Survey design and population

Between June 20 and August 31, 2022, a survey was carried out in 31 provinces, autonomous regions, special administrative regions, and municipalities in China. The survey employed a multistage sampling approach, based on the quota attributes of China’s seventh national census data for cities, including gender, age, and urban-rural distribution. The specific quota method has been previously reported [16].

The present study utilized a face-to-face survey to collect questionnaire data. This study included Chinese participants aged 12 years or older who voluntarily consented to participate, possessed the cognitive ability to understand the questionnaire items, and could complete the questionnaires independently. For participants who possessed cognitive capability but were unable to complete the questionnaires due to mobility constraints, assistance with questionnaire completion was provided by interviewers.

The initial PBICR survey enrolled 21,916 participants. After excluding 2178 participants < 18 years old, a total of 19,738 participants were finally included in this study.

The Ethics Research Committee of the Health Culture Research Center of Shaanxi approved this study (No. JKWH-2022-02). All participants were required to provide informed consent before the collection of data, and the confidentiality of all collected data was anonymously and strictly maintained.

Survey instruments

Based on the prior studies [14, 17,18,19], this study comprehensively encompasses multi-level factors that may be associated with participants’ acceptance level of ACP, based on the socioecological model depicted in Fig. 1. This study included factors across five levels: individual characteristics level (i.e., age group (18–44 years old, 45–64 years old, and ≥ 65 years old), gender (male vs. female), education level (junior high school and below, senior school and middle special school, junior college, and bachelor degree and above), diagnosed chronic disease (no vs. yes), personality traits, self-efficacy, health literacy, depression symptoms, anxiety symptoms, well-being, career status (student, have no job, have a job), and medical insurance type (self-pay, resident basic medical insurance, employee basic medical insurance, and commercial and multiple insurances)); individual behaviors level (i.e., smoking status (no vs. yes), drinking alcohol (no vs. yes), regular exercise (no vs. yes)); interpersonal networks level (i.e., have a spouse (no vs. yes), neighbor relations (a glide rating scale ranging from 1 (very poor) to 7 (very good)), perceived social support, family health, family communication, number of siblings (0, 1, 2, ≥ 3), family social status (a glide rating scale ranging from 1 (lowest) to 7 (highest)), family per capita monthly income (≤ 3000 Chinese Yuan, 3001–6000 Chinese Yuan, and ≥ 6001 Chinese Yuan)); community level (i.e., urban-rural distribution (urban vs. rural)).

Fig. 1

Factors associated with the acceptance of advance care planning based on the socioecological model

Acceptance level of ACP

Before initiating the research, researchers explained ACP to participants. ACP (also known as advance directives, living wills, or healthcare proxies) refers to the process by which competent adults of any age and health status make advance decisions about their future end-of-life medical care preferences based on their circumstances and document and share these preferences with family members and healthcare providers [1]. After ensuring participants’ comprehension of ACP, we assessed their attitudes by asking, ‘What is your acceptance level of ACP?’ The response was rated on a glide rating scale ranging from 0 (not accepting) to 100 (very accepting).

Personality traits

The personality traits of the participants were evaluated using the Big Five Inventory-10 (BFI-10) [20]. The BFI-10 includes five dimensions of personality: extraversion, agreeableness, conscientiousness, neuroticism, and openness. Each item is rated on a 5-point scale, scoring from 1 (totally disagree) to 5 (totally agree). Reverse items are scored from 1 (totally agree) to 5 (totally disagree). Higher scores represent higher magnitudes of personality traits. Given the limited number of items (i.e., two items per dimension) in the BFI-10, Cronbach’s α was not calculated [21].

Self-efficacy

The perceived self-efficacy of the participants was assessed using the New General Self-Efficacy Scale-Short Form (NGSES-SF) [22]. The NGSES-SF consists of three items. Each item is rated on a 5-point scale, scoring from 1 (strongly disagree) to 5 (strongly agree). Total summed NGSES-SF scores ranged from 3 to 15, with higher scores indicating greater perceived self-efficacy. In this study, Cronbach’s α of the NGSES-SF was 0.924.

Health literacy

The health literacy of the participants was measured using the Health Literacy Scale-Short Form (HLS-SF) [23]. The HLS-SF consists of nine items. Each item is rated on a 4-point scale, scoring from 0 (very difficult) to 3 (very easy). Total summed HLS-SF scores ranged from 0 to 27, with higher scores reflecting higher levels of health literacy. In this study, Cronbach’s α of the HLS-SF was 0.937.

Depression symptoms

The depression symptoms of the participants were determined using the Patient Health Questionnaire-9 (PHQ-9) [24]. Each item is rated on a 4-point scale, scoring from 0 (never) to 3 (nearly every day). Total summed PHQ-9 scores ranged from 0 to 27, with higher scores conveying more severe depression symptoms. In this study, Cronbach’s α of the PHQ-9 was 0.918.

Anxiety symptoms

The anxiety symptoms of the participants were appraised using the Generalized Anxiety Disorder-7 (GAD-7) [25]. Each item is rated on a 4-point scale, scoring from 0 (never) to 3 (nearly every day). Total summed GAD-7 scores ranged from 0 to 21, with higher scores denoting more severe anxiety symptoms. In this study, Cronbach’s α of the GAD-7 was 0.940.

Well-being

The well-being of the participants was scrutinized using the World Health Organization Well-Being Index-5 (WHO-5) [26]. Each item is rated on a 6-point scale, scoring from 0 (never before) to 5 (all times). Total summed WHO-5 scores ranged from 0 to 25, with higher scores indicating greater well-being. In this study, Cronbach’s α of the WHO-5 was 0.949.

Perceived social support

The perceived social support of the participants was evaluated using the Perceived Social Support Scale (PSSS) [27]. The PSSS consists of three items. Each item is rated on a 7-point scale, scoring from 1 (strongly disagree) to 7 (strongly agree). Total summed PSSS scores ranged from 3 to 21, with higher scores indicating greater levels of perceived social support. In this study, Cronbach’s α of the PSSS was 0.886.

Family health

The family health of the participants was assessed using the Family Health Scale-Short Form (FHS-SF) [28]. The FHS-SF comprises four dimensions: family/social/emotional health processes, family healthy lifestyle, family health resources, and family external social supports. The FHS-SF consists of ten items. Each item is rated on a 5-point scale, scoring from 1 (strongly disagree) to 5 (strongly agree). Reverse items are scored from 1 (strongly agree) to 5 (strongly disagree). Total summed FHS-SF scores ranged from 10 to 50, with higher scores conveying higher levels of family health. In this study, Cronbach’s α of the FHS-SF was 0.827.

Family communication

The family communication of the participants was assessed using the Family Communication Scale-10 (FCS-10) [29]. Each item is rated on a 5-point scale, scoring from 1 (strongly disagree) to 5 (strongly agree). Total summed FCS-10 scores ranged from 10 to 50, with higher scores denoting greater levels of family communication. In this study, Cronbach’s α of the FCS-10 was 0.966.

Statistical analysis

First, the Kolmogorov–Smirnov test was utilized to assess the normality of continuous variables. Continuous variables manifested non-normal distribution and were depicted using the median and interquartile range (IQR). Categorical variables were represented as numbers and percentages. Second, a random forest regression analysis was utilized to assess the significance of 28 variables derived from the socioecological model. The random forest regression method is a robust machine learning approach that integrates the concepts of ensemble learning and decision trees to produce precise predictions and assess the significance of variables [30]. This method is an expansion of the decision tree algorithm, in which numerous decision trees are generated using random subsets of both the data and features. The forecasts from these individual trees are subsequently combined to generate the ultimate output, resulting in enhanced resilience and predictive accuracy in comparison to a solitary decision tree. A significant benefit of random forest regression lies in its capacity to evaluate the significance of input variables within the model. Through the examination of each feature’s influence on model performance, valuable insights into the intrinsic relationships can be presented within the data [31]. In this study, a random forest regression analysis was utilized to evaluate the significance of 28 factors originating from the socioecological model. This method facilitated the ranking of variables according to their respective impact on the predictive efficacy of the model, thereby offering valuable insights into the primary determinants within the socioecological framework. Third, we integrated the top 50% of important factors from the random forest regression analysis into a univariate generalized linear model to examine the association between the study variables and the acceptance level of ACP. Fourth, study variables that exhibited statistical significance at the P < 0.05 level in the univariate generalized linear model were included in the multivariate generalized linear model for further examination. The variance inflation factor (VIF) test was utilized to identify collinearity, with a max VIF of 2.41, indicating no collinearity.

All two-tailed P < 0.05 was deemed to be statistically significant. All statistical analyses were executed utilizing Stata version 16.0 (StataCorp, College Station, TX, USA) and R software version 4.3.0.