THE EFFECTS OF DIFFERENT OSHA INSPECTION TYPES ON DIFFERENT MANUFACTURING ESTABLISHMENTS,

Transcription

1 THE EFFECTS OF DIFFERENT OSHA INSPECTION TYPES ON DIFFERENT MANUFACTURING ESTABLISHMENTS, JOHN M. MENDELOFF and WAYNE B.GRAY* * John Mendeloff is Professor of Public Management and Policy at the University of Pittsburgh and Wayne Gray is Professor of Economics at Clark University. We are grateful to the National Institute of Occupational Safety and Health (R01-OH ) for financial support, and to Joe DuBois and John Ruser for their help in working with the OSHA and BLS data. The research was conducted with restricted access to Bureau of Labor Statistics (BLS) data onsite at BLS. The views expressed here are those of the authors, and don t necessarily reflect the views of BLS or supporting agencies. 1

2 Abstract We test whether lost workday injuries decline in the years following an OSHA inspection and whether the inspection s effect differs based on the characteristics of the inspection or the establishment, using manufacturing sector data for We find that inspections have larger effects where other accident-prevention mechanisms are weakest. Specifically, we find larger effects for smaller establishments (subject to little or no experience rating under workers compensation) and non-union establishments (where worker involvement and voice are likely to be absent). For inspection characteristics, we find larger effects for inspections that impose penalties, and smaller effects for programmed inspections and (surprisingly) safety inspections. We also find different effects on injuries that involve days away from work, compared to those which only involve restricted activity workdays. 2

3 1. Introduction Congress established the Occupational Safety and Health Administration (OSHA) in 1970 to prevent occupational injuries and illnesses. The latest data, for 2001, indicate that workers in private industry suffered about 5.2 million recordable injuries or illnesses Of these about 1.5 million caused workers to lose days at work. Another subset of a million cases led to restrictions on work activity. (Bureau of Labor Statistics, 2002) In the same year, the estimate of fatal occupational injuries in the private sector was 5270, a figure that includes acute toxic exposures, but few cases of chronic disease, which are often harder to link to workplace exposures. (Bureau of Labor Statistics, 2003a) OSHA may affect injuries and illnesses through several possible mechanisms. The agency enforces a set of safety and health standards and may create new standards. It also provides information to employers and employees, both directly through consultations and training activities and through the provision of educational materials. Most of OSHA s resources are devoted to its enforcement program. Inspections, backed up by the threat of penalties for non-compliance, may push employers to comply with standards or even to improve their overall safety programs. The threat of inspection may also generate compliance actions in order to avoid expected penalties. Even though most workplaces are inspected infrequently, especially in industries with low injury rates, the ability of workers to request OSHA inspections enhances their general deterrent effect. A number of earlier studies of OSHA s impacts used a combined time-series, crosssection, industry level data set to examine whether the industry injury rate was affected by the number of inspections in that industry and the average penalty. (Bartel and Thomas, 1985; Viscusi, 1979a, 1986; Lanoie, 1992) These studies found either that there were no effects or that 1

4 the probability of inspection had a small effect on the industry rate. Three studies used BLS Survey data to examine whether establishments which were inspected early during a calendar year experienced injury rate reductions compared to those inspected late in the year, finding mixed results, with some evidence of significant effects for smaller workplaces. (Smith, 1979; McCaffery, 1983; Ruser and Smith, 1991). Smith (1992) provides a more detailed review of these early studies. In the early 1990s, Scholz and Gray published a series of studies (Scholz and Gray, 1990, 1997; Gray and Scholz, 1991, 1993) which incorporated several improvements over earlier efforts: 1) Instead of aggregated industry-level data, they used a micro-data file created by linking establishment injury data from the Annual Survey of Occupational Injuries and Illnesses, conducted by the Bureau of Labor Statistics (BLS), with inspection data from OSHA s Integrated Management Information System (IMIS). 2) Instead of looking for effects only in the year after the inspection, they followed up to include effects 2 to 3 years after the inspection. 3) Instead of looking only at whether or not there had been an inspection, they also looked at the types of inspections and at any penalties that resulted. Although some prior studies had incorporated some of these features, none had used all of them. The Scholz-Gray studies found that penalty inspections (about 35% of the total) were followed by significant declines in lost workday injuries, ranging from 15% to 22%, depending on the model used. They found no effect of non-penalty inspections. In a recent (2002) study, we presented evidence that the average effectiveness of federal OSHA inspections had declined substantially over time. Using the Scholz-Gray methodology 2

5 and three similar datasets, we found that the impact of a penalty inspection on injuries fell from 15% in to 8% in and to 1% in The effect in the 1990s was not statistically significant. The findings suggested that the impacts might be greater at smaller workplaces (fewer than 250 workers), but the results for this subgroup were also not statistically significant in the 1990s In this paper we focus exclusively on the data from the period, considering OSHA s impact on separate injury types (days away from work and restricted work activity cases) as well as the overall number of lost workdays. The major objective of the paper is to determine whether the impact of an OSHA inspection varies, depending on the characteristics of the inspection or the plant being inspected. Plant characteristics included employment size, unionization status, and industry. Inspection characteristics included motivation (programmed or complaint), type (safety or health), jurisdiction (state plan or Federal OSHA), and whether or not the inspection imposed a penalty. Differences in OSHA s impact across these categories may shed some light on how OSHA inspections could reduce injuries. Differences in inspection impact might also justify some reallocation of OSHA s effort towards the most effective areas. Section 2 of the paper discusses the data used and the analysis and the process of database construction. Section 3 discusses the model and some econometric issues affecting the analysis. Section 4 presents the results, and Section 5 summarizes the findings and concludes. 2. Data The basic sample universe for the data set is all manufacturing establishments in the U.S. We focus on manufacturing workplaces, since they have long been a focus of OSHA activity and are longer-lived and better-defined than workplaces in other sectors (such as construction). This 3

6 is important, since we want to measure the impact of OSHA inspections on injuries over several years. Our earlier study included only the 29 states (and District of Columbia) where federal OSHA directly operates the enforcement programs. In this paper we add establishments in the 21 states where federal OSHA has turned over enforcement authority to state agencies (so-called state plan states), allowing us to test for differences between the two groups of states. The injury information we use is confidential plant-level injury data from the Bureau of Labor Statistics (BLS) Survey of Occupational Injuries and Illnesses. BLS collects data from hundreds of thousands of establishments each year in a stratified sampling process that results in larger establishments being more likely to be in the sample. Since we are looking at changes in an establishment s injuries over time we focus on those establishments that have BLS injury data for consecutive years. This necessarily results in large establishments being over-represented in our data sets, relative to all manufacturing establishments. In addition to injury data, the BLS survey includes name and address information about the establishment. Figure 1 shows the lost workday injury and illness rate (LWD) per 100 full-time manufacturing workers from 1972 to The numbers are based on reporting to the annual Survey of Occupational Injuries and Illnesses, conducted by the Bureau of Labor Statistics (BLS, 2003b). The rate for all LWD cases changed little from 1972 until the early 1990s, except for the expected cyclical changes. Injury rates typically fall in recessions and increase in booms, primarily because of the changes in the number of newly hired, inexperienced workers. However, in the 1990s the overall manufacturing injury rate dropped by about 25% despite continuous prosperity during those years. Figure 1 also shows two subcategories of lost workday rates: cases with days away from work (DAW) and cases with only restricted work activity but no full days away from work (RWA). These rates diverged, with the RWA rate rose 4

7 substantially after the mid-1980s while the DAW rate declined substantially, accounting for the overall LWD decline in the 1990s. Given the differences in trends for the three injury measures, we examine them separately in all of our analyses in this paper. The information on OSHA inspections comes from OSHA s Integrated Management Information System (IMIS), which contains data on all Federal OSHA inspections since the early 1970s. In addition, since 1990 it also includes inspection data from all of the State Plan states (those where the inspection program is conducted by the states themselves under Section 18(b) of the Occupational Safety and Health Act rather than by Federal OSHA). The data include the opening date of the inspection and the name and address of the establishment being inspected, along with the inspection characteristics and results. Our key measure of inspection results is whether or not a penalty was imposed. 1 In our data set we include three types of inspections: programmed inspections, targeted by OSHA based on industry hazardousness; complaint inspections, where OSHA is responding to a worker s complaint; and referral inspections, usually cases where a health inspector makes a referral to a safety inspector or vice versa. These three types include the great majority of all inspections. Through 1997, OSHA targeted its programmed inspections by identifying industries with high injury rates in a particular state. It then chose inspection sites randomly within this industry-state cell, excluding only those workplaces with fewer than 11 workers and those which had a comprehensive OSHA inspection within the previous 2 years. Thus there is no reason to think that programmed inspections are triggered by higher injury rates at the inspected plant There is evidence that complaint inspections take OSHA to establishments that have high LWD rates within their industries (Mendeloff, 1997). However, that could result either because 1 An alternative measure could be whether or not a serious violation was cited during the inspection: during this time period 99% of OSHA inspections with penalties had serious violations, and 93% of inspections with serious violations had penalties, so they are essentially equivalent. 5

8 a high injury rate in that year triggered the complaint or because establishments that generate complaints are disproportionately those that regularly have high rates. Only the former would create a regression to the mean bias affect the validity of the study. We do not have a long enough time series in our sample to examine these patterns. However, if the changes following complaint inspections were due only to regression to the mean rather than to the effect of the inspection, we should not expect to find that inspections with penalties had any more effect than inspections without penalties. The number of injuries would go down merely because it had been unusually high. However, in fact, we do find that complaint inspections with penalties have significantly greater effects than complaint inspection without penalties. The main inspection types excluded from our analysis are follow-up inspections and accident investigations. Follow-up inspections were excluded because they represent the second stage of a single intervention at the establishment and because they are not intended to reexamine other parts of the workplace. Accident investigations were excluded because they inevitably involve reverse causation, with injuries affecting these inspections rather than the other way around. The OSHA and BLS records were linked together, using various establishment characteristics to identify records that referred to the same establishment. These characteristics included name, address, zip code, city, state, employment size, and industry. The matching algorithms were based on a technique developed by Fellegi and Sunter (1969) that calculates the probabilities of two records matching, based on the agreements and disagreements among all of the characteristics. The general matching methodology is explained in more detail in Gray (1996). 6

9 Since our analysis focuses on injury changes, two consecutive years of BLS data are needed to generate one observation for analysis. For the period from 1992 to 1998, we included all establishments with 3 or more consecutive years of BLS Survey data. The resulting data set has 50,276 observations from 16,036 establishments. Table 1 presents summary statistics on the variables in the data set. The average plant has 589 employees and only 23% of the observations come from plants with under 100 employees. These plants are much larger than the average for all OSHA inspections of manufacturing plants during the same period, which average 244 workers and have 68% with fewer than 100 employees. These plants are also more likely to be unionized (42% vs. 23%) and fewer of their inspections are programmed (43% vs. 53%) or safety (57% vs. 66%), though they have a similar probability of imposing a penalty (62% vs. 65%). 2 Half of the plants are located in State Plan states, and about half don t receive any inspections over the period (MISSINSP, which is relevant because we only observe union status for inspected plants). Considering the dependent variables, overall lost workday injuries are declining 3.9 percent per year at these plants. The two injury components show very different experiences, with an annual decline of 7.8% in days away from work cases and an increase of 4.3% in restricted work activity cases. There are also declines in employment (0.8% per year) and hours (0.7% per year) in the typical plant over this period. 3. Model In this paper we model the effects of inspections in terms of two sets of characteristics, features of the establishment being inspected and features of the inspection itself. Under the 2 These numbers are calculated per inspection, not per plant, so don t correspond to the data in Table 1, except for the mean unionization rate, calculated as UNION/(1-MISSINSP), since we only know union status for plants that were inspected at least once in our dataset. 7

10 former, we consider the number of employees in the establishment, whether the workers are represented by a union, and the industry that it is in. Under the latter, we consider whether a penalty was levied, along with the motivation for the inspection (programmed or complaint), its type (safety or health), and whether the plant was located in a Federal OSHA or a State Plan state. We expect that the impact of an OSHA inspection on a particular establishment will depend heavily on that establishment s intrinsic safety level, determined by the firm s demand and supply for safety. This affects how much hazardousness remains for OSHA to influence. In addition, it s possible that some plants may be more or less responsive to OSHA influence. The firm s demand for safety will largely depend on the strength of the incentives provided by existing institutional arrangements. Smaller firms are partially insulated from the financial consequences of injuries by the limited extent of experience rating for them by workers compensation insurers. Larger firms may also be under closer media scrutiny, incurring large public relations costs if they are not perceived as good corporate citizens by their customers. The most trustworthy outcome data, which are for fatalities, do indicate that small establishments have fatality rates that are many times higher than those for large establishments in the same industry (Mendeloff and Kagey, 1991). 3 Larger establishments also tend to have been in existence longer, so they will on average be more likely to have had prior OSHA inspections, probably diminishing the impact of the current inspection. Larger firms and larger establishments are also more likely to have unionized workforces. Unions create a mechanism through which workers can bargain collectively over safety and health conditions. Unlike the market, which gives the greatest weight to the marginal worker, 3 For lost workday injuries, the BLS Survey finds that there is an inverted U relationship of injury rates to establishment size, with the highest rate in the category. The findings for fatalities suggest that the smallest establishments may underreport non-fatal injuries, although other interpretations are possible. 8

11 unions will tend to represent most fully those with the median preferences, who are likely to be older, more knowledgeable, and perhaps more risk-averse. Viscusi found that wage premiums for risky jobs were considerably larger at unionized workplaces (1979b). Weil has shown that unionized workplaces are more likely to be inspected and that inspections there tend to be more intensive taking more time and citing more violations. (Weil, 1991,1996, 2001). 4 Unions also may make workers more knowledgeable about hazards and may increase their willingness to call for OSHA inspections in order to leverage their demands. In manufacturing in , over 30% of complaint inspections were at unionized workplaces, compared to just 15% of programmed inspections. On the supply side, larger firms and establishments are also more likely to employ on-site safety experts, whose presence increases awareness of government rules, reduces some of the marginal costs of meeting them, and also should foster the implementation of effective nonregulatory programs for injury prevention. These safety experts may affect demand if they become in-house advocates for improved safety. Industry factors may also play a role. If injuries in some industries are typically more costly, then the incentive to prevent them will be greater there. Industry-specific technology will potentially affect not only the average cost of injuries, but also the average costs of prevention. For example, safety in outdoor environments (e.g., logging) will be harder to maintain than in the more controllable environment inside a factory. Workers who are widely dispersed may be 4 In inspections in manufacturing from 1996 to 1998, we found that the number of serious violations cited in programmed inspections was larger for unionized workplaces for all size classes; the difference increases in larger size classes. For complaint inspections, the differences between union and non-union workplaces are smaller, and the latter actually have a larger number of serious violations in establishments with fewer than 50 workers. However, it is possible that the higher number of violations cited at unionized workplaces reflects, not a lower level of firm compliance, but either greater OSHA scrutiny or a lower threshold for an OSHA citation. Unions can and often do generate more pressure on employers to reduce workplace hazards. 9

12 harder to monitor. It is also true that OSHA standards may be more relevant to the hazards in some industries than they are to others. The mechanisms used to generate OSHA inspections also interact with some of the characteristics discussed above (Siskind, 1993). As noted earlier, OSHA targeted its programmed inspections based on industry average injury rates within a state. Although large and small establishments had a roughly equal chance of receiving programmed inspections (at least for those with 11 or more employees), complaint inspections tended to be proportional to the number of workers at a workplace. As a result the ratio of annual inspections to establishments in employment size classes ranged from 0.05 for establishments with fewer than 20 workers to 0.74 for establishments with more than 500 workers. Also, as noted earlier, inspections are more frequent at unionized establishments. For all of these reasons, establishments that are large, unionized, or in high injury rate SICs are more likely to have had OSHA inspections. To the extent that OSHA inspections display declining marginal effectiveness, we might expect to find smaller impacts there. 5 The basic model estimated here takes the following form: 1) dlwd t = a t + b 0 IPEN t + b 1 IPEN t-1 + b 2 IPEN t-2 + b 3 IPEN t-3 + c 1 demp it + c 2 dhrs it + SIC2 i + YEAR t + u it The annual log change in the number of lost workday injuries at an establishment is modeled as a function of whether the establishment had been inspected and penalized in the prior three years. For example, for an observation on the change in injuries from 1993 to 1994, IPEN t refers to 5 The long-term effects of prior inspections depend upon whether the hazards that were cited and abated stay fixed. We suspect that when abatement involves changes in equipment, the correction is likely to last longer than when the hazard involved behavioral mistakes. But in general, we would expect at least some degree of declining marginal effectiveness of inspections. Gray and Jones (1991) do report that, at least with respect to citing violations for overexposing workers to toxic substances, inspections after the first one are much less effective. 10

13 inspections with penalties during 1994 while IPEN t-3 refers to 1991 (below we refer to these as IPEN and IPENL3). The dependent variable (dlwd) is the change in the log of the number of injuries, allowing us to interpret coefficients in terms of their impact on the percentage change in injuries. 6 This specification was chosen by Scholz and Gray (1993) after extensive tests with various alternatives. In particular, they find that measuring the dependent variable in logs (rather than changes in logs) results in a very large positive endogeneity bias in the IPEN coefficients, since more hazardous plants tend to get inspected more often. Measuring injuries in a percentage change form greatly reduced this bias in their analysis, and we do the same here. The central explanatory variables are related to whether an establishment had any OSHA inspections with penalties during the year (IPEN), recalling as noted above that we include only programmed, complaint, and referral inspections in our data set. Earlier work with this data set had shown that counting the number of inspections or the number of penalty inspections occurring within the year performed less well than a 0/1 dummy indicating one or more inspections, so we follow the same specification for all of our inspection variables. Control variables included the annual log change in the number of employees (demp) and the annual log change in the number of hours worked (dhrs). The coefficients on both of these variables were expected to be positive, with demp reflecting the role of inexperience for newly hired workers and dhrs reflecting the hazards associated with speeding up production and working longer hours. Other control variables included dummy variables for year and industry, using the 2 digit Standard Industrial Classification (SIC2, with SIC 20 as the excluded category). We explored 6 For all of the variables expressed in logs, we took the log of (x+1) rather than the log of x, in order to avoid taking the log of zero. 11

14 using 4 digit SIC controls in preliminary runs, but they had little effect on the coefficients of the policy variables. In some models, we also tested state dummy variables in order to capture possible differences across states, particularly those arising from change in state workers compensation laws. Including state dummy variables slightly increased the size and significance of the IPENX coefficients (making them roughly 5 to 10% larger than those shown here; results available on request). The results presented here are all based on ordinary least squares regressions. Earlier studies by Gray and Scholz with data from applied a variety of statistical methods to the same basic model, and we have tested these methods with our current data (results available from authors on request). Allowing for autoregressive errors, we find essentially the same pattern of signs and significance shown here, and significant negative autocorrelations in the errors. Using a panel data method due to Chamberlain (1982, 1984), we find larger and more significant impacts of OSHA inspections on reducing injuries than shown here with OLS regressions. Gray and Scholz (1993) also found larger impacts using the Chamberlain method. Unfortunately, our implementation of the Chamberlain method requires a balanced panel of plants. The shrinking sample sizes collected for the BLS survey in the 1990s means that we lose over 80% of our sample when we balance it, and 94% of the plants in the balanced sample have 250 or more employees while our analyses show the strongest results for smaller plants.. 4. Findings The effects of every specification are shown for each of the three injury types: the overall lost workday injuries (LWD), and its two components, injuries with days away from work (DAW), and injuries with restricted work activity (RWA) but no lost time. We begin with 12

15 models of the effect of inspections with penalties and then introduce interactions with a) characteristics of the workplace and b) additional characteristics of the inspection. Overall Impacts The first column of Table 2 shows the effect of inspections with penalties (IPEN) in different years on the percentage change in the total number of lost workday (LWD) injuries for the entire sample of establishments. None of the 4 IPEN variables had a statistically significant impact on LWD injuries. For injuries with days away from work (DAW), only the IPENL1 coefficient was significant. For restricted work activity injuries (RWA), two of the coefficients were positive and significant and one was negative and significant. To provide a simpler way of summarizing the IPEN results across the different lags (especially in later tables, when we are estimating different impacts for different inspection and plant types), we created a variable (IPENX) which takes the value of 1 if there was an inspection with a penalty in any of these 4 years. For the full sample, the IPENX variable had a value of 1 in just over 26% of the observations. Its coefficient is usually close to the average of the IPEN- IPENL3 coefficients. The IPENX coefficient should be interpreted as having a cumulative effect; e.g. a coefficient of implies a total reduction of about 8% over 4 years. The right panel in Table 2 shows the results for IPENX. The significant negative coefficient on DAW injuries is mostly offset by the positive coefficient on RWA injuries, leaving a negative but not significant coefficient on the combined category of LWD injuries (recall that in our sample RWA injuries are about 40% of all LWD injuries). Establishment Characteristics We looked at four characteristics of establishments to see whether OSHA inspections have different impacts at different types of establishments: 13

16 a) Size, measured by number of workers b) Location, whether in states where federal OSHA operates or in State Plan states which operated their own programs with some OSHA oversight. c) Unionization status d) Industry, measured by 2-digit SIC. The left panel of Table 3 includes the first two of these variables. The Small category includes establishments with fewer than 100 workers; the Medium category those with 100 to 249 workers; the Big category those with 250 to 499 workers, and the Very Big category those with 500 or more workers. STPLAN is a dummy variable for one of the 21 State Plan states and STIPENX is the interaction between that dummy variable and IPENX. The results show substantial differences in the impact of OSHA inspections on establishments of different sizes. Because of structure of the dummies, the IPENX coefficients refer to the effect of penalty inspections at an establishment with fewer than 100 workers in a Federal OSHA state. In the first panel, the effect of penalty inspections on LWD injuries in this group of establishments is negative, but not quite significant at the 0.10 level in a two-tailed test. The effect on DAW injuries is significant at the 0.05 level, indicating a cumulative injury reduction of about 16% (4 times -0.04) over 4 years. As in Table 2, we find that the two subgroups of LWD injuries have opposite signs, but here the negative effect dominates. The coefficients on IPEN2X through IPEN4X show that OSHA inspections are progressively less effective in larger establishments. For establishments with over 499 workers, this difference is statistically significant: the net impact of an inspection in a very big plant ( ) is close to zero. 14

17 The coefficients on STIPENX are positive, which indicate smaller impacts of penalty inspections on injuries in State Plan states, but the coefficients are small and insignificant. We also tested whether inspection effects differed among the 7 State Plan states for which we had the most observations. These differences were not statistically significant, so we don t explore them further here. 7 The right panel of Table 3 adds a dummy variable for the presence of a union and a term interacting that dummy with IPENX. Unionized workplaces were 9.7% of Small establishments; 18.4% of Medium; 24.0% of Big; and 32.3% of Very Big (21.8% of the entire sample). In addition, because we can only identify union status at plants which were inspected at some point in our data (from 1990 to 1998), we include a dummy variable (MISSINSP) for uninspected establishments. In this expanded model, we see that the IPENX coefficient grows in size and significance. Now this coefficient refers to a non-unionized workplace with fewer than 100 workers in a Federal OSHA state. The coefficients for the establishment size interactions are not affected by introducing the union variables. The positive union interaction terms (UNIPENX) indicate a smaller impact of OSHA inspections in unionized workplaces, but this effect is only significant for RWA injuries. The control dummy variables measure differences in the trends of injuries across the different groups of establishments. EMP2 through EMP4 show that the LWD and DAW rates at smaller establishments were increasing, and their RWA rates decreasing, relative to larger establishments. STPLAN shows no difference in injury trends between Federal and State Plan 7 State plan states are allowed to use different methods for targeting inspections than federal OSHA does. For example, for some of the period under study the State of Washington targeted its programmed inspections partly on the basis of whether an establishment had recently reported an injury involving more than 7 lost workdays. This could tend to inflate the estimated effect of inspections if injuries at those plants would have declined anyway, due to regression to the mean. 15

18 states. Both the UNION and MISSINSP coefficients show declining RWA injuries during this period. This is estimated relative to a base group of plants that were inspected sometime during the period and were reported as non-union. Uninspected plants also show declining LWD and DAW injuries. In other analyses not shown here, we also tested whether OSHA inspections differed across industries, by interacting 2-digit SIC dummies with IPENX. We found no evidence for significant differences across industries. We also estimated industry coefficients separately for Federal and State Plan states, finding almost no agreement between the two sets (only 8 of the 19 industry interaction coefficients had the same sign and the correlation between the two sets of coefficients was only +0.07). It seems to us that if there were strong industry-specific differences in inspection impacts, we should find more agreement across the two groups of states. Thus we are reluctant to attach much importance to these particular industry interaction coefficients. Inspection Characteristics Now we turn from characteristics of the establishment to characteristics of the inspection. We look at whether the inspection was a programmed inspection or due to a complaint and whether it was a health or safety inspection. We separately estimate effects for penalty and nonpenalty inspections. We also look at the year of the inspection, which might capture changes in inspection policy or practices. The left panel of Table 4 looks at the effects of programmed inspections with and without penalties and complaint inspections with and without penalties. The only statistically significant effect comes from complaint inspections with penalties. In the right-hand panel of Table 4, we look at the effects of health inspections with and without penalties and safety inspections with 16

19 and without penalties. The only preventive effects on injuries came from health inspections with penalties. The finding that health inspections had a bigger average impact on preventing injuries than safety inspections seems counter-intuitive (although a similar result was reported for an earlier period in Gray and Scholz (1991)). Health inspections tend to involve an industrial hygienist, rather than a safety engineer some cross-training allows health inspections to cite safety hazards and vice versa. One possible explanation is that health inspections in this period averaged over twice as much inspector time as safety inspections (Siskind, 2002). Unionized establishments account for a disproportionate share of complaint inspections, so we sought further insight about whether the inspection type effects seen in Table 4 differed by union status. Table 5 interacts the unionization dummy variable with the 4 inspection categories from the left panel of Table 4. Three of the 8 LWD coefficients are significant at the 0.10 level. At unionized workplaces, programmed inspections that do not levy penalties are followed by increases in LWD injuries. In contrast, at non-union workplaces all 4 inspection types are followed by reductions in LWD injuries, with both programmed and complaint penalty inspections having significant effects. Overall, inspections appear to have substantially greater preventive effects at non-union establishments. As in Table 4, inspections which impose penalties show more preventive effects than non-penalty inspections. Table 6 examines similar interactions, focusing on differences in the effects of health and safety inspections. The largest impacts on LWD injuries come from health inspections with penalties, both programmed and complaint. Safety inspections and non-penalty inspections have smaller effects, and non-penalty programmed inspections tend to be followed by increases in injuries. 17

20 Table 7 examines another inspection characteristic - the year in which the inspection occurred. An earlier paper (Gray and Mendeloff, 2002) found declining effects of inspections between those done in the early 1980s and those done in the 1990s. Based on this long-run trend, we expected to find the impact of inspections declining within our time period, with inspections in the late 1990s having a smaller impact than those earlier in the 1990s. Instead, we found greater impacts later in the period. Interacting a trend variable with IPENX generates a negative and statistically significant coefficient for LWD injuries. Interacting specific year dummies with IPENX, we find a decrease in impact between 1993 (the base year) and 1994, followed by steadily increasing impacts through 1998, though the individual year effects are not significant. 6. Discussion and Conclusions Do OSHA inspections lead to reductions in injuries, and does the effect of inspections differ based on observable characteristics of the inspection or the inspected plant? These are the questions that this paper set out to begin to answer, at least for manufacturing in the years from 1992 to It appears that OSHA inspections lead to reductions in lost workday injuries when they occur at small (under 100 workers), non-union workplaces. Over a 4 year period, the estimated reduction in injuries adds up to about 20%. The plant size finding is consistent with the argument that smaller employers will tend to have weaker financial incentives to prevent injuries because of their lesser degree of experience rating under workers compensation. The non-union finding is consistent with their workers having less information about hazards and less ability to exercise voice effectively. Earlier studies (e.g., Smith, 1979) had also found that the largest (indeed the only) effect of inspections was found at smaller plants. Gray and Scholz 18

21 (1991) also found larger effects of inspections in smaller plants, although the effects in large plants were still substantial in their study. Studies of unions have shown that unionized workplaces are more likely to be in compliance with standards (Weil, 1996), consistent with union workers having greater ability to insist on hazard reductions, even without the pressure of an OSHA inspection. We also found evidence that health inspections were, on average, more effective in preventing injuries than safety inspections. Similar results were found by Gray and Scholz (1991) with data from the early 1980s. The robustness of this finding should spur efforts to learn whether the reason is the greater time spent on site by inspectors or relates to the different types of inspection personnel. Our analysis of Federal OSHA and State Plan states does not find any difference in the average effectiveness of penalty inspections. However, we should also note that a) the percentage of State Plan inspections which cite penalties is lower than in Federal States, but b) State Plan states carry out many more inspections than federal OSHA does. On net, the higher inspection rates win out: State Plan states conduct more inspections that levy penalties per 1000 establishments and thus the aggregate impact of specific deterrence might be expected to be larger in those states. However, comparisons of the impact of different inspection strategies cannot focus only on inspections with penalties. It is important to remember that we found that non-penalty inspections were followed by increases in LWD injuries. A possible explanation could be that a clean bill of health from inspectors tends to reduce managerial attention to safety. If this finding is real and causal, any estimates of the overall impact of all OSHA inspections should take into account both reductions and increases in injuries, reducing the overall impact. The fact that 19

22 preventive effects now appear to be limited to the smallest workplaces is troubling in this regard, because those preventive effects will affect relatively few workers, and might be swamped by the non-preventive effects at larger workplaces. This type of composition effect tends to offset the larger cumulative effect of inspections in State Plan mentioned in the previous paragraph, since the percentage of State Plan inspections which cite penalties is lower than in Federal States. If more workplaces are getting a clean bill of health in State Plan states, that could offset the preventive effects of their penalty inspections. The disparity in findings about the effects of penalty inspections on DAW and RWA injuries raises the question of what category, if any, provides a good measure of real changes in workplace risk. There are two broad possible explanations for finding bigger effects for DAW than for RWA. The first is that the different effects are real, i.e., DAW and RWA injuries are distinct types of injuries, and for some reason RWA-type injuries are not being prevented by penalty inspections while DAW-type injuries are being prevented. The second is that there are no real inspection effects, but that having a penalty inspection leads employers to reclassify DAW injuries as RWA injuries, with a greater use of return to work programs. Note that if only the latter explanation holds, we should find no impact of inspections on the aggregate, LWD injuries. It is not evident why the preventive effectiveness of an inspection should be related to how injuries are treated after they occur. For the effect to be real, two conditions need to be met: a) the injuries that lead to restricted work activity are caused by different events than those that lead to staying away from work; and b) inspections could impact the second set of events, but not the first. This scenario seems unlikely and not consistent with the finding that the bulk of 20

23 the difference in the effect of penalty inspections on the two types of injuries arises at unionized workplaces. The alternative explanation is that employers, especially at unionized workplaces, react to penalty inspections by accelerating the substitution of RWAs for DAWs. Why might they do that? Several reasons are possible. First, penalties may heighten employer concerns about inspections and they may (incorrectly) believe that the substitution will reduce the probability that they will be inspected in the future. This belief is incorrect because, to the extent that OSHA uses injury rate data to target inspections, it uses the total LWD rate, not just one of its components. However, it is possible that employers are not aware of the distinction. Second, they may believe that an emphasis on RWA constitutes good practice in loss control. Just as a penalty may prod the firm to upgrade its safety program, so it may also pay more attention to other methods of limiting losses from injuries. Third, employers heavier reliance on restricted work programs may make reporting injuries less attractive to workers. If so, given uncertainty about whether a given injury would be handled as a RWA case, workers might become less likely to report any lost workday injuries. The implication of this reasoning is that growing use of RWA would not only lead to substitution, but also depress reporting of DAW injuries. Conclusions Our results indicate that the effects of OSHA inspections on injuries vary, based both on the characteristics of the inspection and the workplace being inspected. The results are broadly consistent with our hypothesis that inspections will have larger effects where other accidentprevention mechanisms are weakest. Large unionized plants show little effect, while small (under 100 worker) non-union plants show significant reductions in injuries over the three years after an inspection. Inspections which impose penalties are more effective than non-penalty 21

24 inspections, and there is some evidence that injuries rise after non-penalty inspections, which should be taken into account when analyzing the overall effectiveness of OSHA enforcement activity. Not all characteristics matter: inspections carried out in State Plan states seem to be about as effective as those in Federal States, nor do we find evidence for consistently different impacts of inspections at plants in different industries. Some results are surprising (although consistent with earlier research): health inspections are more effective than safety inspections, and we find substantial differences between the impacts of inspections on DAW and RWA injuries. Future research may help clarify these puzzles, as well as pursuing more detailed examinations of the main results of this paper. 22

25 REFERENCES Bartel, Ann P. and Lacy Glenn Thomas, (1985) Direct and Indirect Effects of OSHA Regulation: A New Look at OSHA s Impact, Journal of Law and Economics 28:1-25. Chamberlain, Gary, (1982) Multivariate Regression Models for Panel Data, J. of Econometrics 18:5. (1984) Panel Data, in Zvi Griliches and Michael Intrilligator, eds. Handbook of Econometrics (Amsterdam: North-Holland). Fellagi, Ivan P. and Alan B. Sunter, (1969) A Theory of Record Linkage, J. of the American Statistical Association 64: Gray, Wayne B. (1996) Construction and Analysis of BLS-OSHA Matched Data: Final Report, mimeo, June 22, Gray, Wayne B. and Carol A. Jones (1991) Are OSHA Health Inspections Effective? A Longitudinal Study in the Manufacturing Sector, Review of Economics and Statistics,73: Gray, Wayne B. and John Mendeloff (2002) The Declining Effects of OSHA Inspections on Manufacturing Injuries: 1979 to 1998, NBER Working Paper Gray, Wayne B. and John Scholz (1991) Analyzing the Equity and Efficiency of OSHA Enforcement, Law and Policy 13:185:214., (1993) Does Regulatory Enforcement Work?: A Panel Analysis of OSHA Enforcement, Law and Society Review 27: Lanoie,Paul, (1992) The Impact of Occupational Safety and Health Regulation on the Incidence of Workplace Accidents in Quebec , Journal of Human Resources 27: McCaffery, David P. (1983) An Assessment of OSHA s Recent Effects on Injury Rates, Journal of Human Resources 18: Mendeloff, John M., (1984) The Role of OSHA Violations in Serious Workplace Accidents, Journal of Occ Med 26: , (1997). Do OSHA Inspections Target High Injury Rate Establishments? A Report to the OSHA Office of Statistics, September 12,

26 REFERENCES (cont.) Mendeloff, John M. and Betsy T. Kagey. (1990) "Using OSHA Accident Investigations to Study Patterns in Work Fatalities," Journal of Occ Med 32: Ruser, John W. and Smith, Robert J., (1991) Re-estimating OSHA s Effects: Have the Data Changed? Journal of Human Resources 26: Scholz, John T. and Wayne B. Gray, (1990). OSHA Enforcement and Workplace Injuries: A Behavioral Approach to Risk Assessment Journal of Risk and Uncertainty 3: (1997). Can Government Facilitate Cooperation? An Information Model of OSHA Enforcement, American Journal of Political Science 41: Siskind, Frederick B. (1993), Twenty Years of OSHA Federal Enforcement Data, U.S. Department of Labor, Office of the Assistant Secretary for Policy, mimeo. (2002) "Twentieth Century OSHA Enforcement Data: A Review and Explanation of the Major Trends," U.S. Department of Labor, Office of the Assistant Secretary for Policy, mimeo. Smith, Robert S. (1979), The Impact of OSHA Inspections on Manufacturing Injury Rates, Journal of Human Resources 14: (1992), Have OSHA and Workers Compensation Made the Workplace Safer? in Lewin, D and Mitchell OS, eds., Research Frontiers in Industrial Relations and Human Resources, Industrial Relations Research Association. U.S. Department of Labor (2002), Bureau of Labor Statistics website for December 19, 2002 news release on 2001 fatalities and injuries. Accessed June 15, (2003a), Bureau of Labor Statistics website for Census of Occupational Fatalities. Accessed June 15, (2003b), Bureau of Labor Statistics website for Survey of Occupational Injuries and Illnesses: 1989 to current: Accessed June 15, Viscusi, W. Kip (1979a), The Impact of Occupational Safety and Health Regulation. Bell Journal of Economics 10: (1979b), Employment Hazards: An Investigation of Market Performance (Cambridge: Harvard University Press). 24

27 REFERENCES (cont.) Viscusi, W. Kip (1986), The Impact of Occupational Safety and Health Regulation, Rand Journal of Economics 17: Weil, David, (1991), Enforcing OSHA; The Role of Labor Unions, Industrial Relations. 30: (1996), If OSHA is So Bad, Why Is Compliance So Good? Rand Journal of Economics 27: , (2001), Assessing OSHA Performance: Evidence from the Construction Industry, Journal of Policy Analysis and Management 20:

28 Table 1 Summary Statistics (50,276 observations) Variable Mean (std dev) Description Injury data LWD DAW (0.67) (0.69) Change in log(total lost workday cases) Change in log(days away from work cases) RWA (0.81) Change log(restricted work activity cases) Inspection data IPEN IPENL (0.26) (0.26) Any inspection with penalty this year IPEN lagged one year IPENL (0.26) IPEN lagged two years IPENL3 IPENX (0.27) (0.42) IPEN lagged three years Any inspection this or past three years Plant characteristics DLEMP (0.31) Change in log(employment) DLHRS (0.37) Change in log(hours worked) SMALL (0.42) Employment 1-99 MED (0.42) Employment BIG VBIG (0.43) (0.46) Employment Employment 500+ STPLAN UNION (0.50) (0.41) State Plan state (not Federal OSHA) Union plant (from OSHA inspection data) MISSINSP (0.50) Not inspected (so no union info) Interactions IPENX * plant characteristics IPENXMED IPENXBIG (0.22) (0.23) IPENX * MED IPENX * BIG IPENXVBIG (0.27) IPENX * VBIG STIPENX UNIPENX (0.34) (0.31) IPENX * STPLAN IPENX * UNION 26

29 Table 2 Basic Model Penalty Inspections (50,276 observations; t-statistics in parentheses) LWD DAW RWA LWD DAW RWA Const *** *** *** *** *** *** (-4.03) (-9.58) (6.77) (-4.04) (-9.56) (6.73) IPEN (0.35) (1.03) (1.32) IPENL * (-0.98) (-1.92) (-0.11) IPENL *** (0.47) (-1.47) (2.73) IPENL * ** (-1.82) (-1.35) (-2.35) IPENX * (-0.83) (-1.93) (1.09) DLEMP *** *** *** *** *** *** (24.50) (21.52) (11.01) (24.51) (21.53) (11.02) DLHRS *** *** *** *** *** *** (16.35) (11.70) (9.23) (16.34) (11.81) (9.23) YEAR *** *** *** *** (4.06) (4.12) (0.73) (4.05) (4.13) (0.69) YEAR ** ** *** ** ** *** (-2.22) (-2.39) (-2.65) (-2.25) (-2.42) (-2.69) YEAR *** *** *** *** (-3.12) (-1.25) (-6.34) (-3.14) (-1.28) (-6.36) YEAR * ** *** * ** *** (1.75) (2.07) (-2.93) (1.72) (2.07) (-2.99) YEAR *** *** *** *** (-0.51) (3.31) (-5.41) (-0.53) (3.31) (-5.47) R