This post will demonstrate how the use of a DAG and an adjustment set identifies an opportunity to have done something differently in a recent study. A student recently completed a review of an article for research methods. She did an outstanding job and I had to read the article. The paper is: Can sit-to-stand lower limb muscle power predict fall status? By fall status they mean whether someone has fallen in the past - not whether they will fall in the future, why they call it prediction and not classification is beyond me or the scope of this blog right now.

The authors draw the following conclusion: “Our study clearly indicated that SP (stabilization phase) duration could be a predictor of past falling events, but its power to predict future falls requires further validation.”

But I shall not digress into the many issues packed within that sentence. Today I want to simply focus on one aspect of the underlying association they are attempting to demonstrate, that between LE power (particularly reduced LE power) and Falls, and that if the design considered the DAG of even a small portion of the causal structure it could have been more sufficiently adjusted.

Before getting to the opportunity that a DAG and a priori adjustment that would have led to less bias I want to point out one interesting thing. The student doing the review astutely points out that the true independent variable in this study is “fall status” as they recruit people based on fall status (over 65 fallers, over 65 non fallers, or young). Fallers had fallen within the past 12 months, non fallers had not fallen within the past 12 months. The true dependent variables include a large number of variables collected from force plates during a well designed protocol of sit to stand, the one they were mostly interested in was lower extremity (LE) power but based on their results they quickly change to speed of movement and duration to stabilize. Recall, power is a function of work and time, work is a function of force and distance. Power = (force x distance) x velocity. With sit to stand, velocity is the speed of ascending and stabilizing in the upright position so speed is a fine variable since velocity is in one direction (up).

It is important to point out that while the above variable designation is based on Fall Status -> LE Power, that is not the implicit causal association that the authors are interested in if they are interested in prediction of future events, they are (based on reading the paper) interested in, decreased LE Power -> Risk of Falls. After all, if the causal direction is Fall Status -> LE Power then LE power cannot be used clinically to predict future falls. To predict future falls based on LE power we have to hold the justified true belief (that is it must be true) that LE power is causal of future falls. But in this study the only temporal relationship is one where Fall Status precedes measurement of LE Power, so if any causal association is to be drawn from the association in this study it is that having a history of falling increases risk of decreased LE power.

Now back to the concept of an adjustment set and how it relates to this study. The study found equal normalized strength between the faller and non faller groups, but lower power in the faller group as compared to the non faller group due to the difference in speed (faller group took, on average, more time to rise and stabilize during the sit to stand). They then conclude that stabilization duration is related to fall status, and may account for the relationship between LE power and balance. However, no independent test of balance was conducted so we are left to wonder - as the authors did - whether balance is the important modifier (since unmeasured it is a biasing confounder). Let’s look at the DAG:

dagitty-model

The underlying study purpose was to design a system to collect data and then compare “with clinical balance scales in an attempt to identify fall status predictors.”  The clinical balance scale used was the sit to stand (five times sit to stand test). This is not an independent measure of balance from the other parameters measured in the study that are based off the sit to stand.  The relationship between LE power and fall risk (based on fall status) as represented in the above DAG specifically identifies decreased muscle force to decreased LE strength to decreased LE power to fall risk (assumed causal chain). Since power is both work and velocity then speed of movement is identified and adjusted for in the study (measured). Balance is related to speed of movement (including stabilizing from movement) and independently related to fall risk. By not measuring balance independent of speed of movement (that is adjustment for balance), balance becomes a confounder and without information about it we cannot know for sure what the impact of speed of movement or dec LE power is on fall risk. The authors simply point out that SP may be associated with balance but they cannot say more because balance was not assessed.

With the above DAG in existence (online here) it can be built on further. It can be used by researchers planning to do another study, modified when additional studies are reviewed, debated and discussed. All of the assumptions about causal associations become clear for further study, debate and discussion; and for the clinical practice side of KBP, for considering what to test clinically (abduction) and how to intervene (deduction). On the DAGitty site where the model is posted you can see the adjustment set (upper right), the DAGitty code (lower right) which can be cut and pasted into another DAGitty code window to make your own version for modification and saving if interested. The DAGitty repository is a nice way to share models that have been published, but I am working on another approach to organizing models for use by a broader community of physical therapy researchers using DAGitty for their research that allows searching and browsing of models by topic area, and then going back to DAGitty for development and modification.

As I read more research methods mid terms and come across more studies I will share when I can examples of DAGs and adjustment sets and how their use could be helpful moving forward. After all - there is a reason I identified DAGs and adjustment sets when asked to provide some posts on moving forward.