high

quoting a comment on Caplan and Waters (1999) Verbal working memory and sentence comprehension. by Akira Miyake, Michael J. Emerson, and Naomi P. Friedman

...

Prevalent yet nonoptimal designs and analyses, namely, the use
of highly discrete measures of individual differences (i.e., reading
span scores such as 2.5 and 4.0) and the creation of arbitrary span
groups (i.e., high-span and low-span groups), are the major causes
of low statistical power in experiments cited by C&W. When compared
to more continuous ways of scoring working memory span
tasks (e.g., total number of words recalled), discrete span measures
have lower power because they reduce the variance by not capturing
subtle differences that may exist among individuals with the
same span score.

...


In fact, dichotomizing a continuous variable can reduce
the variance of that variable by 20–67%, with a corresponding loss
of power equivalent to throwing out one- to two-thirds of the sample
(Cohen 1983).