Formalizing intuition - from hazard sketches to Bayesian Networks that reason under uncertainty
Return to the hospital ward you explored in Chapter 1. You drew lines on paper - from chronic understaffing to workload, from workload to burnout, from bullying to psychological distress, and from distress to medication errors. Those lines captured something real: your professional judgement about how psychosocial hazards propagate through a workplace. But now imagine you must advise the hospital's executive team on where to invest a limited wellbeing budget. "Increase supervisor support" costs $200,000 per year across the nursing workforce. "Reduce patient-to-nurse ratios" costs $1.2 million. Both interventions plausibly reduce burnout - but by how much? And does the answer change if bullying is already being addressed through a separate program?
Your hand-drawn web cannot answer these questions. It shows you that connections exist, but not how strong they are, not how they interact, and not how evidence at one node should change your beliefs about another. To move from intuition to intelligence - from "these things are connected" to "here is what we should expect if we intervene here" - you need a formal language. That language is the Bayesian Network.
In Chapter 1, you built hazard webs - informal diagrams linking psychosocial risk factors to mediating processes and health outcomes. In Chapter 2, you articulated what you wished a model could do: handle uncertainty, represent interactions, update with evidence, and support what-if reasoning about interventions. A Bayesian Network (BN) does all of these things, and it begins with a structure you already know how to draw: a directed acyclic graph, or DAG.
A DAG consists of two elements: nodes (representing variables) and directed arrows (representing direct probabilistic influences between them). The word "directed" means each arrow has a source and a target - it points from a cause toward an effect, or more precisely, from a variable that directly influences another. The word "acyclic" means you can never follow the arrows in a circle back to where you started. Time, in a sense, only flows forward (Pearl, 1988).
Consider the three-variable pathway you explored in the hospital scenario: Workload → Burnout → Medication Errors. This tiny DAG makes three claims simultaneously. First, it asserts that workload has a direct probabilistic influence on burnout. Second, it asserts that burnout has a direct probabilistic influence on medication errors. Third - and this is the claim that is easiest to overlook - it asserts that workload does not directly influence medication errors once we account for burnout. Every missing arrow in a DAG is a substantive independence claim (Pearl, 2009). The absence of an arrow from Workload directly to Medication Errors says: "Whatever effect workload has on errors, it operates entirely through the burnout pathway." That might be right or wrong - but the DAG forces you to commit to a position.
In a psychosocial risk BN, each node represents a variable that can take on a finite set of states. The variable Workload might be categorised as High, Medium, or Low. Supervisor Support might be Present or Absent, or graded on a three-level scale. Burnout might be measured as a discretized score from a validated instrument like the Maslach Burnout Inventory. The choice of states is a modelling decision - it determines the granularity of your analysis and the size of the probability tables you will need to populate.
Each arrow encodes a direct probabilistic dependency. In BN terminology, the node at the tail of an arrow is called a parent, and the node at the head is a child. A node with no parents is called a root node - it represents an exogenous factor whose probability distribution is specified unconditionally. In the hospital BN, root nodes might include organisational factors like staffing policy, shift scheduling, or management culture: variables that influence other factors in the model but are not themselves explained by it.
Look at the hazard web you drew in Chapter 1. Pick any two nodes that are not directly connected by an arrow. What independence claim are you implicitly making? Is that claim defensible? Under what conditions might it break down?
The "acyclic" constraint is the most common stumbling block for students formalising psychosocial risk models. In real workplaces, feedback loops are everywhere: burnout leads to reduced performance, which increases workload, which deepens burnout. If we draw arrows in both directions - Burnout → Workload and Workload → Burnout - we have a cycle, and the graph is no longer a DAG.
The resolution is temporal unrolling. Rather than representing "Workload" as a single timeless variable, we distinguish Workload at time t from Workload at time t+1. Burnoutt can influence Workloadt+1, and Workloadt+1 can influence Burnoutt+2, without creating a cycle. This is precisely how dynamic Bayesian Networks handle feedback - a topic we will explore in a later chapter. For now, the key insight is that removing cycles is not a limitation of the formalism; it is a discipline that forces you to be explicit about temporal ordering (Koller & Friedman, 2009).
The widget below lets you practice this transformation. You will take the kind of informal hazard web you built in Chapter 1 and walk through the steps of formalising it into a valid DAG.
A DAG tells you what influences what - the qualitative skeleton of your model. But the skeleton alone cannot answer the hospital executive's budget question. For that, you need numbers. In a Bayesian Network, each node is associated with a conditional probability table (CPT) that specifies the probability of each state of that variable given every possible combination of states of its parents (Pearl, 1988).
Let us build a CPT from scratch using our hospital example. Consider the node Perceived Stress, which has two parents: Supervisor Support (High or Low) and Job Demands (High or Low). Perceived Stress itself can be High, Medium, or Low. The CPT must specify the probability distribution over {High, Medium, Low} stress for each of the four parent configurations:
Each row in this table is a complete probability distribution - the three values must sum to 1.0. The row for (Low Support, High Demands) might look like P(High) = 0.70, P(Medium) = 0.20, P(Low) = 0.10, reflecting our belief that this combination is strongly associated with elevated stress. The row for (High Support, Low Demands) might be P(High) = 0.05, P(Medium) = 0.25, P(Low) = 0.70 - a much more favorable profile.
Notice what the CPT captures that a simple correlation cannot: it specifies how the combination of parent states matters. High demands with high support produces a different stress profile than high demands with low support. This is precisely the interaction effect that García-Herrero et al. (2013) found in their BN analysis of 11,054 Spanish workers: social support did not merely add a constant benefit but modulated the impact of work demands on stress in a configuration-specific way.
One of the most powerful properties of Bayesian Networks is the chain rule factorization. The full joint probability distribution over all variables in the network - which could be astronomically large if specified directly - decomposes into a product of the local CPTs. For a network with variables X1, X2, ..., Xn, the joint probability is:
P(X1, X2, ..., Xn) = ∏ P(Xi | Parents(Xi))
This means that if you specify a CPT for each node - conditioning only on that node's direct parents - the mathematics guarantees a coherent probability distribution over the entire network. A compact set of local probability specifications generates complex global behaviour (Koller & Friedman, 2009). This is why BNs scale to networks with dozens or even hundreds of variables: you never need to directly enumerate all possible combinations of all variables.
For root nodes - variables with no parents - the CPT reduces to an unconditional prior probability distribution. The root node Job Demands might have priors P(High) = 0.40, P(Low) = 0.60, reflecting the prevalence of high demands in the population of interest. These priors can be set from survey data, organisational records, or expert judgement.
Consider a node with 3 parents, each having 3 states, and the child node also has 3 states. How many probability values must you specify in its CPT? (Hint: count the parent configurations, then multiply by the child states minus one, since the last probability is determined by the sum-to-one constraint.) What does this tell you about the practical limits of BN complexity?
The answer - 54 free parameters for a single node - reveals why parsimony in DAG structure matters. Every arrow you add increases the size of the child's CPT exponentially. This is not just a computational concern; it is an epistemological one. Each number in the CPT must come from somewhere - data, literature, or expert judgement. García-Herrero et al. (2012) populated their occupational safety BN using national survey data from over 11,000 respondents, but even with large samples, sparse parent configurations can yield unreliable estimates. The structure of your DAG determines how much information you need to quantify it.
Use the widget below to experience this firsthand. You will set CPT values for a three-node stress network, see how the numbers interact, and watch how changing parent priors ripples through to the child's marginal distribution.
We have said that every missing arrow in a DAG encodes an independence claim. But the DAG actually encodes far richer independence information than just pairwise missing arrows. Through a concept called d-separation, you can read off whether any two variables (or sets of variables) are conditionally independent given any other set - just by examining the graph's topology (Pearl, 1988).
D-separation (where "d" stands for "directional") is a graph-theoretic criterion that determines whether a set of variables Z blocks all paths of influence between variables X and Y. If Z d-separates X from Y, then X and Y are conditionally independent given Z in every probability distribution that is compatible with the DAG. This is a profound claim: the qualitative structure of the graph constrains the quantitative properties of the probability distribution.
To understand d-separation, you need to understand three elementary path structures and how conditioning affects information flow through each. Consider three variables A, B, and C:
Chain (A → B → C): Information flows from A to C through B. If you observe B - that is, if you condition on B - then the path is blocked. Learning about A tells you nothing new about C once you already know B. In our hospital example: Workload → Burnout → Absenteeism. If you already know a nurse's burnout level, learning about their workload gives you no additional information about absenteeism (assuming the DAG is correct).
Fork (A ← B → C): B is a common cause of both A and C. Without conditioning on B, A and C are correlated (because they share a common cause). But if you condition on B, the path is blocked, and A and C become independent. Example: Management Style → Workload and Management Style → Role Clarity. High workload and poor role clarity tend to co-occur - not because one causes the other, but because both arise from the same management decisions. Once you know the management style, the correlation vanishes.
Collider (A → B ← C): Here, B is a common effect of A and C. This structure behaves counter-intuitively: the path between A and C is blocked by default. A and C are marginally independent. But if you condition on B (or any descendant of B), the path becomes active and A and C become dependent. This is the famous "explaining away" effect. Example: Bullying → Psychological Distress ← Work Overload. Bullying and overload may be independent risk factors. But if you learn that a worker is highly distressed and then discover they are not being bullied, you must increase your belief that they are overloaded - learning about one cause "explains away" the other when the effect is known (Pearl, 2009).
The collider effect often surprises people because conditioning on a variable creates a dependency that did not previously exist. Can you think of an example from your own workplace experience where learning about an outcome made two previously unrelated factors suddenly seem connected?
To determine whether X and Y are d-separated given Z, you examine every undirected path between X and Y and check whether every path is blocked. A path is blocked if it contains: (a) a chain or fork node that is in Z (i.e., conditioned on), or (b) a collider node where neither the collider nor any of its descendants is in Z. If every path is blocked, X and Y are d-separated by Z and therefore conditionally independent given Z (Koller & Friedman, 2009).
Spirtes, Glymour, and Scheines (2000) showed that d-separation is not merely a convenient reading tool - it is the foundation of causal discovery from data. Their PC algorithm works in reverse: it tests conditional independence relationships in data and uses the results to constrain which DAG structures are compatible with the observed independencies. The faithfulness assumption - that all conditional independencies in the data correspond to d-separation in the true graph - bridges the gap between statistical patterns and causal structure.
The widget below lets you practice d-separation reasoning on a psychosocial risk network. You will condition on different variables and trace whether information flows between query nodes - building the graph-reading fluency that is essential for interpreting and critiquing BN models.
With structure (the DAG) and parameters (CPTs) in place, a Bayesian Network becomes a reasoning engine. Inference is the process of computing the posterior probability of any variable given observed evidence about other variables. You enter evidence - "this worker reports high job demands and low supervisor support" - and the network propagates that information through the graph to update beliefs about unobserved variables like stress, burnout, and absenteeism risk.
Pearl's (1988) original contribution included an exact inference algorithm for tree-structured networks called belief propagation, in which messages pass between neighboring nodes until the network reaches equilibrium. For more complex networks with multiple paths, the junction tree algorithm converts the DAG into a secondary structure that enables exact computation (Koller & Friedman, 2009). When networks grow very large, approximate methods such as Monte Carlo sampling become necessary.
What makes BN inference particularly powerful for psychosocial risk is its bidirectionality. Evidence can enter the network at any node and propagate in any direction - not just from causes to effects but from effects back to causes. If you observe that absenteeism on a ward has suddenly increased, the network can compute updated probabilities for upstream factors: perhaps workload has increased, or perhaps a bullying complaint has gone unaddressed. This diagnostic reasoning - from symptoms to causes - is exactly what OHS investigators need when trying to understand a psychosocial incident (Sun et al., 2023).
García-Herrero et al. (2012) demonstrated this capability using a BN linking working conditions to occupational accidents through psychological and musculoskeletal disorders. By entering different evidence patterns - high workload, poor ergonomics, low social support - they computed how the probability of an accident changed under each scenario. This allowed them to compare intervention strategies quantitatively: which modifiable factor, when improved, would produce the greatest reduction in accident probability?
Critically, BN inference supports not just observation-based updating but also intervention-based reasoning - what Pearl (2009) calls the "do-calculus." There is a subtle but important difference between learning that supervisor support is high (observational evidence) and setting supervisor support to high through an intervention. In the observational case, high support might be correlated with other favorable conditions (perhaps well-managed departments have both good supervisors and reasonable workloads). In the interventional case, you force support to be high regardless of other conditions - breaking the natural correlations.
We will formalise interventional reasoning more fully in later chapters. For now, the key insight is that the DAG structure tells you when observational and interventional reasoning diverge, and Bayesian Networks give you the computational machinery to handle both.
The most critical question in any applied BN project is not computational but epistemological: who decides what the DAG looks like? In psychosocial risk, there are four major sources of structural knowledge, and best practice uses them in combination (Constantinou et al., 2011).
Domain experts - occupational psychologists, safety professionals, workplace health practitioners - carry rich causal knowledge about psychosocial dynamics. Structured elicitation protocols can extract this knowledge systematically: asking experts to identify variables, specify which variables directly influence which others, and rank the strength of influences. García-Herrero et al. (2013) used expert opinion to construct the initial structure of their occupational stress BN before refining it with data. The hazard webs you built in Chapter 1 are, in fact, informal expert elicitations - your domain knowledge, made visible.
The challenge is that experts can disagree, may confuse direct and indirect effects, and tend to include too many arrows (every factor seems to influence every other). Disciplined elicitation involves asking experts not just "does A influence B?" but "does A influence B directly, after accounting for all other variables in the model?" This is a harder question, but it is the question the DAG demands.
Decades of occupational health research have identified robust pathways: the demand-control-support model (Karasek & Theorell), effort-reward imbalance (Siegrist), job demands-resources theory (Bakker & Demerouti), and many others. These theoretical frameworks, and the empirical studies that test them, provide well-supported arrows for your DAG. Sun et al. (2023) explicitly drew on published psychosocial hazard literature to identify the nodes and edges of their construction mental health BN before quantifying the relationships with survey data.
Standards like ISO 45003 and guidance from Safe Work Australia (SWA) implicitly specify causal claims. When SWA identifies "low supervisor support" as a psychosocial hazard and "psychological injury" as a consequence, it is asserting a causal pathway. When ISO 45003 lists "organisational culture" as a factor that shapes exposure to specific hazards, it is specifying a parent-child relationship. Regulatory frameworks may not use the language of DAGs, but their content can be translated into DAG structure - and doing so makes their causal assumptions explicit and testable.
When sufficiently large datasets are available - survey data, HR records, incident databases - algorithms can learn DAG structure from statistical patterns. The PC algorithm of Spirtes et al. (2000) uses conditional independence tests to eliminate edges and orient arrows. Score-based methods evaluate candidate DAGs against data using metrics like BIC or Bayesian scoring functions. García-Herrero et al. (2012) used data from Spain's VI National Survey of Working Conditions (over 11,000 respondents) to learn and validate their BN structure.
However, purely data-driven structure learning has well-known limitations. Many different DAGs can encode the same set of conditional independencies (a phenomenon called Markov equivalence), and statistical tests can fail with small samples or noisy data. This is why the hybrid approach - starting with expert knowledge and literature, then refining with data - is considered best practice (Constantinou et al., 2011). As Sun et al. (2023) demonstrated in their construction industry study, combining published evidence with expert review and data validation produced a more robust and interpretable network than any single source could provide alone.
Consider the BN you would build for your Chapter 2 scenario. For each arrow in your network, which source of knowledge would you rely on most - expert opinion, published literature, regulatory guidance, or data? Are there arrows where the sources might disagree? How would you resolve such disagreements?
We can now see why the work you did in Chapters 1 and 2 was not merely "background." The hazard identification skills, the understanding of psychosocial dynamics, the ability to trace pathways from organisational conditions to individual outcomes - these are not supplementary to BN modelling. They are modelling. A Bayesian Network is only as good as its structure, and structure comes from domain knowledge. The algorithms are mathematically elegant, but without the substantive understanding of what causes what in workplaces - the kind of understanding that occupational health professionals spend careers developing - the algorithms have nothing meaningful to compute.
This is the fundamental insight of this chapter: Bayesian Networks are not a replacement for expert judgement about psychosocial risk. They are a formalisation of that judgement - a way to make it precise, testable, updatable, and computationally useful. The informal intuitions you have been developing are the raw material; the BN formalism is the precision tool that shapes them into a functioning model.
Let us consolidate what we have learned. A Bayesian Network rests on three pillars:
Together, these three pillars transform the hazard webs you drew in Chapter 1 into functional reasoning engines. A well-specified BN can answer questions that no checklist, no risk matrix, and no simple regression can address: "Given that we observe these specific conditions, what is the probability of this specific outcome, and how would that probability change if we intervened on this specific factor?"
The construction industry study by Sun et al. (2023) illustrates this integration beautifully. Their BN included root nodes for poor physical environment, contract pressure, and lack of coworker support; mediating nodes for emotional demands, work-life conflict, and job insecurity; and outcome nodes for anxiety, depression, and overall mental health. The DAG was informed by literature and expert review. The CPTs were estimated from survey data collected from 186 construction practitioners. And inference revealed that 66% of the sample fell into poor mental health categories - a finding that only emerged from the interaction of multiple risk factors as captured by the network's conditional probability structure.
You now have the language - DAGs, CPTs, d-separation, inference. In Chapter 4, we will put this language to work. You will build your first complete Bayesian Network for a psychosocial risk scenario, populating it with probabilities derived from published research and your own expert judgement. We will also confront the practical challenges of parameterisation: what do you do when the data are sparse, the experts disagree, and the literature only partially covers your context? The answer involves a Bayesian approach to the parameters themselves - learning from data while respecting prior knowledge.
Constantinou, A. C., Fenton, N., & Neil, M. (2011). Incorporating expert knowledge when learning Bayesian network structure: A medical case study. Artificial Intelligence in Medicine, 53(3), 181–204. https://doi.org/10.1016/j.artmed.2011.09.002
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.2012.04.007
García-Herrero, S., Mariscal, M. A., Gutiérrez, J. M., & Ritzel, D. O. (2013). Using Bayesian networks to analyze occupational stress caused by work demands: Preventing stress through social support. Accident Analysis & Prevention, 57, 114–123. https://doi.org/10.1016/j.aap.2013.04.009
Koller, D., & Friedman, N. (2009). Probabilistic graphical models: Principles and techniques. MIT Press. https://mitpress.mit.edu/9780262013192/probabilistic-graphical-models/
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
Pearl, J. (2009). Causality: Models, reasoning, and inference (2nd ed.). Cambridge University Press. https://doi.org/10.1017/CBO9780511803161
Spirtes, P., Glymour, C., & Scheines, R. (2000). Causation, prediction, and search (2nd ed.). MIT Press. https://direct.mit.edu/books/monograph/2057/Causation-Prediction-and-Search
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(4), Article 04023011. https://doi.org/10.1061/JCEMD4.COENG-12905