Graphical Causal Models

The primary explanation for why a knowledge based practice is different from an evidence based practice was presented early on and claims that despite a seeming transitivity:

if evidence, then knowledge

if knowledge, then practice

therefore if evidence, then practice

There are problems with the above deduction. The more I have thought about it the more I have come to understand that the problem with the above is equivocation (the use of ambiguous language to conceal the truth or to avoid committing oneself (my mac’s dictionary)). The ambiguous language being used is knowledge. There is knowledge generated by evidence, and then there is knowledge used in practice. These types of “knowledge” are bridged, but they are bridged with a highly rationale process of model building using causal inferences. The bridge is not simply built by teaching clinicians about statistics and research design, or how to read a systematic review.

Again, the subsequent models - generated by a rationale process - are the bridge.

On one side of the bridge are empirical observations being turned into universals. This is knowledge. But not ready for clinical practice.

On the other side we need to go from causal models to particulars to particular decisions (what to do in a particular clinical case, hence practice).

A knowledge based practice makes the explicit claim that between each side of this process (induction on one side; deduction and abduction on the other side) is the rationale process of causal model building. Here is an attempt to graphically represent this idea:


In this figure we see the process of generating universals through empirical evidence and induction. This knowledge is put into causal models (the bridge). Causal models are then used in the interpretation of particulars which are what is faced in clinical practice (each patient is a particular instance of the population). There are problems with applying sample statistics to a single case. With the above representation we do not have to apply sample statistics to a particular. We have used sample statistics to help develop our causal models. We figure out what to do in practice by inference of cause from particular observed effects (abduction) to figure out what the problem is, and then knowing the cause we practice of inference of hoped for effects by manipulation of causes (deduction) for interventions (actions).  Knowledge for this process are the causal models which are far more complicated structures than the universals generated from the inductive process. So, with a knowledge based practice, knowledge refers to all parts of this process, it is all knowledge, but that means we are adding the causal model knowledge.

To close this loop a bit, with more complexity to the above graphic we see that particulars, practice, universals and causal models all impact the evidence generation process. They provide the structure which we use to make sense of observations (data) as it is collected, even to organize our collection of data (methods).


How is the causal model knowledge different from the knowledge derived from empirical evidence, the universals? It is far more connected and thus to complex for any study to provide direct empirical evidence to support, or at least direct empirical evidence to support decisions about a particular. The name of this post is “Graphical Causal Models.” In upcoming posts I will be starting to go through the process of demonstrating the usefulness of graphical causal models for generating knowledge for practice. When discussing the case for cause (Part 2) and about models, I mentioned the importance of Judea Pearl’s work to KBP. Here, the use of graphical causal models (particular Directed Acyclic Graphs or “DAGs”) is directly attributed to a study of Pearl’s contribution to causation. Such an approach is increasingly common for causal inference in epidemiology and the social sciences. So the left side of the graphic is quite well-developed. KBP faces the challenge of figuring out the use and benefits of graphical causal models for the right side of the above graphic (from causal models to particulars to practice decisions).

I see three major initiatives coming from this framework.

  1. Articulating general principles and particular examples of the interaction between universals and graphical causal models (left side of the graphic)

  2. Articulating general principles and particular examples of the interaction between graphical causal models and particulars (right side of the graphic)

  3. Articulating the consequences of the above for generating clinical practice guidelines and practice recommendations; clinical education; clinical research and research agendas (entire graphic)

These are bold and will most likely take the rest of my career. I am actively seeking collaborators that would like to start meeting to discuss being involved in the above process. If interested, just let me know:

Leave a Comment