This summer students in the UML DPT pathology class are making a DAG per week, related to the materials of the chapter(s) we are covering any given week. So far they are doing an outstanding job (1 DAG submitted / student). This post is to provide additional information to them (and anyone else interested) about particular pragmatic goals of a DAG based on causal structures of pathological processes and clinical reasoning. The founding assumption is that all clinical reasoning is, at it’s core, reasoning about causal structures. The particular pragmatic goal would influence what you include in the DAG from the set of possible causal associations, being basically: {etiologies, risk factors, pathogenesis, pathomechanisms, clinical manifestations, interventions}. There is no need to add to that set “examination” or “evaluation” as these are part of the DAG itself. Meaning, if it is a variable in the DAG it is a piece of information that can be examined (in some way, using some set of sensory perception, even extended sensory perception with imaging, or a stethescope, or an ECG or EMG, then by being in the DAG it is part of examination), and the process of reasoning through the DAG is essentially what evaluation is (that is achieving the initial pragmatic goal).

First some definitions:

Etiologies: Causes of disease / condition (definite or possible causes)dagitty-model

Risk factors: Risks that increase the probability of a disease / condition (i.e. exacerbating factors)


It is certainly a judgement call as to why “Risk Factor” is a risk factor and not an Etiology and vice versa.

A formal definition might be:

P(DiseaseEtiology & Risk Factor) >P(DiseaseEtiology)
and: P(DiseaseEtiology) > P(DiseaseRisk Factor)
A really strict definition could even require:  P(DiseaseRisk Factor) = 0 (in other words, just the risk factor alone does not cause the disease, but then we need to modify the DAG in some way to denote that there is no causal association between Risk Factor and Disease and I am not willing to do that, so I am not a fan of the “strict definition” as proposed. Keep in mind that much time is spent attempting to identify which “risk factors” are the most important, which are true “etiologies”.

Pathogenesis: Mechanistic causal pathway (definite or possible) from etiologies to pathomechanisms of the disease / condition


Pathomechanisms: Mechanistic causal pathways (definite or possible) of the disease / condition that emerge to clinical manifestations


Clinical manifestations: Consequences of the disease / condition, signs, symptoms, structural and functional effects. Note that once I added Clinical Manifestations to the DAG I dropped the variable “Disease” - now the “Disease” is essentially embedded in the pathomechanisms and clinical manifestations (or simply just the clinical manifestations - depending on the disease).


A disease or condition is simply a set of pathomechanisms and clinical manifestations (consequences) - so once we add these both to the DAG, or at least once we add the clinical manifestations, there is no need to continue to have the disease itself. UNLESS, the goal of the DAG is differential diagnosis and you want to track back to a particular diagnosis from amongst a wider set of diseases that have shared clinical manifestations and / or pathomechanisms.

Interventions: Any action taken explicitly to eliminate (or modify) a cause or an effect. The issue is how to represent in a DAG, particularly when the aim of the intervention is to counter assumptions embedded within the graphical causal structure.


Here there are 4 interventions - each attempting to eliminate or modify a cause of effect - at different steps in the causal process. Intervention 1 could be considered prevention - attempting to remove or modify the etiology of a pathological process. We consider this an intervention is the pathogenesis process has started and are not sure whether removing the etiology will stop the subsequent causal chain. But if you have Lyme disease, and you still have the tick in you, your are going to get that tick out even if it will not remove the Lyme disease. Similarly, we recommend people stop smoking prior to getting lung disease, but also after they are diagnosed with lung disease since there is a continuous pathogenesis process associated with continued exposure to the etiology.

A DAG of an intervention could also start with the intervention itself and then model the causal structure of the intervention itself (as opposed to adding intervention to a causal model of the pathological process).


Pragmatic goals related to clinical reasoning include (and I am always open to ideas and opinions about other possibilities):

1.Differential diagnosis – what is the cause of this problem? There are at least two types of such DAGs - based on the two types of different diagnosis I have previous described here.  If reasoning to possible disease states that might be the cause of what you are seeing - in particular to rule out causes outside of your scope of practice (screening and referral) the DAG is going to include information about diseases and clinical manifestations with the goal of helping with an abductive inference to the best explanation:


In this DAG we see that the clinical manifestation set {A, B, D} seems unique to Disease 1. IN fact, since clinical manifestation A is, according to this very simplistic DAG, only associated with disease 1, so that is easy - - perhaps, what if clinical manifestation 1 is data you do not have access to at this point? A blood test, or imaging? Then to rule out disease 1 you may have to rule out other causes of Clinical Manifestation B(that is disease 2 and 3) for example.

The second type is when we know the disease - or at least enough about the disease status - that we are attempting to determine the particular pathomechanism (associated with the patient’s disease status) that is the predominantly responsible for their current limiting clinical manifestations. For example, of all the things that can go wrong with COPD, which of them are the primary issue causing this particular patient’s activity limitation? Keeping with general DAGs here - it is an attempt to discern the particular pathomechanism.


Here you can see I am getting a bit lazy - used the same DAG from above and just replaced Disease (which would be a diagnosis) with “Pathomechanism”. Each disease has a variety of pathomechanisms at play, and variation in those mechanisms results in variations in clinical manifestations, and can be intervened on in different ways.

2.Prognosis – what will happen in the future? Prognosis DAGs will be a topic of an upcoming post on “predictive analytics” - after all, predictive analytics are simply an approach to use data to attempt to predict the future, or at least the probability of certain future events relative to other events. Underlying that data, and the analysis, is a set of assumptions based on a causal structure.

3.Intervention - Can we change the future by doing something? (includes contra-indications)

4.Prevention – Can we change the future by not doing something, or changing how we do something? (includes contra-indications)

Intervention and prevention have been dealt with in a cursory way earlier in this post in the definition of intervention. After I post on predictive analytics and prognostic DAGs I will turn my attention towards a better understanding and explanation of intervention / prevention DAGs. I expect it will take me about a week for the predictive analytics (prognosis) and then another week for intervention. So, for now either use what is above to do you best, or stick with a differential diagnosis based DAG for your journals (if you are taking the class).