Drivers and frictions of workplace accidents: an empirical investigation of cross-country European heterogeneity

ABSTRACT

This paper presents an empirical investigation on the determinants of workplace accidents across Europe and focuses on the extent to which production-system characteristics (employment sectoral risk, size of firms, temporary contracts), business cycle and socio-economic factors (GDP, level of investments, unemployment, education) and other territorial controls (crime index) might account for cross-country heterogeneity. We use Eurostat data, and our panel is composed of 27 European countries over the period 2010–2018. Implementing different functional forms and estimation methodologies (pooled OLS, panel fixed and random effects models, system-GMM and semiparametric fixed effects model), we find robust evidence that productive-system structural characteristics, business cycle controls and the other territorial variables are effective in explaining European cross-country heterogeneity. Moreover, we find evidence of a nonlinear relationship between GDP and occupational accidents.

KEYWORDS:

• Occupational accidents
• European productive-system
• European countries
• system-GMM
• semiparametric fixed effects model

JEL CLASSIFICATION:

• J28
• J83
• I18
• H53

I. Introduction

Workplace safety is a relevant issue that has been increasingly attracting the attention of institutions, labour organizations, researchers, and policymakers (European Commission 2002, 2007; Greer 2018; da Silva and Amaral 2019; Liu 2019; Schulte et al. 2019; Tassinari, Chesta, and Cini 2020). In Europe, even though the number of occupational injuries shows a decreasing trend, cross-country differences in workplace accident rates show a heterogeneous dynamic mainly related to differences in definitions and measurement (Boone and van Ours 2006; Anyfantis, Boustras, and Karageorgiou 2018; Ivascu and Cioca 2019; Lafuente and Daza 2019; Verra et al. 2019). In Italy, from 2008 to 2019, the highest percentage of work accidents occurred in those sectors that are typically considered particularly at risk and that employ vulnerable workers (low-skilled, low-paid, low-educated), such as manufacturing industries, construction, trade, and transport activities (Farina et al. 2018; Lombardi, Fargnoli, and Parise 2019; De Merich et al. 2022). Even though, over the past few years, the total accidents in Italy, like in the other European countries, has decreased (Hämäläinen, Saarela, and Takala 2009)¹, the work-related risks, however, have not been reduced in a uniform way leaving some categories of workers, companies and sectors overexposed to workplace risks (Cioni and Savioli 2016; Giraudo et al. 2018; Zambon et al. 2018; D’Ambrosio, Leombruni, and Razzolini 2022). In Italy, for instance, Central-Northern provinces have a higher proportion of employment in industrial sectors (in comparison to the Southern regions) which pose a higher risk of work accidents/injuries.

The frequency of workplace accidents has been extensively studied in the literature for which, traditionally, it is possible to identify four main groups of factors affecting injuries (see Fabiano, Currò, and Pastorino 2004): i) individual factors related to workers characteristics (age, gender) and experience, ii) job-related factors, organization of work and environmental conditions, iii) technology used, and iv) economic factors, such as general economic conditions, unemployment rate, labour and social insurance legislation, business cycles. Cornelissen, Van Hoof, and De Jong (2017) have provided one of the most comprehensive overviews on the determinants of occupational safety, identifying and clustering several possible determinants of occupational injuries in high-risk industries (i.e. construction, petro-chemistry, warehouses, and manufacturing). The resulting seven clusters constitute a suitable framework to identify the possible determinants of workplace accidents since Cornelissen, Van Hoof, and De Jong (2017) considered both theoretical and empirical studies and included determinants which have received so far little attention in previous models (e.g. external factors).

More recently, Theorell (2022) has edited the Handbook of Socioeconomic Determinants of Occupational Health, collecting a robust evidence on key issues in occupational health research, considering both a macro and a micro-level perspective. Specifically, at macro level, working conditions are put in relation to political change, gender, age, precarious employment, national economy, and retirement. At micro level, instead, the analysis focuses on a wide array of factors, ranging from the social patterns in distributions of psychosocial and physical risks at work in general to physical and socioeconomic issues as noise, shift work, under/overemployment, occupational physical activity, job intensity, digitization in modern work, climate change, childhood determinants of occupational health in adult years, and theoretical models currently used in occupational epidemiology (Theorell 2022).

As emphasized by Cioni and Savioli (2016), workplace accidents are the result of a process that involves different and multifaceted determinants; our work aims to enrich the literature by investigating the key macroeconomic determinants of workplace accidents, both at European and Italian levels, in an attempt to document the relationships between occupational accidents and economic, territorial and production-system factors affecting workplace risk. Differences in occupational safety and health (OSH) performances are not likely to be the result of firms’ intrinsic characteristics alone, but rather of their interaction with social, political, regulatory, and economic contexts which are therefore investigated.

Based on this strand of literature, the aim of the paper is twofold. First, grounding on a model specification that combines productive system characteristics and socio-economic controls, our empirical analysis investigates, at a macro level of analysis, the main determinants of occupational accident rates, aiming to disentangle European cross-country heterogeneity, within the business cycle theoretical framework. Secondly, we deepen the study of the relationship between the European country heterogenous growth trends and occupational accidents, investigating the existence of a nonlinear relationship between accidents’ rate and per capita GDP increases, for which the accidents’ rate might first grow and then fall. Our empirical analysis relies on a panel dataset composed by 27 European countries, for a time span of 9 years (2010–2018), and we employ alternative econometric approaches (OLS pooled, panel random- and fixed-effects model, system-GMM and semiparametric fixed effects model). The remainder of the manuscript is organized as follows. In section II, we introduce the data, while section III presents our econometric methodology. Estimation results are presented and discussed in section IV. Section V outlines our conclusions and policy implications.

II. Data

Our panel dataset is composed by 27 European countries observed for 9 years (from 2010 to 2018). The data are provided by Eurostat, both for the dependent variable and the independent variables, which are described in detail in the following sections. Table 1 briefly illustrates the main characteristics of the quantitative analysis, in which the occupational accident rate is modelled as the dependent variable of an array of regression models involving several regressors (see Table 2). Specifically, the panel analysis focuses on 27 OECD countries and a time span of 9 years (from 2010 to 2018). In the reminder of this Section, we briefly illustrate the dependent variable and the selection of independent variables used in the panel analysis.

Table 1. An overview of the quantitative analysis.

	Description
Dependent Variable	Occupational accidents rate: the ratio, measured in percentage terms, is calculated as the number of serious accidents per 100,000 persons in employment (in total business economy, repair of computers, personal and household goods, except financial and insurance activities). Source: EUROSTAT.
Countries	Austria (AT), Belgium (BE), Bulgaria (BG), Croatia (HR), Czechia (CZ), Denmark (DK), Estonia (EE), Finland (FI), France (FR), Germany (DE), Greece (EL), Hungary (HU), Ireland (IE), Italy (IT), Latvia (LV), Lithuania (LT), Luxembourg (LU), Netherlands (NL), Norway (NO), Poland (PL), Portugal (PT), Romania (RO), Slovakia (SK), Slovenia (SI), Spain (ES), Sweden (SE), United Kingdom (GB)
Time frame	From 2010 to 2018 (9 years)

Table 2. Independent variables description.

Variables	Description	Source	Unit
Employment risk	The indicator is defined as the sum of the employees working in sectors with high accidents risk (mining and quarrying, construction, and transport) per the total persons in employment (in total business economy, repair of computers, personal and household goods, except financial and insurance activities)	Eurostat	Index
Large firms	The indicator is defined as the number of large firms (250 persons employed or more) per total number of firms in the country. Firms are counted in total business economy, repair of computers, personal and household goods, except financial and insurance activities	Eurostat	Percentage
Temporary contracts	Share of the employees from 15 to 64 years of age with a temporary contract	Eurostat	Percentage
GDP per capita	The indicator is calculated as the ratio of real GDP to the average population of a specific year. GDP measures the value of total final output of goods and services produced by an economy within a certain period of time	Eurostat	Euro
Secondary Education.	Share of the population with at secondary education attainment per inhabitants	Eurostat	Percentage
Fixed investment/GDP	Gross fixed capital formation as a percentage of gross domestic product	Eurostat	Percentage
Unemployment	Share of the population of persons from 15 to 74 years of age (16 to 74 years in ES, IT and the UK) unemployed	Eurostat	Percentage
Theft	The indicator is calculated as the number of thefts per hundred thousand inhabitants	Eurostat	Ratio

Dependent variable

The dependent variable in the regression model is measured as the incidence rate of workplace accidents resulting in the number of serious accidents per 100,000 persons in employment. In particular, the numerator is the number of accidents that occurred during the year; the denominator is the reference population (i.e. the number of persons in employment) expressed in 100,000 persons. An accident at work is defined as a discrete occurrence in the course of work which leads to physical or mental harm² If the accident does not lead to the death of the victim, it is called a ‘non-fatal’ (or ‘serious’) accident, and it is intended involving more than three calendar days of absence from work.

Figure 1 shows the average percentage rate of change in OA from 2010 to 2018 for the 27 countries of our panel.

Figure 1. Average annual variation of OA rates (2010–2018).

Source: Eurostat

Aware of the phenomenon of underreporting, the purpose of this indicator is not to give a picture of OSH performance in the countries illustrated, but to provide a descriptive statistical representation of which countries experience a reduction or an increase in the change of OA rates. Based on the evidence, Lithuania, Romania, Latvia, France, Slovakia and Croatia are the countries with the largest average positive percentage change (increase) between the previous and following years among the 27 countries considered. The other countries with an average increase in OAR from 2010 to 2018 are Luxembourg, Hungary, Portugal and Ireland. The countries, on the other hand, with the largest negative percentage change (reduction) rates are Norway, Italy, Czech Republic, Denmark, Netherlands and United Kingdom. Greece, Slovenia, Austria, Finland, Bulgaria, Spain, Estonia, Germany, Poland and Sweden, although to a lesser extent, still exhibit a negative average percentage change from the previous to the following year.

Independent variables

We include a set of explanatory variables, which are divided in three macro-categories: production-system, socio-economic, and other territorial variables. The independent variables were chosen according to those most commonly used in the economic literature on workplace safety, accident and injury research, and they are described below. Table 2 summarizes them.

Production-system variables

A substantial part of the literature focuses on the relationship between workplace accidents (non-fatal) and types of worker contracts Lopez et al. (2008). Dupré (2001) found that, in European countries, the risk of accidents for temporary workers who had been employed for less than two years was particularly high in the construction, health and social sectors. Fabiano et al. (2008) using Italian data for the period 2000–2004, show that workers employed by temporary-help agencies suffer a higher injury frequency index (compared to employees that are directly hired) due to lack of experience, insufficient specific knowledge, and inadequate training. Virtanen et al. (2005) found that temporary workers may have a higher risk of psychological morbidity and work-related injuries as compared to permanent workers. However, Benavides et al. (2006) found that, even though temporary workers seem to have a higher risk for work-related injuries than permanent workers, after controlling for the length of employment the probability of accidents is quite similar in both groups. The firm size is also considered a relevant variable able to exert a significant impact on the level of occupational risk (Fabiano, Currò, and Pastorino 2004) finding an inverse correlation between the frequency index of workplace accident and the firms’ size.

In literature, a great deal of agreement exists on the type of economic sector which plays a determining role in occupational accident risk levels: this is confirmed by the fact that the differences, between sectors, found in this respect persist even when controlling for heterogeneity between clusters of countries (Lenaerts et al. 2022). The sectoral perspective is relevant particularly for studying the risks associated to the physical environment (Leigh et al. 1990; Leigh 1989; Maiti and Bhattacherjee 1999; Maiti, Bhattacherjee, and Bangdiwala 2001; Maiti, Chatterjee, and Bangdiwala 2004; Haslam et al. 2005; Khanzode, Maiti, and Ray 2012, 2011), which depend on the production process, the materials used, the equipment endowment, as well as the activities carried out during the work activity (Eurofound 2012; Wadsworth and Walters 2014; Lenaerts et al. 2022); for those employed in manual work, physical health risks remain the strongest predictor of injuries (Kubicek, Bhanugopan, and O’Neill 2019; Toch et al. 2014).

Intersectoral differences in terms of activities and tasks performed by workers are also directly related to the occupational structure of each sector. In particular, sectors dominated by blue-collar employments, such as construction, agriculture, industry and transport, typically have less secure physical environments; in contrast, sectors dominated by white-collar occupations, such as financial services, education and public administration, have better levels of physical environment (Lenaerts et al. 2022). Therefore, while sectors dominated by occupations with a higher level of education and generally higher labour income are subject to OSH risks more related to psychological factors (such as stress, and anxiety), the blue-collars are exposed to higher accidents’ risk (Dorman 2000).

Socio-economic variables

Since the pioneering study by Kossoris (1938) on the relationship between occupational accidents and the business cycles, several studies (Ussif 2004; Davies, Jones, and Nuňez 2009; Fortin, Lanoie, and Laporte 1996) have found a pro-cyclical relation between the business cycle and the occupational injuries rates, showing that the number of accidents tends to increase during economic upswings and vice versa (Robinson 1988; Nichols 1991; Fabiano et al. 1995); periods of economic slow-downs may accompany a reduction in the number of injuries. Furthermore, macroeconomic indicators, such as gross domestic product (Asfaw, Pana-Cryan, and Rosa 2011) or unemployment rate (Brooker, Frank, and Tarasuk 1997) have been primarily used to study their respective incidence on occupational accidents. Other economic characteristics such as unemployment rate and labour and social-insurance legislation (Blank, Diderichsen, and Andersson 1996) are also considered relevant variables. Some authors have suggested that workers tend to underreport minor injuries during a recession (Boone and van Ours 2006), which reduce the official number of minor injuries that they report in times of crisis.

Territorial variables

The number of thefts is taken as a proxy for the level of crime within a given country. The underlying hypothesis is that a higher level of crime is related to lower OSH compliance by firms. Regulation and compliance might have an impact on OSH levels and this is demonstrated in several empirical studies. Viscusi (1986), for instance, estimates a statistically significant impact of OSH regulation on lost working days due to accidents. Weil (1996), on the other hand, investigates the effect of regulation on firm compliance, as measured by the number of violations of expected standards. In particular, the work investigates the effect of inspections – one of the main control instruments in OSH regulation – which impose fines and penalties for violations. One of the main outcomes of this study shows that compliance decisions would be made on the basis of potential, rather than actual, penalties. Concerning the role of the sequence and frequency of inspections on the compliance of firms with OSH regulations, Ko, Mendeloff, and Gray (2010) find that from the first to the second inspection there is at most a 31% drop in the number of violations (i.e. higher level of compliance), with a smaller drop in the case of serious violations; furthermore, the effect of repeat inspections on serious violations is greater in planned inspections than in other types of inspections.

Table 3 provides an overview of the selected variables and their summary statistics. At first glance, the data are characterized by a high variability.

Table 3. Descriptive statistics.

Variable	Obs.	Mean	Std. Dev.	Min.	Max.
OA Rate (non-fatal)	243	1226.776	887.779	47.97	3458.28
Total OA Rate	243	1229.107	887.657	52.88	3461.02
Employment Risk	243	0.351	0.073	0.192	0.497
Large Firms	243	0.002	0.001	0.001	0.005
Temporary Contracts	243	15.093	9.788	1.8	46.9
Per capita GDP	243	27166.46	18113.98	5050	83470
Secondary Education	243	48.021	11.38	18.8	71.1
Fixed Investments/GDP	242	20.707	3.502	10.77	35.81
Unemployment	243	9.335	4.909	2.2	27.5
Crime Index	243	1490.213	1055.366	220.99	5282.09

The research assumption is that work safety is a complex phenomenon influenced by both micro (individual-related, job-related) and macroeconomic factors (GDP, unemployment), which should be analysed comprehensively. Our main contributions are: i) to extend the literature in this vein by understanding the main macroeconomic determinants of the incidence of occupational accidents; and ii) to test the existence of a non-linear relationship between the occupational accident rate and the GDP level, while controlling for production-system characteristics and unobserved heterogeneity. On this background, we test for two main hypotheses:

-Hp1: Territorial heterogeneity of workplace accidents, at both European and Italian level, can be explained by recurring to a macro level of analysis, within the business cycle theoretical framework;

-Hp2: The relationship between workplace accidents and the GDP level is complex (de La Fuente et al. 2014; Asfaw, Pana-Cryan, and Rosa 2011) and we may expect that it takes a non-linear (and non-monotonic) relationship. Indeed, occupational accidents rate may increase with economic wealth due, for instance, to the higher probability of reporting more injuries (Boone and van Ours 2006; Davies, Jones, and Nuňez 2009), or to the share of new or inexperienced workers in the total labour force increases during economic expansions (Boone and van Ours 2006; Robinson 1988); however, the relationship between workplace accidents and GDP may invert after reaching a certain wealth threshold, while controlling for production-system, institutional and macroeconomic characteristics. Generally speaking, countries with higher per capita income may exhibit higher levels of health and safety workplace measures and regulations, not only because firms will tend to pay more attention to the workers’ safety conditions (i.e. healthier and safer workers can work more efficiently and with higher labour productivity) but also because workers might have a higher perception of the potential risks associated with the probability of occurrence of workplace accidents. One view (Bonamore, Carmignani, and Colombo 2015) is that in time of economic downturn and rising unemployment, tighter firms’ budget constraints might imply cuts in workplace safety-related equipment and measures. Economic expansions are, instead, typically characterized by declining unemployment, improved labour market conditions with higher expenditures on health and safety workplace measures. The quality of the workplace systems might, therefore, improve because of the positive effects that economic upturns have on the investment capacity of the firms. If the non-linear relationship is confirmed, then the GDP coefficient will be positive and the coefficient associated to the square of the GDP variable will be negative. If our hypothesis is correct, then, at aggregate level, we should observe an inverted U-shaped relationship: at lower levels of income, an increase of per capita GDP might initially be associated with an increase of occupational accident rates; but when the level of income is sufficiently high, then further increases of per capita GDP will decrease occupational accidents rate. To the best of our knowledge, the previous literature does not provide any evidence on the existence of this potential non-linearity in estimating the determinants of workplace accidents at macro level; thus, the present study is oriented to fill this gap.

III. Econometric methodology

Since Hartwig’s et al. (1997) cross-country analysis, an extensive stream of literature has deepened the relationship between workplace accidents (or workers compensations) and the business cycle. The main insight in the background of these studies is that macro-economic trends can affect the drivers and frictions of occupational accidents. To substantiate this hypothesis several empirical works (Hartwig et al. 1997; Brooker, Frank, and Tarasuk 1997; Ussif 2004; Boone and van Ours 2006; Davies, Jones, and Nuňez 2009; Asfaw, Pana-Cryan, and Rosa 2011) find evidence of a correlation between occupational accidents rates and major macro-economic indicators (GDP, fixed investments, unemployment rate).

In order to analyse the heterogeneity of the occupational accident phenomenon at European level (Hp1), this paper aims to extend this literature by including, besides to the main business cycle indicators, also the productive system and institutional characteristics of the EU-27 countries considered in our panel. In this context, we begin by testing the presence of a link between the annual occupational accidents and the main countries’ productive characteristics, controlling for business cycle indicators and institutional characteristics. The OLS estimation equation takes the following log-log form:

(1) $\mathit{O}{A}_{\mathit{it}}={\beta }_0+{\beta }_1\mathit{Xps}{c}_{\mathit{it}}+{\beta }_2\mathit{Zb}{c}_{\mathit{it}}+{\beta }_2\mathit{Ut}{v}_{\mathit{it}}+T_t+\mathit{F}{E}_i+{\epsilon }_{\mathit{it}}$(1)

where the subscripts i and t represent, respectively, the country and the time period. All variables are expressed in natural logarithms. OA is the number of occupational accidents rate measured as the number of serious accidents per 100.000 persons in employment, X is a vector of countries’ productive system characteristics (employment risk, large firms, temporary contracts), Z is a vector of business cycle controls (per capita GDP, secondary education, fixed investments over GDP, unemployment), and U is a vector of other territorial variables (crime index). Finally, T and FE are, respectively, years and geographical fixed effects, ε is the time-varying error term which stands for a well-behaved error term distributed IID (0, σ²).³

This first approach provides the estimates of the pooled OLS, and the panel fixed (FE) and random (RE) effects models. In order to choose between the fixed and random estimate, both the Breusch-Pagan Lagrange Multiplier Test (1980) and the Hausman Test (Hausman 1978) reveal that the fixed effects model is the most appropriate one. The risk of hysteresis in the perpetration of accidents at work has raised the attention on the phenomenon that todays’ occupational accidents rate can be related to previous ones. Countries past occupational accidents may indeed affect current workplace injuries. For instance, a countries’ occupational accident rate in the previous year can explain the level of adherence to occupational safety regulations in the previous year, which in turn can explain a low standard in occupational safety policy enforcement that can determine in the following year a higher level of occupational accidents.

For these reasons, the relevance of a dynamic path associated to occupational accidents rates suggests that OLS coefficients could be inconsistent due to the correlation of OA_t-1 and the error term, even implementing a first difference approach. Therefore, we complement the standard panel approach with a dynamic panel data analysis (Arellano and Bond 1991; Arellano and Bover 1995; Blundell and Bond 1998). With the implementation of an auto-regressive approach, we are able to enhance the accuracy of the OLS estimates and to include the entire history and transition over time of our model.

In our case, given that the panel units are larger with respect to the time periods, we opt for the system-GMM estimator (Arellano and Bover 1995) that is more efficient (Baltagi 2005) than the difference-GMM estimator (Arellano and Bond 1991). Our choice is also confirmed by comparing the magnitude of the coefficients of the lagged dependent variable from the difference- and system-GMM, with those obtained from the pooled OLS and the panel fixed effects (P-FE).

We test Hp1 by also dealing with the dynamic and simultaneity problems arising by the hysteresis of occupational accidents. In particular, we estimate the baseline specification model in Equation (1)(1) $\mathit{O}{A}_{\mathit{it}}={\beta }_0+{\beta }_1\mathit{Xps}{c}_{\mathit{it}}+{\beta }_2\mathit{Zb}{c}_{\mathit{it}}+{\beta }_2\mathit{Ut}{v}_{\mathit{it}}+T_t+\mathit{F}{E}_i+{\epsilon }_{\mathit{it}}$(1) through a system-GMM (Arellano and Bond 1991; Arellano and Bover 1995) by also including on the right-hand side the lagged dependent variable. The system-GMM estimation takes the following equation form:

(2) $\mathit{O}{A}_{\mathit{it}}={\beta }_0+{\beta }_1\mathit{O}{A}_{\mathit{it}-1}+{\beta }_2\mathit{Xps}{c}_{\mathit{it}}+{\beta }_3\mathit{Zb}{c}_{\mathit{it}}+{\beta }_4\mathit{Ut}{v}_{\mathit{it}}+{\eta }_i+{\xi }_t+{\epsilon }_{\mathit{it}}$(2)

All variables are expressed in natural logarithms. The subscripts i and t represent the country and the time period, respectively. On the right-hand side of Equation (2)(2) $\mathit{O}{A}_{\mathit{it}}={\beta }_0+{\beta }_1\mathit{O}{A}_{\mathit{it}-1}+{\beta }_2\mathit{Xps}{c}_{\mathit{it}}+{\beta }_3\mathit{Zb}{c}_{\mathit{it}}+{\beta }_4\mathit{Ut}{v}_{\mathit{it}}+{\eta }_i+{\xi }_t+{\epsilon }_{\mathit{it}}$(2) , the explanatory variables are the same of previous Equation (1)(1) $\mathit{O}{A}_{\mathit{it}}={\beta }_0+{\beta }_1\mathit{Xps}{c}_{\mathit{it}}+{\beta }_2\mathit{Zb}{c}_{\mathit{it}}+{\beta }_2\mathit{Ut}{v}_{\mathit{it}}+T_t+\mathit{F}{E}_i+{\epsilon }_{\mathit{it}}$(1) plus the lagged dependent variable (OA_t-1) which was included into the model in order to identify the persistency in the dynamics of occupational accidents; η_i is a country fixed effect and ξ_t a time fixed effect; ε_it stands for a well-behaved error term distributed IID (0, σ²).

The system-GMM estimator (Arellano and Bond 1991; Arellano and Bover 1995) controls for both time-invariant country-specific effects and endogeneity criticalities that arise from the lagged value of occupational accidents rate. The validity of the instruments used are tested by the Sargan test of over-identifying restrictions to examine the consistency of the instruments and by the Arellano and Bond (1991) test for the serial correlation of the disturbances up to the second order.

In the subsequent stage of the analysis, we provide an insight on the link between occupational accidents and economic growth by testing the existence of a non-linear and non-monotonic relationship (Hp2). We believe that advancements in technology (i.e., Industry 4.0) can provide solutions able to enhance workers’ health, safety, and wellbeing, enabling the workforce to shift from low- to high-productivity sectors. Thus, countries that have exploited new technologies displace, with respect to the others, higher levels of GDP potentially justifying a cross-country inverted-U-shaped relationship between OA and GDP. Our hypothesis, therefore, is that an increasing GDP can improve social welfare and the overall security of a society; however, because of the different development stages of one country, the economic and technological structures will be different, and their impacts on occupational accidents may also be different, depending on whether occupational safety and health requirements are regarded as necessary, normal or luxury goods. In this perspective, we start by estimating through pooled OLS, fixed and random effects model the following equation:

(3) $\mathit{O}{A}_{\mathit{it}}={\beta }_0+{\beta }_1\mathit{GD}{P}_{\mathit{it}}+{\beta }_2\mathit{GD}{P}_{it}^2+T_t+{\epsilon }_{\mathit{it}}$(3)

where, all variables are again in natural logarithm and i and t represent the country and the time period, respectively. The other elements of the equations have been set out above and are self-explanatory.

We then estimate Equation (3)(3) $\mathit{O}{A}_{\mathit{it}}={\beta }_0+{\beta }_1\mathit{GD}{P}_{\mathit{it}}+{\beta }_2\mathit{GD}{P}_{it}^2+T_t+{\epsilon }_{\mathit{it}}$(3) by recurring to the following system-GMM estimation:

(4) $\mathit{O}{A}_{\mathit{it}}={\beta }_0+{\beta }_1\mathit{O}{A}_{\mathit{it}-1}+{\beta }_1\mathit{GD}{P}_{\mathit{it}}+{\beta }_2\mathit{GD}{P}_{it}^2+T_t+{\epsilon }_{\mathit{it}}$(4)

Finally, in order to ascertain the robustness of our GDP quadratic specification, we also implement the alternative two-stage semiparametric fixed effects additive model (Baltagi and Li 2002):

(5) $\mathit{O}{A}_{\mathit{it}}={\beta }_0+{\beta }_1\mathit{Xps}{c}_{\mathit{it}}+{\beta }_2{T}_{\mathit{it}}+\mathit{g}\left(\mathit{GD}{P}_{\mathit{it}}\right)+{\nu }_{\mathit{it}}$(5)

Where unlike previous models, GDP enters the model nonlinearly. All variables are again expressed in natural logarithms. The specified additive model is built in order to satisfy the stochastic equicontinuity condition and is pointwise asymptotically normal. Thus, it achieves the standard one-dimensional rate of convergence and has the same asymptotic accuracy as if the nuisance terms were known with certainty. Furthermore, for the parametric component, the estimates are asymptotically normal. Note, we estimate g (∙) using b-spline (k = 4). In the parametric part of the estimation, we include the previous set of controls Xpsc and the time.

IV. Estimation results

As widely explained in the previous section, our preferred econometric methodology is the GMM-system estimator that allows us to control for dynamics in occupational accidents rates and joint endogeneity of the explanatory variables and for country-specific effects. For comparative purposes, we also present results obtained by mean of the pooled OLS estimator, which ignores the presence of country-specific effects and treats all variables as exogenous, and the random effects (RE)/fixed effects (FE) estimators, which controls for the presence of unobserved country-specific heterogeneity. The results of the pooled OLS estimates are presented in Table 4; the panel RE/FE and system-GMM results are displayed in Table 5.

Table 4. Pooled OLS estimations (dependent variable: OA rate).

Variables	Model 1	Model 2	Model 3
Empl. Risk	1.583*** (0.337)	2.321*** (0.310)	2.537*** (0.292)
Large Firms	−0.565*** (0.125)	−0.183 (0.116)	−0.242** (0.109)
Temp. Contracts	1.225*** (0.105)	0.532*** (0.136)	0.531*** (0.127)
GDP_PC		0.905*** (0.125)	0.724*** (0.121)
Secondary Edu.		−0.892*** (0.227)	−0.532** (0.221)
Fixed Investments/GDP		−0.531 (0.326)	−0.598* (0.304)
Unemployment		0.469*** (0.145)	0.500*** (0.136)
Crime Index			0.414*** (0.075)
Year	Yes	Yes	Yes
Cons.	1.903** (0.852)	1.798 (2.273)	−0.731 (2.161)
F-stat	14.42***	23.72***	27.20***
R^2	0.41	0.62	0.67
Countries	27	27	27
Obs.	210	210	210

Standard errors in parentheses: ***p < 0.01, **p < 0.05, *p < 0.1.

Table 5. Panel Fixed/Random effects model and system-GMM (dependent variable: OA rate).

Variables	Model 1 FE	Model 2 RE	Model 3 SYS-GMM
OA_Rate (t-1)			0.267*** (0.022)
Large Firms	−0.397** (0.194)	−0.307** (0.136)	−0.197*** (0.073)
Temporary Contracts	0.276** (0.110)	0.401*** (0.085)	0.008 (0.057)
Per capita GDP	1.016** (0.462)	0.810*** (0.162)	0.089 (0.349)
Secondary Education	0.145 (0.321)	−0.236 (0.256)	−0.244** (0.113)
Fixed Investments/GDP	−0.555*** (0.187)	−0.534*** (0.185)	−0.513*** (0.143)
Unemployment	−0.201* (0.111)	−0.250*** (0.093)	−0.145** (0.065)
Crime Index	0.157** (0.069)	0.215*** (0.066)	0.107** (0.050)
Year	Yes	Yes	Yes
Cons.	1.903** (0.852)	2.651 (1.936)	1.798 (2.273)
F-stat or Wald χ^2	4.38***	151.67***	168.69***
R^2	0.65	0.72
AR(1) Pr > z			0.1036
AR(2) Pr > z			0.2780
Sargan test			19.60
Countries	27	27	27
Obs.	210	210	210

Standard errors in parentheses: ***p < 0.01, **p < 0.05, *p < 0.1.

In Table 4, Model 1 shows the results of the baseline model where occupational accidents rate is regressed only against the vector of the chosen covariates relative to the productive-system variables, i.e. employment risk, large firms, and temporary contracts. In models 2 and 3, the business cycle-related explanatory variables (i.e. per capita GDP, fixed investments/GDP, unemployment) and some other socio-economic controls (i.e. secondary education, crime) are added. Overall evidence, at this very first stage of analysis, confirms the consistent effect exerted by all the different sets of covariates on occupational accident rates, across all the estimates. The obtained positive and highly statistically significant relationships support the argument that higher levels of employment risk, temporary contracts, per capita GDP, unemployment, and crime may accompany an increase in the rate of occupational accidents. Likewise, the obtained negative and highly statistically significant relationships allow us to hold that higher levels of large firms, secondary education, and fixed investments/GDP may drive a reduction in the rate of occupational accidents.

As it is shown in Table 5, the obtained panel fixed/random effects models and the system-GMM results confirm, with a high degree of reliability, that the employed approach to the assessment of the occupational accidents rate is adequate. The negative relationship with the firms’ large size is consistent with the hypothesis by main literature (Fabiano, Currò, and Pastorino 2004) according to which the number of accidents declines with the increase of company’s size. It is often argued (Salminen 1993, 1998) that large companies have fewer accidents since they have more knowledge and better financial resources for workplace health and safety investments. With regard to the level of education, the negative and statistically significant relationship is in line with the relevant literature (Cioni and Savioli 2016).

As we can observe, in the system-GMM model, the coefficient of the lagged dependent variable is positive and highly statistically significant suggesting that occupational accidents rate displays persistence over time, i.e. the dynamics and current complex workplace safety conditions can subsequently develop more accidents. The negative and highly statistically relationship with fixed investments/GDP confirms that occupational accidents are influenced by economic cycles; thus, higher levels of fixed investments/economic upturns might reduce the number of occupational accidents. The inverse relationship with unemployment rate is also consistent with relevant literature (i.e. Ferguson, McNally, and Booth 1985; Rasmussen 1987; Vredenburgh 2002; Boone and van Ours 2006; de La Fuente et al. 2014): the idea behind this finding is that, if unemployment is high, workers are reluctant to report accidents because they fear that employers will hold this against them. From a policy perspective, this result implies that, in times of high unemployment, safety auditors should encourage people to report accidents. Finally, the estimates highlight a positive relationship between occupational accidents rate and crime, thus suggesting, in line with our hypothesis, that countries with a higher level of crime are exposed to lower OSH compliance by firms, resulting in an increase of occupational accidents.

Robustness checks

The findings of our estimations, however, could potentially undermine two main issues. On the one side, the effect over the dependent variable of the differences in the propensity to report occupational accidents across our panel of 27 European countries. In particular, the phenomenon of underreporting could vary by national context and, even though constant over the short-time span of our analysis, could affect the panel random effects model estimates. On the other side, the different sectoral composition of the economic and productive systems of the 27 countries included in our panel could lead to a conflation of industry-specific determinants of workplace accidents.⁴

As pointed out in the literature, underreporting of workplace accidents is a multidimensional issue that is strictly correlated to national administrative requirements, institutional frameworks, economics cycle, and cultural heritage (Boone and van Ours 2006; Hall 2016; Picchio and van Ours 2017). Even though we believe these factors to be structural and rigid over time – i.e. especially in the short time horizon of our analysis -, we provide a robustness check with this respect for the random-fixed effects estimation presented in the previous section, by extending our analysis to fatal accidents. Indeed, including fatal accidents in the measurement of our dependent variable reduces to the minimum the risk of measurement error due to the underreporting phenomenon.

In Table 6, we implement our approach using alternatively as dependent variable the fatal occupational accident rate (model 1 and 2) and the total occupational accident rate (non-fatal plus fatal, model 3 and 4). From our estimations results, it is possible to see that when investigating exclusively on fatal work accident rates our model specification remains overall statistically appropriate and robust both in the panel fixed and random effects model estimations. Moreover, in models 3 and 4, to reduce the extent of the possible influence of underreporting with respect to the baseline estimations, we also consider as dependent variable the overall work accidents rate (non-fatal plus fatal accident rates). The overall estimation results, even in this case, provide a sounding confirmation of the findings presented in Table 5: signs and statistical significance of the estimated parameters hold.

Table 6. Robustness check - Panel Fixed/Random effects model (dependent variable: OA rate – fatal and total workplace accidents).

Variables	Model 1 FE OA_Fat.	Model 2 RE OA_Fat	Model 3 FE OA_Tot.	Model 4 RE OA_Tot.
Large Firms	0.364*	0.474***	−0.258	−0.266*
	(0.224)	(0.152)	(0.191)	(0.143)
Temporary Contracts	0.226*	0.073	0.226**	0.369***
	(0.124)	(0.094)	(0.111)	(0.089)
Per capita GDP	−1.712***	−0.776***	1.212***	0.665***
	(0.475)	(0.173)	(0.460)	(0.170)
Secondary Education	−0.873**	−0.604**	−0.064	−0.276
	(0.383)	(0.290)	(0.317)	(0.268)
Fixed Investments/GDP	−0.146	0.045	−0.514***	−0.454**
	(0.225)	(0.211)	(0.190)	(0.187)
Unemployment	−0.241*	−0.008	−0.220*	−0.333***
	(0.132)	(0.088)	(0.113)	(0.092)
Crime Index	0.169**	0.134*	0.129*	0.176***
	(0.083)	(0.073)	(0.070)	(0.067)
Year	Yes	Yes	Yes	Yes
Cons.	22.677***	12.517***	−5.852	−0.281
	(5.454)	(2.881)	(5.081)	(2.703)
F-stat or Wald χ^2	6.63***	48.72***	3.90***	115.62***
R^2	0.24	0.27	0.58	0.67
Countries	27	27	27	27
Obs.	210	210	210	210

Standard errors in parentheses: ***p < 0.01, **p < 0.05, *p < 0.1.

Moreover, in Table 7, in order to address more carefully the differences in the composition of the productive and economic systems among our 27 European countries, to reinforce the validity of our baseline estimation, we furtherly implement our analysis by focusing on the two sectors that exhibit the highest OA rates (Takala 2019): construction and manufacturing. As shown in Table 7, considering both panel fixed and random effects models, the findings provide a further confirmation (in terms of statistical significance, signs and magnitude) of the results presented in the baseline estimations in Table 5.

Table 7. Robustness check - Panel Fixed/Random effects model (dependent variable: construction OA rate).

Variables	Model 1 FE OA_Const.	Model 2 RE OA_Const.	Model 3 FE OA_Manu.	Model 4 RE OA_Manu.
Empl. Construction	0.571**	0.590***	0.652	0.877***
	(0.223)	(0.210)	(0.398)	(0.309)
Large Firms	−0.398*	−0.320*	−0.342*	−0.444***
	(0.209)	(0.165)	(0.201)	(0.145)
Temporary Contracts	0.256**	0.373***	−0.055	0.210**
	(0.124)	(0.103)	(0.115)	(0.091)
Per capita GDP	0.565	0.720***	0.860*	0.646***
	(0.464)	(0.202)	(0.442)	(0.171)
Secondary Education	0.436	−0.029	0.381	−0.116
	(0.400)	(0.338)	(0.361)	(0.286)
Fixed Investments/GDP	−0.470**	−0.531***	−0.302*	−0.305*
	(0.210)	(0.199)	(0.174)	(0.172)
Unemployment	−0.379***	−0.366***	−0.466***	−0.511***
	(0.118)	(0.098)	(0.111)	(0.096)
Crime Index	0.344***	0.372***	0.245***	0.292***
	(0.082)	(0.078)	(0.077)	(0.074)
Year	Yes	Yes	Yes	Yes
Cons.	−0.023	0.099	−2.318	1.001
	(4.999)	(2.825)	(4.707)	(2.443)
F-stat or Wald χ^2	6.97***	166.55***	9.59***	181.38***
R^2	0.58	0.65	0.33	0.54
Countries	27	27	27	27
Obs.	210	210	210	210

Standard errors in parentheses: ***p < 0.01, **p < 0.05, *p < 0.1.

Occupational accidents and growth

Ultimately, with regard to the shape of the relationship between occupational accident rate and per capita GDP level, we test Hp2, employing the same set of estimation strategies. The existence of a non-linear and non-monotonic relationship will imply that for each unit increase of per capita GDP, workplace accidents will increase at a decreasing rate, generating greater occupational safety gains. Our final estimation results are presented in Table 8.

Table 8. Testing the existence of a non-linear and non-monotonic relationship between GDP and occupational accidents rate (Dependent variable: OA rate).

Variables	Model 1 OLS	Model 2 FE	Model 3 RE	Model 4 SYS-GMM
GDP_PC	19.707*** (1.678)	12.468*** (2.050)	10.663*** (2.238)	6.370*** (1.230)
GDP_PC^2	−0.939*** (0.084)	−0.569*** (0.104)	−0.464*** (0.115)	−0.334*** (0.061)
OA_Rate (t-1)				0.426*** (0.005)
Year	Yes	Yes	Yes	Yes
Const.	−95.840*** (8.340)	−60.609*** (10.146)	−53.090*** (10.992)	−26.330*** (6.146)
F-stat or Wald χ^2	38.77***	9.07***	127.23***	6803.97***
R^2	0.61	0.54	0.59	-
Countries	27	27	27	27
Obs.	210	210	210	210

Standard errors in parentheses: ***p < 0.01, **p < 0.05, *p < 0.1.

We find evidence of an inverted U-shaped relationship between the rate of OA and the real per capita GDP, in a time of transition towards a new technological paradigm in Europe. This is an important consideration if economic growth policies are to be used as a strategy to improve workplace safety and health measures. The implementation of macroeconomic expansionary policies would have a positive impact on workplace environments, increasing investments in technological innovation, new machineries and safer equipment, higher training of workers, enhancing the overall occupational safety levels and the quality of the workplace systems. The results are robust to endogeneity linked to the transitional dynamics and persistency of the OA rate over time. The transition towards the new Industry 4.0 paradigm, currently in progress in the time period of our analysis, generates a skill biased technological change (SBTC) that, if exploited, could provide solutions able to enhance workers’ health, safety and wellbeing, reducing OA rates.

To provide a further robustness check to the reliability of the quadratic specification, we estimate Equation 5(5) $\mathit{O}{A}_{\mathit{it}}={\beta }_0+{\beta }_1\mathit{Xps}{c}_{\mathit{it}}+{\beta }_2{T}_{\mathit{it}}+\mathit{g}\left(\mathit{GD}{P}_{\mathit{it}}\right)+{\nu }_{\mathit{it}}$(5) using a semiparametric fixed effects estimator (Baltagi and Li 2002). Therefore, we estimate the occupational accidents rate allowing the GDP to enter the model nonlinearly. In Figure 2, we plot the nonparametric b-spline estimate of g(GDP); we can observe that the nonparametric estimate (red line) depicts a clear inverted U-shaped curve. This result provides an ‘unconstrained’ sounding reassessment of the non-linear relationship between GDP and occupational accidents rate obtained with the implemented parametric approaches.

Figure 2. Semiparametric Fixed effects model (Baltagi and Li 2002).

V. Conclusions

Our results can give rise to some relevant policy implications. Overall, our cross-country analysis reveals that the heterogeneity of OA rates is significantly correlated to productive systems characteristics, to the business cycle and to some other territorial variables. Based on our findings, the improvement of both macro- and micro-economic conditions should be the priority for new policies on work safety. In particular, policymakers and employers should be aware of the correlation between the business cycle and the occurrence of workplace accidents, which calls the need for additional safety measures during economic expansions. Although definitive conclusions cannot be drawn from the results of this study, some interesting insights do emerge. Firstly, in relation to the production system controls we can highlight the following findings: the higher is the share of large firms, the lower is the OA rates (confirming previous empirical literature); the existence of a positive, even though feeble, correlation between temporary contracts and OA (in line with Cioni and Savioli 2016); the higher is the employment risk attached to the productive sector, the higher is the OA rate. Secondly, in relation to the business cycle controls: the OA rate increases with real per capita GDP (confirming previous empirical literature); the higher is the level of education, the lower is the OA rate (Fotta and Bockosh 2000; Jeong 1999; Kletz 1993; de La Fuente et al. 2014; Cioni and Savioli 2016); the higher is the propensity to invest in fixed assets, the lower is the OA rate (Blank, Laflamme, and Diderichsen 1996; Sari et al. 2004; Asfaw, Pana-Cryan, and Rosa 2011). Finally, the estimations reveal that countries with a higher level of crime generate lower OSH compliance by firms, thus determining an increase in the number of occupational accidents.

Taken all together, our findings imply that OSH targets, besides pursuing the aims established in principle 10 of the European Pillar of Social Rights (Healthy, safe and well-adapted work environment and data protection) are crucial drivers for improving the sustainability and competitiveness of the EU economy.

Our findings call for a main policy recommendation. According to our estimates, fixed capital investments, hazardous and labour-intensive workplaces, and firms’ size, represent the main determinants of OA rates. This implies that the risk of underinvestment in OSH of small and medium enterprises (SMEs), especially during economic and financial downturns – as experienced throughout the 2007–2008, the 2012 sovereign debt, and the COVID-19 pandemic crises – becomes higher and socially concerning. This raises the issue whether or not relying solely on regulatory remedies and sermons is still sufficient. In our view, especially in recession periods, it becomes urgent to implement carrots policies (i.e. direct and indirect public aid schemes) to support OSH investments. In particular, in this vein, Italy is adopting this type of policy option throughout a specific initiative implemented by the Italian National Institute for Insurance against Accidents at Work (INAIL), providing direct financial grants to SMEs for both tangible and intangible OSH investments.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The work was supported by the Istituto Nazionale per l’Assicurazione Contro Gli Infortuni sul Lavoro [Bric Inail Vip-Moving].

Notes

¹ Hämäläinen, Saarela, and Takala (2009) states that although globally the number of accidents has risen slightly, the trend in Europe is declining. One of the reasons for this decrease could be the introduction of stricter occupational health and safety legislation. Moreover, Takala (2019) confirms the downward trend and lists a number of reasons: improvements in technology, processes and methods, better leadership, management and safety and health efforts, a reduction in the number of workers in hazardous industries and changing economic structures. Again, mechanization, automation and prefabrication processes are oriented to create jobs that are less prone to injury.

² It includes all accidents in the course of work, whether they happen inside or outside the premises of the employer, on the premises of another employer, in public places or during transport (including road traffic accidents or accidents in any other mean of transportation) and at home (such as during teleworking); it also includes cases of acute poisoning and wilful acts of other persons. Instead, it excludes: accidents on the way to or from work (commuting accidents); deliberate self-inflicted injuries; occurrences caused solely by a medical condition (such as heart attack or stroke) that occurred during work, i.e. which were not (at least partially) caused by the occupational activity of the victim; accidents to members of the public, for example family members of a worker who are not working; and occupational diseases. https://ec.europa.eu/eurostat/cache/metadata/en/hsw_acc_work_esms.htm..

³ A further approach for dealing with the cross country differences would imply the inclusion of a full set of time/country (or other covariate/country) interaction terms. However, this option is not feasible in our setting due to multicollinearity, as it is confirmed by the VIF test (well above the threshold of 10).

⁴ We would like to thank an anonymous referee for pointing this out.