Extended Summary

Keywords

Occupational Accidents, Workplace Accidents, Accidents at work, Workplace injuries, Determinants, Factors, Cost, Occupational Safety, Occupational Risk, Commuting Accidents, Accident Frequency, Accident Severity

Occupational Accident (OA) in Belgium impose a substantial financial burden. In the literature overview (see 1  Literature Overview), we mentioned that in 2022, costs in the public sector alone were about €10 million. If similar patterns hold in the private sector, total costs could exceed €50 million, underscoring the need for stronger prevention to protect workers and reduce economic loss.

A synthesis of research (see 1.3.7 Summary Overview Determinants) indicates that OA occurrence and/or severity are shaped by temporal variation (years, seasons, months, days, hours, holidays, vacations), sectoral differences, individual factors (age, sex, work experience, nationality, salary level, residence, commuting distance, Personal Protective Equipment (PPE) use, health problems, sleep disorders, substance use, personality traits, years of schooling), job-related factors (occupation, workforce segment, contract type, shift work and long hours, risks), and organisation-related factors (work location, company size, psychosocial risks, prevention investments). Organisation-related factors mainly exert an indirect influence by shaping job conditions and behaviours; job-related factors have a direct impact; individual factors interact with both. An effective prevention strategy must address all three.

While the relationship between hazard identification and risk assessment is well studied, gaps remain in explaining why some risks lead to OA while others do not (see 1.4 Identifying and addressing knowledge gaps on occupational accidents). A promising way forward is to analyse a large, structured dataset containing diverse variables for individuals across occupations both with and without OA. Liantis can compile and analyse such a dataset.

The construction and quality of this dataset are described in detail in the current project’s data quality report (see 2  Data Quality Report).

After identifying the stakeholders in the OA data landscape (see 2.1 On the stakeholder landscape of occupational accident data), we assessed how Liantis data could serve two aims: (1) develop a clear view of the relative importance of determinants of workplace accidents on selected outcomes, and (2) communicate clear messages to employers, employees, and policymakers.

First, we examined the Liantis customer base. A general overview of customers between 2014 and 2023 (see 2.2 A general overview on the employers covered in the different source datasets) shows that Liantis External Service for Prevention and Protection at work (EDPB in Dutch) (ESPP) delivered prevention-related prestations for 69,157 unique employers, and Liantis Payroll Services (PS) calculated wages for 79,723 unique employers. In total, 47,820 employers were mutual Liantis ESPP and PS customers.

Through Federaal agentschap voor beroepsrisico’s (FEDRIS)’ automated messaging flow, Liantis ESPP received 293,938 accident declarations (covering 161,696 employees) from 20,636 unique employers. Of these, 90,619 declarations (about one third) came from 12,407 employers who were mutual Liantis ESPP and PS customers (representing 52,240 employees).

For analyses of OA occurrence, notification records from employees of mutual customers should be considered and related to the entire mutual customer base (including employees without any OA). For analyses of OA severity, declarations from the >161,000 workers who experienced an OA suffice.

The feasibility of using OA notification records was extensively investigated and confirmed in the data quality report.

We conclude (see 2.10 Conclusions about the potential of OA notification records) that the 293,938 raw OA records initially received did not all correspond to unique OA. After validation -conducted with the help of FEDRIS- the number was reduced to 218,034 unique validated OA records (74.18%). After matching with the Liantis database, 216,946 unique OA records (99.5%) were retained for analysis.

These validated notification records are not sufficient on their own. Several additional datasets (see 2.20 Conclusions about the additionally gathered datasets) were collected to enable in-depth analyses of the OA determinants described in the literature overview. We assembled detailed worker-level information -health complaints, risk factors, PPE use, working hours, absences, and wages- as well as employer-level data. We also enriched the dataset with external factors such as weather conditions, public holidays, and Coronavirus Disease 2019 (COVID-19) restrictions, resulting in a comprehensive, multifaceted dataset.

Combined with the validated OA notifications (see 2.10 Conclusions about the potential of OA notification records), these additional datasets (see 2.20 Conclusions about the additionally gathered datasets) enabled an in-depth exploration of underlying patterns and determinants of OA.

The first part of the data analytics report (see 3  Data Analytics Report) acknowledges that the Liantis customer base is not fully representative of all Belgian employers and employees. Many issues can, however, be addressed after appropriate representativeness checks.

From 47,820 mutual employers, we drew a sample of 10,000 (see 3.1 Representativeness of the employer data) representative by Flemish province, sector (excluding B, D, E, O, U, and T), and company size (< 50 vs. \(\geq\) 50 employees). Including a variable in statistical models indicating whether an employer (and its employees) belong to this representative sample helps correct potential interpretation bias.

The statistical methodology section (see 3.2 Statistical methodology) details the analyses. For each outcome describing occurrence (monthly counts of all notifications, accepted cases, commuting OA, workplace OA, and company-level frequency degree; see 3.3 Outcomes) or severity (serious workplace OA, absence length, direct costs, and company-level severity degree; see 3.3 Outcomes), we present descriptive, univariable (see 3.4 Descriptive and univariable analysis of potential determinants), and multivariable analyses (see 3.5 Multivariable analysis of potential determinants).

In the univariable section, determinants identified in the literature are examined one by one. We first present descriptive statistics and, where applicable, include external references (e.g., National Social Security Office (NSSO), FEDRIS). Next, each determinant is modelled against selected outcome variables (see 1.2 Outcomes of occupational accidents). Conclusions are provided per determinant, with the caveat that confounding may exist and findings should be confirmed in multivariable analyses.

However, not all determinants could be examined using this same analytical approach.

Variables such as personality traits, number of years of schooling, shift work, long working hours, safety climate, and safety culture were not investigated any further due to the absence of systematically maintained data for both workers who did and did not experience an OA. For variables like usual professional activity, specific activity at the time of the OA, type of workstation, and subcontracting, data were only partially available and limited to workers who had experienced an OA. These variables were therefore excluded from further analysis. The dataset on personal protective equipment was too limited to be used in modelling.

Some other determinants were explored in preliminary analyses, including descriptive statistics, \(\chi^2\) tests, and selected univariable models. These variables were not included in the multivariable models because access to technical tests such as audiometry (used to investigate hearing quality, see 3.4.1.9 Health problems: quality of hearing), or General Medical Questionnaire (AMV in Dutch) (GMQ) questionnaires (assessing work stress, job satisfaction, and social relationships, see 3.4.3.4.1 Work stress, 3.4.3.4.2 Job satisfaction and 3.4.3.4.3 Work relationships) was not uniform across workers. Results from these tests depend on assigned risk categories, which vary widely and are difficult to compare (see 3.4.2.5.4 Risks (ESPP)). Access to these tests depends on whether a worker attended a preventive medical consultation and results of the GMQ coupled to this consultation are only available for the last two years of the ten-year study.

Temporal patterns and time invested in prevention were examined using dedicated analytical approaches.

For day and hour-level effects, we do not have corresponding exposure data available for workers who did not experience an accident. Therefore, we used a generalized additive model to investigate temporal patterns such as the impact of extreme weather, holidays and vacations, seasonal time changes, the COVID-19 pandemic, and terrorist attacks (see 3.4.5.10 Interplay of temporal patterns). Both commuting and workplace OA drop significantly on Saturdays, Sundays, and during school/legal holidays, reflecting reduced activity on these days. Harsh weather (like snow or freezing temperatures) increases commuting OA more than workplace OA. Stricter COVID-19 measures were linked to fewer OA, especially for commuting. Shifts to and from daylight saving time showed no clear link to more OA.

Regarding time invested in prevention, no data were available on non-Liantis ESPP activities, other financial investments, or individual-level time registration. Therefore, Signal Extraction in Arima Time Series (SEATS) analysis was performed on monthly aggregates of commuting OA, workplace OA, and time invested in prevention of OA (excluding any post-OA timeregistrations) across all Liantis ESPP clients (see 3.4.3.6 Investments in prevention). Time series decomposition using SEATS, revealed distinct seasonal peaks in both commuting and workplace OA, suggesting potential for targeted interventions during high-risk months. The analysis confirmed that the observed patterns are not purely random. Workplace OA show a clear and encouraging downward trend, which appears to coincide with a collective increase in preventive investments across the Liantis ESPP client portfolio. The aggregated monthly data was further examined using ordinary linear regression. A negative relationship was observed -more time invested in prevention was associated with fewer workplace accidents- but no causal conclusions can be drawn from this type of analysis.

In the multivariable section, all available and suitable determinants (excluding those not investigated any further, those explored in preliminary analyses only, and those examined using dedicated analytical approaches) were combined in a multivariable analysis (see 3.5 Multivariable analysis of potential determinants) to examine the selected outcomes individually. A final summary compares the multivariable models (see 3.5.14 Summary overview of the discussed multivariable models). Depending on the model, between 9,822 and 7,938,537 complete-case observations were available, representing 14,198 to 291,661 employees and 2,946 to 16,362 employers. Based on R²-related parameters (marginal, conditional, and ratio), some models performed well (e.g., for occurrence: likelihood of an accepted workplace OA; for severity: absence duration and cost of an accepted workplace OA), while others performed poorly (e.g., for occurrence: likelihood of refusal or the company-year frequency degree; for severity: likelihood of classification as serious or the company-year severity degree).

The multivariable model for the likelihood of notifying an OA (see 3.5.1 Likelihood to notify an OA (hadOA, model 1)) highlights significant effects of individual demographics (age, gender), employment conditions (wage, experience, company size), and temporal factors (year and month). Wage and company size point to structural and economic influences, while the temporal decline may reflect broader labour market trends or policy changes. Accounting for both individual- and company-level variation is essential.

The multivariable model for the likelihood to refuse an OA (see 3.5.2 Likelihood to refuse an OA (statusOA, model 1.1)) shows strong fit at the individual level but weak explanatory power from fixed effects, including insurer identity. Individual-level heterogeneity dominates; company affiliation has no measurable role. The model’s utility for identifying systematic refusal patterns is limited by the absence of OA-specific variables. Apart from 2017, no significant insurer differences were observed, suggesting a broadly level playing field. Ongoing transparency and monitoring remain important.

The multivariable model for the likelihood to experience an insurer-accepted commuting OA (see 3.5.3 Likelihood to experience an accepted commuting OA (comOA, model 2)) indicates that individual-level heterogeneity outweighs wage, time, and other fixed effects. Risk appears to depend more on personal differences than on salary, seasonality, or other characteristics. Some individuals are simply more likely to experience such OA regardless of employer or earnings. Prevention should therefore emphasise individual risk factors, while companies can add value through targeted risk communication, for example, focusing on higher-risk groups with increased efforts in winter and early spring.

The multivariable model for the likelihood to experience an insurer-accepted workplace OA (see 3.5.4 Likelihood to experience an accepted workplace OA (wplOA, model 3)) shows higher risk among men, younger to middle-aged workers, those on full-time contracts, and employees in very large firms, and lower risk among those with longer tenure, white-collar status, certain sectors, many months, and more recent years relative to their reference categories.

The multivariable model for the company-level frequency degree (see 3.5.6 Frequence degree of workplace OA (freqOA, model 4)) indicates that variation is largely random. Although some determinants are statistically significant -insurer, sector, company size, and, among employee fractions, contract nature, workforce segment, and salary level- the practical impact is questionable. With 90.8% unexplained variance, the model has limited practical utility and cautions against using Total Recordable Incident Rate (TRIR)-like measures such as the frequency degree at company level.

The multivariable model for the likelihood to experience a serious workplace OA (see 3.5.7 Likelihood to experience a serious workplace OA (seriousOA, model 5)) suggests that company size and contract nature are associated with classification as serious, but individual-level variation and accident-specific factors are likely more important. The model fits well at the individual level but fixed effects have weak explanatory power. Company affiliation plays only a minor role. The absence of OA-specific variables limits the model’s utility for identifying systematic patterns.

The multivariable models for the number of days absent due to a workplace OA (see 3.5.8 Days absent (validated) due to a workplace OA (nDaysTAO, model 6) and 3.5.9 Days absent (calculated) due to a workplace OA (lengtecor, model 6.1)) show that age, sector, and company size are useful determinants for estimating absence duration, even without OA-specific information. These models have high explanatory power. Company-level heterogeneity matters and must be accounted for. Critical outcomes -final absence duration, total cost, final dossier status, and seriousness- are currently not shared across stakeholders in a structured way, limiting evidence-based prevention and reintegration/return-to-work efforts. Systematic outcome sharing would benefit all parties.

The multivariable model for the direct wage cost of an OA (see 3.5.11 Direct wage cost due to a workplaceOA (costOA, model 7)) shows that individual-level factors (age, experience, workforce segment, salary level, contract nature), employer-level factors (sector, size, insurer), and temporal factors (year, month) significantly determine direct wage costs. Although accident-specific parameters (not included) likely have substantial effects, group-level estimates based on these determinants can predict cost profiles. Company-level heterogeneity must be considered in OAcost estimations.

The multivariable models for the severity degree (see 3.5.12 Severity degree (validated absence) of workplace OA (sevOAval, model 8) and 3.5.13 Severity degree (calculated absence) of workplace OA (sevOAcal, model 8.1)) show that company-specific and yearly severity degrees are highly random (>98% unexplained). While insurer, sector, company size, and other variables may be statistically significant, their explanatory value is limited. These models are scarcely useful in practice and reinforce caution when using such safety measures at company level.

Overall (see 3.6 Strengths and limitations of the study), the study’s strengths include its breadth and scale, adjustments for representativeness, and effective use of mixed modelling. Integrating external data (NSSO, FEDRIS) enables comparison and contextualisation. Including an insurer variable yielded new insights that merit replication. Limitations include the absence of OA mechanism-level analyses. Such data are, de facto, unavailable for individuals without OA, which would have excluded modelling likelihoods as done here. The complete-case requirement leads to waterfall loss, so only subsets can be included depending on the model. Underreporting also remains an issue, though there is no reason to believe this study is more affected than others.

Having achieved the first aim -developing a clear view of the relative importance of determinants on selected outcomes- through the data quality and data analytics reports, we believe the practical compendium (see 4  Practical Compendium) achieves the second aim: communicating clear messages to employers, employees, and policymakers. This chapter summarises key insights for these audiences.

Many insights from this report can be translated into training materials for employers, prevention advisors, HR professionals, and sector-specific stakeholders. Developing concrete recommendations for dissemination and training content (see 5  Recommendations for Knowledge Dissemination and Training) was outside the project’s scope, but the findings provide a strong foundation for future initiatives.

With support from practice experts and training professionals, this knowledge can be converted into accessible formats like workshops, e-learning modules, toolkits, and sector-specific guidance. These materials can help bridge the gap between data-driven insights and everyday prevention practice, ensuring key messages reach those positioned to act.

To maximise impact, training should be tailored to sector-specific needs and roles, aligned with operational realities, enriched with real-world examples, and integrated into existing training frameworks. Effectiveness should be evaluated using well-defined performance indicators. Lead metrics -such as safety training completion, near-miss reporting frequency, audits and inspections, implementation of personal and collective prevention measures, safety engagement, and hazard identification and risk reduction- offer more immediate feedback than lagging metrics like accident rates, and can support a more evidence-based approach to occupational safety and health management.

We hope the report’s recommendations will soon be translated into usable knowledge for the various actors involved. Doing so will strengthen understanding of workplace risks and prevention strategies and contribute to reducing OA and their associated human and financial impacts.

List Of Acronyms

COVID-19

Coronavirus Disease 2019

ESPP

External Service for Prevention and Protection at work (EDPB in Dutch)

FEDRIS

Federaal agentschap voor beroepsrisico’s

GMQ

General Medical Questionnaire (AMV in Dutch)

NSSO

National Social Security Office

OA

Occupational Accident

PPE

Personal Protective Equipment

PS

Payroll Services

SEATS

Signal Extraction in Arima Time Series

TRIR

Total Recordable Incident Rate