Representing interventions in DAGs has proven confusing and prior to teaching this pathology class I had not given the potential source of that confusion much thought. Thrust into the situation has required some thinking. The earlier post delineated at least two ways model interventions with DAGs. It defined intervention as: Any action taken explicitly to eliminate (or modify) a cause or an effect. One DAG approach is simply to insert a DAG where, within a causal structure, some sort of intervention is warranted and possible. I see this approach as being core to clinical reasoning. We identify our interventions by putting together a causal structure of a patient. Much of that structure is framed for us by the underlying known condition (soon I will present an example of COPD). But a patients particular structure will vary from the frame. This approach is also helpful for designing and interpreting studies of interventions on a macro scale.

The second approach is to 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). This approach is helpful for the same two reasons above as we interpret and respond to what actually happens in a patient (or sample of patients) compared to what we expect to happen in a patient (or sample of patients).

I think an item of confusion is with the first approach described above. Once we enter the “intervention” into the DAG we expect changes in the rest of the DAG. In this example I have taken COPD-4 from the Physical Therapy DAG repository on COPD (here) - and expanded it and added two interventions (inspiratory muscle training and supplemental oxygen).


COPD-4 starts as a causal model with the intention of allowing determination of the underlying cause of DOE (dyspnea on exertion) in a patient with COPD. Two interventions are added to the DAG. Inspiratory muscle training (IMT) which addresses the Low Maximal Inspiratory Pressure (Low MIP), and Supplemental Oxygen with addresses Hypoxia. It is confusing because based on DAG structure it is saying IMT causes Low MIP, so we expect an update of Low MIP which would then update the entire DAG. I agree, this is confusing. For these sorts of intervention DAGS there needs to be another connection between the intervention and the variable it intends to effect. For now we will denote this by adding “INT” to the intervention variable name in these types of DAGs.


Keep in minds that DAGs as models are static, they can be dynamic under two situations. One is if we add further information to the variables and their connections (i.e. probabilities, equations, parameters and /or coefficients) and turn them into computer programs for calculating actual values and probabilities. This is possible but not common in our profession (I hope to contribute to this in the future with much more work). It is also not possible with DAGitty, but may be possible with TETRAD (which I am exploring).

Much more commonly, and without really thinking about it, we make them dynamic in our mind - reasoning through them. From the DAG above we reason that if MIP is low, and if there is an intervention for low MIP (that is IMT), then it might work. But it might not work - after all, there are two causes of Low MIP and neither of them (dynamic hyperinflation (DHI)) is going to change with IMT; nor do we know the extent that each contributes to the low MIP in the first place. IMT works on muscle strength (related to deconditioning), it does not reduce DHI. Therefore, if DHI is the primary reason for Low MIP then IMT may not have much of an effect. The DAG also demonstrates why it is perfectly reasonable to expect DOE to continue despite supplemental oxygen. Hypoxia is one of many causes of DOE. Treating hypoxia directly with oxygen is not as effective as treating DHI.

It should be no surprise for you to know that DHI is treated with interventions such as pursed lip breathing, CPAP, BiPAP and mechanical ventilation (increasing order of severity). These interventions are much more effective at reducing dyspnea than supplemental oxygen as they are more central to the root problem and have the potential to influence many more patho mechanisms.

Here is another example from a student in class (with permission - Cristi Diaz Contreras):


It is a DAG of fibromyalgia - going from etiologies (risk factors) to pathomechanisms (broadly depicted) and some clinical manifestations. Then it adds interventions (blue but not outcomes or exposures so no arrow or “I” embedded in them) that address many of the clinical manifestations. This is a nice overview (like the DAG on COPD). Each of these interventions -effect interactions could be expanded for further insight.

So far in my exploring I have only worked with this first type of intervention DAG. Whether I find these most interesting because I have worked with them, or whether I have worked with them because they are most interesting to me I cannot say. Given this is still in development, I expect to have more to post as I learn more and look forward to this semester’s class helping me out with that process….whether they realize it or not.