The OHBWC is the largest of four exclusive state-run WC systems in the US and provides WC insurance for approximately two-thirds of Ohio workers. For the last 15 years, the OHBWC has been supporting several large-scale OSH intervention programs and services designed to improve primary through tertiary prevention at insured employers. For example, in 1999 OHBWC initiated the Safety Intervention Grants Program (SIG) to provide matching funds to insured employers to implement OSH engineering controls and measure effectiveness [Ohio Administrative Code, 2006]. From 1999 to 2009, grants ranged from a 2:1 to 4:1 OHBWC to employer match in funds, with up to $40,000 per grant. Over the history of the program, OHBWC has provided funds for over 1,800 interventions with an average annual total of $3 million in matching funds. As part of the SIG program, OHBWC consultants work closely with employers to develop proposed interventions, submit applications, and verify that they have been implemented fully in the workplace. As a part of the application process, an OHBWC consultant also conducts an onsite comprehensive safety assessment and five-year claims history review, and may also recommend steps to reduce identified hazards.

Beginning in 2003, OHBWC required that participating SIG employers identify WC claims and person-hours that were associated with the employees who would be directly affected by the implementation of the intervention for a two year baseline period pre-intervention. Employers were asked to report on the total person-hours worked by this population, regardless of what task they were performing. Post-intervention, the employer reported quarterly to OHBWC the WC claims and person-hours for employees affected by the intervention for a two year follow-up period. Employers used the date of implementation of the control equipment as the beginning of the post-intervention reporting period. The time gap between the end of the pre-intervention period and the beginning of the post-intervention data reporting periods varied by employer. From 2003 to 2009, the median gap was 4.5 months, and 95% of the gaps were under one year. The employer also reported their yearly average total number of employees (both part- and full-time) as part of the Ohio Quarterly Census of Employment and Wages (QCEW). To participate in the SIG program during the majority of the examined 2003–2009 implementation period, employers were required to have at least one WC claim pre-intervention that could have been prevented by the proposed intervention. Employers after July 2009 were not required to have a pre-intervention claim. Approximately 83.5% of submitted SIG applications were accepted for funding from 2003–2009.

In early 2012, NIOSH acquired data from OHBWC for a census of all 468 employers who were approved for grant support in the period 2003–2009. All employers had implemented the interventions and had completed the necessary follow-up documentation. NIOSH excluded any claims that were disallowed or dismissed by OHBWC (109 claims, 3.8%) and compiled the following 2001–2011 pre-and post-intervention WC metrics for all participant employers for the employees affected by the intervention:

QCEW data were used to create two employer size groups (small, 1–99 employees, large, ≥100 employees) using the yearly average number of part- and full-time employees reported by the employer for the two-year pre-intervention period. These categories were chosen because they produced a fairly equivalent number of employers in each group (small n =242; large n =204). Specific industries were defined by two to three digit North American Industry Classification System (NAICS) codes for employers as reported in the QCEW. NAICS codes were also grouped in NIOSH industry sectors for some analyses.

Intervention categories were consensus-coded by two certified professional ergonomists prior to analysis and without reference to the WC outcomes for each intervention. The coders based decisions on summary narrative information about each intervention (including keywords such as “lift table” and equipment brand names) as well as detailed case study information about the intended prevention purpose of the intervention. Intervention type categories were based partially on modified 2–3 digit source codes from the Occupational Injury and Illness Classification Manual Version 2.01 [Bureau of Labor Statistics, 2012a]. Additional detailed information on the intervention coding structure is provided in the Supplemental Online Material.

Specific interventions were classified into four categories. First, the ergonomic category included interventions (material handling equipment, patient handling equipment, and other ergonomic equipment and tools) intended to reduce risk factors (overexertion, awkward postures, excessive force, contact stress, and vibration) for work-related musculoskeletal disorders [NIOSH, 1997]. Second, the safety category included interventions (scaffolding, slip resistant flooring, and other equipment including specialty saws and machine guarding) intended to reduce acute accidents (caused by falls, struck by/against objects, and caught between/compressed by events). Third, the ventilation category included interventions (stand-alone ventilation systems or machinery and powered equipment with built-in ventilation) intended to reduce airborne contaminants. Finally, the multiple-purpose category included interventions intended to reduce more than one hazard (ergonomic, safety, and/or airborne contaminants).

A basic concern with evaluating WC cost trends over time was that older claims are allowed more time to develop costs than newer claims and therefore tend to have higher reported costs. To make the measurement of costs more consistent across years, claim cost data were limited to paid medical costs and paid indemnity benefits (compensation for lost wages) valued 30 months after January 1 of the calendar year in which the claim occurred. The paid 30-month costs do not represent the ultimate cost of claims. Another measure of claim costs, called factor-adjusted costs, takes into account the ultimate costs of all claims. For example, another study of a large sample of OHBWC claims data determined that the factor-adjusted costs for calendar year 2009 were approximately three times as high as paid 30-month costs [Wurzelbacher et al., 2013].

OHBWC sponsors two main programs that impact the cost of claims reported in their database. The first program allows insured employers to pay first dollar medical costs up to a specified limit for medical-only claims. Only medical paid costs in excess of this limit are reported to OHBWC. A second program allows insured employers to pay first dollar indemnity (wage replacement) costs, which are not reported to OHBWC. Employers that participated in either of these two programs were included in claim frequency analyses, but excluded from the analyses of claim costs. Claims with $0 in total paid costs (not impacted by the two programs above) were included in both the frequency and cost analyses (127 clams, 4.7%).

All outcomes were analyzed with multivariable models to account for temporal trends. For the claim frequency rate analysis, Poisson regression with repeated measures was performed. For the claim cost per FTE analysis, a two-part model was used, which is fully described in the Supplemental Online Material. This type of model is useful in analyzing data with a large proportion of zeroes. In the first part of the analysis, the probability of observing a $0 cost per FTE is estimated, and in the second, the mean cost per FTE is estimated for those employers with cost >$0. For the WC cost per claim analysis, a linear regression model with repeated measures was used. In addition, for the WC cost per claim analysis, since the cost data for this analysis was skewed, it was transformed using the natural log before modeling and reported as geometric means, which generally approximate the median of the distribution. To address taking the natural log of a $0 claim, $5 was arbitrarily added to the cost of every $0 claim. One employer with an extreme outlier claim (that was an order of magnitude larger than the next highest claim) was excluded from both cost analyses. All models controlled for changes in rates over time that may have occurred independently of the interventions. This method measures pre–post changes relative to a time trend that is based on data from program participants rather than a separate comparison group, thus avoiding bias caused by differences in time-invariant characteristics between employers who applied for and received safety grants and those who did not. Differences in employer characteristics that vary over time would also not bias the results, unless the variation is correlated with the intervention, as might be the case if other, unrelated safety efforts tended to be undertaken at the same time as the grant-supported intervention. All analyses were conducted using two models to control for the time trend. The 1st model assumed constant/linear cost trends over time but included a factor that accounted for the random effects of each employer’s paired pre/post data. The 2nd model did not restrict costs to a yearly trend, and instead used dummy variables for each year to allow background changes in costs to vary independently by year. We found only slight differences between the two different model results and report only the 1st model here since it allows us to better summarize the independent time effect on outcomes.

Analyses were conducted initially with all available data and then stratified by employer size. Analyses were then conducted for specific industries, and interventions. Since industry and intervention type were strongly correlated with year, due to the changing emphases of the SIG program, these analyses were conducted without controlling for changes in rates over time that may have occurred independently of the interventions. Additional information on the regression modelling approach is provided in the Supplemental Online Material.

Five post-hoc analyses were conducted to evaluate the possible influence of regression-to-the mean (RTM) effects, employer under-reporting of WC claims for affected employees, and limits to generalizability. To investigate possible RTM effects, a series of post-hoc analyses were conducted [Barnett et al., 2004]. We first compared the claim frequency rate results of high- and low-rate groups (low-rate was defined to be ≤ the median rate) using the same regression model as in the main analysis. The second post-hoc analysis examined the time path of claim rates, cost per FTE, and geometric mean cost per claim at a finer level of detail—annually in the pre-intervention period, and quarterly in the post-intervention period. The objective was to determine whether there was evidence of an unusual spike in rates just prior to intervention, and whether this was immediately followed by a similar decline that could be explained by random variation in rates over time rather than intervention. In this second analysis, since employers reported affected employee data in 2-year sums for the pre-intervention period, an assumption was made that the affected employee hours were divided equally between the two years. Post-intervention claims data were reported by the employer quarterly. In this analysis, all regressions again controlled for repeated measures and changes over time that may have occurred independently of the interventions.

Third, we performed a similar analysis as just described, except examining the entire employer. This allowed us to examine two more years before the 2-year pre-intervention period, as well as two years after the 2-year post-intervention period, to evaluate whether there was a trend upward in the rates and costs of the entire employer before the intervention, that could explain selection into the program. Employer-level data is reported in calendar years, and yearly rates were calculated in relationship to entrance into the program. The first year before entrance into the program (labeled “−1”) was represented by the last calendar year with at least nine months before the application date, and the first year after entrance into the program (labeled “+1”) was represented by the first calendar year with at least nine months after the implementation of the intervention. For many employers, this left one calendar year that represented some mixture of pre and/or post periods and the intervening time between application and implementation. These calendar years were labeled “0”.

To evaluate the possibility that employers may have under-reported WC claims for affected employees post-intervention, a fourth post-hoc analysis was conducted to compare the unaffected employee WC claim frequency rates pre- and post-intervention. One way that employers could have under-reported claims for affected employees is by assigning their claims to the unaffected group. We hypothesized that the effect would be most pronounced for smaller employers where a few misattributed claims could elevate claim rates among unaffected employees to a greater degree. To test this hypothesis, the total number of employees reported by the employer in the QCEW for the pre- and post-intervention periods was multiplied by 2,000 to estimate total annual employee hours, and then the reported affected employee hours were subtracted to estimate unaffected employee hours. The 2,000 hr assumption likely results, on average, in an underestimate of rates, since it does not take part-time work into account. However, this does not bias the estimate of percent change in rates. Poisson regression with repeated measures analysis was again performed to control for changes in rates over time that may have occurred independently of the interventions. Estimation of unaffected employee claim frequency rates was also used to compare claim rates of affected and unaffected employees in the pre-intervention period. This may be useful for understanding how the affected employee group was selected for the intervention.

To evaluate the possibility that participant employers may have had higher or lower injury rates and costs than non-participants (which would limit generalizability), and to examine how any differences from non-participants may have changed over time, we conducted a fifth post-hoc analysis by analyzing the experience modification ratings (EM) of the participant employers. The EM is a traditional insurance metric that compares the actual past WC losses of the employer versus the expected losses for that employer’s specific industry [as defined by National Council on Compensation Insurance (NCCI) class code] in the state of Ohio.

The OHBWC EM is different from most EM formulas used in other states in that much more weight is placed on the cost of WC claims as opposed to their frequency [OHBWC, 2012]. EMs are revised each year on July1, and are based on data for a 4-year “experience” period ending 1.5 years earlier. The EMs were classified into pre- and post-intervention years relative to entrance into the program as described above, based on the last year of the experience period. A series of median, annual EMs was constructed to reflect the relative level and trend in the overall claims experience of grant recipient employers, prior to and following intervention. A second series of EMs was also constructed that was based, where applicable, on EMs that were assigned to employers based on their membership in group plans. While each employer has their own individual EM, some employers have another EM that is assigned to all employers in their group, and this is the one that is used to compute premium payments and may be more influential in employer decision making. Approximately 46% of OHBWC-insured employers participate in group rating or other programs that discount EMs.

All analyses were conducted using SAS^® version 9.3 (SAS Institute, Inc., Cary, NC). This research was approved by the NIOSH Institutional Review Board. The requirement for informed consent was waived because the study involved the analysis of coded and previously collected WC and program evaluation data.

Participant employer sizes ranged widely (1–24 employees, n =59; 25–49 employees, n =73; 50–74 employees, n =65; 75–99 employees, n =46; 100–149 employees, n = 61; 150–199 employees, n =40; 200–249 employees, n = 24; 250 +employees, n =81).

The number of affected employees at each employer also ranged widely (mean =49, SD =188, median =18). The ratio of affected to total employees differed by employer size. In small employers (1–99 employees), the median percentage of affected employees was 35% of the total number of employees. In large employers (100+ employees), the median percentage of affected employees was 15% of the total number of employees.

Table I provides summary statistics on the affected employee WC claim outcomes for the 468 employers from 2001–2011. For every year of intervention, there was a decrease in the rate between the pre- and post-intervention periods. There was also a general tendency for rates to decrease over time independent of intervention, as can be seen by the change over time in pre-intervention rates.

Of the intervention studies [Marras et al., 2000; Owen et al., 2002; Fujishiro et al., 2005; Alamgir et al., 2008; Park et al., 2009] that have investigated the impact of engineering controls on injury rates, few [such as Park et al., 2009 and Fujishiro et al., 2005] directly controlled for the injury trends over time that occurred independently of the interventions. Recent Bureau of Labor Statistics (BLS), NCCI, Washington State, and OHBWC data have demonstrated that overall Occupational Safety and Health Administration (OSHA) recordable injury/illness and WC claim numbers and rates are declining significantly in most industries. For example, a study by NCCI [NCCI, 2010] found 4% average annual declines in the rates of lost-time WC case rates, for a total decline of 57% from 1990 through 2009 while Washington State found declines of 5.9% per year for all claims [Silverstein, 2007]. From 2000 to 2009 the number of claims reported in the OHBWC claims system also dropped by 47.6% [Tarawneh et al., 2013].

There is no clear consensus on why these rates and numbers are declining, but some researchers have suggested this could be due to under-reporting [Morse et al., 2005], economic reasons [NCCI, 2012], or “advances in automation, technology, and production” and “emphasis on workplace safety and loss control” [NCCI, 2010]. Regardless of the cause, it is important in effectiveness research to control for the WC change over time that may occur independently of the evaluated intervention. This study was able to show that over and above the general WC claim decline, total claim rates (medical-only and lost-time claims) declined an additional 66% in the affected employee groups post-intervention. This study also examined the impact of engineering controls in a wide range of employer sizes and provides key evidence that targeted engineering interventions are an effective tactic for reducing WC claims among both smaller (1–99 employees) and larger employers.

Most prior intervention studies have demonstrated intervention effectiveness in a specific industry such as healthcare [Owen et al., 2002; Collins et al., 2004; Fujishiro et al., 2005; Park et al., 2009]. The current study evaluated effectiveness in a wide range of industries (10 major industry sectors). As a whole, this study provides evidence that engineering interventions were implemented successfully in wide variety of applications and industrial settings.

Many of the intervention studies that have been conducted in the past focused on a single type or narrow range of controls such as ergonomic patient handling devices [Owen et al., 2002; Collins et al., 2004; Fujishiro et al., 2005; Park et al., 2009]. The current study demonstrated significant reductions in WC outcomes for a wide range of ergonomic, safety, and multiple-purpose interventions across many industries. For example, ergonomic interventions were utilized most in the Manufacturing and Services (except Public Safety) sectors but were implemented in almost every NIOSH industry sector. The significant reductions for ergonomic and safety equipment were expected, given that these types of interventions target the most common occupational exposure events (such as overexertion, repetitive motion, bodily reaction, slip trip falls, struck by/against, caught in/compressed by), which together account for the majority of injuries/illnesses [BLS, 2012b] and WC costs [Liberty Mutual, 2013]. This provides evidence that many types of specific engineering controls are effective in reducing WC outcomes for affected employees, and that the same intervention type is effective in multiple industrial settings.

By contrast, ventilation controls were associated with a significant post-intervention increase in WC paid cost per affected FTE and WC paid cost per affected employee claim post intervention. This is understandable, given that there are few WC claims related to air contaminant exposures. This finding does not establish that the ventilation controls were not effective in reducing contaminants or improving worker health outcomes. Rather the 2-year post-intervention WC outcomes were not sensitive enough to capture the effectiveness of the ventilation controls and/or the time period selected was too short to observe any measurable change in occupational disease incidence. The negative finding for ventilation controls actually provides support for the integrity of the study since reductions in WC outcomes would not necessarily be expected for this type of workplace change.

The range of effectiveness data by specific industry and intervention (see Tables III and IV and Table SI–Table V) provide useful information for employers to consider when conducting analyses prior to implementing specific engineering control projects. Several cost-benefit calculators have been developed recently [Goggins et al., 2008; IWH, 2009], but a missing parameter is often the expected range of effectiveness for a given industry and intervention. As a reminder, the analyses for specific industries and interventions did not control for the time trends independent of the intervention. The very large effect sizes may have also been influenced in part by RTM effects and other study limitations discussed below. Therefore, for practitioners, it may be most appropriate to focus on the lower bound of the confidence interval as the expected effectiveness of a given engineering control for planning purposes.

The effectiveness of the entire SIG program also provides useful information for insurers who may consider implementing similar matching grant programs. Funding support by insurers for engineering controls is rare. Six state insurance bureaus support some form of matching fund programs, but only four provide funds that can be used for engineering controls (WA, NY, ND, and OH). The others provide matching funds for training programs (MA, OR). No known commercial insurers offer similar programs to support engineering controls. For exclusive state-funds (OH, WA, ND, and WY), the business model is such that the insurer will insure the employer as long as they are in business or the employer does not self-insure. Thus, primary loss prevention may tend to play a greater role in risk management for the exclusive insurer. This study indicates that supported engineering programs such as the SIG should continue to be a focus for such exclusive insurers. For example, OHBWC recently expanded annual funding for the SIG from approximately $4–12 million based in part on the apparent effectiveness of the program.

For US commercial insurers, employers are generally bound to their insurer only by annual policies that must be renewed. Although primary loss prevention is still a focus, most commercial insurers also emphasize loss reduction to control claim costs, risk selection to choose which employers to insure, and experience rating systems to set appropriate premiums and derive profit. Although supported engineering control programs may be less attractive initially to commercial insurers, this study did demonstrate that the claims reductions occurred immediately after implementation of the intervention. This suggests that the return-on-investment may still be attractive enough on a group if not individual employer level from the commercial insurer perspective. It may also be possible for commercial insurers to develop special contracts such that, if an employer chooses to insure with another insurer after one year, that the insured employer must pay back matching funds. Alternative arrangements, such as providing premium discounts for demonstrated engineering control implementation at the insured employer, may also be possible for commercial insurers. Such arrangements would likely require the approval of the WC bureau in the US state where the insured employer is located. State WC bureaus could also possibly provide incentives to encourage insured employers and insurers to implement similar supported engineering control programs.

There are several limitations to the analyses presented in this paper. First, this study used a pre–post design that lacks randomization and a true control group. The drawbacks of this design are mitigated by the fact that interventions occurred over a wide range of dates, such that there is no common, independent change at a point in time that coincides with intervention. In addition, the baseline time trend against which claim rates and costs are measured, is calculated with program participant data rather than with data on non-participants who may have been different. A remaining concern is that the timing of intervention may have coincided with other changes for the affected employee group. The lack of any dramatic changes in claim rate among unaffected employees suggests coincident changes were not great at the overall employer level.

A second main limitation is that employers may have applied for grants at a time when claim rates and costs were high for the affected employee group, due in part to a random upward fluctuation in rates. If so, rates would have tended to fall in the post-intervention period due to a RTM effect rather than intervention effectiveness. The possibility of an RTM effect on the results could have been increased by a rule, in place for most of the 2003–2009 period, that employers in the SIG program had to have at least one WC claim during the pre-intervention period that could have been prevented by the proposed intervention. This possibility was mitigated by assessing WC outcomes for 2-year rather than 1-year periods, pre- and post-intervention.

Full and accurate measurement of the potential contribution of RTM to the observed pre–post decline requires determination of the difference between the normal or expected rate for the affected employees and their actual rate in the pre-intervention period. Ideally, this might be done by examining their trend in rate over many years prior to intervention, or by examining the long-term trend in rates of a closely comparable set of employees. Data of this kind is unavailable in the current study, but some of the post-hoc analyses provide useful evidence.

In the first post-hoc analysis (Figure S1 and Table SV), pre–post-intervention WC claim frequency rate changes were observed to be greater for employers with higher pre-intervention rates. However, this would be expected even in the absence of a RTM impact on the results of the study. Higher pre-intervention rates are always associated in some degree with a greater tendency for subsequent decline, due to the role of random variation in their higher rates. The question is whether the study population contains a balanced representation of groups with random increases and decreases in rate prior to intervention.

The second post-hoc analysis examined the affected employee WC claim frequency rate and cost data divided into yearly and quarterly rates (Table Va, b). This time path of claim rates and costs does not rule out RTM, but neither does it support an RTM explanation of the results. First, there was no notable increase in rates or costs in the year before intervention over the previous year. However, employers may instead have taken a longer view, considering both pre-intervention years, or a longer span of years, and all of these years together may have seen a random increase above the central tendency. We do not have the data to determine whether this was the case. Yet, even if employers did take a longer view, it would still have been logical to expect employers to place the greatest weight on their claims experience in the immediately preceding year. Second, the distinct downward trend in claim rate and cost per FTE in the post-intervention period (Table Vb) is not compatible with the RTM theory, in which rates and costs would have the same expected value in each successive post-intervention period, or track general time trends in risk levels. The initial, post-intervention declines could still have been at least partly due to RTM. However, it would be most reasonable to assume that a program of apparent effectiveness in the second year was also at least somewhat effective in the first year, especially since the type of safety and ergonomic equipment funded by the program would generally be expected to have impacts in less than one year.

There is some evidence to suggest that some employers may have been motivated to participate not by increased claim rates and costs for the affected employees, but by their overall claim rates and costs. First, in the pre-intervention period, overall employee claim rates were 1.7 times higher on average than in the affected group, while the total employee cost per FTE was about the same on average as in the affected group. Second, while the evidence is somewhat mixed, the time path of overall claim rates, cost per FTE, and the EM suggests that employers may have perceived an upward trend prior to the grant. Claim rates (Figure 2) increased from year 4 to year 3 before the grant, but were declining slightly thereafter, before the grant. Cost per FTE increased substantially from year 4 to year 3 before the grant, and then again in the year before the grant, before dropping even more substantially in year 0. This is suggestive of an RTM effect on the overall employer level with respect to costs, although the post-intervention period as a whole does not suggest a downward return to trend, and the year 0 contains a mixture of pre- and/or post-intervention periods as well as the intervening time between pre-and post-intervention periods. The presence of an RTM effect on the employer level does not necessarily imply an RTM effect for the affected employee group, for which we saw a slight decline in cost per FTE in the last pre-intervention year, rather than an increase. However, some of these upward trends occurred previous to the pre-intervention period, and it is possible that these were shared by the affected employee group.

It is notable that the decline in the individual employer EMs was modest after intervention, and that group policy EMs did not decline post intervention. However, EMs are designed to reflect changes in claims experience only partially, due to the small size of the employers, and changes in group policy EMs may also be due to changes in group policy membership and rules governing group policy pricing. In addition, the safety grants would be expected to have a relatively modest observable impact on EMs, since the affected work group was 35% of the workforce in small employers and only 15% in large employers.

It may seem surprising that interventions were not targeted at an employee subgroup with especially high rates and costs. One possibility is that the highest employer risks required more extensive process changes that were not always addressed through the program. Based on OHBWC input, the choice of intervention depended not just on the magnitude of risks but on the feasibility of reducing those risks within the scope of the SIG program (which is focused on placing of commercially available engineering controls within defined employer processes with a matching budget of $40 k or less).

Another possibility is that, since costs per employee were similar, but claim rates lower among the affected employees, the relatively high cost per claim among affected employees could have made them a more visible target for intervention. This visibility could have been heightened if the higher average cost per claim was driven by a few, very expensive claims. If affected employee groups were selected partly on the basis of a few high cost claims, this could have led to an RTM effect with respect to cost per claim, and helped explain why we observed a higher percentage decline in cost per employee than in the claim rate. However, this does not appear likely, since there is little difference between claim rate decline and cost per employee decline when we focus only on more costly lost-time claims. The claim rate decline itself would not have been affected by this selection criterion.

In sum, the time trends in affected employee claim rates and costs, both pre- and post-intervention, did not exhibit patterns that would have been expected if RTM effects were the main driver of the results. In addition, the evidence indicated that affected employees were not selected for participation on the basis of higher claim rates and cost per employee than their peers in the workplace, and that, based on EMs, employers who participated had only slightly higher claim costs and rates than peers in their industry. There remains a question about whether RTM effects might nevertheless have explained part of the observed post-intervention declines, because data is not available to determine whether claim rates and costs in both years prior to intervention were random upward departures from the long term trend. This possibility is reduced, given the relatively low claim rates and similar costs among affected employees, since any random increase would have to have been from a still lower level. On the other hand, the time path of overall claim rates, costs, and EMs shows an increase before the pre-intervention period that could have been shared by the affected employee group. While the affected employee group within the employer may have been targeted in part, because of some highly visible, expensive claims, this is a possible explanation for only a small portion of the post-intervention changes in cost per claim. Thus, it is possible that an RTM effect contributed to the results, although the post-hoc analyses do not, on the whole, provide very much support for the RTM having a substantial influence on the observed decline in WC outcomes post-intervention.

Another potentially important limitation was that SIG employers self-reported WC claims associated with the affected work group pre- and post-intervention. It is possible that the employers’ expectations of intervention success and the desirability of demonstrating success could have influenced their reporting. There may have also been disincentives for reporting in some employers such as “safety bonuses” for supervisors or employees who do not report. One way for this to have occurred is that employers under-reported claims for affected employees after the intervention by assigning some of their claims to the unaffected group. This would have produced post-intervention increases in claim rates among unaffected employees, especially among smaller employers, since the affected group is a greater proportion of their employees. However, post-hoc analysis (Table VI) indicated that the rates for unaffected employees did not change significantly post-intervention, and the rates for smaller employer unaffected employees actually decreased slightly post-intervention. This analysis provides some assurance that the employers were not under-reporting claims for the affected employees.

An additional analysis issue was that the claims being reported were all claims associated with the affected employee group and not just those claims directly preventable by the intervention. This approach tends to reduce the power to detect injury reductions, but the data collection method was equally applied pre- and post-intervention, and there should have been no net bias towards reduction post-intervention. Furthermore, one advantage of having the employer report all claims for the affected work group was that the employer was not being asked to make a decision about whether the intervention could have prevented a particular claim, thus removing another possible source of bias and inconsistency in reporting.

Finally, this study focused only on employers who elected to apply for the grant program and received funding. Thus the study group was self-selected, and the findings may not be generalized to employers who do not seek to participate in programs similar to the SIG. Based on the EM analysis, employers who participated had claim costs that tended to be just slightly higher than other employers in their industry and size class (individual EM averaging 1.04 for the 4-year experience period immediately preceding intervention, see panel b in Figure 2). However, OHBWC personnel involved in the program suggest anecdotally that participating employers tend to be those who participate in other supported programs such as OHBWC consultation services, and have above-average commitment to improving safety. Thus, this study’s results may be most generalizable to employers that have demonstrated some prior safety success and interest.