Reverse propagation and diagnostic inference - reasoning backward from observed outcomes to uncover the psychosocial hazards most likely responsible.
It is a Tuesday morning, and the Director of People and Culture at a metropolitan hospital sits across from three documents. The first is a workers' compensation summary: three psychological injury claims from Ward 7 nursing staff in the past quarter, up from zero in the same period last year. The second is the most recent engagement survey, which shows Ward 7's "feeling supported by my manager" score has dropped from the 70th to the 28th percentile. The third is the ward's absenteeism dashboard, blinking red at 14.2% - nearly double the hospital average.
The symptoms are unmistakable. But the causes are not. Is this a workload problem - perhaps Ward 7 has absorbed patients from a recently closed unit? Is it a leadership problem - a new charge nurse who lacks the skills to buffer staff from organisational pressure? Is it an autonomy problem - a procedural change that stripped experienced nurses of clinical decision-making? Or is it some combination of all three, interacting in ways that no single investigation question can untangle? The director needs more than intuition. She needs a principled method for reasoning backward from observed outcomes to the upstream hazards most likely responsible. That method is diagnostic inference in a Bayesian network.
In Chapter 4, we learned to reason forward through a Bayesian network. We set evidence at root nodes - psychosocial hazards like excessive workload or poor supervisor support - and propagated beliefs downstream to predict the probability of outcomes like stress, burnout, and absenteeism. That forward direction, which Pearl (1988) calls predictive inference, answers the question: if these hazards are present, what outcomes should we expect?
Predictive inference is invaluable for proactive risk assessment, but it is not the reasoning direction that organisations most urgently need when problems have already materialised. When the workers' compensation claims have already been filed, when the engagement scores have already plummeted, when absenteeism has already surged, the question reverses: given these observed outcomes, which upstream hazards are most probably responsible? This is diagnostic inference - also known as reverse propagation or abductive inference - and it is arguably the more practically powerful direction of reasoning for psychosocial risk management (Korb & Nicholson, 2010).
The distinction matters because the two directions of reasoning produce fundamentally different kinds of knowledge. Forward reasoning tells you what to watch for; backward reasoning tells you where to intervene. Forward reasoning supports prevention planning; backward reasoning supports incident investigation and root cause analysis. An organisation that can only reason forward is like a physician who can list the symptoms of every disease but cannot, upon examining a patient, work backward to the most probable diagnosis. As Gámez et al. (2004) observe, abductive inference - finding the best explanation for observed evidence - is the form of reasoning most central to diagnostic problem-solving across disciplines.
Consider the Ward 7 scenario from the opening. If you could only ask three questions in your investigation, what would they be? What implicit assumptions about cause-and-effect are embedded in your choices? Hold your answers in mind - we will return to them once you have the tools of diagnostic inference.
The mathematics of diagnostic inference rest on a single foundational principle: Bayes' theorem. In its simplest form, Bayes' theorem states that the probability of a cause given an observed effect is proportional to the probability of the effect given the cause, multiplied by the prior probability of the cause:
P(Cause | Effect) = P(Effect | Cause) × P(Cause) / P(Effect)
This equation encodes a profound investigative logic. The term P(Cause) represents our prior belief about how likely a particular hazard is before we observe any outcomes - perhaps informed by industry base rates, organisational history, or expert knowledge. The term P(Effect | Cause) is the likelihood - how probable the observed outcome would be if the hazard were indeed present. And P(Cause | Effect) is the posterior - our updated belief about the hazard after taking the observed outcome into account (Pearl, 1988).
In a Bayesian network with multiple interconnected nodes, this single-node calculation extends to a full network-wide belief update. When we enter evidence at an outcome node - for example, setting Absenteeism = High - the network uses the conditional probability tables at every node to propagate that evidence backward through the directed edges. Every upstream node receives an updated posterior probability that reflects the new evidence. The result is a complete diagnostic picture: a revised probability distribution over every potential cause, conditioned on the observed symptoms.
Consider our hospital Bayesian network from Chapter 4. The root nodes represent psychosocial hazards: Workload (High or Normal), Supervisor Support (Poor or Good), and Job Control (Low or Adequate). These feed into intermediate nodes - Stress (High or Low), which combines with Job Control to influence Burnout (Present or Absent) - and ultimately into the observable outcome, Absenteeism (High or Normal).
Before any evidence is entered, each node sits at its prior probability. Perhaps Workload = High has a prior of 0.35, reflecting the base rate across hospital wards. Now suppose we observe that Absenteeism = High. This single piece of evidence triggers a cascade of backward updates. The probability of Burnout = Present increases (because burnout is a strong predictor of absenteeism). The probability of Stress = High increases (because stress feeds burnout). And crucially, the probabilities of the root hazards - Workload = High, Supervisor Support = Poor, and Job Control = Low - all shift upward, because each is a potential upstream explanation for the observed absenteeism.
But the updates are not uniform. The hazards that are connected to absenteeism through stronger conditional probability pathways receive larger probability increases. If the conditional probability tables encode, for instance, that poor supervisor support has a stronger effect on stress than workload does, then observing high absenteeism will increase the posterior probability of poor support more than it increases the posterior of high workload. The network quantifies the relative explanatory power of each upstream hazard - something that informal root cause analysis rarely achieves (García-Herrero et al., 2012).
The real power of diagnostic inference emerges when evidence is entered sequentially. Returning to our scenario: the director first observes Absenteeism = High, and the network updates. She then learns from exit interviews that nurses report feeling highly stressed - she enters Stress = High, and the network updates again. Each new piece of evidence refines the diagnostic picture, narrowing the field of probable causes.
This sequential updating mirrors the structure of real workplace investigations. An investigator does not learn everything at once. First there is the initial report, then the survey data, then the interviews, then the document review. At each stage, the Bayesian network integrates the new information with everything previously known, producing an ever-sharper picture of the most probable causal configuration. As Sun et al. (2023) demonstrated in their Bayesian network study of psychosocial hazards among construction practitioners, this iterative diagnostic process can identify causal pathways that neither raw data inspection nor traditional statistical models would reveal.
Perhaps the most intellectually striking - and practically important - phenomenon in diagnostic inference is the explaining-away effect. First formalised by Pearl (1988) and given its definitive treatment by Wellman and Henrion (1993), explaining away occurs when two or more causes converge on a common effect, creating what is known as a collider (or v-structure) in the network.
The logic is intuitive once grasped, even though the mathematics can be surprising. Consider two potential causes of high stress in a nurse: Excessive Workload and Poor Supervisor Support. In the network, both parent nodes have directed edges pointing to the child node, Stress. Before any evidence is observed, the two parent hazards are marginally independent - knowing the workload level tells you nothing about the quality of supervisor support, because there is no direct connection between them.
Now suppose we observe that Stress = High. By Bayes' theorem, both parent hazards become more probable: high stress needs an explanation, and both workload and support are potential explanations. But here is where the phenomenon becomes powerful. Suppose we subsequently learn that Workload is actually Normal. The high stress still needs an explanation, but one of the two candidate causes has just been eliminated. The probability mass that was allocated to excessive workload as an explanation must now flow to the remaining cause. The posterior probability of Poor Supervisor Support jumps sharply upward - often far above where it was after observing stress alone.
This is explaining away: confirming one cause of an observed effect reduces the posterior probability of alternative causes, and conversely, disconfirming one cause increases the posterior of alternatives (Wellman & Henrion, 1993). The two causes are not marginally dependent - they become conditionally dependent given the common effect. Observing the child node "activates" the collider, creating an informational pathway between the parents that did not exist before.
In everyday workplace investigations, have you seen situations where ruling out one potential cause immediately made investigators more confident in an alternative? The explaining-away effect formalises this intuition - but also reveals situations where our intuitions fail. Can you think of a scenario where two causes might reinforce rather than explain away each other?
The explaining-away effect is not merely a mathematical curiosity - it captures the precise reasoning pattern that workplace psychosocial risk assessors must use when responding to incident reports and survey findings. When an investigator learns that a worker is experiencing high stress, she must consider multiple potential hazards. As she gathers evidence confirming or disconfirming each hazard, her beliefs about the remaining hazards should update accordingly. If she learns that workload is reasonable, her attention should shift more strongly toward supervision quality, role clarity, or other unmeasured factors.
Wellman and Henrion (1993) demonstrate that while explaining away is the default pattern at collider structures, the opposite - positive synergy, where confirming one cause increases the probability of another - can also occur under specific conditional probability configurations. This means investigators cannot simply assume that ruling out one cause always makes others more likely. The direction and magnitude of the update depend on the specific quantitative relationships encoded in the network, reinforcing why formal Bayesian reasoning is superior to informal intuition for complex multi-causal scenarios.
García-Herrero et al. (2012) provide empirical evidence of this phenomenon in their study of Spanish workers. Using a Bayesian network fitted to over 11,000 survey responses, they demonstrated that the diagnostic probability of poor hygiene conditions given an observed psychological disorder changed substantially depending on whether ergonomic conditions were simultaneously observed to be adequate or poor - a direct manifestation of inter-causal reasoning in an occupational health context.
Diagnostic inference gives us updated posterior probabilities for each individual node. But often, an investigator needs something more integrated: the single configuration of all unobserved variables that best accounts for the observed evidence. This is the Most Probable Explanation (MPE) problem - finding the joint assignment to all hidden variables that maximises the posterior probability given the evidence (Gámez et al., 2004).
The distinction between individual posteriors and the MPE is subtle but important. Individual posteriors tell you that Workload = High has, say, a 62% posterior probability, and that Supervisor Support = Poor has a 71% posterior probability. But these are marginal probabilities - each computed independently. The MPE, by contrast, identifies the single joint configuration that is most probable. It might tell you that the most likely explanation for the observed pattern of high absenteeism and high stress is specifically the combination of Normal Workload, Poor Supervisor Support, and Low Job Control - even if, marginally, high workload has a reasonably high individual posterior.
This happens because the MPE accounts for dependencies between variables that marginal posteriors ignore. As Kwisthout (2011) demonstrates, finding the MPE is computationally NP-complete in general, but becomes tractable for networks with low treewidth - which, fortunately, describes many of the Bayesian networks used in occupational health applications. The practical implication is that for the kinds of moderately sized networks typically constructed for psychosocial risk assessment, exact MPE computation is entirely feasible.
The value of the MPE for workplace investigators lies in its ability to generate targeted investigation priorities. Rather than investigating every possible hazard with equal effort, the investigator can use the MPE to identify the specific combination of upstream conditions that most probably produced the observed outcomes, and allocate investigation resources accordingly.
Solé et al. (2017), in their comprehensive survey of root cause analysis techniques, position Bayesian MPE inference as the leading probabilistic approach to fault diagnosis in complex systems. They contrast it with traditional deterministic root cause analysis methods - such as the "5 Whys" or fishbone diagrams - which lack a principled mechanism for quantifying the relative probability of different causal hypotheses. A Bayesian network with MPE inference does not merely list possible causes; it ranks them by posterior probability, providing a quantitative basis for prioritising investigation actions.
This quantitative ranking becomes especially valuable when investigation resources are constrained - as they almost always are. If the MPE indicates that the combination of Poor Supervisor Support and Low Job Control explains 73% of the posterior probability mass, while the combination involving High Workload explains only 12%, the investigator knows to prioritise interviews about management practices and decision-making authority over workload audits. The investigation becomes targeted rather than exhaustive.
Traditional root cause analysis often settles on a single "root cause." How does the MPE framework differ from this practice? What are the advantages of identifying a joint configuration of multiple upstream variables rather than a single root cause?
The diagnostic reasoning capabilities described in this chapter are not merely academic - they map directly onto regulatory requirements that Australian organisations must meet. Safe Work Australia's (SWA) model Code of Practice for managing psychosocial hazards at work requires employers to identify psychosocial hazards, assess the associated risks, and implement control measures. Critically, it also requires ongoing monitoring and review, including the investigation of incidents and complaints related to psychosocial hazards.
ISO 45003:2021, the international standard for managing psychosocial risks at work, goes further. Section 8.2 explicitly addresses the need for organisations to investigate psychosocial risk-related incidents to identify contributing factors and underlying causes. The standard calls for a systematic approach to investigation that considers "the range of contributing factors" - not just proximate triggers but the upstream organisational conditions that created the environment for harm.
Bayesian diagnostic inference formalises and improves upon the investigative processes these standards envision. Where traditional investigation relies on linear causal narratives ("the worker was stressed because the workload was high"), a Bayesian network captures the multivariate reality: multiple hazards contributing simultaneously through overlapping causal pathways, with the relative contribution of each quantified by posterior probabilities. Sun et al. (2023) explicitly frame their Bayesian network model of psychosocial hazards in construction as a tool for precisely this kind of multi-causal diagnostic investigation, demonstrating how backward inference through the network can identify which combinations of hazards (poor physical environment, contract pressure, lack of support) most probably underlie observed mental health outcomes.
The diagnostic picture produced by reverse propagation does more than satisfy regulatory requirements - it guides intervention design. If the posterior probabilities and MPE point strongly toward poor supervisor support as the primary explanation for observed outcomes, the organisation knows to invest in leadership development, supervisory training, and support structures rather than (or in addition to) workload redistribution. If the MPE identifies low job control as a critical co-factor, intervention should include restructuring decision-making authority.
This evidence-based approach to intervention targeting represents a significant advance over the common organisational practice of implementing generic wellbeing programs (yoga classes, resilience training, employee assistance programs) that address symptoms rather than causes. Bayesian diagnostic inference forces the organisation to confront the specific upstream conditions that produced the observed harm, making it much harder to deflect responsibility onto individual worker resilience. The network's posterior probabilities constitute a quantitative argument for where organisational change is most needed - and where it will have the greatest impact.
Powerful as diagnostic inference is, it comes with important limitations that the responsible practitioner must understand. First, the quality of diagnostic reasoning depends entirely on the quality of the Bayesian network itself - its structure and its conditional probability tables. If the network omits an important causal pathway (for example, if bullying is a significant cause of stress on Ward 7 but is not represented in the network), reverse propagation will distribute the explanatory probability mass among the causes that are represented, potentially inflating their posterior probabilities and leading the investigator astray.
Second, the MPE identifies the most probable explanation, not the definitive truth. A configuration with 40% posterior probability is "most probable" relative to alternatives, but there is still a 60% chance that a different configuration is correct. Investigators should consider not just the MPE but the second and third most probable explanations, and design investigations that can discriminate among them. Kwisthout (2011) notes that the gap between the first and second most probable explanations is itself informative - a large gap suggests high diagnostic confidence, while a small gap suggests that further evidence is needed before committing to a causal narrative.
Third, diagnostic inference assumes that the evidence entered is accurate. In psychosocial risk assessment, "evidence" often comes from self-report surveys, interviews, and manager observations - all of which are subject to reporting bias, social desirability effects, and measurement error. A network that receives inaccurate evidence will produce inaccurate posteriors. Sensitivity analysis - examining how much the posteriors change under different evidence assumptions - is an essential complement to diagnostic inference in practice.
Finally, there is the computational consideration. While exact inference is tractable for the relatively small networks typical of psychosocial risk models, larger and more complex networks may require approximate inference methods. Gámez et al. (2004) review both exact and approximate algorithms for abductive inference, and Kwisthout (2011) provides complexity results that help practitioners understand when exact methods are feasible and when approximations are necessary.
Let us return to where we began. The Director of People and Culture has three pieces of evidence: high absenteeism, declining support scores, and a cluster of psychological injury claims. Rather than launching an unfocused investigation or - worse - defaulting to a generic resilience program, she enters this evidence into the hospital's psychosocial risk Bayesian network.
The network propagates backward. The posterior probability of Poor Supervisor Support rises to 0.78 - the highest of any upstream hazard. High Workload rises to 0.47, above its prior but not dramatically so. Low Job Control rises to 0.61. The MPE identifies the most probable configuration: Normal Workload, Poor Supervisor Support, and Low Job Control - suggesting that the primary issue is not that nurses are doing too much, but that they are unsupported and disempowered while doing it.
Armed with this diagnostic picture, the director designs a targeted investigation. She conducts structured interviews focused on supervisory behaviours and decision-making authority. She reviews whether recent procedural changes have removed clinical judgement from experienced nurses. She examines the ward's management structure for gaps in support. Each new piece of evidence is entered into the network, further refining the posteriors and sharpening the MPE.
The result is an investigation that is efficient, evidence-based, and aligned with the requirements of ISO 45003. It targets the specific upstream conditions most probably responsible for the observed harm, and it generates specific, actionable recommendations for organisational change. This is the power of diagnostic inference: not just knowing that something went wrong, but understanding - with quantified confidence - why.
We have now mastered both forward and backward inference in Bayesian networks. But where do the numbers come from? In Chapter 6, we turn to the critical question of parameterisation - how to populate a Bayesian network's conditional probability tables using a combination of empirical data, expert elicitation, and published research. We will learn structured expert elicitation protocols suitable for psychosocial risk domains where data is sparse, and confront the challenge of translating qualitative workplace knowledge into the quantitative probabilities that make inference possible.
Gámez, J. A., Mateo, J. L., & Puerta, J. M. (2004). Abductive inference in Bayesian networks: A review. In Advances in Bayesian networks (pp. 101–120). Springer. https://doi.org/10.1007/978-3-540-39879-0_6
García-Herrero, S., Mariscal, M. A., García-Rodríguez, J., & Ritzel, D. O. (2012). Working conditions, psychological/physical symptoms and occupational accidents: Bayesian network models. Safety Science, 50(9), 1760–1774. https://doi.org/10.1016/j.ssci.2012.04.005
Korb, K. B., & Nicholson, A. E. (2010). Bayesian artificial intelligence (2nd ed.). CRC Press. https://doi.org/10.1201/b10391
Kwisthout, J. (2011). Most probable explanations in Bayesian networks: Complexity and tractability. International Journal of Approximate Reasoning, 52(9), 1452–1469. https://doi.org/10.1016/j.ijar.2011.08.003
Pearl, J. (1988). Probabilistic reasoning in intelligent systems: Networks of plausible inference. Morgan Kaufmann. https://www.sciencedirect.com/book/9780080514895/probabilistic-reasoning-in-intelligent-systems
Solé, M., Muntés-Mulero, V., Rana, A. I., & Estrada, G. (2017). Survey on models and techniques for root-cause analysis. arXiv preprint arXiv:1701.08546. https://arxiv.org/abs/1701.08546
Sun, C., Hon, C. K. H., Way, K. A., Jimmieson, N. L., Xia, B., & Wu, P. P.-Y. (2023). A Bayesian network model for the impacts of psychosocial hazards on the mental health of site-based construction practitioners. Journal of Construction Engineering and Management, 149(3). https://doi.org/10.1061/JCEMD4.COENG-12905
Wellman, M. P., & Henrion, M. (1993). Explaining "explaining away." IEEE Transactions on Pattern Analysis and Machine Intelligence, 15(3), 287–292. https://doi.org/10.1109/34.204911