Allen, Harris PhD; Woock, Christopher PhD; Barrington, Linda PhD; Bunn, William MD
Questions about the relative benefits and costs of aging workers loom large for employers as they face incentives to retain these workers stemming from trends in the labor market. On one hand, factors like workforce demographics (eg, baby boomers advancing toward retirement) and changes in the nature of work itself (eg, its increasingly specialized nature) coupled with the advantages that older workers often bring to the job (eg, greater experience, expertise and organizational knowledge) are generating a new impetus for hiring and holding on to this segment of the workforce. On the other hand, the prospects of older workers incurring excessive direct and indirect costs have become a compelling deterrent. The conundrum these competing considerations pose is coming into clearer focus for many employers as they endeavor to maintain or improve their organizational financial bottom.
Of particular interest in this equation are the indirect costs—especially those stemming from the productivity losses associated with the onset of adverse health conditions, workplace injuries and illnesses, and other work limiting disabilities. The direct costs arising from claims-tracked health care and disability utilization—and how younger and older employees compare in this regard—are a reasonably well-known quantity. What are less well-known are the indirect costs, especially those resulting from health-related productivity loss, and how older employees fare versus their younger counterparts in this regard. The fact that employers are increasingly subscribing to the view that indirect costs are both the “elephant in the room” (ie, the largest part of the total health cost burden) and manageable (ie, the potential for which is now just beginning to be explored) is only fueling the need to better understand this component.1
This study takes up the issue of age differences in indirect costs with a case study comparing younger and older employees at a heavy manufacturing company with regard to the productivity loss they incur as a result of working overtime. Overtime often functions as a tool for business operations, a key element in the strategy which a company uses to manufacture and deliver its goods and services so as to succeed in the marketplace. As such, it provides fertile ground for exploring this issue.
Consider, for example, companies which experience cyclical demand for their goods and services. They are often structured such that the more their employees work longer hours, the more efficient and profitable their operations become. They thus have the incentive to promote the use of overtime among their workforce. But, is it also the case that overtime spawns greater rates of adverse events that generate productivity loss and thereby new indirect costs? If so, are these increased rates of a magnitude that, once documented, they raise the need to rethink the use of overtime in business operations?
It was this issue that led Navistar, Inc, headquarters for International Truck and Engine Corporation, a global manufacturer of medium and heavy trucks and mid-range diesel engines based in Chicago, IL, to sponsor research published last year in this journal and elsewhere2–4 that focused on the health, safety, and productivity impact of overtime among its US workforce. The research team included experts from the United Auto Workers and the National Institute for Occupational Safety and Health, and its study findings were unexpected. The vast majorities of employees working overtime were able to do so without incurring increased rates of injuries, illnesses, or impaired productivity. Where such rates did occur, they were confined to certain subgroups of employees doing certain types of work in certain settings at certain levels of longer workhours. Moreover, overtime proved much less predictive of adverse outcomes than prior health and other antecedents to hours worked.
Although this earlier study did not include in its analysis tests for age effects, it is quite plausible that younger and older employees will register differences in the impact of working overtime on health, productivity, and safety outcomes. Three lines of research provide grist for hypotheses in this regard. First is the considerable work done confirming the declines in physical health, functioning, and capabilities that attend advance in age.5 These studies suggest a first hypothesis that when older employees work longer hours they will be more likely to incur injuries, illnesses, or impaired productivity.
Conversely, a number of recent studies have documented improved mental health and well-being with advancing age.6 This leads to such correlations as better knowing one's limitations enabling older employees to work “smarter not harder”. Older employees also often have more control over the types of positions and kinds of duties they assume by virtue of seniority. Such factors give rise to a second hypothesis: As employees age, they will find ways to neutralize any greater risk of injuries, illnesses, and/or impaired productivity that might logically accompany declines in their physical health, functioning, and capabilities when they work overtime. With respect to the risk of injuries, illnesses, and/or impaired productivity, this neutralization may lead to no differences between age groups. It could also lead to actual decreases in overtime-related adverse event rates with advancing age. Either way, the result is the obverse of the prediction that adverse impacts of overtime increase as employees get older.
A third line of research supports what in effect is a blend of these two hypotheses. Older employees record fewer disability episodes than younger employees, particularly those resulting from a prominent cause of this type of loss, musculoskeletal injuries. But, once on disability, the amount of time they require on average to return to work is longer.7,8 Thus, whether it be because they have less risky job positions (because of seniority) or because they more aware of what they can and cannot do in whatever job they have, older workers are less likely to go on to disability—an edge readily attributable to better mental health. On the other hand, once on disability, they take longer to recover, presumably because of the declines in physical functioning that invariably attend aging.
The study reported here provided a forum for testing these two hypotheses in the context of new analyses conducted on the same data set as the earlier study. With the cosponsorship of The Conference Board, this study examined the same issues but expanded on the previous work by explicitly highlighting age as a moderating variable.
In so doing, they address the kinds of questions that will only become more urgent with time for companies like Navistar given their future workforce imperative. Do older workers incur greater rates of indirect costs than younger workers? If so, do these greater rates occur across the gamut of types of productivity loss? Or is the type of greater loss they incur as a function of longer workhours more confined? That is, are older employees at risk for only certain types of productivity loss when doing certain types of jobs at certain levels of longer work hours? By grappling with these questions now, companies like Navistar will be better positioned to cope with the unique challenges of a workforce segment that will become increasingly important to their success in the marketplace.
Materials and Methods
To conduct this study, the project team accessed an integrated database whose development Navistar had previously sponsored to monitor the burden of disease.9–12 This disease management initiative included an educational intervention mounted during summer 2001 using a quasi-experimental/control, pre/postdesign, with surveys fielded in May and September of 2001. Survey results were supplemented with archival data on workers compensation (WC) and disability from May 2000 to April 2002. With a few modest enhancements, this database was converted to monitor overtime impact for the previously reported study and for the age comparisons in this study.
The original project's approach to data collection was “general population” in orientation: All employees at each site were eligible for inclusion regardless of clinical status. The survey was roughly 100 items in length and covered a wide variety of topics, including health status, chronic disease, and presenteeism. Its administration was conducted electronically for some groups and manually in on-site kiosks for others. Both delivery approaches garnered overall response rates in excess of 50%. The measures of adverse events were captured from administrative databases: Navistar's WC and short-term disability (STD) database and its payroll database. The measures drawn from these databases covered all employees.
The study population on which this database was developed consisted of all active 10,000+ employees at each of six Navistar locations in Illinois and Indiana. This population averaged over 46 years of age and was roughly 80% men. Nearly two-thirds held an hourly versus a salaried position.
The longitudinal sample on which the analyses herein were conducted consisted of all employees reporting nonmissing data on all the survey and administrative measures in the integrated database (n = 2746). The representativeness of this panel vis-à-vis its source population has been well established.10
Table 1 highlights age and the average workweek hours for spring 2001 in a breakdown of the sample by various personal and job characteristics. The top panel of the table splits age apart into nine groupings by 5-year increments, starting with 29 or younger and proceeding upward (ie, 30 to 34; 35 to 39) to 65 or older. These groupings were selected as the point of departure based on the observation that public policy based on age is often predicated on such 5-year increments. The bottom half of the table splits workhours into four groupings: those working <40 hours per week, 40 to 48 hours, 48.01 to 59.99 hours and 60+ hours. The second and third of these served to differentiate what company policy considered normal working hours from overtime (herein labeled “moderate overtime)”. The fourth (60+) used a common definition found in the literature to identify those working “extended overtime.”
As shown, as age increased, this sample consisted of more men and more likely to be compensated on an hourly as opposed to salaried basis, but less likely to be exempt from overtime. In contrast, job type distinctions (ie, the proportion of skilled or production job posts) stayed constant across the age range, whereas tenure with the company and the length of the workweek were significantly greater in the older age groups.
Similarly, as workhours per week increased, the sample in general became older and more men and more likely to be hourly but also to be exempt from being paid overtime. When those working less 40 hours per week were excluded, these patterns became less pronounced but still remained significant. Interestingly, the percentages of those in skilled and production positions, and length of time at the company, were virtually equivalent across the various levels of workhours.
Three categories of measures were used in this analysis: average workhours, antecedents (factors whose origins preceded any effects stemming from the number of hours worked), and outcomes. Age—scored continuous in its original form—was treated as an antecedent in this classification.
The workhour measures were those used for Table 1. They blended data drawn from survey self-reports and payroll/administrative data. They included four dummy variables (coded 0,1) assessing different levels of average hours per week recorded during the April to June 2001 period: <40, 40 to 48.0, 48.01 to 59.99, and 60+.
Five categories comprised this category: demographics (age and gender), job characteristics (compensation, company tenure, and job type), prior work-related illnesses and injuries, prior health (health status and disease), and health risks (smoking and overweight). The demographic and job characteristic data were drawn from Navistar's payroll database. Separate counts of work-related acute, musculoskeletal, and other injuries and of disability episodes incurred during the prior year were developed from the company's workers comp/disability database. The health and health risk behavior data came from the May survey. The health measures included a count of preexisting diseases reported in response to a 21-condition and scales of prior physical health and mental health formed from survey responses to four health status items taken The health status measures referenced the 4 weeks before the May survey.
This set included 10 health, safety, and productivity measures, all of which are introduced in Table 2. Three health measures came from the 9/01 survey: two assessed physical health and mental health during the prior 4 weeks and were based on a repeat assessment of the same three health status items appearing in the May survey, whereas the other was a repeat count of the same list of chronic conditions.
Four safety measures included counts of any STD incidents or any acute, musculoskeletal, or other injuries incurred during May 1, 2001 to April 30, 2002 (the latter developed from the WC data). In addition, three self-report measures from the second (9/01) survey captured health-related limitations during the 4 weeks before the survey and herein classified as measures of presenteeism (ie, impaired on-the-job work performance). The nonphysical measures in this category included an item designed to capture health impact on the time, mental/physical and output demands of work, and an item rating health impact on overall work effectiveness. The physical measure included an item on difficulties with bending and twisting at work. For more information on these measures, see Allen et al2 and Bunn et al.10
The analyses were conducted in two stages. The first served to determine the key cut points on the age continuum to identify groups for the age comparisons. The second was designed to compare the groups so identified in multiple group models specified to test for differences in the relationships between overtime and study outcomes.
As mentioned above, because public policy initiatives are often based on age groupings spanning 5-year periods, the starting point for this first stage divided the age continuum into nine categories, starting with 29 or younger and proceeding upward in 5-year increments (ie, 30 to 34; 35 to 39) to 65 or older. Then, an iterative, building-up strategy was undertaken to distinguish those neighboring groups that could be collapsed together with no significant loss of predictive information. The objective was to differentiate these groups from those neighboring groups that needed to remain separate, for which collapsing would result in a significant loss of predictive information.
Toward this end, each of the nine groups was first compared with its neighbor as dummy variables in multivariate regression analyses. Those groups emerging significant were then explicitly compared with the 60+ age group again via multivariate regression. These two steps were designed to identify the minimum number of age groupings whose differentiation and cut points were justifiably rooted in the empirical results.
The final step in this sequence consisted of an exploratory regression of the full set of study outcomes on dummy variables reflecting the minimum age groupings obtained from these analyses. This step examined direct effects only and made no adjustments for potential confounds. It as such provided a “ballpark” benchmark for interpreting the results when age differences in overtime impact on these same outcomes were examined.
This second stage was likewise undertaken in several steps. First, a multiple group hierarchical structure, reflecting the “direct” and “indirect effects” model reported in our previous work2 segmented for each of the identified age groups, was estimated. The series of models tested to estimate this structure made certain adjustments to the previous model in order to maximize the current test (eg, adding Company Tenure as an antecedent), but otherwise retained the former model's three-step causal structure.
In the context of this structure, evaluation of the two hypotheses encompassed not only the direct impact of all predictors on the study outcomes but also the indirect or mediated impact of antecedents on the outcomes via the workhour measures. This required inspection not only for the direct effects of all antecedents and workhour predictors on the study outcomes, but also the effects of the antecedents on workhours as well as the total effects (both direct and indirect) of all predictors on the outcomes. All such parameter estimates were examined for their import and implications for the two hypotheses.
The estimates obtained from this first series of models were calculated on all employees in the panel, thereby producing “main effect” comparisons between older and younger groups at the aggregate level. The prior study had also, however, provided evidence of systematic differences between certain subgroups, eg, hourly versus salaried employees.
The next step for this stage, therefore, focused on the possibility that the impact of age differed at various levels of various antecedent measures. First, specific interaction terms crossing age with the antecedent measures were formally tested in the context of multivariate regressions. Then, more multiple group models comparing the three age groups were run, with each series devoted to each subgroup associated with the antecedent characteristic that emerged significant from the interaction tests with age.
The model fitting sequence followed here has been described elsewhere.2 The tests of the multiple group formulations, however, entailed certain additional steps. First, the “freed” version of the model, in which all causal structure parameter estimates across all tested groups were allowed to be free for estimation, was identified. Then, various estimates reflecting hypothesized predictions were constrained to be equal and the model reestimated. Significance tests conducted on these constraints examined whether a significant loss in predictive information obtained when the constraints in question were made. A test result was not significant indicating that the constraint could be made with no appreciable loss information, ie, the groups in question did not significantly differ on the parameter. Conversely, a significant result indicated that the loss in information was significant, signifying that the groups in question did indeed differ on the parameter.
In this context, assessment of the parameter estimates occurred in three ways. The first involved inspection standardized beta coefficients for key causal structure parameter estimates and their corresponding critical t-ratios. The standardized betas and their ratios enabled comparisons of the relative strength and magnitude of estimates when contrasted across groups across multiple group models.
Second were the contrasts of nested models, which obtained when parameter(s) were dropped and a given model reestimated. Each of these modifications yielded a variant nested within the given model, and the nested sequence allowed χ2 difference tests to be calculated to ascertain the contribution of the parameter(s) to the full model.
The third method pertained specifically to the constrained parameters and was a special case of method #2. The net effect of constraining a parameter to be equal across two groups was the dropping of a freed parameter (ie, the loss of 1 df) from the overall model. This created a nested variant and thus enabled the contribution of the freed parameter to be assessed by examining differences between the nested models.
Here, however, the approach went one step further by taking into account the multiple correlation across of the constrained estimates. The multivariate test involved a forward stepwise procedure, by which the constraint with the largest univariate χ2 was entered first. Next, the constraint adding the next most to the value of the χ2 was added to the test, etc. This process continued until cumulative contribution of the constraints reached the 0.05 level. All constraints included up to this cutoff point were judged significant. All constraints added after the 0.05 cutoff were not judged significant. In accordance with the previous study, the STATA software program was used for data management and for all above “direct effect” (ie, A → B) only analyses.13 The EQS program was used to assess all “direct and indirect effect” (ie, A → B → C) models.14,15
From Hypotheses to Testable Predictions
Our two originating hypotheses were translated into the following predictions. For hypothesis #1—as age increases, employees will incur more injuries, illnesses or impaired productivity—increasing rates of adverse consequences (ie, worsened health outcomes; greater work limitations; more injuries or disability episodes) with advancing age would be deemed as grounds for not rejecting the hypothesis. For hypothesis #2—as age increases, employee will incur fewer injuries, illnesses or less impaired productivity—patterns in which older age groups recorded rates of adverse consequences that were less than younger groups would be taken as grounds for not rejecting the hypothesis.
This section begins with tests that explored the direct impact of age on the study outcomes both as a continuous variable and as a collection of variables stratifying the age continuum to determine how to score age. It then moves to findings for the three age group mo/aggregate sample model, then to the results for the three age groups segmented by the one characteristic yielding a noteworthy interaction with age. We conclude with the summary results from the nested model comparisons.
The discussion for each model focuses first on the effects exerted by the two levels of overtime measured in the study: moderate and extended (48.01 to 59.99 and 60+ workhours per week, respectively). The focus then turns to effects associated with various parts of the model that reflected effects possibly associated with past overtime worked at some past point. It culminates by singling out effects associated with various “controls” or “antecedents” that had a bearing on the group comparisons. Throughout, the emphasis is on evidence differentiating hypotheses #1 and #2.
The “possibly a function of overtime” portion of this discussion merits special comment. This area included the effects linked with working less than 40 hours per week (both as dependent variable and predictor) and with prior acute, musculoskeletal and “other” work injuries as well as STD episodes (as predictors). No data were available to link these variables to long workhours. Thus, workers could have been working less than 40 hours per week or have incurred prior injuries or disabilities for reasons other than overtime. Herein, we nonetheless follow the precedent of our previous article in ascribing these linkages to overtime (with the appropriate qualifications). That is, if the employee was working less than 40 hours at the time of the analysis or had recorded previous injuries or disabilities, it is assumed that previous overtime was the culprit. This assumption enabled this exercise to be as generous as reasonably prudent in making attributions to overtime impact when determining the viability of the two hypotheses.
Grouping Age for the Comparisons
The tests for determining where to collapse the age categories proceeded in two steps. The first consisted of multivariate regressions, with the results for each presented in Table 3. Each regression focused on one of the nine categories comprising the initial classification for age (see Table 1), with the category in question scored as a dummy variable (1 = the category itself; 0 = other). The analysis regressed the set of 10 outcomes on this variable and confined the analytic sample to this category of employees plus the employees in the adjoining age category. The coefficient in this context registered whether the two categories in question differed. A result that was not significant indicated no and meant that the two categories could be collapsed with no real loss in prediction. A significant result indicated yes and meant that the two categories should remain distinct.
As shown, the following pairs recorded significant (P < 0.05) differences in this analysis: 40 to 44 versus 45 to 49; 45 to 49 versus 50 to 54; and right at the cutoff, 50 to 54 versus 55 to 59. Under the preceding logic, all age groups 40 to 44 and below could be combined, and all age group 55 to 59 and above could be combined, with no loss in predictive power.
The second step involved a similar analytic approach and was undertaken primarily to explore further the robustness of keeping the 50 to 54 and 55 to 59 age groups distinct from each other and from the upper age groups. Here, multivariate regression was again conducted but with two differences. The first had to do with which groups were compared. In this case, each of the following age groups were pitted against the 60+ age group as the hold-out group: 44 or less, 45 to 49, 50 to 54, and 55 to 59. The second difference concerned how measures were grouped for the tests. For the pairings of each of these four age groups with the 60+ group, three omnibus F statistics were calculated: one for the paired groups on the three health measures, the second for the paired groups on the three productivity measures, and the third for the paired groups on the four safety measures. This made for a total of 12 omnibus F tests, and to help guard against a spurious interpretation, a conservative bonferroni adjustment was applied to the significance criterion (in effect, decreasing its associated P-value from 0.05 to 0.004).
As shown in Table 4, the <45 and 45 to 49 age groups, respectively registered clearly significant differences relative to the 60+ age group on all 10 measures on all three omnibus F tests. This was not so for the 50 to 54 and 55 to 59 age groups. The 50 to 54 group reported significant differences on the health measures (both the individual and omnibus tests), but their differences on the productivity set were mixed (with two of three measures not significant and the omnibus test result equivocal) and their differences on all safety measures were not significant. For the 55 to 59 age group, only one measure in each of the health and productivity sets produced a significant result, and the two corresponding omnibus F tests were equivocal, whereas the safety measures all recorded were not significant.
Considered altogether, these findings indicated that a further collapsing of the age groups at the older end could be achieved with no substantial loss to the integrity of the age comparisons. That is, the 50 to 54 and 55 to 59 age groups could be combined with the 60+ age group, with the most consistent and noteworthy differences on the age continuum remaining intact. As such, the costs of retaining a more differentiated set of age groups were outweighed by the benefits of settling with a more parsimonious typology. Based on this reasoning, the following age groups were finalized: <45, 45 to 49, and >49.
Two additional probes explored the repercussions of this scoring decision. The first deployed the three group taxonomy of age to develop a comparative look at how older employees compared to the younger two age groups on health, productivity, and safety outcomes. These tests were “direct effects only” multivariate regression of the 10 study outcomes on two dummy variables (one each for the two younger age groups), with no controls for antecedents or for hours worked, as reported in Table 5. The omission of a term for the oldest age group meant that the latter was the “hold-out” group for these tests.
As shown, the oldest age groups reported significantly better mental health outcomes and fewer musculoskeletal injuries than the youngest age group—both results replicating previously reported findings in the literature. But the older age group also recorded significantly poorer physical health outcomes, more diseases, more limitation on all three measures of presenteeism, and more acute injuries and STD episodes.
This result provided context for interpreting the group comparisons that followed.
The second probe examined the interaction of age, scored as a three-level measure to reflect these groups, with workhours in predicting the 10 study outcomes. This analysis included main effects for age and workhours, and an interaction term crossing the two variables, in a multivariate regression with no other statistical controls. The main effects for both measures proved significant (for age: F(10,2481) = 2.72; P < 0.003; for workhours: F(10,2481) = 1.94; P < 0.036), but their interaction did not (F(10,2481) = 1.48; P < 0.139). This second result meant that explicit tests of the interaction of age with workhours could be excluded from the modeling that followed.
Figures 1 and 2 show the hypothesized models that guided the tests during stage two. Tables 6 and 7 focus for the three age group/aggregate sample model whereas Table 8 reports the age × antecedent interaction results. Tables 9 and 10 focus on the three age group/hourly only model and Tables 11 and 12 focus on the three age group/salaried-only model. Table 13 reports overall goodness fit and χ2 difference tests for these models.
Figure 1 depicts the tested variables and their hypothesized relationships in the core hypothesized structure examined across all stage two models (see Allen et al2 for further details describing this structure). Figure 2 illustrates how the multiple group model testing procedures expanded on the causal portion of this hypothesized core structure.
As shown in Fig. 1, the measurement portion of the core structure treated all but one measure as separate measured variables (denoted by rectangles). This one exception—a latent factor (denoted by an oval) labeled presenteeism/nonphysical—was confirmed by factors analyses (not reported) to underlie overall work effectiveness and two items comprising the Capacity to Meet Nonphysical Work Demands Scale (one on meeting Time demands; the other on meeting mental/interpersonal demands). The results indicated that the other productivity self-report, capacity to bend and twist, was best treated as a separate observed variable. The measurement structure for stage 2 thus reduced the number of study outcome measures from 10 to 9.
Although this core structure retained the essential formulation tested in the prior study,2 it was complex, especially so given that it was to be assessed in the context of multiple group models. In an attempt to reduce this complexity, we conducted some additional factor analyses to explore for the presence of other latent factors. To retain the longitudinal element, these analyses were conducted separately for the nine study outcomes, the three workhour measures, and the 14 exogenous independent variables. In each instance, however, the results (also not reported) failed to uncover even one factor with an eigen value of 1.0 or more, a conventional criterion for factor structure. The findings, as such, confirmed the need to retain the core structure in this context.
As for the causal portion of this structure, three dummy workhour variables (with the 40- to 48-hour group treated as the “hold-out”) were placed at the center. They were predicted by, on the left, demographic/job/health characteristics that were antecedent to the workhour measures. They predicted, placed to the right, the study outcome measures that spanned timeframes clearly subsequent to the workhour measures. This structure specified that workhours could affect the study outcomes directly and also mediate the indirect effects of the antecedent characteristics on the study outcomes. It also posited direct effects that bypassed workhours by linking antecedents to the outcome measures.
As described, our procedures in effect assessed this core structure against the sample input covariance matrix computed for each group tested. Every effort was made to replicate the testing of this entire structure on each group. The only adjustments needed were when a specific group contained no nonmissing values for a given measure. In such instances, the variable in question and all of its hypothesized (measurement and causal) parameters vis-à-vis the other variables in the model were dropped. Otherwise, the core causal structure was retained intact in its entirety.
Three Age Group Model (M3_1)
Table 6 gives the results for the three workhour dependent variables across the three age groups (<45, 45 to 49, >49) obtained in the three age group/aggregate sample model. Table 7 gives the results of the nine health, productivity, and safety outcomes. To hone in on the issue at hand, these tables report the standardized beta coefficients for each group—the test result, which as stated above, is explicitly framed to permit comparison of a given parameter estimates across groups in this model and across models. These tables denote all significant (P < 0.05) parameter estimates with a “@”.
In addition, Tables 6 and 7 give the significance results of the constraint tests (ie, the tests constraining parameters to be equal across specified groups). The notation is framed to have the reader start with the left-most column and work right. The second column from the left is designated column “a” while the third column from the left—or in this case the column to the far right—is designated column “b.” For any given parameter, an “a” superscript in the column just to the right of the numeric entry indicates that this group's estimate of the parameter significantly differed (P ≤ 0.05) from group “a's” estimate. Similarly, a “b” superscript just the right of the numeric entry indicates that this group's estimate of the parameter significantly differed from group “b's” estimate. The absence of an “a” or “b” superscript means that this group's estimate did not significantly differ from the other group(s)' estimate.
When comparing group estimates of a given effect (eg, the impact of extended overtime on physical health outcomes for the <45 vs >49 age groups), our interpretation was driven by two criteria: 1) whether one or more of the parameter estimates were statistically significant (ie, accompanied in the tables by a “@”) and 2) whether the associated constraint tests indicated a significant increment in the overall χ2 when the parameters were constrained to be equal. Each type of criteria was considered as notable in its own right. The evidence for a meaningful difference was considered strongest, however, when both criteria signified statistical significance.
For reference, Appendices A and B give the unstandardized beta coefficients, the standard errors, and the critical t-ratios for the estimates in Tables 1 and 2, respectively. The input covariance matrices and the full set of “freed” and “constrained” results (for this model and the two models that follow) are available on request from the author.
Working 60+ hours as a predictor exerted differential impact between the age groups, but this differential was limited in scope (Table 7 and Appendix B). Relative to <45 employees, the oldest group recorded greater rates of presenteeism (nonphysical) limitations on both criteria. Moreover, both this group and the 45 to 49 group posted significant rates of acute work injuries, compared with the estimate for the <45 group that was not significant, although the corresponding constraint tests were not significant. In addition, whereas working 60+ hours was significantly linked to better mental health outcomes at the aggregate level, this linkage did not surface for specifically any of the three age groups when the latter were looked at separately. All of these findings were taken as consistent with the hypothesis (ie, #1) that adverse rates increase with advancing age.
On the other hand, several findings here were more consistent with hypothesis #2. The >49 group did not record significant rates of adverse outcome differences on either criteria for any other six study outcomes (Table 7 and Appendix B). That is, its physical and mental health outcomes, its disease count, its capacity to bend and twist, and its rates of STD episodes, and musculoskeletal injuries proved no worse than either other age group. Moreover, whereas the STD episode rate for the oldest group was not significant, the STD rate for the 45 to 59 group was significant. Although the corresponding constraint was not significant, this particular finding suggested an outcome for the oldest group that was not as adverse as that for the middle age group.
In addition, those in the middle age group working 60+ hours recorded no significant differences with either other age group on any of the other study outcomes, with one exception: its mental health outcomes were significantly less favorable than the <45 groups, but this was also the case relative to the >49 age group. Otherwise, its rates on the other two health outcomes, all four safety outcomes, and the two productivity outcomes all proved statistically the same (ie, not worse) as the youngest (ie, <45) group.
Working 60+ hours, in turn, was best predicted by exempt status, prior diseases, and prior other injuries (Table 6 and Appendix A). Although exempt status' ties to 60+ work hours was significant for each age group, this linkage was significantly greater among 45- to 49-year olds than among >49-year olds. The oldest age group also registered less of linkage between prior diseases and prior other injuries, on one hand, and 60+ hours on the other. Thus, of all the possible predictors of 60+ hours, the only ones to emerge significant either were a job characteristic (exempt status) and a prior health and safety rate whose impact registered in a direction that had no particular adverse implication for the oldest age group. This pattern, too, was more consistent with hypothesis #2 than hypothesis #1.
On the other hand, moderate overtime (48.01 to 59.99 hours) was predicted by gender, exempt status, job type, prior physical health, prior disease count, prior other injuries and STD episodes, and company tenure (Table 6). Here, although not uniform, the effects overall conveyed a pattern indicative of better capacity on part of the younger groups to be working at this level of hours. The prior physical health of the 45 to 49 group was significantly more positive than the >49 group, whereas the <45 group alone registered a negative link between prior STD episodes (although not significantly more so than the older age group). Although the <45 age group also posted significantly more prior diseases than the >49 age group, this pattern was more consistent with hypothesis #1 than #2.
Turning next to the “possibly a function of overtime” sector, those working less than 40 hours a week and 49 years of age or older recorded more prior STD episodes than the youngest age group (Table 7 and Appendix B). But, their prior physical and mental health, their number of prior diseases, and their rates of prior acute, musculoskeletal and other work injuries were statistically the same as both younger age groups. This pattern, viewed on the whole as such, was more consistent with hypothesis #2 than #1.
Also, providing more support for hypothesis #2 was the dearth of any significant effects conveying greater negative linkages between prior acute, musculoskeletal, and other injuries, and prior STD episodes, on one hand, and health and productivity outcomes on the other. Prior acute injuries did significantly predict more diseases and greater presenteeism (nonphysical) limitations in the postperiod, but among <45-year olds and not among the two older age groups. Thus, the prior safety profile of the two older age groups did not exert the kind of greater (negative) impact on these two areas of outcomes that one would expect to see if past deleterious effects of overtime had been operative.
Where prior safety health profiles were in fact operative in registering notable age differences was on the three injury and STD episode outcome measures (Table 7 and Appendix B). The two older age groups posted significant linkages between more prior STD episodes and more STD episode outcomes. They also posted links between more prior acute and musculoskeletal injuries on one hand, and more musculoskeletal injury outcomes on the other. On both these dependent measures, the corresponding estimates for the <45 age group were not significant, although the corresponding constraint comparisons were not themselves significant with one exception: Fewer prior “other” injuries predicted musculoskeletal injuries among the <45-year olds than >49-year olds.
These findings were indicative of greater support for hypothesis #1, but they were counterbalanced by other findings. Prior musculoskeletal and other work injuries, and prior STD episodes showed much stronger positive associations with acute injury outcomes among the <45-year old group than among the >49-year old group (Table 7 and Appendix B). Likewise, prior acute injuries were significantly predictive of more musculoskeletal injuries among the two younger age groups but not among the >49-year olds, whereas prior musculoskeletal injuries significantly predicted STD episode outcomes in the <45 group but not the older groups. As such, even in the safety area, the overall record was mixed, with younger employees reporting stronger possibly related overtime effects on certain measures whereas older employees reported stronger possibly related overtime effects on other measures.
The above results provide some evidence in support of the hypothesis that older employees sustain greater rates of adverse events due to decline in health and functioning. This evidence was manifest in the several instances in which the >49-year old group—relative to 45 to 49 year olds and particularly the <45-year old group—recorded greater levels of association between overtime, or factors that could theoretically be linked to past overtime on one hand, and study outcomes.
Nevertheless, the frequency of linkages tying greater rates of adverse events to 45- to 49-year olds, and particularly to <44-year olds, than the 50+ age groups was striking. What we did not clearly obtain was the concentration of effects linking overtime or plausibly related overtime factors to adverse outcomes that would be needed to fully confirm hypothesis #1. On balance, although the evidence supported both hypotheses, the evidence in both cases was circumscribed to certain areas of the model.
Age X Antecedent Interactions
Did the impact of age vary across the levels of any of these personal/job characteristics enough to merit special treatment in the group comparisons? Table 8 addresses this question, with the main effect of age being compared with the main effects of gender, compensation type, exempt status, job type, company tenure, and the interactions of age with each of these characteristics. As shown, the one interaction to emerge significant (P < 0.05) in this analysis was age by compensation type. All other interaction terms proved not significant.
This result attested to the interaction of age with compensation type—the effect of age clearly differed depending on whether the employee had hourly or salaried pay status. It confirmed the need to incorporate tests that embraced this interaction in the following group comparisons. This was achieved by running separate group models of the three-step causal structure for hourly and salaried employees.
Three Age Group/Hourly Model (M3H_1)
Tables 9 and 10 report the parameter estimate results for this model. Table 9 gives the prediction results for the three workhour measures whereas Table 10 gives the prediction results of the nine health, productivity, and safety outcomes. The format follows that of Tables 6 and 7 with two exceptions. First, the parameter estimates reported for each dependent variable include two categories of measures: 1) the three workhour measures regardless of the significance of the estimate (for Table 10); and 2) those parameter estimates from the remaining set of predictors which either yielded a significant (P ≤ 0.05) estimate for one of the groups or which yielded a significant (P ≤ 0.05) increment in the multivariate constraint tests (for Tables 9 and 10). Those predictors that did not yield estimates meeting either of these criteria are not reported.
Second, exempt status was dropped from the predictive model at the outset of the analyses for this section because of company policy which precludes hourly employees from having exempt status. As a consequence, this measure had no chance to appear as a predictor in either Table 9 or 10.
Among hourly employees, 60+ workhour weeks were associated with greater presenteeism (nonphysical) limitations among >49-year old employees (Table 10) than <45-year olds on both the estimate significance and constraint test criteria. Sixty+ hours was also linked to worse physical health outcomes in the elderly group vis-à-vis the youngest group. The >49 group also recorded significant rates of acute injuries and “other” work injuries, whereas the <45 group did not. Although the corresponding constraint tests for these latter two effects did not yield significant increments, all of these effects were all consistent with hypothesis #1.
At moderate levels of overtime, the evidence in support of hypothesis #1 was unequivocal in two instances: The youngest age group (<45) posted significantly lower rates of acute injuries than the two older groups (Table 10). The middle age group (45- to 49-year olds) also reported a significantly lower disease count and presenteeism (nonphysical) limitations than the oldest age group (>49).
With respect to the prediction working 60+ hours, those aged >49 recorded significantly more favorable prior mental health and fewer prior other work injuries than those age <45 (Table 9). Both effects, if anything, however, suggested more support for hypothesis #2 than #1. On the other hand, among those working 48.01 to 59.99 hours, those aged <45 recorded a significantly negative effect for prior STD episodes whereas the older two age groups did not, and those aged 45 to 49 reported a significant effect for prior other injuries, whereas the other two age groups did not—both effects favoring hypothesis #1. Still, above and beyond these latter two effects, none of the predictors registering significant prediction of the two levels of overtime did so in a way that supported hypothesis #1.
Moving to the “possibly a function of overtime” sector, among those working <40 hours, none of the study outcomes registered significant differences for or between any of the three age groups (Table 10). As for predicting <40 hours, both the <45 and the 45 to 49 age groups posted fewer prior “other” work injuries than the >49 group via the constraint tests—findings that favored hypothesis #1. But, these differences were not accompanied by significant group estimates per se, which meant that only one of two criteria for conveying a meaningful difference was met in this instance.
The two older age groups posted significant linkages tying more prior STD episodes with more STD episode outcomes and more diseases in the postperiod, whereas the youngest age group did not (Table 10). The <45 versus >49 group difference was significant. The >49 age group posted linkages tying more prior musculoskeletal and other work injuries to musculoskeletal injury outcomes, and more prior acute injuries to other injury outcomes, whereas the younger two age groups did not. All of these effects added evidence in support of hypothesis #1.
But, even here, this evidence was again circumscribed. The link tying more prior acute injuries to poorer physical health outcomes was significant for the <45 age group, but not the other two age groups. The <45 age group also posted significant linkages tying more acute injuries to greater presenteeism (nonphysical) limitations and more prior “other” injuries and STD episodes to more acute injury outcomes. The two older age groups posted neither linkage (Table 14). And, an effect linking fewer prior acute injuries with the capacity to bend and twist obtained for the >49 age group, but not the two younger age groups. These latter results were more in line with hypothesis #2.
Hourly employees, considered alone, generated more support for hypothesis #1 than the aggregate sample. This support was most pronounced with respect to extended overtime, although it also extended to moderate overtime and the “possibly a function of overtime” sector. Yet, it was still far from uniform. For example, the results for the 60+ workhours measure as predictor failed to support hypothesis #1 on five of nine study outcomes, as was the case for the results for moderate overtime as predictor for six of nine study outcomes. More generally, evidence for hypothesis #2 obtained on several measures in this model.
Three Age Group/Salaried Model (M3S_1)
Tables 11 and 12 report the parameter estimate results for this model and follow the format used immediately above for the three age group/hourly model, with two differences. First, exempt status appears as a predictor in these tables because a) company policy compensates some segments of the salaried workforce on an exempt basis (46% of salaried employees in the sample had exempt status), thereby enabling this measure to be included in the analysis, and b) it met one or both of the two criteria for being included as a significant predictor in the reported estimates.
Second, several measures contained no instances of occurrences during the study period: for the 45 to 49 age group: acute, musculoskeletal, and “other” work injuries, and prior acute and “other” injuries; for the >49 age group: acute and “other” work injuries, and prior “other” injuries. All such measures were dropped from the analyses and not reported in what would otherwise have been their appropriate places in the tables.
Across all study outcomes for which comparisons on age could be made (eight from the list of nine used in the study), the effects of moderate and extended overtime showed no significant differences between any of the age groups (Table 12). Moderate overtime was significantly tied to more musculoskeletal injuries for the >49 group, whereas the corresponding effect for the <45 group was not significant. But the corresponding constraint test did not produce a significant increment, meaning that only one of two criteria for a meaningful difference was met. Aside from this one instance, hypothesis #1 received no support when the effects of the two levels of overtime were compared and contrasted across the study outcomes.
Having exempt status was significantly tied to working 60+ hours for all three age groups, none however significantly more so than any other (Table 11). In fact, of all the potential predictors of extended overtime, only one cast a significant effect for any of the three age groups: More prior disease was linked to 60+ hours among <45-year olds. But again, the corresponding constraint tests between the youngest age group and the two older age groups on this measure were not significant. None of these effects were suggestive of a differential level of support for hypothesis #1. Although they did not affirm hypothesis #2, the larger message is that there was no evidence to support the notion that working extended overtime was somehow facilitated by prior differences in health.
With respect to the prediction of moderate overtime (Table 11), having exempt status was again a significant predictor for all three age groups, in fact, exerting the greatest impact of all predictors in the model, although none significantly more so for any one age group. Being overweight was significantly more linked to working 48.01 to 59.99 hours among the >49 age group than the 45 to 59 age group, but here the effect was U-shaped because the <45 age group also recorded a greater linkage than the middle age group.
There was also some evidence of prior health differences between age groups at this level of overtime (Table 11): The <45 age group reported more prior diseases than the >49 age group (via the constraint criteria only), but also recorded a significant linkage with better prior mental health (via the significant estimate criterion but not the constraint criterion). Here, the first effect could be seen as support for hypothesis #2, whereas the second effect could be seen as support of hypothesis #1, but in both cases the support was equivocal.
Turning to the “possible effects of overtime” sector, working less than 40 hours a week showed no differences as a predictor across the age groups on any of the study outcomes tested (Table 12). On the other hand, prior STD episodes significantly predicted <40 workhours, but this effect surfaced among the youngest age group (<45), not either of the older age groups. Moreover, the prior physical health of 45 to 59 years olds working <40 hours was significantly more positive than the <45 age group. Both of these latter effects were suggestive of support for hypothesis #2, but only equivocally so.
As for the prediction exerted by prior work injuries and prior STD episodes on the study outcomes, the most striking finding was the general absence of any consistent pattern of differences between age groups (Table 12). Having fewer acute injuries was significantly linked with better physical health outcomes among the >49 age group, but not so for the <45 age group. Although the corresponding constraint test did not produce a significant increment, this result could be seen as supporting hypothesis #1. But, conversely, having fewer STD episodes was linked with better mental health outcomes among the <45 age group, but not for either of the two older age groups. Although the corresponding constraint tests were not significant in either of these cases as well, these findings could nonetheless be more suggestive of support for hypothesis #2. But these latter effects were the exception: The larger pattern was one of no differences between the age groups.
The contrast of the results of the salaried-only age group comparisons versus those for the hourly only age group comparisons was striking. Whereas the evidence linked overtime to increased rates of adverse consequences for hourly employees—albeit confined primarily to the 60+ workhour level and just a subset (not all) of the study outcomes—there were no comparable trends remotely paralleling these results among salaried employees. Even with controls for exempt status, length of company tenure, job type, and even prior health, no notable adverse consequences accompanied overtime for the salaried portion of our analytic sample.
Nested Model Comparisons
The overall goodness of fit χ2 statistics and model difference tests provided a summary picture of the above three sets of model results. The top panel of Table 13 reports the goodness of fit information for the fully “freed” and “constrained” variants for M3_1, M3H_1, and M3S_1. The bottom panel records the χ2 difference tests when the nested constrained variants were compared with their respective “freed” model. These tests recorded the significance of the decrement in overall goodness of fit that obtained when the parameters reflecting the entire causal structure of pairs of specified age groups were constrained to be equal. Significant χ2 statistics in this regard indicated that the two groups of parameters were better left freed because of the gains of information that resulted. In contrast, χ2 that was not significant indicated that the parameters could be constrained to be equal with no real loss in information explained.
Because these were large models conducted on relatively large sample sizes, the conventional χ2 statistic is supplemented with additional goodness of fit indices that were less affected by such characteristics. These latter indices provided measures of the extent of the change in fit that accrued with the tests.
As shown, the three age groups all differed significantly from one another in overall model results at the aggregate level. But, closer inspection reveals that these differences by age were much more pronounced among hourly employees pronounced than among salaried employees. Although the <45 and 45 to 49 salaried groups recorded a significant overall χ2 result, the differences between the oldest age group and the two younger age groups were modestly significant in one case (<45 vs >49) and not at all significant in the other (45 to 49 vs >49). These results speak forcefully to the absence of overriding age-related effects involving older, salaried employees and provide a compelling counter to the idea that adverse rates increase with advancing age because of some universal age-related declines in health/functioning in this portion of the workforce.
The published literature on overtime has yet to produce a study designed to see if patterns between workhours on one hand and health, safety, and productivity outcomes on the other hand, differ for younger versus older employees. This dearth of research stems from a precedent, established early on in this research, that in effect abstained from systematic study of group differences in the impact of long work hours. Although it would seem a straightforward proposition, the idea that employees with differing characteristics doing differing types of tasks in differing work settings might be more or less prone to adverse health, productivity or safety consequences has yet to be examined in a systematic way.
In the context of what amounts to an empirical vacuum, public policy has developed in this area with far-reaching implications. No less august a body than the European Union has issued guidelines for workers in its member countries that place constraints on workhours per week.16 These guidelines are across the board—they do not differentiate by industry type, by employee/job characteristic, or any other factor along these lines.
Although these guidelines do allow for exceptions in implementation, we have argued elsewhere that they appear to be overly conservative in today's increasingly competitive marketplace.2–4 If, indeed, it turns out that only employees with certain characteristics doing certain types of work in certain work environments are at elevated risk when working certain levels of longer hours, then it would appear that, at a minimum, a much more carefully calibrated strategy is in order. The prospects for this strategy will be maximized if it is predicated on a relatively precise grasp of when and under what circumstances overtime becomes detrimental to employee health, safety, and productivity.
This study takes up this theme by asking, do rates of adverse health, productivity, and safety consequences—and the repercussions they have for differing indirect costs—vary by age? What do data say really happens as employees get older? Is there the kind of profound and consistent shift toward greater rates of adverse outcomes—and indirect costs—with advancing age that would justify concerns about employability? Or are younger employees just as likely or more likely to incur outcomes that reflect impairment—deteriorating health, limitations in on-the-job performance, rates of injuries, and the like—in ways that would argue against concerns about advancing age and employability?
Our results exhibited support for the proposition that increasing age leads to greater rates of adverse consequences (hypothesis #1). But, these increases were largely confined to hourly employees working extended overtime (60+ hours) and occurred on only four of the nine study outcomes examined in this study. Although there were occasional instances of increased age effects at the level of moderate overtime (48.01 to 59.99 hours) and some support among effects reflecting the possible adverse effects of past overtime, the preponderance of the evidence in these latter spheres was much more in line with either the rejection of hypothesis #1 or the nonrejection—acceptance—of hypothesis #2. The findings supporting the proposition that adverse heath, productivity, and safety outcomes are the same across age groups—or actually decrease with increasing age—occurred in every series of models tested.
Perhaps the most compelling test result pitting the two hypotheses concerned the summary χ2 difference tests. If the “increasing adverse outcomes with advancing age” hypothesis were to have held—certainly true to the spirit if not the letter of policy initiatives such as the European Union's Working Time Directive—age group differences would at the very least needed to have been observed in these tests regardless of extraneous characteristics. Yet, this was not the case. Among salaried employees, the older group manifested nonmeaningful differences with the two younger groups—that is, they for all intents and purposes proved statistically the same across the collective space of all workhour/adverse outcome effects analyzed in this study—a strong affirmation of hypothesis #2.
Thus, older employees may have displayed markedly poorer physical health outcomes, more diseases, more limitation on all three measures of presenteeism, and more acute injuries and STD episodes when pitted directly (without statistical adjustment) against younger employees. But, these differences did not carry through—at least in any wholesale fashion—to the impact of overtime. That is, most of the older employees working overtime were able to do so without registering the greater proclivities toward adverse consequences that they accrued by virtue of advancing age alone. When working longer hours, they in effect would appear to have made the adjustments that enabled these proclivities to be neutralized and, indeed, to have done so in manner that resulted in less productivity loss—and by implication, reduced indirect costs—on certain outcomes. It took narrowing the lens to a certain type of employee (ie, hourly) working certain levels of long hours (ie, extended overtime) before a notable portion of these age-based proclivities obtained. These proclivities were largely if not entirely muted with respect to a whole other type of employee (ie, salaried) and those working at more moderate levels of overtime.
The results herein, as such, lead to the conclusion that, at the very least, a differentiated approach to age with respect to working long hours is warranted. More research is needed to develop specific recommendations for such an approach. This work should proceed, however, in view of the fact that the current study was not without its own limitations. These limitations included design flaws such as the retrospective, secondary character of this analysis and the shortness of the length of the exposure period. A prospective study design with primary data collection centered on the impact of workhours as the primary study issue would improve the capacity for causal inference. As for exposure to workhour impact, the exposure period may not have been long enough to give adverse consequences a chance to register. Lengthening the observation period, and systematically testing for critical thresholds of time for the long workhour effect to develop, would strengthen confidence in the study results.
More generally, the survey for this study focused only on the assessment of average hours worked in a typical week. Yet, work schedules at the sponsor company differed in ways other than simply the extent of overtime, including use of double shifts extended hour shifts (eg, 10 to 12 hour shifts vs 8-hour shifts) and blends of extended hour/normal 8-hour shifts). Numerous studies have reported associations between alternative variations in work schedules and adverse health consequences.17 Future work should broaden the measurement of work schedules to test the effects of such variations.
Fourth, the need for more systematic study of potential covariates frequently called for in past studies was applicable here as well. By incorporating and controlling other relevant measures, future studies would yield a richer context for causal inference. A fertile focus in this regard would appear to be the interplay among job demands, autonomy, and social support, whose moderating impact on the effects of workhours has been documented elsewhere.18 Also, this study's findings would gain further weight if replicated on other employee populations linked other industries in other countries.
A final issue has to do with this study's treatment of indirect costs. This treatment reported estimates of nonfinancial indirect costs and explored their implications. It did not try to generate financial cost estimates, primarily because of limitations in the data. Specifically, although the prevalence and magnitude of limitations at work could be assessed, measures of lost compensation were not available. Similarly, although the probability of adverse events such as workplace injuries and disability episodes could be estimated, measures of the length of lost or restricted time and the associated lost compensation were not available. A published finding cited at the outset of this article—that older employees record fewer disability episodes but once on disability require more time before they return to work—underscores the need for future studies to include such empirical distinctions if the intent is to be more definitive about financial indirect costs.
These limitations notwithstanding, this study provides a new empirical map for guiding employers and policymakers on issues concerning age, overtime, and health, productivity, and safety outcomes. Most importantly, the data suggest older workers can be retained in positions involving overtime and not pose an increased risk of productivity loss and, quite likely indirect costs, but the scope of their activities and roles should be informed by empirical research. Going forward, initiatives that are shaped by studies that address the above limitations will be well-positioned to advance organizational competitiveness and profitability in ways that genuinely address the concerns of labor, management, and government on the subjects of age, quality of life, and productivity.
This study was funded by an unrestricted educational grant provided by The Conference Board, New York, NY. The authors are solely responsible for all findings and conclusions presented in this article.