Super interesting new paper that tackles an important problem: How do we know if we're measuring variables we care about at the "right" times?



Authors = @Oisin_Ryan_ and Ellen Hamaker (sorry if I missed their twitter!) https://psyarxiv.com/2ambn/ 

Most lagged network/SEM models can give you super different answers about which variables are "important" depending on how often you measure them

This paper outlines how you can get nearly the opposite conclusions if you measure certain variables every half hour vs. every hour

The authors argue these models are way too dependent on the sampling rate (ie how often we ask people to answer our questions) for drawing practical conclusions about which variables are actually more or less important

They propose continuous-time models as a potential solution, as these models inherently assume/explicitly model that all variables are potentially influencing each other... well, continuously across time

Authors: "Current anxiety levels in part determine feelings of stress a second, a minute, ten minutes and an hour from now to different degrees, meaning that there are likely different lagged relationships between those variables over a whole range of different time-intervals"

Through lots of math (primarily calculus?) they are able to define new centrality measures that are determined over a specific time-interval

These statistics could give us a sense of which variables we might want to target depending on the time course we care about most

These statistics can be "direct" (e.g., > anxiety over a specific time interval makes you uniquely more physically uncomfortable) or "total" (e.g., the implications of > anxiety on all variables in the system over a specific time interval -> less physically uncomfortable)

The authors make the argument that we should probably consider the "total" effects as more important, especially if we're trying to identify intervention targets

My conclusions:
- Using models that take time into account more realistically could help us make better decisions on how to intervene
- These kinds of models force us to think about *when* and for *how long* we want our interventions to affect variables we care about