Forward inference - pushing probability from causes to effects to predict workplace health outcomes
It is early June, and the director of nursing at a metropolitan public hospital is staring at a workforce planning dashboard. Three things are happening simultaneously. First, a severe influenza season has arrived two weeks earlier than predicted, spiking patient admissions by 30%. Second, a hiring freeze imposed last quarter means the ward is operating with 15% fewer registered nurses than the staffing model recommends. Third, the experienced charge nurse on the night shift - widely regarded as the emotional anchor of the team - has just submitted a transfer request, citing her own exhaustion. The director needs to answer a concrete question before Friday's executive meeting: given these conditions, what is the probability that clinical burnout across the ward will exceed the threshold at which patient safety incidents historically spike? She does not need a vague sense that things are bad. She needs a number, a defensible estimate, and a clear picture of which inputs are driving that number the most.
This is the forward question - the question of predictive inference. We know the current state of the causes. We want to know what those causes predict about the effects. This chapter teaches you how Bayesian Networks answer it.
In Chapters 2 and 3, we constructed the architecture of Bayesian Networks: directed acyclic graphs encoding qualitative relationships between variables, and conditional probability tables (CPTs) quantifying those relationships. The architecture alone, however, is inert. A network sitting on a desk with no observations entered is like a weather model with no satellite data - structurally sound but operationally useless. The moment we enter evidence - an observation about the actual state of one or more nodes - the network comes alive. Probabilities ripple outward from the observed nodes through the conditional probability tables, updating our beliefs about every other variable in the system.
When that evidence flows from causes toward effects, from parent nodes toward their descendants, the process is called forward propagation or, more formally, predictive inference (Koller & Friedman, 2009). This is the mode of reasoning that dominates practical risk management. A safety regulator asks: if this mine operates 12-hour rotating shifts in a remote location, what is the expected probability of depressive symptoms among workers? A startup founder asks: if we restructure the engineering team while simultaneously asking everyone to absorb new roles, what happens to psychological injury risk? In each case, the reasoning direction is the same - from known upstream conditions to predicted downstream outcomes.
Forward propagation is not merely intuitive reasoning dressed up in mathematics. It is a precise computational procedure that handles complexity our intuitions cannot. Humans are notoriously poor at combining multiple uncertain inputs. We tend to anchor on one salient factor and neglect others. We underestimate the compounding effect of simultaneous hazards. We struggle to propagate uncertainty through chains of more than two or three variables (Arora et al., 2019). The Bayesian Network, by contrast, propagates evidence through every path simultaneously, weighting each configuration of parent states by its probability, and producing an updated distribution at every descendant node. The landmark algorithm that makes this tractable was developed by Lauritzen and Spiegelhalter (1988), who showed that local computations on the graph structure - exploiting conditional independencies - can replace the exponentially expensive computation of the full joint distribution.
Before reading further, consider our hospital director's situation. She faces three simultaneous changes: reduced staffing, increased patient acuity (emotional demands), and loss of a key supervisor. If you had to estimate the probability of high burnout on the ward, would you combine these three factors additively (each adds a fixed increment to burnout risk) or do you suspect the combination is worse than the sum of its parts? Hold your intuition - we will test it shortly.
Evidence entry is the act of clamping a node to an observed state. When we observe that staffing is low, we replace the prior probability distribution over Staffing Level - which might have been P(Adequate) = 0.60, P(Low) = 0.40 - with a certainty: P(Low) = 1.00. This is called hard evidence. In some applications, we may instead enter soft evidence (also called virtual evidence), where we update the distribution without collapsing it entirely - for example, shifting to P(Adequate) = 0.20, P(Low) = 0.80 because the report is preliminary. For now, we work with hard evidence because it makes the propagation mechanics transparent.
The critical point is that entering evidence at one node does not directly change the CPTs anywhere in the network. The conditional probability tables are fixed parameters of the model. What changes is the effective distribution flowing into downstream CPTs. When Staffing Level is clamped to Low, the CPT for Workload still contains both rows - P(Workload | Staffing = Adequate) and P(Workload | Staffing = Low) - but only the second row is now active, because the first row is multiplied by P(Staffing = Adequate) = 0.
Consider the simplest case: Staffing Level → Workload. With no evidence, the marginal distribution over Workload is computed by marginalization - summing over all possible states of the parent, weighted by the parent's probability:
P(Workload = Excessive) = P(Excessive | Adequate) × P(Adequate) + P(Excessive | Low) × P(Low) = 0.20 × 0.60 + 0.75 × 0.40 = 0.42
Now enter evidence: Staffing = Low. The computation collapses:
P(Workload = Excessive | Staffing = Low) = 0.75
The probability of excessive workload jumps from 0.42 to 0.75 - a substantial shift driven by a single observation. This is the simplest possible forward propagation: one parent, one child, one piece of evidence.
Now consider the Stress node in our hospital network. Stress has three parents: Workload, Emotional Demands, and Supervisor Support. Its CPT has 2 × 2 × 2 = 8 rows (for binary variables), each specifying P(Stress = High | specific combination of parent states). When we have evidence on some parents but not others, we must marginalize - sum over all configurations of the unobserved parents, weighting each configuration by its probability:
P(Stress = High | evidence) = Σw,e,s P(Stress = High | W=w, E=e, S=s) × P(W=w | evidence) × P(E=e | evidence) × P(S=s | evidence)
This formula is the engine of forward propagation in networks with multiple parents. It respects every possible configuration. When all three parents are observed, the sum collapses to a single term - the CPT row corresponding to the observed configuration. When none are observed, all eight terms contribute, weighted by prior probabilities. When some are observed and some are not, we get a partial collapse - the observed parents lock in their states while we sum over the unobserved ones.
This is precisely the pattern that Mariscal et al. (2013) employed when they built a Bayesian network from Spanish working conditions survey data to study occupational stress. Their model included intellectual demands, overwork, shift length, and social support as parent nodes feeding into a stress outcome. By entering different evidence configurations - high overwork with low social support, for example - they produced forward-propagated stress probabilities that quantified how combinations of hazards interact. The CPT values in their model were learned from 11,054 survey responses, grounding the propagation in empirical data rather than expert guesses.
García-Herrero et al. (2017), working with healthcare professionals specifically, found that forward propagation through their network revealed a crucial asymmetry: emotional demands exerted a greater influence on stress probability than family demands, even though both were significant. This is exactly the kind of insight that marginalization delivers - by comparing the propagated output under different evidence configurations, we can identify which parent nodes carry the most weight in the CPT.
In a network where Stress has three binary parents, there are 8 possible parent configurations. If you observe all three parents, you read one row of the CPT. If you observe none, you weight all 8 rows by their joint probability. What happens if you observe two parents and leave one unobserved? How many CPT rows contribute to the computation? (Answer: 2 - one for each state of the unobserved parent, each weighted by that state's probability.)
Let us walk through the full hospital network that our nursing director faces. The network has eight nodes arranged in a causal chain with convergent structures:
The CPT values used here are calibrated to reflect occupational health literature. The baseline (no evidence) probability of high burnout is approximately 35% - consistent with prevalence estimates in healthcare settings (Shi et al., 2019). Under adverse conditions, we expect this to rise substantially.
The interactive engine below lets you explore this network yourself. Enter evidence at any root node, and watch the consequences propagate forward through every descendant.
If you experimented with the propagation engine, you likely noticed something important: the combined effect of multiple adverse conditions is not simply additive. Setting Staffing to Low raises burnout probability somewhat. Adding High Emotional Demands raises it further. But adding Low Supervisor Support on top of the first two produces a jump that exceeds what you would predict by simply adding the individual effects. This is the signature of interaction effects encoded in the CPT - the conditional probability of high stress given all three adverse inputs (0.90) is higher than what a linear combination of the individual effects would yield. Bayesian Networks capture these interactions naturally because the CPT represents the full conditional distribution, not a linear approximation of it.
Returning to our nursing director's scenario: Staffing = Low, Emotional Demands = High (flu season means more patient distress, more families in crisis, more end-of-life conversations), and Supervisor Support = Low (the experienced charge nurse has departed). Propagating through the network, P(Stress = High) rises to 0.90 - the worst row of the CPT. This cascades: P(Burnout = High) reaches approximately 0.70, and P(Turnover Intent = High) climbs to roughly 0.52. In other words, the model predicts that more than half the ward will be contemplating leaving.
Under Safe Work Australia's (SWA) Code of Practice for managing psychosocial hazards, employers must eliminate or minimise psychosocial risks so far as is reasonably practicable. A model showing 70% burnout probability under current conditions constitutes evidence that the risk is not being adequately managed. More importantly, the model can show which intervention - increasing staffing, reducing emotional load through support programs, or replacing supervisory capacity - would produce the largest reduction. We return to this prioritization logic in the sensitivity analysis section below.
In the technology startup scenario, role ambiguity and organisational change serve as root hazards feeding into job insecurity, which in turn affects psychological distress and, ultimately, psychological injury risk. When a startup pivots its product strategy (Organisational Change = High) while simultaneously asking engineers to absorb design responsibilities they were never hired for (Role Ambiguity = High), the forward-propagated probability of psychological injury rises sharply. The compounding pattern is similar to the hospital case: the CPT for job insecurity under simultaneous high role ambiguity and high organisational change produces a probability that exceeds the sum of the individual contributions.
The mining scenario illustrates a different kind of compounding. Remote location isolation and shift work are the root hazards. Torquati et al. (2019), in their meta-analysis of seven longitudinal studies encompassing 28,431 workers, found that shift work was associated with a 33% increase in the risk of depressive symptoms (ES = 1.33, 95% CI = 1.02, 1.74). The mechanism runs through circadian disruption, impaired sleep quality, and reduced opportunities for social connection. In a remote mining camp, the social isolation amplifies the shift work effect: workers on rotating 12-hour shifts in a remote location cannot access the community supports that partially buffer shift workers in urban settings. The CPT encoding this interaction produces a depression probability under combined remote isolation and shift work that is substantially higher than either factor alone would predict.
In all three scenarios, the combined effect of multiple hazards exceeded what you might predict by considering each hazard independently. Why does this matter for regulatory compliance? Consider the SWA requirement to assess risks from psychosocial hazards - would a checklist approach that evaluates each hazard in isolation capture the interaction effects that Bayesian Networks reveal?
Reading about forward propagation is one thing. Developing accurate intuitions for how probabilities flow through a network is another. Research on probabilistic reasoning consistently shows that people underestimate the impact of converging evidence - multiple parents pointing to the same adverse state - and overweight single dramatic inputs while neglecting indirect paths (Arora et al., 2019). The following exercise is designed to calibrate your intuitions by giving you immediate feedback on your estimates.
If you completed all eight challenges, the widget identified patterns in your estimation errors. The most common bias is underestimating compound effects - when two or three parent nodes are simultaneously set to adverse states, most people predict a stress probability lower than the CPT actually yields. This is not a character flaw; it reflects well-documented limitations in human probabilistic reasoning. The Bayesian Network compensates for exactly this limitation, which is why computational propagation is not merely a convenience but a genuine decision-support tool (Koller & Friedman, 2009).
Forward propagation answers the question what will happen? But for risk management, an equally important question is which input matters most? This is the domain of sensitivity analysis - the systematic study of how changes in input variables affect output probabilities.
The foundational methodology was developed by Laskey (1995), who showed that sensitivity values - formally, partial derivatives of output probabilities with respect to CPT parameters - can be computed analytically within the Bayesian Network framework. The practical implication is powerful: we can rank every input variable by its influence on any chosen output, identifying the leverage points where intervention would produce the greatest reduction in adverse outcome probability.
The simplest form of sensitivity analysis, and the one most accessible to practitioners, is parameter sweeping: for each parent or ancestor node, we fix that node to its best state and compute the output, then fix it to its worst state and compute the output. The difference - the swing - measures the node's influence. Nodes with large swings are high-leverage targets for intervention. Nodes with small swings, no matter how intuitively alarming, are poor candidates for resource allocation because changing them would not substantially alter the outcome.
Safe Work Australia's Code of Practice requires that employers implement controls that are reasonably practicable. "Reasonably practicable" involves weighing the likelihood and degree of harm against the availability and cost of controls. Sensitivity analysis directly informs this weighing. If the model shows that Supervisor Support has a swing of 0.35 on burnout probability while Staffing Level has a swing of 0.20, then investing in supervisor training and support programs is a higher-leverage intervention than hiring additional staff - even if hiring is the more obvious response. The model provides evidence for prioritizing interventions based on predicted impact, not on intuition or convention.
García-Herrero et al. (2017) demonstrated this approach with their healthcare worker network. By systematically varying each parent node, they found that emotional demands had the largest sensitivity value for the stress outcome - larger than working hours, larger than family demands, and larger than control over work pace. This finding directly informed their intervention recommendations: addressing emotional demands through debriefing programs and emotional support training should take priority over workload reduction programs, because the model predicted a larger outcome shift per unit of resource invested.
The standard visualization for sensitivity analysis is the tornado diagram - a horizontal bar chart where each bar represents one input variable, the bar length represents the swing (output change from best to worst input state), and bars are sorted from largest to smallest. The resulting shape, widest at the top and narrowing toward the bottom, resembles a tornado. The variables at the top of the tornado are the ones that matter most. The variables at the bottom are those whose state barely affects the output - they may be important for other outcomes, but for the chosen target, they are not leverage points.
If you explored the tornado diagram with Burnout as the target, you likely found that Emotional Demands and Supervisor Support produced larger swings than Staffing Level. This might seem counterintuitive - surely staffing is the most important factor in a hospital? - but it reflects the network structure. Staffing affects Burnout only indirectly, through Workload and then through Stress. Emotional Demands and Supervisor Support affect Stress directly, and the CPT values give them substantial weight. The lesson is important: the most obvious hazard is not always the highest-leverage intervention target. The network reveals influence structures that intuition alone cannot reliably identify.
We will return to sensitivity analysis in Chapter 7, where it becomes the foundation for intervention prioritization - converting model-derived leverage points into actionable risk control strategies ranked by expected impact.
Imagine you are presenting the tornado diagram to a hospital executive committee. The diagram shows that Supervisor Support has the largest swing on burnout probability. The CEO asks: "So you're saying we should invest in supervisor training instead of hiring more nurses?" How would you explain that the sensitivity analysis identifies leverage, not exclusivity - that the highest-leverage intervention should be prioritized first but does not render other interventions unnecessary?
A natural concern at this stage is: how much should we trust the propagated numbers? The outputs are only as reliable as the CPT values, and those values were either learned from data or elicited from experts - both processes introduce uncertainty. Laskey (1995) addressed this concern directly by showing that sensitivity values tell us not only which inputs influence the output most, but also which CPT parameters the output is most sensitive to. If the output is highly sensitive to a particular CPT entry, then the precision of that entry matters a great deal, and we should invest effort in validating it. If the output is insensitive to a CPT entry, then even substantial errors in that entry will not meaningfully change the prediction.
This creates a practical workflow for model validation. After building the network structure and populating initial CPT values (from literature, expert judgement, or organisational data), the modeller runs a parameter sensitivity analysis. The parameters with the highest sensitivity values are flagged for additional validation - perhaps by collecting targeted organisational data, consulting additional domain experts, or comparing against published epidemiological estimates. Parameters with low sensitivity values can remain at their initial estimates without significant concern.
Shi et al. (2019), in their study of correction workers' burnout, demonstrated a complementary approach: they learned the entire network structure and CPT values from data using the Bayesian search algorithm, then used the learned model's forward propagation outputs to identify the most probable scenario linking psychosocial variables. The data-driven CPT values provided a level of empirical grounding that expert-elicited values alone could not achieve, though the approach requires sample sizes sufficient to estimate the CPT entries reliably - a constraint that limits its application in small organisations.
Forward propagation is not an academic exercise. It produces actionable risk intelligence - quantified predictions that can be directly mapped to regulatory obligations and organisational decisions. Consider three concrete applications:
Arora et al. (2019), reviewing Bayesian Networks as risk prediction tools in health research, emphasised that BNs offer a distinctive advantage over regression models in this context: they naturally handle individual-level prediction. A regression model estimates an average effect across the population. A Bayesian Network, by entering specific evidence about this ward, this team, this shift schedule, produces a prediction tailored to the specific configuration of conditions - precisely what a risk manager needs.
We close this chapter by acknowledging a tantalizing limitation. Everything we have discussed flows in one direction: from causes to effects. We enter evidence at parent nodes and read updated probabilities at descendant nodes. This is the predictive question: given what we know about conditions, what outcomes should we expect?
But there is another question - arguably more urgent in the aftermath of an adverse event: given that burnout has occurred, what were the most likely causes? This is the diagnostic question, and it runs backward through the network, from effects to causes. A worker has been diagnosed with a psychological injury. The regulator needs to determine which workplace conditions most likely contributed. The same network, the same CPTs, and the same propagation machinery can answer this question - but the reasoning direction is reversed, and the mathematics require Bayes' theorem in its full glory.
Forward propagation tells us what will happen. Diagnostic inference tells us why something happened. Together, they constitute the complete inferential capability of Bayesian Networks. But they are different cognitive operations, with different practical applications and different interpretive pitfalls. The diagnostic question - backward propagation - is the subject of Chapter 5.
Forward propagation answers "what will happen?" - but after a psychological injury has already occurred, the pressing question reverses: "what caused it?" In Chapter 5, Diagnosing Causes, we flip the direction of inference. Using the same networks and the same CPTs, we enter evidence at outcome nodes and propagate backward to identify the most probable root causes. You will discover that Bayes' theorem transforms the same conditional probability tables into a powerful diagnostic engine - and that diagnostic reasoning introduces interpretive subtleties that predictive reasoning does not.
Arora, P., Boyne, D., Slater, J. J., Gupta, A., Brenner, D. R., & Druzdzel, M. J. (2019). Bayesian networks for risk prediction using real-world data: A tool for precision medicine. Value in Health, 22(4), 439–445. https://pubmed.ncbi.nlm.nih.gov/30975395/
García-Herrero, S., Mariscal, M. A., Guadix, J., & Ritzel, D. O. (2017). The influence of recognition and social support on European health professionals' occupational stress: A Demands-Control-Social Support-Recognition Bayesian network model. BioMed Research International, 2017, 4519529. https://pubmed.ncbi.nlm.nih.gov/29250540/
Koller, D., & Friedman, N. (2009). Probabilistic graphical models: Principles and techniques. MIT Press. https://mitpress.mit.edu/9780262013192/probabilistic-graphical-models/
Laskey, K. B. (1995). Sensitivity analysis for probability assessments in Bayesian networks. IEEE Transactions on Systems, Man, and Cybernetics, 25(6), 901–909. https://doi.org/10.1109/21.384252
Lauritzen, S. L., & Spiegelhalter, D. J. (1988). Local computations with probabilities on graphical structures and their application to expert systems. Journal of the Royal Statistical Society: Series B, 50(2), 157–194. https://doi.org/10.1111/j.2517-6161.1988.tb01721.x
Mariscal, M. A., Herrera, S., García-Herrero, S., Rubio-Romero, J. C., & Guadix, J. (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
Shi, Y., Bainbridge, H. T. J., & Leung, A. S. M. (2019). Correction workers' burnout and outcomes: A Bayesian network approach. Journal of Occupational Health Psychology, 24(1), 100–112. https://pmc.ncbi.nlm.nih.gov/articles/PMC6352158/
Torquati, L., Mielke, G. I., Brown, W. J., Burton, N. W., & Kolbe-Alexander, T. L. (2019). Shift work and poor mental health: A meta-analysis of longitudinal studies. American Journal of Public Health, 109(11), e13–e20. https://doi.org/10.2105/AJPH.2019.305278