Causal Pathway

Causal Pathway in Health Technology Assessment and Modelling

Economic evaluations rarely observe a technology's effect on final outcomes directly, so they rely on an assumed causal pathway running from a treatment, test or policy, through changes in the body or in care, to the outcomes that matter to patients, such as survival and quality of life. Each link in that route is a cause of the next, so an effect that travels along it reaches the outcome only if every link holds. Health technology assessment relies on it whenever evidence stops short of the final outcome and a model has to carry the effect along the remaining links. Surrogate endpoints, linked evidence for diagnostic tests and the conceptual models behind economic evaluations are all ways of stating, testing or filling in a causal pathway.

Links, mediators and final outcomes

A pathway is made of variables joined by arrows, running from the intervention or exposure to the final outcome. Any variable between them that carries part of the effect is a mediator. Hernán and Robins give the case of aspirin, which lowers the risk of heart disease because it reduces platelet aggregation, so platelet aggregation is a mediator of that effect.

If aspirin acts only through platelet aggregation, then among people with the same level of platelet aggregation, knowing whether a person took aspirin adds nothing to the prediction of heart disease. Conditioning on the mediator blocks the flow of association along the path. A pathway can also branch, with part of the effect passing through one mediator and part through others or through routes nobody has measured. Mediation analysis splits a total effect into the parts carried by each route.

Pathway variables have to be kept apart from confounders. Hernán and Robins note that the traditional definition of a confounder required that the variable does not lie on a causal pathway between treatment and outcome. A confounder links treatment and outcome through a common cause, and adjusting for it removes bias. A mediator sits on the path itself, and adjusting for it removes part of the effect being estimated, so placing each variable correctly is part of any causal inference about costs or outcomes.

Drawing the pathway as a graph or logic model

Writing the pathway down exposes its assumptions to challenge, and tells the analyst which variables must be measured. Hernán and Robins define a causal directed acyclic graph (DAG) as variables joined by arrows with no directed cycles, in which a missing arrow means no direct causal effect and every common cause of two variables on the graph is also drawn.

For complex interventions, where the pathway runs through people and services as well as biology, Rohwer and colleagues developed two templates for systematic reviews and HTAs, one describing the system in which participants, intervention and context interact, and a process-orientated logic model that displays the processes and causal pathways leading from the intervention to multiple outcomes. A causal loop diagram differs from both because it is built around feedback loops, which a DAG rules out by definition.

The pathway inside an economic model

Every health economic model contains an assumed causal pathway, even when it is not drawn. The ISPOR and SMDM task force reported by Roberts and colleagues separates problem conceptualisation, which turns knowledge of the health care process or decision into a representation of the problem, from model conceptualisation, which represents those parts with a particular analytic method. It recommends an explicit process, such as expert consultation, influence diagrams or concept mapping, so that the model structure reflects current disease knowledge. Tappenden's disease process model sets out the disease-specific events and processes in which the decision problem sits, and his list of issues to consider in model design asks what evidence links intermediate and final outcomes when trials do not report final outcomes.

The NICE manual (PMG36, section 4.6.3) asks for the conceptual modelling behind the model structure to be transparent and justified, and states that prior use of a structure in published models or earlier NICE submissions is not enough. A pathway that leaves out a route by which the technology affects costs or outcomes is a source of structural uncertainty, and parameter sampling will not reveal it.

Surrogate endpoints as a point on the pathway

A surrogate endpoint is a measurement taken partway along the pathway and used in place of the final outcome. Prentice proposed a strict criterion, under which the rate of the true endpoint at any follow-up time must be independent of treatment once the history of the surrogate is known. In pathway terms, the surrogate must carry the whole effect of treatment. The NICE Decision Support Unit's Technical Support Document 20 (TSD 20) notes that Fleming and DeMets described scenarios, depending on how a treatment acts, in which this full mediation fails, and that satisfying Prentice's criteria does not guarantee a causal relationship between treatment effects on the surrogate and on the final outcome.

The Cardiac Arrhythmia Suppression Trial is the standard warning. Patients whose ventricular ectopy after myocardial infarction could be suppressed were randomised to active drug or placebo, to test whether suppressing ectopy reduced sudden death. Among 1,498 patients in the encainide and flecainide comparisons, 43 on active drug and 16 on placebo died of arrhythmia, and both drugs were discontinued because of excess mortality. The surrogate moved in the right direction while the final outcome moved in the wrong one.

Because one study cannot show that the pathway holds for other treatments, TSD 20 favours a meta-analytic approach, in which associations between treatment effects derived from randomisation are average causal effects. Daniels and Hughes, as reported in TSD 20, call a relationship strong when it has a zero intercept and a non-zero slope, so that no effect on the surrogate implies no effect on the final outcome.

How NICE grades the evidence for a pathway

The NICE manual (PMG36, section 4.6.6) adopts three levels of evidence for surrogate relationships from Ciani and colleagues. Level 3 is biological plausibility, level 2 is a consistent association between surrogate and final outcome, usually from epidemiological or observational studies, and level 1 is evidence from randomised controlled trials that the technology's effect on the surrogate corresponds to a commensurate effect on the final outcome.

Section 4.6.7 asks for good evidence that the relative effect on the surrogate predicts the relative effect on the final outcome, preferably from a meta-analysis of level 1 evidence using the bivariate methods in TSD 20. Sections 4.6.8 and 4.6.9 add that validation is specific to the population and technology type, and that carrying a relationship to a different population, class or mechanism of action needs thorough justification. Section 4.6.11 requires the uncertainty in the relationship to be quantified and included in probabilistic sensitivity analysis. In practice, Taylor and Elston found that 4 of the 35 UK HTA reports from 2005 and 2006 that addressed treatment effectiveness with a cost-effectiveness model, about 10 per cent, based that model explicitly on a surrogate outcome, with considerable variation in the strength of evidence and in how transparently the relationship was quantified.

Linked evidence for diagnostic tests

Diagnostic tests rarely change health directly, so their value lies at the end of a longer pathway. Sections 4.6.12 and 4.6.13 of the NICE manual explain that test accuracy statistics are intermediate measures a model can use to predict later health outcomes, and that the benefits of testing generally arise from the treatment or prevention that follows. Section 3.3.2 prefers end-to-end studies for diagnostic technologies and a linked evidence approach when they are insufficient, and section 4.6.14 asks the analyst to specify the links, such as those between diagnosis, treatment and final outcomes, and to review data on each.

Merlin and colleagues, describing the Australian framework for codependent tests and medicines, called this building a chain of argument by linking different types of evidence. They noted that a trial would capture the entire causal pathway, including unexpected and unknown effects, which linked evidence cannot. Evidence on diagnostic accuracy is only the first link.

Worked example with a three-link diagnostic pathway

All numbers in this example are illustrative. A new test for a condition with 20 per cent prevalence has sensitivity 0.90 against 0.75 for the current test, and specificity 0.92 against 0.95. Patients who test positive are treated. Treatment lowers the five-year risk of a serious event from 0.30 to 0.18 in people with the condition, and each event avoided gains 2.5 QALYs. Treating a false positive costs 0.05 QALYs through side effects.

$$\Delta Q = N p \, \Delta Se \, \Delta r \, q - N (1 - p) \, \Delta Sp \, h$$

where $\Delta Q$ is the QALY gain per $N$ people tested, $p$ is prevalence, $\Delta Se$ is the gain in sensitivity, $\Delta r$ is the absolute risk reduction from treatment, $q$ is the QALY gain per event avoided, $\Delta Sp$ is the loss in specificity and $h$ is the QALY loss per false positive treated.

Step 1 (accuracy link). Of 1,000 people tested, 200 have the condition. True positives rise from $TP_{0} = 200 \times 0.75 = 150$ to $TP_{1} = 200 \times 0.90 = 180$, so 30 extra. Among the 800 without it, false positives rise from $FP_{0} = 800 \times 0.05 = 40$ to $FP_{1} = 800 \times 0.08 = 64$, so 24 extra.

Step 2 (treatment link). With $\Delta r = 0.30 - 0.18 = 0.12$, the extra true positives avoid $E = 30 \times 0.12 = 3.6$ events.

Step 3 (outcome link). The gain is $Q_{E} = 3.6 \times 2.5 = 9.0$ QALYs, less $H = 24 \times 0.05 = 1.2$ QALYs of harm, so $\Delta Q = 9.0 - 1.2 = 7.8$ QALYs per 1,000 tested.

Step 4 (costs). The new test costs 30 GBP more per person and treatment 500 GBP per patient, so for the 54 extra positives the extra cost is $\Delta C = 1{,}000 \times 30 + 54 \times 500 = 57{,}000$ GBP, ignoring savings from avoided events. The ICER is $ICER = 57{,}000 / 7.8 = 7{,}308$ GBP per QALY.

Step 5 (a weak link). Extra cases found by a more sensitive test may be milder than those in the treatment trials. If their risk falls by only 0.06, the gain becomes $Q_{E} = 30 \times 0.06 \times 2.5 = 4.5$ QALYs, so $\Delta Q = 4.5 - 1.2 = 3.3$ and $ICER = 57{,}000 / 3.3 = 17{,}273$ GBP per QALY.

The accuracy evidence is unchanged. The result moves because the treatment link came from patients who may differ from those the new test adds, and section 4.6.13 of the NICE manual asks that the populations in accuracy studies be comparable with those in the evaluation of the technology that follows the test.

Boundaries with nearby ideas

A causal pathway describes how an effect is produced. A care pathway or treatment pathway describes the order of services or therapies a patient receives, which need not be causal. An intermediate outcome is a variable on the pathway, and it becomes a surrogate endpoint only when it stands in for the final outcome and the evidence supports that use.

Limitations and common misreadings

A diagram does not test itself. The arrows in a DAG or a disease process model are assumptions, and the analysis inherits any error in them. Merlin and colleagues note that modelling used to explore uncertain links may itself be prone to oversimplification and bias.

1. Higher test accuracy is a health gain in itself. Accuracy is only the first link, and in the worked example the same accuracy gain is worth 7.8 or 3.3 QALYs per 1,000 tested depending on the treatment link.
2. Adjusting for intermediate variables gives a cleaner effect estimate. Conditioning on a mediator removes the part of the effect it carries, so the result answers a different question.

Sources

• Hernán MA, Robins JM. Causal Inference: What If. Boca Raton: Chapman & Hall/CRC; 2020. Online edition of 19 August 2026, chapters 6 and 7. https://miguelhernan.org/whatifbook
• Rohwer A, Pfadenhauer L, Burns J, et al. Logic models help make sense of complexity in systematic reviews and health technology assessments. Journal of Clinical Epidemiology. 2017;83:37-47. https://doi.org/10.1016/j.jclinepi.2016.06.012
• Roberts M, Russell LB, Paltiel AD, et al. Conceptualizing a model: a report of the ISPOR-SMDM Modeling Good Research Practices Task Force-2. Value in Health. 2012;15(6):804-811. https://doi.org/10.1016/j.jval.2012.06.016
• Tappenden P. Conceptual modelling for health economic model development. HEDS Discussion Paper 12/05. University of Sheffield; 2012. https://eprints.whiterose.ac.uk/id/eprint/74464/
• National Institute for Health and Care Excellence. NICE technology appraisal and highly specialised technologies guidance: the manual (PMG36). 2022, last updated 13 October 2026. Sections 3.3 and 4.6. https://www.nice.org.uk/process/pmg36
• Bujkiewicz S, Achana F, Papanikos T, Riley RD, Abrams KR. NICE DSU Technical Support Document 20: multivariate meta-analysis of summary data for combining treatment effects on correlated outcomes and evaluating surrogate endpoints. Version 2. Decision Support Unit; 2019. https://www.sheffield.ac.uk/nice-dsu/tsds
• Prentice RL. Surrogate endpoints in clinical trials: definition and operational criteria. Statistics in Medicine. 1989;8(4):431-440. https://doi.org/10.1002/sim.4780080407
• Ciani O, Buyse M, Drummond M, Rasi G, Saad ED, Taylor RS. Time to review the role of surrogate end points in health policy: state of the art and the way forward. Value in Health. 2017;20(3):487-495. https://doi.org/10.1016/j.jval.2016.10.011
• Echt DS, Liebson PR, Mitchell LB, et al. Mortality and morbidity in patients receiving encainide, flecainide, or placebo: the Cardiac Arrhythmia Suppression Trial. New England Journal of Medicine. 1991;324(12):781-788. https://doi.org/10.1056/NEJM199103213241201
• Taylor RS, Elston J. The use of surrogate outcomes in model-based cost-effectiveness analyses: a survey of UK Health Technology Assessment reports. Health Technology Assessment. 2009;13(8). https://doi.org/10.3310/hta13080
• Merlin T, Farah C, Schubert C, et al. Assessing personalized medicines in Australia: a national framework for reviewing codependent technologies. Medical Decision Making. 2013;33(3):333-342. https://doi.org/10.1177/0272989X12452341