Cohen U1, U2, and U3
Introduction
Cohen (1988, p. 23) provided three measures that relate to Cohen's d.
- \(U_1\), is (supposedly) the proportion of non-overlap between distributions
- \(U_2\), is (supposedly) the proportion of overlap between distributions
- \(U_3\), is the proportion of one group's scores below the mean of another group
\(U_1\) and \(U_2\) are probably the least used of these three, since most likeley Cohen was wrong in his calculation (Grice & Barrett, 2014). A nice interactive visualisation of the relation between Cohen \(U_3\) and the Common Language Effect size, can be found at rpsychologist. It actually correct's Cohen U.
Obtaining the Measure
with Python
Jupyter Notebook: ES - Cohen U (P).ipynb
with stikpetP
To Be Made
without stikpetP
To Be Made
Formulas
The following formulas are used (Cohen, 1988, p. 23):
\(U_3 = \Phi\left(d\right)\)
\(U_2 = \Phi\left(\frac{d}{2}\right)\)
\(U_1 = \Phi\left(\frac{2\times U_2 - 1}{U_2}\right)\)
Symbols used:
- \(d\), Cohen's d value
- \(\Phi\left(\dots\right)\) the cumulative density function of the standard normal distribution
Interpretation
As mentioned in the introduction each of the U's has a different meaning. As for rules-of-thumb, we can convert each of the U's back to Cohen d. After the conversion, any of the rules-of-thumb from Cohen d could be used.
Click here to see the formulas for the conversion
\(d = \Phi^{-1}\left(U_3\right)\)
\(d = 2\times\Phi^{-1}\left(U_2\right)\)
\(d = 2\times\Phi^{-1}\left(\frac{1}{2 - U_1}\right)\)
Google adds