Now that the definition of an adjustment set has been presented we can turn back to the model presented earlier. The model is presented in a way that you can make changes to it at DAGitty’s model repository here. For the ease of reading the blog I will also provide a picture.

dagitty-model

As a reminder, on the DAGitty repository you can manipulate this graph. You can move things around (does not change the causal structure), you can add variables and connect them, delete variables, delete connections (all of which changes the causal structure). You can also change focus by identifying different or additional “exposure” variables and “outcome” variables. The selection of an exposure variable (above it is Functional Status) and outcome variable (above HRQOL) allows DAGitty to identify, based on the causal structure, which variables would need to be adjusted in order to make a causal inductive inference from the exposure to the outcome. Adjusting basically means knowing about it or somehow controlling for it. The variables that must be adjusted in order for an unbiased causal inductive inference are the adjustment set.

What is an unbiased causal inductive inference? It is a causal inductive inference that has no alternative explanations. A biased causal inductive inference is one that is not true or has potential alternative explanations.

In the above graphic the adjustment set includes both “Social” and “Symptom Status”. These variables all directly impact HRQOL and Functional Status. If we did not adjust for them we would have no way of knowing whether the observed relationship between Functional Status and HRQOL was due to Social or Symptom Status. If we adjust for Social and Symptom Status we now know about Health Perception since Health Perception comes from Social and Symptom Status. Mind you - the value of these adjustment sets are entirely predicated on the validity of the underlying causal structure.

If you open the link to the DAGitty model you can scroll over the Social and Symptom Status variables and press your “a” key on your keyboard to “adjust” for those variables and see how the graphic changes. When you do this all paths change from red to black (unbiased). If not already mentioned - a red path is a biasing pathway; and it is biasing our ability to make a causal inductive inference from the set exposure to the set outcome variable (in this example from the Functional Status to HRQOL variables).

It is now important to point out that these adjustment sets are for inductive inferences. They are useful when considering how unstructured observations of particulars may be biased when attempting to generate universals (i.e. considering the evidence). They are very useful for planning structured observations (research) of particulars in order to generate unbiased inductive inferences to universal statements of what is known (generating the evidence). Whether considering or generating evidence, evidence when unbiased can lead to expansion of our knowledge. That knowledge (causal models) then can be used to consider or generate more evidence for the generation of more knowledge (expanded causal models). In other words, adjustment sets are a really important aspect of the left side of the KBP graphic (below) discussed in this post.

KBP2

If these adjustment sets are for causal inductive inferences what about deduction and abduction inferences? That is what I will work on next.