Scheduling a Healthier Workforce

(1)

Scheduling a Healthier Workforce

Exploring unhealthy

roster

characteristics from historic data

MSc. Thesis

M.H.W. Cornel

S2738929

10-06-2020

Rijksuniversiteit Groningen Faculty of Economics and Business MSc. Technology and Operations Management

(2)

2

Preface

(3)

3

Abstract

A healthy workforce is important from a societal, organizational, and employee perspective since an unhealthy workforce leads to sickness absence resulting in increased costs from all perspectives. Current literature concludes that shift work is a major cause for an unhealthy workforce. However, less is known about the role of detailed roster characteristics in this relationship. A datamining approach using historic roster and absence data of a major Dutch airline is used to explore potentially harmful roster characteristics. First, classification models using time series of roster data were made to classify short-term sickness absence, which showed up to 64% of predictive accuracy. Moreover, it was found that the results heavily depended on the type of rosters. Likewise, conditional logistic regression models showed the same dependency on the type of rosters. More specifically, the mean duration of shifts is the most consistently associated harmful roster characteristics for rosters without night shifts, whereas the number of rotations is this for rosters including night shifts. Besides, the length of the included rosters also proved to be an important factor in explaining sickness absence. Because this thesis focused on discovering new associations between rosters and health, the exact underlying mechanisms are still unknown. Therefore, this thesis mainly provides directions for further research for both practice and academic rather than direct suggestions for roster improvements.

Keywords: shift work, detailed roster characteristics, health effects, sickness absence,

(4)

4

Introduction

According to data from World Population Prospects: the 2019 Revision, by 2050, one in six people in the world will be over age 65 (16%) compared to one in 11 in 2019 (9%). Expectations for Europa and Northern America are even worse, where one in four persons could be aged 65 or over by 2050 (WHO, 2019). Population aging effects various areas of society, of which the production and services system is one. The challenge for the production and services industries boils down to the fact that more work has to be done with fewer people (Lisenkova, Mérette and Wright, 2013). Consequently, personnel will become one of the scarcest, and thus valuable, resources for organizations. Therefore, it has never been more important to improve employee health and reduce sickness absence without worsening business performance. To accomplish this, practices in which these aspects are integrated need to be reevaluated. One of these is personnel scheduling.

Personnel scheduling, or rostering, is the process of constructing work timetables for its staff so that an organization can satisfy the demand for its goods or services (Ernst et al., 2004). It is extremely difficult to find good solutions to these highly constrained and complex problems and even more difficult to determine optimal solutions (Ernst et al., 2004). Many constraints and objectives are considered in the scheduling process, which could be related to health and/or sickness absence. However, in the extensive literature review of Van den Bergh et al. (2013), only one out of the set of 291 articles about personnel scheduling emphasizes the health and safety issues of workers. Likewise, empirical research has shown that the effect of rosters on the employees’ health and sickness absence only receives very little attention in the scheduling process in practice (De Snoo, Van Wezel and Jorna, 2011).

(7)

7

only found strong evidence between fixed night shifts and sick leave and acknowledge the lack of detail on roster characteristics in studies (Merkus et al., 2012). These limitations may be part of the reason why up to this point scheduling theory and occupational health have not yet successfully found each other. To gain a better understanding of the influence of rosters on health the following question is stated: What roster characteristics are detrimental to

health?

The contribution of this thesis is threefold. To the best knowledge of the researcher, the use of detailed roster characteristics and data mining is limited in the current literature on this subject. First, using detailed roster characteristics contributes to the explanation of inconsistencies between studies and may identify relationships that are more detailed. Second, the use of data mining explores the possibilities and value of using this method in this field of research. Third, the findings of this thesis could be tested in organizations and thereafter used to predict health consequences and eventually improve rosters. Therefore, this thesis also has a practical contribution for organizations using shift work rosters and/or experiencing high rates of sickness absence.

(8)

8

2

Theoretical Framework

2.1 Personnel Scheduling

2.1.1 Scheduling classifications

The first classifications used for personnel scheduling by Baker (1976) and Van den Bergh et

al. (2013) displays a general overview of the kind of scheduling problems. The first is shift

scheduling in which different shifts have to be scheduled across a daily planning horizon. The second, day off scheduling, has to deal with an inconsistency between the facility’s operating week and employee’s working week. The third, which is tour scheduling, is a combination of the first two classifications and the most relevant in this case for two reasons. First, more and more industries and companies are operating 24/7, also known as the “24 Hour Society” (Härmä and Ilmarinen, 1999), and thus are dealing with tour scheduling problems. Secondly, both shift work and irregular rest days are known to harm employee health (Costa and Sartori, 2007; Härmä and Kecklund, 2010) and thus tour scheduling problems are hypothesized to encounter the most health-related problems.

Another classification for personnel scheduling is used by Ernst et al. (2004). This classification is more process orientated and defines personnel scheduling as a sequence of modules. According to Ernst et al. (2004), the scheduling process starts with demand modeling. The next two modules are days off scheduling and shift scheduling. However, when considering tour schedules, these modules are combined. This is because of both modules must be considered together since a tour schedule is a combination of a days off schedule and shift schedule. Therefore, the next module, namely line of work construction, is then executed. This module constructs a roster, which considers both the demand pattern and complexities opposed by the organizational context. The next modules are the task assignment and the assignment of individual staff to the rosters. Those ensure that the staff with the right skills are linked to the lines works or rosters. In this view, the line of work construction is most interesting, since different goals and complexities are balanced in this module, which mainly determines the roster and thus its characteristics.

2.1.2 Schedule product

(9)

9

an overview of the first four weeks of several common shift systems. Whereas working in the day shift system is logically considered day work, working in some other kind of shift system is considered shift work. The exact definition of shift work will be elaborated on in the next section. As can been seen in this figure, rosters can differ in the type of shifts that are worked, the number of shifts of the same type worked consecutively, the sequence in which different types of shift follow each other and the different days and number of days an employee is free from work. Although this is not visible is this simple representation of several rosters, shifts could also have different durations and start times. All of these elements are considered as roster characteristics in this thesis. Since there are many different combinations of roster characteristics possible, many different rosters exist.

In order to create an acceptable roster, many complexities must be taken into account. Whereas some complexities are very straightforward to implement, such as the work rules, some are more problematic. For instance, employee fairness and preferences. A question such as "What is fair?" is difficult to answers and preferences may largely vary per person. The same vagueness is observable when asking: “What is an unhealthy roster?”. To overcome such problems as much as possible, the relationships between rosters, health, and absenteeism will be explored and explained first. Thereafter roster characteristics could be identified that may harm the health and increase the absenteeism of employees.

2.2 Rosters, Health, and Sickness Absence

Based on recent review articles it can be stated that shift work, opposed to regular workdays, can be linked with a higher risk of negative health outcomes for physical, mental and social

(10)

10

issues (Kecklund and Axelsson, 2016; Arlinghaus et al., 2019; Moreno et al., 2019; Zhao et al., 2019). Although the current literature mainly focuses on the negative health effects of shift work, for the overall picture it is important to mention that positive effects are also linked to shiftwork (Keller, 2009). Examples of this are working a second job or extra free time during day hours. Nevertheless, this thesis will build on the negative effects of shifts work. Therefore, several problems are identified in the literature that made it difficult to provide statements about specific harmful roster characteristics and establish causal relationships up to this point.

First, many potential factors could lead to a higher risk of negative health outcomes besides effects from rosters (Moreno et al., 2019). For instance, genetic characteristics or type of job. Second, the term “shift work”, which is often used, has a broad meaning and is hard to conceptualize; besides shift work is often used as a binary variable (Zhao et al., 2019). Most studies only distinguish between either working solely day shifts or working in some kind of shift system. Therefore using this term make it hard to identify cause-effect relationships, since many different shift systems exist as was shown in the previous section (Moreno et al., 2019). Third, systematic reviews and meta-analyses show inconsistency between studies (Kecklund and Axelsson, 2016; Arlinghaus et al., 2019). For instance, Vyas et al. (2012) found that shift work is related to myocardial infarction and ischemic stork, whereas Frost, Kolstad and Bonde (2009) concluded that limited evidence existed for a causal relationship between shift work and heart diseases. Consequently, relationships and pathways between health and rosters need careful interpretation and must be seen as potential explanations rather than proven facts.

2.2.1 Shift work

As noted before shift work is often used as the independent variable and, for already mentioned reasons, this concept needs more clarification. An early definition by Costa (2003) states the following: “We use the term ‘shift work’ to refer in general to a way of organizing

daily working hours in which different persons or teams work in succession to cover more than the usual 8 h day, up to and including the whole 24 h.” More recently Arlinghaus et al. (2019)

adopted a definition from Costa (2010) concluding on the earlier definition saying: “Shift work

(11)

11

Consequently, by defining “standard” also “non-standard” working hours are associated with shift work. Although those concepts largely overlap, they are not the same. Non-standard working hours also include on-call work, part-time work, and flexible working hours (Costa, 2010), which are not necessarily shiftwork when considered solely, but in combination may be important factors. An example of this is being on-call during the day shift (between approximately 7:00 to 18:00) and begin on-call during evening or night shift (between 18:00 and 7:00).

Since the term shift work is initially useful in this research for identifying possible relationships between rosters and health effects in literature, a broad definition is suitable to lower the chance of accidentally excluding relationships. Therefore, the following definition based on the literature discussed above will be used. Shift work involves working at least outside typical or “standard” working hours, which are defined as Monday to Friday, between approximately 08:00 and 18:00h.

2.2.2 Health outcomes

The literature distinguishes roughly three categories of increased risks of health-related outcome variables. Most often, these risks are express using the relative risk or odds ratio. The relative risk compares the risk of an event or outcome between two groups in which one is exposed to some kind of negative factor whereas the other groups is not. The odds ratio compares the presence to the absence of a certain exposure given that the outcome is already known. In this section, the exposure is mainly some type of shift work and the outcome is some kind of health effect or sickness absence.

(12)

12

exposure is also important. Increased odds ratios (OR) are also found for type-2 diabetes (OR=1,09), with a greater risk for males (OR=1,37) and rotating shifts (Gan et al., 2015). The increased risk for rotating shifts and high relative risks associated with night work is in line with the assumption that specific roster characteristics may be more important in determining the health effect than others. Lastly, obesity is also associated with shift work (van Drongelen

et al., 2011; Proper et al., 2016). However, the lack of case-control studies on this outcome

makes it impossible to determine the increased risks.

The second category is the risk of mental health issues. Moreno et al. (2019) identified two meta-analyses relating night work to depression (RR =1,42-1,43) (Angerer et al., 2017; Lee et

al., 2017). Zhao et al. (2019) did a more specific literature review explicitly focused on mental

health issues. They found that the majority of studies focus on general mental health (psychological distress) and/or depression. Moreover, their review demonstrated reasonable evidence in the existing literature between shift work measured as a broad binary indicator and mental health, but a lack of consistency exists when studies are categorized into specific shift work types. Whereas the evidence for working night shifts for several years and mental health problems is strong, only a little evidence exists for other types of shift work such as rotating shifts. Again, this suggests that one should look at specific roster characteristics instead of shift work as a broad variable.

(13)

13

conflicts exist as the number of different family situations. This also makes it difficult to conclude on increased risks between shift work and social health that can be generalized.

2.2.3 Sickness absence

Merkus et al. (2012) did the first literature review on the relationship between shift work and sickness absence. Instead of using sickness absence, they used the term sick leave, which is defined as: “absence from work that is attributed to sickness by the employee and accepted

as such by the employer” (Merkus et al., 2012). This definition is adopted for sickness absence,

because of its clarity and ease of implementation in a measure. They only found an increased risk (RR=1,30) between permanent evening work and sick leave in female healthcare. However, most studies did not include information about the continuity of the shift systems (including weekends or not) and at best only reported if the rosters were rotating or not. Therefore, they conclude that their findings call for research where specific and more roster characteristics are assessed.

(14)

14

person. In conclusion, these articles suggest that psychical, mental, and social health issues could explain the relationship between shift work and absenteeism, but the effects may be small and different per person. Therefore, in line with Marmot et al. (1995), sickness absenteeism seems to be an appropriate combined measure for physical, mental, and social health when more specific measures are unavailable. Consequently, sickness absence will be used as the measure for health in the thesis.

Lastly, sickness absenteeism is also interesting even if it fails to measure health-related issues at all for whatever reason. Namely, sickness absenteeism is associated with high financial costs for both employer and society (Merkus et al., 2012). The employer is often obligated to pay salary costs for absentees as well as salary for the replacement. Besides the costs associated with productivity lost and reduced quality should be expected. From a society's point of view, long-term sick leave is associated with a lower probability to return to work, leading to financial deprivation and social isolation.

2.2.4 Pathways between rosters and health

To test what specific elements of a roster are harmful in practice, specific roster characteristics, which are theoretical harmful must be identified first. These roster characteristics should link with higher risks of adverse health outcomes and thus a higher probability of sickness absence. To accomplish this, the pathways explaining why rosters are detrimental to health will be discussed first. These pathways will not be tested empirically, but it will be used to identify which and argue why specific roster characteristics may result in health issues and thus absenteeism. A three-step model is constructed based on existing literature describing the process between shift work and health in a clear and basic way. The underlying idea of the model is based on Puttonen, Härmä and Hublin (2010), who argue that shift work disrupts a rhythm leading to different kinds of stress. These stressors then result in different health issues. Examples of the relationships and symptoms for each step will be given, these are however not comprehensive lists.

Shift work and disturbed rhythms

(15)

15

which regulates the sleep/wake cycle and biological processes (Costa, 2003; Puttonen, Härmä and Hublin, 2010; James et al., 2017; Copertaro and Bracci, 2019). Being active when the circadian processes want you to be resting, causes a disturbance that hinders biological processes and the length and quality of sleep. The social rhythm refers to a stable rhythm identified in western industrialized societies in which social and family activities mainly take place during evenings and weekends (Arlinghaus et al., 2019). In general, this will be disturbed when people are working shifts (Costa, 2003). However, this is not or to a lesser extent true for all people. For example, people without partners or children experience a different social rhythm than most of the population.

Disturbed rhythms and stress

The disturbance of the rhythms causes one or more kinds of stress. The first kind of stress is psychosocial. This is caused by both the disturbance of the circadian and social rhythm in several ways. For instance, the disturbance of the circadian rhythm may increase the perception of insufficient recovery from work, which can lead to thoughts of stress (Puttonen, Härmä and Hublin, 2010). Besides, the disturbance of the social rhythms can cause limited time for social activities resulting in work-life conflicts, which could also result in increased stress levels (Arlinghaus et al., 2019).

The second kind of stress is behavioral stress. This is caused by the disturbance of the circadian rhythm. An example of a phenomenon belonging to this category is less and disrupted sleep (Åkerstedt, 2003a; Folkard, Lombardi and Tucker, 2005; Van Dongen and Dinges, 2005). The displacement of sleep may also influence the meal timing and in turn changes the hormonal balance leading to less nutritious and more food intake (Lowden et al., 2010; Cain et al., 2015). Therefore it is not surprising that relationships between night shift work and higher body mass index are found in previous research (Morikawa et al., 2007; Zhao, Bogossian and Turner, 2012).

(16)

16

2013), desynchronization of the HPA-axis (Kalsbeek et al., 2012), and atherosclerosis (Haupt

et al., 2008).

Additionally, stressors may also interact with each other. Psychosocial stress may influence shift workers to sleep less or develop unhealthy habits such as smoking. Besides, psychosocial stress factors are also associated with increased physiological stress. For example, a higher risk of atherosclerosis (Eller et al., 2009). Excessive sleepiness caused by prolonged shortened sleep, in turn, might contribute to adverse physiological changes (Puttonen, Härmä and Hublin, 2010). At the same time, this excessive sleepiness could increase the perception of insufficient recovery leading to increase psychosocial stress.

Stress and health

The last steps in our model explain how different variants of stress could cause increased risks of adverse health effects. Relating to psychosocial stress work-life conflicts badly influence social health, since it is harder to maintain social relationships when no valuable social time is available (Arlinghaus et al., 2019). Besides, prolonged exposure to stress could lead to depression and thus increased risks of mental health issues (Angerer et al., 2017). In the category of behavioral stress, bad eating habits and insufficient sleep are associated with metabolic syndrome (Spiegel, Leproult and Van Cauter, 1999; Ha and Park, 2005; Esquirol et

al., 2009). Over time this could lead to type 2 diabetes (Knutsson and Kempe, 2014), which is

also associated with a higher risk of cardiovascular diseases (CVD) and thus psychical health.

(17)

17

A general model describing how shift work influences health is visually displayed in figure 2.2. Because of the “simplicity” of the model, it is impossible to say something about the size of the effects and the causality. Although this model largely neglects important mediating and moderating factors such as demographics, job type, and social-economic status, it is suitable for this thesis. The model mainly shows that harmful roster features should disturb at least the social or circadian rhythm, resulting in adverse health outcomes and thus absenteeism. This logic allows us to identify harmful roster characteristics in the next part.

Figure 2.2 Pathways leading from shift work to health issues.

2.3 Roster Characteristics Related to Health

Now a theoretical explanation for the relationship between shift work and health is provided. Therefore, it is possible to reason why certain roster characteristics discussed in the first section of this chapter could be detrimental to health.

(18)

18

Looking at shift characteristics independently from the kind of shift, shift length, and the time between shifts are important characteristics (Tucker and Folkard, 2012). Long shifts could cause insufficient rest time and limit the time to engage in social activities (Philip Tucker and Folkard, 2012). Besides long days from home mean less time at home and thus higher chances of quick changeovers (<11 hours between shifts). These quick changeovers are linked to adverse effects relating to the disruption of the circadian rhythm (Åkerstedt and Kecklund, 2017). However, longer shifts also result in compressed working weeks, since the total number of days worked is smaller if more hours are worked on each day (Tucker and Folkard, 2012). Compressed working weeks increase the successive days off, meaning more time available for social activities for instance through longer weekends or more free nights. Therefore, it can be concluded that longer shifts and thus less time between shifts not necessarily have the same adverse effect on one’s social rhythm as it has on one’s circadian rhythm.

Another important factor affecting the time between shifts is the rotation speed and direction of different shifts (Tucker and Folkard, 2012). The first question is whether shift should rotate at all or people should have fixed shifts. Fixed day shifts are unsurprisingly most favorable. However, the beneficial effects of fixed night shifts depend on the fact if people can adjust their circadian rhythm completely. Research on this subject suggests that people are not able to adjust their rhythm to receive any benefits from it (Tucker and Folkard, 2012). Besides, people tend to change back to a "day routine" during their days off and thus not switch to a complete night rhythm.

(19)

19

two reasons. First, it increases the time between shifts and thus limits quick changeover. Second, people tend to find it easier to delay sleep than to forward sleep. The fact that the circadian rhythm is slightly longer than 24 hours explains this preference (Tucker and Folkard, 2012).

The timing of changeovers could also disrupt both the social and circadian rhythm (Åkerstedt, 2003a). Early morning changeovers make it impossible to execute family life in the morning. However, starting earlier also means being home earlier enabling people to engage in social life. The same paradox is notable for the circadian rhythm. Starting earlier disrupts the rhythm of the morning shift, but starting later disrupts the rhythm of the prior night shift in a continuous system. Starting the morning shift at 7:00 is assumed to balance the positive and negative effects for both shifts (Åkerstedt, 2003a).

Some shift systems must compromise between the total weekly workhours and other roster features. For instance, a better pattern of rotations. Although more hours in one week mean more resting hours in other weeks, extreme variations of this distribution are unadvisable. Disruption of the circadian rhythm is prolonged opposed to an equal weekly distribution of the hours. Besides, there are various situations possible where this could influence one's social rhythm. For instance, working in the weekends to compensate for the free days during the week in the other week.

2.4 Conceptual Model

(20)

20

characteristics are known as possible confounders between shift work and sickness absenteeism (Van Drongelen et al., 2017). Therefore, they are added as control variables in this conceptual model. This model is visualized in figure 2.3. The roster characteristics mentioned in figure 2.3 are only theoritical linked to sickness absence. Therefore, the goal of the rest of this thesis is to first explore the different elements of this model en thereafter link these roster characterisitics to sickenss absence and health using emperical data.

(21)

21

3

Methods

3.1 Research Design

For answering the research question, quantitative data analysis was chosen as the appropriate research design. More specifically, a data mining approach using historical roster and absence data was chosen. Data mining, which is also known as knowledge discovery in databases (KDD), is a process of extracting information and knowledge hidden in data by approaches such as modeling and rule generation using computer technologies (Shao, Rowe and York, 2008). Extracting new information and knowledge matches the goal of this thesis. Since a large amount of roster and absence data is needed to find patterns, a company is required that can provided this data. The engineering and maintenance division of KLM airlines is willing to provide the required data.

Fayyad, Piatetsky-Shapiro and Smyth (1996) provide a process for KDD using nine steps, which is displayed in figure 3.1. The first step is already executed by the development of the theoretical framework. Step 2 to 4 discuss the used data and data preparation. Step 5 to 7 describe the core of data mining which is the identification of patterns in the data. The last two steps of the KDD process are covered by reporting and discussing the findings. Since KDD is an iterative process, in each iteration most steps are repeated. Therefore, for clarity and understandability reasons it was chosen to report on the methodological choices and action per main iteration chronologically. First, the development of the main dataset and exploratory analysis is elaborated on. Thereafter, two different analyses of specific datasets are discussed. However, before that the company profile of the E&M division of KLM airlines will be elaborated on first.

(22)

22

3.2 Company Profile

Since 2004, KLM Group is part of the Air France-KLM Group. KLM and KLM Cityhopper are the main part of the KLM-group, whereas Transavia and Martinair are wholly owned subsidiary companies (KLM, 2017). In 2019 KLM-Group had an operating result of 853 million euros, achieved a turnover of 11.1 billion euros, served 44 million customers, and flew to a total of 171 destinations (KLM, 2020). All of this is accomplished with a fleet of approximately 240 airplanes. To keep the fleet in a safe and operating state maintenance is needed, which is among other things organized through an extensive personnel-scheduling process.

The KLM E&M division is responsible for the fleet related to KLM airlines, which consists of 123 airplanes for the long and medium-haul (KLM, 2020). E&M is split into three maintenance units: Airframe, Component Services, and Engine Services. The personnel scheduling for the complete maintenance division is done by the Labour Time Management department. Currently, the total number of maintenance staff is approximately 3800 of which around 60% engage in some kind of shift work. The average sickness absenteeism used in calculations during the scheduling process is 6%, but higher percentages are known for specific departments. This makes the KLM E&M division a suitable object for this research.

3.3 Exploratory Analysis

3.3.1 Creating a target dataset

The planning software of KLM was used to gather the data for this research. Only this source of secondary data was used. This data contains the roster, roster mutations, and absence data of each employee that works based on a roster in the E&M division between 2012 and 2019. Since companies are obligated by law to collect this data and payments are based on this data, it is assumed as reliable as possible.

(23)

23

leave days, for each employee from 2012 to 2019. The fourth dataset contains personal and work-related characteristics for each employee active from 2012 to 2019.

3.3.2 Data cleaning and pre-processing

Because the data was collected from a professional planning system, the data is relatively clean. However, the following is done to create the main datasets. The first 10 rows of each dataset can be found in Appendix A.

Roster and absence data

Rows containing missing data were deleted first. This data is missing since almost empty rows are created in the system for the complete year new employees are linked to a roster instead of only for the period they are employed at KLM. Shifts taking place on two days, like night shifts, have multiple rows for each part of the shift that took place during a different day. Likewise, if an employee worked overtime or was partly absent during a shift, a new row is constructed with a separate starting time and duration. Therefore, rows were combined to ensure that every employee has a single row of complete data in every dataset for each day he/she was employed at KLM during that period. Shifts taking place at two days were linked to the date where the majority of the shift took place.

Employee and work characteristics

For every employee, it is known when they are born, when they started working at KLM, when they started working at the E&M division and their sex. Besides, for every employee it was determined which years they worked, and if they worked part-time that year. All other personal related data was deleted and each employee received a unique random number. Lastly, employees who left KLM before 2012 are delete, since no roster or absence data is known for these employees.

3.3.3 Data reduction and projection

Since sickness absence and general roster characteristics are the main concepts of interest in this part of the analysis, the data were transformed into measures for these concepts.

(24)

24 General roster characteristic

Three broad categories of rosters can be distinguished in the roster data as designed by the planners. Since the rosters between these groups are relatively different and inside these groups are relative similar it is interesting to separate groups for the analyses. The first category is “day workers”. Day workers only work morning or day shifts and no other shifts during the week. The second category is "day/evening workers". Logically they only work (extra-long) evening and morning/day shifts, which could be during both the weeks and weekends. The last category is "shift workers". They work a combination of at least three different shifts including night shifts during both the weeks and weekends. These shifts could be super early morning, early morning, day, afternoon, evening, extra-long evening, and night shifts.

Sickness absenteeism

(25)

25

3.3.4 Data mining methods

To summarize the main datasets, statistics such as means, standard deviations, percentiles, and correlations were determined. Besides, results were visualized using bar charts, histograms, and percentile graphs.

3.4 Time Series Analysis

To the knowledge of the writer, no study exists that uses a time series representation of a roster to study the relationship between rosters and sickness absence. Therefore, classification models were constructed to investigate the predictive capabilities of time series of roster data.

3.4.1 Creating target datasets

Based on the early discussion about the length of sick spells and the exploratory analysis it is chosen to construct datasets with only short sick spells (<= 21 days) as a measure of sickness absence for this analysis. Causes for absence that could be affected by roster characteristics are assumed covered best by using this threshold. Moreover, Ropponen et al., (2019) showed that this kind of sick spells could be partially explained using a relatively short analyzed roster period of 1 to 12 months. Lastly, considerably more cases of short sick spells exist in our database compared to long sick spells, which is beneficial for constructing classification models.

For each case of short-term sickness, a control case is needed, which is a time-series that not results in sickness absence. This ensures an equal amount of cases for both outcomes in the model. To select those control cases, the same case-crossover design is used as in Ropponen

et al., (2019) that is first introduced by (Maclure, 1991). For instance, if the 90 days before the

(26)

26

occurred during both windows and the roster data was taken from the actually worked rosters.

A dataset with windows of 90 days with sick spells of smaller or equal to 21 days was used as the benchmark model. Thereafter, windows of 30 days and 150 days were used to examine the effect of the length of the window. Moreover, the benchmark dataset was split based on the roster worked by the employee and age (< or > than 50 years) for both sick spell lengths based on the results of the exploratory analysis. Lastly, all models were also constructed for the length of a sick spell of less or equal to 7 days to examine the effects of the sick spell length. An overview of different datasets used with their sizes are given in table 3.1.

Table 3.1 Total number of cases per datasets used in the time series analysis.

Sick spell

length Employees included in the dataset 90 days Window length 30 days 150 days

<=

21

days

All sick spells 24632 (Benchmark) 57998 12106 Day Rosters 9492

Day Evening Rosters 8174 Shift Rosters 6966 Older than 50 years 14098 Younger than 50 years 10534

<=

7

days

All sick spells 17336 40038 8720

Day Rosters 7576 Day Evening Rosters 5042 Shift Rosters 4718 Older than 50 years 9670 Younger than 50 years 7666

The datasets are based on the earlier constructed datasets that were already cleaned and pre-processed. Therefore, this phase was passed for this analysis.

3.4.3 Data reduction and projection

Since this analysis focusses on rosters represented as time series, each roster needed to be converted to a meaningful time series representation. One of the most accurate presentations

(27)

27

would be to indicate for a small interval if an employee has worked or not. However, this results in relatively long time series, limiting the possibility to include many cases due to the limited amount of computing power and efficient algorithms. Therefore, rosters were converted to time series by indicating for each day how many hours are worked. This approach ignores the time of the day a shift is worked, i.e. does not distinguish between shift types. However, it still represents information about the length of the shifts and gives an indication of the time between shifts, the distribution of work hours, and the speed of rotation (number of successive shifts). Moreover, this information is very accurately represented in the time dimension, compared to models that require information to be summarized over the time dimension.

3.4.4 Data mining methods K nearest neighbor

K nearest Neighbor models were used to investigate the predictive ability of time series. The K nearest neighbor algorithm is chosen because this model is proven very effective for time series classification problems in many domains when using dynamic time warping as a distance measure (Petitjean et al., 2014). Dynamic time warping is further discussed below.

Since the working of algorithms is not the focus of this study and many explanations are widely available, the model will not be discussed in depth. However, the following approach was used to construct and test the models. An optimal number of neighbors K was searched for each model using a 5-fold cross-validation approach on 80% of the dataset. Thereafter, the model was trained using the same 80% of the dataset with the optimal found K. Lastly, the model is tested using the remaining 20% of the dataset and the results were compared to the cross-validated model to control for overfitting. This was done using the Python library SciKit-Learn (Pedregosa et al., 2011).

Dynamic time warping

(28)

28

Euclidean distance between the time series may be very large (Keogh and Pazzani, 2000). DTW overcomes this problem by minimizing the total distance between time series where every point can be linked to any point of the other time series. The Python library dtaidistance was used to perform the DTW for the KNN model (Craenendonck and Ma, 2019). This library was chosen since it has a fast C implementation of the DTW algorithm.

3.5 Conditional Logistic Regression Model

Compared to the time series analysis, which uses relative raw data, a model using constructed measures could result in another outcome than only the predictive capability. A variant of a regression model could use these measures, which represent a specific roster characteristic, to define the relationship between this roster characteristic and sickness absence. Therefore, conditional logistic regression models were constructed using measures for each roster characteristics.

3.5.1 Creating target datasets

The same case-crossover approach that is discussed before is used for this analysis. However, as is required in a matched case-control design like this one, the case and control windows of a sickness spell are explicitly matched to each other. In this way, every sickness absence case has exactly one control case per dataset. Thereafter, it is possible to compare each case window with each control window to determine if one or more roster characteristics are more occurring during the case window than during the control window. It could then be hypothesized that certain roster characteristics are likely to cause sickness absence.

(29)

29

Table 3.2 Total number of cases per datasets used in the conditional logistic regression.

Sick spell

length Employees included in the dataset 28 Window length 84 140

<=

21

days

Day Workers 22382 9552 4934

Day Evening Workers 24772 9832 4760

Shift Workers 12544 7164 3732

<=

7

days

Day Workers 8687 7600 3948

Day Evening Workers 15492 6238 3182

Shift Workers 8380 4820 2592

These datasets are based on the earlier constructed datasets that were already cleaned and pre-processed. Therefore, this phase was passed for this analysis.

3.5.3 Data reduction and projection Initial measures

Based on the theoretical framework measures were developed for all roster characteristics. Härmä et al. (2015) also recommends a list of tested measures for studies using objective roster data. The measures for shift length, shift types, the time between shifts, and part of the measures of rotation speed and distribution of work hours discussed here, were also proposed by Härmä et al. (2015). Although Härmä et al. (2015) also acknowledge the importance of shift rotations, they are not included as explicit measurements. Since the importance of these measures in literature is often stressed, measures were included for those variables.

(30)

30

Collinearity

When rosters are made, they must comply with numerous rules and regulations. These regulations impose constraints on the rosters and therefor certain patterns and combinations of shifts are more likely than others, which is also proven by (Ferguson, 2019). For instance, a roster with longer series of shifts is likely to have less rotations than a roster with short series given that all else is equal. Consequently, high degrees of collinearity are expected for this set of measures. It is known that inadequate detection and handling of collinearity will lead to misinterpretation of the results. Based on Vatcheva and Lee (2016), the correlations and variance inflation factor (VIF) of the measures were assessed using cut-off points of correlation coefficient > 0,8 and VIF > 10. Based on these cut-off points, measures were removed or transformed. Thereafter, as suggested by Vatcheva and Lee (2016), initial

1_{TBSS is the time between shift starts}

Roster

characteristic Measures Further clarification Shift types

• Percentage of shift type of all shifts.

• Percentage of weekend days worked

Shift types: early morning, morning, day, afternoon, evening, extra-long evening and night

Shift length • Mean duration of shifts • Percentage of long shifts of total shifts

Long: > 10 hours

Time between shifts

• Mean time between shifts • Percentage of quick returns • Percentage of free days of

total days

Time between shift and a free day or separated by free day not included.

Quick return: <= 11h between shifts

Direction of rotations

• Frequency of forward rotations between and inside series

• Frequency of backward rotations between and inside series Forwards inside: TBSS1_{> 25h} Backwards inside: TBSS < 23h Forwards between: TBSS – (#days off * 24) > 25h Backward between: TBSS – (#days off * 24) < 23h Rotation speed

• Mean (sub)series length • Frequency of long

(sub)series • Frequency of short

(sub)series

Short series ≤ 5 shifts Long series > 5 shifts Short subseries ≤ 3 shifts

Long subseries > 3 shifts

Distribution of work hours

• Coefficient of variation of working hours per week • Percentage of long working

weeks of total weeks

(31)

31

conditional logistic regression models were built and the correlations between the independent variable coefficients and the effects of leaving or more variables out of the model were assessed. This was done using the datasets with sick spells of less or equal than 21 days and a window of 84 days. Table 3.4 displays an overview of all the initial and new measures and indicates which measure are excluded due to collinearity for the final analysis. At least one measure is included for each roster characteristic mentioned in the conceptual model. Appendix B contains the VIFs before and after the exclusion of measures. The mean values of the measures for each model can be found in appendix C.

3.5.4 Data mining methods

(32)

32

Table 3.4 Overview of the measure used in the final conditional logistic regression.

Roster

Characteristics Measures Included Reason

Shift type

% Early morning shifts Yes % Morning shifts Yes

% Day shifts No Since the first 7 shift types add up to 100% by definition, one must be excluded. Comparable with the use of dummy variables in regressions.

% Afternoon shifts Yes % Evening shifts Yes % Extra-long evening shifts Yes % Night shifts Yes % Weekend shifts Yes Shift length Mean duration Yes % Long Shifts Yes

Time between shifts

Mean time between shifts Yes % Quick Returns Yes

% Free Days No

High degree collinearity with rotations. Free days often occur between different shift types and rotations do this too. Therefore, they are related with could lead to misinterpretation of the results.

Direction of rotations

Forwards rotations inside

series No

All types of rotations are highly related to each other at KLM, since changes in one direction often result in a rotation in the other direction. Therefore, they are transformed into one measure.

Backward rotations inside

series No Forwards rotations between

series No Backward rotations between

series No Total number of rotations Yes

Rotation speed

Mean series length Yes

Mean subseries length No _{All rotation speed measures are highly correlated with} each other at KLM. For instance, more short subseries often results in longer series. Therefor chosen to use the mean series length as the only measure for this roster characteristic. Short subseries No Long subseries No Short series No Long series No Distribution of work hours

CV hours per week No

A high degree of collinearity with both rotation speed and direction. Long series in week one result in more short series in the following weeks. This is because of the average number of contract hours per week.

(33)

33

4

Results

The results of the three analyses are presented below in a comparable way to the methods, i.e. per main iteration. First, the results of the exploratory analysis are presented. By doing so, more information about the different elements of the conceptual model is provided. This information is then used in making decisions for and interpreting the results of the other analyses. Thereafter, the results of the time series are discussed, which give more insights in the predictive capabilities of the rosters. Lastly, the results of the conditional logistic regression are presented with the aim to determine which roster characteristics are more likely to influence sickness absence and health.

4.1 Exploratory Analysis

4.1.1 Employee and work characteristics

Table 4.1 contains characteristics of the working population of roster-based employees at the E&M division per year. In general, the population is relatively old and consists of mostly men. In 2017, a change in the population is visible. When comparing the histograms in figure 4.1 of age, years at KLM and years at E&M of 2012 and 2019, it is visible that older and more experienced employees have left the KLM/E&M division and new and younger employees have joined the E&M division.

Table 4.1 Employee characteristics of the KLM E&M division per year.

A second pattern is identified in the correlations in table 4.2. It shows a strong relationship between age and year working at KLM/E&M indicating that employees stay relatively long at KLM. Therefore, only age will be elaborated on further in the analyses instead of age, years at KLM and years at E&M, since roughly the same pattern can be expected.

(34)

34

Figure 4.1 Histograms of the distribution of employee characteristics.

P < 0.001 Age Years at _KLM Years at _E&M

Age 1,00 0,85 0,74

Years at KLM 0,85 1,00 0,80

(35)

35

4.1.2 Employee rosters

Figure 4.2 shows the number of employees per roster type for all the years. On average 60% works according to a shift roster or

day-evening roster and 40% works according to a day roster. Since each group is considerably large and the roster are very different for each group, it is logical to perform further analyses per roster type. Figure 4.3 visualizes the relationships

between employee and work

characteristics and employee roster types.

Looking more closely at the ratio of roster types of 2012 comparted to 2019 suggests a slightly more equally division of the roster types over the age groups in 2019. Moreover, figure 4.3 shows that women compared to men mainly work according to day rosters between 2012 and 2019. Part-time workers are also more often day workers between 2012 and 2019, although this difference is, compared to the division based on gender, considerably smaller.

Figure 4.2 Number of employees per roster type per year.

(36)

36

4.1.3 Sickness absence

Total sickness absence

The mean number of working days at which a form of sickness absence occurred over the examined years is 8,7% (STD = 17,1%). Full day sickness absence occurred on average 5,8% percentage (STD = 11,3%) and non-full day sickness absence occurred on average 2,8% (STD = 10,1%). When including non-working days, the mean sickness percentages are respectively 8,6% (STD = 17,0%), 5,7% (STD = 11,3%) and 2,8% (STD = 10,1%). Figure 4.4 shows time series of the sickness absence on working days. For full-day sickness absence, a strong seasonality is visible, whereas this seasonality is only limited visible for non-full days of sickness absences. Therefore, it can be assumed that full day sickness absence has a more general cause and it shall therefor be considered as the only kind of sickness absence of interest. The visualization in figure 4.5 of the mean

full day sickness absence per person per year in percentiles shows that the mean sickness absence per person is very unequally distributed. This suggests that it is possible that part of this unequally distribution is caused by the effects of rosters.

When the relationships between sickness absence and employee and work characteristics are examined, a clear relationship between age and mean sickness absence is found. Making age a variable of interest for further analyses. Figure 4.6 shows this relationship between age and sickness absence for 2019. Bar charts for the other years and tables on the relation of other employee and work characteristics and sickness absence can be found in appendix D.

Figure 4.4 Time series of sickness absence percentage

(37)

37 Sick spells

The time series graphs of the number of sickness spells in figure 4.7 show the same seasonality as the time series of the mean sickness absence, especially for short sick spells. The start date of each spell is used as the date indicator. When long sick spells are defined as longer than 14 days of successive full day sickness, the time series of long sick spells still shows signs of seasonality. However, if the threshold is increased from 14 to 21 days, almost no signs of seasonality are shown for long sick spells. Therefore, using 21 days as the initial threshold between short and long sick spells seems to be suitable to distinguish the more regular short-term sickness absence from long-short-term sickness absence that reflects chronic morbidity.

Figure 4.7 Time series of the number of sick spells for different thresholds

Table 4.3 is constructed using the 21 days threshold to give more insights into the different types of sick spells. Most interestingly, it shows that short sick spells contribute to almost half of the total days of sickness absence. The other half of the days consist of long sick spells, which is interesting since the number of long sick spells is only one-tenth of the number of short sick spells.

(38)

38

Table 4.3 Statistics per kind of sick spell using the 21 days threshold to distinguish short and long.

Sickness absence per roster type

Since three different groups of rosters can be distinguished, it is interesting to explore their relationship with sickness absence. As can be seen in table 4.4, day workers have on average the lowest sickness absence percentage during each year. Which is logical since this type of roster is considered as least harmful. Following this same logic, day evening workers should have a lower sickness absence percentage than shift workers. However, this is clearly not the case.

Therefore, this analysis suggests two main points, which are further explored in the other analyses. First, rosters could explain sickness absence, since considerable differences are found in the percentage of sickness absence for each type of roster. Further analyses should provide insights in the predictive power and the impact of other important factor, such as the length of the included roster. Second, although this analysis in general supports current literature that shift work is more harmful than day work, it also suggests that some roster characteristics are more important than others are. Otherwise, shift workers should have a higher percentage of sickness absence than day evening workers have. Further analyses should therefore be done to discover which roster characteristics are the most important ones.

Table 4.4 Sickness absence percentage per roster type per year for full day sickness absence

Type of Sick Spell Mean Length STD Count of

(39)

39

4.2 Time Series Analysis

Table 4.5 displays the accuracy of all the constructed KNN models. The benchmark model has an accuracy of 60%, which means that 60 percent of all outcomes were predicted correctly. Using randomly guessing in a model with two outcome variables would result in a mean accuracy of 50%. Therefore, the benchmark model performs on average 20% better than randomly guessing. The models using sick spells of day workers performed

overall the best and have an accuracy of 64%, whereas the shift worker models performed the worse with 53% accuracy. Overall, the accuracy seems to differ slightly between the sick spell lengths and the window lengths. Therefore, these factor needs to be considered in the next analysis, in which relationships between rosters characteristics and sickness absence are further explored. Since the age of the workers have no effect on the accuracy of the model, this factor will be excluded in the next analysis.

4.3 Conditional Logistic Regression

For each roster type, different roster characteristic measures were associated with short-term sickness absence using a conditional logistic regression with a significance level of 5%. This resulted in odds ratios, which describe the increased risk of short-term sickness absence associated with a unit increase of a measure compared the same roster without this one-unit increase. The found associations are described per roster type.

In the datasets containing only the day workers, most of the significant effects were found in the datasets using a 140 days window for both sick spells lengths of ≤7 and ≤21 days (see table 4.6). Increased odds ratios were found for the mean duration of shifts (OR=5,794 and OR=4,515, respectively), mean time between shifts (OR=3,226 and OR=2,873), and mean series length (OR=1,690 and OR=1,547). Here, it must be noted that the results for the mean

Employees included

in the dataset _{90 days 30 days 150 days}Window

Sic k sp ells < = 21 d ay

s All sick spells 60% 59% 61% Day Workers 64% Day Evening Workers 59% Shift Workers 53% Older than 50 years 60% Younger than 50 years 60%

Si ck sp el ls <= 7 days

(40)

40

time between shifts are very counterintuitive. The mean duration of shifts also showed an increased OR for the datasets using an 84 days window for both sick spell lengths (OR= 1,627 and OR=1,633). Considerably smaller associations were found for the total number of rotations in two datasets using a sick spell length of ≤21 and for the mean time between shifts in the dataset using a sick spell length of ≤7 and a window of 28 days. More interestingly, a decreased OR was found for working weeks >48H. However, the mean value of this measure was very low indicating that working weeks like this only very rarely happen. Lastly, all measures that showed at least one significant association using any window length were found using the 140 days window for day workers.

The results of day evening workers datasets show a similar pattern for both sick spells lengths of ≤7 and 21 days (see table 4.7). Slightly increased ORs were found for morning (OR=1,009 and OR=1,008, respectively), evening (OR=1,015 and OR=1,013) and extra-long evening shifts (OR=1,021 and OR=1,014). Since day shifts are omitted from the model due to perfect collinearity, these ORs must be interpreted as compared to regular day shifts. Therefore, for instance an increased odd of 0,9% of short-term sickness absence of ≤7 days is associated with every percent increase of morning shifts compared to day shifts in a 28-day window. Moreover, for all but one dataset increased ORs between 1,004 and 1,035 are associated with the percentage of weekend shifts. Stronger associations were found for the mean duration of the shifts (OR=1,211-2,202) in all but one dataset, percentage of long shifts (OR=1,015-1,053) in three datasets, the total number of rotations (OR=1,026-1,092) in three datasets, and mean series length (OR=1,119 and OR=1,137) in two datasets. Again, decreased ORs were associated with working weeks of >48H and >42,5H (OR=0,842-0,921). Lastly, all measures that showed at least one significant association using any window length were found using a 28 days window for the day evening workers.

(41)

41

(42)

42

Table 4.6 Odds ratios with 95% confidence interval per measure for each sick spell and window length combination for day worker. Significant results (p < 0,05) in bold.

Sick spell length <=21 Sick spell length <= 7

Window = 28 Window = 84 Window=140 Window = 28 Window = 84 Window =140

95% CI 95% CI 95% CI 95% CI 95% CI 95% CI

Measures OR Lower Upper OR Lower Upper OR Lower Upper OR Lower Upper OR Lower Upper OR Lower Upper % Early morning shifts 0,949 0,851 1,059 0,000 0,000 0,000 0,532 0,062 4,585 0,955 0,856 1,066 0,000 0,000 0,000 0,546 0,063 4,718 % Morning shifts 1,009 0,994 1,025 1,051 0,961 1,150 1,001 0,957 1,046 1,017 0,998 1,037 1,054 0,958 1,161 0,990 0,938 1,045 % Afternoon shifts 0,982 0,952 1,013 1,010 0,912 1,119 1,097 0,891 1,351 0,968 0,930 1,008 0,979 0,878 1,093 1,041 0,815 1,331 % Evening shifts 0,156 0,002 10,918 0,002 0 4,12E+12 0 0 1,29E+16 0,12 0 1140,549 0,002 0 1,3E+14 0 0 3,48E+19 % Extra-long evening shifts 0,211 0 2,01E+08 0,001 0 5,1E+180 0 0 1,16E+80 0,185 0 1,44E+16 0,001 0 1,4E+232 0 0 7,02E+87 % Night shifts 1,001 0,976 1,027 0,817 0,614 1,085 0,792 0,505 1,242 0,997 0,967 1,029 0,820 0,610 1,102 0,708 0,325 1,539 % Weekend shifts 1,007 0,992 1,022 1,047 0,990 1,107 1,019 0,929 1,118 1,018 0,999 1,037 1,059 0,990 1,133 1,043 0,926 1,175 Mean duration of shifts 1,013 0,935 1,097 1,633 1,163 2,293 4,515 1,863 10,944 1,025 0,925 1,135 1,627 1,102 2,401 5,794 2,141 15,679

(43)

43

Table 4.7 Odds ratios with 95% confidence interval per measure for each sick spell and window length combination for day evening worker. Significant results (p < 0,05) in

bold.

95% CI 95% CI 95% CI 95% CI 95% CI 95% CI

(44)

44

Table 4.8 Odds ratios with 95% confidence interval per measure for each sick spell and window length combination for shift worker. Significant results (p < 0,05) in bold.

95% CI 95% CI 95% CI 95% CI 95% CI 95% CI

(45)

45

5

Discussion

To gain a better understanding of the conceptual model, first an exploratory analysis was done. Since most of those results were used for making decisions for the more in-depth analysis, those results will not be discussed further. Analyses that are more detailed were done to explain the main hypothesis from the exploratory analysis and current literature that different kinds of shift work have different effects on sickness absence and health. This should answer the research question: What roster characteristics are detrimental to health? Therefore, a time series analyses and conditional logistic regression were performed using historical absence and roster data. The results, implications, and limitations will be discussed separately for each analysis.

5.1 Time series analysis

5.1.1 Discussion of the results

First, KNN clustering models were build using a time series of the worked hours per day. For the model including all types of roster, an accuracy of 60% was found. Suggesting that at least 60% of the sickness absence cases can be predicted using roster information, which is an increase of 20% compared to random guessing. Moreover, it was found the accuracy depended heavily on the type of roster and an accuracy of even 64% was achieved for day roster. Therefore, the results show that roster are very useful to predict short-term sickness absence. Since no other similar research is done in this field, only a benchmark with the already mentioned 50% resulting from random guessing can be done. Still, two reasons are identified why the accuracy of this model could be improved.

First, the used time series may not have been the best representation of the information, which could explain sickness absence. The time series contained no information about the type of shifts or exact rest time between shifts, which both have been associated as problematic roster characteristics using objective and self-reported data (Åkerstedt and Kecklund, 2017; Van Drongelen et al., 2017).

(46)

46

Time series of shift rosters have relatively irregular shapes compared to time series of day rosters. Therefore, a change in a shift roster, and thus a disturbance of the social or circadian rhythm will not be as well notable as a change in the day roster. In other words, irregularities are easier detected in day rosters. This also explains the highest accuracy for day rosters datasets. Predicting sickness absence for a set of rosters that are very similar with more detailed roster information, such as the used conditional logistic regression, is therefore likely to achieve a higher accuracy.

Lastly, the found accuracies might be closer to the maximum achievable accuracy than is expected at the first sight. Even when the used time series was a perfect representation for one or more roster characteristics, this could only result in an increased risk for short-term sickness absence. It is still possible that an employee would stay healthy although the worked roster was considered unhealthy. Besides the opposite could also hold, which means that someone would be absent even though the worked roster is considered healthy (Bourbonnais

et al., 1992). Therefore, a high accuracy for this type of analysis is very unlikely in explaining

sickness absence by rosters.

5.1.2 Limitations and recommendations

(47)

47

5.2 Conditional logistic regression

5.2.1 Discussion of the results

Compared to the time series analysis, the conditional logistic regression resulted in more information about harmful roster characteristics and thus provided direct answers to the research question. However, only a limited amount of the studies on detailed roster characteristics and sickness absence are done. Moreover, most of them use samples that consists of mostly females, e.g. nurses (Bourbonnais et al., 1992; Dall’Ora et al., 2019; Ropponen et al., 2019, 2020). Consequently, the opportunities to compare the result directly with other studies is limited. Besides, the possibilities of random effects even though a significance level of 5% is used is present, since the data is not gathered in a controlled environment. Because of this, most attention is paid to the patterns of results rather than single measures, since they are less likely to be influenced by random effects. In conclusion, the discussion below will mainly focus on theoretical explanations suggested by the patterns of the results of the conditional logistic regression, which are useful to build hypotheses and for further research to investigate.

Logically the shift type measures showed no association for the day workers. Nevertheless, for the day evening and shift workers, the shift types included in the shift systems showed in general positive association compared to day shifts with sickness absence for relative short window lengths. This in line with Niedhammer, Chastang and David (2008) who found an association between different shift types and sickness absenteeism for men in his research. Moreover, for the day evening and shift workers, positive associations were found for the percentage of weekend shifts. This agrees with the finding of Åkerstedt and Kecklund (2017) who found a negative association between social life and working in weekends and Jacobsen and Fjeldbraaten (2018) who states that work-life conflict might be a reason for sickness absence. Overall, these findings support the line of reasoning that shift types other than day shifts during the week disturb the social and/or circadian rhythm, which could lead to adverse health effects resulting in sickness absence.

Scheduling a Healthier Workforce