Class 7

The Intervention Calculus

From probabilistic insight to defensible action - learning the algebra of workplace change

The Gap Between Knowing and Doing

Everything we have built across the first six chapters of this course - the directed acyclic graphs, the conditional probability tables, the diagnostic reasoning, the calibrated assessments - has operated on what Judea Pearl calls the first rung of the Ladder of Causation: the level of association, of observing and updating beliefs (Pearl & Mackenzie, 2018). When we entered evidence that a nurse reported high workload and watched beliefs about psychological distress update via Bayes' theorem, we were asking an associational question: given what I see, what should I believe?

But the committee at Meridian Hospital is not asking what to believe. They are asking what to do. And this seemingly small shift - from passive observation to active intervention - requires us to climb to the second rung of the Ladder: the rung of intervention. Pearl's fundamental insight, developed formally in Causality (Pearl, 2009a), is that no amount of observational data alone can answer interventional questions. The mathematics of "seeing" and the mathematics of "doing" are fundamentally different operations, and confusing them is one of the most consequential errors in applied risk management.

This chapter introduces what we will call the intervention calculus - not Pearl's full formal do-calculus (which requires notation most practitioners will never need), but the intuitive reasoning framework that allows psychosocial risk professionals to use Bayesian networks as intervention planning tools. By the end of this chapter, you will be able to take a calibrated BN, simulate the effects of proposed interventions, rank those interventions by cost-effectiveness, and determine whether you should act now or gather more information first.

---

Observation versus Intervention: The See/Do Distinction

Let us return to the Meridian Hospital network. One of the nodes is Supervisor Support, which has two parent nodes: Management Commitment and Supervisor Workload. When management is committed to staff wellbeing and supervisors themselves are not overwhelmed, supervisor support tends to be high. Supervisor Support, in turn, causally influences Psychological Distress and Team Cohesion downstream.

Now consider two scenarios. In the first, we observe that Supervisor Support is high - perhaps through a staff survey. In the second, we intervene to make Supervisor Support high - perhaps by implementing a structured supervisor training program with protected time for staff check-ins. In both cases, the value of the Supervisor Support node is set to "High." So why would the downstream effects differ?

Why observation and intervention diverge

When we observe that Supervisor Support is high, Bayes' theorem updates our beliefs in all directions - not just downstream but also upstream. Observing high support is evidence about the parents of that node: it becomes more likely that Management Commitment is strong (because committed management tends to produce supportive supervisors) and less likely that Supervisor Workload is crushing (because overwhelmed supervisors struggle to provide support). These updated beliefs about parent nodes then propagate through their own downstream pathways, creating a cascade of revised probabilities throughout the network.

When we intervene to set Supervisor Support to high, something categorically different happens. Our intervention is an external force acting on the system. We are not learning that the natural dynamics of the workplace produced high support; we are forcing it. The intervention tells us nothing about Management Commitment or Supervisor Workload - those variables could be in any state. As Pearl (2009a) formalised, intervention on a variable severs the causal arrows coming into that variable. The node's value is no longer determined by its parents; it is determined by us.

Pearl (2009b) emphasised this point as foundational: "causal relations cannot be expressed in the language of probability alone - any mathematical approach to causal analysis must acquire new notation" (p. 99). That new notation is the do-operator. The probability of psychological distress given that we observe high supervisor support is written P(Distress | Support = High). The probability of distress given that we intervene to make support high is written P(Distress | do(Support = High)). These are different quantities, computed by different procedures, and they can yield very different numbers.

Graph surgery: the mechanics of simulating interventions

The computational procedure for simulating an intervention in a Bayesian network is elegantly simple - so simple that it has earned the vivid name graph surgery (Pearl, 2009a). To compute the effect of intervening on variable X:

1. Delete all incoming arrows to node X. This "mutilates" the graph, severing X from its parents. The node is no longer a function of its natural causes; it is now under external control.
2. Set X to the desired value. Fix the node's probability table so that it takes the intervention value with certainty (e.g., Supervisor Support = High with probability 1.0).
3. Propagate beliefs through the modified network using standard Bayesian updating. Because the incoming arrows to X have been removed, information does not flow backward through X to its former parents.

The result is the post-interventional distribution - the probability distribution over all variables in the network that would obtain if we performed the intervention. This is what Pearl calls the "truncated factorization" of the joint distribution: we take the original factorization of the joint probability (the product of all conditional probability tables) and simply delete the factor for X, replacing it with a point mass at the intervention value (Pearl, 2009a).

Gonzalez-Jimenez et al. (2022) demonstrated this approach in a risk-assessment context, constructing a BN with 41 causal factors and computing intervention effects using graph surgery. Their key finding echoes our pedagogical point: "associative reasoning answers 'how does new evidence change beliefs about Y?' while intervention reasoning answers 'how does doing X change Y?'" (p. 4). The difference, they showed, was not merely academic - it produced materially different risk estimates and therefore different intervention recommendations.

The widget below lets you experience this distinction directly. Use the Meridian Hospital network to compare what happens when you observe versus intervene on Supervisor Support.

---

From Single Interventions to Ranked Priorities

The see/do distinction is critical, but it is only the first step. A psychosocial risk committee rarely faces a binary choice. The Meridian Hospital team has at least six modifiable hazard nodes in their network: Workload, Supervisor Support, Role Clarity, Job Control, Social Support, and Management Commitment. Each could be the target of an intervention. The committee needs to know: which intervention produces the greatest reduction in our target outcome?

The answer requires systematically applying graph surgery to each modifiable node, computing the post-interventional probability of the target outcome (say, Psychological Injury), and comparing the results. This procedure - what we might call an intervention sweep - yields a ranked list of intervention targets ordered by their causal impact on the outcome of interest.

The intervention sweep procedure

For each modifiable node X_i in the network:

1. Perform graph surgery: remove incoming arrows to X_i.
2. Set X_i to its best feasible state (e.g., Workload = Manageable).
3. Propagate and record the post-interventional probability of the target outcome.
4. Compute the absolute risk reduction (ARR): baseline probability minus post-interventional probability.
5. Restore the original network and repeat for the next node.

This is exactly the approach demonstrated by Mohammadfam et al. (2017), who constructed a workplace safety BN and systematically computed the impact of improving each modifiable factor on overall safety behaviour. Their results showed that the nodes with the strongest statistical associations were not always the best intervention targets - a finding that only becomes visible when you use intervention reasoning rather than associational reasoning. Similarly, García-Herrero et al. (2013) used a BN modelling occupational stress to quantify how social support interventions reduce stress caused by high demands, demonstrating the practical value of intervention simulation in psychosocial risk contexts.

Introducing cost: from effectiveness to cost-effectiveness

A ranked list of interventions by risk reduction is valuable, but it is incomplete. In every real organisation, interventions compete for finite resources. A program that reduces psychological injury risk by 12 percentage points but costs $500,000 may be less attractive than one that reduces risk by 8 percentage points at $50,000. This is where cost-effectiveness analysis enters the intervention calculus.

Gavious et al. (2018), in their systematic review of the economic evaluation of occupational safety and health interventions, found that economic evidence for evaluating psychosocial interventions "remains thin" but that "economic incentives are a key decision factor for employers" (p. 226). The gap they identified - between the recognised need for cost-effectiveness data and the scarcity of such data - is precisely the gap that BN-based intervention simulation can help fill.

The cost-effectiveness ratio for an intervention is computed as:

Cost-Effectiveness Ratio = Intervention Cost ÷ Absolute Risk Reduction

Lower ratios indicate better value: less money per unit of risk reduction. When we plot each intervention as a point in cost–effectiveness space (cost on the x-axis, risk reduction on the y-axis), a Pareto frontier emerges - the set of interventions for which no other single intervention is both cheaper and more effective. Interventions not on the frontier are said to be dominated: there exists at least one alternative that is superior on both dimensions. Rational resource allocation should focus on the frontier.

Building an intervention portfolio

In practice, organisations rarely implement a single intervention. They assemble portfolios - combinations of interventions targeting different nodes in the causal network. The power of the BN framework is that it allows us to simulate combined interventions: perform graph surgery on multiple nodes simultaneously and propagate the joint effect. The cumulative risk reduction from a portfolio is typically not the simple sum of individual reductions (because interventions may share downstream pathways), but the BN computes the correct combined effect automatically.

This portfolio approach aligns directly with the hierarchy of controls mandated by the Safe Work Australia Code of Practice and the systematic approach recommended by ISO 45003. Both frameworks emphasise that psychosocial risk management requires addressing hazards at multiple levels - not placing all organisational eggs in one interventional basket.

The widget below lets you conduct a full intervention sweep, incorporate costs, identify the Pareto frontier, and build an intervention portfolio for Meridian Hospital.

---

When the Best Action Is to Investigate: The Expected Value of Information

There is an intervention the committee has not yet considered - one that does not appear in any hazard node or budget line. It is the intervention of gathering more information. Sometimes the most valuable thing an organisation can do is not act, but learn. The question is: how do we quantify the value of learning?

This is the domain of expected value of information analysis, formalised in the foundational work of Raiffa and Schlaifer (1961). The core idea is breathtakingly practical: before spending money on an intervention, ask whether spending a smaller amount on investigation would change which intervention you select - and if so, how much better your decision would be.

Expected Value of Perfect Information (EVPI)

The Expected Value of Perfect Information (EVPI) for a given node answers the question: if we could learn the true state of this variable with absolute certainty, how much would that improve our expected outcome? More precisely, EVPI is the difference between the expected value of the best decision made after learning the node's true state (averaged over all possible states) and the expected value of the best decision made with current uncertainty.

Consider Role Clarity in the Meridian Hospital network. Currently, based on survey data from six months ago, the committee estimates Role Clarity is Low with probability 0.6 and High with probability 0.4. Their current best intervention plan (the portfolio from the Priority Ranker) yields an expected psychological injury rate of 18%. But:

• If Role Clarity were known to be Low (probability 1.0), the optimal portfolio might shift - perhaps adding a role-clarification workshop - and yield an expected injury rate of 15%.
• If Role Clarity were known to be High (probability 1.0), the optimal portfolio might drop the workshop (saving money) and still yield 17%.

The EVPI is the expected improvement: (0.6 × improvement if Low) + (0.4 × improvement if High). If this number is large, it means the committee's decision is sensitive to Role Clarity, and investing in a quick assessment of role clarity before acting could substantially improve outcomes. If EVPI is near zero, the current uncertainty about Role Clarity does not matter much - the same portfolio is optimal regardless.

Expected Value of Sample Information (EVSI)

EVPI represents an upper bound - the value of perfect knowledge, which is rarely attainable. In practice, we collect imperfect information: a survey with measurement error, a focus group with selection bias, an audit with limited scope. The Expected Value of Sample Information (EVSI) quantifies the value of such imperfect evidence. As Raiffa and Schlaifer (1961) formalised, EVSI is always less than or equal to EVPI, and the difference depends on the diagnostic accuracy of the data collection method. If a survey reliably distinguishes Low from High role clarity (high sensitivity and specificity), EVSI approaches EVPI. If the survey is noisy, EVSI may be much lower.

The practical decision rule is simple: if EVSI exceeds the cost of the investigation, investigate before acting. If EVSI is less than the investigation cost, act now with current knowledge. This transforms the seemingly subjective question "should we collect more data?" into a quantified comparison.

Connecting investigation to action

The expected value of information framework creates a natural bridge between the diagnostic reasoning of Chapter 5 and the implementation planning of Chapter 8. Chapter 5 taught us how to update beliefs when evidence arrives; this chapter teaches us which evidence to seek based on its decision-relevant value. And the answer is not always the evidence that reduces the most uncertainty in a general sense (Shannon entropy, for instance), but the evidence that most changes what we would do.

This is a subtle but crucial distinction. A risk professional might want to investigate the node with the widest uncertainty bands, reasoning that "we know the least about this." But EVPI analysis often reveals that the most decision-relevant investigation targets a node with moderate uncertainty whose true state would tip the balance between two close-ranked interventions. The value of information is decision-relative, not absolute.

The widget below lets you explore this directly. Based on your portfolio from the Priority Ranker, determine which unobserved node has the highest EVPI - and compare that to your intuitive guess.

---

Putting It All Together: The Intervention Calculus Workflow

We can now articulate a complete workflow for converting a calibrated Bayesian network into a defensible intervention plan. This is the intervention calculus in practice:

1. Establish the baseline. Using the calibrated BN (with all current evidence entered), record the probability of the target outcome. For Meridian Hospital, this is P(Psychological Injury = Yes) = 0.34.
2. Identify modifiable nodes. Distinguish between nodes that represent hazards amenable to intervention and those that represent fixed features of the environment (e.g., hospital type, patient acuity mix).
3. Conduct the intervention sweep. For each modifiable node, perform graph surgery, set the node to its best feasible state, propagate, and record the post-interventional outcome probability. Compute absolute risk reductions.
4. Incorporate costs. Attach estimated intervention costs to each modifiable node. Compute cost-effectiveness ratios. Identify the Pareto frontier.
5. Build the portfolio. Select a combination of non-dominated interventions that fits within the budget. Simulate the combined effect using simultaneous graph surgery.
6. Assess information value. For each unobserved node with substantial remaining uncertainty, compute EVPI. If any EVPI exceeds the cost of investigation, investigate before finalising the portfolio.
7. Document and justify. Present the ranked interventions, their expected effects, and the rationale for each selection - grounded in the causal structure of the network, not merely in statistical associations.

This workflow does something that no purely statistical or purely qualitative approach can achieve: it quantifies the causal impact of proposed actions, accounts for confounding structure, incorporates resource constraints, and identifies when more information is needed. It is, in essence, the decision-support architecture that frameworks like ISO 45003 call for but rarely specify how to implement.

Common pitfalls to avoid

Even with the intervention calculus in hand, several traps await the unwary practitioner:

• Confusing association with intervention. The entire first section of this chapter exists because this error is ubiquitous. Always use graph surgery for intervention reasoning; never substitute observational conditioning.
• Ignoring interactions between interventions. The risk reduction from intervening on two nodes simultaneously is generally not the sum of individual reductions. Always simulate the combined portfolio, not just the sum of parts.
• Assuming all states are equally achievable. Setting Workload to "Low" in the model is easy; achieving it in a hospital emergency department is hard. The feasibility of the target state must be assessed alongside the cost.
• Neglecting implementation fidelity. Graph surgery sets a node to its target value with certainty. Real interventions achieve partial, probabilistic shifts. A more realistic simulation would set the intervention node to, say, 80% High and 20% Moderate, reflecting expected implementation quality.
• Treating EVPI as a reason to delay indefinitely. There is always a node with non-zero EVPI. The decision to investigate must compare EVPI to investigation cost and the cost of delayed action. Sometimes acting under uncertainty is better than waiting for perfect knowledge.

---

The Ethical Imperative of Causal Reasoning

There is a deeper reason why the see/do distinction matters beyond computational accuracy: it carries ethical weight. When an organisation chooses an intervention based on observational associations alone - "supervisor support is correlated with lower distress, so let's mandate supervisor check-ins" - they may be investing resources in an action whose apparent effect is largely confounded by management commitment. The intervention fails, workers continue to suffer, and the organisation concludes that "psychosocial interventions don't work." This is not merely a waste of money; it is a failure of the duty of care owed to workers.

Pearl and Mackenzie (2018) argued that the inability to distinguish observational from interventional reasoning has been "the main obstacle to a science of cause and effect" (p. 28). In workplace health and safety, the stakes of this obstacle are measured in human suffering. The intervention calculus is not an academic exercise - it is a tool for ensuring that limited resources are directed toward actions that will actually change outcomes, as determined by causal structure rather than spurious correlation.

The Safe Work Australia Code of Practice requires that psychosocial hazard controls be "effective" and "suitable." A BN-based intervention calculus provides a principled, transparent, and auditable method for demonstrating that proposed controls are expected to be effective - and for identifying when they are not.