Risk Analysis, Vol. 35, No. 10, 2015

DOI: 10.1111/risa.12364

Workers’ Risk Tolerance and Occupational Injuries Monica Galizzi∗ and Tommaso Tempesti

This study explores the relationship between individuals’ risk tolerance and occupational injuries. We analyze data from a national representative survey of U.S. workers that includes information about injuries, risk tolerance, cognitive and noncognitive attributes, and risky behaviors. We measure risk tolerance through questions regarding individuals’ willingness to gamble on their lifetime income. We estimate zero-inflated count models to assess the role played by such measures on workers’ recurrent injuries. We discuss some implications of our results for future research and occupational safety policies. Our results highlight the concurrent and changing role played by individual, work, and environmental factors in explaining recurrent incidents. They show that risk tolerance affects recurrent injuries, although not in the direction that proponents of the concept of proneness would expect. Our measure of risk aversion shows that individuals who are somewhat more risk tolerant have fewer recurrent injuries than those who are risk averse. But the estimated relationship is U-shaped, not monotonic and, therefore, not easy to predict. At the same time, we find that individuals’ “revealed risk preferences”—specific risky behaviors—are related to higher injury probabilities. Demanding working conditions, measures of socioeconomic status, health, and safety problems experienced by workers during their youth remain among the most important factors explaining the phenomena of recurrent injuries. So our results contribute also to the important debate about the relationship between health and socioeconomic status. KEY WORDS: Accident proneness; count data; determinants of health; occupational injuries; risk aversion; risk tolerance

1. INTRODUCTION

sult in a work absence, restriction of work activity, or transfer to another job. Other cases may not involve any lost days. The most recent BLS estimates for 2013 reported incidence rates of 3.5 nonfatal injuries and illnesses per 100 equivalent full-time workers in all industries including state and local government. Among these, 1.8 cases referred to serious incidents that involved days away from work, job transfers, or restrictions.(2) While these numbers confirm the declining trend in the number of nonfatal occupational incidents reported by the BLS since 1992, they hide an important fact: the high costs that such injuries cause. For 2011 the National Safety Council estimated that the cost of work injuries amounted to $188.9 billion.(3) These expenses were carried by workers, employers, and third parties in terms of wage and lost productivity, medical expenses, administrative expense, and employers’ uninsured costs.

The purpose of this study is to assess the relationship existing between workers’ risk tolerance and a specific labor market outcome: occupational injuries. According to the U.S. Bureau of Labor Statistics (BLS), an “occupational injury is any wound or damage to the body resulting from an event in the work environment” and such event may have “either caused or contributed to the resulting condition or significantly aggravated a pre-existing condition.”(1) Some of these cases may be fatal. Some may reDepartment of Economics, University of Massachusetts Lowell, MA, USA. ∗ Address correspondence to Monica Galizzi, Department of Economics, University of Massachusetts Lowell, One University Avenue, FA 302 J, Lowell, MA 01854, USA; tel: (978) 934-2790; fax: (978) 934-3071; [email protected].

1858

C 2015 Society for Risk Analysis 0272-4332/15/0100-1858$22.00/1 

Workers’ Risk Tolerance and Occupational Injuries

1859

These costs have increased over time. In fact, the latest Council estimates suggest a 12% increase in real terms over the last decade. While a rich body of epidemiological research has focused on the determinants of on-the-job incidents, the economic literature has mainly focused on the economic outcomes suffered by injured workers.(4–7) Few economic studies have explored the characteristics of workers who are being injured, and, among these, who suffer recurrent injuries or workers’ compensation claims.(8–10) To the best of our knowledge, no other studies have explored the potential relationship between the phenomenon of recurrent occupational injuries and individual propensity toward risk. Our focus is especially on recurrent injuries: in our study this refers to a worker getting injured more than once. These events are not necessarily due to repetitive tasks or exposures or to the same type of occupational injury,1 even though they certainly can be.(2) Among all injuries, whether or not job related, recurrent cases account for the majority of days in hospital care and of medical costs.(11,12) We study data from the U.S. National Longitudinal Survey of Youth 1979 (NLSY79). We find that the number of injuries follows a U-shape with respect to risk tolerance, with the most risk-tolerant workers having a similar injury rate to the most risk-averse category. Injuries are instead consistently positively and significantly related to different risky behaviors. Measures of socioeconomic status and health or job safety problems experienced by workers during their youth remain among the most important factors explaining the phenomena of recurrent injuries.

(generally between 20% and 50% depending on the study(10) ) ended up filing for additional workers’ compensation claims in the years following their originally reported occupational injury. Similar results are found also in studies that have made use of survey data.(10,13,14) This result is not new. In fact, almost 100 years ago, Greenwood and Woods(15) had already found that a relatively small proportion of individuals employed in a British factory accounted for the majority of incidents. They attributed this to workers’ personality characteristics. This finding is often considered the originator of the concept of incident proneness, that is, the tendency of some individuals to experience more incidents than otherwise identical people.(16) This concept came under attack during the 1970s. It was accused of leading to the attitude of “blaming the victim” that had developed within the occupational safety community. This attitude resulted in dismissal of the role played by dangerous working conditions in determining injuries.(17) Critics argued that the statistical evidence of incident proneness in large samples was very weak.(18) Furthermore, a valid test of incident proneness implied the ability to control for the exposure to risk, for unknown or nonobservable personal and nonpersonal factors, and for potential underreporting of injuries.(19) It was argued that given equal individual initial liability, an incident could alter the probability of subsequent injuries.(20) As a result of these arguments the research focus shifted toward the study of the role played by specific organizational characteristics. These characteristics included hazardous working conditions,(21) safety practices,(14) managers’ own behaviors and practices toward safety,(22,23) financial resources, use of overtime,(21) shifts, subcontracting,(13) and temporary jobs,(24) as well as the degree of workload and job security. These occupational features have been found to play both a direct and indirect role in the probability of work incidents. In fact, challenging working conditions or temporary jobs can lead to workers’ stress,(25,26) increased effort to secure rehiring, or lower investment in human capital (including safety training) by companies. This can increase employees’ vulnerability to additional occupational injuries.(13,14,24,26) During the last few years, however, the concept of incident proneness has witnessed a resurgence. This is partly due to the accumulating evidence that some individuals experience more injuries than we would expect based on a purely random distribution of events.(16) At the same time, the research

2. LITERATURE REVIEW 2.1. “Accident Proneness” Versus “Accident Liability”2 The majority of studies that have examined the phenomenon of recurrent occupational injuries have made use of workers’ compensation data. They have found that a large percentage of claimants 1 According

to the BLS, the leading causes of nonfatal occupational injuries that require time off work are overexertion and bodily reaction, contacts with objects, and falls. Sprains, strains, and tears account for 40% of total cases.(95) 2 Despite the common use of the word accident in the majority of studies we are reviewing, in our analysis we are only using the term incident or injury to be consistent with the recommendation of the medical and epidemiological research community.(96)

1860 community has increased its ability to collect data with richer individual information and measures of personality traits. This facilitates the task of identifying which attributes—innate or learned—may render some individuals more likely to become involved in incidents than others. In this sense, despite the continuous use of the term proneness, most new studies fall under the study of “accident liability.”(19) In this context, environmental, organizational, and human attributes are seen as complementary,(13) potentially changing over time interactively.(20) For example, recurrent injuries are related to physical disabilities(14) and poor mental health.(27) Individuals with higher scores of self-reported cognitive failures (in perception, memory, and motor function) are also more likely to experience recurrent workplace incidents. This happens when these individuals are put under job-related stress.(28) Psychological stress can be caused not only by occupational stressors,(13,26) but also by family stressors such as poor housing conditions,(13) changes in family members’ health, and unemployment.(29) In this context, cultural norms may also play a large role in affecting individuals’ responses to occupational safety and experiences with injuries.(30) For example, men may experience more occupational injuries not only because of the types of jobs and industry that employ them,(31) but also because of their dismissal of safety practices due to their enacting “working-class masculinity,”(32,33) or their tolerance for risk.(34,35) It has also been established that work injuries are related to risky behaviors such as alcohol,(36) tobacco,(14,37) and narcotics abuse.(38,39) To the best of our knowledge, however, no study has explicitly tested whether differences in risk tolerance are reflected in different likelihood of occupational incidents. 2.2. Workers and Their Attitudes Toward Risk The economic literature has studied the hypothesis of workers’ heterogeneous preferences toward occupational risk in the context of the theory of compensating wage differentials. Such theory predicts that in a competitive environment where workers have complete information about risks and where moves across jobs are costless, employees who accept more dangerous jobs will be compensated with extra pay. Most of the empirical research on this topic has focused on proving the existence of such compensating wage differentials. However, very little research has been conducted to assess whether workers’ differences in risk aversion result indeed

Galizzi and Tempesti in experiencing different occupational incidents. This lack of analysis is largely due to the difficulty of measuring workers’ risk aversion. In our study, we analyze the relationship between the likelihood of recurrent work injuries and a measure of risk tolerance, originally developed by Barsky et al.,(40) through questions regarding individuals’ willingness to gamble on their lifetime income. Depending on which (hypothetical) gambles a respondent is willing to accept or reject, we can categorize respondents in four categories of increasing risk tolerance (decreasing risk aversion). The lottery questions also allow us to distinguish the role of beliefs from the role of risk aversion in a respondent’s choices. Indeed, someone may engage in a behavior that is considered risky either because she is more tolerant of risk, or because she is less aware of the consequences of such behavior. By explicitly specifying the probability distribution of the uncertain outcomes, the lottery questions allow us to keep beliefs constant and so derive a better measure of risk aversion.(41)3 Some studies show that these lottery questions are able to predict risky behavior such as stockholding, self-employment, and timing to marriage.(40,42,43) 2.3. Risk and “Revealed Preferences” Risk measures based on lottery questions and self-rating scales both share a limitation. Individuals may report that they would choose A over B, although when faced with a real choice between A and B, they may end up opting for B. Economists have therefore also focused on a different approach. For example, Leigh(44) used a composed measure of “risk avoidance.” This was constructed around nine different observed workers’ choices about car maintenance, car and health insurance, use of seat belts, smoking, and personal savings. Such a measure was designed to capture risk and time preferences. He found it to be negatively related to the probability of taking risky jobs, but also determined by family economic background and race. Viscusi and Hersch(45) measured risk tolerance toward occupational health using the preferences that individuals reveal by undertaking one specific type of risky behavior: smoking. Their findings indicate that smokers—and therefore individuals with higher tolerance toward 3 However,

in Section 5 below we discuss how the lottery questions may still not be able to fully control for the role of beliefs in a respondent’s willingness to take risk.

Workers’ Risk Tolerance and Occupational Injuries

1861

health risks—are indeed also more prone to suffer occupational injuries. Smokers also face worse labor market opportunities, which lead them to lower earnings compared to nonsmokers. Deleire and Levy(31) and Grazier and Sloane(46) follow yet another approach when looking for proxies of risk aversion. They use workers’ family structure as an exogenous determinant of individuals’ aversion to risk. Their results show that single parents and married women with children are indeed the most averse to risk as suggested by their choice of safer jobs. Differences in occupational hazard also explain part of the occupational gender segregation, with female workers concentrated in safer jobs.4 In our study, we are contributing also to this body of literature by testing the relationship between risky behaviors and the likelihood of recurrent work injuries after controlling for different job characteristics.

juries in the workplace. For this reason, we limit our analysis to a subset of 6,680 individuals who worked for at least one week between 1986 and 2000, answered the risk tolerance questions in 1993, and who participated in all survey rounds from 1988 to 2000 (the period when workplace injury information was collected).6 Among this subgroup, 33.5% of individuals (n = 2,237) experienced at least one occupational injury, for a total of 3,747 incidents. Among those who reported at least one injury, 40% (n = 901) reported two or more occupational incidents during the surveyed years (Table I).

3. THE DATA This study makes use of data contained in the NLSY79, a nationally representative panel survey of U.S. workers sponsored by the Bureau of Labor Statistics (BLS), U.S. Department of Labor. A total of 12,686 men and women were first interviewed in 1979 about their employment experiences when they were between the ages of 14 and 22. Between 1988 and 2000, the survey also collected information about job-related injuries and illnesses.5 From 1988 to 2000, a total of 3,280 NLSY79 individuals reported that they suffered at least one occupational injury. In the years 1993, 2002, 2004, 2006, and 2010, individuals also answered questions designed to capture their risk tolerance. The survey also includes information about cognitive and noncognitive attributes and risky behaviors. Our analysis requires all the available information about potential recurrent occupational in4 These

authors study this effect of jobs’ physical risks on the Duncan index of dissimilarity. However, their analyses cannot explain whether risk affects segregation through personal preferences or discrimination. 5 In 1988, individuals were asked whether they had had an incident at any job that resulted in an injury or illness during the “past 12 months.” Answers indicated that only 5% of all the reported incidents had happened in 1987. In the following survey rounds, the same question referred to the time “since the last interview.” In our subsequent analysis we refer only to “injuries” and do not distinguish between injuries and illnesses because only 8% of all the recalled incidents resulted in illnesses. Such percentage is slightly higher than the 5.1% found by the BLS, according to which the most frequent of illnesses are skin disease, hearing losses, respiratory conditions, and poisoning.(2)

3.1. The Demographic, Occupational, and Personality Characteristics Table I describes some individual characteristics for our larger subsample as of 1988, the first year that respondents were asked about occupational injuries. Descriptive results are reported after being adjusted by survey weights that account for the representativeness of individuals of different races continuously surveyed between 1988 and 2000. Such weights allow the answers to be considered as national totals: 79% of workers were white, 15% were black, and 6% were Hispanic. In 1988, the typical respondent was 27 years old. As expected, those who sustained an injury were typically male and had less education and less tenure on the job. Workers who reported more injuries over the years were working also longer hours or not on fixed shifts as of 1988. We exploit the richness of the NLSY79 data to extract additional information about many potential determinants of injuries. The early rounds of the NLSY79 enable us to calculate whether any of the following conditions had happened before respondents turned 23 years old:7 a total family income above or below the poverty level; suffering any work limiting health conditions; and exposure to unsafe (unhealthy or dangerous) jobs at an early age.8 6 Richer details about the representativeness of the subsample are

contained in Ref. 10. calculating the variables capturing early socioeconomic and health status we chose to study records until age 22 because this was the highest age reported by the oldest surveyed individuals in 1979, the first round of the NLSY79. Also in our models we do not include age as a regressor because of its very limited variation in the NLSY79: in 1988 respondents’ age ranged between 23 and 32 years. 8 The respondent was asked whether her job was dangerous or unhealthy. If she responded yes to either question, then we create the dummy for “unsafe job.” 7 In

1862

Galizzi and Tempesti Table I. NLSY 1988–2000 Subsample of Workers as of 1988 (Weighted Summary Statistics) No Occupational Injury (n = 4,443) Mean

Only One Occupational Injury (n = 1,336) Mean

Recurrent Occupational Injuries (n = 901) Mean

Age White Black Hispanic Female

27 0.79 0.15 0.06 0.56

27 0.79 0.14 0.07 0.44

27 0.83 0.11 0.06 0.39

Cognitive skills Education < High school High school > High school AFQT

0.09 0.41 0.50 53

0.14 0.48 0.39 48

0.14 0.59 0.26 46

Noncognitive skills Rotter Rosenberg Before 1988 In poverty before age 23 Health limitation before age 23 Dangerous/unhealthy jobs before 1983

8 23

9 22

9 22

0.36 0.16 0.46

0.37 0.19 0.57

0.39 0.23 0.69

As of 1988 Married Children Tenure (weeks) Weekly worked hours Fixed shifts Collective bargaining

0.55 0.48 152 41 0.86 0.11

0.51 0.52 146 43 0.84 0.13

0.49 0.56 136 44 0.81 0.20

Risk categories 1993 1: Most risk averse 2: Moderately risk tolerant 3: Strongly risk tolerant 4: Very strongly risk tolerant

0.45 0.14 0.18 0.23

0.46 0.12 0.18 0.24

0.50 0.09 0.16 0.24

0.24 0.45

0.29 0.46

After first injury Lost wages Filed for workers’ compensation

Note: Individuals who participated in all survey rounds from 1988 to 2000 and who had some weeks of work during the same period and answered the risk tolerance questions in 1993.

We use years of education and the Armed Forces Qualification Test (AFQT) as a measure of cognitive skill. This is expressed as a percentile score.9 As to noncognitive skills, we use the Rotter scale of locus of control. The higher this scale, the lower the control the individual believes to have over her own life. We also use the Rosenberg Self-Esteem Scale from 1980. Respondents were asked 10 statements about 9 In

2006 the NLSY renormed the AFQT percentiles controlling for age. We use this renormed version of the AFQT. More details on the construction of the AFQT are available here: http://nlsinfo.org/content/cohorts/nlsy79/topical-guide/ education/aptitude-achievement-intelligence-scores#asvab, accessed January 9, 2015.

themselves. The higher the score, the higher their self-esteem. To facilitate interpretation of the coefficients, in our regressions we standardize the AFQT and Rotter and Rosenberg scales to have mean zero and standard deviation equal to one.10

10 See:

http://nlsinfo.org/content/cohorts/nlsy79/topical-guide/ attitudes (accessed January 9, 2015) for more information about the Rotter and Rosenberg scales. It is worth noting that the original Rotter scale was based on the answer to 60 questions while the NLSY retains only four of the original questions. While this scale has predictive power, for example, it correlates well with education, the internal consistency of the scale is quite low (Cronbach’s α is 0.36).

Workers’ Risk Tolerance and Occupational Injuries 3.2. The Measures of Risk Tolerance In 1993, each respondent was asked the following question: Suppose that you are the only income earner in the family, and you have a good job guaranteed to give you your current (family) income every year for life. You are given the opportunity to take a new and equally good job, with a 50–50 chance that it will double your (family) income and a 50–50 chance that it will cut your (family) income by a third. Would you take the new job?

If she answered “no” (or “yes”), she was then asked whether she would accept a job with a 50–50 chance that it will double her family income and a 50–50 chance that it will cut her family income by 20% (or 50%). So there are four possible patterns of response and we construct four categories of risk tolerance, 1–4, with 4 indicating the highest level of risk tolerance (Fig. 1). The most risk-averse respondents reject both lotteries while the most risk-tolerant ones accept both. Table I shows that in our NLSY79 subsample almost half of individuals report the lowest level of risk tolerance. In our analysis we focus on the risk-tolerance variable from 1993. This is the earliest risk-tolerance variable available. We use the risk variables from other years, 2002, 2004, 2006 and 2010,11 to confirm the robustness of our regressions. We also use another more general measure of risk aversion that the NLSY79 collected in 2010. That year the respondent was asked to rate herself on a 0–10 rating scale, where 0 means “unwilling to take any risks” and 10 means “fully prepared to take risks.” We used this question when asked in the specific domain of health choices. Finally, our last measures of risk aversion are based on information about risky behaviors. The assumption is that, other things equal, more risktolerant individuals will engage in more risky behaviors such as smoking or using drugs. In this regard, we use a set of dummy variables capturing respondents’ consumption of cigarettes, drugs, and alcohol. 4. REGRESSION ANALYSIS Our data cover workers who had experienced at least one work-related injury and workers who never experienced one. Therefore, the data allow us to study factors that may explain recurrent episodes of occupational injuries through a regression model

1863 in which the dependent variable is the expected number of occupational injuries for each individual over 13 years (between 1988 and 2000). We use information about individual and occupational characteristics as of 1988 as explanatory variables. Also, because occupational injuries can only occur when workers are employed, our model accounted for different “exposure” in terms of the number of weeks during which each person was working during the survey years 1987 and 2000. Our dependent variable—the number of workrelated injuries—assumes values bounded from below by zero. Moreover, there is a preponderance of zero counts as the majority of the surveyed individuals did not report any occupational incidents (Table I) and it is likely that recurrent incidents are correlated (“overdispersion”). An appropriate model in our case is the zero-inflated count model.(10) Such a model estimates simultaneously a binary probability model (a logit) to determine whether a zero or a nonzero outcome occurs, and a count model (the negative binomial) to estimate what predicts the frequency of positive outcomes. 4.1. The Risk/Job Lottery Measure and the Recurrence of Occupational Injuries We start with a basic model where our covariates include dummies for each level of risk tolerance, a female dummy, and race/ethnicity dummies.12 The results are shown in two columns under Model 1 (Table II). In the left column are the results of the logit model where the outcome of interest is whether a worker has zero injuries or not (zero injuries constitute a “success,” i.e., are coded with 1 in this logit model). The positive, albeit insignificant, coefficient for female suggests that women have a higher chance of not being injured (i.e., a higher chance of being in the zero injury group). Race dummies are not significant. Finally, risk tolerance does not affect the probability of not being injured. As to the count model, the effect of gender is in line with the logit results: conditional on having a chance of getting injured, women have fewer injuries. Blacks have a lower expected number of injuries. As to our regressor of interest, workers with a moderate risk tolerance have significantly lower counts of 12 Following

11 The question asked in 2010 was rephrased to address the concern

about “status quo bias.”(40)

Chapter 2 of the NLSY79 User Guide,(97) we do not weigh our regressions, but instead use dummy variables for the black and Hispanic oversamples.

1864

Galizzi and Tempesti

“Are you willing to keep your current income or accept a loƩery with 50% chance that income is doubled and a 50% chance that you lose…” Most risk averse 1/5 of income?

No

Yes

Fig. 1. Risk-tolerance levels based on NLSY79 1993 lottery question.

No 1/3 Yes of income?

No

1/2 of income? Yes

Most risk tolerant

injuries than those who are most risk averse (again, conditional on having a chance of getting injured). However, this result is not monotonic, as workers with higher risk tolerance do not have significantly lower counts of injuries than those who are most risk averse. We also graph the predicted injury counts by risk-tolerance levels in Fig. 2 (top panel). Notice that this prediction incorporates both the results of the logit model and of the count model.13 From the graph it appears that the number of injuries follows a U-shape with respect to risk tolerance. As we will see, this pattern is very robust to other specifications. In Model 2 (Table II) we also control for cognitive and noncognitive skills by adding dummies for completing high school, more schooling than high school, the standardized AFQT, the standardized Rotter scale of control, and the standardized Rosenberg scale of self-esteem. Relative to workers 13 According

to the Stata 12 Manual for the zinb command, the predicted number of events for individual j is given by (1 − p_j)(exp(x_j b) × exposure_j), where p_j is the predicted probability of a zero outcome and exposure is the number of units of time j has been exposed to the risk of injury (in our case, the number of weeks worked). The means in Fig. 2 are computed using the command “margins.” This command sets the risk-tolerance level to category 1 for all workers, leaving the other x_j variables unchanged. Then it computes the predicted injury count for each worker according to the formula above and finally it averages these counts across all workers. This gives the predicted count of injury for risk category 1 graphed in Fig. 2. The predicted counts for other risk categories are computed analogously.

with less than a high school diploma, more educated workers have a higher chance of having zero injuries (though these coefficients are not significant) and a lower injury count. AFQT has a significant impact on the chance of not getting injured, but no impact on the injury count. The locus-of-control and selfesteem scales have no effect on either the odds of not being injured or the injury counts. Relative to Model 1, the female dummy has a moderately significant negative effect on the odds of not getting injured (and so a positive effect on the odds of being injured). However, the coefficient on the female dummy in the count model increases in magnitude. This second effect dominates the first one so that the predicted number of injuries for female is 0.48 versus 0.65 for males, with the difference being significant at the 5% level.14 As previously found in Fig. 2, the predicted number of injuries still follows a U-shape with respect to risk tolerance. As in our previous study,(10) in Model 3 (Table II) we also add dummy variables for having experienced the following conditions before age 23: work-limiting health conditions, living in a household in poverty, and working in a unsafe job.15 These covariates try to control for workers’ early experiences that may be correlated with their risk aversion, their 14 These predictions are obtained with the Stata margins command.

The significance test is obtained with the Stata command margins, dydx(female). 15 We have added a dummy that is equal to 1 if the information about early job safety is missing and 0 otherwise.

Workers’ Risk Tolerance and Occupational Injuries

1865

Table II. Count Models of Recurrent Injuries Model 1 No Injury

Count

Model 2 No Injury

Count

Most risk averse Moderately risk tolerant Strongly risk tolerant Very strongly risk tolerant Female

0.80** 0.18 −0.79 14.08

−0.25*** −0.09 −0.09 −0.13*

−0.34 0.16 −0.09 −0.67*

−0.24*** −0.03 −0.02 −0.33***

Count

No Injury

Count

−0.23 0 −0.05 −0.05

−0.22** −0.04 −0.03 −0.28***

−0.56 −0.19 −0.09 −1.35***

−0.24*** −0.09 −0.05 −0.20***

−0.18** −0.24***

−3.37 0.38

−0.15** −0.17**

Reference group 0.2 −1.3

0.03 −0.16***

−0.85 −0.09

−0.16** −0.19***

Less than high school High school More than high school Standardized AFQT Standardized Rotter Standardized Rosenberg Health issues before age 23 In poverty before age 23 Unsafe job before 1983 Missing unsafe job dummy Tenure (weeks) Collective bargaining Weekly worked hours Missing hours dummy Fixed shifts Occupation dummies Industry dummies Constant Observations

No Injury

Model 4

Reference group

White Hispanic Black

Model 3

−0.88 −0.2

Reference group −0.15** −0.45*** −0.03 0.02 0.02

7.73 20.96 0.71** −0.25 0.16

−15.45

−6.69***

−22.42

6,680

−6.34*** 6,254

12.38 13.7 0.45 −0.32** 0.09 −0.86 −0.17 −0.5 −0.34

−14.48

−0.08 −0.36*** 0 −0.01 0.03 0.27*** 0.12** 0.23*** 0.13

−6.60*** 6,254

−0.67 −0.08 −0.65 −0.43*** 0.35 0.03 −0.1 0 0.2 0.03 −0.07 0.28*** −0.75** 0 −0.35 0.21*** 0.39 0.05 0 −0.00*** −0.98* 0.12* ** −0.04 0.00** 0.52 0.31*** 0.24 −0.18*** Yes Yes Yes Yes −12.82 −7.12*** a 5,361

Note: *p < 0.10, **p < 0.05, ***p < 0.01. Two-sided tests. Results from zero-inflated binomial models with a logit for the probability of no injury (no injury = 1, at least one injury = 0) and a negative binomial for the count of injuries. We include two dummy variables to capture cases where the unsafe job dummy and weekly worked hours were missing. a The number of observations drops mainly because of missing values in the occupation and industry dummies.

occupational choices, and with their injuries. All of these dummies have the expected sign in both the logit and the count models, but only the coefficients in the count model are significant. For instance, having early health issues reduces (nonsignificantly) the chances of not getting injured and significantly increases the count of injuries (conditional on having a chance of getting injured). The predicted number of injuries still follows a U-shape with respect to risk tolerance and the results for the other covariates are similar to Model 2. Our models so far do not account for the characteristics of the workers’ jobs. This is because risk tolerance may affect the job, industry, and occupation

in which a worker is employed. For example, workers with a higher risk tolerance may end up working in jobs with a higher risk of injury. If one includes job characteristics among the covariates, then this possible effect of risk tolerance is partialled out and the coefficients on the risk-tolerance dummies may underestimate the effect of risk tolerance on injuries. In Model 4 (Table II) we present the results of the zeroinflated count model when controlling for the characteristics of the worker’s job in 1988, industry dummies and occupation dummies.16 We use aggregated 16 We

impute the missing values for the weekly worked hours variable using the mean of the variable for the nonmissing values. As

1866

Galizzi and Tempesti

Fig. 2. Predicted counts by risk-tolerance levels Note: The predictions are obtained from Model 1 (top panel) and 4 (bottom panel) in Table II, which use the zero-inflated count model for count of injuries.

one-digit occupations and industries according to the 1970 Census classification.17 We find that working fixed shifts does reduce the injury count while working longer hours increases it. Having a contract regulated by a collective bargaining agreement is associated with an increase in the injury count, possibly reflecting the fact that workers in riskier jobs tend to unionize more.(47) Unionized workers may also feel more comfortable voicing their safety concerns or reporting their injuries.(48) As to the results for the occupation and industry dummies, the two riskiest occupations are operatives and laborers and the two riskiest industries are public administration and minregressor, we also add a dummy that is equal to 1 when weekly hours are missing in the original data. The sample size is smaller than in the other models due to missing values on the occupation and industry dummies, which we do not impute. 17 The occupations are: Professional and Technical Workers; Managers and Administrators, except Farm; Sales Workers; Clerical and Unskilled Workers; Craftsmen Workers; Operatives; Laborers; Service Workers. The industries are: Agriculture; Mining and Construction; Durable Manufacturing; Nondurable Manufacturing; Transportation, Communication, and Public Utilities; Wholesale and Retail; Finance, Insurance, and Real Estate; Business, Repair, Personal, Entertainment, and Recreation Services; Professional Services; Public Administration. We use one-digit occupation and industries dummies because the zero-inflated negative binomial model did not converge when we use the many dummies at the two-digit level. However, we also tried a simple OLS regression of injuries on all the regressors in Model 4 but using two-digit occupation and industry dummies. We still obtained the same results as those reported in Table II.

ing and construction.18 The pattern of predicted injuries relative to risk-tolerance levels is represented in Fig. 2 (bottom) where again a U-shape occurs. Even though our focus is not on compensating wage differentials, we have added the log of hourly wage to the regressors of the model with job covariates (Model 4 of Table II). We still find a U-shaped pattern of injuries relative to risk tolerance. The estimated coefficient on the log of wage is negative and significant, indicating that higher wage is correlated with lower injuries. This result is not uncommon in the literature.(49) It is consistent with the hypothesis of safety being a normal good. Therefore, we run the same regression as above but replace wage with family income as a regressor: our results do not change. In addition, in the United States, workers’ compensation benefits vary by state and, in theory, more generous workers’ compensation benefits could induce ex-ante moral hazard behavior(50) and affect our

18 Note

that these are the riskiest occupations and industries after the effect of other regressors is controlled for. These findings are consistent with general BLS injuries statistics.(2) Also note that many workers in the public administration industry are firemen or policemen. When we run the simple OLS regression of injuries on all the regressors for Model 4 but using two-digit occupation and industry dummies, we find that the riskiest occupations are textile operatives and protective service workers (firemen, policemen, etc.) and the riskiest industries are transportation services, local public administration, and machinery manufacturing.

Workers’ Risk Tolerance and Occupational Injuries estimated risk-tolerance coefficients.19 However, our results do not change when we include state dummies to account for possible systematic differences across U.S. states in injuries, workers’ compensation rules, and risk tolerance.20 Finally, Spivey(43) argues that married status may be itself a function of risk. So we have also rerun our regressions by adding a dummy for being married and also one for having children.21 We still find a U-shaped pattern of predicted injuries with respect to risk tolerance. Our results indicate that being married reduces the expected count of injuries while having a child increases it. The results for children are somewhat surprising because we would have expected more prudent behavior among people who had children.(31,51,52) A possible explanation may have to do with the fatigue that is associated with parenthood and with the positive relationship existing between fatigue and likelihood of occupational injuries.(53,54) A possible concern is that our risk-tolerance measure was collected in 1993 while our reported on-the-job injuries were recorded in the 1988–2000 period. An early injury may have affected attitudes toward risk in 1993, possibly biasing our estimates of the effect of risk tolerance on injuries. To lessen this concern, we analyze also a subsample of workers who had not experienced any injury before 1994. Obviously, the sample size is smaller; the range of the counts of injuries also becomes smaller (range 0–4). The regression results are shown in Table III.22 As can be seen in Fig. 3, the U-shape of injuries with respect to risk tolerance characterizes this subsample as well. The results highlight the significant role played by education and early socioeconomic and health factors in determining recurrent injuries. As an additional robustness check, we rerun our Model 3 (Table II) using as risk-tolerance

1867 Table III. Count Models of Recurrent Injuries, After 1993 Sample Zero Outcome

Count Model

Count of injuries after 1993 Most risk averse Moderately risk tolerant Strongly risk tolerant Very strongly risk tolerant Female White Hispanic Black Less than high school High school More than high school Standardized AFQT Health issues before age 23 In poverty before age 23 Unsafe job before 1983 Missing unsafe job dummy Constant Observations

Reference group −1.13 1.28 1.03 −13.36

−0.29 0 0.06 −0.27**

Reference group −1.49 0.84

0.03 −0.15

Reference group −19.83 0.05 0.9 −0.54 0.12 0.19 −0.54 −1.92

−0.25* −0.49*** 0.01 0.33*** 0.22** 0.17* 0.15 −7.07*** 4,724

Note: *p < 0.10, **p < 0.05, ***p < 0.01, two sided-tests. Results from zero-inflated binomial models with a logit for the probability of no injury (no injury = 1, at least one injury = 0) and a negative binomial for the count of injuries. We include one dummy variable to capture cases where the unsafe job dummy was missing.

19 For

a discussion of the effect of difference in workers’ compensation rules, see also Galizzi.(10) 20 We have also rerun our regressions in Table II with the svy Stata command to correct the standard errors for geographic clustered sampling (we thank an anonymous referee for this suggestion). Our results are robust also to this specification. 21 All the results about two-digit industries, state dummies, and marital status are available from the authors. 22 When including the variables for noncognitive skills, the model did not converge and so we omit them from the regression. We have also tried adding as a regressor each of these variables at a time. In this case the model did converge, and we still find a U-shaped pattern of predicted injuries relative to risk tolerance.

Fig. 3. Predicted counts by risk-tolerance levels for sample after 1993. Note: The predictions are obtained from the zero-inflated count model in Table III. The model is applied only to those workers who had no injury before 1994.

1868

Galizzi and Tempesti

Fig. 4. Predicted counts by risk-tolerance levels with respect to health. Note: The predictions are obtained from the zero-inflated count model. The independent variables are all the ones in Model 3 of Table II with the exceptions of the risk-tolerance measure. Instead of using the risk-tolerance variable based on lottery questions from 1993, here we use the self-reported scale of risk-taking attitude with regard to health. This scale was collected in 2010. It is a 0–10 scale, with higher values indicating higher risk tolerance. We use the levels of this scale in our regression.

measures the ones based on the lottery questions asked in 2002, 2004, 2006, and 2010. Again, the Ushape found above appears when using measures from these years.

4.2. Using Other Measures of Risk Tolerance Dohmen et al.(55) show that the lottery question we have used so far is a good predictor of financial risk-taking behavior (e.g., holding stocks), but not of risk-taking behavior in other contexts, such as healthrelated decisions. A possible reason for this is that attitude toward risk is context specific: some individuals may be more willing to take risks in a certain area, for example, financial decisions, but not in others, for example, health. Dohmen et al.(55) show that a more general measure of risk tolerance (a 0–10 self-rating scale described below) is instead a better predictor of risk-taking behavior. The NLSY79 data contains also these 0–10 scales about self-reported willingness to take risks that were collected in 2010. We use them in some of our specifications. We keep all the regressors from Model 3 in Table II substituting the risk-tolerance measure from 1993, first with the general risk-taking scale from 2010, and then with the health-specific risk-taking scale, also collected in 2010. We categorize these scales by creating 10 dummies for each level. None of the coefficients on these dummies is significant and so we do not find any clear association between these last risk-taking scales and the number of injuries. The results for health choices

are graphically represented in Fig. 4. The results for the general scale are also very similar. We also follow the “revealed preference” approach and use as measures of risk tolerance the following variables: a dummy for smoking at least 100 cigarettes before 1992; a dummy for having used cocaine at least once by 1988; a dummy for having used marijuana at least once by 1988; a dummy for having had six or more drinks at least once in the month before the 1988 interview. In this case, the zero-inflated count model has a variance matrix that is nonsymmetrical or highly singular. Even if the model converges, the standard errors of the coefficients cannot be computed. This outcome is often found in this type of model.(56) For this reason, we use an ordinary least squares (OLS) regression. We use the number of injuries in the 1988–2000 period as dependent variables. We use all the regressors from Model 3 in Table II but substituting the risk-tolerance measure from 1993 with each of the dummy for risky behavior in turn (for a total of four different OLS regressions). We also add the exposure variable (weeks worked from 1987 to 2000) to the regressors. According to the regression results, cigarette, marijuana, and cocaine use are all associated with a higher count of injuries at the 1% significance level. High consumption of alcohol is significantly associated with a higher count of injuries at the 10% significance level. Fig. 5 graphically represents the results from using each of these variables in turn in our OLS regressions.

Workers’ Risk Tolerance and Occupational Injuries

1869

Fig. 5. Predicted counts by risky behaviors. Note: Each graph is from a different regression. The predictions are obtained from a linear regression model with count of injuries as the dependent variable. The independent variables are all the ones in Model 3 of Table II with the exceptions that: (a) the self-reported risky behavior dummy is used instead of the self-reported risk-tolerance variable; (b) the exposure variable (i.e., weeks worked between 1987 and 2000) is also included among the independent variables. Ninety-five percent confidence intervals are reported.

5. DISCUSSION 5.1. The Contribution of Our Study The goal of this study was to analyze the phenomenon of recurrent on-the-job injuries by focusing on the role played by individual risk aversion. To the best of our knowledge, this represents the first analysis on the topic that makes use of a survey measure that was specifically designed to assess individuals’ risk tolerance. As in Galizzi,(10) this is also one of the few studies that explore the problem of recurrent work injuries by making use of a national sample of individuals, by following individuals over more than 10 years, and by comparing individuals who suffered occupational injuries with others who did not. This is a key feature that is often missing in most of the analyses that have focused on the phenomenon of recurrent injuries and have used workers’ compensation administrative or only injured workers’ survey data. In addition, compared to previous studies on the topic, ours has two additional strengths: collection of our longitudinal data started when most of the surveyed individuals were teens. Therefore, the available information also captures workers’ early socioeconomic, health, and employment statuses. So our results contribute also to the important debate about the relationship between health and socioeconomic status.(57,58) Also, the richness of the original data permits us to analyze our specific labor market outcomes—injuries—as if

determined not only by a variety of measures of risk aversion (including measures of “revealed” preferences), but also by workers’ cognitive and noncognitive abilities. These last variables are becoming increasingly important in labor economics research.(59)

5.2. The Role of Risk Tolerance and Other Personal Traits We study the occupational injuries reported by surveyed workers between 1988 and 2000. Our main focus is the role played by a measure of risk tolerance that was collected in 1993: a job/lottery question. We find that this measure affects significantly the count of injuries. However, the pattern of the estimated parameter is not monotonic and assumes a U-shape that is consistent across all our different model specifications and a variety of robustness checks. These include estimating the effect of risk propensity only among those individuals who experienced a first work incident in the years that followed the first survey round when the risk-tolerance question was asked. Therefore, our result about risk tolerance cannot be used to support the controversial hypothesis of injury proneness. If anything, the workers who are somewhat more risk tolerant than the most risk-averse ones are characterized by safer work experiences.

1870 Even though we include a vast array of controls, it is possible that our estimated nonmonotonic effect of risk tolerance on injury is due to individual heterogeneity that we are not able to measure. We know from different theories of risk perception that attitudes toward risk can be affected by cognitive factors (familiarity and dread of risk(60) ), by an ongoing interplay between analytic reasoning and emotions,(61) by social and cultural influences,(62,63) and by previous exposure to risk.(35) Our data do not permit a way to control for such a variety of potential explanatory factors. For example, our risk-tolerance scale may also capture a person’s pessimism and so the answers to the lottery questions may actually distinguish three groups of respondents. The first group could be composed by the most risk-averse individuals23 whose refusal to take any risk may be due to their pessimism about their prospects.24 The second group, the moderately risk tolerant (category 2 in our Fig. 2), may be composed by those individuals who are partly more optimistic about their prospects and so are more comfortable with taking some risk. The third group may be the most risk-tolerant group (categories 3 and 4 in our Fig. 2) that indeed tends to take more risks. The pessimism of the first group may actually be justified if these individuals correctly anticipate that they are more likely to experience unfortunate events, such as injuries. For example, a pessimistic answer may actually proxy some features of a worker’s environment that make injuries more likely but that our other regressors do not capture.25 Indeed, studies have shown that individuals who experience negative emotions are at more risk of onthe-job injuries.(64) Analogously, the higher risk tolerance of the moderately risk averse may capture the fact that their environment is less likely to lead to unfortunate events such as injuries. Finally, injuries 23 This

is the top group in Fig. 1 and category 1 in Figs. 2 and 3. lottery questions specify the probabilities of each outcome. As we have argued in Section 2.2, if respondents take these probabilities at face value, then different respondents will give different answers to the lottery questions because of their different risk tolerance but not because of differences in their beliefs about the probability of a risky outcome. However, a pessimist subjectively revises upwards the unfavorable outcome’s probability that the lottery question objectively specifies (50% in our particular case; see Fig. 1).(98,99) If this is the case, the lottery questions do not allow us to fully disentangle the role of risk tolerance from the role of beliefs in a respondent’s answers. 25 We use the term “environment” in a broad sense here to refer to not only an individual’s community and physical environment but also to her workplace and personal circumstances.(100) 24 The

Galizzi and Tempesti may increase for the most risk-tolerant individuals because, whether they are optimists or not, they tend indeed to take more risks. While the above scenario is merely speculative, we have regressed our measures of the respondent’s control over her own life (the Rotter scale) and of self-esteem (the Rosenberg scale) on the risktolerance scale and we have found that the moderately risk averse (category 2 in our graphs) have the most control over their lives and the highest selfesteem, with these differences being significant.26 This area deserves further research but, at a minimum, our results suggest that studies making use of risk-tolerance measures similar to ours should allow for their effects to be not linear.27 We are also aware that our risk aversion measure may be subject to criticism. In fact, although other studies have found it to be predictive of other labor market and personal decisions,(42,43,65,66) it is a measure that may be more suitable to assess financial decisions. It may not be the best instrument to capture a general coefficient of people risk aversion, and may fail to appropriately measure risk aversion in the more specific context of health decisions.(55) When we test our model with a more general measure of risk aversion in health decisions that was collected by the NLSY79 in 2010, we still find no evidence that more risk-tolerant individuals have had a higher number of occupational injuries in the past. However, the interpretation of this last finding is complicated by the fact that such new risk-tolerance measures were collected many years (between 10 and 22) after the injuries were reported. We find a positive relationship between risk tolerance and probability of injuries, however, when we measure such attitudes in terms of “revealed preferences” through the analysis of different risky behaviors. These results are consistent with what was found in previous studies.(44,45) But when we interpret these findings it is important to notice that behaviors such as smoking, drinking, and drugs consumptions may indicate both an attitude toward risk and a health 26 We

use a simple OLS regression of the Rotter scale on our four levels of risk tolerance. We do the same for self-esteem. Results are available from the authors. 27 Some studies use dummies for risk tolerance and this allows for the effect of risk tolerance on the outcome of interest to be not linear.(43) But other studies compute a measure of risk aversion based on the NLSY79 respondents’ answers to the lottery questions and assuming a specific functional form for the utility function. They then enter this measure linearly in their regressions.(42)

Workers’ Risk Tolerance and Occupational Injuries

1871

condition that may affect workers’ performances and put them at higher risk of injuries. Furthermore, they are behaviors that can be influenced by the work environment.(67) In terms of cognitive skills we find that a higher level of education does not prevent occupational injuries but significantly reduces the likelihood of recurrent ones. This could indicate that more educated individuals are better able to process information about risk after a first exposure, or face better accommodation or greater choice of future jobs after an injury. We find that AFQT is not consistently associated with injuries across various specifications. As far as noncognitive skills are concerned, the psychological and safety research has found that the personality factors of conscientiousness and agreeableness are negatively associated with occupational incidents,(68) while factors such as extraversion and neuroticism have the opposite effect.(69,70) In our data we have two measures of noncognitive skills, the Rosenberg scale of self-esteem and the Rotter scale of locus control. Both do not play a consistently significant role across all our model specifications. This result contradicts what was found in several previous studies that reported a positive association between higher locus control and occupational injuries.(69) Our result could be explained by the fact that the Rotter measure has been found not to perform well when administered among blue-collar workers.(69) Furthermore, the NLSY 79 asked only a reduced version of the Rotter instrument while the strongest relationships between locus of control and on-the-job injuries have been found in studies that have used the more refined Safety Locus of Control Scale(71) that was designed to specifically address problems of occupational safety. The lack of significance of the Rosenberg scale of self-esteem may be due to the fact that our regressions control for several job attributes. Studies have shown that stress caused by the imbalance between work capacity and work demands can produce lower self-esteem(72) and cognitive failures(28) that lead to work injuries. Finally, despite our improved ability to control for some cognitive and noncognitive traits, the results of this study confirm some findings reported in previous research about recurrent injuries(10) and high utilizers of healthcare.(73,74) We find that demanding working conditions (long hours, not fixed shifts) matter, but so does the socioeconomic status experienced by workers in their youth, before any occupational injury was reported: early health problems, early household poverty, and early exposure to un-

safe jobs. These last characteristics may suggest a pathway into a segmented labor market, a labor market that individuals with low socioeconomic status enter early(75) and where “bad” jobs are associated with higher risk of recurrent work incidents and, possibly, lower wages.(37,45) Recurrent injuries may be one more mechanism through which low family income and teens’ poor health affect individuals’ future health and economic well-being. In a country that is searching for strategies to reduce overall medical costs it is critical to better understand the early social determinants of the phenomenon of recurrent injuries and to develop strategies to prevent them. This last implication is further reinforced by the fact that our result about the negative effect of hourly wages on recurrent injuries contradicts the hypothesis of compensating wage differentials. 5.3. Implications for Future Research and Interventions Overall, our results do not support the hypothesis of injury “proneness” but are more in line with an explanation of recurrent occupational injuries based on the hypothesis of “liability,”(19,20) that is, the concurrent and changing role played by individual, work, and environmental factors in explaining recurrent incidents. Clearly, the interpretation of our results needs to be tested with further research that will overcome some of the limitations of our analysis. This topic needs to be studied with measures designed to capture risk aversion in the specific context of health choices. We need data that will permit us to study these measures before individuals start cumulating labor market experience, and data that will permit us to assess how such measures change over time. We need to study further both how risk aversions affect workers’ occupational choices and how features of the work environment can change risk aversion over time.(76) This is information we can obtain by developing rich longitudinal surveys such as the NLSY79. However, this in an example of a research area that would greatly benefit by the combined analysis of quantitative and qualitative data.(77) The psychological and cultural influences on workers’ risk aversion, choices, and abilities to control their jobs are more likely to be expressed by individuals through personal interviews, focus groups, or worksite observations. Furthermore, studies conducted with primary data in specific industries may also provide richer insight about

1872 workers’ knowledge of specific sectors or jobs risks, their attitude toward them, their ability to manage them,(78) to avoid them—or—not given their financial needs, and, finally, their willingness and ability to reduce them through safety practices.(79) In terms of potential strategies to reduce recurrent injuries, our results suggest that measures of risk tolerance would not be appropriate selection tools for hiring companies that want to rely on ability and personality tests to identify workers at greater risk of recurrent injuries.(80,81) At the same time, our evidence attributes a stronger predictive power to individuals’ “revealed risky preferences,” that is, specific risky behaviors that we find to be related to higher injury probabilities. But the existing relevant literature has produced mixed evidence about the effectiveness of alcohol and drug screening of employees.(82) Furthermore, drugs and alcohol screening raise severe privacy concerns. Our additional results show that heavy workload (long hours and not fixed shift), early risky jobs, early low socioeconomic status, and early health problems are very important determinants of recurrent injuries. Thus, the focus should shift toward policies that more carefully monitor working conditions, employers’ and employees’ adherence to occupational safety regulations, and the implementation of appropriate rehabilitation programs. Indeed, evidence has shown that rehabilitation treatments can be very effective to diminish the risk of new injuries and return to employment workers who have suffered recurrent injuries.(83,84) Finally, our results suggest another potential strategy to specifically prevent recurrent injuries. Our findings suggest that risk aversion may play a different role in affecting the probability of any injury or several recurrent ones. This result needs to be validated by future research but it could indicate that the experience of a first injury somehow changes workers.(79) Some employees may feel more vulnerable to incidents and make less risky choices.(85) Others who are already in risky environments may not care about the new evidence or addition of risk. They may react emotionally and exhibit greater risky behavior.(86,87) Such changes could be interpreted also in the context of the emerging literature that is showing how personality traits such as agreeability or conscientiousness can lead to different outcomes after negative job events,(88) and to changes in judgments of responsibility.(89) The “external” causal attribution of work injury, that is, blaming the employer or other environmental factors, can also lead

Galizzi and Tempesti to changes in safety consciousness and increases in exposures to risk.(35,90–92) This suggests the need to introduce safety systems into workplaces where managers and workers not only work together in designing safety practices,(91,93) but also meet to reflect on the causes that have led to the work incident. Possibly this should happen on an external site with the mediation of an external safety specialist.(22,94) In conclusion, to further advance our understanding of recurrent injuries, and to increase our confidence in making policy recommendations to reduce the costs of recurrent work injuries, we need further research in two main directions. We need to better identify what are the very early determinants of risk aversion, behaviors, and work-related health problems. We also need to further investigate what happens to employees after a first injury both in terms of their psychological reaction, and in terms of their relationship with their employers and co-workers.

ACKNOWLEDGMENTS AND DISCLAIMERS We thank Les Boden, Laura Punnett, two anonymous reviewers, and the participants of the Harvard Center for Risk Analysis March 2014 “Conference on Risk, Perception and Response” who offered many useful comments on this study. We thank the Harvard Center for Risk Analysis for its financial support. The findings and conclusions of this article are those of the authors and do not imply endorsement by any component of Harvard University or other sponsors of the Conference. Comments should be directed to the authors. This research was conducted with restricted access to Bureau of Labor Statistics (BLS) data. The views expressed here do not necessarily reflect the views of the BLS. REFERENCES 1. U.S. Bureau of Labor Statistics, U.S. Department of Labor. Occupational Safety and Health Definitions. Available at: http://www.bls.gov/iif/oshdef.htm, Accessed January 9, 2015. 2. U.S. Bureau of Labor Statistics, U.S. Department of Labor. Employer-Reported Workplace Injuries and Illnesses. Available at: http://www.bls.gov/news.release/pdf/osh.pdf, Accessed January 9, 2015. USDL-14–2183 ed.; 2014. C . 2013 Edition. Itasca, 3. National Safety Council. Injury Facts  IL, 2013. 4. Reville RT. The impact of a disabling workplace injury on earnings and labor force participation. Pp. 147–173 in Haltiwanger JC (ed). Contributions to Economic Analysis,

Workers’ Risk Tolerance and Occupational Injuries

5. 6. 7. 8. 9.

10. 11.

12. 13. 14.

15.

16. 17. 18. 19. 20. 21. 22.

23.

24. 25.

Vol. 241. Amsterdam, New York, and Oxford; Elsevier Science, North-Holland; 1999. Boden LI, Galizzi M. Income losses of women and men injured at work. Journal of Human Resources, 2003; 38(3):722– 757. Galizzi M, Zagorsky JL. How do on-the-job injuries and illnesses impact wealth? Labour Economics, 2009; 16(1):26–36. Butler RJ, Johnson WG, Baldwin ML. Managing work disability: Why first return to work is not a measure of success. Industrial and Labor Relations Review, 1995; 48(3):452–469. Campolieti M. Recurrence in workers’ compensation claims: Estimates from a multiple spell hazard model. Journal of Risk and Uncertainty, 2001; 23(1):75–94. Du J, Leigh JP. Incidence of workers compensation indemnity claims across socio-demographic and job characteristics. American Journal of Industrial Medicine, 2011; 54(10): 758–770. Galizzi M. On the recurrence of occupational injuries and workers’ compensation claims. Health Economics, 2013; 22(5):582–599. Jansson B, Stenbacka M, Leifman A, Romelsjo¨ A. A small fraction of patients with repetitive injuries account for a large portion of medical costs. European Journal of Public Health, 2004; 14(2):161–167. Gawande A. The hot spotters. New Yorker, January 24, 2011. Kirschenbaum A, Oigenblick L, Goldberg AI. Well being, work environment and work accidents. Social Science & Medicine, 2000; 50(5):631–639. Gauchard GC, Mur JM, Touron C, Benamghar L, Dehaene D, Perrin P, Chau N. Determinants of accident proneness: A case-control study in railway workers. Occupational Medicine, 2006; 56(3):187–190. Greenwood M, Woods HM. A report on the incidence of industrial accidents upon individuals with special reference to multiple accidents. Report no. 4 ed. London: Industrial Fatigue Research Board, 1919. Visser E, Pijl YJ, Stolk RP, Neeleman J, Rosmalen JGM. Accident proneness, does it exist? A review and meta-analysis. Accident Analysis and Prevention, 2007; 39(3):556–564. Sass R, Crook G. Accident proneness: Science or nonscience? International Journal of Health Services, 1981; 11(2):175–190. Cameron C. Accident proneness. Accident Analysis & Prevention, 1975; 7(1):49–53. Bernacki EJ. Accident proneness or accident liability: Which model for industry? Connecticut Medicine, 1976; 40(8):535– 538. McKenna FP. Accident proneness: A conceptual analysis. Accident Analysis and Prevention, 1983; 15(1):65–71. Leigh JP. Individual and job characteristics as predictors of industrial accidents. Accident Analysis and Prevention, 1986; 18(3):209–216. de Koster RBM, Stam D, Balk BM. Accidents happen: The influence of safety-specific transformational leadership, safety consciousness, and hazard reducing systems on warehouse accidents. Journal of Operations Management, 2011; 29(7):753–765. Sadiq A, Graham JD. Organizational Risk Perceptions of Disaster: Do Risk Managers Matter? Paper presented at the “Risk, Perception, and Response” Conference at the Harvard School of Public Health, March 2014. Guadalupe M. The hidden costs of fixed term contracts: The impact on work accidents. Labour Economics, 2003; 10(3):339–357. Roberts JT. Psychosocial effects of workplace hazardous exposures: Theoretical synthesis and preliminary findings. International Journal of Industrial Ergonomics, 1993; 40(1):74– 89.

1873 26. Abbe OO, Harvey CM, Ikuma LH, Aghazadeh F. Modeling the relationship between occupational stressors, psychosocial/physical symptoms and injuries in the construction industry. International Journal of Industrial Ergonomics, 2011; 41(2):106–117. 27. Cherry N, Burstyn I, Beach J. Mental ill-health and second claims for work-related injury. Occupational Medicine, 2012; 62(6):462–465. 28. Day AJ, Brasher K, Bridger RS. Accident proneness revisited: The role of psychological stress and cognitive failure. Accident Analysis & Prevention, 2012; 49:532–535. 29. Kang Y, Hahm H, Yang S, Kim T. Application of the life change unit model for the prevention of accident proneness among small to medium sized industries in Korea. Industrial Health, 2008; 46(5):470–476. 30. Kouabenan DR. Role of beliefs in accident and risk analysis and prevention. Safety Science, 2009; 47(6):767–776. 31. DeLeire T, Levy H. Worker sorting and the risk of death on the job. Journal of Labor Economics, 2004; 22(4):925–953. 32. Paap, K. Working Construction: Why White Working-Class Men Put Themselves—and the Labor Movement—in Harm’s Way. Ithaca and London: Cornell University Press, ILR Press, 2006. 33. Veevers JE, Gee EM. Playing it safe: Accident mortality and gender roles. Sociological Focus, 1986; 19(4):349–360. 34. Savage I. Demographic influences on risk perceptions. Risk Analysis, 1993; 13(4):413–420. 35. Landry LG. Preventing occupational injuries: Women’s perception of risk from musculoskeletal exposures. AAOHN Journal, 2006; 54(2):75–83. 36. Dawson DA. Heavy drinking and the risk of occupational injury. Accident Analysis & Prevention, 1994; 26(5):655– 665. 37. Hersch J, Viscusi WK. Smoking and other risky behaviors. Journal of Drug Policy, 1998; 28(3):645–661. 38. Macdonald S, Hall W, Roman P, Stockwell T, Coghlan M, Nesvaag S. Testing for cannabis in the work-place: A review of the evidence. Addiction, 2010; 105(3):408–416. 39. Badri A, Nadeau S, Gbodossou A. Proposal of a risk-factorbased analytical approach for integrating occupational health and safety into project risk evaluation. Accident Analysis & Prevention, 2012; 48:223–234. 40. Barsky RB. Preference parameters and behavioral heterogeneity: An experimental approach in the health and retirement study. Quarterly Journal of Economics, 1997; 112(2):537–579. 41. Flores M, Loomes G. Within-Person and Between-Sample Sensitivity to Dimensions of Risk. Paper presented at the “Risk, Perception, and Response” Conference at the Harvard School of Public Health, March 2014. 42. Ahn T. Attitudes toward risk and self-employment of young workers. Labour Economics, 2010; 17(2):434–442. 43. Spivey C. Desperation or desire? The role of risk aversion in marriage. Economic Inquiry, 2010; 48(2):499–516. 44. Leigh JP. Who chooses risky jobs? Social Science & Medicine, 1986; 23(1):57–64. 45. Viscusi WK, Hersch J. Cigarette smokers as job risk takers. Review of Economics and Statistics, 2001; 83(2):269–280. 46. Grazier S, Sloane PJ. Accident risk, gender, family status and occupational choice in the UK. Labour Economics, 2008; 15(5):938–957. 47. Nichols T. The Sociology of Industrial Injury. London: Mansell, 1997. 48. Fenn P, Ashby S. Workplace risk, establishment size and union density. British Journal of Industrial Relations, 2004; 42(3):461–480. 49. Pouliakas K, Theodossiou I. The economics of health and safety at work: An interdisciplinary review of the theory and policy. Journal of Economic Surveys, 2013; 27(1):167–208.

1874 50. Butler RJ, Worrall JD. Claims reporting and risk bearing moral hazard in workers’ compensation. Journal of Risk and Insurance, 1991; 58(2):191–204. 51. Umberson D. Family status and health behaviors: Social control as a dimension of social integration. Journal of Health and Social Behavior, 1987; 28(3):306–319. 52. Wang X, Kruger DJ, Wilke A. Life history variables and risktaking propensity. Evolution and Human Behavior, 2009; 30(2):77–84. 53. Cooklin A, Giallo R, Rose N. Parental fatigue and parenting practices during early childhood: An Australian community survey. Child: Care, Health and Development, 2012; 38(5):654–664. 54. Swaen GM, Van Amelsvoort LG, Bultmann U, Kant IJ. Fatigue as a risk factor for being injured in an occupational accident: Results from the Maastricht cohort study. Occupational and Environmental Medicine, 2003; 60(1):i88–i92. 55. Dohmen TJ, Falk A, Huffman D, Schupp J, Sunde U, Wagner GG. Individual Risk Attitudes: New Evidence from a Large, Representative, Experimentally-Validated Survey. CEPR Discussion Papers 5517, 2006. 56. Deb P, Manning W, Norton E. Minicourse: Modeling Health Care Costs as Counts. ASHEcon-Cornell University Conference, June 2010. 57. Smith JP. The impact of childhood health on adult labor market outcomes. Review of Economics and Statistics, 2009; 91(3):478–489. 58. Fletcher JM, Sindelar JL, Yamaguchi S. Cumulative effects of job characteristics on health. Health Economics, 2011; 20(5):553–570. 59. Heckman JJ, Stixrud J, Urzua S. The effects of cognitive and noncognitive abilities on labor market outcomes and social behavior. Journal of Labor Economics, 2006; 24(3):411– 482. 60. Slovic P. Perception of risk. Science, 1987; 236(4799):280– 285. 61. Slovic P, Finucane ML, Peters E, MacGregor DG. Risk as analysis and risk as feelings: Some thoughts about affect, reason, risk, and rationality. Risk Analysis, 2004; 24(2):311– 322. 62. Dake K, Wildavsky A. Theories of risk perception: Who fears what and why? Pp. 41–60 in Burger EJ, Jr. (ed). Daedalus, Vol. 119, no. 4. Ann Arbor: University of Michigan Press, 1993. 63. Rippl S. Cultural theory and risk perception: A proposal for a better measurement. Journal of Risk Research, 2002; 5(2):147–165. 64. Iverson RD, Erwin PJ. Predicting occupational injury: The role of affectivity. Journal of Occupational and Organizational Psychology, 1997; 70(2):113–128. 65. Light A, Ahn T. Divorce as risky behavior. Demography, 2010; 47(4):895–921. 66. Di Mauro C, Musumeci R. Linking risk aversion and type of employment. Journal of Socio-Economics, 2011; 40(5):490– 495. 67. Sorensen G, Barbeau E, Hunt MK, Emmons K. Reducing social disparities in tobacco use: A social-contextual model for reducing tobacco use among blue-collar workers. American Journal of Public Health, 2004; 94(2):230–239. 68. Cellar DF, Yorke CM, Nelson ZC, Carroll KA. Relationships between five factor personality variables, workplace accidents, and self-efficacy. Psychological Reports, 2004; 94(3):1437–1441. 69. Hansen CP. Personality characteristics of the accident involved employee. Journal of Business and Psychology, 1988; 2(4):346–365. 70. Maruˇsic A, Musek J, Gudjonsson G. Injury proneness and personality. Nordic Journal of Psychiatry, 2001; 55(3):157– 161.

Galizzi and Tempesti 71. Wuebker LJ. Safety locus of control as a predictor of industrial accidents and injuries. Journal of Business and Psychology, 1986; 1(1):19–30. 72. Enache RG. Burnout syndrome and work accidents. Procedia—Social and Behavioral Sciences, 2013; 78:170– 174. 73. Mautner DB, Pang H, Brenner JC, Shea JA, Gross KS, Frasso R, Cannuscio CC. Generating hypotheses about care needs of high utilizers: Lessons from patient interviews. Population Health Management, 2013; 16:S-26–S-33. 74. Ponzer S, Bergman B, Johansson LM, Brismar B. Patients with recurrent injuries—Psychosocial characteristics and injury panorama. European Journal of Emergency Medicine, 1999; 6(1):9–14. 75. Frone MR. Predictors of work injuries among employed adolescents. Journal of Applied Psychology, 1998; 83(4):565–576. 76. Cho IS. Are Risk Attitudes Fixed Factors or Fleeting Feelings? Iowa State University, Department of Economics, Staff General Research Papers, 2013. 77. Galizzi M, Miesmaa P, Punnett L, Slatin C. Injured workers underreporting in the health care industry: An analysis using quantitative, qualitative, and observational data. Industrial Relations: A Journal of Economy & Society, 2010; 49(1): 22–43. 78. Eklof M, Torner M. Perception and control of occupational injury risks in fishery—A pilot study. Work Stress, 2002; 16(1):58–69. 79. Nelkin D, Brown M. Observations on workers’ perceptions of risk in the dangerous trades. Science, Technology, & Human Values, 1984; 9(2):3–10. 80. Burkolter D, Kluge A, Sauer J, Ritzmann S. The predictive qualities of operator characteristics for process control performance: The influence of personality and cognitive variables. Ergonomics, 2009; 52(3):302–311. 81. Finomore V, Matthews G, Shaw T, Warm J. Predicting vigilance: A fresh look at an old problem. Ergonomics, 2009; 52(7):791–808. 82. Cashman CM, Ruotsalainen JH, Greiner BA, Beirne PV, Verbeek JH. Alcohol and drug screening of occupational drivers for preventing injury. Cochrane Database of Systematic Reviews, 2009; 15(2):CD006566. 83. Garcy P, Mayer T, Gatchel RJ. Recurrent or new injury outcomes after return to work in chronic disabling spinal disorders: Tertiary prevention efficacy of functional restoration treatment. Spine, 1996; 21(8):952–959. 84. Evans TH. Recurrent versus nonrecurrent work-related injuries in chronic disabling spinal disorder patients presenting for the tertiary rehabilitation treatment: A psychosocioeconomic comparison. (back pain) Dissertation Abstracts International: Section B: Sciences and Engineering, Voll. 60(7– B), pp. 3561. ProQuest Information & Learing [Dissertation], US, 2000. 85. Gollier C, Pratt JW. Risk vulnerability and the tempering effect of background risk. Econometrica, 1996; 64(5):1109– 1123. 86. Kahneman D, Tversky A. Prospect theory: An analysis of decision under risk. Econometrica, 1979; 47(2):263–291. 87. Liu L, Rettenmaier AJ. Effects of mortality risk on risktaking behavior. Economics Letters, 2007; 94(1):49–55. 88. Boyce CJ, Wood AM, Brown GD. The dark side of conscientiousness: Conscientious people experience greater drops in life satisfaction following unemployment. Journal of Research in Personality, 2010; 44(4):535–539. 89. Rickard LN. Perception of risk and the attribution of responsibility for accidents. Risk Analysis, 2014; 34(3):514– 528. 90. Niza C, Silva S, Lima ML. Occupational accident experience: Association with workers’ accident explanation and definition. Safety Science, 2008; 46(6):959–971.

Workers’ Risk Tolerance and Occupational Injuries 91. Attwood D, Khan F, Veitch B. Can we predict occupational accident frequency? Process Safety and Environmental Protection, 2006; 84(3):208–211. 92. Gonc¸alves SMP, da Silva SA, Lima ML, Melia´ JL. The impact of work accidents experience on causal attributions and worker behaviour. Safety Science, 2008; 46(6):992–1001. 93. Geller SE. Psychological science and safety: Large-scale success at preventing occupational injuries and fatalities. Current Directions in Psychological Science, 2011; 20(2):109– 114. 94. Mbaye S, Kouabenan DR. How perceptions of experiencebased analysis influence explanations of work accidents. Journal of Safety Research, 2013; 47:75–83. 95. U.S. Bureau of Labor Statistics U.S. Department of Labor. Non-Fatal Occupational Injuries and Illnesses

1875

96. 97. 98. 99. 100.

Requiring Days Away from Work. Available at: http://www. bls.gov/news.release/pdf/osh2.pdf, Accessed January 9, 2015. Davis RM, Pless B. BMJ bans “accidents”. BMJ, 2001; 322(7298):1320–1321. Zagorsky JL. NLSY79 User’s Guide. Center for Human Resource Research, Columbus, OH Ohio State University, 1997 Diecidue E, Wakker PP. On the intuition of rank-dependent utility. Journal of Risk and Uncertainty, 2001; 23(3):281– 298. Gumen A, Ok E, Savochkin A. Decision-making under subjective risk: Toward a general theory of pessimism. Mimeo, unpublished, 2012 Lerner JS, Keltner D. Fear, anger, and risk. Journal of Personality and Social Psychology, 2001; 81(1):146–159.