Compact Introduction to Human Error

Stu Moment
University of Illinois Human Factors Division
©2008, Stu Moment
Stu Moment web page

APA reference: Moment, S. L. (2008). A Compact Introduction to Human Error, University of Illinois Human Factors Division Proceedings, Retrieved [month] [day],[year], from http://www.humanfactors.illinois.edu/research/HumanElementArticles/CompactIntroToHumanError/

Copyright notice: This publication may be used, reproduced, printed and redistributed for personal, academic, research or non-commercial purposes as long as 1) it is not modified, 2) credit is attributed to the Human Factors Division at the University of Illinois at Urbana-Champaign and the author, and 3) the copyright notice and this notice are reproduced on any copies. If you have any questions regarding distribution of this paper contact the author.

Human Error is relatively new concentration of Human Factors. Until recently, it’s causal [construct], analysis and intervention methods have been more art than a computational science, yet advances in the science occur as qualified factors are identified and organization-wide analysis are created.

The application of analytical methods in human error evaluation is in its infancy. Tools are being utilized and more important, data is being collected. Explanations of the underlying causes of human error at the individual level, developed over the last 60 years, show stability and growth on a relatively solid foundation of knowledge. Philosophies of human performance in a job environment have been refined to a point where most contributions to an error may be mapped, thus giving today’s safety systems designer and analyst useful structures and methods.

The definition of human error appears simple, errors caused by humans rather than machines. But such a definition is too shallow for fruitful analysis and corrective prescription. 50 years ago human error was often described just at the level of the human directly involved with error. One example of this assignment is Coast Guard investigations of ferry operators. The investigations assigned the cause of most errors as, inattentiveness, poor judgment or negligence by crew members. Modern definitions look to failures at any point in a system within an organization, from micro-task decisions and task implementation, personal monitoring systems, to team member assignments and actual coordination of middle organizational decisions relating to the implementation of people and equipment chosen for task accomplishment. The systems themselves are scrutinized for their tolerance to human frailty. These systems are designed by humans. Human error can occur in any stage of a task or strategy where a human is involved.

How should we look at error? In Sheridan’s introduction to a famed Rasmussen article he states, "Error is simply a difference between an actual state and a desired state” (Sheridan 2003). But important errors are ones which adversely affect direct personal well being, the environment, or a system which society considers important. In all but the most basic systems, a performer will never achieve an optimal state. Yet adverse states will occur when a threshold of performance is not met or when multiple performance thresholds produce an adverse condition.

Many references to human error are connected with high-profile catastrophes. Nuclear power plant accidents, oil rig fires and chemical plant disasters are very apparent to the public and their investigation [solicits] human error analysis.

Public transportation accidents are more constant examples of errors which are in the public spotlight. The recorded investigation of railroad accidents in the United States began in 1911. Aircraft accident investigations became formal in the 1920’s. In 1939, New Jersey created the “Traffic Bureau” which was charged with, “compiling accident statistics, conducting studies of congestion, accident causes, and the effectiveness of current safety measures.” The Occupational Safety and Health Act of 1970 mandated reporting of at work accidents. After the 1999 report, To Err is Human: Building a Safer Health System, The Federal Government specified programs to improve error analysis in our health systems. States are starting similar programs.

In all of these examples, government is the investigative and analyzing agency. Internal company investigations have historically been for the purposes of protecting the company in the case of law suit. But due to competitive safety, internal economics and reduced potential of claimed negligence, both from the standpoint of the accident itself and from the standpoint of practiced safety emphasis, more companies are analyzing incidents internally. The “Company Safety Officer” has become a standard position in many larger institutions.

Many references to human error are connected with high-profile catastrophes but the real distribution of accidents are reflected in OSHA documents. Agriculture, carpentry and non-commercial transportation including driving and boating, lead the list when data is grouped by industry. General at work accidents, not associated with units main purpose, account for a surprisingly large number of injuries when data is not grouped by industry. (U.S. Department of Labor 2007)

The notion of human error should extend beyond tangible incidents. Economic planning and military command errors are as valid a target for analysis as a medical diagnosis error.

“Safety culture has previously been defined as the enduring value and prioritization of worker and public safety by each member of each group and in every level of an organization. It refers to the extent to which individuals and groups will commit to personal responsibility for safety; act to preserve, enhance and communicate safety information; strive to actively learn, adapt and modify (both individual and organizational) behavior based on lessons learned from mistakes; and be held accountable or strive to be honored in association with these values.” (von Thaden T. L., Gibbons, A. M. 2008)

The ability to communicate safety emphasis can easily be improved in any organization but many classic studies in organizational psychology illustrate organizational behavior phenomena which become barriers to commitment and risk assessment in organizations. Ambiguity in job definitions and safety personnel’s power, is aggravated by lack of knowledge, resources or a perceived high cost of system fixes.

At the individual practitioner’s level, worry of legal exposure or job repercussions will limit the flow of information on all aspects of organizational safety except in the case of a mandatory accident report.

vonThaden, Kessel and Ruengvisesh divide their safety culture questionnaires and their analysis of safety culture into general groups of formal safety program, informal aspects of safety, organizational commitment, operations personal and overall safety (von Thaden, T. L., Kessel, J., Ruengvisesh, D. 2008). A subcategory within “formal safety program” is “reporting system”. The questionnaire solicits attitudes about fear of negative repercussions, ease of use and the respondent’s opinion of other employee’s use of incident reports.

The acceptance and use of a good reporting system, by all levels of employees, may not be an indicator of decent safety climate in itself, but is a necessary prerequisite to establishing data collection. Data collection is the weak point of human error analysis.

The psychological constructs of limited human performance benefited from many studies in the 1950’s and 60’s. Human Error studies prominent in the 1980’s include studies of basic notions of cognitive failures and add notions of error systems and error taxonomies.

Egon Brunswik set a foundation for the analysis of human decision making for tasks requiring evaluation. Brunswik’s lens model relates human judgements to cue availability, cue utilization and cue validity in a probabilistic fassion (Brunswik 1955). Brunswik’s works, as adapted by cognitive engineering researches, became the core of many models of human performance.

Kenneth Hammond applied Brunswik’s notions in human performance studies. Hammond's 1972 article, Cognitive Control, gives us many considerations to include in analysis of judement/decision making tasks (Hammond 1972).

Donald Norman’s often quoted 1980 article, Errors in Human Performance, expands on the notion of mistakes vs. slips, a dichotomy which also appears in many modern analysis (Norman 1980). Norman’s approach to Human Error goes beyond just those observable behavior types. Norman observes that accidents are difficult to categorize and that the human operator is working in a system.

Jens Rasmussen’s often quoted 1983 article, Skill, Rules, and Knowledge; Signals, Signs, and Symbols, and Other Distinctions in Human Performance Models, presents three categories of human behavior, skill-based, rule-based and knowledge-based, which make their way into modern error analytical systems (Rasmussen 1983).

David Meister’s 1989 article, The Nature of Human Error, expands on the system’s role in error (Meister 1989). While not prescribing a specific taxonomy, Meister describes many pieces of error systems from which a taxonomy for a specific negative performance can be constructed.

Perhaps the most popular human error writer, James Reason, combines much of the earlier notions of both the errors themselves and the systems in which humans work, into a the general but complete “Swiss Cheese” model through which errors are allowed to occur, in his 1990 book, Human Error (Reason 1990).

There are many other names which should be included in this evolution of Human Error study. Also, there are many new names presenting credible expansion of generic Human Error Analysis as well as people expanding the study into their specific industry, an extremely important lateral development since no single analysis method fits many systems.

A quick mention of a few contributing sub-disciplines is important since productive error analysis depends on this root-level knowledge as applied within identified human/organization/equipment system types as well as the design of improved systems. Organizations need to staff or consult people with root-level knowledge in areas appropriate to their systems, often independent of the group chosen to develop analysis systems. These contributing sub-disciplines are not mutually exclusive. Some disciplines gain from or build on findings in other disciplines.

Basic Cognition
The implications of cognitive man go beyond just the notion that we know and think as opposed to exhibit programmed reactions. Our ability to make decisions involve cognitive principles of perceptions and reasoning. Bloom’s cognitive level taxonomy of knowledge, comprehension, application, analysis, synthesis and evaluation give considerations applicable to specific job demands and the systems designed to help man meet those demands. Research examining the transition of a human in a specific task, from evaluative to rote reaction responses, has implications where forcing evaluation is always necessary, even when diversions from normal are seldom encountered.

Working memory
Research has found that we have very limited working memory. Working memory is the memory directly involved with a micro-analysis in a task. Working memory is also used to retrieve information from long term memory. System designs may allow working memory to be overloaded. Performance degradation during overload has been measured in many studies. Conditions which force overload have also been studied.

Attention
In their book, Applied Attention Theory, Wickens and McCarley present chapters, among others, on attention control, information sampling, resources and effort, time sharing, interruptions and task management (Wickens and McCarley 2008). The findings in the study of attention apply directly to most task systems and improvements made in such systems can contribute greatly to error reduction.

Schema (development and activation)
Schema is a human’s collection of information and practiced evocation of knowledge, behavior or required action on a particular [topic]. A person’s schema which surfaces at a particular time is dependent on the environment and actions which bring that schema out.

Early schema research came from two different disciplines, learning psychology and personality disorder psychology. Studies in learning have concentrated on how best to learn. Those studies better define the notation that new learning is attached to some old knowledge. Individual areas of practiced knowledge are part of the whole set of schemas, termed schemata.

Studies in personality disorders have successfully measured schema in areas some academicians term as personality factors. Personality factors are formed in much the same way as knowledge schema, that is, by attaching new information to information currently in mind and associated, at the time, with a certain state or assessment of the environment. Like knowledge schema, different [formations] of a topic related schema are brought out by different activators. For the sake of incident analysis, we will include personality factors as schema assessments.

Schema may not be ignored. It affects many task types. Schema that is difficult to measure can still be included in error analysis and system design. After-the fact references to schema effects will acknowledge its importance to some systems. Desired schema activators can be programmed into systems.

Some writers introduce the “Nature of Human Error” in terms of unwanted occurrences just waiting to happen. The fallible human does little faultless activity while getting ready for work, eating, communicating with a spouse, driving to work or while working. Errors may range from sub optimal activities such as driving a route which adds miles or time, to activity with the potential of a negative consequence such as forgetting to relocate flammable liquids to a safe place after unloading.

In [contrast] with such a bleak look at human fallibility, we humans perform rather well at work. In the work place, the human is placed into a system which controls behavior relevant to a specific role. A work role places the human into a structure which directs important activities in a micro-world, that is a relatively small number of relevant occurrences, and usually, a relatively small number of desired responses. Systems designs have evolved to reduce the potential for error in desired responses. Some systems evolved informally and some by a modern concerted effort.

The system must be designed to reduce errors specific to the task-type/error-type. In addition, a good system design should account for the performance modifiers of physical or mental personal states and variable environmental conditions.

The latent conditions approach to human error potential may best fit in task environments with established rules or procedures and few evaluative portions of the task.

Whereas “fallibility” is a term usually attached to the imperfect human, “latent conditions” describe potential imperfections in a system which allow errors to occur. However one desires to think of fallibility and latency, the importance behind the concepts are to block possible paths of error.

James Reason characterized the “trajectory of accident opportunity” with his “Swiss cheese” model of accident causation (Reason 1990). This model separates the performance of unsafe acts from psychological precursors and organizational defenses. In order for an accident to occur, the trajectory of accident opportunity passes though holes in the cheese slices. The unsafe act is an observable error such as choice of a wrong procedure, task miss-prioritization, or a violation of established methods. Reason’s psychological precursors or preconditions include failure to use safety equipment, personal stress and many other conditions which modify performance. Reason’s defensive systems include among other things, engineered safety features and safety equipment requirements, rules and procedures, training, drills etc. (Reason 1997)

The Department of Defense’s HFACS analysis system is a descendent of Reason’s 1990 Swiss cheese model (Department of Defense 2005). Four levels where failures have occurred are analyzed; organizational influences, unsafe supervision, preconditions for unsafe acts and the unsafe acts themselves. This adaptation of Reasons architecture was developed by Wiegmann and Shappel and published in their book, A Human Error Approach to Aviation Accident Analysis, The Human Factors Analysis and Classification System (Wiegmann and Shappel 2003).

Reason’s latest organization of error analysis, simplifies the cause assignment and investigation into three levels; the organization, error provoking conditions and the unsafe acts. (Reason 1997) Reason’s newest analytical structure does more than reorganize the four tier model which evolved into HFACS. Reason uses the three tier model for both analyzing causes and initiating investigation into errors.

As we organize the elements that allow human error, we will mix many author’s arrangements of error factors and place these elements into a refined three level structure, the system level, error potential modifiers, and the act task-type/error-type. This arrangement does not define a taxonomy, but instead defines a structure to build taxonomies specific to an organization. Within the structure, mathematical relationships may be added for help with system design.

For the purpose of system identification, we will create a new acronym for this 3 level system, the Human Environment Diagnostic/Design System (HEDS). Again, this system is not a new format for error evaluation, just a customizable mix and modification to previous systems. Multiple factors may be viewed in parallel for any given incident. Intervention strategies are directly derived from this analysis.

A systems approach to task accomplishment does differ from the HFACS/Swiss cheese model in one important way. Whereas "defenses to latent conditions" appear the be the theme of the HFACS/Swiss cheese model, a more positive "locating potential for improved performance" may be integrated into a systems approach.

The cognitive engineering approach may best serve task environments with many evaluative tasks.

Cognitive engineering models vary, but include common components of the environment, human perceptions and the achievement desired of the human. Judgments are influenced by cues obtained from the environment. Cue availability as well as the practitioner’s assessment of their ecological validity affect judgment. Technological aids which give a practitioner more information and do so in a non-confusing way are offensive tools for error reduction.

Decision making will also be influenced by many system level factors as well as what we will term, error potential modifiers. Also the desired short term goal of the human’s behavior may not be as well specified as in a task structure of established rules and procedures. Many judgments may be part of a series of judgments required to obtain a goal.

The effects of the practitioner’s desired outcome adds another complicating factor to the strategy of ecological improvement. The practitioner may not know when to use a defensive mode of decision making, that is, when to minimize the probability of adverse effects as opposed to maximizing the probability of positive effects.

Systems vary with the complexity of the process, the potential for error and the cost of error. Systems level implementations may take a variety of forms from technology/equipment specifications, monitoring systems, specified procedures, checklists, team member requirements, training and re-training requirements, etc. These systems should be implemented with a knowledge of, and connection to, task types and task duration.

Simple tasks may require fewer systems. A carpenter may be working in simple systems. A carpenter may be required to use shields on a cut-off saw. A carpenter may not be allowed to work unless another employee is present. A carpenter’s task may involve many short duration tasks, some of a repetitive nature and some with variety.

A medical practitioner may be involved in complex systems with sophisticated equipment requirements, strict crew-member requirements and constant monitoring requirements. The duration of a medical task may have considerable variance. Monitoring the results of a procedure may cover a large time frame. Parts of a procedure may appear rote. Other parts may require considerable evaluation and knowledge search.

The safety systems designer must design systems to improve task performance but must always assume imperfect performance [caused] by either common task errors or [magnified] by error potential modifiers.

Reason’s psychological precursors were presented in the previous section. Wiegmann & Shappel referred to this layer as “preconditions for unsafe acts” and add to this category, personal subfactors such as personal readiness and the environmental factors of both physical and technological environment. Wiegman and Shappel divide “condition of operators” into adverse mental, physiological and physical/mental states (Wiegmann and Shappell 2003).

When creating categories for a particular organization, the safety system designer may need to place similar sounding factors at more than one of the three simplified levels. Technology/equipment specification normally belongs at the system level yet a valid reference to how an individual interfaces with the technological environment (or its failure) can be made in this Error Potential Modifier section.

Other factors which may be placed with specificity in this section include time pressure, [importance] pressure, attention predisposition, procedural memory activators, attention overload and relevant schema base/activation variables which are not part of the system design.

Basic error types identified by Rasmussen are skill-based, rule-based and knowledge-based. DOD HFACS uses a similar approach to the classification of errors types. The error types are skill-based, decision-making or perceptual.

Because DOD HFACS was designed for the analysis of aviation errors, many organizations will find other categories useful. Errors of omission may be an oft used category for tasks with multiple analytical stages. Attention error breakdowns could be useful in automated systems. Ambiguity caused errors are also found where equipment malfunctions are disbelieved or crew member authority is poorly defined.

Basic error types, as defined by Reason, are slips/lapses, rule based mistakes and knowledge based mistakes. Reason incorporates these types in his “unsafe acts” slice of the Swiss cheese model with subdivisions of unintended action and intended action. Unintended actions include slips and lapses. Intended actions include mistakes and violations. Different systems based on the Swiss cheese/HFACS model of analysis subdivide violations into one or more categories depending on the cause and organizational acceptance of the violation. Table 1 illustrates error types in the unsafe acts section of DOD HFACS.

Task types, defined in a systems description, should be connected to error types in a good analysis system.

Due to differences in system level structures, no existing analysis system gives a perfect fit for their individual error situations. The individual analyst must outline the peculiarities of the system and tasks in a particular job and either adapt currently available systems or invent new analysis.

Human Error analytical techniques may be divided in to predictive models and post-incident investigation. Predictive models include probabilistic risk assessment models and cognitive engineering models. Predictive models will not be covered in this article but it is important to note that predictive models will improve as post-incident investigation models create data relevant to them.

Categorizing
Classic categories often do not contribute directly to error analysis but may be useful when error types are examined by category. The classical categories in the medical field: diagnosis, drug and medical procedures will probably reveal different error systems, modifiers and types, when data is retrieved by individual category.

Besides classical categories, other descriptive categories can be built into report and investigation forms. Certain questions on the circumstances surrounding the incident become natural categories on which to examine for differences when report quantities become substantial.

Root Cause
Root cause analysis has its roots in the analysis of machine failures. In the 1970’s and 1980’s, most documented applications of root cause analysis were made in product quality assurance and equipment failures. By the early 1990’s root cause analysis was extended to cover issues beyond equipment. In his 1991 article, In Search of the Root Cause, John Dew professed “Systemic issues concern how the management of the organization plans, organizes, controls, and provides quality assurance and safety in five key areas: personnel, procedures, equipment, material, and the environment.” (Dew 1991). But use of root cause analysis in human affected systems may be limited to the human's duties in the system as opposed to the causes of suboptimal human performance. The Joint Commission on the Accreditation of Healthcare Organizations (JCAHO) avoids the issue and notes “A root cause analysis focuses primarily on systems and processes, not on individual performance.” (JCAHO 2009)

Some credible applications of root cause(s), in support of human error analysis, can be made, mostly, to discover seminal events which, when similar seminal events occur with [high] frequency, will pinpoint an area of needed system improvement. A seminal event is an event which, an analyst decides, begins a chain of events which leads to an error. While root cause analysis may have a place in human error investigation, it contributes little to the examination of simultaneous holes in systems and defenses against probable modifiers. Root cause analysis usability is further diminished when multiple parallel conditions combine to enable an error. Despite the limitations of root cause analysis, a good investigation system will allow a remitter or investigator to label a contributing element as “seminal”, thus allowing future inspection of common seminal events as data size grows.

Other Industry Inspired Methods.
Traditional Methods other than root cause have made their way into human error literature. Fault Tree Analysis, Probability Tree Analysis and Markov Reliability Analysis have provided quantitative reliability and error analysis in machine systems. While, like root cause analysis, they are not built on the human error foundations of knowledge, their application to systems involving humans can produce helpful task design considerations.

Multi-tier Models
Multi-tier type models like HFACS and the simplified three-tier models, have been described in previous sections. These type of analysis are built on the human error foundations of knowlege. Multi-tier analysis, which tightly connects the levels of error, will provide useful data for analytical methods currently available or to be developed in the future. Vertical connection refers to the grouping of system/environment, modifiers and error types for each error noted in an incident, and, keeping them separate from the factors associated with other errors found in the same incident. Table 2 contrasts poor factor assignment from vertically connected factor assignment. Vertically connected analysis can better pinpoint acceptable system improvements and intervention strategies.

Cognitive Task Models
Cognitive control architectures relate the human’s assessment of the environment to judgments, decisions or actions. Ecological effects are only [crudely] involved in current frameworks of error analysis, yet may be the foremost contributor to error assessment when knowledge-based errors occur.

Two mainstream cognitive assessment models, the lens model and the ACT-R model enable quantitative assessment of error contribution to tasks. Like the traditional industrial reliability models, their application can be made to local tasks in a system.

Cross-Mapping
The use of any human error analysis method may be mapped to other methods if there is relational identification to a case number in all descriptions, categorical, hierarchical, seminal, system reliability or cognitive task models. When using hierarchical models, inclusion of comments associated with a particular vertical column, will aid transfer of finished analysis to improved systems where new factors are included.

Correct intervention strategies will be found directly in a well designed analysis system. The costs of each error reduction strategy or contributors to that strategy will limit the choices.

Safety system design (or engineering in systems where mathematical relationships are developed) allows for many solutions.

Training is often mentioned as a cure for problems but training is suboptimal or ineffective as a band-aid for systems which are in need of improvement. The following is a short-list of possible improvements.

Safety analysis, system design and error reduction go hand in hand with a well designed analysis system. Conversely, without a good analysis system, error improvement is much less efficient.

Brunswik, E. (1955) Representative Design and Probabilistic Theory in a Functional Psychology, Psychological Review, Vol 62(3), 193-217.

Hammond, K.R., Summers, D.A. (1972) Cognitive Control, Psychological Review, Vol 79(1), 58-67

Meister, D. (1989), The nature of human error, IEEE Global Telecommunications Conference, Communications Technology for the 1990s and Beyond, volume 2, 783-786

Norman, D.A. (1980), Errors in Human Performance, University of California, San Diego, Center for Human Infromation Processing Report No. 8004

RASMUSSEN, J. (1983), Skill, Rules, and Knowledge; Signals, Signs, and Symbols, and Other Distinctions in Human Performance Models, IEEE Transactions on Systems Man and Cybernetics, Vol. 13(3) 257-266

Reason, J. (1997), Managing the Risks of Organizational Accidents. Aldershot: Ashgate Publishing Limited

U.S. Department of Defense (2005). DoD HFACS, Attachment 1, Retrieved November 3, 2008, from http://www.safetycenter.navy.mil/hfacs/downloads/hfacs.pdf

U.S. Department of Labor, Bureua of Labor Statistics (n.d.). Fatal occupational injuries by industry and event or exposure, Retrieved October 12, 2008, from http://www.bls.gov/iif/oshwc/cfoi/cftb0223.pdf

Wickens, C. D., McCarley, J. S. (2008), Applied Attention Theory, Boca Raton: CRC Press.

Wiegmann, D. A., & Shappell, S. A. (2003), A human error approach to aviation accident analysis: The human factors analysis and classification system, VT: Ashgate.