10.1177/0021886304263849ARTICLETHE JOURNAL OF APPLIED BEHAVIORAL SCIENCEMarch 2004Edmondson / DETECTION OF HUMAN ERROR
Learning From Mistakes Is
Easier Said Than Done
Group and Organizational Influences
on the Detection and Correction of Human Error
Amy C. Edmondson
This research explores how group- and organizational-level factors affect errors in
administering drugs to hospitalized patients. Findings from patient care groups in two
hospitals show systematic differences not just in the frequency of errors, but also in the
likelihood that errors will be detected and learned from by group members. Implications
for learning in and by work teams in general are discussed.
The evening nurse reported for work at 3 p.m. in the surgical intensive care unit of
and began her first round. Checking on a patient admitted to the
unit more than 24 hours earlier after a successful cardiac operation, she noticed that the
bag of medication hanging upside-down in the intravenous drip was not heparin—a
clot-preventing blood thinner routinely administered after heart surgery—but was
This project was funded by the Agency for Health Care Policy and Research as part of the Prevention of Drug
Related Complications Study at the Harvard School of Public Health. The author benefited from the support
and advice of Lucien Leape, David Bates, and David Cullen (the three principal investigators of the study),
and two other members of the project team, Kathy Porter and Martha Vander Vliet. Andy Molinsky contrib
uted enormously to this article by conducting much of the qualitative research. Clayton Alderfer and two
anonymous reviewers provided many helpful suggestions that greatly improved the final version of this arti
cle. Finally, I am indebted to Richard Hackman who provided invaluable advice and feedback at each stage
of this research.
Amy C. Edmondson is a doctoral candidate in the joint program in organizational behavior at the Harvard
Business School and Department of Psychology. Her current research explores issues related to organiza
tional learning and work team effectiveness.
THE JOURNAL OF APPLIED BEHAVIORAL SCIENCE, Vol. 40 No. 1, March 2004 66-90
© 2004 NTL Institute
instead lidocaine. Lidocaine, an anesthetic and heart rhythm stabilizer, is not likely to
harm a patient for whom it is not prescribed; however, the absence of heparin might
have been fatal. Fortunately the patient suffered no ill effects from the error in this
case, which subsequently was investigated by the author as part of a larger interdisci
plinary study seeking to understand causes of these kinds of errors in hospitals.
The average hospitalized patient receives 10 to 20 doses of medication each day
and stays for 5 or 6 days (Leape et al., 1991), thus risking exposure to errors such as
this about 85 times. Bates, Boyle, Vander Vliet, Schneider, and Leape (1995) recently
found an average of 1.4 medication errors per patient per hospital stay, with 0.9% of
these errors ultimately leading to serious drug complications. In an earlier population
study, 0.35% of 80,000 patients in New York State Hospitals suffered “a disabling
injury” caused by medications during hospitalization (Leape et al., 1991). These dif
ferent frequencies—0.35% to 140%—reflect a range of drug error phenomena, from
infrequent but disabling injuries to less consequential but far more frequent errors in
medication dosage or timing. One fact is clear: Mistakes in administering drugs in hos
pitals occur and some patients are harmed by these errors.
Patient injuries related to drugs are called adverse drug events, or ADEs. ADEs are
classified here as either preventable (the result of human error) or nonpreventable (not
involving human error, as in an unpredictable allergic reaction) (Leape et al., 1991).
The present study explores underlying causes of preventable errors in drug adminis-
tration. As these errors occur in organizational contexts, this study examined organiza-
tional and group influences on the process of administering drugs to patients, toward a
goal of supplementing existing research that has focused on classifying types of errors
and identifying which individuals or professions are more likely to make them.
The Heparin Event
To illustrate how organizational systems affect error rates, consider further the hep
arin case. First, we might ask who was responsible for this life-threatening mistake.
The answer is not completely straightforward. Between the point at which the operat
ing surgeon prescribed the routine administration of heparin and the evening nurse dis
covered the error, no fewer than six health care professionals were responsible for the
patient’s care. The perfusionist, a medical technician in the operating room, hung the
wrong bag in the intravenous drip. In an interview with the author, he claimed to have
pulled the bag off the shelf where heparin belongs, and he noted that the bags look
alike. (Did pharmacy mistakenly stock lidocaine where heparin should be? Perhaps,
but should not the perfusionist have checked the bag carefully without assuming it to
An anesthesiologist wheeled the patient to the surgical intensive care unit and failed
to notice the error. Then, another nurse, on duty the afternoon the patient arrived in the
intensive care unit, was responsible for checking medications. How could she have
missed the perfusionist’s mistake? Easily, according to cognitive psychologists. The
human tendency to perceive what one expects to see rather than what is actually there
is a well-documented psychological phenomenon (Norman, 1980, 1981; Reason,
1984; Rumelhart, 1980). A heparin bag hanging by the bedside of postcardiac surgery
Edmondson / DETECTION OF HUMAN ERROR 67
patients is utterly routine—making it all too easy for caregivers to assume the presence
of heparin and miss the error. This nurse was replaced by another at 11 p.m., and
another at 7 a.m. the next morning. All of them could have detected the mistake before
the next evening nurse with whom this episode opened came on duty.
Had the patient experienced an adverse drug event, it would have been technically
difficult to assign blame to one person, for it was in the job descriptions of several indi
viduals to avoid and to check for such errors. Drug administration is a collective under
taking—inviting the question of whether drug error rates can be explained as a func
tion of group or hospital unit properties, rather than focusing exclusively on individual
characteristics such as ignorance, fatigue, or carelessness. Moreover, in terms of pre
venting ADEs, individual errors cannot ever be completely eliminated. Indeed, even
when everything “goes right” in the hospital setting, errors undoubtedly have occurred
without notice or consequence. When an ADE caused by human error is documented,
that error is likely to have involved multiple errors—such as the repeated errors of fail
ing to detect and correct the discrepancy, as described above. Secondary errors of not
noticing are as critical as primary errors that start a potentially harmful chain of events.
In summary, the system of drug administration in a modern hospital is complex,
involving multiple handoffs in the journey from physician decision making all the way
through to the receipt of a medication by a patient. Bates, Leape, and Petrycki (1993)
have identified 10 points at which an error can occur (or be caught): (a) physician pre-
scription, (b) initial delivery to a unit secretary who (c) transcribes the order, which
then (d) must be picked up by a nurse who (e) verifies and transcribes again and (f)
hands off to the pharmacist who (g) dispenses the medication and (h) sends it back to a
nurse who (i) administers to a patient who (j) receives the drug. The present study
explores organizational influences on the execution of these loosely coupled tasks, all
of which take place within the context of hospital patient care “units.” In the next sec-
tion, I review approaches to understanding errors and accidents in organizations at
three levels of analysis, to set a context for the design of the present study.
THE SOCIAL PSYCHOLOGY OF ERRORS AND ACCIDENTS
Medical researchers, psychologists, and organizational theorists have conceptual
ized errors and accidents in at least two different ways, and their research has investi
gated both the causes of errors and the efficacy of preventive strategies. One approach
focuses on the individual and the other on the role of the system in which individuals
operate in inducing or preventing accidents. A third perspective, proposed here, inte
grates system and individual levels of analysis by focusing on the work group as the
point where organizational and cognitive effects meet and play out in enabling or pre
In general, medical researchers have tended to emphasize the role of individual
caregivers in studies of errors, and to blame physiological or educational deficits for
68 THE JOURNAL OF APPLIED BEHAVIORAL SCIENCE March 2004
adverse events (e.g., Lesar et al., 1990; Melmon, 1971). Many studies have been con
ducted to detect the frequency of errors among hospitalized patients and to identify
responsible individuals. For example, physicians have been identified as more often
responsible for such mistakes than nurses, pharmacists, or other personnel (Bates
et al., 1993). Suggested strategies for prevention include individual-focused devices
such as computerized order entry to prevent errors caused by poor handwriting and
physician remedial education (Classen, Pestotnik, Evans, & Burke, 1991; Cohen,
1977; Massaro, 1992).
Psychologists have offered both cognitive and affective explanations for human
error. According to schema theory, perceivers’expectations, or frames, have the power
to steer attention away from actual visual data, enabling perceptual processes to con
struct images consistent with expectations (Rumelhart, 1980). If we see what we
expect to see, we can make mistakes such as administering lidocaine instead of hepa
rin. Similarly, well-learned activities can be carried out without conscious attention,
allowing us to make odd slips such as putting a cereal box in the refrigerator (Norman,
1981; Reason, 1984). This automatic quality is indeed present in some drug errors; “I
wasn’t thinking” or “I can do [a particular task] in my sleep” are frequent phrases used
by nurses interviewed, describing how an obviously wrong medication was given or
not noticed. Such cognitive explanations for error have no need for Freud’s “hidden
impulses” to generate “slips” (Goleman, 1985); however, the unconscious effects of
emotions such as anger or anxiety also can induce error, by distracting people’s atten-
tion from the task at hand. Although the current social psychological literature empha-
sizes cognitive explanations for human behavior (Fiske & Taylor, 1991), in examining
group-level influences below, we reconsider the role of emotions.
Certain sociologists and organizational theorists have focused on the properties of
systems in understanding error (e.g., Perrow, 1984). Rather than trying to explainwhy
slips are made by individuals, this approach examines the design of systems and how
systems give rise to human error. The nature of the system both influences the actions
of individual operators and determines the consequences of errors. Perrow (1984)
describes a “normal accident” as a predictable consequence of a system that has both
interactive complexity and tight coupling. Interactive complexity is characterized by
irreversible processes and multiple, nonlinear feedback loops. Interactively complex
systems thus involve hidden interactions; the consequences of one’s actions cannot be
seen (Perrow, 1984). Giving a patient the wrong medication has this quality; immedi
ate feedback is typically not present. Tightly coupled systems have little slack; actions
in one part of the system directly and immediately affect other parts. The system of
drug administration to hospitalized patients offers considerable interactive complex
ity; however, because the procedures linking medications to patients are loosely cou
pled, failures in a part of the system can be caught and corrected without causing harm.
Although the modern hospital thus does not fit Perrow’s worst case scenario, the
interactive complexity of medications does create considerable potential risk for
Edmondson / DETECTION OF HUMAN ERROR 69
In Perrow’s model, individual human error is taken as a given; the critical question
is, under what conditions is the ever-present potential for human error dangerous?
When do simple slips trigger an irreversible chain of events—by virtue of properties of
the system—that ends in disaster? He challenges “the ready explanation of operator
error” (Perrow, 1984, p. 26) and proposes that certain systems are accidents waiting to
Just as the design of a system can invite accidents, straightforward solutions—
implemented to prevent specific kinds of errors that have occurred in the past—can
lead to unintended negative consequences. Organizational systems tend to resist
straightforward solutions to problems (Forrester, 1971). When only the superficial
symptoms of complex problems are addressed, the underlying problem typically
remains unsolved, and even can be exacerbated if the solution feeds into a vicious
cycle (such as providing food as direct aid, which relieves starvation but perpetuates
the problem of population growth in inhospitable climates; Senge, 1990). This per
spective suggests that strategies such as remedial education of physicians who have
written incorrect prescriptions in the past, or use of warning stickers to remind care
givers of allergies, may have limited effectiveness in preventing drug errors, as they
fail to consider the nature of the system that consists of many individuals and technical
systems interacting in administering drugs to patients.
Organizational systems also transmit broader social forces that affect attitudes and
behaviors related to error—including social forces that preclude “embracing error”
(Michael, 1976) and thus inhibit learning. A widely held view in society of error as
indicative of incompetence leads people in organizational hierarchies to systemati-
cally suppress mistakes and deny responsibility (Michael, 1976). Hierarchical struc-
tures thus discourage the kind of systematic analysis of mistakes that would allow peo-
ple to better design systems to prevent them.
In summary, systemic approaches expand the scope of analysis of errors and acci-
dents beyond the idiosyncrasies of individual behavior. Moreover, they introduce the
variable of how errors that have been made are treated in the system, and how systems
for error prevention can be designed. Some of these theories focus on technical and
structural features of macro systems that primarily influence the outcomes of error and
thwart simple preventive strategies; others focus on societal forces that shape attitudes
toward error. However, these theories are less useful for understanding variance in
behavior within a given organization or system. Thus we turn to the social psychology
of groups for insight into how small social systems such as work teams may operate to
prevent or catch errors.
Individual skills, motivation, and cognition are imperfect. Organizational systems
are inevitably flawed. Hackman (1993) proposes that, facing this ever-present poten
tial for error, teams in organizations can act as “self-correcting performance units.”
Members of a superb team have a way of coordinating tasks, anticipating and respond
ing to each other’s actions, and often appearing to perform as a seamless whole. Con
sider, for example, the study by Foushee, Lauber, Baetge, and Acomb (1986) on the
70 THE JOURNAL OF APPLIED BEHAVIORAL SCIENCE March 2004
effects of fatigue on flight crew errors. These researchers found, to their surprise, that
crews who had just logged several days flying together (fatigue condition) made sig
nificantly fewer errors as teams than well-rested crews who had not yet worked
together. As expected, the fatigued individuals made more errors than others; however,
functioning as teams they were able to compensate for these, presumably because they
were better able to coordinate and to catch each other’s mistakes.
Such differences in group behavior and performance have been examined by two
research traditions in the social psychology literature. One research tradition has
focused on social, affective, and unconscious influences on groups and their members,
whereas another tradition is rooted in cognition, goals, and structures.
The former tra
dition finds its origins in the late 19th-century work of Gustave Le Bon, who proposed
that unconscious processes of a crowd or group could manifest themselves in the
actions of individual members (Alderfer, 1987). Similarly, the Hawthorne studies
reported that emotions and tacit group norms could exert a greater influence on perfor
mance than working conditions or economic incentives (Roethlisberger & Dickson,
1939), and sociotechnical theorists later described the importance of a strong social
unit in motivating workers (Rice, 1958; Trist & Bamforth, 1951). Finally, the inter-
group perspective has explored how membership in identity groups (such as gender or
race) or organizational groups (such as rank or function) can affect communication
and motivation in task groups (Alderfer, 1987), which could lead to errors due to lack
of coordination. When members of a work team communicate across tacit boundaries
imposed by rank or identity group, this can inhibit the transfer of valid data (Argyris,
1985). Along these lines, nurses and physicians working as part of the same team (hos-
pital unit) face identity group boundaries confounded with status differences that can
affect within-team communication and thereby influence the process of administering
drugs to patients.
The other research tradition emphasizes intellectual formulations such as structure
and design, minimizing attention to emotional processes. The emphasis is on identify
ing conditions that help teams work together and solve problems (e.g., Hackman &
Morris, 1975; Maier, 1967), such as early research that examined what leadership style
would enable positive group outcomes (Lewin, Lippitt, & White, 1939). Recent work
in this tradition has focused on the importance of practice (Senge, 1990) and members’
interpersonal skills (Helmreich & Foushee, 1993) as influences on team performance.
The structure of the team, the design of its task, and the supportiveness of the organiza
tional context (reward, information, and educational systems) have also been central
concerns in this tradition (Hackman, 1987). Both research traditions share the under
standing that a group is more than the sum of its members and that group-level phe
nomena exist and influence task performance (Alderfer, 1987; Hackman, 1990).
Although the design of the present study drew almost exclusively from the latter tradi
tion, the results, as discussed more fully below, suggest that the story of drug errors
cannot be told fully without attention to the constructs of the former.
Hospital units are one kind of work group, and group research may help explain dif
ferences in error rates. Hospital units are primarily managed and staffed by nurses,
whereas physicians and pharmacists affiliated with each unit interact with the nursing
teams. Preliminary observation revealed considerable flexibility in carrying out work
Edmondson / DETECTION OF HUMAN ERROR 71
processes; units vary in the way work is done and in how people work together. When
and by whom the various tasks will be accomplished is left largely up to the discretion
of the unit members, suggesting that critical performance outcomes such as the preva-
lence of drug errors can vary. Based on this perspective, the present study focused on
hospital units as the unit of analysis in attempting to gain insight into causes and pre-
vention of medication errors. The guiding question—rather than what causes people
to make mistakes—was thus, are some work groups better than others at catching and
correcting human error before it becomes consequential? Further, are such groups
better able to learn from the inevitable errors that do occur and to avoid making the
same ones in the future?
Previous research suggested that error rates vary widely across units within the
same hospital (Bates et al., 1993). Similarly, unit error rates obtained in the current
study of drug complications range from 2.3 to 23.7 errors per thousand patient days.
(Table 6, discussed in the Results section, displays the data for each unit.) What
accounts for these differences? One possibility is that unit characteristics (such as
team stability, norms, and work structure) influence error rates. Starting with the prop
osition that team behaviors are influenced by organizational context, team leader
behaviors, task design, resource adequacy, and team composition (Hackman, 1987),
the present study tests the hypothesis that error rates vary with these unit characteris
tics. Figure 1 depicts a model of proposed influences on, and outcomes of, hospital
unit work processes. In this model, unit outcomes such as error rates and members’
assessments of unit performance are group-level characteristics.
The research question guiding the present study was, are differences in work group
(unit) properties associated with differences in error rates? And do members’ percep
tions of how their unit performs and of the quality of unit relationships vary with the
“hard” data of drug error rates?
Thus the study explores the extent to which certain
72 THE JOURNAL OF APPLIED BEHAVIORAL SCIENCE March 2004
FIGURE 1: Model of Unit-Level Influences on Drug Error Rates and Other Performance
group-level properties contribute to understanding differences in drug error rates in
The study used a comparative nonexperimental research design. Eight hospital unit
teams were randomly selected for study from two urban teaching hospitals affiliated
with the same medical school. No known differences in unit composition, skill, pro
fessionalism, or workload existed between the two hospitals.
Three parallel data collection activities were conducted in the units studied. Inde
pendent of the present study, potentially harmful drug-related errors were identified
over a 6-month period by trained medical investigators through daily chart review,
daily informal visits to each unit to inquire about unusual drug events, and a confiden
tial system to allow unit members to report incidents in writing. Previous research
(Bates et al., 1993) found this combination of activities to uncover more errors than
Data were collected for four variables related to patient adverse drug events
(ADEs), each expressed as a number of incidents per thousand patient days: (a)
nonpreventable ADEs (adverse drug events unrelated to error), (b) preventable ADEs,
(c) potential ADEs (or PADEs) (consequential errors that did not harm patients despite
having the potential to cause injury, such as the heparin incident described above), and
(d) interceptions (errors caught and corrected before reaching patients). The first two
variables (nonpreventable ADEs and preventable ADEs) were measured by a combi-
nation of patient chart review and voluntary reporting. The second two variables
(PADEs and interceptions) could only be measured through voluntary reporting of
events by unit members. To capture the construct of human error, the primary depen
dent variable for the present study (“detected error rates”) is the sum of preventable
ADEs and potential ADEs.
Work Group Measures
To address the main research question of how unit properties influenced these out
come variables, two independent methods were employed (survey and observation) to
develop a full picture of how the hospital units studied functioned and how they dif
fered. First, the author developed a survey to assess the social and organizational prop
erties of hospital units, based on a prior instrument designed to study the performance
of cockpit crews (Hackman, 1990). The new survey measures include leadership
behaviors of nurse managers, organization context (adequacy of training, information,
and equipment), team characteristics (stability, composition, quality of unit relation
ships, and performance outcomes), and individual satisfaction and motivation. Figure
1 depicts relationships among these variables.
Next, another researcher, blind to both error rates and survey results, observed
nursing teams in the eight units for several days each over a 2-month period; his goal
Edmondson / DETECTION OF HUMAN ERROR 73
was to observe behavioral dynamics in support of the larger project goal of under
standing conditions surrounding drug errors.
He interviewed all nurse managers, sev
eral nurses, and members of the support staff in each unit, using a semistructured inter
view format that included open-ended questions to elicit interviewees’ own
descriptions of how mistakes are handled and perceived in their unit.
Eight units in two urban hospitals were included in the study. Five units from one,
referred to as Memorial Hospital, include three intensive care units and two general
care units. The three units from the other, referred to as University Hospital, are all
general care units. Two hundred and eighty-nine surveys were distributed to nurses,
physicians, and pharmacists identified by hospital personnel as full- or part-time
members of the eight units. Fifty-five percent (159) of the surveys were completed and
Surveys were filled out at the end of the second month of the 6 months dur
ing which drug errors were tracked.
Survey items include unit descriptions such as “This unit operates as a real team” as
well as individual satisfaction items such as “I am satisfied with the amount of pay I
receive.” (Table 2 shows items for selected scales as illustration.) Each descriptive
statement is followed by a 7-point Likert-type scale (7 = strong agreement,1=strong
disagreement), except for the items describing nurse manager behaviors, which are
followed by a 5-point scale ranging from never to frequently. As shown in Table 1, the
individual items combine to create 12 variables in five categories: leadership behavior,
organization context, unit characteristics, unit outcomes, and individual satisfaction.
(These categories map to the research model in Figure 1.) Single-item variables
include members’ perceptions of the frequency of drug errors, of the degree of report
ing of such errors, and of the consequences of making a mistake in their unit.
Preparatory analyses assessed (a) the adequacy of the psychometric properties of
the new survey instrument, and (b) whether the survey scales were meaningful group-
level variables. As the research questions examine relationships between unit attrib
utes and unit error rates, it is important to establish that the survey variables in question
are meaningful as unit-level properties (Kenny & LaVoie, 1985). Spearman rank order
correlations were then used to examine substantive relationships between unit proper
ties and detected error rates, interceptions, and nonpreventable ADEs. Finally, qualita
tive data were analyzed independently to understand ways in which unit climates and
nurse behaviors differ, and units were ranked according to openness in discussing mis
takes, a variable that surfaced during observation and analysis as differing across units.
These qualitative results subsequently were compared to the quantitative survey and
drug error data.
74 THE JOURNAL OF APPLIED BEHAVIORAL SCIENCE March 2004
Survey psychometrics. Analysis of the psychometric properties of the survey
results yielded satisfactory results, demonstrating the validity and internal consistency
reliability of the scales as shown in Table 1. Internal consistency reliabilities range
from 0.67 for team composition to 0.90 for each of the three nurse manager behaviors.
Unit-level properties. To assess the degree to which the organizational attributes
measured by the survey are meaningful as group-level variables, two complementary
measures of within-unit agreement about unit properties were computed. Measures of
within-unit agreement also provide an indication of interrater reliability in measuring
unit characteristics with this survey. The first, the intraclass correlation coefficient
(ICC), uses one-way analysis of variance (ANOVA) to compare between- and within-
unit variance. The second, the interrater reliability coefficient (IRR) derived by James,
Demaree, and Wolf (1984), compares actual variance to a measure of “expected vari
ance” to assess within-group agreement, without between/within comparisons.
ICCs for survey variables assessing unit characteristics were examined first; Kenny
and LaVoie (1985) maintain that an ICC greater than zero indicates that a variable is
meaningful at the group level. In many organizations, however, institutional forces
restrict variance, leaving strong similarities across work groups, which suggests that
allowing the magnitude of between-group differences to determine whether a variable
Edmondson / DETECTION OF HUMAN ERROR 75
Summary Statistics for Survey-Based Measures
Scale Mean SD Reliability
Nurse manager direction setting 3.78 0.86 .90
Nurse manager coaching 3.60 0.88 .90
Nurse manager external relations 3.79 0.88 .90
Supportiveness of organization context 4.96 0.94 .74
Team composition 4.43 1.15 .67
Team stability 4.76 0.95 .68
Unit performance outcomes 5.06 0.93 .77
Quality of interpersonal relationships 4.12 0.84 .80
Internal motivation 6.14 0.71 .70
General satisfaction 4.63 0.57 .73
Satisfaction with work relationships 5.74 0.66 .68
Satisfaction with growth opportunities 5.51 0.85 .72
“Drug-related errors occur frequently in this unit” 2.88 1.25
“In this unit, drug-related errors are always reported” 4.43 1.54
“If you make a mistake in this unit it is held against you” 3.24 1.20
is meaningful at the group level may not always be sensible in real organizational set
tings. For example, in these data, several variables designed to measure unit-level
attributes have very low positive ICCs, such as team composition (.08) and unit perfor
mance outcomes (.06). Does this indicate lack of within-team agreement? Perhaps
not. IRRs for the same variables (.80 and .88, respectively) provide a measure of
within-group agreement that is unaffected by between-group similarities, and these
values suggest high levels of within-team agreement (see Table 3). As many organiza
tional features are similar across units, the critical sources of variance among units
may be largely due to those features that are noticeably different, such as leadership
behaviors of nurse managers.
The IRR is appealing as an absolute measure of within-unit agreement; however, its
expected variance term is derived based on an assumption that survey responses are
rectangularly distributed, thus posing a different limitation. Positive leniency in
responding to items about unit performance is likely, which limits variance and can
inflate estimates of agreement (see James, Demaree, & Wolf, 1984).
bias makes it difficult to assess absolute levels of unit agreement. However, examining
both coefficients shows which variables have greater between-group differences while
76 THE JOURNAL OF APPLIED BEHAVIORAL SCIENCE March 2004
Survey Items for Selected Scales
Selected Scale Items
Nurse manager direction
The nurse manager takes initiatives to establish strong standards of medical
excellence and professionalism in this team. The nurse manager sets clear
goals and objectives for this team. The nurse manager is clear and explicit
about what he or she expects from unit members. The nurse manager ac
tively encourages nurses in this unit to stretch their level of performance.
Nurse manager coaching The nurse manager takes initiatives to build the unit as a team. The nurse
manager actively coaches individual unit members. The nurse manager
shares leadership with other experienced members of the unit. The nurse
manager is an ongoing “presence” in the unit—someone who is readily
Quality of interpersonal
Members of this unit care a lot about it and work together to make it one of
the best in the hospital. Working with the members of this unit is an ener
gizing and uplifting experience. Some people in this team do not carry
their fair share of the overall workload. (R) Every time someone attempts
to straighten out a team member whose behavior is not acceptable, things
seem to get worse rather than better. (R) There is a lot of unpleasantness
among members of this team. (R)
Recently, this unit seems to be “slipping” a bit in its level of performance
and accomplishments. Patients often complain about how this unit func-
tions. (R) The quality of care this unit provides to patients is improving
over time. Drug-related errors occur frequently in this unit. (R) This unit
shows signs of falling apart as an organization. Attending physicians often
complain about how this unit functions. In this unit, drug-related errors
are always reported.
NOTE: (R) = reverse scored.
having similar within-group agreement levels. To facilitate comparison of the two
kinds of unit-agreement coefficients, ICCs and median IRRs are shown for eight vari-
ables in Table 3.
The ICC’s between/within comparison is useful in showing which variables stand
out as attributes that vary across, and hence best distinguish among, units. For exam-
ple, nurse managers’ direction setting (.22) has the highest ICC in the sample, reflect-
ing the greater variance between units in perceived behavior of nurse managers, com-
pared to variables describing less tangible unit characteristics such as performance
outcomes (ICC = .06). ICCs for each of the three leadership scales are significant
(p < .001). This result is consistent with the author’s on-site observations of the nurse
managers’behaviors; each unit has a different nurse manager, in some cases with strik
ingly contrasting management styles. In contrast, the variable “unit performance out
comes” has low between-unit variance (ICC = .08) despite having the same level of
within-unit agreement (IRR = .88) as nurse manager direction setting.
In sum, IRRs and ICCs provide different, complementary measures of within-
group agreement. In these data, the IRRs reveal the high within-unit agreement levels
for variables, such as unit performance outcomes, which lack high between-unit dif
ferences and thus have low ICCs. Together these coefficients allow us to have confi
dence that these measures are meaningful as group- or unit-level variables.
RESULTS AND DISCUSSION
As discussed above, this study employed two distinct methods to investigate orga
nizational factors that may account for variance in drug error rates across hospital
units. In this section, I first review the quantitative results, and then present results
from the qualitative research. These two lenses provide distinct and complementary
pictures of the phenomenon.
Edmondson / DETECTION OF HUMAN ERROR 77
Intraclass Coefficients (ICC) and
Interrater Reliability Coefficients (IRR) Compared Across Units
ICC Median IRR
Supportiveness of organization context .12* .85
Nurse manager direction setting .22* .88
Nurse manager external relationships .21* .89
Nurse manager coaching .18* .82
Team composition .08* .80
Team stability .10* .83
Unit performance outcomes .06 .88
Satisfaction with growth opportunities .04 .91
“Drug-related errors occur frequently” .08* .67
*One-way ANOVA for each of these variables by unit is significant (p < .05).
Relationships Between Error Rates and Unit Characteristics
Correlational analyses of the relationships between unit characteristics and error
rates yielded unexpected results. We expected to find higher error rates to be associ
ated with lower mean scores on perceived unit performance, quality of unit relation
ships, and nurse manager leadership behaviors. Exactly the opposite result was found.
As shown in Table 4, detected error rates are strongly associated with high scores on
nurse manager direction setting (r = .74), coaching (r = .74), perceived unit perfor
mance outcomes (r = .76), and quality of unit relationships (r = .74), with p < .03 in
each case. These relationships between nurse manager behaviors and the independ
ently collected error rates are noteworthy, and at first glance an odd result indeed. Do
better coached teams make more mistakes?
An alternative interpretation of these findings also merits careful consideration: In
certain units, the leaders may have established a climate of openness that facilitates
discussion of error, which is likely to be an important influence on detected error rates.
78 THE JOURNAL OF APPLIED BEHAVIORAL SCIENCE March 2004
Organizational Influences on the Detection of Error:
Spearman Rank Order Correlations Between
Survey Variables and Drug Error Variables
Detected error rates and . . .
Nurse manager coaching .74*
Nurse manager direction setting .74*
Unit performance outcomes .76*
Quality of unit relationships .74*
Willingness to report errors .55
Mistakes (are not held against you)
Intercepted errors and . . .
Nurse manager coaching .71*
Nurse manager direction setting .83*
Unit performance outcomes .71*
Quality of unit relationships .76*
Willingness to report errors .62
Mistakes (are not held against you) .45
Nonpreventable drug complications and . . .
Nurse manager coaching –.10
Nurse manager direction setting –.09
Unit performance outcomes .11
Quality of unit relationships –.07
Willingness to report errors .09
Mistakes (are not held against you) .07
NOTE: All correlations are between unit means for each survey variable and the dependent drug error
a. The survey item that reads “If you make a mistake in this unit it is held against you” is reverse scored in
analyzing the data.
*p < .03, two-tailed.
Awareness of the labor intensiveness and difficulty of tracking drug-related errors—
and of how easy and natural it is for human beings to underreport error—suggests that
measuring errors in each unit is not a trivial undertaking. The difficulty of assessing
actual error rates is shown in a recent study in which a hospital in Salt Lake City was
able to increase the number of ADEs identified forty-fold after instituting a new com
puting system to predict and track errors (Evans et al., 1992). The magnitude of this
increase indicates how few of the errors made in hospital units are reported, and sug
gests also that variance in error rates caused by differences in patient severity and com
plexity across general and intensive care units can be overwhelmed by the influence of
lack of reporting of errors (see Note 10). In short, detected error rates are a function of
at least two influences—actual errors made and unit members’ willingness to report
errors. Moreover, in organizational settings in which errors are consequential, willing
ness to report may be a greater influence on the error rates obtained than is variance in
actual errors made. These observations suggest that positive correlations between
error rates and nurse manager coaching, perceived unit performance, and quality of
unit relationships may be explained by examining the role of members’ willingness
and ability to catch and report drug errors.
The Role of Willingness to Report Errors
Higher detected error rates in units with higher mean scores on nurse manager
coaching, quality of unit relationships, and perceived unit performance may be due in
part to members’ perceptions of how safe it is to discuss mistakes in their unit. Several
other survey variables provide support for this hypothesis. First, a variable labeled
“willingness to report errors” was computed for each unit; this is the difference
between two items means—”in this unit, drug errors are always reported” and “drug
errors occur frequently in this unit.” This difference score, which ranges from –4 to 5
in these data, measures perceived willingness to report errors controlling for perceived
frequency of error. Detected error rates then were found to be correlated with willing
ness to report errors (r = .55). Second, a single-item variable measuring unit members’
perceptions that making a mistake in their unit will not be held against them is also cor
related with detected error rates (r = .44; see Table 4). Moreover, one-way ANOVA
reveals significant between-unit differences for both of these single-item variables
(p < .05), suggesting that unit climates vary significantly in perceptions of the risk of
Examining relationships between these two survey variables and another depen
dent variable, the number of interceptions in each unit, provides additional support
for the hypothesis that willingness to report error influences ability to detect errors
in a unit (see Table 6 for means and standard deviations of interceptions andother
drug error data). Recall that the concept of self-correcting performance units high
lights the role of catching errors before they become consequential; the number of
interceptions in a unit provides an indication of unit members’ attentiveness to their
interdependence in caring for patients, as well as a measure of their ability to function
as a self-correcting performance unit. Analysis of the data shows that units with a
greater number of interceptions (team self-correcting behaviors) also tend to have rel
Edmondson / DETECTION OF HUMAN ERROR 79
atively higher scores on nurse manager direction setting (r = .83), coaching (r = .71),
performance outcomes (r = .71), and quality of unit relations (r = .76), with p < .03 in
each case. Further, correlations between interceptions and willingness to report (r =
.62) and between interceptions and unit tolerance of mistakes (r = .45) suggest that
interceptions are more prevalent in units in which members are less concerned about
being caught making a mistake. Figure 2 depicts a model that summarizes these rela-
tionships. The model proposes that nurse manager leadership behaviors—especially
related to how mistakes are handled—create an ongoing, continually reinforced
climate of openness or of fear about discussing drug errors. The more readily errors
are reported and discussed, the more willing unit members may be to report error in
the future, and the more they may believe that making a mistake will not be held
against them. A feedback loop indicates how these perceptions contribute to a self-
perpetuating cycle of learning, or else of defensiveness.
Finally, if willingness to discuss mistakes is indeed an important influence on
detected error rates, then nonpreventable ADEs should not be correlated with the unit
characteristics discussed above. The results in fact do reveal a consistent lack of asso
ciation between nonpreventable ADEs and unit variables such as nurse manager direc
tion setting (r = –.09), coaching (r = –.10), unit tolerance of mistakes (r = .07), and
willingness to report errors (r = .09). Although error rates show significant correla
tions with these unit characteristics, there is simply no relationship between nonerror
drug complications and these survey variables.
Error Suppression in Authoritarian Units
In the above discussion, I offer an explanation for the unexpected reversal of the
direction of the correlations between drug error rates and unit characteristics, and I
support this with additional quantitative results that are consistent with this interpreta
Fortunately, however, there are other independent data available that corrobo
rate the proposed phenomenon.
80 THE JOURNAL OF APPLIED BEHAVIORAL SCIENCE March 2004
FIGURE 2: Model of Influence of Willingness to Report Error on Detected Error Rates
Qualitative data gathered by a researcher blind to the quantitative results (for both
drug and survey data) support the hypothesis that these units vary in terms of openness
in discussing mistakes (see Note 5). The social climate was noticeably different across
the eight units, including differences in nurse-physician relationships, and appeared to
be influenced by nurse managers whose behavioral styles varied widely. Several
behavioral patterns related to mistakes were observed and described in interviews. In
analyzing these data, the qualitative researcher identified several variables that distin
guished among units, such as unit climate (blame oriented vs. learning oriented),
openness, nurse manager attire, nurses’ trust in their nurse manager, and perceived
supportiveness of both nurse manager and peers. He then ranked each unit as high,
medium, or low on openness, and arranged the eight units in a table from most authori
tarian to most open. Table 5 displays these results, which are strikingly consistent with
quantitative results, as juxtaposed in Table 6. Only one unit is noticeably out of place,
Memorial 4, which has fewer errors than six of the other units along with a nurse
manager who is perceived as relatively open and accessible.
A more in-depth look at four of the units will illustrate the phenomenon of authori
tarian suppression of error that I propose may explain much of the variance in the
dependent variable of detected error rates. These units illustrate four distinct leader-
ship styles and climates; they range from a highly supportive to a highly authoritarian
environment, and each illustrates a different way of dealing with mistakes. Differences
in willingness to collaborate across professions are also evident, which is highly rele-
vant to catching and correcting drug errors, as will be illustrated below. I start with a
description of the most open unit, then proceed to the most authoritarian, and then fill
in the middle with two other units.
This unit is run by a nurse manager described uniformly by nurses, physicians, and
others interviewed as highly accessible and as inspiring top performance. Observed on
the job wearing blood-stained scrubs, she is a hands-on manager who actively invites
questions and concerns. Her own descriptions of the unit reveal a high degree of
respect for her subordinates: “The nurses problem solve together . . . they have good
relationships.” Further, she reports that “attending physicians” on the unit are “respect
ful of nurses’ expertise,” and that relations between nurses and physicians working in
the unit are smooth. Other nurses concur, with one describing relations between physi
cians and nurses as “more collaborative here than in other units.” By volunteering
these observations, several nurses reveal that physician-nurse relationships are salient
to them and suggest that harmonious relations are not the norm in the hospital. As will
be illustrated below, these relations are relevant to correcting drug errors.
In an interview, the nurse manager explains that a “certain level of error will occur,”
so a “nonpunitive environment” is essential to deal with this error productively. All
three staff nurses interviewed offered descriptions of how mistakes are handled that
suggest a nonpunitive environment has indeed been established. One nurse describes a
recent drug error, saying “there is no punishment; you just let the doctor know and fill
out an incident report.” Another nurse says that there is an “unspoken rule here to help
Edmondson / DETECTION OF HUMAN ERROR 81
Units Ranked According to Analysis of Qualitative Data
Memorial 1 University 1 University 3 Memorial 2 Memorial 4 Memorial 5 University 2 Memorial 3
High High Medium/High Medium/High Medium Medium/Low Low Very Low
Learn Tends toward
Learn Neutral Blame, fear Blame, fear Neutral Blame
“Nurse vs. man-
views of nurse
25%/75% 50%/50% 50%/50% 10%/90% 0%/100% 50%/50% 10%/90%
hands on vs.
Hands on Hands off, but
Hands on Hands on Hands on, but
Hands off Hands off Hands off, and
Scrubs Scrubs Usually scrubs Scrubs Business suit/
Scrubs or dress Business suit
views of staff
“They are too
hard on them
are capable and
and don’t take
should talk to
each other first
to me with a
Nurses feel a lot
hard on them
into [my] ‘the
“Nurses are al
as kids need
treats nurses in
same way, pays
tion to report
Staff’s views of
A superb nurse
not a boss”
“A rule person,”
“She talks to
want to cover
“Tends to blame
you the silent
“Treats you as
guilty if you
make a mis
you like a two-
Staff’s views of
cause of the
afraid to tell
because we get
ing our work
“Med errors are
not a big deal
people get in
don’t want to
ment is un
“You get put on
Units Ranked According to Quantitative Data (detected error rates) and
Juxtaposed With Independent Qualitative Ranking From Table 5
Memorial 1 University 1 University 3 Memorial 2 Memorial 4 Memorial 5 University 2 Memorial 3
Detected error rates
23.68 17.23 13.19 11.02 10.31 9.37 8.6 2.34
Interviewer’s overall rating
on openness High High Medium/High Medium/High Medium/Low Low Medium Very Low
a. All preventable ADEs and potential ADEs per 1,000 patient days. Mean detected error rate = 11.97 interceptions per 1,000 patient days (SD = 6.33); mean interceptions =
3.30 interceptions per 1,000 patient days (SD = 2.03); mean nonpreventable ADEs = 7.03 interceptions per 1,000 patient days (SD = 4.75).
each other and check each other,” and a third volunteers, “people feel more willing to
admit to errors here, because [the nurse manager] goes to bat for you.” Interestingly,
the nurse manager explained that “nurses tend to beat themselves up about errors; they
are much tougher on themselves than I would ever be,” revealing a managerial philoso
phy that renders a punitive stance unnecessary—as well as unproductive in serving the
larger goal of learning and improving.
A recent prevented drug error in this unit illustrates how errors can be intercepted
by the collaborative tendencies described above. A nurse on duty noticed that a partic
ular medication order appeared to be too high a dose. She telephoned the physician on
call at home, who confirmed the nurse’s concern and agreed with her recommendation
to cut the dose in half, and a potential ADE was avoided. This interdependence and
communication across professional group boundaries facilitates catching errors and is
not in evidence in all of the units, as will be seen below.
Interviews and observation in another unit in Memorial Hospital reveal a very dif
ferent climate. One nurse told us that the unit “prides itself on being clean, neat and
having an appearance of professionalism.” The nurse manager is dressed impeccably
in a business suit and has discussions with unit nurses behind the closed door of her
office. Several nurses volunteered that making a mistake here means “you get in trou-
ble,” and one nurse describing an incident in which she had hurt a patient while draw-
ing blood said that the nurse manager made her feel like she was “on trial; it was
degrading, like I was a two-year-old.” She continued, “I’ll probably get in trouble for
telling you this.” Many referred to the need to place the blame for mistakes. Another
nurse in this unit explained that Memorial Hospital “doesn’t support nurses; doctors
condescend, and they bite your head off if you” make a mistake. Moreover, “nurses are
blamed for mistakes”—revealing a degree of tension between professional groups in
this unit that leads to resentment and blame across group boundaries.
In an interview, the nurse manager was tense and agitated, and expressed anger
about having to “discipline the interns for repeatedly leaving their room messy.” Later,
walking by an interview between the researcher and a pharmacist, she comments with
out warmth or visible humor, “it looks like you’re plotting a revolution.” We also noted
that she required us to wear badges at all times, whereas in all other units except one,
badges were never mentioned, and wearing them was not required. This too was illus
trative of intergroup tensions; as researchers we felt less welcome here than in other
Another unit, rated as high to medium on openness based on the qualitative data,
illustrates a different way of being open. Without featuring the active coaching by the
nurse manager that was evident in Memorial 1, this unit is characterized by an appar
ently very high willingness to speak openly about mistakes and other threatening
issues. The nurse manager also wears scrubs, which she explicitly mentioned in an
84 THE JOURNAL OF APPLIED BEHAVIORAL SCIENCE March 2004
Edmondson / DETECTION OF HUMAN ERROR 85
interview as a way of “building commonality” with nurses. She also describes efforts
to engage unit members in team building and quality improvement processes, which
themselves “require open discussion of mistakes.” In discussing drug errors, she ech
oes an earlier comment from the nurse manager in Memorial 1 by saying that nurses
are “much harder on themselves” than she would ever be on them.
Nurses in this unit describe their manager as “distracted,” and “not quite focusing
on you when you’re talking to her,” but she “helps at the drop of a hat with unpleasant
tasks.” One nurse reports that the nurse manager “is cool about incident reports,” and
that “you feel bad already and are not made to feel worse.” Another says, “medication
errors are not a big deal around here,” and “problems exist here, but we have very can
did conversation about them.” She also notes that another unit (not one of those studied
here) is “very blame oriented, [unlike this one].” Indeed, observations of behavior in
this unit reinforce these descriptions, as nurses were seen talking openly and audibly in
candid conversations about errors, with no apparent concern whether they could be
overheard by the nurse manager. Members of Memorial 2 do not seem to convey the
same spirit of collaboration and trust that characterizes Memorial 1. Instead, they
describe a distracted but decidedly nonpunitive manager who—even though she does
not actively “go to bat” for them—is completely approachable and not a source of
threat. Their candid, nonfearful behavior supports this view.
Finally, a unit at University Hospital, rated as medium to low on openness based on
the qualitative data, is run by another apparently authoritarian nurse manager who is
described by nurses interviewed as “an authority not a coworker.” Her office is some-
what removed from the unit, and according to nurses interviewed, the “door is always
closed.” As in Memorial 3, she consistently wears a business suit rather than nursing
scrubs, and is viewed by unit nurses as “uncomfortable to deal with” and “inconsis
tent” in being supportive. In an interview, she explains that mistakes should be learn
ing experiences, but observes that “people are nervous about being called into the prin
cipal’s office” to talk to her about them. One nurse reports that the nurse manager has a
tendency to blame individuals for mistakes, and that “people don’t advertise error
here; if there’s no adverse event, then don’t report it.”
The behavior of nurses in this unit is noticeably less collaborative and supportive
than in Memorial 1 or 2. Those interviewed describe “backstabbing” and “cliques,”
and during observations on the unit, one nurse was overheard making a mean-spirited
comment behind the back of another nurse while explaining why she refused to help
the other nurse.
Interviews and observations in the eight units reinforce quantitative findings that
shared perceptions about the consequences of making mistakes influence the climate
and reporting behaviors within a unit team. A picture emerges from the qualitative
results described above that is consistent with the earlier interpretation of the unex
pected quantitative results. It appears that nurse manager behaviors are an important
influence on unit members’ beliefs about the consequences and discussability of mis
takes. In addition to the influence of what is said by the nurse manager, the ways past
errors have been handled are noticed, and conclusions are drawn, which then are
strengthened by ongoing conversations among unit members. In this way, perceptions
may become reality, as the perception that something is not discussable leads to avoid
ance of such discussions. These kinds of perceptions, when shared, contribute to a cli
mate of fear or of openness, which can be self-reinforcing, and which further influ
ences the ability and willingness to identify and discuss mistakes and problems. These
climates are characterized in part by the nature of relationships within and between
professional identity groups.
IMPLICATIONS AND CONCLUSIONS
This article reports a positive answer to the question guiding this study; differences
in unit properties do appear to be associated with differences in error rates. However,
the relationship discovered here was an unexpected one, suggesting that a primary
influence on detected error rates is unit members’ willingness to discuss mistakes
openly. Thus a model is proposed in which leadership behavior influences the way
errors are handled, which in turn leads to shared perceptions of how consequential it is
to make a mistake. These perceptions influence willingness to report mistakes, and
may contribute to a climate of fear or of openness that is likely to endure and further
influence the ability to identify and discuss problems.
In this research, as in organizational life, actual errors are confounded inextricably
with detected errors. Despite its thoroughness, the process of documenting errors is
still partly dependent upon organizational members’ willingness and ability to detect
and report errors within their units. Thus there are at least two sources of influence on
the dependent variable of detected errors—actual differences in errors made and will
ingness to expose them—and disentangling the relative contribution of each source is
not feasible in most naturalistic organizational settings. Organizational characteristics
influence both of these behavioral outcomes; however, conditions that foster making
errors are likely to be different from those that foster catching, correcting, discussing,
and learning from errors. We can speculate that actual medication errors may be lower
in units that perform as more tightly coordinated teams and have more accessible, open
nurse managers; however, from these data we cannot support this hypothesis directly.
What we do learn is that willingness to report errors varies systematically with per
ceived openness of unit leaders, and we can speculate that these attributes may over
whelm differences in actual error rates. This speculation is supported by other research
discussed above that reveals that the rate of detection of actual drug errors is orders of
magnitude below 100%.
These findings provide evidence that the detection of error is influenced by organi
zational characteristics, suggesting that the popular notion of learning from mistakes
faces a management dilemma. Detection of error may vary in such a way as to make
those teams that most need improvement least likely to surface errors—the data that
fuel improvement efforts. This has important implications for quality improvement
86 THE JOURNAL OF APPLIED BEHAVIORAL SCIENCE March 2004
efforts that rely upon work teams’ participation in detecting and correcting error, by
suggesting that there may be barriers that prevent some (more than other) teams from
doing so. Communication failures caused by intergroup tensions may also affect the
ability of teams to discuss and correct mistakes. Michael (1976) has suggested that
embracing error is feasible when organizations reward such behaviors, and this study
is supportive of that proposition. Organizational and group interventions may be
needed to encourage detection and discussion of error, although strategies for accom
plishing this are beyond the scope of this article.
Finally, this research suggests that the group level of analysis offers a useful per
spective for investigating the phenomenon of errors in organizations. Research at the
individual level of analysis tends to suggest educational and technical interventions to
reduce the incidence of errors, whereas researchers at the systems level of analysis
warn of the perverse effects of targeted technical solutions. Given that human error
will never disappear from organizational life, an important management issue thus
becomes the design and nurturance of work environments in which it is possible to
learn from mistakes and collectively to avoid making the same ones in the future. This
research contributes to this goal by pointing to conditions at the group level that may
influence the degree to which errors are caught and corrected by work teams.
1. This study involved two urban tertiary care hospitals; University Hospital and Memorial Hospital are
2. The Prevention of Drug Complications Study at the Harvard School of Public Health was funded by
the Agency for Health Care Policy and Research. Physicians, nurses, and social psychologists collaborated
to explore the phenomenon of drug errors from diverse perspectives.
3. I thank Clayton Alderfer for this insight, which has helped me to view this study in a new way.
Although the research design relied on the latter (structural/intellectual) tradition, the results required me to
take a new look at the former (emotional/unconscious) tradition.
4. It is important to note that this “hard” measure consists of those drug errors identified by the research
team, and thus is not one and the same as “actual error rates.” More will be said about this distinction in the
Results and Discussion section.
5. I am indebted to Andy Molinsky who conducted most of the on-site interviews and observations; his
qualitative data are summarized in Tables 5 and 6 and form the basis of the four individual cases described in
the Results section.
6. The respondents include 123 nurses, 19 physicians, and 4 pharmacists. Nurses are full-time members
of the eight units; pharmacists and physicians who work with or rotate through the units were asked to
judge—as relative outsiders—how units that they know well function. Thus, although 77% of the surveys
returned and analyzed were from the full-time unit staff, responses from nurses and physicians were similar;
we tested for differences across these groups and found none.
7. Adam Galinsky’s assistance in administering the survey and his endless supply of ideas and enthusi
asm during this project are acknowledged with great appreciation.
8. In the case of positive leniency, the denominator (expected variance) will be too large to represent
accurately the response patterns reasonably expected for this survey, such that James’s formula (unity minus
actual variance divided by expected variance) will produce an inflated IRR.
Edmondson / DETECTION OF HUMAN ERROR 87
9. The IRR also differs from the ICC in that it provides individual coefficients for each unit, enabling
10. Because patients in oncological and intensive care units receive more medications each day than
patients in general medical or surgical units, several of the physicians involved in the larger drug study sug
gested that I adjust the dependent variables for exposure to medications, to make these values more compara
ble across units. Moreover, the positive correlations may indicate simply that ICUs have more actual drug
errors while also having better coordinated teams and more active nurse manager coaches. Thus, following
this suggestion, I reexamined the correlations between unit survey variables and drug rates, using drug error
data adjusted for exposure. (Each dependent variable was divided by mean number of drugs per patient per
day in that unit and multiplied by the mean across units, thus deflating ICU values and inflating values for
general care units.) The result, rather than weakening the reported correlations, was an overall strengthening
of the relationships, as follows:
Detected Error Rates and . . . Correlation Correlation Using Adjusted Data
Nurse manager coaching .74* .76*
Nurse manager direction setting .74* .88*
Unit performance outcomes .76* .74*
Quality of unit relationships .74* .88*
Willingness to report errors .55 .69
Mistakes (are not held against you) .44 .67
*p < .03.
Intercepted Errors and . . . Correlation Correlation Using Adjusted Data
Nurse manager coaching .71* .69
Nurse manager direction setting .83* .81*
Unit performance outcomes .71* .64
Quality of unit relationships .76* .67
Willingness to report errors .62 .62
Mistakes (are not held against you) .45 .33
*p < .03.
The results reported in the article use the original detected error data, as the phenomenon discussed is that
of reporting differences across units. With this perspective, each instance of a nurse willingly reporting a
drug error is an illustration of proactive, learning-oriented behavior, which contributes to the prevention of
future errors. The actual count of the number of times such a behavior occurs in each unit is thus a better
dependent variable for this study than the more conservative adjusted error data.
The principal investigators of the drug study also suggested that I do the above analyses using only
PADEs, which can be only obtained through voluntary reporting, instead of using the current dependent vari
able of total error rates (PADEs plus preventable ADEs, the latter measure being obtained by chart review
and voluntary reporting). Separating PADEs and ADEs illustrates the relatively stronger relationships
between unit characteristics and those errors only voluntarily reported, supporting the interpretation that unit
differences in willingness to report errors accounts for much of the variance in detected error rates. More
over, the existence of (weaker) correlations between unit characteristics and ADEs (obtained in chart review
and voluntary reporting) support the hypothesis of a relationship between error reporting and error rates.
88 THE JOURNAL OF APPLIED BEHAVIORAL SCIENCE March 2004
PADEs and . . . Correlation Correlation Using Adjusted PADEs
Nurse manager coaching .79* .81*
Nurse manager direction setting .76* .93*
Unit performance outcomes .64 .69
Quality of unit relationships .62 .74*
Willingness to report errors .33 .50
Mistakes (are not held against you) .33 .52
*p < .03.
ADEs and . . . Correlation Correlation Using Adjusted PADEs
Nurse manager coaching .31 .24
Nurse manager direction setting .41 .38
Unit performance outcomes .52 .64
Quality of unit relationships .48 .41
Willingness to report errors .45 .50
Mistakes (are not held against you) .45 .50
Alderfer, C. (1987). An intergroup perspective on group dynamics. In J. W. Lorsch (Ed.), Handbook of orga-
nizational behavior. Englewood Cliffs, NJ: Prentice Hall.
Argyris, C. (1985). Strategy, change, and defensive routines. Boston: Ballinger.
Barker, K. N., & McConnell, W. E. (1962). Detecting errors in hospitals. American Journal of Hospital
Pharmacy, 19, 361-369.
Bates, D. W., Boyle, D. L., Vander Vliet, M. B., Schneider, J., & Leape, L. L. (1995). Relationship between
medication errors and adverse drug events. Journal of General Internal Medicine, 10, 199-205.
Bates, D. W., Leape, L. L., & Petrycki, S. (1993). Incidence and preventability of adverse drug events in hos
pitalized patients. Journal of General Internal Medicine, 8, 289-294.
Classen, D. C., Pestotnik, S. L., Evans, R. S., & Burke, J. P. (1991). Computerized surveillance of adverse
drug events in hospital patients. Journal of the American Medical Association, 266(20), 2847-2851.
Cohen, M. R. (1977). Medication error reports. American Journal of Hospital Pharmacy, 12(1), 42-44.
Evans, R. S., Pestotnik, S. L., Classen, D. C., Bass, S. B., Menlove, R. L., Gardner, R. M., & Burke, J. P.
(1992). Development of a computerized adverse drug event monitor. In Proceedings of the Annual Sym
posium on Computerized Applications for Medical Care, pp. 23-27.
Fiske, S. T., & Taylor, S. E. (1991). Social cognition. New York: McGraw-Hill.
Forrester, J. W. (1971). Counterintuitive behavior of social systems. Technology Review, 73(3), 52-68.
Foushee, H. C., Lauber, J. K., Baetge, M. M., & Acomb, D. B. (1986). Crew factors in flight operations III:
The operational significance of exposure to short-haul air transport operations (Technical Memoran
dum No. 88342). Moffett Field, CA: NASA-Ames Research Center.
Goleman, D. (1985). Vital lies simple truths: The psychology of self-deception. New York: Simon &
Hackman, J. R. (1987). The design of work teams. In J. W. Lorsch (Ed.), Handbook of organizational behav
ior. Englewood Cliffs, NJ: Prentice Hall.
Hackman, J. R. (1990). New directions in crew-oriented flight training. Unpublished proceedings of the
ICAO Human Factors Seminar, Leningrad.
Hackman, J. R. (1993). Teams, leaders, and organizations: New directions for crew-oriented flight training.
In E. L. Wiener, B. G. Kanki, & R. L. Helmreich (Eds.), Cockpit resource management. Orlando, FL:
Edmondson / DETECTION OF HUMAN ERROR 89
Hackman, J. R., & Morris, C. G. (1975). Group tasks, group interaction process, and group performance
effectiveness. In H. H. Blumberg, A. P. Hard, V. Kent, & M. Davies (Eds.), Small groups and social inter
action (Vol. 1). Chichester, UK: Wiley.
Helmreich, R. L., & Foushee, H. C. (1993). Why crew resource management? Empirical and theoretical
bases of human factors training in aviation. In E. L. Wiener, B. G. Kanki, & R. L. Helmreich (Eds.),
Cockpit resource management. Orlando, FL: Academic Press.
James, L. R., Demaree, R. G., & Wolf, G. (1984). Estimating within-group interrater reliability with and
without response bias. Journal of Applied Psychology, 69(1), 85-98.
Kenny, D. A., & LaVoie, L. (1985). Separating individual and group effects. Journal of Personality and
Social Psychology, 48(2), 339-448.
Leape, L. L., Brennan, T. A., Laird, N., Lawthers, A. G., Localio, A. R., Barnes, B. A., Hebert, L., Newhouse,
J. P., Wyler, P. C., & Hiatt, H. (1991). The nature of adverse events in hospitalized patients: Results of the
Harvard Medical Practice Study II. New England Journal of Medicine, 324(6), 377-384.
Lesar, T. S., Briceland, L. L., Delcoure, K., Parmalee, J. C., Masta-Gornic, V., & Pohl, H. (1990). Medication
prescribing errors in a teaching hospital. Journal of American Medicine, 263, 2329-2334.
Lewin, K., Lippitt, R., & White, R. (1939). Patterns of aggressive behavior in experimentally created social
climates. Journal of Social Psychology, 10, 271-299.
Maier, N.R.F. (1967). Assets and liabilities in group problem solving: The need for an integrative function.
Psychological Review, 74, 239-249.
Massaro, T. A. (1992). Introducing physician order entry at a major academic medical center: II. Impact on
medical education. Academic Medicine, 68, 25-30.
Melmon, K. L. (1971). Preventable drug reactions: Causes and cures. New England Journal of Medicine,
Michael, D. N. (1976). On learning to plan and planning to learn. San Francisco: Jossey-Bass.
Norman, D. A. (1980). Post-Freudian slips. Psychology Today, 13, 42-50.
Norman, D. A. (1981). Categorization of action slips. Psychological Review, 88, 1-15.
Perrow, C. (1984). Normal accidents. New York: Basic Books.
Reason, J. (1984). Lapses of attention in everyday life. In R. Parasuraman & D. R. Davies (Eds.), Varieties of
attention (pp. 515-549). Orlando, FL: Academic Press.
Rice, A. K. (1958). Productivity and social organization: The Ahmedabad experiment. London: Tavistock.
Roethlisberger, F. J., & Dickson, W. (1939). Management and the worker. Boston: Harvard University Press.
Rumelhart, D. E. (1980). Schemata: The building blocks of cognition. In R. J. Spiro, B. C. Bruce, & W. F.
Brewer (Eds.), Theoretical issues in reading comprehension: Perspectives from cognitive psychology,
linguistics, artificial intelligence and education (pp. 33-58). Hillsdale, NJ: Lawrence Erlbaum.
Senge, P. (1990). The fifth discipline: The art and practice of the learning organization. New York:
Trist, E. A., & Bamforth, K. W. (1951). Some social and psychological consequences of the Longwall
method of coal-getting. In D. S. Pugh (Ed.), Organization theory (pp. 393-419). London: Penguin.
90 THE JOURNAL OF APPLIED BEHAVIORAL SCIENCE March 2004