Do longer working hours lead to more workplace injuries? Evidence from Australian industry-level panel data.

Abstract

Using Australian industry-level data on weekly hours of work and frequency of new workers' compensation claims for work-related accidents over the 1990s, we examine the relationship between working time and work-related injuries is examined. Results using panel data techniques suggest that there is no relationship between working time of full-time workers and workplace safety performance. This finding is in contrast to cross-sectional evidence presented by previous researchers showing significant effects of working time on safety performance. Evidence is found in this study, however, that increased working time of part-time employed persons is associated with a greater rate of workplace injuries.

Introduction

Understanding the determinants of work-related injuries is critical to the implementation of safe and productive work practices. One potentially important determinant is working time--the length of time worked per shift, week, or indeed any other period of time. For example, fatigue associated with long hours of work may increase the likelihood of accidents, and exceptionally long hours may also result in injuries associated with breaching physical endurance limits. While it stands to reason that a relationship between working time and safety does exist, there is limited understanding of the extent and nature of the relationship. For example, it is not clear what effect working an additional hour per week has on the likelihood of injury.

In this study, we examine the relationship between working time and safety using industry-level data disaggregated at the Australia and New Zealand Standard Industrial Classification (ANZSIC) two-digit level over the period 1991-92 to 1999-2000. These data comprise industry-level information on weekly hours of work and the number of workers' compensation claims for work-related accidents. Two specific issues are examined. First, the relationship between mean weekly hours of work and work-related injuries, and whether this differs between part-time and full-time workers, are investigated. Second, the relationship between the incidence of extreme, and more specifically, very long, weekly hours of work and work-related injuries is examined.

Attention is restricted to these two dimensions of working time largely because of the limitations of available data. We acknowledge that other dimensions to working time are potentially important for their effects on workplace safety, and these two measures may not capture these dimensions. For example, the length of time worked without breaks within a shift, the length of the shift and the number of days worked in succession are all likely to be important, and are also likely to be imperfectly correlated with mean weekly hours of work and the proportion of workers in the industry working very long hours.

Data availability similarly dictates the decision to use industry-level data. Ideally, individual-level data on working hours and the occurrence of workplace injuries for a representative sample of workers would be used to investigate the issue, one advantage of which would be that examination of how the effects of working time on safety differ by industry, occupation, age, educational attainment, sex and other personal characteristics would be possible. Suitable individual-level data are not available in Australia, however.

Despite the limitations of the available data, treating them as an industry-level panel data set (a time-series of cross-sectional data on industries) can deliver valuable insights into the relationship between working time and workplace safety. In fact, the application of panel data methods to the data represents the main contribution of this study. Previous studies of this issue, whether using individual-level or industry-level data, have employed cross-sectional regression methods (for example, Kriebel (1982), Leigh (1982, 1986), Curington (1986) and Wooden (1990, 1998)). (1) Such methods are susceptible to spurious inferences deriving from unobserved characteristics. Specifically, even when variables for observed characteristics that are likely to impact on safety are included in cross-sectional regression models, unobserved factors that affect safety and are also correlated with working time are likely to remain.

Panel data econometric methods, widely regarded as powerful and robust methods for investigating a large range of issues (see, for example, Wooldridge (2002)), provide a potential alternative approach to cross-sectional methods. They are able to control for (time invariant) unobserved characteristics, and are therefore not open to the criticism that these unobserved characteristics are driving an apparent association between the variables of interest. The intuition for the approach, in the current context, is that the effect of working time on safety is identified by changes in measures of working time and safety from year to year within each industry. Put plainly, if there is a relationship between working time and safety, then a change in working time in an industry from one year to the next should, all else equal, cause a change in workplace safety performance. While panel methods have been applied to industry-level data in a number of previous studies (for example, Hubert and Pain (2000), Lui et. al. (2000) and Driver et. al. (2001)), none of these studies has examined the relationship between working time and workplace safety.

As the preceding discussion suggests, an important motivation for the current study is doubt about the validity of findings by previous research using cross-sectional approaches. Previous studies using industry-level data have essentially used differences in working hours and safety performance across industries to identify the relationship, attempting to control for other sources of differences in safety outcomes by including variables for other observable characteristics of industries. While differences in focus, data and variable definitions make generalisations difficult, it is reasonable to characteristics these studies as finding a negative relationship between working time and safety performance: that is, increased working time leads to more workplace injuries.

The current study calls these findings into question. No evidence is found that increased working time of full-time workers is associated with an increased rate of workplace injuries, suggesting that results of previous studies may derive from unobserved differences across industries that happened to be correlated with working hours--for example, the extent of manual labour used in the industry and the inherent danger of the work. Evidence is found, however, that increased working time of part-time workers does increase the injury rate.

Data

Safety outcomes

Safety outcomes are measured using data on new workers' compensation claims in each financial year over the period 1991-92 to 1999-2000, produced by the National Occupational Health and Safety Commission (NOHSC). (2) These data are compiled from claims received from insurance companies, self-insurers and some government departments for workers' compensation made under the various Commonwealth, State and Territory workers' compensation Acts. Claims covered are those that resulted from fatalities, permanent disability or temporary disability involving absence from work of at least five days. In addition to excluding claims for injuries resulting in absences from work of less than five days, the NOHSC data also exclude:

* claims for injuries or diseases occurring on journeys to or from work;

* cases compensated under legislation for specific groups of workers that is separate from the general Commonwealth, State and Territory workers' compensation legislation;

* injuries suffered by military personnel within the defence forces;

* cases not explicitly acknowledged as being work-related injuries; and

* most occupational injuries to the self-employed.

The time frame 1991-92 to 1999-2000 is largely dictated by the available data on workers' compensation claims. National data on work-related injury claims do not exist for years before 1991-92, while the most recent year for which national data currently exist is 1999-2000. (3) Indeed, the data available over the period 1991-92 to 1999-2000 are not entirely complete, with claims for Victoria and the ACT excluded for all years in order to produce a consistent measure of claims over time. The reasons for these exclusions are that data for the ACT are not available in any year, while the Victorian claims data are not available in 1991-92 and 1993-94, and for other years exclude claims involving absences from work of between five and ten days. The available NOHSC data, however, do not allow us to exclude Victoria in 1992-93. The analysis therefore excludes the 1992-93 financial year.

A further source of inconsistency over time is that changes to the system for classifying industries occurred in 1994-95. Industries were previously classified according to the Australian Standard Industrial Classification (ASIC), but have since been classified according to the Australia and New Zealand Standard Industrial Classification (ANZSIC). Although most of the 53 industry subdivisions are the same, there are some important differences. Details on the attempts made to achieve concordance between the ASIC and ANZSIC categories are provided in an appendix.

In addition to the problems in constructing the variable 'new workers' compensation claims', it should also be acknowledged that the variable is itself not a perfect measure of workplace injuries. First, only injuries that give rise to the compensation claims described above are included, implying measured workplace injuries diverge from actual workplace injuries. Importantly, the degree of divergence between the number of claims and the number of actual workplace injuries is likely to vary across industries. This is because of both differences in the types of injury incurred--that is, some industries will have a higher incidence of injuries that do not give rise to included workers' compensation claims--and differences in the propensity to lodge a claim for a given injury. For example, more highly unionised industries probably have higher rates of claims, all else being equal, because workers are likely to be better informed about their rights.

A second important limitation of using data on the number of compensation claims is that no allowance is made for the severity of the injury. It is arguable that some adjustment for severity is important. For example, an injury leading to permanent work incapacity should perhaps be given more weight than an injury leading to a one-week absence from work and no long-term adverse effects. The available data do not, however, allow us to do this.

Rather than examine the number of claims, the focus is on the frequency of claims, defined as the number of claims per one million hours worked in the industry. This frequency variable is preferable to the actual number of claims, since it adjusts for the number of hours that workers are 'at risk' of injury in each industry (i.e., the number of hours worked in the industry). The effects of differences in industry sizes are therefore eliminated. Expressing claims relative to hours worked also has the advantage of reducing sensitivity to the exclusion of groups of workers. For example, as noted above, most claims for self-employed workers are excluded, so that the hours worked by these self-employed workers are also excluded in estimating the claims frequency.

Working Time

Statistics of hours of work in each two-digit industry for each year come from unpublished ABS data. These data relate to August of each year, with the focus on mean weekly hours of work and the proportion of employees working long hours, defined as 50 or more hours per week. Managerial employees are excluded from all data used in the analysis on the ground that managerial hours are unlikely to affect the frequency of workplace injuries. (4) The data also permit distinguishing between full-time and part-time workers, and therefore allow us to examine separate hours measures for these two groups, as well as to condition on the proportion of employees who are part-time.

As indicated in Section 1, these measures may not capture dimensions of working time that may affect workplace safety. These include the length of time worked over different time frames, such as a single day, a month or a year, as well as other working time patterns, such as the extent and nature of shift work, the length of time between breaks on shifts, the number of days worked in succession and interactions between these dimensions (for example, the length of the shift combined with the time of day worked). Ideally, the effects on workplace safety of these other dimensions of working time would be examined, but this is not possible with the available ABS data on working hours in each industry. (5) The measures used do, however, provide important information on the length of working time of employees in each industry, and it is therefore legitimate to investigate their implications for the rate of occurrence of workplace injuries.

The Relationship Between Working Time and Workplace Safety Methods

Previous research using industry-level data to investigate the relationship between working time and workplace safety has adopted a cross-section regression approach (for example, Kriebel (1982), Curington (1986), Wooden (1990, 1998)). The intuition is that if industries with higher injury rates tend to have higher (or lower) working times, then we should find a relationship between the two. The equations estimated are of the form:

E([c.sup.i])= [alpha] + [beta][H.sub.i] (1)

where E([c.sub.i],) is the expected injury rate of industry i (as measured workers' compensation claims frequency), [H.sub.i], is either mean weekly hours of work in industry i or the proportion of employees working long hours in industry i, and where estimation is over all I industries in one year.

This approach has the problem that differences between industries other than differences in working time may be responsible for differences in injury rates. For example, inherently dangerous industries may tend to have more persons employed on a full-time basis, perhaps because costs of required safety training are too prohibitive to warrant employment on only a part-time basis. In such industries, the higher mean working hours do not cause the higher rate of injury; rather, a high injury rate and long working hours both derive from the more dangerous nature of the work. Cross-sectional estimates of the effects of working hours on safety outcomes may therefore reflect the effects of (unobserved) differences, such as the inherent danger of the work, rather than the effects of working hours differences themselves.

In principle, it is possible to control for other sources of inter-industry variation in injury rates and thereby accurately identify the effect of working time using cross-sectional methods. Indeed, previous studies of workplace safety have included in the models estimated variables for a number of characteristics other than working time. In practice, however, the problem of not controlling for other sources of variation in injury rates is unlikely to be completely resolved by the inclusion of additional variables. This is because some sources are likely to be unobservable (such as the use of technology to improve safety) and therefore variables for these sources of inter-industry variation in injury rates will not be available from any data source.

This problem motivates the use of panel data regression methods--specifically, 'fixed effects' panel models, which can control for differences across industries in unobserved characteristics (which are treated as 'fixed effects'--that is, effects that are constant over time). These models essentially identify the relationship between working time and safety by comparing the within-industry changes over time in working hours and safety outcomes, and can be formally described by the equation:

E([c.sup.it)= [alpha]+[beta][H.sub.it]+[v.sub.i]+[[eta].sub.t] (2)

where c and H are as defined in Equation (1). This shows that [beta] measures the variation in c associated with variation in H (across I industries and T years), controlling for that variation in c associated with industry ([v.sub.i]) and time-period (year, [[eta].sub.t]). Time-invariant unobserved industry effects are captured by v, (industry effects that are 'fixed' over time). For example, v can capture differences across industries in the innate danger of the work involved. The model also controls for economy-wide 'year effects', which are captured by [[eta].sub.t] (year effects that are 'fixed' over industries). For example, changes to government policy may change the workers' compensation claims frequency over time, and this effect is controlled for in Equation (2), to the extent that it is the same for all industries. (6)

The intuition for the fixed effects model given by Equation (2) is perhaps best understood by means of comparison with a cross-sectional approach. Cross-sectional models compare combinations of working time and safety outcomes across industries at a point in time. For example, a finding that, all other observed characteristics constant, industries with higher working times tend to have higher injury rates is interpreted as evidence that increased working time causes more workplace injuries. By contrast, the fixed effects panel model identifies the relationship between working time and safety by comparing the within-industry changes over time in working hours and safety outcomes. Specifically, variations over time in each industry's claims frequency and working hours, controlling for economy-wide changes over time in the claims frequency, are used to identify the relationship between the two. For example, if industries with the largest increases in working time tend to have the largest increases (or smallest decreases) in workplace injuries, this is interpreted as evidence that increased working time causes more workplace injuries.

Implicit in the fixed effects model is the assumption that changes in nonworking-time industry characteristics that impact on workplace safety (and which are 'unobserved' in our model) are not correlated with changes in working time (Wooldridge (2002), Ch. 10). (7) Put another way, changes in the composition of industries over time in terms of characteristics that impact on workplace injury rates, such as occupational mix, unionisation rates, female employment shares, firm sizes, shift work patterns and work experience and educational attainment levels of workers, are assumed to be uncorrelated with changes in working time if they impact on safety. A closely connected assumption is that there is no systematic association between differences across industries in the impact on safety performance of economy-wide changes over time (for example, due to regulatory or technological changes) and changes in working time in each industry.

This so-called 'fixed effects' assumption does create a potential source of bias in estimates of the effects of working time on safety. Adding variables for industry characteristics other than working time could mitigate concerns of bias from this source, but available data do not permit this. (8) While it is difficult to anticipate the magnitude and direction of the bias, one suspects that, to the extent that any bias is present, it is towards overestimation of the (negative) effect of working time on injury rates. This is on the basis that compositional changes likely to increase work-related injury rates, such as reduced unionisation, are perhaps more likely to be associated with increased working time; while there appears to be no a priori reason for believing that any differences across industries in the effects of economy-wide changes are correlated with changes in working time. This is, of course, speculative, and the magnitude and direction of bias from the fixed effects assumption is indeterminate. Compared with a cross-sectional approach, however, a panel approach arguably involves less restrictive assumptions, in that only if changes in unobserved characteristics impacting on safety are correlated with changes in working time does bias from unobserved characteristics arise. In particular, fixed industry characteristics that impact on safety performance are controlled for, whereas they are not in a cross-sectional approach. In any event, it is incontrovertible that a panel approach involves different assumptions to a cross-sectional approach, and therefore at a minimum delivers a fresh perspective on the issue.

Results

Table 1 contains information on the characteristics of the data used in the analysis, presenting descriptive statistics in each year for each variable used. The top panel shows that there is substantial variation across industries in safety performance, as measured by the workers' compensation claims frequency. In every year, the highest claims frequency is at least twenty times the lowest claims frequency. There is also significant variation over time, with an overall trend decline in the claims frequency evident: the mean claims frequency across the 53 industries declines from 21 claims per million hours worked in 1991-92 to 14 in 1999-2000.

The remaining panels of Table 1 contain descriptive statistics for the measures of working time used in the estimated models. Notable changes over time include increases in mean weekly hours of full-time employees, the proportion of employees working long hours and the proportion of employees employed part-time. Increased variation across industries in weekly hours of work, as measured by the coefficient of variation and the difference between the minimum and maximum, is also apparent.

Cross-Sectional Evidence

Cross-sectional evidence is presented to illustrate the value of the panel data methods used in this article, and also to highlight the potential for other studies to make incorrect inferences on relationships of this kind because of the failure to take into account other (unobserved) sources of differences across industries. Table 2 presents coefficients estimates where each industry subdivision is assigned the mean value for each variable over the period 1991-92 to 1999-2000 (excluding 1992-93). They show a positive effect on claims of hours worked. The estimates imply that each one-hour increase in mean hours increases the workers' compensation claims frequency by 0.65, or that a one percentage point increase in the proportion working long hours increases the claims frequency by 0.28. Although we do not control for other sources of differences in claims rates across industries (for example, the unionisation rate, the fatality rate and the age composition of employment), this is broadly consistent with other studies that do attempt to control for these differences. For example, Wooden (1998) found that each additional hour of paid overtime was associated with an increase in the claims frequency of 2.9. While paid overtime is clearly a different quantity from the working time variables employed in this study, Wooden's finding of a positive association between working time and injury rates is consistent with the cross-sectional evidence presented in Table 2. (9)

The cross-sectional results therefore imply that, although safety performance improved (as evidenced by Table 1), greater improvements would have been achieved had working time not increased, because of the positive relationship between workplace injuries and hours of work apparent from the cross-sectional data. As mentioned, however, the critical concern is that there are other differences across industries that impact on claims frequency and are correlated with working time. Failure to control for these other differences may be leading to incorrect inferences.

Panel Data Evidence

The fixed effects models estimated control for both other (unobserved) sources of differences in workers' compensation claims frequencies across industries that are stable over time and other sources of change in claims frequencies over time that are the same across all industries.

Estimates from three specifications are reported in Table 3, with the first two specifications being analogous to the two specifications estimated using the cross-sectional approach. To consider these two specifications first, the estimates imply that, in contrast to the cross-sectional results, there are no statistically significant effects on the frequency of compensation claims of either mean hours or the incidence of long hours, although a weakly significant (at the 10% level) negative effect of the incidence of long hours on the claims frequency is evident. Thus, once controls are employed for unobserved differences across industries (that do not change over time) and for the effects of economy-wide changes over time, we find no significant role for working time.

Table 3 also presents estimates for a third specification which allows for working time of part-time employees to impact on safety differentially to working time of full-time employees, by controlling for the proportion of the industry employed part-time and by distinguishing mean part-time hours from mean full-time hours. Specifically, the following four variables for working time are included:

* proportion of employees employed part-time;

* mean hours of part-time employees;

* mean hours of full-time employees; and

* proportion of all employees working more than 49 hours per week.

The results show no significant effects of the proportion employed part-time, mean hours of full-time workers, or the incidence of long hours of work. Interestingly, however, there does appear to be a 'role' for part-time employment: increases in the mean hours of part-time workers are associated with increases in the claims frequency. It is curious that it is not the rate of part-time employment itself that is associated with an increased claims frequency, but rather the hours of work of those part-time employees that are in the industry. Explanations of the effects associated with part-time work therefore need to focus not on differences between part-time workers and full-time workers, but rather on differences between part-time workers by working time. For example, it may be that higher part-time working hours per week are associated with longer individual shifts and/or shift work. There may also be an interaction effect operating between lower training levels of part-time workers and longer working time--adverse safety consequences of inadequate training may, for various reasons, grow at an increasing rate as working time increases.

In interpreting the finding with respect to working time of part-time workers, it should be noted that the association identified in this analysis is between mean working time of part-time workers in the industry and injury rates among all workers in the industry. It is therefore possible that the coefficient estimate for mean hours of part-time workers could be picking up higher injury rates among all workers due to other working patterns (unobserved in our data) that are associated with longer part-time hours. For example, it may be that longer part-time hours are correlated with late-night work or a general tendency towards long individual shifts of all persons in the industry. Nonetheless, irrespective of its source, the finding of an association between working time of part-time workers and work-related injury rates would seem to be significant.

The contrast of these findings with Wooden (1998) is particularly striking. Wooden, using the same data on workers' compensation claims, but over the period 1991-92 to 1993-94, found significant effects on claims frequencies of both paid overtime and long hours of work. His analysis is cross-sectional, and although he controlled for other sources of variation in claims across industries, the panel data results presented in this article suggest that he was not able to control for all other sources of variation in claims (that are correlated with working hours). That is, the implication of the panel data results presented here is that the highly significant effects of paid overtime and long hours on claims that Wooden found are the result of differences across industries, other than working time, that are correlated with both claims frequency and working time.

Conclusion

The evidence from Australian industry-level panel data is that hours worked per week by full-time workers do not affect safety performance. This represents a contrast with results from cross-sectional studies of this issue. For example, Wooden (1998) finds that a one-hour increase in mean paid overtime (which necessarily applies only to full-time workers) increases the workers' compensation claims frequency by 2.9 per million hours worked. The implication of this study is that it is other differences between industries, correlated with both safety and working hours, that are responsible for the observed association between hours and safety at the cross-sectional level. For example, industries that have high usage of physical labour may tend to have both more claims and higher mean hours of work for full-time workers. Thus, it may be that, rather than longer hours causing higher claims, the physical nature of the work causes both higher claims and longer hours of work.

It is, on reflection, perhaps unsurprising that working time of full-time workers is found not to impact on safety outcomes. While it is almost certain that extremely long working hours are detrimental to safety, it is not clear why this effect should operate at less extreme hours--for example, working 40 hours versus 35 hours per week. Yet, this type of effect has indeed been asserted by the existing research, which has found an increase in working time increases the likelihood of injury per hour worked at all levels of working time. It is important to emphasise that the findings of this study do not imply that working time of full-time workers has no impact on work place safety. Rather, there is no evidence of an impact at the level of variation in weekly hours observed at the industry level. Thus, it cannot be inferred from the results obtained here that working very high hours--to the point where physical fatigue develops--does not increase the risk of injury; nor do the findings of this study imply that other dimensions of working time, such as shift work incidence and patterns, number of days worked in succession, length of time between shift breaks, and so on, have no impact on workplace safety.

The fixed effects model does identify one effect of working time on safety performance, namely, that increased working time of part-time workers increases the injury rate per hour worked. There would seem to be value in future investigation of the reasons for this association. Irrespective of the reasons, however, the implication of this finding would seem to be that policies aimed at reducing workplace injuries should pay particular attention to part-time workers.

References

Australian Bureau of Statistics (ABS) (2003), 'Employment by Gender and Industry', obtained from www.abs.gov.au/ausstats February 24, 2003.

Curington, W.P. (1986) 'Safety Regulation and Workplace Injuries,' Southern Economic Journal, vol. 53, pp. 51-71.

Dawson, D., McCulloch, K. and Baker, A. (2001), Extended Working Hours in Australia: Counting the Costs, Report prepared for the Queensland Department of Industrial Relations.

Driver, C., Temple, P. and Urga, G. (2001), 'The Influence of Uncertainty on Investment in the UK: A Macro or Micro Phenomenon?', Scottish Journal of Political Economy, vol. 48, pp. 361-382.

Haenecke, K., Tiedemann, S., F. Nachreiner, F. and Grzech-Sukalo, H. (1998), 'Accident Risk as a Function of Hours at Work and Time of Day as Determined from Accident Data and Exposure Models for the German Working Population, Scandinavian Journal of Work and Environmental Health, vol. 24, pp. 43-48.

Heiler, K., Pickersgill, R. and Briggs, C. (2000), Working Time Arrangements in the Australian Mining Industry, International Labour Office Working Paper No.162, Geneva.

Hubert, F. and Pain, N. (2000), Inward Investment and Technical Progress in the UK Manufacturing Sector, OECD Economics Department Working Papers No. 268.

Kriebel, D. (1982), 'Occupational Injuries: Factors Associated with Frequency and Severity,' International Archives of Occupational and Environmental Health, vol. 50, pp. 209-218.

Leigh, J.P. (1982), 'Are Unionised Blue-Collar Jobs More Hazardous Than Non-Unionised Blue-collar Jobs?', Journal of Labor Research, vol. 3, pp. 349-357.

Leigh, J.P. (1986), 'Individual and Job Characteristics as Predictors of Industrial Accidents', Accident Analysis and Prevention, vol. 18, pp. 209-216.

Loundes, J. and Wooden, M. (2001), How Unreasonable Are Long Working Hours?, Melbourne Institute Working Paper No. 1/02.

Liu, X., Siler, P., Wang, C and Wei, Y. (2000), 'Productivity Spillovers from Foreign Direct Investment: Evidence from UK Industry Level Panel Data', Journal of International Business Studies, vol. 31, pp. 407-425.

Nachreiner, F., Akkermann, S. and Haenecke, K. (2000), 'Fatal Accident Risk as a Function of Hours into Work', in Hornberger, S., Knauth, P., Costa, G. and Folkard, S. (eds), Arbeitswissenschaft in der betrieblichen Praxis 17: Shiftwork in the 21st Century, pp.19-24, Peter Lang, Frankfurt.

National Occupational Health and Safety Commission (NOHSC) (1998a), Compendium of Workers' Compensation Statistics, Australia, 1995-96, AGPS, Canberra.

NOHSC (1998b), Compendium of Workers' Compensation Statistics, Australia, 1996-97, AGPS, Canberra.

NOHSC (1999), Compendium of Workers' Compensation Statistics, Australia, 1997-98, AGPS, Canberra.

NOHSC (2000), Compendium of Workers' Compensation Statistics, Australia, 1998-99, AGPS, Canberra.

NOHSC (2002), Compendium of Workers' Compensation Statistics, Australia, 1999-2000, AGPS, Canberra.

Wooden, M. (1990), 'Factors Associated with Workplace Accidents: Evidence From the 1983 Health Survey', The Journal of Occupational Health and Safety Australia and New Zealand, vol. 6, pp. 97-102.

Wooden, M. (1998), 'Factors Associated with Inter-Industry Differences in Workers Compensation Claims Rates', The Journal of Occupational Health and Safety Australia and New Zealand, vol.14, pp. 349-357.

Wooldridge. J. (2002), Econometric Analysis of Cross Section and Panel Data, MIT Press, Cambridge, Massachusetts.

Worksafe Australia (1993), Estimates of National Occupational Health and Safety Statistics, Australia, 1991-92, AGPS, Canberra.

Worksafe Australia (1994), Estimates of National Occupational Health and Safety Statistics, Australia, 1992-93, AGPS, Canberra.

Worksafe Australia (1995), Estimates of National Occupational Health and Safety Statistics, Australia, 1993-94, AGPS, Canberra.

Roger Wilkins *, Melbourne Institute of Applied Economic and Social Research, University of Melbourne 

Endnotes

* Thanks to Mark Wooden for helpful comments. All errors are the responsibility of the author.

(1) It should also be noted that there is a significant body of what might be characterised as 'medically-based' research, investigating the links between workplace safety and fatigue and sleep deprivation, both of which may be correlated with working time. See Dawson et al. (2001) for a review of this literature. Dawson et al. also document two studies (Haenecke et al. (1998) and Nachreiner et al (2000)) that examine the closely related issue of the relationship between length of shift (an important dimension of working time) and fatality risk. Both studies find that working beyond the ninth hour increases the fatality risk.

(2) The sources for these data are Worksafe Australia (1993, 1994, 1995), the NOHSC website (http://www.nohsc.gov.au) and NOHSC (1998a, 1998b, 1999, 2000, 2002).

(3) The potential exists to utilise state-based data over longer time frames, but aggregation to produce national estimates is not viable due to significant gaps in the data and differences in the information collected across states.

(4) While managerial employees are typically at low risk of physical injury, they are arguably at greater risk of stress-related conditions. But because our interest is in workplace injuries, and less than 4 per cent of workers' compensation claims are for mental disorders (see http://www.nohsc.gov.au for details on claims by injury type), the focus on working time of non-managerial employees appears justified. Other occupational groups could also possibly be excluded on this basis (e.g. professionals), but more ambiguity arises for these groups than for the managerial group. Consequently, no other occupational groups are excluded.

(5) Heiler et al. (2000) explore some of these dimensions of working time for a sample of 180 coal and metalliferous mines in Australia in 2000. Although these authors discuss the potential implications of working time patterns for workplace safety in the mining industry, they do not have data on safety outcomes to enable any conclusions to be drawn about the relationship between working time patterns and workplace safety.

(6) The models estimated are sometimes called 'two-way fixed effects' models, in reference to effects across both industries and years. The other main class of panel data models is the 'random effects' group of models. These models require stronger distributional assumptions than fixed effects models, in essence assuming that unobserved characteristics are not correlated with the dependent variable. These models are therefore liable to the same criticisms as cross-sectional models, and for this reason the current study focuses on fixed effects models. The benefits of random effects models are that, compared with fixed effects models, the stronger assumptions 'buy' more precise estimates; while compared with cross-sectional models, they utilise more information--both the variations across and within industries (versus only variation across industries). All else equal, the preferred approach is the fixed effects panel data model. (Random effects models were in fact estimated in this study, and inferences found to not be sensitive to the imposition of the additional assumptions.)

(7) The term 'fixed effects' which is used to describe these models was presumably coined to capture the notion that unobserved characteristics that impact on the dependent variable are assumed to be constant over time. The term is therefore a misnomer, since this is a stronger assumption than is actually required by the model. Specifically, we do not require that unobserved characteristics affecting the dependent variable be constant over time; rather, we only require the assumption that changes in these characteristics be uncorrelated with changes in the independent variable(s).

(8) Wooden (1998), who uses cross-sectional methods to examine the relationship between claims frequency and industry characteristics in Australia, draws on data sources (namely, the 1991 Census, the 1993 Survey of Education and Training, the 1989-90 Australian Workplace Industrial Relations Survey and the 1993 Working Arrangements Survey) that do not permit the time-series variation in the variables that is required for the addition of the variables to add value in panel data models.

(9) Wooden also includes a variable 'proportion of employees working in excess of 49 hours per week', but given the inclusion of the variable 'number of hours of paid overtime', he interprets this as primarily reflecting the effects of working unpaid overtime. Wooden actually finds a negative effect of this variable on the frequency of injuries, which, he argues, reflects a selection effect--workers work long 'unpaid' hours only if the work is relatively safe.

Table 1: Characteristics of the data

              1991-   1993-   1994-   1995-   1996-
               92      94      95      96      97

No. of obs.      53      53      53      53      53

                                  Safety

Frequency of new workers' compensation claims per million hours worked
Mean           20.9    18.4    20.4    16.7    14.5
Coef. of      0.785   0.560   0.775   0.557   0.552
variation
Minimum         2.1       2     3.2     1.9     0.9
Maximum        85.9    47.7   105.2      41    34.7

                               Working-time

Mean weekly hours of all employees
Mean           35.2    35.5      36    36.3    36.7
Coef. of      0.131   0.141   0.144   0.138   0.147
variation
Minimum        17.7    19.6    18.5    19.1    20.5
Maximum        43.5    47.9      48    46.1    54.8

Proportion of all employees working more than 49 hours per week (%)
Mean           13.2    14.5      17    17.5    18.1
Coef. of      0.561   0.503   0.535   0.543   0.569
variation
Minimum           0     4.8     5.1     5.2     4.2
Maximum        34.5      39    40.6    43.5    57.7

Proportion of employees working part-time (%)
Mean           16.5    18.1    18.5    18.4      18
Coef. of      0.982   0.902   0.905   0.907   0.866
variation
Minimum           0       0       0       0       0
Maximum        72.6    74.7      71    71.2    64.2

Mean weekly hours of part-lime employees
Mean           16.5    16.6      17    17.1    16.2
Coef. of      0.285   0.277   0.259   0.281   0.278
variation
Minimum           0       0       0       0       0
Maximum          32      24    28.6    26.9    26.9

Mean weekly hours of full-time employees
Mean           39.3    40.1    40.6    40.9    41.1
Coef. of      0.069   0.080   0.079   0.064   0.085
variation
Minimum        33.8      34    34.7    34.1    35.7
Maximum        47.4    49.9    52.4    48.4    54.8

              1997-   1998-   1999-    All
               98      99     2000    years

No. of obs.      53      53      53     424

                             Safety

Frequency of new workers' compensation claims per million hours worked
Mean           14.9    13.1    13.9    16.6
Coef. of      0.638   0.557   0.734   0.699
variation
Minimum         0.3     1.8     1.2     0.3
Maximum        54.1    36.7    72.1   105.2

                          Working-time

Mean weekly hours of all employees
Mean           36.3    36.4    37.1    36.2
Coef. of      0.135   0.151   0.164   0.144
variation
Minimum        20.3    16.4    18.2    16.4
Maximum        47.8    50.4    50.3    54.8

Proportion of all employees working more than 49 hours per week (%)
Mean           17.3    17.3    19.3    16.8
Coef. of      0.497   0.491   0.544   0.542
variation
Minimum         3.6     2.4     3.8       0
Maximum        43.3    43.7    50.6    57.7

Proportion of employees working part-time (%)
Mean           17.9    18.3    18.4      18
Coef. of      0.916   0.961   0.852   0.912
variation
Minimum           0       0       0       0
Maximum        74.7    84.3    65.7    84.3

Mean weekly hours of part-lime employees
Mean             17    16.4    16.9    16.7
Coef. of      0.282   0.335   0.331   0.293
variation
Minimum           0       0       0       0
Maximum          32      32    30.7      32

Mean weekly hours of full-time employees
Mean           40.6    40.9    41.5    40.6
Coef. of      0.074   0.071   0.104   0.081
variation
Minimum        34.3    36.8    35.3    33.8
Maximum        48.7      53    57.6    57.6

Table 2: Estimates of the effects of working time on the frequency of
workers' compensation claims--Cross-sectional evidence (mean values
of variables across all years)

                                Specification (1)   Specification (2)

Constant                             -6.923            *** 11.893
                                     (9.702)             (3.092)
Mean hours                          ** 0.650
                                     (0.266)
Proportion working long hours                            * 0.280
(%)                                                      (0.166)
Adjusted R squared                   0.087               0.034

Notes: Number of observations is 53. The dependent variable
is the frequency of new workers' compensation claims per one
million hours worked. Standard errors are reported in parentheses.
* indicates the coefficient estimate is statistically significantly
different from zero at the 10% level, ** indicates significance at
the 5% level and *** indicates significance at the 1% level.

Table 3: Estimates of the effects of working time on the
frequency of workers' compensation claims--Fixed effects
models

                      Specification   Specification   Specification
                           (1)             (2)             (3)

Constant              22.217 ***      22.506 ***      21.028 ***
                         (5.639)         (1.279)         (7.998)
Mean hours               -0.039
                         (0.158)
Proportion working                       -0.125 *        -0.108
long hours (%)                           (0.073)         (0.105)
Proportion working                                       -0.115
part-time (%)                                            (0.095)
Mean hours of part-                                     0.235 ***
time workers                                             (0.086)
Mean hours of full-                                      (0.018)
time workers                                             (0.222)
R squared within          0.195           0.202           0.220
R squared between         0.105           0.053           0.028
R squared overall         0.049           0.028           0.083
Rho                       0.724           0.735           0.721
Number of observations      424             424             424

Notes: The dependent variable is the frequency of new workers'
compensation claims per one million hours worked. These results
exclude data from the 1992-93 financial year. Standard errors
are reported in parentheses. * indicates that the coefficient
estimate is statistically significantly different from zero at
the 10% level, ** indicates significance at the 5% level and
*** indicates significance at the 1% level. 'R squared within'
is the proportion of the variation in claims frequency within
industries (across years) 'explained' by variation (within
industries) in the independent variables. 'R squared between'
is the proportion of the variation in claims frequency across
industries 'explained' by variation (across industries) in the
independent variables. 'R squared overall' is the proportion
of the variation in claims frequency 'explained' by the model.
'Rho' is the proportion of the variation in claims frequency
'explained' by variation in claims frequencies across
industries.