A hybrid STAMP-fuzzy DEMATEL-ISM approach for analyzing the factors influencing building collapse accidents in China

To explore the factors influencing recent construction collapse accidents, this study utilizes a sample of 355 reports on building collapse accidents from 2012 to 2022. The investigation employs the systems-theoretic accident modeling and processes (STAMP) model to retrieve 22 key causal factors of accidents from the physical, operational, managerial, and supervisory layers. Subsequently, an improved decision-making trial and evaluation laboratory (DEMATEL)-interpretive structural modeling (ISM) method is used to analyze the relationships and strengths of these influencing factors, providing a comprehensive understanding of the logical connections between the causes of building collapse accidents. The results indicate that the deep-rooted causes of building collapse accidents are primarily lax safety management at the enterprise level and the exchange of interests at the government regulatory level, which in turn affect workers at the operational level and the physical aspects of accidents on-site. Furthermore, integrating the STAMP model and the triangular fuzzy DEMATEL-ISM model overcomes the limitations of the traditional STAMP model, allowing for a more focused identification of key factors.

of 98 people.At 7:14 p.m. on March 7, 2020, a building that was part of the Jiaxin hotel collapsed in Quanzhou, Fujian Province, resulting in 29 deaths and 42 injuries, with a direct economic loss of 57.94 million yuan.The occurrence of building collapse accidents will have not only a great impact on economic development but also a negative impact on social stability 8 .
To effectively prevent building collapse accidents and ensure the safety of construction projects, it is crucial to analyze the causal factors of these accidents and to examine the interrelationships between these factors.To do so, we improve the systems-theoretic accident modeling and processes (STAMP) accident causality model based on the needs of analyzing and investigating building collapse accidents.The modified STAMP model is utilized to systematically analyze a large number of building collapse accidents through modular solidification.Subsequently, an improved decision-making trial and evaluation laboratory (DEMATEL)-interpretive structural modeling (ISM) method is employed to analyze the interrelationships between modules within the building collapse accident system.Ultimately, a hybrid model that is capable of providing a comprehensive analysis of the building collapse accident system as a whole is formed.

Literature review
The analysis of accident influencing factors is mostly based on accident cause theory 9,10 .The development of accident cause models involves single-factor, double-factor, multifactor and complex systems 11 .A single-factor model includes the theory of accident frequency tendency proposed by Farmers and Chambers.This theory holds that an accident is caused by some personnel who are characterized by frequent accidents, and it completely denies the impact of the objective environment and equipment on the accident.The main theories of two-factor models are the accident causal chain theory proposed by Heinrich (1941) and the accident causal chain theory proposed by Byrd (1966) 12.13 .The theories of these models hold that the occurrence of accidents is influenced by not only human factors but also management factors.In multifactor models, the accident causal chain theory proposed by Surry (1969) adds institutional and social factors 14 .In recent years, popular accident cause theories based on complex systems have included the STAMP model proposed by Leveson (2003) and a nonlinear accident cause model based on the system safety-functional resonance analysis method (FRAM) proposed by Hollnagel (2012) 15,16 .At present, the analysis of construction accidents mainly targets large-height falling accidents, and few scholars have analyzed the causes of collapse accidents.Zhang et al. (2019) used four contemporary popular accident causal models, i.e., the STAMP, AcciMap, human factors analysis and classification system (HFACS), and 2-4 models, to analyze a typical collapse accident, and they found that in accident analysis, the four models have different focuses 17 .Kale et al. (2021) analyzed 723 accidents caused by ditch collapses and pointed out that more than 50% of the collapse accidents were caused by a lack of protective measures 18 .Huang et al. ( 2021) discussed China's construction safety under COVID-19 and analyzed collapse accidents from five perspectives, i.e., contractors, organization and management, technical methods, participants and interactive feedback, from the perspective of the STAMP model 19 .Yan and Kim (2018) proposed a stage model of building collapse accidents and divided building collapse accidents into an incubation period, an outbreak period, a transmission period and a final stage 20 .Okunola (2021) interviewed 42 people using a random sampling method and found that the main causes of building collapse accidents in Nigeria were institutional failure and human factors 21 .
Based on the literature analysis above, many causality models applied in accident analysis tend to have their own focuses.For instance, the HFACS model primarily emphasizes the analysis of human factors, while the 2-4 model places greater emphasis on accidents caused by management failures.This difference in focus leads to a lack of comprehensiveness and detail in the analysis of accident causes 22 .The STAMP model represents the pathways of accidents using a systemic network, which aligns with the complexity of accidents.It can provide a detailed analysis of individual accidents.However, there are still two main shortcomings.First, due to the lack of a modularized classification of causal factors, its efficiency in analyzing a large number of accident statistics is not high.Second, it uncovers too many causes, making it challenging to establish the logical relationships and strengths of the mutual influences among factors.To address these limitations, we conduct a statistical analysis of a large number of building collapse accidents.The modularization of the STAMP model is solidified into four aspects: the physical, operational, managerial, and supervisory layers of building collapse accidents.This solidification enables a rapid analysis of the causes of multiple accidents.Furthermore, we incorporate the fuzzy DEMATEL-ISM method with the STAMP approach to more effectively identify the key factors contributing to accidents.

Methodology STAMP
Compared with traditional analysis methods for accident cause chains, the advantage of the STAMP model lies in the fact that it fully explains the impact of interaction defects between complex system components on system safety 23 .The main reason is that the model has three basic structures: security constraints, a hierarchical safety control structure and a process model.The hierarchical safety control structure can be established based on the confirmation of hazards and safety constraints 24 .The structure has top-down characteristics.From the bottom, the safety constraints and inappropriate control behaviors of each layer in the entire hierarchical safety control structure can be identified in turn so that a dangerous defect control process can be confirmed.The interaction relationship between each control level is represented by the process model.The general process model includes a controller and control process, an output and feedback process between them, process input and output, and the interference of external information.The deficiency lies in the fact that the STAMP model emphasizes the depth and breadth of accident analysis.Therefore, the analysis requires practitioners to have very high professional background knowledge.Additionally, the analysis process is time consuming and inefficient and lacks modular cause factor classification, which is not conducive to the statistical analysis of a large number of accidents 25 .Therefore, we solidify the STAMP model and build an accident control model from four levels, physics, operation, management and supervision, to make the model more conducive to large-scale accident analysis.

Fuzzy DEMATEL
The DEMATEL is a method of systematic factor analysis using graph theory and matrix tools 26 .Through an analysis of the logical relationship and direct influence matrix among the factors in a system, the influence degree of each factor on other factors and the affected degree can be calculated to calculate the center degree and cause degree of each factor.This method makes full use of the experience and knowledge of experts to deal with complex social problems, especially systems with uncertain relationships between factors that are more effective.

ISM
ISM is an analytical method that explains the logical structure between influencing factors.By constructing an adjacency matrix between influencing factors, the adjacency matrix is added to the identity matrix according www.nature.com/scientificreports/ to Boolean operation rules.After continuous iteration, it is equal to the initial adjacency matrix and forms the final reachability matrix.The reachability set, antecedent set and common set among the influencing factors are obtained using the reachability matrix, and finally, the hierarchical structure of the system is divided.

STAMP-Fuzzy DEMATAL-ISM combination
The STAMP model is an accident cause model for complex systems that can clarify the accident causes within and between organizations from the perspective of control.However, when using the STAMP model to analyze accidents, the influence intensity of accident influencing factors cannot be obtained, which makes it impossible to grasp the key points when formulating relevant preventive measures.To overcome this defect, we introduce the DEMATEL-ISM method to analyze the interaction between factors in complex systems.However, the direct impact matrix in the DEMATEL-ISM model is usually divided into a direct relationship or no direct relationship (1 or 0), which is not suitable for accurate scoring by experts 27 .In some collapse accidents, there is no clear division between the causal factors; that is, their boundaries are ambiguous.Therefore, we introduce triangular fuzzy numbers to improve the method of obtaining the direct influence matrix in the DEMATEL-ISM model to make the obtained results more representative.The specific idea is shown in Fig. 3. Step

Identify influencing factors
The STAMP method is used to determine the factors of a building collapse accident S = {S 1 , S 2 , . . ., S n }.
Step 2 Analyze the effects between indexes The interaction strength between indicators is difficult to directly represent using data.Therefore, this paper uses an expert scoring method to evaluate the influence degree between indicators.The expert scoring value is divided into five levels: very high (VH), high (H), low (L), very low (VL), and no effect (NE) (Table 1).Due to the subjectivity and fuzziness of expert scoring, triangular fuzzy numbers are used for transformation.Based on the correspondence between linguistic phrases and the membership function of triangular fuzzy numbers provided by Kuzu (2021), the five levels of expert scoring are transformed into triangular fuzzy numbers 28 .For the evaluation index, it is assumed that there are m experts scoring.The triangular fuzzy number corresponding to the influence degree of the i-th index evaluated by the k-th expert on the j-th index is Step 3 Construct and normalize the direct impact matrix We use an improved converting fuzzy data into crisp scores (CFCS) method to convert triangular fuzzy numbers into clear numbers.The clear value of the influence degree of the i-th index on the j-th index can be obtained based on the four-stage CFCS algorithm: (1) Standardization Table 1.Transformation relationship of triangular fuzzy numbers.

Influence degree Triangular fuzzy number
Very High (VH) (0.75, 1, 1) The The standardized parameter p is calculated and used to standardize the matrix Z and obtain the matrix Q: Step 4 Obtain the comprehensive influence relation matrix where I is a unit matrix.
Step 5 Conduct comprehensive impact relation matrix analysis The influence degree ( r i ), affected degree ( c j ), center degree ( O i ) and cause degree ( P i ) of each factor are calculated.
Step 6 Calculate overall impact relationship matrix U Step 7 Calculate reachability matrix K First, we calculate threshold λ.The value of λ is the sum of the mean and standard deviation of all elements of matrix F. The value rule of index k ij in the i-th row and the j-th column of the reachability matrix is as follows: where k ij = 1 indicates that factor s i influences s j and k ij = 0 indicates that factor s i has no influence on s j .
Step 8 Computation of the reachability set, antecedent set, and common set If the common set satisfies C(s i ) = R(s i ), the influencing factor s i contained in T(s i ) is the current lowest element.Then, the rows and columns corresponding to the elements contained in the bottom layer can be deleted from the reachability matrix.The steps above are repeated until all factors are stratified.
Step 9 Draw an ISM multilevel ladder diagram based on the stratification of influencing factors

Data sources
Based on the websites of the MHURDC, safety management offices, and provincial work safety supervision and administration offices as well as the portals of relevant departments searched on the internet, 355 investigation reports on building collapse accidents from 2012 to 2022 were statistically analyzed [29][30][31] .Each report included the basic situation of the accident, the accident process and rescue, the accident cause, an accident responsibility analysis, handling suggestions, and accident prevention and rectification measures.
From the perspective of geological factors, these accidents covered a total of 28 provincial-level administrative regions (over 80% of the geographical area of China) from 2012 to 2022.The urban distribution is concentrated in the eastern part of China, as the eastern region is more developed.From the perspective of the degree of casualties in accidents, there were 9 extremely large accidents (with a death toll of ≥ 10 people), 111 major accidents (with a death toll of ≥ 3 people), and 235 general accidents (with a death toll of < 3 people).The types of projects include both construction operations (new construction, expansion, renovation projects) and those that occur during use (maintenance, demolition, renovation, etc.).The types of collapse accidents also include typical collapse accidents such as earthwork collapses, scaffolding collapses, formwork collapses, demolition project collapses, and building collapses.

Using the STAMP model to discover the causal factors of collapse accidents
Construction engineering safety control is the basis for analyzing accident causes using the STAMP model.Based on the actual operation of China's construction projects, the safety control structure is divided into a supervisory layer, a managerial layer, an operational layer, and a physical layer (Fig. 4).
(1) The supervisory level includes government supervision departments and supervision units.A government supervision department controls the safety management of the construction industry as a whole through legislation and by issuing norms and standards.A supervision unit accepts the entrustment demand of a construction unit and carries out safety production supervision and hidden accident danger rectification at a construction site.(2) The managerial level includes all units participating in a construction project and adopts an organizational form that establishes a project department to make all participating parties coordinate with and restrict each other and to control the safety and progress of the project.A construction unit is the investor and manager of a construction project, which realizes the constraints by describing the needs to other participating units and receiving feedback from them.The survey and design units accept the entrustment needs of a construction unit, and they provide it with true and accurate survey and design documents.The construction unit controls the whole process of project progress, quality management and safety management at a construction site.(3) The operational layer includes the operators and team leaders at a construction site.The operators accept the instructions and constraints of a construction unit and construct based on the construction organization plan and safety rules and regulations 32 .At the same time, the construction unit restricts the operators through on-site inspection, training and education.(4) The physical layer includes mechanical equipment, materials and the construction environment 33 .Frontline employees may encounter interference from various factors in the physical layer during actual construction operations.
By analyzing the investigation reports on 355 building collapse accidents and combining them with the control structure chart of building collapse accidents, starting from the physical layer, operational layer, managerial layer and supervisory layer, 22 main influencing factors as well as the level and main performance of each factor are finally obtained (Table 2).

Using the triangular fuzzy DEMATEL model to analyze the interrelationships between the causal factors of collapse accidents
We invited seven experts to score the influence strength among the 22 factors and constructed a corresponding evaluation score matrix.Based on the corresponding relationship in Table 1, a triangular fuzzy evaluation www.nature.com/scientificreports/variable matrix was obtained, and the results were calculated based on solution steps (2)-( 5) of the triangular fuzzy DEMATEL method.The initial direct matrix Z (Appendix 1) of building collapse accidents was calculated using formulas (1)-( 8).The comprehensive action matrix F of building collapse accidents was calculated using formulas ( 9)-(11) (Appendix 2).
The influence degree, affected degree, center degree and cause degree of the factors causing a building collapse accident were obtained using formulas ( 12)-(15) (Table 3).The higher the centrality value is, the greater the impact of representative factors on building collapse accidents, and vice versa.There are positive and negative cause degrees.A positive cause degree indicates a stronger influence on other causal factors, whereas a negative cause degree indicates that the causal factor is strongly affected by other factors.
Based on the data in Table 3, we used the quadrant determination method to draw a causal relationship diagram with centrality and causality as the horizontal and vertical axes, respectively, and we marked the position of each element.The intersection of the horizontal and vertical coordinates is O (1.71, 0) (1.71 is the average of the sum of the center degrees).The 22 factors identified were plotted in four quadrants based on the quadrant determination method described above (Fig. 5).

Hierarchical analysis of the causal factors of building collapse accidents based on ISM
Based on the analysis of the causality and centrality of building collapse accidents using the DEMATEL method, ISM was further used to explore the hierarchical structure and overall influence relationship of the causal factors of building collapse accidents.Using formulas ( 16)-( 17), we calculated the reachability matrix K of the causal factors of building collapse accidents (Appendix 3).
Based on the reachability matrix K of the causal factors of building collapse accidents, we used formulas ( 17)- (19) to calculate reachability set R(S i ), antecedent set A(Si) and common set C(Si).The reachability set is the set of the 1 row where the causal factor is located, the antecedent set is the set of the 1 column where the causal factor is located, and the common set is the intersection of the reachability set and the antecedent set (Table 4).
Using the R(Si) = C(Si) criterion, the first layers of the causal factors of building collapse accidents are S 6, S 7 , S 9 , and S 11 .The first-level factors are removed from the reachability set, antecedent set and common set.Then, based on this criterion, the second-level factors of the causal factors of building collapse accidents are S 1 , S  S 5 , and S 8 .By iteratively repeating this method, a multilevel ISM model of the causal factors of building collapse accidents is finally obtained (Fig. 6).

Result analysis
By using the DEMATEL method, the influence degree, affected degree, cause degree and center degree of the factors causing building collapse accidents are determined, and the key factors are identified.In addition, a multilayer hierarchical structure model of building collapse accidents is constructed using the ISM method to reveal the direct, indirect and deep-rooted causes of building collapse accidents.We obtain the following results: (1) Analysis of the influence degree and affected degree.
As shown in Table 3, the most influential factor is loose safety management (S 17 ) at the managerial level, followed by untimely rectification feedback (S 18 ), ineffective inspection work (S 19 ), benefit exchange behavior (S 20 ), and dereliction of duty in supervision (S 21 ) at the supervisory level.In addition, the factors insufficient safety awareness (S 12 ) and insufficient safety investment (S 16 ) at the managerial level are highly influential.The most affected factor is error of judgment (S 11 ), followed by illegal command (S 10 ) and insufficient safety training (S 13 ).Poor health (S 9 ), illegal work (S 7 ) and extreme weather (S 5 ) also have a high influence degree.These results show that in building construction, the control structures that are easily affected are mainly at the operational layer and the physical layer.These factors are mainly affected by corporate management and government regulation.By strengthening the supervision of various government departments and improving the level of enterprise safety management and investment in safety production, managers can increase the safety management of construction sites, discover potential safety hazards in time, and employ front-line operators in accordance with safety regulations to ensure the standardization and stability of the support system.
(2) Analysis of the center degree and cause degree.
The top three factors of the center degree are loose safety management (S 17 ), illegal command (S 10 ) and untimely rectification feedback (S 18 ).These three factors are the most important in the causal system of building collapse accidents and play the most obvious role in the occurrence of such accidents.Therefore, to prevent the occurrence of building collapse accidents, we need to focus on controlling the occurrence of these three factors.Figure 5 shows that the center degree and cause degree of factors S 17 , S 18 and S 19 in the first quadrant are high, indicating that they are the key factors affecting building collapse accidents and need to be treated and improved first.Factors S 5 , S 7 , S 9 -S 11 and S 13 in the second quadrant have a high center degree but a low cause degree.They are supportive factors and play an auxiliary role in the model.There are five factors in the third quadrant, S 1 -S 4 and S 6 .Both their cause degree and centrality are low, they are independent factors, and they are vulnerable to other factors.Factors S 8 , S 12 , S 14 -S 16 and S 20 -S 22 are in the fourth quadrant.Although the center degree of these factors is slightly lower, the cause degree is higher, which means that they are core problem elements.Although they are not key factors affecting building collapse accidents, they have a strong impact on other factors.Therefore, these factors still need to be considered.
(3) Analysis of the ISM hierarchy.Based on Fig. 6, the causal factors of building collapse accidents include surface causes, intermediate causes and deep-rooted causes, which are divided into eight levels.Among them, S 6 , S 7 , S 9 and S 11 are the surface causes of building collapse accidents.The factors of other layers are directly or indirectly related to these four factors, which shows that these problems can be solved by improving the factors of other layers.For example, we can   ) and dereliction of duty in supervision (S 21 ).In addition, the interaction of S 12 , S 13 and S 16 will strengthen the influence of the intermediate layers on the surface layer.Although the causal factors of the intermediate layers are not direct influencing factors or deep-rooted influencing factors, they also indirectly affect the occurrence of building collapse accidents, which cannot be ignored.
The causal factors (layers 6-8) located in the deep-rooted layers of the system include four factors: S 17 , S 18 , S 19 and S 20 .The deep-rooted influencing factors affect the surface influencing factors by affecting the intermediate factors.The lowest of the deep-rooted influencing factors involve two aspects: loose safety management (S 17 ) at the managerial level and benefit exchange behavior (S 20 ) at the supervisory level.These two factors jointly impact the timeliness of rectification feedback from government departments (S 18 ) to affect the intermediate influencing factors.In addition, S 17 will also directly affect the unreasonable construction plan at the intermediate layers (S 14 ), and S 20 will directly affect dereliction of duty in supervision (S 21 ) to exert an impact on the surface causal factors.

Discussion
With the continuous innovation of technology and the continuous development of the economy, in terms of both the building volume and technological complexity, the risks in the construction process have increased significantly compared to any previous period, resulting in higher levels of accidents.The increase in the complexity of accident causes has led to deviations in the applicability of traditional accident cause theories.Traditional accident cause theories divide the causes of building collapse accidents into human, equipment, environmental and managerial factors 34,35 .Although this method can comprehensively identify the causal factors of accidents, it is too loose for controlling accidents.The STAMP method based on system theory is based on the control structure, which allows a site manager to determine the cause of a collapse accident more logically.
For each system safety constraint, determining whether the high-level system has assigned responsibility for executing the constraint to the lower components in the security control structure or whether the high-level system has sufficient control to ensure that the assigned security constraint is implemented in the lower components is the key to analyzing the cause of the accident 36 .Through the dynamic perspective of the STAMP model, it was found that the lack of communication and coordination among various levels of collapse accidents as well as feedback loops (safety inspection reports, maintenance feedback, etc.) are key factors leading to incorrect decision-making or inappropriate control 37 .The upper level is not aware of the true situation of the lower level.As a result, it is impossible to determine whether the system's safety margin is sufficient, and as a result, local adjustments made by superiors are no longer effective.
In the study of building collapse accidents, most of the research results point to worker violations as the direct cause of such accidents 38,39 .Clearly, this problem is also a key point that we obtain from the DEMATEL-ISM model.However, the more organizational reasons underlying this problem are the key to preventing collapse accidents.Loose enterprise safety management will affect managers' failure to pay attention to the management and safety training of front-line personnel.In addition, enterprises will not pay attention to investing in safety.Another problem that is easy to ignore is a problem in the regulatory process, especially the exchange of interests between government departments and enterprises 40 .
This interest exchange includes two forms.First, enterprises influence the government's regulatory decisions through their own influence.However, based on the actual situation in China, local governments still play a leading role in the exchange relationship between local governments and enterprises, which can lead to laziness in enterprise safety work and dependence on government regulatory authorities.However, with the deepening of reform and the acceleration of market-oriented processes, the influence of enterprises is also increasing.Second, not all levels of government departments have established a complete management mechanism for the safety supervision of construction projects 41 .The combination of multiple departments, such as safety production supervision and management, quality and technical supervision, environmental protection, land and resources, urban and rural planning, and fire protection departments, makes it difficult to form an effective and diverse collection, resulting in management loopholes and delayed supervision of hidden danger rectification.
To prevent and control construction collapse accidents, targeted preventive and control measures should be proposed by considering the attributes of various levels of factors.Resolving problems in the operational layer relies on other factors.For instance, in addressing the issue of front-line workers engaging in unauthorized operations, the managerial layer should organize safety education and training programs to enhance these workers' operational competence and safety awareness.Regarding problems in the physical layer, the supervisory layer should intensify technological innovation and research and development efforts to enhance construction process standards.Each unit within the managerial level should also implement the system of safety production responsibility to fundamentally ensure the level of project safety management.The supervisory layer should enforce organizational construction standards, strengthen law enforcement efforts, and intensify penalties for violations to address the deep-rooted causes of accidents.

Conclusions
Based on investigation reports on 355 typical building collapse accidents, this paper uses the STAMP method to extract the causes of such accidents from the four control levels of the physical layer, field operational layer, enterprise managerial layer and government supervisory layer.Through a literature analysis and questionnaire

Figure 1 .
Figure 1.Safety accident statistics for housing and municipal engineering in China, 2000-2022.Data source: Ministry of Housing and Urban-Rural Development of China (MHURDC).

Figure 3 .
Figure 3. Analysis logic of a construction accident based on the STAMP-DEMATEL-ISM method.

Figure 4 .
Figure 4. Control structure of a building collapse accident.

CFigure 5 .
Figure 5. Quadrant distribution of the center degree and cause degree of building collapse accidents.

Table 2 .
Analysis of the factors causing building collapse accidents based on the STAMP model.S ij is the symbol for the causal factors of building collapse accidents.

Table 3 .
Influence degree, affected degree, center degree and cause degree of building collapse accidents.

Table 4 .
Reachability set, antecedent set and common set of building collapse accidents.ISM hierarchy of the causal factors of building collapse accidents.organizesafety education and training through enterprises to ensure that staff are familiar with safe operation technologies and procedures and have sufficient safety knowledge and safety awareness to prevent accidents caused by illegal operations.We can also prevent accidents by strengthening the supervision of safety management departments to ensure that staff operate or command based on the specifications.There are 14 factors in the intermediate layers (layers 2-5) of the system, namely, S 1 -S 5 , S 8 , S 10 , S 12 -S 16 , S 21 and S 22 .Among them, the lowest contributing factors are insufficient safety awareness (S 12 ), inadequate safety training (S 13 ), insufficient safety investment (S 16