Updating the "Risk Index": A systematic review and meta-analysis of occupational injuries and work schedule characteristics.

Fatigue is a major risk factor for occupational 'accidents' and injuries, and involves dimensions of physical, mental, and muscular fatigue. These dimensions are largely influenced by temporal aspects of work schedules. The "Risk Index" combines four fatigue-related components of work schedules to estimate occupational 'accident' and injury risk based on empirical trends: shift type (morning, afternoon/evening, night), length and consecutive number, and on-shift rest breaks. Since its first introduction in 2004, several additional studies have been published that allow the opportunity to improve the internal and external validity of the "Risk Index". Thus, we updated the model's estimates by systematically reviewing the literature and synthesizing study results using meta-analysis. Cochrane Collaboration directives and MOOSE guidelines were followed. We conducted systematic literature searches on each model component in Medline. An inverse variance approach to meta-analysis was used to synthesize study effect sizes and estimate between-studies variance ('heterogeneity'). Meta-regression models were conducted to explain the heterogeneity using several effect modifiers, including the sample age and sex ratio. Among 3,183 initially identified abstracts, after screening by two independent raters (95-98% agreement), 29 high-quality studies were included in the meta-analysis. The following trends were observed: Shift type. Compared to morning shifts, injury risk significantly increased on night shifts (RR = 1.36 [95%CI = 1.15-1.60], n = 14 studies), while risk was slightly elevated on afternoon/evening shifts, although non-significantly (RR = 1.12 [0.76-1.64], n = 9 studies). Meta-regressions revealed worker's age as a significant effect modifier: adolescent workers (≤ 20 y) showed a decreased risk on the afternoon/evening shift compared to both morning shifts and adult workers (p < 0.05). Number of consecutive shifts. Compared to the first shift in a block of consecutive shifts, risk increased exponentially for morning shifts (e.g., 4th: RR = 1.09 [0.90-1.32]; n = 6 studies) and night shifts (e.g., 4th: RR = 1.36 [1.14-1.62]; n = 8 studies), while risk on afternoon/evening shifts appeared unsystematic. Shift length. Injury risk rose substantially beyond the 9th hour on duty, a trend that was mirrored when looking at shift lengths (e.g., >12 h: RR = 1.34 [1.04-1.51], n = 3 studies). Rest breaks. Risk decreased for any rest break duration (e.g., 31-60 min: RR = 0.35 [0.29-0.43], n = 2 studies). With regards to time between breaks, risk increased with every additional half hour spent on the work task compared to the first 30 min (e.g., 90-119 min: RR = 1.62 [1.00-2.62], n = 3 studies). Rest break duration and interval seem to interact such that with increasing duration, the time between breaks becomes irrelevant. The updated "Risk Index". All four components were combined to form the updated model and the relative risk values estimated for a variety of work schedules. The resulting "Risk Map" shows regions of highest risk when rest breaks are not taken frequently enough (i.e. <4 h) or are too short (i.e. <30 min), when shift length exceeds 11 h, and when work takes place during the night (particularly for >3 consecutive night shifts). The "Risk Index" is proposed as an empirical model to predict occupational 'accident' and injury risk based on the most recent data in the field, and can serve as a tool to evaluate hazards and maximize safety across different work schedules.


Introduction
The prevention of 'accidents' 1 and injuries in the workplace is a major concern of workers and their employers. An important step in their prevention is the identification of modifiable risk factors, and several fatigue-related factors have been identified in recent decades (Williamson et al. 2011). Fatigue can be defined as "a biological drive for recuperative rest" and includes the sub-factors of sleepiness, and of mental, physical and muscular fatigue (Williamson et al. 2011). For workers, fatigue can be affected by the physical and mental efforts of their work tasks and also by components of their schedules, such as the timing of shifts and the number of hours worked (Folkard and Åkerstedt 2004).
Combining several such fatigue-related aspects of work schedules, Folkard and Lombardi proposed a "Risk Index" model in 2004 to predict real-world occupational 'accident' and injury risk (Folkard and Lombardi 2004). The "Risk Index" is based upon empirical trends associated with the following work schedule characteristics: the type/ timing of a shift (e.g., morning, afternoon/evening, and night shift), the number of consecutive shifts (e.g., working four night shifts in a row from Monday night to Friday morning), the shift length (e.g., 8-h vs. 12-h shifts), and the timing of rest breaks at work (e.g., minutes since last break, or time on task).
For these work schedule components, consistent trends have been identified in previous research forming the original "Risk Index": the relative risk for occupational incidents (i) increased by 18% on the afternoon/evening shift and by 30% on the night shift, relative to that on the morning shift; (ii) was 6% higher on the second night, 17% higher on the third night, and 36% higher on the fourth night, relative to that on the first night shift; a similar yet weaker trend was found for morning shifts with 2%, 7%, and 17% increased risk, respectively; (iii) showed a 13% increase for 10 h shifts and 27% increase for 12 h shifts, relative to 8 h shifts; and (iv) increases approximately linearly with time spent on task, such that risk had doubled after 90-119 min, compared to the first 30 min Lombardi 2004, 2006).
Since the first introduction of the "Risk Index" in 2004, additional peer-reviewed published studies have been conducted, providing updated information on the relationship between 'accident' and injury risk and the above-described variables. Thus, the goal of this study was to update the "Risk Index" by synthesizing and integrating the past research progress in a systematic review of the peer-reviewed literature and re-quantifying the risk estimates using meta-analysis.

Materials and methods
In conducting and reporting this review, we followed Cochrane Collaboration directives and guidelines for meta-analysis of non-randomized, observational studies in epidemiology (MOOSE) (Stroup et al. 2000).

Electronic database
In collaboration with an expert on systematic reviews from the Harvard Francis A. Countway Library of Medicine, a highly sensitive search strategy was developed to identify all relevant studies capturing various types of injuries and 'accidents' as well as the four work schedule characteristics of interest, namely type, length and consecutive number of shifts, and on-shift rest breaks (the exact search strategy is provided in Supplement A). The search strategy combined two primary concepts: outcome variable ('injury/accident/safety') and exposure variables ('shift timing/length/number/rest breaks'), linked by the Boolean operator 'AND' and including a variety of key terms (controlled vocabulary and free text words) within each concept separated by the operator 'OR'. The occupational nature of outcomes was implied by the exposure search terms used for shift timing, length and number, and added explicitly for rest breaks. No filters were applied, hence the search was not limited to a particular publication language or study type (i.e. human studies). The last search was conducted in Medline (Vienna, Austria) on 4th April 2016 yielding 3,171 hits.

Searching other sources
We i) hand-searched all issues of Chronobiology International from 2015 to 2017 (Vol. 32-34), ii) examined bibliographies of systematic and non-systematic review articles identified through the electronic data base search (n = 41), and iii) screened reference lists of all articles eligible for full text screening (n = 257) to find additional relevant studies.

Outcome variable
The outcome of interest was a work-related unintentional injury (of any severity) or 'accident' to the worker. 'Accidents' included non-injurious incidents, such as falls and exposure to bloodborne pathogens or body fluids. For this review, we excluded motor vehicle crashes, commercial driving and commuting 'accidents', as well as studies reporting errors, near-misses/near-'accidents', work performance, heat or cold stress, chronic musculoskeletal disorders, repetitive strain injuries and cumulative trauma disorders, and injuries resulting from violence against the worker (i.e. assaults, homicides).

Work schedule characteristics
We were specifically interested in the timing and length of a work period, the number of successive shifts, and the interval between or length of on-shift rest breaks. Studies were eligible if they reported at least one of the four exposure variables.
Shift type. Timing of a work period classified as 'morning/day' shift, 'afternoon/evening' shift and 'night/graveyard' shift.
Shift number. The number of shifts worked in a row with periods off duty in between, e.g., working four morning shifts on four consecutive days. Accordingly, a double shift of, e.g., 24-h does not qualify as two consecutive 12-h shifts.
Shift length. The duration of a work period reported as time into shift, daily work hours or scheduled shift length (irrespective of when the 'accident' or injury happened during that shift).
Rest breaks. The total added duration of rest breaks or time interval between rest breaks taken while at work.

Type of studies and worker populations
Original articles from observational studies (casecontrol, cross-sectional, and retrospective and prospective cohort studies) were eligible. Case studies, case reports, laboratory studies or intervention studies were not included. Review articles were used for reference list screening to identify additional studies. We did not limit the eligibility of original studies to a specific type of industry or occupation, minimum sample size, or workers' age and gender. These variables were deemed potential effect modifiers and later extracted to account for heterogeneity between studies.

Screening process
Abstract screening For the abstract screening, three study investigators independently assessed abstracts and titles in pairs of two (DF-DAL and DF-JW). Disagreements were resolved between the involved raters after discussion in a consensus meeting.

Full text screening
In the first step, the selected full texts were preassessed by DF excluding studies that did not report the relevant exposure (i.e. weekly work hours) and/or outcome (i.e. repetitive strain injury). In the second step, three authors independently evaluated the remaining studies in pairs of two (DF-DAL and DF-JW) for inclusion in the systematic review. The third step involved the evaluation of all identified full text articles by the three authors in pairs of two (DF-DAL and DF-JW) for their inclusion in the meta-analysis. Eligible for meta-analysis were studies that (i) indicated a risk estimate (odds ratio, relative risk, incident rate ratio) or frequencies for the calculation of the risk estimate, (ii) reported exposure in adequate categories for synthesizing (e.g., shift length in hourly intervals vs. binary cut-off), and (iii) controlled to some extent for a-priori risk (i.e. risk for injury might not be constant across the morning, afternoon/evening and night shift if staff size varies across the 24-h day), e.g., studies that reported the number of injuries along with the number of workers or controlled for weekly work hours in the regression model.

Data extraction
Data for the meta-analysis and meta-regression (risk estimates with their 95% confidence intervals, raw frequencies for calculation of risk estimates and sampling variance, and potential effect modifiers) were extracted by DF and crosschecked by JW. Potential effect modifiers included: country and year of data collection, study design, adjusted vs. unadjusted risk estimate, sample age and gender ratio, sample size, industry, type of injury/ incident and report of injury/incident (selfreported, medically verified).

Data handling and statistical analyses
Different types of risk estimates were included, such as odds ratios (ORs), relative risks (RRs) and incident rate ratios (IRRs), and hazard ratios (HRs). HRs were not pooled with ORs, RRs or IRRs, but used in separate meta-analyses including only studies with HRs. By pooling ORs, RRs and IRRs, we did not introduce a relevant artificial bias to our results because occupational injuries and 'accidents' are a rare outcome, and when injury incidence is rare the ORs approximate the RRs. In case no risk estimate (OR, RR, IRR) was provided but the study was considered eligible, reported raw frequencies were used to calculate relative risks. Only logarithms of risk estimates were used in the meta-analysis. Sampling variance was calculated from 95% CIs, respectively, from sample sizes according to the following formula (variance of logRRs): 1 n exposed=injured À 1 n total=exposed þ 1 n unexposed=injured À 1 n total=unexposed All risk estimates are expressed relative to their work schedule component reference: morning shift (shift type), 1st shift (shift number), 8-h shifts (shift length), and 1st interval/no rest break (rest breaks), respectively.
We conducted an inverse variance approach to meta-analysis in R using the package "metafor" (Viechtbauer et al., 2010). As only observational studies were included, we did not assume a common, fixed effect across studies but computed a random-effects model allowing variation to stem from both, within and between studies, thus estimating an average effect. Heterogeneity was quantified using the I 2 statistic. Publication bias was assessed by Egger's regression coefficient and visual inspection of the funnel plot. The results on heterogeneity and publication bias are reported in Supplement B1. In order to derive a risk estimate for the overall shift length, as opposed to the hour at work when the incident happened, the hourly risk estimates were integrated by additional meta-analyses, e.g., the meta-analytical risk estimate for a 7-h shift length is based on 7 estimates (hour 1-7) x 7 studies = 49 risk estimates in total. Meta-regression analyses using potential effect modifiers were computed to explain heterogeneity (between-studies variance) for the component shift type, including sample mean age and gender ratio, country of data collection, year of publication, adjustment vs. non-adjustment of risk estimate, number of incidents in the study, study design, type of industry/occupation, injury/'accident' type and method to assess injury/'accident'. To predict relative risk in particular work schedules, the "Risk Index" components were combined multiplicatively based on each component's best fit. For instance, a polynomial curve was fitted to the estimates for shift number (1st, 2nd, 3rd, 4th shift in a row), and predictions for the 5th shift were extrapolated using the polynomial equation.

Selection process of included studies
Of 3,183 initially identified studies, 97 were selected for a systematic review, and 29 were included in the meta-analysis ( Figure 1). In summary, this process was as follows. Initially, two independent raters screened all 3,183 abstracts (n = 1,590 by DAL-DF and n = 1,590 by JW-DF) showing almost perfect inter-rater agreement (99% and 95%, respectively). All disagreements (n = 95) were resolved by consensus after discussion resulting in the inclusion of 257 studies. The reference lists of these 257 studies were screened for potentially relevant articles based on their titles, and n = 13 additional studies were identified resulting in n = 270 studies in total for full text retrieval.
The selected full texts of the 270 studies were pre-assessed by DF excluding n = 173 studies that clearly failed to report any exposure variable (e.g., weekly work hours) and/or outcome variable (e.g., chronic musculoskeletal disorders). Another criterion for exclusion was failure of the study to report on the relationship between injuries/'accidents' and work schedule characteristics. Two independent raters assessed the remaining 97 studies (n = 48 by DAL-DF and n = 49 by JW-DF) showing moderate to substantial inter-rater agreement (83% and 75%, respectively). All disagreements (n = 20) were resolved between the respective two raters in a consensus meeting resulting in the selection of n = 29 studies for the meta-analysis. The majority of reasons to exclude articles from the meta-analysis were that either no risk estimate was provided or could be calculated from counts data, data used in different studies were from the same cohort, and that exposure categories were defined as too broad for our purposes (e.g., shift length: 8 h vs. >8 h). In addition, n = 23 studies failed to account for an a priori risk, e.g., if the number of individuals at work is not constant over the 24-h day, and were excluded since injury risk cannot be compared across shifts in an unbiased manner. Accordingly, if staff sizes varied, but were known and reported, the respective study was included.

Characteristics of included studies
Of the 29 studies included in the meta-analysis (Table 1)   n = 4, 12 h, and > 12 h: n = 3), and 5 studies included data for rest breaks (interval between rest breaks: n = 3, rest break duration: n = 2). Six studies reported risk estimates for more than one work schedule component (Levin et al. 1985;Ogiński et al. 2000;Quaas and Tunsch 1972;Smith et al. 1994;Tucker et al. 2003;Violanti et al. 2012).
The 29 studies were published between 1962 and 2014, and comprised a total of 1,410,004 incidents. Mean age was 33.8 years (based on 15 studies that provided information) and 66% of subjects were male (based on 16 studies). The majority of studies were primarily conducted in Europe (n = 13) and the US (n = 12), with the remaining 4 studies being from Canada (Wong et al. 2011), China (Lombardi et al. 2014), Russia (Vinogradova et al. 1975), and New Zealand (Fransen et al. 2006). Studies designs varied, and included 3 cohort studies (2 retrospective and 1 prospective) (Ayas et al. 2006;Violanti et al. 2012;Wong et al. 2011), 4 case-crossover studies (Arlinghaus et al. 2012(Arlinghaus et al. , 2014Lombardi et al. 2003;Tucker et al. 2006: Study 1 and 2), and 22 cross-sectional studies. Workers were primarily employed in the industrial sector (n = 12), ranging from mining to engineering and paint manufacturing, and in the healthcare sector (n = 4; Ayas et al. 2006;Horwitz and McCall 2004;Macias et al. 1996;Neuberger et al. 1984). One study was among police officers (Violanti et al. 2012), and 12 studies comprised samples of diverse occupations and industries.
The majority of studies (n = 21) did not specify the exact nature of the work injury (e.g., "injurious accidents", Tucker et al. 2003) or reported across a wide variety of injury types (e.g., McCall et al. 2007). However, 2 studies specifically investigated percutaneous injuries (needlesticks, cuts, Ayas et al. 2006;Neuberger et al. 1984), one the exposure to hazardous body fluids (Macias et al. 1996), and one the occurrence of burn injuries (Horwitz and McCall 2004). Injuries resulting from ladder falls and traumatic injuries to the hand were the focus of 1 (Arlinghaus et al. 2012) and 3 studies (Lombardi et al. 2003(Lombardi et al. , 2014Tucker et al. 2006: Study 2), respectively. Surveillance methods used to collect and verify work injuries were also diverse: 16 studies made use of hospital/company records or register data, 6 studies utilized compensation claims data, and 3 used self-reports (Ayas et al. 2006;Fransen et al. 2006;Niedhammer et al. 2008) (n = 4 made no indication). Injury severity ranged from minor incidents (e.g., Neuberger et al. 1984) to severe injuries (e.g., Lombardi et al. 2014) and fatal 'accidents' (e.g., Nachreiner 2000) but the majority included work injuries of any severity. More details on the included studies (e.g., work schedule characteristics such as shift start and end times) can be found in Supplement B2.

Meta-analysis
In the following, we report the main findings of the meta-analysis for each work schedule component. Assessments of heterogeneity (between-studies variance I 2 ) and publication bias (funnel plots and Egger's regression) are described in Supplement B1.
Shift type (Figure 2a and supplement C, fig. S1) The relative risk of an occupational injury or 'accident' was not elevated on the afternoon/evening shift compared to the morning shift (RR = 0.97 [95%CI = 0.63-1.49], n = 11 studies), but increased by 33% on the night shift (RR = 1.33 [0.98-1.80], n = 16 studies). Relative risk estimates ranged from 0.38 to 2.97 for the afternoon/evening shift, and from 0.36 to 3.72 for the night shift. Removing the highest and lowest estimates did not change the overall relative risk on the afternoon/evening shift (RR = 0.97 [0.79-1.18]) but further increased risk on the night shift (RR = 1.35 [1.15-1.60]). Figure 2a) To examine the heterogeneity among studies in the component shift type, we conducted meta-regression analyses to examine potential effect modifiers extracted from the 16 included studies. Among all tested variables, effect modification was significant only by workers' age (≤ 20 yrs vs. > 20 yrs) for afternoon/evening shifts (but not night shifts). Adolescent workers (≤ 20 yrs) showed a lower relative risk on the afternoon/evening shift compared to the morning shift (RR = 0.52 [0.28-0.98], n = 2 studies, Horwitz and McCall 2005;McCall et al. 2007) and compared to adult worker's risk (RR = 1.54 [0.72-3.29], p = 0.04, n = 3 studies, Horwitz and McCall 2004;Smith et al. 1994;Violanti et al. 2012). Occupational injury risk for adolescent workers on the night shift appeared higher than that on both, morning and afternoon/ evening shifts (RR = 1.16 [0.12-11.40], n = 2 studies, Horwitz and McCall 2005;McCall et al. 2007), but lower than that of adult workers on the night shift (RR = 1.62, [1.27-2.07], n = 7 studies, Ayas et al. 2006;Fransen et al. 2006;Horwitz and McCall 2004;Niedhammer et al. 2008;Smith et al. 1994;Violanti et al. 2012;Wong et al. 2011), although non-significantly (p = 0.60). Given this effect modification, we

Morning
Evening Night

Risk Ratio
Risk Ratio  Figure 2. Results of the meta-analysis for work schedule components shift type (a), shift number (b), hours at work (c) and shift length (d), and rest break duration (e) and interval (f). Panel a shows two risk estimates for afternoon/evening shift (gray bars with and without stripes) and night shift (black bars with and without stripes): the bars without stripes (on the left) depict the estimate for all studies including adolescent and adult workers, while the bars with stripes (on the right) represent the risk estimates for adult workers only (>20 y), excluding younger workers (≤20 y). The inlay in panel a shows the effect modification of shift type by workers' age: risk for adolescent workers (≤20 y) is significantly decreased on afternoon/evening shifts as compared to both morning shifts and adult workers (>20 y). Reference categories are: morning shift (shift type), 1st shift (shift number), 8th hour, respectively, 8-h shift (shift length), and no break (rest break duration) and first 30 min since last break (rest break interval). Note that the estimate for >60 min in panel e is based on a single study (Lombardi et al. 2014). excluded the two studies with adolescent worker samples and re-calculated the relative risk estimate for shift type, resulting in RR = 1.12 [0.76-1.64, n = 9 studies] on the afternoon/evening shift and RR = 1.36 [1.15-1.60, n = 14 studies] on the night shift.

Risk Ratio
Because the number of studies was small for three of the four work schedule components (shift type, shift number and rest breaks, n = 5-8 studies), meta-regressions were computed for shift type only (n = 16 studies overall).
Shift number (Figure 2b and supplement C, fig. S2) Compared to the first shift in a block of consecutive shifts, relative risk rose exponentially for morning shifts ( 14-1.62]; n = 8 studies) while risk for afternoon/evening shifts appeared unsystematic. The trends observed for morning shifts and night shifts were however non-significant except for a significantly increased risk on the fourth night shift by 36%. Removing the highest and lowest estimates did not change the findings.
Shift length (Figure 2c,d and supplement C, fig. S3) Compared to the 8th hour at work, relative risks were not significantly elevated for hours 1 to 9, but appeared slightly increased between the 2nd and 5th hour on duty. However, beyond the 9th hour, risk rose in an exponential fashion and almost tripled after the 12th hour at work (RR = 2.73 [2.02-3.69], n = 3 studies (Åkerstedt 1995;Hänecke et al. 1998;Nachreiner 2000)) ( Figure 2c). These trends were mirrored when looking at shift length (e.g., using meta-analysis to synthesize the risk estimates for hours 1 to 8 into an 8-h shift length, the estimates for hours 1 to 10 into a 10-h shift length, etc.): relative risk for shifts longer than 12 h increased by 34% (RR = 1.34 [1.04-1.51]) (Figure 2d).

Rest breaks
The included studies investigated two important aspects of on-shift rest breaks: total duration of rest breaks (n = 2 studies, Arlinghaus et al. 2012;Lombardi et al. 2014) and interval between any two rest breaks (also referred to as: time on task or minutes since last break; n = 3 studies, Tucker et al. 2003Tucker et al. , 2006: Study 1 and 2).  Figure 3. Interaction of rest break duration and interval. A logarithmic curve and an exponential curve were fitted to the components rest break duration and rest break interval, respectively. Based on the resulting best-fit equations, relative risk estimates for duration (ref.: 0.2 h) and interval (ref.: 2h) were multiplied to examine their interaction. Panels a and b show the same results but from a different angle, i.e. panel b is rotated clockwise by 90°, to better illustrate our finding: with increasing rest break duration (in 0.2-h increments), the impact of time between breaks diminishes. Note that rest break duration refers to the total added duration of rest breaks at work, not to the duration of single breaks.

Interaction of rest break duration and interval
We multiplied the risk estimates for duration and interval of rest breaks, plotting their interaction in Figure 3. Regarding relatively short rest breaks (i.e. total added duration < 1 h), the rest break interval has a pronounced effect on injury risk: the more  Figure 4. Towards a "Risk Map". The relative risks (RRs) for an occupational 'accident' or injury are predicted for multiple work schedules using the updated "Risk Index". Reference category is one 8-h morning shift with a rest break duration of 15 min in total and 4h between breaks (labeled "ref"). The RRs for a rest break interval of 2h, 4h, and 6h are shown in panels a-f, g-l, and m-r, respectively. Within those panels, the upper ones depict rest break durations of 30 min (i.e. panels a-c, g-i, and m-o) and the lower ones durations of 15 min (i.e. panels d-e, j-l, p-r). Columns show different shift types (morning, afternoon/evening, night shift; note that estimates excluding adolescent workers (n = 2 studies) were used due to the observed effect modification by workers' age) with increasing numbers of consecutive shifts from left to right (1st, 2nd,. . .5th). Shift length is shown from top to bottom, ranging from 8-h shifts to >12-h shifts.
frequent the breaks, the lower the relative risk. Yet, with increasing duration the impact of the interval diminishes, with practically no impact when total added rest break duration is~2 h.

The updated "Risk Index"
We multiplicatively combined the risk estimates from the meta-analysis to estimate the relative risk across various work schedules (note that for shift type, the estimates excluding adolescent workers (n = 2 studies) were used due to effect modification by workers' age). The predicted values are presented in the form of a "Risk Map" (Figure 4) displaying different colors for different risk values (i.e. ranging from 'green' for RR≤1.0 to 'dark red' for RR≥2.0). This map illustrates that, compared to a single 8-h morning shift (rest break interval = 4h, total added rest break duration = 15 min), the relative risk of a work injury predicted from the Risk Index is low (RR<1.2, green regions) mostly for morning and afternoon/evening shifts shorter than 11 h, when duration of rest breaks is at least half an hour in total. The relative risk goes up substantially (RR>1.5, orange-red regions) when rest breaks cannot be taken frequently enough (i.e. < 4 h) or are too short (i.e. < 30 min), when shift length exceeds 11 h, and when work takes place during the night. When all these factors are combined, the model predicts the relative risk for a work injury to be more than twofold (dark red regions). Working 4 or even 5 shifts in a row are below a relative risk of 1.5 (greenyellow regions) when worked during the day with sufficient (i.e. ≥ 30 min) and frequent (i.e. ≤ 4 h) breaks at work but this threshold is quickly exceeded for more than 3 consecutive night shifts, irrespective of rest breaks and shift length.

Discussion
The current study updates the trends in the relative risk of occupational incidents (i.e. 'accidents' or injuries) for the components of work schedules initially reported in 2004 by Folkard and Lombardi (Folkard and Lombardi 2004). The original components were updated conducting systematic literature searches and meta-analysis, and then used as input parameters into an updated "Risk Index" model to estimate the risk associated with a work schedule. This summary estimate combines the relative risk for the type of shift (e.g., morning, afternoon/evening, and night shift), the number of consecutive shifts (e.g., working four night shifts in a row from Monday night to Friday morning), the shift length (e.g., 8-h vs. 12-h shifts), and the timing (e.g., minutes since last break) as well as duration of rest breaks at work. Using Cochrane Collaboration directives and MOOSE guidelines for meta-analysis of non-randomized, observational studies in epidemiology, 3,183 abstracts were identified and after screening by two independent raters (95-98% agreement), 29 high-quality studies published between 1962 and 2014 were included in the meta-analysis. The following overall key trends were observed: With regards to the type of shift, injury risk increased by 36% on night shifts compared to morning shifts, while injury risk on afternoon/evening shifts increased by 12%, although non-significantly. The number of hours on duty significantly increased injury risk after the 9th hour in an exponential fashion and almost tripled it after the 12th hour at work, resulting in a 34% increase for shifts longer than 12 hours. With regards to the number of consecutive shifts, risk was not significantly elevated for consecutive morning shifts but injury risk rose by 36% on the 4th consecutive night shift on duty. The significant protective effect of rest breaks was apparent for both, duration and timing of rest breaks: risk of a work-related injury decreased by more than half for rest breaks of any duration, and for time between breaks, risk increased with every additional half hour spent on the work task compared to the first 30 min. Both of these rest break results were based on only a few studies and should be interpreted with caution.
Overall, the updated risk estimates are consistent with the original estimates, substantiating the validity of the "Risk Index" and the consistency in research on work 'accidents' and injuries. However, for three of four components the number of studies included was still relatively low, continuing to indicate a strong need for further well-designed studies in these areas. Most new studies identified were for shift type; yet, recent studies allowed for the addition of the new component rest break duration, not included in the original "Risk Index". Combining the two components duration and timing of rest breaks in the new model suggests that length seems more important than frequency. However, since the individual studies used total duration of rest breaks, not duration of single rest breaks, it may be that the combination of short and frequent breaks is most effective to minimize risk; still, the number of rest breaks adequate for a given shift length remains unclear. Recent studies in professional truck drivers suggest that for a driving period of 10 hours, a third rest break does not further decrease crash risk Xie 2014a, 2014b).
Meta-regressions conducted for shift type revealed significant effect modification by worker age, in that studies of adolescent workers (aged ≤20 years), showed lower injury risk on the afternoon/evening shifts, compared to both morning shifts and adult workers. This result is based on two studies of adolescent workers, thus clearly more studies are needed on younger workers to establish reliable estimates of injury risk across components of work schedules. Yet, the notion that the afternoon/evening shift involves lower risk for younger than older workers is consistent with what is known about sleep and circadian rhythms in adolescents. The circadian system actively synchronizes (entrains) to the 24-h day via environmental signals of light and darkness. Individuals entrain differently depending on exogenous (i.e., light exposure) and endogenous (i.e., circadian response characteristics) factors that produce different phenotypes, known as chronotypes, reflecting peaks and troughs of rhythms in physiology, cognition, and behavior (including sleep) that occur at different times for different individuals (earlier for early types, later for late types). Late chronotypes sleep longer and better than early types on both afternoon/evening shifts and night shifts (Fischer et al. 2016;Juda et al. 2013). Adolescents and young adults 15-25 years of age tend to fall asleep and wake up later than any other age group (Fischer et al. 2017;Roenneberg et al. 2004). Accordingly, since on average adolescents are later chronotypes than adults, they may be better rested and aligned on the afternoon/evening shifts leading to a relative decrease in work-related injury risk. However, this is purely speculative, and another explanation could be that adolescent workers tend to work shorter shifts due to legal regulations (e.g., 16-18 year olds must not work more than 8 h/ day and 40 h/week in the US). Hence, the decreased risk for adolescents on the afternoon/ evening shift might be due to its shortness and not its timing. Similarly, less work load or less hazardous work tasks carried out by younger workers might also account for some of the decreased injury risk.
Overall, heterogeneity (between-studies variance) was large, as is expected for non-randomized, observational studies. Removal of outlier estimates did not significantly reduce the I 2value, indicating that they were not the source of the heterogeneity. Only shift number was sensitive to outlier removal, reducing heterogeneity to nonsignificance. The lower outlier estimate was from the study by Violanti and colleagues (2012), who consistently reported lower estimates on all following shifts compared to the first one. This study was the only one focusing on police officers, who often have irregular work schedules due to overtime or extra shifts with rotating days off (Vila 2006). Without sufficient rest periods between duties, officers might begin their first shift already fatigued, resulting in the 'first-day-back' effect and thus higher injury rates compared with subsequent shifts (Vila and Kenney 2002).
Publication bias was present showing a tendency toward a lack of smaller studies reporting larger risk estimates. This is inconsistent with the expectation that smaller studies with smaller, nonsignificant effect sizes do not get published as often, thus causing bias toward effect over-estimation in the meta-analysis (meaning in this study the estimates for shift length and shift number may have been under-estimated). Shift type showed a higher spread among larger studies, contrary to the expectation that smaller studies usually scatter more but this appears to be due to the two studies on adolescent workers.
The updated estimates to predict relative risk for a work injury were used in the resulting "Risk Map", which shows distinct regions of high and low risks associated with work schedules. The derived predictions are based on independently assessed estimates and their assumed interaction is modeled multiplicatively, e.g., we assumed that shift length and shift type interact in such a way that, e.g., 12-h night shifts have an even higher risk than would be expected from 12-h length and night work alone. The multiplicative model has been previously independently validated, showing good predictive power for aggregated data (Greubel and Nachreiner 2013). Given that the updated "Risk Index" is based on a higher number of studies as well as more recent research, we would expect the updated model to be even more accurate. According to our model, injury risk is clearly highest for night work, and increasing rest break duration and frequency or limiting work hours and number of consecutive shifts seems effective to a certain extent only. However, if no such limitations are in place, risk quickly doubles illustrating that even though night shifts remain relatively hazardous, risk can at least be contained with rest breaks and limited work hours.
Although one could assume a continuation of the exponential trend in risk for subsequent shifts, shift adaptation effects could also occur and thus stable or even lower risk values could ensue beyond the 4th shift. Another limitation of this study is the overall small number of studies available for meta-analysis in general and meta-regression in particular. More research is needed to identify factors that modify the risk for a work injury given the same schedule. Age, gender, and occupation seem promising candidates, but at this point no reliable estimation is possible. Among other potential factors, we propose hours of sleep and preferred sleep timing (chronotype) for future work on advancing the "Risk Index". Both sleep duration and chronotype may affect injury risk through individually varying levels of sleepiness and/or alertness. More research is needed on rest breaks, particularly on the number, place (e.g., work station/break room) and nature (e.g., eating/exercising/napping).
By synthesizing and integrating the results of a systematic review of the peer-reviewed literature, and re-quantifying the model, we propose the "Risk Index" as a tool for researchers and practitioners to estimate work-related 'accident' and injury risk based on empirical trends, thereby maximizing safety across different work schedules.