Read on Twitter
(1/n) Friday 
:
Are you normalizing your biomechanical data to your participant's bodyweight? Or some other variable? You probably shouldn't be (emphasis on probably; always depends on research Q)
Why? Come with me down my latest
and let me know what you think.
: Are you normalizing your biomechanical data to your participant's bodyweight? Or some other variable? You probably shouldn't be (emphasis on probably; always depends on research Q)
Why? Come with me down my latest
and let me know what you think.
(2/n) When you divide some measure by another measure you create a ratio. This gives us things like varus torque PER UNIT body weight*height or ground reaction force PER UNIT body weight. You get the idea...
(3/n) If u r using a ratio to examine associations btwn the ratio's numerator and an outcome WHILE CONTROLLING FOR the ratio's denominator, you probs shouldn't do that. This is the most common app of ratios I see in biomech so that's where I'll focus for the rest of this thread.
(4/n) Sticking a covariate (thing you're controlling for) in the denominator of a fraction makes two main assumptions...
(5/n) Assumption 1: the two measures you are ratio-ing are linearly related (i.e., ratios cannot handle non-linear relationships). If they aren't, then ratios alter the relationships between your variables of interest. See Figure 2 from Douglas Curran-Everett's 2013 paper 
(6/n) Assumption 2 (also depicted in Everett's figure): Not only must your measures be linearly related, but you should also expect that, if either measure is 0, then the other measure is also zero (i.e., the best fit line of their scatterplot passes through the origin).
(7/n) If those assumptions weren't scary enough, think of ratios as a specific form of interaction in your regression model. @TenanATC said it best in his short thread below. It was his work on acute:chronic workload that started me on this journey https://twitter.com/TenanATC/status/1270204931589443586
(8/n) Does an "inverse interaction" make theoretical sense? Heck if I know, that's for you to decide for your specific research question(s). But you should be confident that it does before using a ratio.
(9/n) What are the consequences? How much harm can ratios cause? Well, if your two measures are normally distributed, then their ratio CANNOT be normal (cite math). This affects all statistical tests that assume normality (i.e., probably the test you're using)
(10/n) Additionally, ratios sometimes induce spurious (i.e., fake/not real/meaningless) correlations btwn the ratio and other variables in your model.
(11/n) Lastly, and IMO most importantly, ratios are more difficult to interpret and relay to non-biomechanists. Ever try explaining a 70 Newton*meter varus torque to a coach or player? Now try explaining what "percent bodyweight*height" means.
(12/n) Interested in reading further into ratios and normalized data? Here are two good places to start: Curran-Everett (2013) and Allison (1995), linked below. Thanks for reading.
https://europepmc.org/article/med/8574275
https://journals.physiology.org/doi/full/10.1152/advan.00053.2013
https://europepmc.org/article/med/8574275
https://journals.physiology.org/doi/full/10.1152/advan.00053.2013
