Beyond root cause analysis: How variation analysis can provide a deeper understanding of causation in complex adaptive systems

Abstract

Root cause analysis (RCA) is a recognised approach to understanding causation of adverse events across high-risk industries including healthcare. These methodologies were developed during the early part of the twentieth century when the workplace could be understood as a series of linear processes. Within a complex system these approaches offer limited insight, which has since been recognised within healthcare literature. This paper proposes an approach to understanding of causation that addresses Hollnagel’s ‘hypothesis of different causes’ and integrates Safety I and Safety II approaches. This develops Stretton’s Lilypond Model to conceptualise the relationship between work-as-imagined and work-as-done within a complex system where individual adaptations and variations can be analysed. Understanding variation in such a way creates a shift in methodology from a deterministic to a probabilistic approach, which is more appropriate for understanding causation within complex systems.

Keywords

Root cause analysis incident reporting patient safety organisational learning risk management

Introduction

Root cause analysis (RCA) has become a ubiquitous part of patient safety management systems, particularly when responding to adverse events. The goal of RCA is “to find out what happened, why did it happen and what do you do to prevent it from happening again.”¹ This approach is not specific to healthcare and exists across many high-risk industries. The Health & Safety Executive define the root cause to be “an initiating event or failing from which all other causes or failings spring. Root causes are generally management, planning or organisational failings”.²

There are a number of accepted methods for RCA. NHS Improvement encourage an approach adopting The 5 Whys, where “by repeatedly asking the question ‘why?’ (use five as a rule of thumb), you can peel away the layers of a problem to get to the root cause.”³ The Five Whys approach was initially developed by Sakichi Toyoda at Toyota and has since been widely adopted into systems of engineering, quality control as well as safety processes.

RCA has been subject to criticism within the healthcare literature. Peerally et al. identify eight challenges for RCA.⁴ These include the questionable quality of some investigations, political hijack and confusion about blame. The authors conclude that “RCA is a promising incident investigation technique borrowed from other high-risk industries, but has failed to live up to its potential in healthcare… A key problem with RCA is its name, which implies a singular, linear cause.⁴” Furthermore, Wiig et all argue “healthcare investigation practice could therefore experiment more with variety in theoretical and methodological approaches to strengthen investigation quality.”⁵ This author agrees with the suggestion that “it’s time to step it up”⁵ within healthcare investigations, I would also argue that the limitations of linear understanding applied to complex systems is not an industry specific issue. Furthermore, RCA and the application of 5 Whys is not fit for purpose when understanding safety outcomes in complex adaptive systems and patient safety understanding will continue to be impaired whilst such approaches are advocated.

There have been other attempts to address the shortcomings of RCA. RCA² shifts to a systems based approach and accepts that there is seldom one root cause in a complex system.⁶ There is also a commitment to action to help organisations learn and improve from the analysis. These are important improvements to RCA, but the mechanism for understanding causation is still incomplete. This paper shall apply understanding of Complexity Science more comprehensively, as well as furthering the shift towards Safety II mindset, to create a new approach to understanding causation in complex adaptive systems such as healthcare.

Causation within safety II

In recent years, there has been an emerging shift in the understanding of workplace safety. Organisations are more aware of Safety I and Safety II approaches.⁷ The movement towards Safety II has created friction with traditional approaches of causation analysis. Hollnagel cites this as “a 'hypothesis of different causes', namely that the causes or 'mechanisms' of adverse events are different from those of events that succeed.⁷” This is a critical flaw of Safety I methodologies. In Safety I, only poor outcomes are investigated and analysed to identify future learning. But we now recognise that the system that creates adverse events also creates desirable outcomes. And in fact, for any system to survive, more of the outcomes must be successful than fail. Consequently, the approach to understanding causation must be applicable to all outcomes, not just a select few failings. Classical models such as The Swiss Cheese Model and Domino Theory are only applicable to describing causation of adverse events and do not provide the opportunity to consider both Safety I and Safety II thinking.^7,8 RCA identifies what initiated a failure, but fails to acknowledge that it is the same management, planning and organisational factors that can achieve overwhelming success.

There are new conceptual models for understanding organisational safety performance, such as Stretton’s Lilypond Model.⁹ The Lilypond provides a conceptual framework that integrates both Safety I and Safety II principles as well as recognising the behaviour and interactions of those working within complex systems. It also enables analysis of a complex system from both telescopic and microscopic viewpoints. All organisational outcomes are represented as lily pads on the surface of the pond and have a coloration according to the desirability of the outcome. Successful outcomes are represented by beautiful lily pads, and failings are black or ugly lily pads. The stem of each lily pad is a causational journey through time as the lily pad grows, irrespective of the outcome, and as such provides a more appropriate framework for developing an understanding of causation within complex systems (Figure 1).

The plan view of the Lilypond evolved from Heinrichs Safety Triangle¹⁰ to include Safety II perspectives of successful as well as adverse events. The model moves away from using a false dichotomy of desirable and undesirable outcomes to a non-monochrome view where learning from events without prejudice from their classification becomes a founding principle. The stem or causational journey applies to a process of a specific work activity. Within each work activity, such as the treatment of patients with suspected urinary tract infections (UTI’s), there would be a range of outcomes that continually change and would be demonstrated by constant fluctuation in the lily pad coloration.

The lily pad and stem represent the process of organisational output, the size and coloration of the pad symbolising the frequency and desirability of outcomes. Larger pads represent more numerous events. Brighter coloration represents increasingly more desirable outcomes. This view of the lilypond therefore is not static as there is no stasis in organisational performance. Instead the plan view of the Lilypond continually waxes and wanes with ever changing coloration and size of lily pads. This more accurately allows for both the order and disorder of complex adaptive systems to be recognised.

The stem in its entirety should be considered representative of how that organisational output occurs. Within this there will be many individual occurrences of this process being undertaken. We could consider it in the same manner as a stretch of motorway where a large number of drivers attempt the same journey, with ever changing lane changes in reaction to traffic flow. It is this causational journey that we can analyse within the Lilypond Model, but to do so important work regarding systems thinking also needs to be considered and developed.

Figure 1.

Causation within The Lilypond Model.

Understanding different types of human work

Hollnagel identifies the need to move away from Scientific Management Theory (SMT) and understand different types of work.¹¹¹², SMT was developed in the early twentieth century and demonstrated how a breakdown of tasks and activities into a linear process could be used to improve work efficiency. This became the basis for production lines and other complicated workplace systems. The synergy therefore with the understanding of causation by applying the Toyota based 5 Whys is clear. The processes of RCA and the 5 Whys were designed to work for a linear production environment. This is very different to the complex healthcare systems in which we work today, and therefore these investigative approaches are not appropriately transferrable. Utilising these linear-based investigative processes places an emphasis on what would be ideal, or work-as-imagined, “thus provided the theoretical and practical foundation for the notion that Work-As-Imagined was a necessary and sufficient basis for Work-As-Done.”¹¹

Classical models of causation are predicated upon the principle of work-as-imagined. This linear approach is a concept of what should happen based on presumptions of the working environment and processes. Behaviours and decisions made at the point of work are often different from what was envisaged, especially when the underlying presumptions that predicated the work-as-imagined turn out to be inaccurate or false. Work-as-done acknowledges that people innovate and amend practices frequently. These adaptations at the point of work, the variation, is integral to understanding causation.

This shift in analysing work activities is considerable and important. When considering the completion of a single work activity, work-as-done could be viewed as a linear process, which may or may not be the same as work-as-imagined. When repeated, the process of this individual work activity becomes increasingly varied, and consequently work-as-done becomes less linear when not taken as an isolated, single work activity.¹³ Shorrock also identifies work-as-prescribed and work-as-disclosed (Figure 2).¹⁴ Work-as-prescribed is the formalisation of work-as-imagined. Work-as-disclosed is what we say and how we write about work. Whilst these observations about types of work are valid and useful, the best learning can be achieved by focusing on the work that actually happens, work-as-done, as well as understanding the work that it was thought would happen, work-as-imagined.

Figure 2.

Different types of human work.

When work-as-imagined and work-as-done are treated as discrete or estranged processes they will remain so, even if at time they are perfectly aligned. This limits the opportunities to learn from outcomes as much as we ought to, as it is the relationship between these types of work that offers greatest insight into understanding causation. Suggestions to focus solely on work-as-done is misguided, as work-as-done does not occur without work-as-imagined.¹⁵ Work-as-imagined creates the frame for the work undertaken and establishes the expected outcome and process. Behaviours observed as work-as-done occur as a result of trying to make work-as-imagined a reality. These types of work do not just overlap at times. They are indelibly linked throughout and it is not possible to develop an understanding of “why did that decision or behaviour make sense at the time” without understanding how the relationship between these types of work created the environment in which the decisions were taken.¹⁶

In complex systems the outcome is unknown, which is why work-as-done offers valuable insights concerning patient safety outcomes. Feedback to the system can come in the form of prior knowledge; how the work-was-done previously or by colleagues. It can also be the notion of what should be achieved and how; work-as-imagined. This is why it is the relationship between work-as-imagined and work-as-done that is integral to fully understanding causation and these cannot be treated as separate entities during any investigative processes. Rather than the distinction between known and unknown futures, our analysis benefits from understanding the work-as-done present within the context of aiming for an idealised future. Often the work-as-done adaptations or variations occur as someone is trying to help direct the present towards the idealised future of work-as-imagined. It is only possible to fully understand the past events by recognising the role that the idealised future had in shaping the environment or decisions that were made.

The Lilypond Model can be used as an approach to causation analysis. This requires a microscopic view of the lily pad stems, rather than the macroscopic viewpoint presented previously. The stems of the lily pads consist not only of work-as-imagined, or just work-as-done, but both. Indeed, the additional labels of human work could also be considered. The cross section of the stem, at any point in time of its development, would display all of the circles of human work, running alongside each other and overlapping and separating to varying degrees throughout the length of the lily pad stem as shown in Figure 3.

Figure 3.

Work-as-imagined and work-as-done within the lilypad stems of the Lilypond Model.

Causation within complex systems

Having identified the conceptual flaws of using the RCA process in complex systems, we have established clear principles which any causational model should accommodate. These are counter to traditional approaches for understanding patient safety outcomes.

Outcome classification – any model and process used must be equally applicable to all organisational outcomes, and not only suitable for examining adverse events.

Weighting of systems and processes – both work-as-imagined and work-as-done must be considered within the analytical process, as well as taking the effort to understand the relationship between them.

The growth of the stem of the lily pad provides the concept of time as a process is undertaken, as well as being useful as an analogy for the different types of human work. We can then consider the varying alignment of these different types of work with each other, throughout the length of the stem of the lily pad. It is possible for a fully automated process, such as car manufacture, to be repeated many times identically. It is not possible for humans to perform to such consistency, especially in complex systems. The same individual repeating complex tasks he is highly skilled at, is unlikely to repeat the process exactly. The addition of multiple agents of a complex system will make this element of variation certain, as a result of crowd effects creating emergent phenomena.¹⁷ One person’s decision, action or execution will impact on another person’s subsequent decisions, actions or executions.¹⁷

The variation between work-as-done and work-as-imagined is a result of the interactions, including micro-interactions, identified within a complex system.¹⁸ Many of these demonstrate a deviation from work-as-imagined but will not necessarily have an impact on the overall outcome. This would be considered non-causal variation. For example, during a consultation, a GP may decide to prescribe a first line formulary antibiotic for treatment of classically described cystitis without having proven objective evidence of urinary tract infection. Work-as-imagined might expect that a mid-stream urine sample would be collected, dipped, and if appropriate, sent for microscopy and culture. In the work-as-done world this may not happen for numerous and variable reasons. In this worked example, the patient has bacteriuria, sensitive to the antibiotic prescribed. The patient experiences an acceptable outcome and the variation from work-as-imagined has no significant impact. It is non-causal.

There will be some variation from work-as-imagined that significantly changes the outcome. This is causal variation. In the example above, the failure to objectively investigate the cause of the patient’s symptoms might mean that the opportunity for early identification of a multi-resistant organism causing urinary tract infection is missed. The patient is prescribed an antibiotic to which the bacteria is resistant and subsequently becomes more unwell with pyelonephritis, requiring admission to secondary care for inpatient treatment. In this example, the variation from work-as-imagined impacts on the outcome.

In both of these scenarios, feedback will impact on future behaviours. In the first instance, the GP may decide in future similar situations, it may not be necessary to obtain an MSU as their experience has demonstrated that this management can led to an acceptable outcome, despite having deviated from work-as-imagined. Conversely, in the second scenario, this GP may or may not be aware of their patient having deteriorated. In a system that had effective multi-channel communication systems the feedback in the system would enable the GP would be able to reflect and learn from their decisions and in future adopt practice more closely aligned with what work-as-imagined would expect. If the system did not effectively use this source of feedback, however, the GP would be unaware of what happened to the patient and not be provided with the opportunity to change their future practice.

Both these examples would apportion some liability applying Safety I methodology, such as RCA, to the individuals due to their deviation from known rules or procedures. This binary approach evidently doesn’t work in complex systems. The separation between the two parts of the stem is not necessarily an indication of poor professional behaviour. People often innovate or adapt from the prescribed system in order to effectively achieve the ideal outcome. A violation or deviation from work-as-imagined is not therefore inherently bad or undesirable. The learning and improvement within the system is more likely to be gained by conscious effort applied to understanding why the separation occurs, what the outcome of that deviation is and why the environment made the decision acceptable in the context within which it was made. This investigative process can be best achieved by developing a truly Just Culture.¹⁹

For example, in the event that a patient didn’t attend an appointment, the system would expect the GP to move onto their next appointment and for the patient to make a new appointment. The GP though has a longstanding rapport with the patient that doesn’t attend and is very familiar with both the health and behavioural history with appointments seldom missed. Concerned, the GP calls the patient who fails to answer the phone. A home visit is arranged, and police contacted whereby the patient if found lying on the floor with a hip fracture. The decision to not strictly follow the systemic expectation enabled the GP to save the patient’s life. Having recognised this successful adaptation and significant variation between work-as-imagined and work-as-done, the question for system thinkers is to consider how work-as-imagined could shift to reduce such variation and make the system safer for future events.

Developing causal and non-causal variation

By embracing the two principles identified earlier, a more effective approach to understanding causation is possible. This can be done by recognising the importance of variation within the lily pad stems of Stretton’s Lilypond Model.

There could be a stem within the Lilypond that represents the treatment of patients with suspected UTIs. The work-as-imagined part of the stem would be consistent and rigid, representing the established working protocol. The work-as-done part of the stem would unpredictably and continually diverge and align with the work-as-imagined part of the stem, and this movement would differ each time the work is undertaken. The cross section of the stem would have therefore a circular work-as-imagined and a circular work-as-done strand, which continually overlap and separate throughout the length of the stem. At times there will be more overlap or convergence between the types of work, at others the variation would be signified by a greater distance between them (Figure 4). The stem therefore consistently shifts in nature and shape throughout time. Consequently, any causal analysis process that tries to offer definite and fixed outcomes should be treated with caution. An approach that offers analysis based on a probabilistic rather than deterministic nature would be more appropriate within complex systems. Rather than find a “singular, linear cause,” we move to ask ‘of all the variations that have occurred, which were the ones that have most likely and most significantly contributed to this outcome?’⁴

Figure 4.

Work-as-imagined and work-as-done within The Lilypond Model.

Applying this concept remains subjective in nature, in the same way in which SCM and other models of causation are. The difference from previous processes of causation analysis is that the focus on what is being analysed and in what context it has occurred, is a complete shift in safety thinking. It may be possible however to more from a purely conceptual framework. To this end the work of David Smith is of particular interest. Smith has adopted a sophisticated mathematical modelling analysis using online complex adaptive systems with multiple agents.²⁰ His modelling has enabled to help predict behaviour within complex systems, creating corridors (Future-Casts) along which the system is likely to move. The width of these corridors is called Characteristic Stochasticity, and their average direction called Characteristic Direction. This is an approach that is non-deterministic and instead indicates the probability of future order or disorder occurring within a complex system, “without knowing what individual objects are each doing – he can produce such corridors into the future… with remarkable accuracy.”¹⁷ Smith’s work begins to bridge the conceptual framework with an empirical one, providing a potential methodology for identifying which variations are causal and most consequential for organisational learning.

The Lilypond Model conceptualises an organisations safety performance in line with Safety II principles. The macroscopic view offers many opportunities to understand, learn and improve. The microscopic view developed within this paper extends the application of the model. This new view of causation analysis allows the true nature of behaviours within complex adaptive systems to be understood. Importantly, the extension of the model does not jeopardise its accessibility but does increase its applicability as well as consequential patient safety learning.

Applying the principle of causal and non-causal variation analysis within the workplace should not require any greater degree of technical understanding than RCA or RCA². It is the understanding of behaviours within complex adaptive systems and the consequential shift in focus that will create substantial change in understanding of causation. An investigation applying the concept of Causal Variation Analysis (CVA) would follow a five-step process to achieve full understanding of the nature of the causation stem as outlined in Figure 5. The process maintains the important commitment to action that RCA² introduces.

Figure 5.

Causal Variation Analysis in practice.

Conclusion

The classical models of accident causation do not provide the necessary framework for effective causation analysis within complex systems. The Lilypond Model integrates Safety II principles, allowing scope to consider complexity in a system. The multi-stranded probabilistic lily pad stem extends this framework further. It incorporates both work-as-imagined and work-as-done, plus the other types of work if so wished, as well as a framework for understanding the relationship between these ideas. It is in these relationships that the most insightful lessons can be learned. Applying this model will enable organisations to more effectively understand the causes of their organisational outcomes in a manner in which previous models do not accommodate and identify the adaptations that most significantly impact on future organisational performance.

Footnotes

Author contributions and guarantor

All work (both conceptual in developing concept, and in writing this manuscript) has been completed by the primary and corresponding author, Paul Stretton. As guarantor, I accept full responsibility for the work, and have controlled the decision to submit for publication.

Acknowledgements

Many thanks to Dr Catherine Stretton for her comments on the initial concept and feedback on multiple drafts of this manuscript.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Paul Stretton

References

Uberoi

Swati

Gupta

, et al. Root cause analysis in healthcare. Apollo Med 2007; 4: 72–75.

Health and Safety Executive. Root cause analysis; literature review, www.hse.gov.uk/research/crr_pdf/2001/crr01325.pdf (2001, accessed 3 May 2020).

ACT Academy for NHS Improvement. Quality, service improvement and redesign tools: root cause analysis using five whys, https://improvement.nhs.uk/documents/2156/root-cause-analysis-five-whys.pdf (2018, accessed 20 February 2020).

Peerally

Carr

Waring

, et al. The problem with root cause analysis. BMJ Qual Saf 2017; 26: 417–422.

Wiig

Braithwaite

Clay-Williams

It’s time to step it up. Why safety investigations in healthcare should look more to safety science. Int J Qual Health Care 2020; 32: 281–284.

National Patient Safety Foundation. RCA² improving root cause analyses to actions to prevent harm, www.ashp.org/-/media/assets/policy-guidelines/docs/endorsed-documents/endorsed-documents-improving-root-cause-analyses-actions-prevent-harm.ashx (2015, accessed 19 November 2020).

Hollnagel

Safety-I and Safety-II. The past and future of safety management. Boca Raton: CRC Press, 2014.

Hollnagel

Resilience engineering: a new understanding of safety. J Ergonom Soc Korea 2016; 35: 185–191.

Stretton

An integrated model of safety II principles in the workplace. A quantum shift in patient safety thinking. J Patient Safety Risk Manag 2020; 25: 85–90.

10.

Heinrich

HW.

Industrial accident prevention: a scientific approach. New York: McGraw-Hill, 1931.

11.

Hollnagel

Wears

Braithwaite

From safety I to safety II: a white paper. The resilient health care net: Published simultaneously by the University of Southern Denmark, University of Florida, USA, and Macquarie University, Australia, 2015.

12.

Taylor

FW.

The principles of scientific management. New York: Harper, 1911.

13.

Seidling

Lampert

Lohmann

, et al. Safeguarding the process of drug administration with an emphasis on electronic support tools. Br J Clin Pharmacol 2013; 76: 25–36.

14.

Shorrock

The varieties of human work, https://humanisticsystems.com/2016/12/05/the-varieties-of-human-work/ (2016, accessed 3 March 2020).

15.

Chartered Institute of Ergonomics and Human Factors. Achieving sustainable change: capturing lessons from COVID-19, https://ergonomics.org.uk/Common/Uploaded%20files/Publications/Sustainable-Change/CIEHF-Achieving-Sustainable-Change.pdf (2020, accessed 3 March 2020).

16.

Horsley

Safety II: implications for staff wellbeing, www.cicm.org.au/CICM_Media/. CICMSite/CICM-Website/Resources/Welfare%20SIG/Safety-II-Implications-for-Staff-Wellbeing- FINAL_1.pdf (2019, accessed 23 November 2020).

17.

Johnson

Simply complexity: a clear guide to complexity theory. London: Oneworld Publications, 2007.

18.

Woodward

Moving towards a safety II approach. Journal of Patient Safety Risk Management 2019; 24: 96–99.

19.

Stretton

Just culture: why it isn’t just and how it could be. British Journal of Healthcare Management 2020; 26: 304–310.

20.

Smith

DMD

Johnson

NF.

Predictability, risk and online management in a complex system of adaptive agents, https://core.ac.uk/download/pdf/44113917.pdf (2006, accessed 1 March 2020).