**3. Research Methods**

### *3.1. Unsafe Behavior Data Analysis Based on Roof Accidents*

The number of roof accidents is very large, it is very difficult to integrate the detailed case data of each accident. In order to ensure the scientific nature of the study, the selection of cases follows two principles: accident integrity and case authority. The main sources of accident cases are Internet, accident analysis report, coal mine typical accident compilation, and so on. The application of data analysis methods has the following advantages: the large amount of data, rather than the selection of random samples, greatly reduces the impact of random events on the overall conclusion; the intrinsic characteristics associated with behavior, not just superficial causality, can be studied; the processing speed is fast, and valuable information can be quickly obtained in a short period of time [23].

Unsafe behavior refers to the behavior of the person who has caused the accident or may cause the accident. There are many causes of unsafe behaviors, including individual factors, psychological factors, organizational factors, environmental factors, and so on. Different researchers have different views on the classification of unsafe behaviors. Unsafe behaviors in a coal mine mainly refer to the "three disobeying" of coal mine safety production [24]. "Three disobeying" is the general term of

disobeying command, disobeying operation, and disobeying labor discipline. The elimination of "three disobeying" has always been an important issue in the safety managemen<sup>t</sup> of all industries, especially coal mining enterprises and other high-risk industries. In order to fundamentally explore the objective rule of unsafe behaviors and the complex relationship between its internal factors, many unsafe behaviors need to be objectively analyzed. Previous coal mine accidents can provide enough data for the analysis of miners' unsafe behaviors, but the unsafe behaviors that lead to different types of accidents are different in nature. To eliminate the heterogeneity of a large number of unsafe behavior data, we selected unsafe behavior leading to roof accidents in coal mines as the research object [25].

A total of 331 roof accidents were collected from 1983 to 2014 in China, including 8 major accidents, 159 major accidents, and 164 general accidents. All unsafe behaviors of miners causing roof accidents are collected from the accident report. For example, a roof accident occurred in a mine in Pingdingshan city, Henan province, which directly resulted in a death, several serious injuries, and a total economic loss of about RMB 585,000. According to the time of this accident, four unsafe behaviors could be analyzed successively, expressed as follows: "the top plate was out of the slag and was not withdrawn in time," related to the coal miner; "there are no timely measures to prevent potential safety hazards," related to the field commanders; "there are no perfect operation procedures or safety technical measures established," related to the middle managemen<sup>t</sup> staff; and "the arrangemen<sup>t</sup> of roadway is in a steep-inclined coal seam," related to the senior managemen<sup>t</sup> staff.

According to the data boundary divided by seven dimensions, 1215 data points were obtained. The data of miners' unsafe behavior was transformed from unstructured text records into structured data. A preliminary single-dimensional statistical analysis of the data points was performed, as shown in Table 1.


**Table 1.** Statistics of a single dimension.

Notes: RL: Risk Level; P: Position; UA: Unsafe action; BI: Behavior individual; BT: Behavior trace; BP: Behavior property; T: time.

According to the description of seven dimensions of miners' unsafe behaviors, interactive analyses can be performed on any two or more of them. The interaction analysis between different variables can map out different practical meanings, the purpose of which is to explore the deep regularity of the unsafe behavior data of miners and to provide the basis for determining the targeted intervention nodes. Through the Apriori algorithm [26], unsafe action is the consequent of the study, and other dimensions are antecedents. The relationship between information and unsafe action in each dimension is discussed (the association between specific unsafe action and time dimension is not obvious enough and is not considered for the time being), as shown in Table 2.



Support is the ratio of the number of consequents and antecedents to the total data set in the database, and the probability of their occurrence is determined. Confidence is the ratio of the support for the occurrence of a consequent and an antecedent to the support for the consequent. It is used to denote the probability of an antecedent derived from an association rule under the condition of a consequent. Lift is the ratio of the confidence of "antecedent → consequent" to the support of the consequent. It reflects the size of the consequent influenced by the antecedent [27]. Minimum support and minimum confidence were 8% and 30%, respectively, and this information was used to obtain effective strong association rules. The results of the analysis are shown in Figure 2.

### *Int. J. Environ. Res. Public Health* **2019**, *16*, 422

**Figure 2.** Association rule results of the miners' unsafe behavior.

### *3.2. The Miners' Unsafe Behavior Targeted Intervention Node Establishment*

The key work behaviors are determined by the frequency, work position, work time, risk level, etc. caused by the "three disobeying" in different types of work. Following the principles of reliability, measurability, controllability, observability, etc., according to the comprehensive distribution characteristics of each dimension of "three disobeying" for various types of work, several "three disobeying" behaviors are selected as the key behaviors of various types of work. These key behaviors are targeted at the first type of intervention nodes for unsafe behavior intervention. In the process of practical safety management, it is expected to improve the effect of safety managemen<sup>t</sup> by focusing on this kind of key work behavior. It is expected that behavior correction can be achieved by interfering with these key work behaviors, and the incidence of unsafe behavior can be greatly reduced [28].

A single-dimensional feature-targeted intervention node is based on the distribution characteristics and proportion of each dimension to guide the intensity of intervention and the allocation of managemen<sup>t</sup> resources in the process of security management. In terms of time, for example, more unsafe behavior happened in January, March, and August. Safety managemen<sup>t</sup> can target these months to strengthen the observation and intervention of unsafe behavior and can directly and effectively reduce the number of unsafe behaviors. Based on the association analysis of each dimension of unsafe behavior, the multi-dimensional characteristic-targeted intervention node was located according to their deep connection and rule, which can provide guidance and direction for practical safety management. A list of three types of intervention nodes based on miner's unsafe behavior-targeted intervention was finally formed through data and theoretical analysis for use in this study, as shown in Table 3.



### *3.3. The Targeted Intervention Node Evaluation*

The purpose of the evaluation of targeted intervention nodes is to verify the effectiveness of each intervention node and the application value of targeted intervention nodes in the intervention of unsafe behavior of miners. In this study, the structural equation model (SEM) was constructed to evaluate the targeted intervention nodes [29]. As a general framework of statistical analysis, structural equation model (SEM) is widely used in data analysis [30,31].

### 3.3.1. Variable Division and Formulate Hypothesis

In order to evaluate the effectiveness of three unsafe behavior-targeted intervention nodes, another latent variable "intervention effect" was introduced. Unsafe behavior incidence (I1), safety managemen<sup>t</sup> efficiency (I2), and safety climate improvement degree (I3) were selected as the observed variables of latent variable "intervention effect" [32–34]. K1–K5, S1–S6, and M1–M7 in the list of the targeted intervention nodes were the observed variables of the "key work behavior node," the "single-dimensional feature node," and the "multi-dimensional feature node," respectively.

In the process of coal mine safety management, in order to effectively reduce the incidence of roof accidents and improve the targeting of workers' behavior intervention, this paper analyzes and summarizes three kinds of targeted intervention nodes: key work behavior, single-dimensional features, and multi-dimensional features. The practical significance of the three factors and their impact on the effect of intervention were analyzed, and the following hypotheses were formulated:

H1. Key work behavior positively influences the intervention effect.

H2. Single-dimensional features have a direct influence on the intervention effect.

H3. Multi-dimensional features positively influence the intervention effect.

These three types of nodes also affect each other, and the targeted intervention on a single class of nodes will also have an impact on other types of intervention nodes. For example, a single-dimensional feature intervention node may result in a change in the incidence of key work behavior, so the following assumptions are formulated:

H4. Single-dimensional features positively influence key work behavior.

H5. Multi-dimensional features have a direct influence on key work behavior.

H6. Multi-dimensional features have a direct influence on single-dimensional features.

### 3.3.2. Construct Model and Questionnaire Design

Based on the following hypotheses, an initial model of miners' targeted unsafe behavior intervention node evaluation was established, as shown in Figure 3.

**Figure 3.** Initial model of miners' targeted unsafe behavior intervention node evaluation.

According to the latent variable and observation variable, the questionnaire was designed, and the experience of other scholars was consulted and verified by experts. The questionnaire's results were compiled with the five-point Likert scale (1–5 represents total disagreement, basic disagreement, partial agreement, basic agreement, and complete agreement, respectively) [35]. This assessment method produces better data distribution. The questionnaire consists of three parts: (1) general information and background information about the interviewees; (2) questions about the role of various targeted intervention nodes in the implementation of behavior interventions; (3) questions about the intervention effect on unsafe behavior through targeted intervention nodes.

A random sampling method was used to survey the employees in a coal mine of the Pingdingshan Coal Industry Group [36]. This coal mine is located in the northwest of Pingdingshan District, Henan province. It was founded in 1956 and put into production on 31 December 1958. It included 4468 registered workers, 376 managemen<sup>t</sup> personnel, 281 professional and technical personnel, 12 senior titles, and 116 intermediate titles. The questionnaire survey was conducted in this coal mine where several roof accidents resulting in huge casualties and property losses had occurred. A total of 260 questionnaires were sent out, and 248 responses were received. The recovery rate was 95.38%. Among the collected questionnaires, 237 were valid, and the effective rate was 95.56%. Among the interviewees, there were 42 people with a bachelor's degree or above, accounting for 16.15%, and 185 coal miners, accounting for 71.15%; the number of safety administrators at all levels was 38, accounting for 14.62%, and 94, accounting for 36.15%, had worked for more than 15 years.

In order to ensure the applicability and validity of the data in the questionnaire, reliability analysis and validity analysis of the questionnaire data were carried out [37]. In this study, Cronbach's α was used to measure the reliability of setting latent variables [38]. Generally, the higher this coefficient, the higher the reliability. In the exploratory study, results of the questionnaire with a reliability up to 0.7 were acceptable. The data from the questionnaire were input into SPSS 17.0 software (International Business Machines Corporation, Armonk, NY, USA), and the calculated Cronbach's α coefficient showed that the questionnaire data had good reliability, as shown in Table 4.


**Table 4.** Reliability test results for latent variables.

Validity analysis tests the degree to which the questionnaire reflects the objective reality. Analysis of questionnaire validity was performed via KMO and the Bartlett sphericity test [39]. Under the standard condition, when the KMO is greater than 0.7 and the concomitant probability of the Bartlett sphericity test is less than significance level 0.001, the questionnaire has high validity. The calculated result was KMO = 0.837 > 0.7, and the significance probability of the statistical value of Bartlett sphericity test was less than 0.001, indicating that the index data were normally distributed. This shows that the questionnaire has high validity. According to the above analysis, the questionnaire has good reliability and validity, and the data obtained are suitable for factor analysis.
