Paper: Occupational accidents model based on risk–injury affinity groups

Occupational accidents model based on risk–injury affinity groups

運用風險工傷的關連性族群來建立職業事故模型

Source：Safety Science 49 (2011) 306–314

Authors：Juan Carlos Conte , Emilio Rubio , Ana Isabel Garcia , Francisco Cano

Abstract

This article sets outs a generalized utility model for the diagnosis and prediction of accidents among the Spanish workforce. Based on observational data classified into a risk–injury contingency table (19×19), we have summarized the accident rate of all Spanish companies over an 11-year period (75,19,732 accidents).

By using correspondence analysis a structure composed of three axes can be obtained, the combination of which identifies three separate risk and injury groups, which we use as a general Spanish pattern. The relationships of greater affinity or likelihood amongst the risks and injuries identified in the pattern facilitate decision-making at the risk-assessment stage in Spanish companies. Each risk–injury group has its own characteristics, interpretable within the phenomenological framework of the accident.

The main advantage of this model is its potential application to any other country and the feasibility of contrasting results from different countries. One limiting factor, however, is that the model currently lacks a common classification frame for risks and injuries which would enhance this contrast. The aim of this model is to automatically manage work-related accidents at a national level..

目的是想設計出一個用來偵測與預測工作場所事故的模式，利用西班牙的所有行業在11年間發生的7519732件事故，運用19個風險類別對19個傷害類型的矩陣進行對應分析（correspondence analysis），找到與鑑別出三大風險與工傷特性的族群(單位是acsom)；提供政府主管機關可針對這三大個性不同的族群訂定不同的管制重點。

備註：

對應分析方法的參考資料

參考1

參考2

（功力已經差不多都歸還給老師了，如果要用的話，要再花點些時間溫習）

1.      Introduction

Our aim is to identify the real risks from a historical record of accidents and to summarize them in a contingency table.  Obtain the criteria needed for their assessment and prioritization from a mathematical–statistical analysis. 

想找出不同行業的風險與傷害特性，進而提供作為政府管制的優先順序安排（只是想回答=>哪些是高風險與需要被優先關注的產業）；邏輯一樣是鑑別、評估與排序

The main handicap in the standard evaluation of the ‘‘assumed’’ risks identified in different jobs within a company is that the risks are treated as isolated, independent events, which may or may not affect individuals (Conte et al., 2007).  Characterizing an accident based on a risk, before the accident has actually taken place, is of little use at present, as it is subject to the fundamental premise of uncertainty.  Thus, once the assumed risk has been identified, it is not possible to establish with certainty if and when it will occur, what the resulting injury will be, or its level of severity.  Moving from the rate of accidents (population) to the actual accident itself (the individual), adds a high level of randomness to the evaluation techniques used: the classic criterion variables have no proven role in the identification of risk (Körvers and Sonnemans, 2008). Consequently, only individual, technical criteria prevail in selecting the specific risk value.

說出了目前風險評估（矩陣法）的罩門：

把各類風險當成個別、相互獨立的因素       

沒有分析的層次與因果邏輯

用總體的統計資料來推估個體發生類似意外的機率，然而以個體的角度而言，類似的意外事故發生機率/嚴重度往往必須從相關的Detection/Protection來推估才比較合理

解決的對策：是用失誤樹或事件樹的方法來評估

但後遺症當然就是麻煩與耗工

借用FBD的概念來分析

Papazoglou and Ale (2007) presents a logical probability model based on the analysis of a functional block diagram (FBD). An interesting feature of this diagram is that, for evaluating the various factors associated with risk it offers a methodology, a deductive development of the logical structure of and a characterization of each relationship. It should not be confused with the classic block diagrams or event diagrams used in reliability analysis, as these use events with two possible states, while the FBD can handle multi-state events. An FBD is formed of various hierarchical levels in which the characteristics of the previous level are subdivided into other, more specific characteristics, as the order of the level increases.

本文的核心概念=>

把造成事故的風險與事故造成的傷害(類型)視為一體(acsom)

試圖用對應分析（correspondence analysis）找出，就國家層別當中所有產業的各項工傷事故當中的acsom的樣貌

Define a general model to understood the accident pattern, yardstick or standard for contrast of any given country, whose properties are applicable to any company within that country.  We call this ‘‘acsom’’, an acronym for ‘‘accidents soma’’ (body of accidents).

Conceptually, acsom represents an equilibrium diagram of accidents.

Taking each accident as a compound event consisting of a risk–injury pair, for each accident we identify its risk–injury (RI) type.  The combination of all these ‘‘RI’’ pairs for any given country constitutes its acsom-G, which is presented as an accident-rate offset model: it covers all the productive sectors, that is, all the positive and negative typological anomalies that characterize each area of activity. When these anomalies are put together, they combine with one another, and offset each other. This produces a matrix diagram (RI) and marginal profiles (R and I), which are used as an equilibrium standard. The local patterns, or acsom-S, of each branch of productive activity, are interpreted in the same way.

By means of correspondence analysis, we present a global model (acsom-G) for accidents; its underlying data structure is made up of three groups of risks and injuries. We have assigned colours to these three groups (red, yellow and green) in order to visually identify them. The colours do not indicate the level of severity of each group, but rather the features associated with the frequency of occurrence.

The contingency table obtained (Table 1) shows three key elements:

the total value, the marginal profiles, and the central body of the table, or the matrix. Each of these identified elements can be analyzed independently by using different methodologies.

2.      Materials

資料來源：We have considered all the reports of labor accidents notified over 11 years (7519,732 accidents), registered (Ministerial Order 16-12-1987, BOE 311, of 29th December) and published by the Spanish Ministry of Labor (Secretaría General Técnica, Subdirección General de Estadísticas Sociales y Laborales).

19(Risk)*19(Injuries)的矩陣分類：

資料的結果(矩陣表)：

3. Methods

可參考上述連結與文章中的說明

而圖1/2應該是事後驗證結果合理性的後製

表達與呈現三個族群(綠、黃、紅)各自的風險與傷害（在19個維度/eigenvalue）的分布情形

4. Results

直接使用19*19矩陣跑對應分析出來的結果，第一個群組(Green)如圖3

The greatest affinities are R10(projection of fragments or particles) with I11 (objects in the eyes); R14 (exposure to thermal contact) with I13 (burns), and the point of transition I15 (environmental hazards or effects) owing to phenomenological affinity although, because of the ambiguity in the definition of I15 it is also related to other risks; R15 (exposure to electrical contacts) with I13 (burns) and I17 (electric shocks); R16 (exposure to chemical contact) with I14 (poisoning and intoxications) and I18 (radiation poisoning); R17 (exposure to radiation) with I12 (conjunctivitis) and I18 (radiation poisoning) and R18 (explosions and fires) with I14 (poisoning and intoxications).

作者看圖的說故事、自圓其說，

從個人實務經驗的角度來看以上說明覺得意義不大（感覺這些關聯性理所當然、沒有意外的發現）。

Fig. 4 represents the group formed by four risks and five injuries.

The highest affinities are between R1 (fall of persons from different levels) and R2 (fall of persons from the same level) with I6 (concussion and internal trauma) and R6 (treading on objects) with I2 (dislocations) and I3 (twists, sprains and strains) and R13 (overstraining) with I4 (back pain) and I5 (slipped disc).

Fig. 5 represents this red group, formed by nine risks and six injuries. The highest affinities are between R3 (fall of objects or collapse), R4 (fall of objects during handling or manipulation), R7 (colliding with immobile objects) and R8 (colliding with moving objects) with I9 (superficial trauma) and I10 (bruises, contusions and crushing). R5 (detachment) with I10 (bruises, contusions and crushing); R9 (bruises, contusions and cuts by objects or tools) and R11 (trapping by or between objects) with I7 (amputations and loss of an eye) and I8 (other injuries); R12 (accidents caused by moving machinery or traffic) with I1 (fractures) and I19 (multiple injuries) and R19 (accidents caused by living beings) with I9 (superficial trauma).

Next, we represent the summarized table composed of the three centroids (Table 4 and Fig. 6), that the new correspondence analysis characterizes with two-dimensions, dim1* = X with eigenvalue k1 = 0.71531 and dim2* = Y with eigenvalue k2 = 0.28518, which represent 100% of the total variance. The three groups are a necessary and sufficient condition for representing the initial table.

最後圖6結果最好看

5. Discussion

In a logical model, occupational risk is modeled through a general FBD where the undesirable health consequence is decomposed to ‘‘dose’’ and ‘‘dose–response’’; ‘‘dose’’ is decomposed to ‘‘center event’’ and ‘‘mitigation’’; ‘‘center event’’ is decomposed to ‘‘initiating event’’ and ‘‘prevention’’. This generic FBD can be transformed to activity, specific FBDs which together with their equivalent event trees are used to delineate the various accident sequences that might lead to injury or death. The methodology and the associated algorithms have been computerized in a program.

理論上，職業傷害應該是作業風險的結果函數=>知道一家公司的作業特性與風險值，就可以計算得出這家公司各種職業傷病的或然率公式

The factorial model considers two base variables called ‘‘risk’’ and ‘‘injury’’, which bear some relationship to the first-level variables of the logical model, ‘‘dose’’ and ‘‘dose–response’’. In addition to these two variables, other variables are considered that are not reflected in this article, such as ‘‘situation’’, to characterize the work environment, the ‘‘duration of absence from work’’ and the ‘‘injured part of the body’’, aimed at characterizing the frequency of severity. New software has been developed, called Bioin, which enables preventive action planning with respect to the frequency of the various types of accidents that occur within a company. It also allows Delta data and lists of situations based on Eurostat criteria, used in Spain since 2003 to characterize accidents, to be translated to ILO data.

作者也知道這樣太簡化因果關聯，忽略了許多背景的情境，還有組織管理的因素（還好！）

`Regarding the groups, group-1, or the green group, includes all those risk and injury variables related to the appearance of historically-recent technological problems, the industrial revolution and scientific and technical development (Baram, 2009; Rasmussen, 1997). Group-2, or the yellow group, contains all those risk and injury variables related to evolving biomechanical problems (Nachreiner et al., 2006). Group-3, or the red group, contains all those risk and injury variables concerning technical–cultural problems (Guldenmund, 2000), related to the evolution of their activity.

以上這樣三群的分類，沒有實務上的意涵（難以對症下藥，還不如所謂物理性、化學性與生物性危害因子的分類）

總結心得與感想：

這是一篇看統計技巧取勝的paper，折磨刑求數據資料的技巧凶狠

但看此一題目對照其內容，則無可厚非（沒有定出太吹噓管理涵義的標題）

結果應用到實務上沒有什麼太大的貢獻與意義（換言之，文章的內外部效度與貢獻都不大；但運用對應分析來玩工安的data mining還頗為新鮮）

文章的罩門在於：Garbage in Garbage out=>

邏輯上各種作業環境與危害，對應所產生的傷害間的因果關聯應該要很明確

所以那「Spanish companies over an 11-year period (75,19,732 accidents). Spanish companies over an 11-year period (75,19,732 accidents).」應該要被clean清洗/篩選過，把一些很瞎的事故紀錄報告整理掉，或許最後結果還會再更好看一些。

另外就此一題目而言，如有要有所突破與對實務界有所助益而言，一個可以試試看的idea是：用事業單位來和這興風險與危害跑對應分析=>看看不同行業的職業傷病問題到底是出在本質風險高、亦或管理水準/技術投入不足。
最後就研究目的與方法而言，如果想找出事故因果之間的潛在關聯性，同樣的資料可以用老師推薦的Text Mining方法(傳統的Data Mining用資料是數據型態，這個方法可以針對文字型態的資料進行分析)