65 OSHA Enforcement and Workplace Injuries a Behavioral Approach to Risk Assessment_tcm296-282256

8/10/2019 65 OSHA Enforcement and Workplace Injuries a Behavioral Approach to Risk Assessment_tcm296-282256

http:///reader/full/65-osha-enforcement-and-workplace-injuries-a-behavioral-approach-to-risk-assessmenttcm296- 1/23

Journal of Risk and Uncertainty, 3:283-305 (1990)0 1990 Kluwer Academic Publishers

OSHA Enforcement and Workplace Injuries:A Behavioral Approach to Risk AssessmentJOHN T. SCHOLZDepartment of Po litica l Science , State University of New York, Stony Brook, NY 11794-4392

WAYNE B. GRAY*Department of Economiq Clark University, Worcester, MA 01570

Key words: regulation, enforcement, OSHA, deterrence, safety

Abstract

We develop a model of risk assessment hat incorporates assumptions from the behavioral theory of the firminto conventional expected utility models of compliance, and test the model using data on injuries and OSHAinspections for 6842 manufacturing plants between 1979 and 1985. Four hypotheses are supported-the spe-

cific deterrence effect of an inspection, the importance of lagged effectsof general deterrence, the asymmetri-cal effectsof probability and amount of penalty on injuries, and the tendency of injury rates to self-correct overa few years.The model estimates that a 10% increase in enforcement activities will reduce injuries by about 1%for large, frequently inspected firms. Prior analyses reporting lower impacts (Smith, 1979; Viscusi, 1986a) arereplicated to distinguish between sampling and modeling differences. The results suggest that further compli-ance theory needs more detailed models of risk-assessment processes to be tested on samples of firms mostaffected by enforcement.

Risk assessment has played a central role in theories of enforcement and complianceat least for the last century, since Bentham focused attention on the deterrence effects ofbeing caught and punished. With the development of the expected utility model by vonNeumann and Morgenstern (1947) and its application to criminology (Becker, 1968;Erlich, 1973) enforcement research has taken great strides in modeling and verifying therole of risk assessment in compliance decisions, particularly in the domain of compliancewith regulatory policies.

Most empirical studies have investigated the deterrence hypothesis that greater en-forcement activity (primarily arrests, inspections, citations, and penalties) increases theexpected value of punishment for noncompliers, which in turn increases the motivationto comply. The results of these studies have not consistently demonstrated the expected

*This project would not have been possible without the cooperation of the Bureau of Labor Statistics and theOccupational Safety and Health Administration. Special thanks are due to William Eisenberg at BLS andFrank Frodyma and Joe Dubois at OSHA. We are particularly indebted to John Ruser of BLS, who performedthe data merging and solved numerous problems. The project was partially funded by NSF grant SES8420920.These individuals and institutions do not necessarily support the conclusions in this article.



284 JOHN T. SCHOLZ/WAYNE B. GRAY

linkage (see Nagin et al., 1978, for an early review), although the insignificance of effects

is sometimes interpreted as a sign of ineffective or inadequate enforcement rather thanof a weak theory (Smith, 1976; Downing and Hanf, 1983).This article applies a model of risk assessment to compliance decisions in which the

risk of being penalized by the Occupational Safety and Health Administration (OSHA)affects the firms decisions that are relevant to risk of injuries for workers in the firm. Ourstudy combines elements of previous empirical research on OSHA enforcement that wasbased either on cross-sectional plant-level data (Smith, 1979; McCaffrey, 1983) or timeseries annual data aggregated by industry (Viscusi, 1979, 1986a; Bartel and Thomas,1985) or injury category (Mendeloff, 1979), all of which are reviewed in Viscusi (1986b).As with other deterrence studies, these studies have found mixed evidence about theimpact of enforcement actions on workplace injury rates, which compliance with OSHAstandards is supposed to reduce.

Our research emphasizes a process-oriented model of risk analysis based on the be-havioral theory of the firm (Cyert and March, 1963), in which behavior is explained notonly in terms of expected utility, the basis of prior OSHA research, but also in terms ofmanagerial attention to risks. We find evidence supporting four hypotheses.

1. Firms monitor their injury experience, with unexpected changes in injury rates bring-ing corrective measures to reestablish prior levels of safety risk.

2. Firms monitor OSHA enforcement activity relevant to their particular circumstances,and respond in ways that decrease injury rates when perceived enforcement riskincreases. A 10% increase in enforcement activity reduced injuries by about 1% in oursample, but this adjustment occurs over several years.

3. Firms respond in ways that reduce injury rates when OSHA assessespenalties againstthem. This specific detemence effect occurs even when controlling for the expectedpenalty associated with general deterrence, suggesting the importance of surprisein focusing managerial attention on enforcement risk.

4. Firms monitor the two dimensions of expected penalty (probability and amount)

independently, and respond more to changes in probability than to changes in averageamount of penalty.

We find OSHA enforcement to be more effective in reducing injuries than has beenfound in previous studies, and we attribute this to two factors. First, our sample iscomposed of the kind of firms with which OSHA enforcement is most concerned: larger,frequently inspected firms with higher accident rates than average firms. By minimizingthe number of firms for which OSHA is of little concern, we increase the analytic powerto capture enforcement effects.

Second, the plant-level data set for the 1979-1985 period provides richer informationthan has been available to other studies, thereby allowing us to include in our modelmore details of decision making under risk. The unique data set obtained by mergingOSHA enforcement records and Bureau of Labor Statistics (BLS) injury data for 6842manufacturing plants allowed us to account for individual plant injury experiences overthe seven-year period and to test the four hypotheses listed above.



OSHA ENFORCEMENT AND WORKPLACE INJURIES 285

Our discussion is organized into seven sections. The first develops hypotheses about

plant decision making and risk based on the behavioral theory of the firm (Qert andMarch, 1963). The second section describes the data set, and the third and fourth discussestimation methods. We present our results in section 5, compare these findings with ourreplication of other studies in section 6, and summarize the results in the conclusion.

1. Injuries, enforcement, and firm behavior

The level of risk of injuries at a plant depends on a variety of factors, including thetechnology in use at the plant (manufacturers of lumber and wood products averaged11.1 lost workday incidents per 100 workers in 1979, compared with 5.6 for manufactur-ing as a whole), the size and quality of the plants work force (more workers, less experi-enced workers, and more tired workers are associated with more injuries), and the qual-ity of supervision. Most of these factors are conceded to be beyond the direct influence ofOSHA enforcement, at least in the short run. Consequently, most OSHA studies havefocused on factors most likely to respond to changes in enforcement, primarily expendi-tures made by the plant to increase safety, both through physical capital (safer equip-ment) and human capital (better management of risk, more safety training for workers).

The expected utility model that informs most empirical studies assumes that the levelof risk in a plant reflects the optimal level of safety expenditures, which is determined bythe marginal cost of the expenditure relative to the expected savings from accidents beingprevented. OSHA inspections, citations, and penalties are assumed under specified con-ditions to increase incentives for safety expenditures that reduce the expected penaltiesfor noncompliance. A detailed presentation of this theoretical model and an empiricallytestable derivation that provides the starting point of our analysis can be found in Viscusi(1979; see also Bartel and Thomas, 1985).

Behavioral decision theory research has found considerable evidence that individual

decision making under risk and uncertainty deviates consistently from behavior definedas optimal in expected value terms, and has investigated various heuristics or decisionaids that account for systematic deviations (see Shoemaker, 1982, Kahneman, Slavic, andTversky, 1982). The overall significance of such heuristics for the study of aggregate firmbehavior, and thus for the existing studies of OSHA enforcement, is not yet clear. Wesuspect that models of the firm that analyze comparable heuristics in firm decisionprocesses will lead to more powerful explanations of firm behavior in response to en-forcement and worker injury risks.

The behavioral theory of the firm (Cyert and March, 1963) provides a framework forconsidering decision processes within the firm. The theory includes four major concepts:quasi-resolution of conflict (addressing multiple goals sequentially rather than simulta-neously); uncertainty avoidance (short-run reaction to feedback rather than long-runplanning); problemistic search (solving particular problems, rather than general optimi-zation); and organizational learning (adaptation of goals and attention rules as the envi-ronment changes).




These concepts are based on observations of business decision making processes, in

particmar observing that firms behavior deviates systematically from optimal perfor-mance (which would simultaneously maximize expected profit over all possible behav-iors) because of limitations on the firms decision-making ability. Attention is focused ineach period on that area in which the firms performance fell furthest below expecta-tions. This sort of behavior, called putting out fires, has been examined in more math-ematical detail by Radner (1975): He showed that this is an effective strategy for survival(if there are any effective strategies), and that it tends to keep the firms performance ondifferent areas close together.

Our model of the decision process affecting accidents incorporates four hypothesesimplicit in the behavioral model that extend the basic model described in Viscuisi (1979,1986a). First, an unexpected increase in accidents will cause managers to pay moreattention to safety. This should lead to a reduction in accidents in later years until thefirms attention turns back to other areas. Similarly, a lower rate of accidents than usualshould lead to less attention and the possibility of rising accidents in later years. Al-though managerial attention and safety expenditures are not measured directly in thedata, the implication is that surprising changes in the number of accidents over timeshould be negatively correlated, which implies that the error term in the equation esti-mating current risk should be negatively correlated with past error terms.

Second, several years could be required to observe the full effect on injuries of changes

in OSHA enforcement. This is due to the time needed for organizational learning: thefirms decision processes are only modified slowly, as the firm learns of the changes in itsoperating environment. The more peripheral the information to primary organizationalprocesses, the longer the lag between environmental changes and responses by thefirm. Previous models (Viscusi, 1979) have also tested for delayed effects of enforce-ment attributed to the lag between safety expenditures and their impact on injuries,which would predict permanent reductions in accidents. Alternatively, if firms respondedto OSHA inspections with more transient changes in operations and administration,more immediate temporary reduction in risks could be followed by later increases, even-

tually returning the firm to the prior level of risk as management turns its attentionelsewhere.Third, an actual penalty imposed on the firm can result in reduced injuries quite apart

from the effect related to the firms subjective probability estimate about the likelihoodof being penalized. Particularly for firms having problems in areas other than safety,behavioral theory suggests that a penalty might increase managerial attention to safetyissues, ust as a sudden increase in injury rates might do so. In the expected utility model,the actual penaltys effect would suggest a Bayesian learning process in which the occur-rence of an inspection with penalty produces a correction in prior estimates of theprobability of being pena1ized.l Deterrence analyses of criminology and deviance behav-ior have long distinguished between general deterrence-the effect of an act of legalpunishment on the subsequent behavior of the general populace-and specific deter-rence-the effect of the punishment on the subsequent behavior of the individual beingpunished. These analyses suggest that specific deterrence could be considerably greaterthan the general deterrence associated with changes in the aggregate level of enforcement.




Fourth, marginal changes in either the probability of a penalty or the average penalty

amount may have quite different general deterrence effects on accident rates, dependingon which of the two is most salient to the firms monitoring of the external environment.It may be that firms have good information on the number of similar firms that arepenalized, but not on the amount of penalties (or conversely may pay more attention to afew extremely large penalties). Expected utility theory generally treats this issue in termsof risk preferences that affect how changes in probability will be weighted in comparisonto changes in the value of outcomes. The issue is important in determining the optimalpolicy mix of wide-ranging coverage versus intensive inspections. Consequently, ourmodel represents general deterrence by separate measures of the probability and aver-age amount of penalties.

2. Data description

The data set assembled for this project combines information over time on both accidentrates and OSHA enforcement, data that were not available at the plant level for previousstudies. A data set produced by the BLS that contained plant-level accident records from1979-1985 was merged with the OSHAs Management Information System (MIS) filecontaining enforcement actions for all plants during the same period. The BLS filematched records from the BLS Annual Survey for all plants with data for each year from1979 to 1985, based on a common identification number available in the annual files(Ruser and Smith, 1988). All plants in this file that were located in the 28 states withfederal OSHA enforcement covered by OSHAs MIS were then matched with theOSHA enforcement file.

Since no common identification number was available in both OSHA and BLS datasets, we employed a sophisticated record-matching program based on the technique ofFellegi and Sunter (1969) as described in Gray (1987). Both data sets contain various

characteristics of the plant, including firm name, address, zip code, city, state, employ-ment, and industry. These characteristics were used to match plants in one data set toplants in the other, based on the probability of agreement on particular variables.2 Toprotect the confidentiality of firms in the BLS Annual Survey, all merging operationswere done at BLS and identifiers of each plant were removed, thereby limiting our abilityto add data not available in either data source.

The final data set consists of 6842 plants with annual data from 1979 through 1985. Foreach year, we know employment and hours worked, as well as the number of lost work-day injuries and the total number of lost workdays. Each OSHA inspection of the plantduring the 1979-1985 period is recorded, including information about the kind of inspec-tion and the citations and penalties assessedas a result of the inspection.

The plants in the data set are not representative of the manufacturing sector, as can beseen by the comparisons in table 1. The BLS surveys are based on stratified randomsamples that oversample large plants, and so the plants included in seven consecutivesurveys are considerably larger than the typical manufacturing plant. They averaged 523




Table 1. Comparison of sample with national manufacturing sector

Sample(1979)

Sample National(1979-1985 manufacturing sectoraverage) (1979)

Number of plantsNumber of employeesAverage employees per plantAverage lost workday injury rateNumber of lost workday injuriesNumber of lost workdaysAverage injuries per plantAverage lost workdays per plantNumber of inspectionsNumber of inspections w/ penaltyTotal penaltiesAverage inspections per plantAverage inspections w/ penalty per plantAverage penalty per inspectionProbability of inspectiondProbability of inspection w/ penaltyd

6,842 6,8423575,394 3,271,318

523 4796.97 6.02

171,333 132,3052,484,704 2,073,126

25 19363 303

3,458 2,5981,145 790

$1,722,973 $691,657.51 .38.17 .12

$498 $269.27 .26.13 .lO

349,913a18,510,49Sa

545.9b

1,243,000b18,998,C00b

454

28,293c9,453c

$10,543,99OC.08.03

$ 373

Sources:a. Census of Manufactures, 1977.b. Occupational Injuries and Illnesses in 1979: Summary (BLS: April, 1981).c. OSHA Management Information System.d. Differs from average inspections by eliminating multiple inspections of the plant within the year.

workers in 1979, compared with 54 workers for all manufacturing plants. The averagelost workday incidence rate in 1979 was 6.97 for the sample, compared with 5.9 formanufacturing as a whole.

The plants in the sample are relatively heavily inspected by OSHA, with 27% of theminspected in 1979, compared with 8% for all manufacturing plants. Furthermore, plantsin the sample represent almost 20% of the employees in manufacturing in 1979, andaccount for an even greater percentage of accidents in the manufacturing sector. Inshort, the sample represents a considerable if not necessarily representative proportionof OSHA? enforcement effort in manufacturing. Since plants in our sample face greaterenforcement pressure, we would expect them to pay more attention and be more respon-sive to OSHA enforcement than the typical manufacturing firm. Our sample thereforeprovides more analytic power than a representative sample for analyzing firm responsesto enforcement.

3. Estimation procedure

To estimate the response of firms to risks of enforcement and of injury, we use measuressimilar to those in previous studies of OSHA that were available in our plant-level




Table 2. Variables used in analysis

VariableSample (Std.mean dev.) Description

1. Injury measu res

Injuries

%CHG Injuries

19.3 (40.4)

- ,051 (.80)

Average Injuries 19.4 (37.8)

Lost Workdays 303 (731)

%CHG Lost Workdays - ,036 (1.0)

2. Probability of inspe ction with penalty

Inspection with Penalty

Predicted Probability

%CHG Predicted Probability

Industry Probability of Penalty

3. Amount of penalty

Penal@+

Predicted Penaltya

%CHG Predicted Penalty

Industry Average Penaltya

4. Size of firm

Hours of Work

%CHG Hours

Employment

%CHG Employment

.099 (.29)

.106 (06)

- ,010 (.02)

.038 (.03)

.600 (1.8)

5.962 (.31)

- ,086 (.17)

4.833 (.69)

926 (1912)- ,019 (.25)

479 (982)

- .019 (.22)

Total plant-year observations = 48,794

Number of lost workday injurie s

Percentage change in Injuries:

Injuriesr - InjurieS- t

(Injuriest + Injuries,- t)/2

Average injurie s in past two years:(Injuriesr- t + Injuries( -2)/2

Number of lost workdays

Percentage change in Lost Workdays:

Lost Workdaysr - Lost Workdays, - r

(Lost Workdays, + Lost Workdays+ t)/2

Inspection with penalty during year (dummy variable forsample firms)

Predicted probability of inspection with penalty (basedon table 3 coefficients)

Percent change in Predicted Probability [Predicted

Probabilityt - Predicted P robab ility-i]

Industry p robability of inspe ction with penalty (inspec-tions with penalty/establishments in industry, based on

national totals from OSHA MIS, aggregated by two-digitSIC industry)

Log of total pena lties assesse d against firm during year

Predicted log of penalties assessed when p enalties wereimposed (based on table 3 coefficients)

Percentage change in Predicted Penalty a s calculated bydifference in logs of predicted penalty [PredictedPenal@ - Predicted Penalty-t]

Industry average log (penalty) asse ssed if inspection with

penalty occurred (two-digit SIC nation al totals fromOSHA MIS)

Hours worked during year (in thousands)

Percentage change in Ln (Hours):

Ln(HourQ - Ln(Hour+ 1)

(Ln(Hours,) + Ln(Hour+ i))/2

Average employment during year

Percentage change in

Ln(Employment,) - Ln(Employmen& 1)

(Ln(EmploymenQ + Ln(Employment,- t))/2

a. Note that Penalty is averaged over all plant-year observations in the data set, but Predicted Penalty andIndustry Average Penalty are based only on inspections with penalty.



290 JOHN T. SCHOLZiWAYNE B. GRAY

data set. Summary statistics and descriptions of each variable are provided in table 2. As

in most studies, two different measures of injury risk are used, one reflecting frequency(annual number of lost workday injuries) and one reflecting severity (annual number oflost workdays) of industrial accidents in a plant in the designated year.

The frequency or severity of accidents in a given year (Ai,) is hypothesized to dependon the hazardousness of this particular plant (Ai), the number of hours worked at theplant (Hit), the experience level of the workers at the plant (Q& the expected enforce-ment faced by the plant (X,,*) and the attention to safety given by the plant manage-ment (&), as well as an unexpected residual component (&). Expressing this as a linearregression gives us

In order to remove the plant-specific hazardousness, we consider the first-differencedform of this equation (replacing Ail with aif, and so on, to reflect the movement tochanges, rather than levels, of variables):

In fact, we use the proportional rather than the simple change in each variable, because

the estimated coefficients obtained in the simple-change form of the regression are verysensitive to a few outliers (plants with very large changes in injuries). Switching to aproportional-change form reduced the influence of these outliers. The use of changerather than level as a dependent variable also minimizes the endogeneity problemcaused by the relationship between inspections and injury rates: as noted in the nextsection, penalty inspections are strongly related to injury levels, but not to changes inlevel. This supports the plausible argument that inspectors tend to use penalties on moredangerous plants but take little account of recent changes injury rates?

Of the variables in (2), we measure a andh directly, while the others are proxied by

variables available in the data set. Changes in worker experience, q, is measured bychanges in the number of workers at the plant (with a negative sign, since new workershave less experience). There are two components to expected enforcement, x*, corre-sponding to deterrence theorys two components of expected penalty: the probability of(an inspection with) penalty and the expected level of penalty. Both are transformed tothe change form, as discussed in the next section.

We use two proxies for s, the changes in attention paid to safety by plant management.The specific-deterrence shock effect of having received a penalty from OSHA after arecent inspection of the plant is represented by the inspection history of the plant(Inspection with Penalty, a dummy variable equal to 1 in a year in which an inspectionwith penalty occurred).4 The feedback effect of having had an unexpectedly high numberof injuries in the recent past is represented by an autoregressive structure for the errorterms in (2). In other words, that component of change in the risk measure that couldnot be attributed to the measured variables (the estimation error) constitutes a surprise,and the effect of past surprises on current risk measures is reflected in the autoregres-sive errors.




Since we expect the impact of enforcement and surprise to affect risk measures over

several periods, these variables are included with multiple lags. The data rejected distrib-uted lag models (that could represent Bayesian learning processes), so we entered laggedvariables directly in the equation. Three period lags have been used in other studies(Viscusi, 1979), provide sufficient time to account for organizational delays, and do notexhaust the full seven years of the data set.5 The control variables representing hoursworked and quality of the work force were not lagged, since they affect the risk measuredirectly.

The final form of the equation to be estimated (with the firm subscript left out forpresentational simplicity) is

%CHG RISK, = PO + 2 Bri% CHG Predicted Probability-ii=O

+ Zpz%CHG Predicted Penalty,-i

3

+ @s$nspection with Penalty-i

+ Bh%CHG Hour+ + Bs%CHG Employment,

+ f (jj*YEARi + f&i*SICi + Vti=83 i=21

3

withv, = et + Cai*vt-i..I=1

Here et is assumed to be an independent, normally distributed series with a mean ofzero. The general deterrence variables and firm characteristics are entered in percentagechange form (designated by %CHG), and the dummy variables for Inspection with Pen-alty, Year, and SIC are not modified. As noted in the variable descriptions in table 2, thelogged form is used for Penalty, Hours, and Employment. Separate estimations weremade for each of the risk variables, %CHG Lost Workday Injuries and %CHG Lost

Workdays.The autoregressive coefficients (ai) address our first hypothesis, which predicts that

they will be negative and sum in magnitude to less than one (to make the feedback modelstable). Enforcement is expected to affect accidents with a substantial lag (hypothesis 2)both general (pii and Bz) and specific deterrence (psi) measures are included (hypoth-esis 3) and the general deterrence effect is separated into predicted probability andpredicted amount of penalty (hypothesis 4). Industry and year dummies are included tocontrol for any systematic changes in injuries along those dimensions. We expect Bii, Bz,and psi to be negative, and control-variable coefficients B4 and Bs to be positive.

4. Estimating predicted probability and amount of penalty

Deterrence is dependent on the firms expectation that it will be penalized if it does notcomply. Generally, empirical studies have relied on aggregate enforcement measures as


http:///reader/full/65-osha-enforcement-and-workplace-injuries-a-behavioral-approach-to-risk-assessmenttcm296 10/23

292 JOHN T. SCHOLi7WAYNE B. GRAY

proxies for general deterrence, since changes in aggregate penalties are presumably

linked to changes in a firms expectations about penalties. Our study is able to model thislinkage explicitly, making use of the enforcement experience of firms as recorded in thedata set to predict the expected probability and amount of penalty in terms of bothindustry-level enforcement measures and firm-level characteristics. For the experienced,routinely inspected firms that dominate our sample, the resultant predictions providericher proxies for an individual firms expected probability and amounts of penalty than isprovided by aggregate enforcement measures alone.(j

The equation used to estimate the probability of penalty reflects the role of OSHAscurrent level of enforcement as well as the size of the firm and its injury experience in therecent past:

Inspect w/ Penaltyi, = pa + BrIndustry Probabilityi, + BzIndustry AveragePenal&+ ps[Injuriesi,- r + Injuriesi,-$2 + &% CHGInjuries+ r 85+ BsHoursi, i- & lTlplOpel& + 1 P-/jYearj -l- eip (4)

j=80

The unit of analysis is the plant. The two industry-level measures of OSHA enforce-ment are based on national aggregates for the plants industry, with Industry Probability

measuring the number of inspections with penalty per firm in the industry and (the logof) Industry Average Penalty measuring the average amount of penalty for inspectionsthat imposed a penalty. The size of the firm is represented by Hours and Employment,both entered in log form. Since OSHA targeting and penalizing decisions may reflect theplants injury experience, we include measures of both the level and change of injuries inthe recent past.

The same equation is estimated for the total annual penalty (in log form), with OLSused to estimate coefficients for Penalty (for inspections that had penalties) and probitfor the dichotomous Inspections with Penalty variable. These coefficients are then used

to calculate the predicted penalty and predicted probability for each firm, which are usedas independent variables in the final analysis. These results are presented in table 3. Asmentioned earlier in explaining why change rather than level was used for the primaryanalysis, note that the Average Injuries variable is highly significant in predicting Inspec-tions with Penalty, while the Change in Injuries is not.

5. Discussion of results

Table 4 reports the basic estimations of enforcement effects on lost workday incidentsand lost workdays. As expected, the independent variables explain less of the variance inlost workdays (6%) than of the variance in lost workday incidents (12%). Both equationssupport the same conclusions about the importance of the behavioral theory of the firmand about the impact of enforcement actions on accidents, although the estimated ef-fects are slightly different in the two equations.




Table 3. Equation used to predict probability and amount of penaltya

Predicted Predictedprobability penalty

Intercept ,041 (6.2) 3.803 (4.9)

Average Injuries .OOO50 (18.4) .000001 (.004)Change in Injuriey- tl JO24 (1.3) ,048 (1.7)Industry Probability ,793 (14.6) 2.096 (3.0)

Industry Average Penalty .0087 (2.7) ,058 (1.2)

Hours of Work - ,039 (- 16.6) ,167 (1.7)

Employment ,044 (15.2) -.0035 (-.03)

Year Dummies

1980

1981

1982

1983

1984

1985

No. observations 47,894 4,735

Mean (dependent variable) .0989 6.07

F-test 172.1 52.2R2 - ,115

- ,011 (-2.7) -.173 (-3.6)

- ,010 - 2.2) - .385 ( - 6.2)

-007 (-1.3) -.535 (-6.9)

- ,016 ( 3.0) - .436 ( - 6.0)

- ,010 (- 1.9) - ,509 ( - 7.2)

-.008 (-1.8) - ,424 ( - 7.0)

a. Based on estima tion of Inspection with Penalty (probit) and Penalty (OLS on nonzero penalties) (t-statisticin parentheses)

5.1. Self-cowecting mechanisms

One of the most striking and robust findings in this and all other estimations we ran is thestrong tendency of surprises in the accident rate to be compensated for and to return to

zero within three years. The autoregressive coefficients q, 012,and or3 are consistentlynegative and highly significant, and their sum approaches - 1. The estimated impact onaccidents decreases for more distant shocks, consistent with the assumption that recentshocks are most important in driving firm behavior. In the estimation based on Percent-age Change in Injuries, any change in a given year is compensated by a 49% change in theopposite direction in the following year, a 32% change in the second year, and a 13%change in the third year. The effect of the residual (unexplained) variation is almost fully(94%) compensated during the next three years. This process alone explains more vari-,ante than the other variables combined, raising the explained variance from 12% for allindependent variables to 29% for the autoregressive estimate for injuries, and from 6%to 27% for lost workdays.

It should be noted that this autoregressive process is not explained by a stochasticprocess in which an unusually high number of accidents in one year is followed by aregression to the mean in the next year, since such a process would not produce multiyearcorrelations among residuals. Furthermore, the autoregressive process associated with




Table 4. Estimated impact of enforcement (Maximum Likelihood estimates, t-statistic in Darentheses)

%CHG in Injuries %CHG in Lost WorkdaysSurprise (autoregressive errors)

One-year lag (cul)Two-year lag (a~)Three-year lag (cq)

- .489 ( - 49.8) - .548 (- 63.8)- ,316 (- 29.5) -.329 (-31.1)- .127 (-9.8) - .141 (- 11.4)

General deterrence

%CHG in Predicted Probability

Year tYeart - 1Yeart -2

%CHG in Predicted PenaltyYear tYeart - 1Yeart - 2

1.208 (4.6)- 1.357 ( - 5.6)-.591 (-2.3)

- .897 ( - 7.8)- .294 (- 3.8)- .381 (-4.9)

1.023 (3.0)- ,840 (-2.6)-.401 (-1.2)

-.446 (-3.6)-.140 (-1.5)- .280 (- 2.9)

Specific deterrence

Inspection with PenaltyYeart

Yeart - 1Yeart - 2Yeart - 3

- .036 (- 2.4) -.OOl (-.05)- ,049 (-3.1) - ,058 ( - 2.7)- ,043 ( - 2.8) - ,043 (-2.1)- .006 ( - 0.4) - ,006 (-0.3)

Firm characteristics

%CHG in Hours%CHG in Employment

,672 (16.0) ,546 (10.3),516 (13.1) .467 (9.0)

Year dummies

1983 ,354 (9.3) ,160 (3.6)1984 .356 (9.7) .198 (4.4)1985 .574 (11.2) ,331 (5.4)

Tbo-digit SIC (19 dummies) (19 dummies)

Intercept - 0.442 (- 11.3) - .24.5 ( - 5.2)

No. observations 27,368 27,368

Mean (dependent variable) - ,046 - ,026R2 (with autoregression) ,289 ,274R2 (regression only) ,123 .063




safety investments, a variable we could not control for directly, would generally lead to

positive rather than negative coefficients as one years investments continue to reducerisk over several years. We believe the self-correcting process is consistent with the behav-ioral model suggestion that a year with a surprisingly high number of injuries inducesincreased attention to plant safety, which in turn reduces injuries in following years.

5.2. Lagged effects of general deterrence

The results in table 4 demonstrate that both the expected probability and the amount of

penalties exert an overall permanent negative effect on accident rates.7 The total effectof a unit increase in predicted probability, obtained by adding the coefficients for thecurrent and lagged values, equals - .75 for injuries and - .21 for lost workdays, with thecomparable figures for predicted penalty being - 1.57 and - .87, respectively. The sig-nificant positive coefficient in both equations for the current value of predicted probabil-ity is unexpected, but may be due to the inclusion of both measures of general deterrence(when the equation is estimated without predicted penalty, the troublesome coefficientapproaches zero). The insignificance of some lagged terms in the lost workday equa-tion may also be explained by the multicollinearity between these general deterrencemeasures.

The length of time required for changes in enforcement to produce their full effect islonger than had been found previously. While other enforcement studies have foundone-year lags between enforcement and effect (Viscusi, 1986a; Scholz, 1987), our esti-mates suggest that the effect continues into later years. Studies using only a single periodmay not capture the full effect of enforcement changes.* This length of delay is consistentwith the assumption that implementation of changes to affect accidents takes consider-able time. The fact that the coefficients do not become positive in longer lags indicatesthat the initial effect does not decay over time, so the effects appear to be due to perma-nent safety investments rather than quick administrative fixes. Our attempts to force thelagged coefficients to follow some smooth decay pattern over time were rejected by thedata in favor of the separate (and fluctuating) coefficients reported here.

5.3. Specific deterrence

The results in table 4 confirm the behavioral hypothesis that the surprise involved in theactual imposition of a penalty has an effect on behavior over and above the generaldeterrence effect of expected probabilities and amounts of penalties. In both equations,the primary effect occurs in the first and second year after an inspection, confirming thelong period of time over which enforcement effects must be measured. Although theeffect remains negative in the third year after an inspection, the coefficient is not signif-icant in either equation. The effect of the Inspection with Penalty variable in the currentyear is insignificant for lost workdays.



296 JOHN T. SCHOLZiWAYNE B. GRAY

As with general deterrence, this pattern is most consistent with permanent invest-

ments rather than transient administrative changes. It does not reflect a pattern ofconsistent decay associated with a Bayesian adjustment process. These results are quiterobust to changes in the measures of general enforcement used, the lag lengths for theenforcement variables, or the inclusion of other controls. On the other hand, the magni-tude of the specific deterrence effect is relatively small when compared to the generaldeterrence effect, as illustrated in table 5 and discussed in the next section.

5.4. Asymmetrical effects of probability and amount of penalty

Expected utility theory generally assumes that the expected penalty for noncompliance(the predicted probability multiplied by the predicted penalty) captures all relevant in-formation for decisions under uncertainty. However, when we added this expected pen-alty variable to the regression in table 4 (in results not shown here), the coefficients forpredicted probability and predicted penalty remained much the same, and the expectedpenalty variable was only significant in one lagged term. The superior performance of theindependent components (probability and penalty) over the product is consistent withthe behavioral theorys suggestion that the probability and the amount of penalty affectaccident rates independently, as they would if monitored and fed into the decision pro-cess ndependently.9

As noted earlier, the impact of changes in probability and in amounts of penalty maynot be symmetrical-that is, a 10% increase in either variable could have the sameimpact on expected value but different impacts on accident rates. Table 5 compares theimpact on injuries of a 10% increase in inspections with a 10% increase in the averagepenalty (assuming that the additional inspections, on average, are as likely to imposepenalties as current inspections). The calculation of impacts follows the two steps used inour estimation procedure. First, the impact on predicted probability and predicted pen-alty is calculated separately for an increase of 10% in the Industry Probability and

Industry Average Penalty variables, based on the coefficients in table 3. Second, the

Table 5. Impact of policy changes on injury measures (Effect of a 10% change in enforcement variables)

General deterrence Specific deterrence Total effect

Predicted Predicted InspectionEnforcement Measure Probability penalty with penalty

I. Lost Workday InjuriesIndustry Probability of Inspection - .22% - 1.26% - .13% - 1.61%

Industry Average Penalty- .06% - .87% - .93%

II. Lost WorkdaysIndustry Probability of Inspection - .07% - .70% -.ll% - .88%Industry Average Penalty - .02% - .48% - SO%




calculated changes in predicted probability and penalty are multiplied by the appropri-

ate coefficients in table 4, giving the separate impacts of probability and penalty on therisk variable reported in table 5. These effects can be summed for the combined generaldeterrence effect. Table 5 also estimates the specific deterrence effect, assuming a 10%increase in actual inspections, and sums the separate effects to give the total effect oninjuries.

The results confirm that changes in probability and penalty are not symmetrical. Thesimulated increase in inspections reduced injuries and lost workdays more than a com-parable increase in penalty. These results are consistent with the general conclusionfrom most empirical research on deterrence (Lempert, 1982; Nagin, 1978) as well asViscusis (1986a) study of OSHA, which finds that the probability of being punished hasa greater effect on compliance than does the amount of the punishment. The results areless consistent with standard assumptions that firms on average are risk-neutral or risk-averse, since only risk-seeking firms would be expected to respond more to probabilitychanges than to penalty changes. Thus, different capabilities for monitoring informationrelated to the probability and amount of penalty may provide a more reasonable expla-nation than differences in risk preference.

On the other hand, specific deterrence has considerably less impact on injuries thangeneral deterrence. The relatively minor role of the shock of being penalized suggeststhat the firms we studied do have relatively effective means of monitoring OSHA en-

forcement and drawing appropriate implications about the risks of penalty. Once thesegeneral effects are controlled for, the remaining surprise of being penalized accounts foronly a small proportion of the effect on injuries.

6. The impact of OSHA enforcement on injuries

Estimates of OSHA enforcements impact on injuries in this and comparable studiesprovide a relatively consistent picture of small but significant effects, with the estimates

from our model and our data higher than estimates reported in other studies. If wecompare studies finding significant results in the percentage reduction in injury measurebrought about by a 10% annual increase in enforcement, impacts range from no signih-cant impact in McCaffrey (1983) and Bartel and Thomas (1985) to ranges of .15% to.36% in Viscusi (1986a), .20% in Smith (1979), .48% to .73% in Cook and Gautschi(1981) to our estimates in table 5 of .5% to 1.6%. Several explanations might account forour higher estimates:

1. The inclusion of both specific and general deterrence effects2. The inclusion of three-year lagged effects3. Our use of inspection with penalty rather than inspection4. Our sample, which overrepresents larger, more dangerous, and more heavily in-

spected plants5. Our model specification, which differs from those used in earlier studies.




We have explored these explanations empirically by using our data set to replicate two

previous studies as closely as possible, one measuring specific deterrence and the othermeasuring general deterrence.

6.1. Specific deterrence

To estimate the impact of an inspection on subsequent injuries in the inspected plant,Smith (1979) used an ingenious research design in which plants that were inspected latein the year (November-December) were used as a comparison group for plants in-spected earlier in the year (March-April). If inspections had an immediate impact oninjuries due to the immediate abatement of violations, this impact would show up in theinjury rates of the group inspected earlier, but not in those of the group inspected later inthe year. The problem of endogeneity of inspections is minimized by comparing the twogroups of inspected plants, eliminating plants inspected in the previous year, and includ-ing other factors influencing injury rates (prior injury rate, employment changes, and sizeand industry dummies). The data source was the annual BLS surveys used in our study,but from an earlier period.

In table 6, Smiths results are compared both with McCatieys (1983) and with ourreplications. Whereas Smith found significant effects from inspections in 1974, McCaf-

frey found no significant results for 1976. Our results, based on inspections during theTable 6. Comparison using Smiths specific deterrence estimates: Impact of inspections on lost workday injuryrates (t-statistics n parentheses)

Smitha McCatTreyb1974 1976

Our data

1982-1985 1982-1985

Specific deterrence

Early inspection - .63 .25 .21 -

(-2.9) (9) (7)Early inspection with penalty - - - -.22

(-.4)Injury Rate _ 1 ..51 .63 ii5 .70

(26.6) (32.9) (34.9) (20.7)

Employment&mploymen& - 1 2.19 49 .72 - .14(3.7) 2.0) (2.7) ( - 4

R2 .35 .44 .55 .58

N 2362 1990 1208 415

Notes: All regressions include industry and size-class dummies. Regressions on our data also include yeardummies to preserve comparability while making use of the multiple years n our data.a. Source: Smith (1979).b. Source: McCaffrey (1983).




1982-1985 period of our previous analysis, are comparable to McCaffreys. As he sug-

gested, the easily accomplished reductions in risk that OSHA inspections could imposemay have already been implemented in 1976, leaving more complex issues of risk reduc-tion less amenable to quick fixes.

What is it that accounts for the difference between the significant specific deterrenceeffect we found with our model (table 4, column 2) and the insignificant effect in table 6,since both analyses are based on the same time period, data set, and measure of risk?The final column in table 6 indicates that limiting inspections to those in which a penaltywas mposed (as we did in our model) at least produces the expected negative impact oninjuries, although still not at significant levels. Given that the impact of an inspection isspread over a three-year period (table 4) the most plausible explanation is that thesingle-year design is too short to capture significant effects of OSHA. Cook and Gautschi(1981) also found significant specific deterrence effects in a study designed to analyzelong-term impacts of enforcement actions. Our finding that specific deterrence contrib-utes a relatively small amount to the overall impact of OSHA (table 5) also makes thelack of significant effects in short-term studies more likely.

6.2. General deterrence

Viscusis (1986a) study analyzes the general deterrence effect on an industrys injury rateof the total annual inspections and penalties in the given industry, and uses the prioryears injuries, several production measures, and year and SIC dummies to control forother impacts on injuries. To replicate Viscusis analysis with our data set, we aggregatedmeasures of injuries, hours of work, and employment in the data set to the two-digit SIClevel, added national aggregate measures of overtime, percentage of female workers,and percentage of production workers, and used Viscusis estimation procedure to mea-sure the impact of national OSHA enforcement activities on aggregate lost workdayinjuries in our sample (using Viscusis log-odds form of the dependent variable).

As in Viscusis study, independent regressions were estimated for inspections and penal-ties. We also estimated regressions using inspection with penalty and average penalty vari-ables comparable to those in our model. Both current and lagged values of the enforcementvariable were included in each regression. The resulting estimates for inspections werecomparable to those reported in Viscusi (1986a, table 2) in terms of signs and significance ofthe control variables and the explained variance (r2 = .98), so in table 7 only the relevantenforcement coefficients were compared with Viscusis results (1986a, table 3).

Table 7 provides two comparisons between Viscusis and our analysis. Taken together,they indicate that the different samples and different measures of enforcement contrib-

ute most to our higher estimation of general deterrence impacts, with the greater detailin our model contributing a relatively smaller amount.

The first two columns provide a direct comparison of differences in coefficients forViscusis model obtained from his national data and our sample data. Using Viscusis



300 JOHN T. SCHOLZWAYNE B. GRAY

Table 7. Comparison using Viscusis general deterrence estimates: Impact of aggregate enforcement mea-sures on log-odds of Lost Workday Injuries (r-statistics in parentheses)

CoefficientsImpact of 10% increase in

enforcementa

National datab Sample dataViscusi model Behavioral

model(1973-1983) (1980-1985) national sample (sample)

InspectionsYear tYeart - 1

-.16% - 1.27% NA9.61 (1.26) -.898* (-3.38)

- 16.64* (- 2.00) ,028 (0.20)

Penalty (per worker) - .04% - .09% NAYear t ,016 (0.94) - .00003 (- 0.01)Yeart - 1 - ,026 ( - 1.52) - .031 ( - 0.77)

Probability (inspections with penalty)Year t NAYeart- 1 NA

NA -.81% - 1.48%-3.27* (-2.55)

0.89 (0.93)

Average penalty per inspection with penalty NA - .66% - .93%Year t NA - .057* ( - 5.47)Yeart- 1 NA - ,009 (- 0.99)

Notes: The coefficients for each enforcement variable are estimated in separate regressions. Following Viscusi(1986a), each regression also includes year and industry dummies, as well as percent female, percent produc-tion workers, average hours per week, average overtime hours, and the change in employment. An asteriskindicates significance at the .05 level; NA indicates not available.a. Since the Viscusi model uses a semi-log form, the percentage impact of a 10% increase in enforcement oninjuries is given by 100x (Coefficientr + Coefficient,- 1)x (0.1 x mean of enforcement variable). Only theaverage penalty model is in log-log form, for comparability with the behavioral model. For average penalties,then, the impact of a 10% increase is simply 10x (Coefficien& + Coefficientr- 1). Values for the behavioralmodel are taken from the general deterrence calculations in table 5, excluding the impact of specificdeterrence.

b. Source: Viscusi (1986a), table 3.

inspection and penalty-per-worker measures, both data sets indicate that only inspec-tions have a significant impact on injuries. However, the variables from our model corre-sponding to deterrence theorys concepts of probability (inspections with penalty) andamount of penalty (average penalty) are both significant for the sample data. The signif-icance of the current rather than the lagged value in the sample data indicates onepossible difference in data sets, although the significance of multiple lags rather thancurrent values in behavioral-model estimates on the sample data suggests that this dif-ference might not occur if more lagged values were included.

The difference between national and sample data becomes clear when the estimatedmagnitude of effects (columns 3 and 4) are compared. For the two variables used byViscusi, the negative impact on injuries of a 10% increase in enforcement actions is eighttimes greater for inspections ( - .16% versus - 1.27%) and twice as great for penalty per




worker (- .04% versus - .09%) in the sample as in the national data. Substituting the

deterrence variables for Viscusis variables in the sample data set produces more balancedestimates of the impact of inspections with penalty (- .Sl%) and average penalty(- .66%).

Finally, comparing the impact of these variables as estimated by the Viscusi and be-havioral model on the sample data set, we see that the behavioral model estimates aresomewhat greater (- .81% versus - 1.48% for inspections with penalty, - .66 versus- .93% for average penalty). However, the difference between these estimates is lessthan the difference between estimates on the two data sets using the Viscusi model,indicating that the more OSHA-relevant sample provides the primary difference in esti-mated impacts. The richer detail in our model, made possible by the plant-level informa-tion in our data set, accounts for a smaller portion of the difference in estimated impacts.

Regardless of which model one uses on the sample data set, a 10% increase in eitherof the general deterrence variables is estimated to decrease lost workday injuries by .66%to 1.48%. In actual terms, a 1% reduction in injuries for our sample translates to 1.9injuries per plant, one injury per 2500 workers, or 1300 injuries per year. A 10% increasein enforcement in 1979, the first year of our sample period, would amount to 2800inspections, 950 inspections with penalty, or $1 million in penalties. These estimates areclearly not generalizable to the entire manufacturing sector, which would presumably beaffected at a lesser rate. The sample is important in its own right, however, since itaccounts for around 20% of the total work force in manufacturing and over 10% of allindustrial accidents.

Furthermore, the differences between plants in the sample and in the total manufac-turing sector (table 1) suggest three factors that may account for the greater effectivenessof enforcement in the sample. Our sample plants are larger, somewhat more dangerous,and more heavily inspected than the average manufacturing plant. Larger firms may bemore attentive to OSHA enforcement, since they are more visible. The concentration ofworkers makes an attractive target for inspections, particularly in larger firms with injuryrates above the average of manufacturing. Furthermore, larger firms may have more

adequate managerial and investment resources to control the risks that are brought totheir attention during inspections. Cook and Gautschi (1981) found size to be related toenforcement effectiveness; significant effects of inspections were found only for firmswith over 200 employees, with even greater effects for firms with over 300 employees.

Finally, the relatively high inspection probability in our sample (over 25% annually,compared with less than 8% for all firms nationally) may be important to increase thesensitivity of a firm to enforcement activities, since more intense scrutiny of criticalsubgroups within the industry may enhance enforcement effectiveness (Scholz, 1984a). Iffrequency of inspection is important, plant size becomes doubly important, becauselarger plants allow OSHA to inspect each firm more frequently for a given ratio ofenforcement resources per worker. The sample accounted for a 20% share of employ-ment in manufacturing in 1979, but for only 12% of all inspections in manufacturing,despite the unusually high frequency of inspections. Multiple inspections of a plant




during a given year appear to be a less effective concentration of resources-in explor-

atory anaIyses, we found muhiple inspections to have no additional impact on injuryrates. However, some minimal frequency of inspection is likely to prove critical to effec-tive enforcement.

7. Conclusions

We find evidence that OSHA enforcement has a significant impact on injuries in asubstantial portion of the manufacturing sector. The number of lost workdays and thenumber of Iost workday injuries decrease significantly after increases in general enforce-ment and after specific contacts with enforcement agencies, even though (1) complianceis only indirectly related to accidents (Mendeloff, 1979), (2) expenditures on compliancemay compete with more productive expenditures to improve safety (Bartel and Thomas,198.Q (3) OSHA resources do not permit extensive monitoring of most workplaces(Smith, 1976), and (4) OSHA penalties are relatively small compared to compliancecosts. Enforcement effects are relatively modest, as other studies have found; a 10%increase in enforcement would reduce injuries by around 1% for the large, frequentlyinspected firms represented in our sample.

In explaining how firms deal with risks associated with industrial accidents and OSHAenforcement, several extensions to the expected utility model were used, based on thebehavioral theory of the firm. These extensions, including the self-correcting feedbackmechanisms that focus managerial attention on risks, the relatively long lags betweenenforcement changes and changes in injury risks, the difference in effects between theexpected probability and the expected amount of a penalty, and the independent (spe-cific deterrence) effect of inspections, all proved significant in our model. Although thebehavioral hypotheses we have developed and tested fall considerably short of a well-developed theory of compliance, they suggest that further extensions to the basic ex-pected utility model of deterrence can contribute to an understanding of how firms

respond to accident and enforcement risks.A richer model of compliance may help improve the effectiveness and efficiency ofenforcement strategies. Just as undue reliance on simple microeconomic models in otherpolicy domains has led to myopic policy advice (Stern, 1985), reliance on unduly simpledeterrence models limit the enforcement debate to a relatively narrow spectrum of thepractical concerns facing enforcement officials. For example, given OSHAs normal levelof activities and the responses of our sample firms, our results suggest that increasing thenumber of penalties has about a 50% greater effect in reducing accidents than a compa-rable percentage increase in the average amount of penalty. This implies that OSHAmight increase its impact by shifting its resources from the more intensive inspectionsthat are required in order to impose high penalties to more frequent inspections thatimposed some penalty. Further research may be able to clarify threshold levels for thefrequency of inspections and the amount of penalty that could be used to increase theeffective deployment of enforcement resources.




The efficiency of safety standards in reducing accidents, while important, may be less

important than the need to focus the firms attention on safety problems. We cannot saywhether the decreases in accidents found in this study are efficient (from societys pointof view), but to the extent that the firms safety expenditures were suboptimal because ofinattention, the decreases in accidents may be due more to focusing the firms efforts onan effective risk-reduction program than to safety improvements directly related to com-pliance with regulations. If the behavioral model proves to be as powerful as thesepreliminary results suggest, then the conventional wisdom on the role of regulation andenforcement in the economy will need to be reevaluated to include this attention-correction function (Scholz, 1984b).

Notes

1. Viscusi (1986b) argues that higher penalties for repeated violations may provide an alternative reason forfirms to make investments in compliance (and safety) after an inspection that they would not have beenwilling to make before being inspected.

2. Those cases where it was not clear whether the records were properly matched were hand-checked.Hand-checking, used to examine the matches on two state samples, ndicated that our error rates for falsematches and missed matches were less than 1%. To ensure that no plant in the final set contained

ambiguous matches, 198 plants were dropped from the original file.3. For purposes of comparison, we note the distinctions that led us to use a different formulation than that ofViscusi (1979,1986a). First, our fixed effects efer to plant rather than industry-specific characteristics, sowe use the first-differenced form to control plant-level effects n addition to the industry and year dummyvariables used by Viscusi. Second, we use the measure of inspections with penalty, rather than just inspec-tions, because it corresponds to the theoretical concern with the probabilityofbeing penalized (see note 4).Third, as noted in the text, endogeneity between inspectionswith penalty and injury rates would be a majorproblem if we did not use the change form for the dependent variable. The alternative of using instrumen-tal variables to predict enforcement measures would not have allowed us to distinguish between generaldeterrence (the predicted enforcement) and specific deterrence (the actual event), and becomes morecomplex because of the laged variables. Finally, using the lagged dependent variable as an independentvariable (rather than as part o f the dependent variable) would cause problems in our plant-level data set,because we assume (and demonstrate) that the estimation error follows an autoregressive process. OLSwould then overestimate the coefficient of the lagged dependent variable and underestimate the remainingcoefficients of independent variables (Ostrom, 1978). In addition, the autoregressive process representsour theoretical interest in feedback on the change variable, so we use the change form and model theautoregressive process instead of using the lagged dependent variable.

4. For theoretical and empirical reasons,we focus on the occurrence of a penalty rather than on an inspectionor on the amount of penalty. On theoretical grounds, inspections that do not result in penalties are notlikely to focus additional managerial attention on accidents. In fac t, they may lead to more complacencyamong management toward safety although good inspectors presumably also point out potential safetyproblems that management may have overlooked. From the viewpoint of behavioral decision theory,managerial attention will increase only if the amount of penalty is higher than some threshold amount. Wetested the hypothesis that any penalty is sufficient to attract attention. See Klepper and Nagin (1989) for asimilar conclusion about criminal penalties and tax evasion. Empirically, preliminary analyses found thatinspections without penalty had less robust effects han inspections with penalties, and that the amount o fpenalty did not explain any more variance than the dummy variable representing inspections with penalties.The single indicator was selected to avoid multicollinearity problems, and also because of the straightforward




interpretation of the coefficient for the dummy variable. It should be noted that inspections without penaltymay have more significant effectson firms inspected less frequently than those in our sample.

5. The results are not sensitive to varying the lag lengths from l-4 years on either the autoregressive processor the enforcement variables.

6. Estimates of our general model were run using two-digit SIC industry aggregates in place of the predictedvalues. The results indicated a significant effect only for the current measure of the number of penalties.We believe that the more significant results obtained in table 4 for predicted values indicate the greateramount of information that is contained in the predicted values, which therefore represent a closer proxyto the firms expectations. The use of predicted or aggregate proxies for the test of general deterrence hadalmost no effect on the coefficients or significance of the specific deterrence measures.

7. These predicted values are not instrumental variables in the usual sense, since we include the actualenforcement experience of the plant in our measures of specific deterrence. They are intended to measurethe firms expectation of the OSHA enforcement that it will face (general deterrence). Consequently, we donot adjust the error of estimation.

8. Viscusi (1979) tested for lags up to three-period, but found no effects.The differences between his findingsand ours are probably due to the nature of our data set, which includes firms more likely to be impacted byOSHA enforcement and which utilizes firm-level data.

9. For a discussion of the differences between additive and multiplicative theories of risky decisions asapplied to the compliance domain, see Casey and Scholz (1989). Viscusi suggests that the disparity ineffectsbetween amount of penalty and inspections (with or without penalty) may be due to costsother thanpenalties associated with inspections.

References

Bartel, Ann, and Lacy G. Thomas. (1985). Direct and Indirect Effectsof OSHA Regulation,JoumalofLawand Econo mics 28,l-26.

Becker, Gary S. (1968). Crime and Punishment: An Economic Approach,Journal of Politica l Economy 76,169-217.

Casey, Jeffrey,and John T. Scholz. (1990). Boundary Effects of Vague Risk Information on Taxpayer Deci-sions, forthcoming inOrganizational Behavior and Human Decision Processes.

Cook, William N., and Frederick H. Gautschi III. (1981). OSHA, Plant SafetyPrograms, and Injury Reduc-tion, Industrial Relations 20, 245-257.

Qert, Richard, and James G. March. (1963).A Behavioral Theoryof the Firm. Englewood Cliffs,NJ: PrenticeHall.

Downing, Paul B., and Kenneth Hanf (eds.).(1983). International Com parisons of Pollu tion Enforcement.Boston: Kluwer-Nijhoff Publishing.

Erlich, Isaac (1973). Participation in Illegitimate Activities: A Theoretical and Empirical Investigation,Journal of Political Economy 81,521-565.

Fellegi, Ivan l?, and Alan B. Sunter. (1969). A Theory of Record Linkage,Journal of the American StatisticalAssociation 64,1183-1210.

Gray, Wayne B. (1987). OSHA MIS Matching Techniques, mimeo.Kahneman, Daniel, Paul Slavic, andAmos Tversky. 1982). Judgement under Uncertainty: Heuristics and Biases.

New York: Cambridge.

Klepper, Steven, and Daniel Nagin. (1989). Tax Compliance and Perceptions of the Risks of Detection andCriminal Prosecution, Law andSociety Review 23,209-240.

Lempert, Richard. (1982). Organizing for Deterrence, Lnwand Society Review16,513-568.McCafIrey, David. (1983). An Assessment of OSHAs Recent Effects on Injury Rates, Journalof Human

Resources 18. 131-146.



OSHA ENFORCEMENT AND WORKPLACE INJURIES 30.5

McCaffrey, David.(1982). OSHA and the Politics of Health Regulation. NewYork: Plenum Press.Mendeloff, John. (1979).Regulating Safety: An Econom ic and Political Anaiysis of Occupational Safely and

Health Policy. Cambridge, MA: MIT Press.Nagin, Daniel. (1978). General Deterrence: A Review of the Empirical Evidence. In Al Blumstein, J. Cohen,

and Daniel Nagin (eds.), Deterrence and Incapacitation: Estimating the Effect of Criminal Sanctions onCrime Rates.Washington, DC: National Academy of Sciences.

Ostrom, Charles W. (1978). Time Series Analysis: Regression Techniqu es.Beverly Hills: Sage.Radner, Roy. (1975). A Behavioral Mode1 of Cost Reduction,Bell Journal of Econo mics 6;196-215.Ruser, John W., and Robert S. Smith. (1988). The Effect of OSHA Records Check Inspections on Reported

Occupational Injuries in Manufacturing Establishments,Journa l of Risk and Uncertuiniy1, 415-435.Schoemaker, Paul J. (1982). The Expected Utility Model: Its Variants, Purposes, Evidence and Limitations,

Journal of Econom ic Literature 20,529-563.

Scholz, John T. (1984a). Cooperation, Deterrence, and the Ecology of Regulatory Enforcement,Law andSociety Review 18, 601-646.

Scholz, John T. (1984b). Reliability, Responsiveness,and Regulatory Policy,PublicAdministration Review 44,145-153.

Scholz, John T. (1987). Bureaucratic Environments and Regulatory Enforcement Effectiveness:A Theory ofCooperative Enforcement, paper delivered at APSA Annual Meeting.

Smith, Robert S. (1979). The Impact of OSHA Inspections on Manufacturing Injury Rates,Journal ofHuman Resources14,145170.

Smith, Robert S. (1976). The Occupational Safeg and Health Act: Its Goals and its Achievements.Washington,DC: American Enterprise Institute.

Stern, Paul C. (1986). Blind Spots in Policy Analysis: What Economics Doesnt Say about Energy Use,Journal of Policy Analysis and Management 5,200-227.

Viscusi, W. Rip. (1979). The Impact of Occupational Safetyand Health Regulation,BeilJoumalofEconom-its 10,117-140.

Viscusi, W. Kip. (1986a). The Impact of Occupational Safety and Health Regulations, 1973-1983,RandJournal of Econo mics 17,567-580.

Viscusi, W. Rip. (1986b). Reforming OSHA Regulation of Workplace Risks. In L. Weiss and M. Klass (eds.),Regulatory Reform: What Actua lly Happened.Boston: Little, Brown and Co.

von Neuman, John, and Oskar Morgenstem. (1947). Theory of Games and Econom ic Behavior,2nd edition.Princeton: Princeton University.

65 OSHA Enforcement and Workplace Injuries a Behavioral Approach to Risk Assessment_tcm296-282256

Documents