted outreach and peer influence (51).

This study highlights that vaccine hesitancy in older adults is influenced by a complex interaction of psychological, perceptual, and structural factors. This involves more than just altering personal perspectives; it requires a comprehensive transformation of the environment in which older adults engage in health decision-making. Future initiatives should focus on restructuring systems to address the unique requirements of aging populations and reinforce vaccination as a standard practice within society. This is the area where ongoing public health innovation is essential.

Limitations

We should acknowledge several limitations. First, the cross-sectional design restricts causal interpretation. Although we observed strong associations between psychological, perceptual, and demographic characteristics and vaccine hesitancy, the directionality and temporal sequence of these relationships remain uncertain. Longitudinal studies are essential to capture how vaccine attitudes evolve, especially in response to policy shifts, epidemic dynamics, or targeted interventions.

Second, our sample was recruited via non-probability convenience sampling. Although we included participants from 15 provinces across eastern, central, and western China to capture diversity, this method means the findings may not be fully representative of the entire older adult population. While the overall size (N = 647) demonstrated sufficient statistical power for our regression model, it may be considered limited for machine learning. Validation using larger and more heterogeneous populations is needed to improve model generalizability and assess potential variations across regions and subcultures.

Secondly, our model did not account for several influential factors previously shown to shape vaccination, such as exposure to misinformation [55], and digital media or social networks influences [56]. The lack of these variables may diminish the model’s explanatory ability. Subsequent research should integrate these elements to more accurately represent the intricacies of the social and informational context influencing vaccine decisions.

Additionally, the logistic regression model employed in this study concentrated on the primary effects of the predictors and did not analyze any interaction terms. Although the random forest model inherently captures such interactions, future research could benefit from explicitly examining theoretically relevant interaction terms within a logistic regression framework. This may enhance comprehension of the influence of synergistic effects on vaccine hesitancy and may offer more precise recommendations for targeted intervention measures.

Finally, the data collection period, occurring from January to March 2023, coincided with the immediate aftermath of China’s sudden adjustment of the zero-COVID strategy. This time was marked by policy changes, public uncertainty, and heightened infection risk, which may have momentarily influenced individuals’ perceived urgency or motivation to receive vaccinations. The occurrence of “post-policy volatility” may generate behavioral noise, thereby affecting the stability and understanding of observed patterns. While our findings align with broader national trends, they may partially reflect transient dynamics rather than enduring hesitancy profiles.

Conclusion

This study showed that in post-zero-COVID China, older adults’ COVID-19 vaccine hesitancy is primarily driven by psychological factors, including low confidence and high perceived constraints, alongside contextual factors such as regional disparities and fear of the disease. Our study successfully demonstrated that an interpretable machine learning process, combining feature selection with predictive modeling, could accurately identify these key drivers with high performance. These findings provide a data-driven evidence base for developing specialized public health strategies to improve vaccine uptake in this at-risk population.

Authors’ contributions

Enming Zhang: Conceptualization; Data curation; Formal analysis; Project administration; Visualization; Writing - Original Draft. Zhengyue Dai: Conceptualization, Data curation; Formal research; Writing - Original Draft. Shuhui Shang: Data curation; Writing - review & editing. Xiaolong Wang: Data curation; Writing - review & editing. Jiale Hu: Supervision. Xian Zhang: Supervision. Daqiao Zhu: Conceptualization, Supervision. Qiong Fang: Conceptualization, Supervision. All authors approved the final version.

Funding

This work was supported by the 3-year action plan for the construction of Shanghai’s public health system (2020–2022), academic leaders cultivating project [grant number GWV-10.2-XD33], an Innovative research team of high-level local universities in Shanghai [grant number SHSMU-ZDCX20212801].

Data availability

The datasets used and analyzed during the present study are available from the corresponding author at the reasonable request.

Declarations

Ethics approval and consent to participate

The study has been approved by Shanghai Jiao Tong University School of Public Health and Nursing Research Ethics Committee approval (SJUPN-202018) and all methods were performed in accordance to the Declaration of Helsinki. All data were collected anonymously following informed consent, and participation was voluntary.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.