Estimating individual occupational risk using registration data

Safety Science 82 (2016) 95–102

Contents lists available at ScienceDirect

Safety Science journal homepage: www.elsevier.com/locate/ssci

Estimating individual occupational risk using registration data P.H.G. Berkhout ⇑, M. Damen RIGO Research, Amsterdam, The Netherlands

a r t i c l e

i n f o

Article history: Received 26 September 2014 Received in revised form 5 June 2015 Accepted 26 August 2015

Keywords: Occupational accident rate Occupational risk assessment Registration data Negative binomial regression Business cycle effects Job-worker matching

a b s t r a c t In this paper the occupational risk of individual employed workers is assessed using merged registration data. The analysis features a negative binomial count data regression model for statistical inference. In the model the hazard rate is determined by variables characterizing the personal background and working conditions, such as age, sex, tenure, contract type, working hours and sector. The estimation results can be used to assess individual occupational risk and identify high-risk groups and branches of industry. The main contribution of the paper to the literature is the fact that calculations include the entire population at risk and are based solely on registration data on individuals, jobs, firms and accident casualties. Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction In many developed countries improved working conditions over the past decades have resulted in declining numbers of occupational accidents. In the U.S workplace fatalities have been reduced since 1970 by more than 65 percent and occupational injury and illness rates have declined by 67 percent. At the same time, U.S. employment has almost doubled. Worker deaths in the U.S. are down from on average about 38 worker deaths a day in 1970 to 12 a day in 2012. Worker injuries and illnesses are down from 10.9 incidents per 100 workers in 1972 to 3.4 per 100 in 2011 (United States Department of Labor, 2014). The incidence rate of time-loss injuries per 100 workers across all jurisdictions in Canada has steadily declined in all years since 1996 (Gilks and Logan, 2010). In Italy the occupational accidents showed a decline of 14.5% in absolute values and a decrease of 21.1% in incidence rates between 2001 and 2008. In 2001 47.4 workers out of every 1000 workers experienced an occupational accident, while in 2008 this number decreased to 37.4 workers, due mainly to an 8.3% increase in overall employment (Inail, 2008). In Portugal the total number of reported occupational accidents decreased by about 4.4%, falling from 248,097 accidents a year in 2002 to 237,222 accidents in 2004 (IGT, 2014). In Australia between 2002 and 2012 workplace fatalities dropped by 42% and work-related ⇑ Corresponding author at: RIGO, Postbus 2805, 1000CV Amsterdam, The Netherlands. Tel.: +31 6 19661176; fax: +31 20 6276840. E-mail addresses: [email protected], [email protected] (P.H.G. Berkhout). http://dx.doi.org/10.1016/j.ssci.2015.08.013 0925-7535/Ó 2015 Elsevier Ltd. All rights reserved.

injuries by 28% (SWA, 2014). In the present paper figures are presented for The Netherlands between 1999 and 2011. Risk assessment is a cornerstone of the European approach to prevent occupational accidents and ill health. European legislation with respect to risk assessment is grounded on the Framework Directive 89/391, which has been transposed into national legislation of member states. The European Agency for Safety and Health at Work (EU-OSHA) suggests a stepwise approach to risk assessment: (1) identifying hazards and those at risk; (2) evaluating and prioritizing risks; (3) deciding on preventive action; (4) taking action and (5) monitoring and reviewing. Clearly, the process of risk assessment builds on identifying hazards and the workers at risk in the initial step. This paper sets out to quantify the occupational risk of all registered employed workers, based on data collected for registration purposes and conventional statistical methods. Usually statistics on occupational accidents are presented on the aggregate level. Risk assessment of individual workers is complicated, since it requires detailed information on outcome variables of all workers at risk, not only the accident casualties. Information gathering is usually initiated when occupational accidents occur; the counterfactual – workers who under the same circumstances did not get involved in an accident – is not observed. For instance, Rivas et al. (2011) analyze with the help of datamining techniques 62 questionnaires to be completed whenever an accident occurred. However, the question is whether inferences drawn from a sample of accidents only can be extrapolated to the workers who did not get involved in an accident. In experimental lingo this is referred to as the external validity, see for instance

96

P.H.G. Berkhout, M. Damen / Safety Science 82 (2016) 95–102

Shadish et al. (2002). In sample surveys of the population at risk generalization is usually not a problem. But sample surveys are usually too small in relation to the probability of observing enough accident casualties in the survey sample. Empirical literature on occupational accidents in The Netherlands centers around two long-term projects: the Occupational Risk Calculator (ORCA) and the Storybuilder database. The ORCAmodel assists companies, government or industry representatives to assess the occupational risks for individual workers and for sections of the workforce and identify cost effective risk reduction strategies (Ale et al., 2006). The ORCA-model uses logical bow-tie models and probability influencing entities (PIEs) as its fundamental building blocks (RIVM, 2008). Storybuilder is a software tool and a database at the same time. It supports occupational and major accident analysis, investigation, inspection and identification of ways to improve safety. The database contains serious accidents reported to and investigated by the labor inspectorate in the period 1998–2009. The database includes the information collected from elaborate case studies of the reported occupational accidents, such as the hazards and the failing of barriers essential to the causation of the accident (Baksteen et al., 2007; Bellamy et al., 2006, 2007, 2008). The accident casualties analyzed in the present paper are exactly the same persons as are stored in the Storybuilder database. However, the database could not be used in the analysis presented here, as the required accident identification key was not made available. In the present paper we will demonstrate that plausible estimations of individual occupational risk can be extracted from detailed personal information collected for registration purposes only. From here on, we will refer to such information as ‘registration data’. The negative binomial count data model is used to estimate the individual rate of accident occurrence per working day. To this end, recorded information on casualties of serious occupational accidents was merged to registration data representing the entire population at risk. In the model occupational risk may vary for individual workers as a result of differences in their human capital (age, gender, flex worker, nativity, tenure, hours worked per week) and the hazardous nature of the work (instrumented by 262 sector fixed-effects covering the entire economy). The research was funded by the Dutch National Institute for Public Health and the Environment RIVM). Since the available data cover a time span of thirteen years (1999–2011), the analysis is not only focused on cross-sectional explanation of occupational risk. In addition, time trends and business-cycle effects were identified. A declining number of accident casualties in this decade was in fact one of the main reasons why this study was initiated. Estimation results show a decline of 27% in the average rate of accident occurrence in time. This raises the question whether occupational safety in The Netherlands has truly improved to such extent. We conjecture that structural changes in labor market supply and demand have contributed significantly to changes in the observed average accident rate. For one, ageing of the worker population pushes the average rate upward (since occupational risk tends to rise with age for workers older than 30). The size of this ageing effect is considerable, due to an increase in the average age of the worker population from 36.3 in 1999 to 39.5 in 2011. On the other hand, a hazard-biased destruction of jobs over time has changed the composition of the job-pool in favor of the less hazardous jobs. A lower average rate of accident occurrence would be the result. Roughly estimated this composition-effect accounts for one third of the total decline. All in all, a ‘true’ decline in the accident rate of 20% remains, implying a considerable improvement in occupational safety in The Netherlands in the past decade. The paper is structured as follows. In Section 2 the observed individual occupational risk is considered within the conceptual

framework of labor market matching of workers to jobs. Data, definitions and a formal description of the empirical model are presented in Sections 3 and 4. In Section 5 estimation results are presented. Section 6 concludes. 2. Conceptual framework 2.1. Labor market matching of individual skills to job requirements In this paper the rate of occurrence of serious occupational accidents – the accident frequency – is analyzed within the human capital framework of labor market matching of workers to jobs. In the matching framework jobs require skills and workers provide these skills. In the hiring process employers try to fit the best workers to their vacant jobs. Each and every worker-job match results in an accident rate specific to that particular match. Intrinsic to any job there is a hazard potential related to the tasks performed. Some of these hazards may be confined by legislation and safety measures, but in practice most jobs will still represent a residual hazard potential. For the employer searching for a worker on the job this hazard potential is given, since he is responsible for the firm’s safety management system and therefor aware of safety measures taken. In our framework the remaining hazard potential is not affected by the worker selected for the job; it is an unobserved (latent) variable. One might say that in case an accident occurs, part of the potential’s hazardous energy is released and observed. The hazards require skills to handle them. These hazard handling skills are supplied by the workers. Anything contributing to the ability of handling hazards successfully can be considered: vitality, experience, responsiveness, knowledge etc. We refer to the aggregate all of such characteristics as the ‘hazard handling skills’. In the framework of matching workers to jobs these skills can be treated as any other requirement for the job. Workers are heterogeneous in their skills endowment, jobs are heterogeneous in their hazard potential. Since occupational accidents involve costs (and accident prevention reduces these costs), the employer treats the worker’s hazard handling skills as any productive skill. Depending on relative prices, there’s a trade-off between the cost evading hazard handling skills on the one hand and productive skills incurring revenue on the other hand. Should accidents incur high expected costs employers will emphasize hazard handling skills in the selection of workers. Vice versa, hazard handling skills will have low value to the employer if expected costs of accidents are low. Expected costs are defined by actual costs multiplied by the accident probability. Given a worker-job match, the job’s latent hazard potential is transformed into an observed occupational risk. From the worker’s perspective the occupational risk he will be exposed to depends on which job he selects from the job offers he receives. From the employer’s perspective the occupational risk related to his jobs varies with the matched workers. Hazard handling skills are instrumented by age, sex, nativity, hours worked, tenure and type of contract. On the demand side fixed-effects for 262 sectors of industry are estimated, representing job average hazard potentials. In the present analysis the individual occupational risk resulting from a specific job-worker match is defined as the rate of accident occurrence per working day. For the sake of brevity we will refer to this rate as the ‘incidence rate’ or ‘accident rate’. 2.2. Structural changes in labor supply and demand The above matching framework is used to explain changes in the population average over time. That is, developments on the macro level are related to mechanisms operating on the micro level

P.H.G. Berkhout, M. Damen / Safety Science 82 (2016) 95–102

of matching workers to jobs. We will assume matching technology did not change during the thirteen years of observation. With the matching technology given and fixed relative prices of hazard handling skills and productive skills, two different ways to change the average accident frequency on the demand side of the market are identified. First, the safety conditions in existing jobs may be affected by technological progress (e.g. improved gear and materials), investments in the firm’s safety culture and regulation by authorities. These factors affect the hazard potential in existing jobs. One might say they are associated with the ‘intrinsic’ or real occupational safety. Second, on the demand side there is on-going dynamic process of destruction and creation of jobs. This process in itself may be hazard-biased in a sense that it is selective with respect to hazard potential. When for instance due to technological progress relatively hazardous job are destroyed and relatively safer job are created, the composition of the job-pool changes over time. All else held constant, this will be revealed in a declining average hazard rate. However, such a composition effect can also contain a spurious improvement of occupational safety. For example, when relatively hazardous jobs are shifted to self-employed individuals the average hazard potential in the entire job population declines, but occupational safety from the national perspective remains unaltered (the hazardous work is still done, but not by employed workers). In this respect one might argue that outsourcing tasks to countries with low labor costs does not change occupational safety in global perspective; it only reallocates hazards geographically. Thus, as long as the destructed jobs are associated with hazardous work that has become redundant, the job-pool composition-effect will represent a real change in occupational safety. If destruction of jobs is associated with a reallocation of hazardous tasks (e.g. from employed workers to self-employed workers), the job-pool composition-effect is in essence no more than a spurious change in occupational safety. On the supply side of the market the matched worker population is equipped with hazard handling skills determining how much of the hazard potential is transformed into observed accident frequencies. Clearly, the amount of these skills present at any moment in time is subject to changes. One interesting hypothesis in this respect is for instance the effect of an ageing worker population on the accident frequencies. Presumably, as workers get older their hazard handling skills seem to deteriorate. The Dutch data show a steadily with age rising accident frequency for workers older than 30. In theory, this suggests that ageing of the worker population induces a higher accident frequency.

3. Data 3.1. Registered population at risk The population at risk is defined as all employed workers of 16 years and older registered to be living in The Netherlands. Self-employed individuals and foreigners not formally registered are excluded from the analysis. Micro register data on the population at risk was supplied by Statistics Netherlands (CBS). Information of several registrations is combined in the so-called SSB, which is the Dutch abbreviation of ‘social statistical micro data’. The SSB is in essence based on the municipal registrations of inhabitants, registrations of social insurances of employed workers and income taxes. The following information was used:
date of birth; gender; first and second generation foreign descent;
beginning and end dates of jobs; contract hours; type of contract (flex workers)
sector of industry (distinguishing 262 sectors).

97

Individuals are uniquely identified by a personal code derived from the citizen service number (BSN). Sectors of industry are identified using the International Standard Industrial Classification of All Economic Activities (ISIC). This is the international activity classification of the United Nations (United Nations, 2008). In the present paper the first three ISIC-digits allow differentiating between 262 sectors of industry. In order to ensure the privacy of casualties and workers, micro data never leave the closed computer environment of Statistics Netherlands. The unique personal identifier is a meaningless number in the world outside this environment. 3.2. Registered accident casualties The Dutch labor inspectorate (I-SZW) registers the casualties of reportable serious occupational accidents. These accidents are serious in the sense that they lead to death, permanent injury or hospitalization of the casualty. They constitute an estimated 1% fraction of all occupational accidents including the non-serious accidents (Bellamy et al., 2014). Traffic accidents on the road, in the air and on sea or waterways are not included in this registration. Furthermore accidents with dangerous sub-stances (e.g. fireworks, asbestos, radioactive materials) and natural resources (gas) are excluded. All these excluded accidents are investigated by other specialized inspectorates. The I-SZW registration however covers the vast majority of serious occupational accidents in The Netherlands. From all registered casualties in 1999–2011 only the employed workers were selected. Self-employed individuals and collateral casualties (e.g. passers-by) were excluded from the analyses. A number of studies suggest that recorded numbers of accident casualties may actually underestimate the true number of nonfatal occupational injuries, see for instance Leigh et al. (2004) and Rosenman et al. (2006). Such underestimation occurs when organizations fail to record employee injuries and illnesses or when employees fail to report injuries and illnesses occurring at work (Probst and Estrada, 2010). Although reporting of accidents to the inspectorate is mandatory/compulsory in The Netherlands, underreporting may occur when a serious accident is erroneously (or deliberately) not reported to the inspectorate. The amount of underreporting is estimated to range from 1000 to 1500 serious accidents per year (Schouten et al., 2008). This amounts to say 30–40% of the yearly average of serious accidents investigated by I-SZW or say 25% of the ‘true’ number of accidents. However, the estimate by Schouten et al. should be treated with caution, since it is based on a comparison with registered casualties of a small sample of first-aid centers unevenly spread across the country. Since there is no evidence of a selection bias due to underreporting, we will treat the registered casualties as a representative 75%-sample of the true casualty population. Therefore, in the following we will report accidents rates resulting after multiplication by 1.33 of the observed rates. 3.3. Merging data Data was assembled by one-to-one merging of several registers by means of a unique identifier of individuals: the citizen service number (BSN). The BSN-code of the casualties was not registered by I-SZW but derived by CBS on the basis of date of birth, gender and address (zip-code and house number). In 1999–2011 a total of 30,424 casualties were registered in all inspected serious accidents in The Netherlands. The citizen service number of 6353 (20.9%) of these casualties could not be identified, mostly due to lacking or incorrect personal information. Also casualties who were not formally registered as inhabitants or as tax-paying workers

98

P.H.G. Berkhout, M. Damen / Safety Science 82 (2016) 95–102

could not be identified. The remaining 24,071 (=30,424–6353) identified casualties compare very well to the total number of casualties counted in the Storybuilder database, in which all serious accidents are bow tie modelled by hand (Bellamy et al., 2006, 2007, 2008). This suggests that the quality of gathering of personal information by I-SZW may fall short when it turns out during the course of the investigation that the incident does not formally qualify as a serious accident. In general, identification failure does not give rise to a selection bias, with the exception of lethal accidents. Since the count of lethal accidents is wellestablished in Storybuilder we estimate a loss of say 10 out of the expected number of 75 deaths per year. From the 24,071 identified casualties another 2842 were excluded because they were not identified as members of the population at risk (employed workers). The final sample of occupational accident casualties consists of 21,229 individuals. 4. Definitions and empirical model 4.1. Definitions The time axis is measured in months, resulting in 156 time periods t in the years 1999–2011. In each t the actual number of workers – on average 7.5 million workers per period – is selected using the dates marking the beginning and ending of a job. The duration of exposure to occupational hazards per month is measured in full time (8 h) working days according to the contract, denoted as dit. For part time workers the working hours are transformed to full time working day equivalents. For instance, a half-time job of four days per week adds up to two full time working days. Workers with flexible hours are separated from other workers. Exposure is corrected for weekend days, varying length of the calendar months and Christian and national holidays. Sick leave and vacations could not be corrected for. The job hours of workers with two (or more) jobs at the same time are added up. Should these jobs be in separate sectors of industry, the sector of the largest job is assigned to the smaller ones. In each period t an indicator Ct expresses the state of the Dutch business cycle (DNB, 2013). Indicator Ait is defined as:
Ait ¼

1 member i of the population identified by data merging as casualty int 0 otherwise ð1Þ

where discrete indicator Ait expresses whether member i of the population at risk was registered as an accident casualty in period t. With on average 7.5 million workers per period we observe approximately one billion Ait in 1999–2011. In order to bring down this number to manageable size, groups of population members are defined. Each group g contains all population members i in period t who are homogenous with respect to characteristics vector Xi. This vector includes the variables: age, gender, native versus non-native workers, flex versus normal workers, part time workers, new recruits and 262 sectors of industry. This results in on average 90,000 homogeneous groups g in any period t. Note that regression-wise no information is lost since there are no variables left to differentiate within the groups. Within groups g the total number of accidents Ngt in period t is defined as:

Ngt ¼

X

Ait

ð2Þ

i2g

where Ait as defined in Eq. (1) and subscripts t, g and i refer to time period, group and individual in the group, respectively. The group’s total amount of exposure to hazards Dgt measured in full time working days is calculated as:

Dgt ¼

X dit i2g

ð3Þ

where Dgt and dit express the number of exposed working days of group g and individual i in time period t. Since Ngt is a nonnegative integer count it can be regressed on explanatory variables by means of a count data model. 4.2. Empirical model The literature offers a wide variety of count data models to regress a count variable on a set of independent variables (see for example Cameron and Trivedi, 1998). A well-known example is the Poisson model. The applicability of this model however is limited due to the fact that it imposes equi-dispersion, which means that the conditional mean and variance are equal. This is a rather strong assumption. A more flexible count data model is obtained when unobserved heterogeneity is introduced in the Poisson intensity parameter k. The most commonly used extension to the Poisson model is the negative binomial model (negbin), which results when k is mixed with a gamma distribution. The negbinmodel imposes over-dispersion, meaning that the conditional variance exceeds the mean. The model contains the Poisson model as a directly testable special case. In the negbin-model the probability that Ngt accidents are observed for group g in period t is written as:



PðN gt ¼ yÞ ¼

Cðq1 þ yÞ qkgt Cðq1 ÞCðy!Þ 1 þ qkgt

y



1 þ qkgt

q1

ð4Þ

where Ngt as defined in Eq. (2), y is the realized number of accidents and kgt refers to the above defined Poisson intensity parameter. Subscripts g and t refer to groups and time. The non-negative parameter q captures the degree of over-dispersion. If q converges to zero, the model reduces to the Poisson model. We will use the negbin type I model, which is parameterized such that q is scaled by kgt. This means that q varies across individuals and that – conditional on covariates – the count variance is a linear function of the count mean. In the present application of the model the intensity parameter k is be interpreted as the mean rate of occurrence of serious occupational accidents per 8-h working day. In short we will refer to k as the ‘incidence rate’ of getting involved in an occupational accident. Incidence rate k is modelled to vary with group characteristics X and may vary in time. Time variation may be due to structural tendencies (as a result from changes in the matching of supply and demand on the labor market), seasonal patterns and the business cycle. In order to assure a positive value of k, it is common practice to model k as follows:

kgt ¼ exp ðZ 0 bÞ

ð5Þ

where for the sake of brevity all explanatory variables are expressed by one matrix Z and all estimated coefficients are collected in one parameter vector b. In essence, the matrix multiplication Z0 b is a linear function adding up all estimated effects. Due to the exponential function these additive effects translate into multiplicative effects on incidence rate k. For example, let bm be the estimated coefficient for male workers; bm is one element of the vector b. The effect exp (bm) then expresses the relation between the incidence rate of male and female workers. An estimate of bm = 0.69 would imply that the incidence rate of occupational accidents among male workers is twice the incidence rate of females, since exp(0.69) = 2. The effect exp(bm) is often referred to as the ‘hazard ratio’, in this particular example: the male/female hazard ratio. In the model b-coefficients were estimated for three age intervals: 16–19, 20–29 and 30–64. The effect of job tenure is captured by a b-coefficient for ‘new recruits’. New recruits are defined as workers not registered to have been working in the preceding year. Three job size intervals are specified: part-time workers of 1–49% and 50–89% of a working week and full-timers (90–100%). The

99

P.H.G. Berkhout, M. Damen / Safety Science 82 (2016) 95–102

model also includes 262 b-coefficients for sectors of industry, one for male workers, one for flex workers and one for non-native workers. A linear time trend is included by counting the years as of 1999. A seasonal pattern in the incidence rate is allowed for by b-coefficients for the twelve months of calendar (with January as a reference). However, these estimates are likely to be biased by the fact that season-biased sick leave and vacations cannot be corrected for in the observed exposure. For instance, since many workers spend most of their time off during the summer holidays, the observed exposure is probably overestimated in July and August leading to a (spurious) decline in the incidence rate in these months. Therefore, any observed seasonal pattern in the incidence rate is obscured by unobserved patterns in absence from work due to sickness and vacations. We incorporated the seasonal b-coefficients primarily to correct for these unobserved patterns and refrain from exhibiting the estimated coefficients, since they have no interpretable meaning. The model is estimated using the nbreg procedure in statistical package STATA (version 12). It takes the merged data of workers and accident casualties, aggregated into homogenous groups, as input. The estimated b -coefficients are the output of the model. In total 329 b-coefficients were estimated. Given these estimates the incidence rate of any individual worker can now be calculated by adding all b-coefficients relevant for the worker and then taking the exponent of their sum. 5. Estimation results In this section the results of the estimated negbin-model are discussed. The estimated model results in 286 coefficients of which 262 are fixed-effects for sectors of industry. We will address these sector effects separately by means of descriptive statistics in subsection 5.3. In subsections 5.1 and 5.2 we will discuss the coefficients reflecting differences in age, gender and nativity and the worker-firm relationship. Subsection 5.4 focusses on the question to what extent the results can be interpreted as an achieved improvement in Dutch occupational safety. 5.1. Worker heterogeneity The average incidence rate of serious occupational accidents in The Netherlands in 1999–2011 per 8-h working day is estimated at 1.57  106. That means 1.57 serious accidents on one million full time working days. The incidence rate varies strongly over the population at risk. The distribution of incidence rates is skewed to the right, showing a relatively large amount of mass beneath the mean. The median is less than half the mean: 0.6  106. When we compare the incidence rates at the 1st and 99th percentiles of the worker population, we find a ratio of approximately 700. That is, the incidence rate of the workers at the right tail of the distribution facing the highest risk is 700 times higher than the incidence rate of those in the left tail of the distribution, where occupational risk is very low. This means that the risk determinants incorporated in the model (age, sex, nativity, worker-firm relationship and sector) can account for large risk differentials within the worker population. A further improvement of the model could be achieved if workers within the sectors are differentiated according to the hazard potential of their jobs, for instance by means of the International Standard Classification of Occupations (ISCO). This information was not available in our data. The male–female hazard ratio is estimated at 4.2, implying male incidence rates of on average say four times the female incidence rate. It is likely that this large hazard ratio is to some extent due to selection of male workers in jobs with high hazard potential

(in for instance construction work and industry). A remarkable age pattern is observed in the incidence rate. An initial specification of the model with separate fixed-effects for all ages revealed that the incidence rate peaks at the age of 19 and subsequently declines with age in the interval 20–29. For workers older than 30 the incidence rate increases steadily as they grow older. On the basis of this observation separate linear effects were estimated for the intervals 16–19, 20–29 and 30–64. Results are displayed in Table 1. As Table 1 shows, the incidence rate increases at a rate of 9 percent per year for workers younger than 20. The hazard ratio for workers of 16 and 19 years of age equals 1.42, implying a 41 percent higher rate for the 19-year old workers compared to workers 16 years of age. There are two hypothetical explanations for this remarkable rise in the accident rate within the youngest age interval. First, workers of 16 and 17 years of age are kept away from tasks (such as working with machines) by protective legislation. This suppresses their accident rate compared to workers of 18 and 19 years of ages. Secondly, young students in low vocation education are formally employed by firms but spend half of their contractual working hours in school. As their education evolves, their actually worked hours converges to full time hours. This suppresses their true exposure to hazards at younger ages, resulting in low estimated hazard rates. Since trainees and their true working hours cannot be identified in our data, this explanation could not be elaborated on in more detail. Presumably, effects of dual vocational education is minimal for young workers older than 20, since most programs end before 20. For workers 20–64 year of age the incidence rate is U-shaped. In the first ten years the incidence rate shows a 19% decline from age 20 to 29. Subsequently it gradually rises as workers grow older. From age 30 to 64 an increase of 23 percent is estimated. The estimated effects are significant beyond statistical doubt. The average incidence rate for all age cohorts in the years 1999–2011 is depicted in Fig. 1. The hazard ratio of non-native workers in The Netherlands equals 1.3, implying a significantly higher accident rate among workers with at least one parent born in a foreign country. In a preliminary specification of the model (not presented in this paper) an effect was tested for second generation non-natives. For those workers the incidence rate turned out to be approximately equal to the incidence rate of Dutch natives. From this one may conjecture that later generations of non-native workers: (1) end up in less hazardous jobs than earlier generations, and/or (2) are better equipped with hazard handling skills. The literature provides little solid empirical evidence supporting relatively high incidence rates of foreign/migrant workers. In Berkhout et al. (2014) an analysis of foreign accident casualties in the Netherlands shows that 15% of all casualties was foreign, whereas their share in the total number of worked days is estimated at 5–10%, depending on how one deals with multiple nationalities. They show that most of the differential can be accounted for by selection of foreigners in high-risk sectors. High accident rates of foreigners prevail in some high-risk sectors, such as the construction industry. In Dressler (2012) it is argued that there is a relation between occupational risk and the extent to

Table 1 Estimated age effects.

Age 16–19 Age 20–29 Age 30–64

b

Std error

0.088 0.023 0.006

0.028 0.005 0.001

Hazard ratio Year-to-year

Age interval min to max

1.09 0.98 1.01

1.41 0.81 1.23

100

P.H.G. Berkhout, M. Damen / Safety Science 82 (2016) 95–102

5.3. Sector hazard ratios

incidence rate (x 10-6)

2.5

2.0

1.5

1.0

0.5

0.0 15

25

35

45

55

65

worker age (years) Fig. 1. The average incidence rate and worker age (1999–2011).

which workers are familiar with safety regulations, cultural settings and the dominant language at the workplace. In the present framework this means that migrant workers are on average less endowed with hazard handling skills. On the other hand, it is sometimes claimed that cultural differences will force foreign workers to work more safely than their local counterparts, grounded by a sense of responsibility towards the family and their financial needs (Grøn and Cruz, 2012).

5.2. Worker-firm relationship Whether the worker-firm relationship is associated with incidence rate differentials is tested by three variables: job size (i.e. hours worked per week), flexible hours and tenure. Estimates indicate that ceteris paribus the incidence rate tends to decrease with increasing job size. That is, part time workers have higher accident rates than full timers, see Fig. 2. The hazard ratio of flex-workers is estimated at 1.3, indicating that workers flexibly affiliated to the firm tend to have higher occupational risk. A job tenure effect was tested by separating new recruits – those entering the labor force for the first time or after an absence of at least one year – from the more experienced workers. Effects in the three age intervals 15–19, 20–29 and 30–64 were estimated, none of which were found significant by conventional standards. This means we find no evidence of higher occupational risk for new recruits in general.

1.00

Sectors of industry were defined using the first three digits of International Standard Industrial Classification of All Economic Activities (ISIC). This results in 262 sectors for all of which a fixed-effect is estimated. For all sectors the hazard ratio with respect to sector Manufacture of Communication Equipment (ISIC-code 26.3) is calculated. This sector is chosen as reference since its average accident rate is closest to the total weighted population average. Due to limitations of space, the results for all 262 sectors are not presented here. They are available upon request. By definition the sector hazard ratios express average job hazard potential differentials, controlled for worker selection. However, sector differentials will in fact be larger if male workers are selected into hazardous jobs. That is, if jobs with high hazard potential are mostly done by male workers. If that would be the case, the estimated male–female hazard ratio merely reflects this selection effect rather than a gender difference in hazard handling skills. A high percentage of male workers in a sector then presumably signals the presence of hazardous jobs. Therefore an alternative sector hazard ratio incorporating sector differences in male–female composition is presented. This correction inflates the standard deviation of the sector hazard ratios from 1.4 to 1.7. The maximum hazard ratio increases from 6.5 to 7.9. Descriptive statistics of the 262 gender-corrected sector hazard ratios are displayed in Table 2. The sectors Manufacture of Weapons & Ammunition (25.4) and Other First Processing of Steel (24.3) show the highest hazard ratio and can be regarded as the most risky sectors to work in. The top 10 high-risk sectors features three branches of the metals industry, two mining branches and two branches closely related to construction: Demolition & site preparation (43.1) and Manufacture of articles of concrete, cement and plaster (23.6). Various branches in the construction sector show hazard ratios ranging from 4.0 to 5.2. Occupational accidents are unevenly distributed over the sectors of industry. This can be shown graphically by means of a Lorenz curve (Lorenz, 1905), which is commonly used in economics to analyze inequality in the distribution of wealth and income. Inequality can be measured by the Gini-coefficient, calculated as the ratio of the area enclosed by the curve and diagonal and the total area beneath the diagonal. For the sector distribution of occupational accidents in The Netherlands the Gini-coefficient is estimated at 0.37. In Fig. 3 the Lorenz curve of occupational accidents is depicted. Along the horizontal axis all 262 sectors are sorted from left to right with increasing numbers of occupational accidents. The percentage on the horizontal axis expresses the accumulated share in total

Table 2 Sector hazard ratios (male–female composition included). Descriptives Average Standard deviation Minimum Maximum N (sectors)

hazard ratio

0.75

0.50

0.25

0.00

half time Fig. 2. The job size hazard ratio.

full time

1.5 1.7 0.0 7.9 262

Percentiles

Hazard ratio

p(10%) p(25%) p(50%) p(75%) p(90%)

0.12 0.30 0.75 2.20 4.10

Sectors with highest hazard ratios (w.r.t. ISIC 26.3: Manufacture of Communication Equipment) Manufacture of weapons and ammunition (ISIC 25.4) Other first processing of steel (ISIC 24.3) Casting of metal (ISIC 24.5) Manufacture of basic precious and other non-ferrous metals (ISIC 24.4) Mining of stone, sand and clay (ISIC 08.1) Demolition and site preparation (ISIC 43.1) Materials recovery (ISIC 38.3) Manufacture of articles of concrete, cement and plaster (ISIC 23.6) Sawmilling and planing of wood (ISIC 16.1) Other mining and quarrying (ISIC 08.9)

7.9 7.8 6.9 6.4 6.2 6.1 6.1 6.0 6.0 6.0

P.H.G. Berkhout, M. Damen / Safety Science 82 (2016) 95–102

employment. The vertical axis shows the corresponding share in the total number of occupational accidents. For example, the depicted Lorenz curve shows that 60% of the employed workers in the least risky sectors suffer say 30% of all accidents. From Fig. 3 we have inferred that 36 percent of all serious occupational accidents are registered in just a handful of sectors representing 14 percent of total employment. These sectors (in the upper right corner of Fig. 3) are: Sector (ISIC-code)

Share in accidents

Temporary employment agencies job pools (ISIC 78.2) Roofing and other specialized construction (ISIC 43.9) Construction installation (ISIC 43.2) Construction of buildings (ISIC 41.2); Manufacture of structural metal products (ISIC 25.1); Other specialized wholesale (ISIC 46.7); Building completion (ISIC 43.3); Construction of roads, railways, bridges (ISIC 42.1) Freight transport by road (ISIC 49.4); Cleaning activities (ISIC 81.2)

8.8% 4.6% 4.0% 3.6% 2.7% 2.6% 2.5% 2.4% 2.4% 2.2%

Temporary employment agencies are at the top of the list. This implies that temporary workers can be considered a high-risk group of workers that may need special attention in preventive policy measures. 5.4. Effects of structural changes in labor supply and demand and the business-cycle The observed population average incidence rate in Netherlands decreased from approximately 1.57  106 in 1999 to 1.14  106 in 2011. This amounts to a total decline of 27% in twelve years. Holding all variables in the model constant as of 1999, the average incidence rate of the 1999-population in 2011 is estimated at 1.25  106. This is equal to a decline of 20% in twelve years. So in the absence structural changes in labor demand and supply the population average incidence rate would have decreased by 20% between 1999 and 2011 (from 1.57  106 in 1999 to 1.25  106 in 2011) instead of the observed 27%. We consider this 20% decrease as the achieved ‘true’ improvement of occupational safety in 1999–2011.

100%

share in occupational accidents

90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 0%

20%

40%

60%

80%

100%

share in total employment Fig. 3. Lorenz curve of occupational accidents in The Netherlands.

101

The difference of 7% between the observed decline (27%) and the decline on the basis of safety improvement (20%) can be attributed to structural changes in labor demand and supply. The model estimates in the present analysis allow for a breaking down of the decline in the Dutch incidence rate into factors driven by business cycle effects and changes in labor supply and demand. On the supply side ageing of labor supply affects the hazard rate. As was shown in Fig. 1 the accident rate slowly rises with age; apparently hazard handling skills of workers gradually deteriorate when they grow older. Between 1999 and 2011 the share of workers older than 30 increased from 67.4% to 72.4%, resulting in a rise of the average worker age from 36.3 to 39.5. Since the incidence rate of workers older than 30 increases 0.6% (see Table 1) each year they get older, we estimate an ageing effect of approximately 2%, calculated as 0.6%  (39.5 36.3). In other words, without ageing the decline in average incidence rate would have been approximately 29% instead of the observed 27%. Although we cannot rule out other serious changes in labor supply than ageing, this result suggests that structural changes in the composition of labor demand may also have had an impact on the observed population average incidence rate. The 9% difference adjusted for ageing effects (calculated as 29–27%) may be interpreted as such a composition effect. That is, one third (9 out of 27) of the observed decline in the incidence rate may be explained by shifts in the composition of labor demand. For instance, hazard-biased jobs destruction shifting hazardous tasks away from the population of employed workers under consideration. Some of these tasks may have been destructed forever, others are still performed by self-employed workers or may be outsourced abroad. Davies and Jones (2005) observe a similar composition effect explaining a declining accident rate in the U.K. A business cycle effect was tested by incorporating a monthly national business cycle indicator in the model. The results reveal a positive significant effect. That is, in an economic boom the accident frequency in The Netherlands is 4–6% higher than in a slump. For the U.K. Davies and Jones (2005) also claim that the rates of major injury follow a pro-cyclical pattern over the course of the business cycle. They conjecture this pro-cyclical pattern to be related to changes in the incidence of new hires over the business cycle.

6. Concluding remarks This paper sets out to quantify the occupational risk of all registered employed workers, based on merged registration data. By means of a negative binomial count data model cross-sectional differences as well as time trends are analyzed. In cross-sectional perspective the model predicts a hazard ratio between the 1st and 99th percentile of the worker population of approximately 700 on the basis of age, sex, nativity, worker-firm relationship and sector. A 27% decline in the incidence rate over the years is partly explained by structural changes in labor supply and demand (e.g. ageing and hazard-biased jobs destruction). In addition, the Dutch incidence rate reveals a pro-cyclical pattern over the course of the business cycle. All in all, a considerable ‘unexplained’ improvement of occupational safety in The Netherlands remains. In The Netherlands the use of registration data for statistical purposes has become common practice in the past decade. In this study a registration of casualties of serious occupational accidents was merged to the entire population of employed workers. The analysis combines data from individuals, jobs, firms and casualties. As the paper shows, the registration data can be transformed into accurate policy information when the proper statistical tools are utilized. Accuracy of the results would improve considerably if a job classification could be made available. But all in all, the paper shows that the incidence rate of individual workers can be

102

P.H.G. Berkhout, M. Damen / Safety Science 82 (2016) 95–102

assessed effectively by combining data from separate registrations. We consider this the main contribution of the paper to the literature. Finally, we have shown that occupational accidents may be analyzed within the standard labor economics framework of matching skills (human capital) to job requirements. The skills needed to handle a job’s hazard potential – say the hazard handling skills – may be treated as any productive revenue generating skills workers are endowed with. Since it is practically impossible to measure the hazard potential of jobs, we believe the focus of attention should be on the actual hazard rate which results from particular job-worker matches. The paper shows that by using complete registrations we are able to identify the effects of structural changes in labor supply and demand on occupational safety in national perspective. References Ale, B.J.M., Bellamy, L.J., Papazoglou I.A., Hale A.R., Goossens L.H.J., Post J., Baksteen, H., Mud M.L., Oh J.I.H., Bloemhoff A., Whiston, J.Y., 2006. ORM: Development of an integrated method to assess occupational risk. In: International Conference on Probabilistic Safety Assessment and Management, May 13-19, 2006, New Orleans, ASME, New York, 2006, ISBN 0-7918-0244. Baksteen, H., Mud M.L., Bellamy L.J., 2007. Accident Analysis using Storybuilder: Illustrated with overfilling accidents including Buncefield, UK. Bellamy, L.J., Oh, J.I.H., Ale, B.J.M., Whiston, J.Y., Mud, M.L., Baksteen, H., Hale, A.R., Papazoglou, I.A, 2006. Story-builder: The New Interface for Accident Analysis. In: Proceed-ings of the 8th International Conference on Probabilistic Safety Assessment and Management [PSAM], May 13–19, 2006, New Orleans, ASME, New York, 2006, ISBN 0-7918-0244. Bellamy, L.J., Ale, B.J.M., Geyer, T.A.W., Goossens, L.H.J., Hale, A.R., Oh, J., Mud, M., Bloemhof, A., Papazoglou, I.A., Whiston, J.Y., 2007. Storybuilder: a tool for the analysis of accident reports. Reliab. Eng. Syst. Safety 92 (2007), 735–744. Bellamy, L.J., Ale, B.J.M., Whiston, J.Y., Mud, M.L., Baksteen, H., Hale, A.R., Papazoglou, I.A., Bloemhoff, A., Damen, M., Oh, J.I.H., 2008. The software tool storybuilder and the analysis of the horrible stories of occupational accidents. Saf. Sci. 46 (2008), 186–197. Bellamy, L.J., Manuel, H.J., Oh, J.I.H., 2014. Investigated serious occupational accidents in The Netherlands, 1998–2009. Int. J. Occup. Safety Ergon. 20 (1), 19–32. Berkhout, P.H.G., Damen, M., Bellamy, L.J., Mud, M., Manuel, H.J., Oh, J., 2014. Serious occupational accidents of non-Dutch workers in The Netherlands. In: R.D.J.M. Steenbergen, PH.A.J.M. van Gelder, S. Miraglia, A.C.W.M. Vrouwenvelder (Eds.). Safety Reliability and Risk Analysis: beyond the horizon.

Cameron, A.C., Trivedi, P.K., 1998. Regression Analysis of Count Data. Cambridge University Press, New York. Davies, R., Jones, P., 2005. Trends and context to rates of workplace injury, Warwick Institute for Employment Research, University of Warwick, research report, pp. 386. DNB, 2013. DNB-business cycle indicator. . Dressler, D., 2012. Increased occupational risks among migrant workers in Austria, ISSA, WOS 2012. EU-OSHA, 2014. . IGT, 2014. http://www.igt.gov.pt/ (in Portuguese: General Labour Inspectorate statistics). Gilks, J., Logan, R., 2010. Occupational Injuries and Diseases in Canada 1996–2008: Injury Rates and Cost to the Economy, Human Resources and Skills Development Canada, July 2010. . Grøn, S., Cruz, G., 2012. National differences in safety practice, Centre for Maritime Health and Safety, University of Southern Denmark, University of the Phillipines. WOS 2012. Inail, 2008. Annual report. Leigh, J.P., Marcin, J.P., Miller, T.R., 2004. An estimate of the U.S. government’s undercount of non-fatal occupational injuries. J. Occup. Environ. Med. 46, 10– 18. Lorenz, M.O., 1905. Methods of measuring the concentration of wealth. Publications of the American Statistical Association, vol. 9(70), pp. 209–219. Probst, T.M., Estrada, A.X., 2010. Accident under-reporting among employees: testing the moderating influence of psychological safety climate and supervisor enforcement of safety practices. Accid. Anal. Prev. 42 (2010), 1438–1444. Rivas, T., Paz, M., Martin, J.E., Matias, J.M., Garcia, J.F., Taboada, J., 2011. Explaining and predicting workplace acci-dents using data-mining techniques. Reliab. Eng. Syst. Safety. 96, 739–747. RIVM, 2008. The quantification of occupational risk; The development of a risk assessment model and software. RIVM-report 620801001/2008, WORM Metamorphosis Consortium. Rosenman, K.D., Kalush, A., Reilly, M.J., Gardiner, J.C., Reeves, M., Luo, Z., 2006. How much work-related injury and illness is missed by the current national surveillance system. J. Occup. Environ. Med. 48, 357–365. Schouten, M.J., Faas A., Hoeben J., 2008. Achtergronden dodelijke en ernstige arbeidsongevallen op basis van in 2007 afgesloten ongevalsonderzoeken, Arbeidsinspectie, December 2008 (in Dutch). Shadish, W.R., Cook, T.D., Campbell, D.T., 2002. Experimental and Quasiexperimental Designs for Generalized Causal Inference. Houghton Mifflin Company, Boston New York. STATA, 2014. . SWA, 2014. Key Work Health and Safety Statistics Australia 2014, . United Nations, 2008. International Standard Industrial Classification of All Economic Activities, Statistical Papers, Series M, revision 4. United States Department of Labor, 2014. .