Loomis D, Marshall SW, Wolf SH, Runyan CW, Butts JD. Effectiveness of Safety Measures Recommended for Prevention of Workplace
Homicide. JAMA. 2002;287(8):1011-1017. doi:10.1001/jama.287.8.1011
Author Affiliations: Department of Epidemiology (Drs Loomis and Marshall and Ms Wolf), Injury Prevention Research Center (Drs Loomis, Marshall, Runyan, Butts and Ms Wolf), and Department of Health Behavior and Health Education (Dr Runyan), University of North Carolina, Chapel Hill; and Office of the Chief Medical Examiner, Department of Pathology, University of North Carolina and North Carolina Department of Health and Human Services, Chapel Hill (Dr Butts).
Context Homicide is the second leading cause of death on the job for US workers.
Government agencies recommend that employers prevent violence against workers
by adopting interventions originally designed to prevent robbery, but the
effectiveness of these interventions is unknown.
Objective To investigate the effectiveness of existing administrative and environmental
interventions recommended for preventing workplace homicide.
Design, Setting, and Participants Population-based case-control study of North Carolina workplaces where
a worker had been killed between January 1, 1994, and March 31, 1998, identified
through a statewide medical examiner system (cases; n = 105) and an industry-matched
random sample of workplaces at risk during the same period, selected from
business telephone listings (controls; n = 210).
Main Outcome Measure Risk of death of a worker due to homicide.
Results Among environmental interventions, strong and consistent reductions
in the risk of a worker being killed on the job were associated with bright
exterior lighting (odds ratio [OR], 0.5; 95% confidence interval [CI], 0.3-1.0).
Among administrative interventions, the largest beneficial effect was for
staffing practices that prevented workers from being alone at night (OR, 0.4;
95% CI, 0.2-0.9). Keeping doors closed during working hours was also associated
consistently with substantially reduced risk (OR, 0.4; 95% CI, 0.1-1.1) but
was not statistically significant. Combinations of 5 or more administrative
measures were associated with significantly lower levels of risk (OR, 0.1;
95% CI, 0.0-0.5).
Conclusions We found evidence suggesting that eliminating solo work at night could
reduce the risk of homicide for workers. Keeping doors closed and using bright
exterior lighting or combinations of administrative interventions also appear
to be beneficial, but there was no evidence of effectiveness for a number
of other recommended measures.
Homicide is the second leading cause of death on the job for US workers,1 with an average of 20 fatal assaults each week.2 A large proportion of all workplace homicides result
from robberies in retail businesses.1 The strategies
government agencies have recommended to prevent homicide in the workplace3- 5 consequently draw largely
on descriptive epidemiologic studies6- 10
and research on robbery by social scientists and industry consultants.11 The recommended measures focus primarily on retail
businesses where money transactions take place3,4,12
and include both environmental design measures, such as increasing the visibility
of persons inside the workplace or installation of alarms and surveillance
devices,13 and administrative or behavioral
measures, like training workers to respond in threatening situations.14 These strategies are based on those adopted by retailers
to make robbery less attractive to potential perpetrators, and which some
studies indicate are effective for that purpose.11,13
Cook15 hypothesized that robbery-prevention
measures should also prevent many workplace homicides. However, very few subsequent
studies have examined this hypothesis or assessed the effectiveness of measures
that have been recommended to prevent workplace homicides. To investigate
the value of these robbery-prevention measures for preventing workplace homicide,
we conducted a case-control study of homicides in North Carolina workplaces
The study population and data collection methods have been described
in detail elsewhere.16 Briefly, the units of
study were workplaces, rather than workers, because we sought to investigate
features of workplaces, like staffing policies and security devices, rather
than behaviors or attributes of individual persons. Cases were North Carolina
businesses or agencies where an employee or proprietor was killed on the job
between January 1, 1994, and March 31, 1998, identified through North Carolina's
statewide medical examiner system. Employers and work locations were identified
from the medical examiner's report or follow-up telephone calls with law enforcement
officers. Workplaces in all industry categories except agriculture, law enforcement,
and the armed forces were included in the study.16
Procedures pertaining to human subjects were reviewed and approved by the
institutional review board of the University of North Carolina School of Public
Controls were workplaces sampled from North Carolina businesses and
agencies contained in a compilation of business telephone listings, American
Business Lists.17The risk set for a given case
included all establishments in this database that were in operation in the
month of the case event (the index month). Controls were interviewed at the
ratio of 2 per case, but to counter anticipated losses due to nonworking telephone
numbers, ineligibility, and refusals, we randomly selected 10 potential controls
for each case, individually matched by broad sector of industry using 1-digit
Standard Industrial Codes.18
After sending an introductory letter, we attempted to contact the manager,
operator, or owner of each workplace by telephone to arrange an interview;
oral agreement to participate was accepted as consent. The owner or manager
in the index month was the preferred informant for both cases and controls,
but if that person could not be interviewed and the workplace was a case,
we followed all available leads to locate a knowledgeable informant, including
coworkers of the victim or police officers who had investigated the homicide.16 For controls, whose numbers were large, we made 6
attempts to conduct an interview for each workplace. If we could not reach
a qualified informant in that number of attempts or consent was refused, we
skipped to the next potential control.
The questionnaire developed for the study was designed to elicit detailed
information about workplaces and an array of factors potentially related to
workers' risk of being killed on the job. Following a classic industrial hygiene
model, preventive measures were divided into environmental control measures
related to workplace design and administrative-control measures related to
work practices and policies. In addition to inquiring about measures to prevent
robbery and violence, the instrument covered operational and physical aspects
of the workplace, its surroundings, and the demographic characteristics of
the workforce. For cases, we requested information about workplace characteristics
in the month the homicide event occurred (the index month). For controls,
we requested the same information for the index month of the matched case.
The instrument was pretested with a sample of potential controls and workplace
homicide cases not included in the study.
Interviews were conducted by telephone by trained, experienced interviewers
and lasted 23 minutes, on average. The goal was to collect data as soon as
possible after identification of a case and its controls, but because the
subject matter was sensitive and the cases were subject to medical and legal
investigation, there was always a lag of at least 3 months between the case
event and the interview.
To supplement the interviews, we obtained information on county size
and urbanization from the 1990 US Census.19
For county-level crime statistics, we obtained data from the North Carolina
Uniform Crime Reporting Program, 1993-1997.20
Analysis of the data sought to quantify the protective effect of interventions
that have been recommended to reduce the risk of homicide in workplaces. Logistic
regression was used to estimate the association of each intervention with
the occurrence of homicide in the workplace via the exposure odds ratio (OR)
and its 95% confidence interval (CI). Variables were coded so that ORs less
than 1.0 indicated a beneficial effect of the intervention.
Each measure was initially examined in isolation using a simple logistic
model. To control for the influence of other workplace and community characteristics,
we also used multiple logistic models to estimate ORs adjusted for potentially
confounding factors that were strongly associated with homicide risk in a
previous analysis of the data. Briefly, the factors from 9 groups of variables
related to workplace location, employer characteristics, and workforce characteristics
that were most strongly associated individually with the risk of homicide16 were entered into a multivariable predictive model
before adding indicators for interventions. The factors considered for control
included the following: (1) being located in an urban county or a county with
high crime rates, (2) belonging to an industry associated with high risk in
previous studies21 (taxicab services, bars
and nightclubs, restaurants and prepared food vendors, grocery and convenience
stores, and gasoline stations), (3) being located in a residential or industrial
area, (4) being open Saturdays, (5) having a high proportion of male or minority
workers, and (6) having moved or opened within 2 years. County characteristics
did not significantly improve the fit of the models and were deleted from
the model, but the other predictors remained. Indicator variables for the
presence of interventions were then added singly to this model, which was
used for all adjusted analyses.
Despite efforts to obtain complete data for each workplace, some missing
values were present and were treated as missing in the analysis. Analyses
were conducted using both unconditional logistic regression and conditional
logistic regression, which preserves the individual matching of cases and
controls. Egret analytical software (Cytel Software Corp, Cambridge, Mass)
was used to fit both classes of models. The results were essentially identical
regardless of the method, but some multivariable models failed to converge
in conditional form because of the smaller cell sizes imposed by matched analysis.
As a result, we report only the results from unconditional logistic models.
Interviews were completed for 105 of the 119 workplaces that were eligible
to be included as cases according to medical examiner's records. Five of the
119 eligible cases were used for a pilot test of the questionnaire and were
excluded from the study, however, and 9 of the 114 workplaces remaining were
dropped from the analysis because no informant could be interviewed.
We initiated attempts to contact 505 potential control workplaces and
succeeded in reaching 344, of which 311 were found to be eligible. Of these
eligible workplaces, interviews were completed for 210 (68%), while respondents
at 101 workplaces (32%) refused. There was no significant difference in response
by industry among cases or controls. The median interval between the homicide
and the interview was 13 months, with a minimum of 82 days; and 4 interviews
required 4 years to complete.
Informants reported preventive measures in almost all workplaces: 99%
had administrative-control measures and 96% had environmental-control measures,
while only 1 workplace had neither type.
Workplaces with a physical barrier between workers and the public were
40% less likely to experience a homicide (OR, 0.6; 95% CI, 0.3-1.3) relative
to workplaces without such features. However, few other environmental control
measures were associated with noteworthy reductions in the risk of homicide
when considered without adjustment for other factors (Table 1).
After adjustment for high-risk industries, location, Saturday hours,
workforce composition, and recent changes in location, bright lighting outside
(OR, 0.5; 95% CI, 0.3-1.0), security alarms (OR, 0.5; 95% CI, 0.2-1.0), and
cash drop boxes that workers could not open (OR, 0.5; 95% CI, 0.2-1.3) were
also associated with notable reductions in risk (Table 1).
Although most individual environmental-control measures were not associated
with reduced risk of homicide, workplaces with at least 5 of these measures
were less likely to have had a homicide than those with none of the measures
In general, administrative control measures were more consistently associated
with reductions in the risk of homicide, compared with environmental measures
(Table 2). When considered without
taking other factors into account, keeping entrances locked or closed during
working hours, making special arrangements with a law enforcement agency,
psychological screening of job applicants, depositing cash proceeds in a bank
daily, having more than 1 worker on duty, and workers not being alone at night
were all associated with noteworthy reductions in risk, with ORs ranging from
0.3 to 0.6. The protective effects of more than 1 worker being present (OR,
0.4; 95% CI, 0.2-0.7) and workers not being alone at night (OR, 0.3; 95% CI,
0.1-0.5) were statistically significant.
Adjustment for the array of external factors considered previously did
not produce major changes in the ORs for most administrative measures (Table 2). The adjusted ORs for closed entrances
(OR, 0.4; 95% CI, 0.1-1.1) and worker training programs (OR, 0.6; 95% CI,
0.3-1.3) suggested slightly stronger protective effects of these measures,
however, while the beneficial effects of arrangements with a law enforcement
agency and having multiple workers on duty were somewhat diminished (Table 2).
Signs advertising limited cash on hand and, after adjustment, sign-in
procedures for visitors and the use of guards were associated with notable
increases in risk.
The majority of workplaces (81 cases and 203 controls) had at least
5 administrative-control measures. This number of measures was associated
with a substantial and statistically significant reduction in the risk of
homicide (OR, 0.1; 95% CI, 0.0-0.5 after adjustment). Data for workplaces
with 1 to 5 administrative measures could not be analyzed separately because
of small numbers.
Because many of the interventions considered here were initially developed
with the intention of preventing robbery rather than worker injuries, we also
conducted an exploratory analysis that took the circumstances of the case
events into account. Cases and their controls were subdivided according to
whether the homicide that defined the case occurred in the course of a robbery
(n = 60) or in association with another kind of event (n = 42). All of the
nonrobbery homicides were related to disputes, and 3 cases with unknown circumstances
For environmental measures, there was no consistent evidence of a stronger
beneficial effect in robbery-related incidents (Table 3). Administrative measures appeared more successful in preventing
robbery-related homicides (Table 4).
The ORs for dispute-related homicides, however, were greater than 1.0 for
several administrative measures. Because these analyses were based on subgroups,
CIs for the ORs were noticeably wider than when all workplace homicides were
Several administrative measures involving work practices or operations
appeared to be beneficial for preventing worker homicides: keeping entrances
closed or locked, depositing cash daily, psychological screening or criminal
background checks of prospective employees, and having more than 1 worker
present, especially at night, were all associated with 30% to 70% reductions
in the odds of a workplace experiencing a homicide in both crude and adjusted
analyses. Environmental changes in workplace design were generally not associated
with reduced risk of fatal violence as actually implemented. After statistical
adjustments to take other workplace characteristics into account, however,
beneficial effects were indicated for bright lighting, alarms, and secure
cash storage devices.
A recent review of research evaluating workplace robbery-prevention
programs identified 26 pertinent studies, the collective results of which
suggest that, overall, these interventions are effective in preventing robbery.13 None of the studies directly evaluated the risk of
employee injury or death in relation to the presence of work interventions,
however. Our study, which did examine the success of interventions for preventing
fatal violence against workers, suggests mixed success for interventions that
have been recommended to prevent violence against workers.
Our findings that the risk of homicide in workplaces may be reduced
by having at least 2 workers on duty are relevant to current public policy.
The state of Florida mandated environmental and administrative control measures
in selected convenience stores through legislation adopted in 1990 and 199222; the 1992 law required stores open at night to have
at least 2 clerks on duty and institute environmental safety measures. The
"2-clerk" provision has been controversial because of the additional cost
to employers and inconsistent empirical support based on earlier, ecological
Our findings suggest that having more than 1 worker present may be an effective
strategy for preventing robbery-related worker fatalities. However, we also
observed increased risk of non–robbery-related homicides in workplaces
where informants reported more than 1 worker "usually" on duty. The numbers
of nonrobbery homicides were not sufficient to investigate these incidents
in detail, but explanations should be sought in appropriately designed studies.
The federal Occupational Safety and Health Administration (OSHA)4 and California's OSHA program (CalOSHA)12
have also issued safety guidelines for late night retail businesses that combine
elements of the Florida laws with other environmental and administrative control
measures. Among the measures recommended by these agencies, our findings indicate
some benefits for bright outdoor lighting, limiting nonemployees' access with
barriers or closed or locked doors, and safe cash-handling procedures, including
drop safes and frequent bank deposits. In this study, security guards, which
the Florida law requires for some businesses, and inquiring about job applicants'
criminal records also appeared to be beneficial, but only in preventing robbery-related
Several other recommended or legally required interventions, including
improved visibility of the work area from outside, video surveillance cameras,
observation mirrors, posted notices of limited cash on hand, and training
to prepare workers for robberies, did not appear to be effective in preventing
robbery-related homicides, however. Odds ratios for several of these measures
were greater than 1.0, an observation that may be explained by inability to
completely control for characteristics of the workplace, such as neighborhood
quality or prior experience with crime, which are related to the risk of homicide
and also influence employers' decisions to implement interventions. It is
also possible, however, that some interventions may be counterproductive if
they advertise the presence of valuable assets but do not constitute a credible
deterrent from robbers' perspective.
Epidemiologic research on workplace violence generally has not differentiated
homicides due to robbery and other causes,6- 10,25
and some agencies have issued recommendations for prevention that do not take
the full range of circumstances into account.3
Our data suggest, however, that different strategies may be needed to control
workplace violence that is not related to robbery.
Forty percent of the deaths defining the case population in this study
were associated with disputes. These events include both work-related disputes
between the victim and a coworker, manager, or client, and disputes involving
domestic partners or family members. Measures intended to make workplaces
unattractive targets for robbery4 generally
had little effect in preventing other kinds of violence. However, closed and
locked doors, which control access by nonemployees, and bright lighting, visibility
from outside, and alarms, all of which increase the chance that an assailant
would be observed, did offer some protection against dispute-related violence.
These findings support CalOSHA guidelines that recommend limiting nonemployee
access and use of alarms to prevent nonrobbery violence against workers.12
This study had several limitations. Among them is that the data were
collected retrospectively through interviews that, because of ongoing legal
investigations and the sensitivity of the subject matter, were typically conducted
1 to 2 years after the events of interest had taken place. It is probable
that informants' recall of past events was imperfect. They may, for example,
have been unable to recall all the safety measures that existed at the time
of the case event, and some measures they did report might not have been functional
or present at that time. While the duration of recall was similar for cases
and controls, it is possible that the quality of recall differed for cases,
which had experienced a traumatic criminal event, and controls, which had
not. Differential reporting of exposure could have resulted if informants
for case workplaces tended to overreport the presence of safety measures,
relative to control informants, because of heightened recall of events associated
with the homicide, or out of desire to avoid blame for the event. Such a tendency
would be most likely to produce bias toward the null. This hypothetical source
of bias, if it were indeed present, might be partially offset by an unplanned
feature of the data collection methods. We interviewed investigating police
officers as proxies for 40% of cases because no other informant was available.16 Officers were generally quite familiar with the circumstances
of the homicides, including preventive measures that were present, but would
presumably not have any motivation to overstate the use of interventions.
The duration of recall was similar for police officer proxies and workplace
informants, but officers sometimes could not provide information on operational
details of the workplace, such as the business hours or the number and sex
of employees, which were treated as covariates in adjusted analyses.
While we attempted to differentiate interventions that were useful for
preventing robbery-related and non–robbery-related violence, the study
design was not ideal for this purpose. We classified cases according to the
circumstances that led to the fatal event, but did not have parallel information
for controls, which could only be subdivided according to the circumstances
of their matched cases.
Despite having included all the cases available in a large state over
5 years, our ability to evaluate some aspects of workplace violence prevention
was still limited by small numbers. Some recommended interventions, such as
bulletproof partitions and comprehensive robbery-prevention packages, were
encountered too infrequently for meaningful statistical analysis. Small numbers
also made effectiveness difficult to evaluate when multiple interventions
were present simultaneously.
Finally, experience in North Carolina workplaces may not be applicable
to some other areas. While North Carolina is the 11th largest state, its largest
cities have fewer than a million inhabitants and some features of the politicolegal
system are atypical. For example, although liquor stores have been a focal
point for other research on robbery prevention, the sale of liquor is a state-controlled
monopoly in North Carolina.
Our research provides evidence that restricting nonemployees' access
to workplaces by keeping doors closed or locked and increasing visibility
with bright exterior lighting are effective in preventing both robbery-related
and non–robbery-related homicides in a wide range of industries. Eliminating
solo work at night also appears to be strongly beneficial for preventing robbery-related
violence. Otherwise, the study failed to provide a strong, unambiguous endorsement
of any single safety measure among those that have been required or recommended
by government agencies and industry groups for preventing workplace homicide.
Having multiple administrative control measures was, however, associated with
a significant reduction in risk.
Measures for preventing workplace violence require further evaluation.
Because of the diversity of the settings in which workers may be subjected
to threats of violence, future studies should consider a range of industries
and geographic areas. Further methodological improvements, including randomized
prospective trials and consideration of different types of violent events,
should be incorporated to improve the quality of measurement and control for
other determinants of crime.