Abstract: Maritime accident refers to an unexpected and abnormal event on a ship, which often leads to various consequences, such as casualties, ship damage and property loss. Due to the complex environment of the sea lanes, the evolution of maritime accidents is affected by a variety of factors, often leading to consequences of varying degrees of severity. Although the international maritime authorities have made great efforts for transportation safety, but the risk of accidents in the sea transportation lanes still exists. Therefore, it is important to explore the factors affecting the severity of accidents in sea lanes in order to provide timely and effective emergency response and to reduce the damage caused by accidents.
A review of the relevant literature shows that numerous influencing factors on the severity of maritime accidents are currently being explored. However, considering that maritime accidents may occur under different conditions, this leads to heterogeneity in the nature of accident dynamics, as well as to the fact that some specific factors influence the consequences of accidents to different degrees or even in opposite directions. Existing approaches to studying the severity of maritime accidents always mask some of these underlying relationships, resulting in little effect on reducing unobserved heterogeneity. For this reason, this paper analyzes the factors affecting the severity of maritime accidents on the basis of existing studies, taking full account of the heterogeneity of maritime accident data.
In this paper, a two-step approach combining latent class clustering and mixed logit modeling is proposed to explore the influencing factors on the severity of maritime accidents. First, latent class clustering is used to classify maritime accidents into different homogeneous clusters; Then, a mixed logit model is used to model each cluster and the full data separately to analyze the influence of the influencing factors on the severity of accidents. The important variables of the modeling are explained through a combination of the estimated parameters and the associated marginal effects. In addition, the hidden influencing variables are revealed by comparing the results estimated by the mixed logit model using clustering and without clustering (full data).
An empirical study is conducted with data information extracted from accident investigation reports released by the China Maritime Safety Administration. The computational results show that: 1)Based on the estimated results, some important factors affecting the accident severity can be found. For example, the accident type of self-sinking, compared with other accident types, has a significant effect on the accident severity of slight severity, severe and very severe, indicating that this variable is a very important variable for the influence of accident severity. 2)It can be found that the analysis of maritime accidents based on heterogeneous data may hide some important influencing factors. For example, liquid cargo ships, ship ages 6-10 and 11-15 years, normal loading conditions, winds of 5-7, poor navigational environment, and low vessel traffic are not statistically significant in the full data model. However, based on the modeling in clustering these variables affect minor severity accidents to different degrees. 3)Clustering-based modeling can reveal changes in the probability of accident severity of influencing variables on different specific situations. For example, the type of vessel involved in an accident is a fishing vessel, the probability of causing a minor severity accident based on the full data model will increase by 3.5%, while the probability of causing a minor severity accident based on the clustering 1, clustering 2, and clustering 4 models will increase by 11.3%, 12.6%, and 5.7%, respectively. 4)The clustering model can even reveal differences in the direction of the influence of certain variables on the severity of an accident. For example, less vessel flow is shown to reduce accident severity in the Cluster 3 model, while the opposite result is shown in the Cluster 4 model.
In summary, a two-step approach combining latent class clustering and mixed logit modeling can have great potential in explaining the sources of heterogeneity, and the results from specific estimation of the factors affecting the severity of maritime accidents can support decision-making in maritime emergency response.

Key words: waterway transportation; heterogeneity; accident severity; latent class clustering; mixed logit

摘要： 海上通道事故的发生及演化受到多种因素的影响,常常导致不同严重程度的后果。考虑到海上事故数据固有的未观测到的异质性,以往研究的海上事故严重程度的方法总是会掩盖其中的一些潜在关系。本文提出了一种结合潜在类别聚类和混合logit模型的两步方法来探究海上事故严重程度影响因素。首先,采用潜在类别聚类将海上事故划分为不同的同质聚类;然后,采用混合logit模型分别对每一个聚类及全数据进行建模,分析影响因素对事故严重程度的影响。以从中国海事局发布的事故调查报告中提取的数据进行实证研究。计算结果表明,将海上事故事先分割成相对同质的聚类有助于揭示隐藏在全数据模型中的一些重要因素;另外也证明了所提出的方法能够在解释异质性来源方面具有很大的潜力;具体估计得出的影响海上事故严重程度因素的结果可以为海上应急响应决策提供支持。

关键词: 水路运输, 异质性, 事故严重程度, 潜在类别聚类, 混合logit