# Non-parametric effect size

Effect size, as estimated by Cohen's $\delta$, is:

$d=\frac{\hat x_1 - \hat x_2}{\hat s_{pooled}}$

Since the numerator is an estimate of difference in location, and the denominator is an estimate of the spread, I have always thought it reasonable to replace the top with difference in medians and the bottom with interquartile range:

$m=\frac{Median(X_1) - Median(X_2)}{IQR(X_1\ and\ X_2)}$

But I cannot find any references to justify this. Does this sound reasonable?

The type of application I'm looking for here is the often encounter situation where I'm doing test of location with small data ($n<20$). I compute the t-statistic and generate its null distribution with permutation. This gives me nice p-values. But for numerical estimation of effect magnitude, I thought this $m$ quantity may be helpful in describing the size of the effect magnitude.

• I haven't heard of this either but if you're in a field of study where medians and IQR are nearly universally used there might be little wrong with this approach. Some simulations could easily demonstrate so. But don't call it anything like d. Given that the question is about the reasonableness of the approach you need to add why you do it. – John Nov 2 '14 at 20:02