Median of the shifts vs. shift of the medians, Part 2: Gaussian efficiency

In the previous post, we discussed the difference between shifts of the medians and the Hodges-Lehmann location shift estimator. In this post, we conduct a simple numerical simulation to evaluate the Gaussian efficiency of these two estimators.


We consider two samples of equal size \(n\): \(x = \{ x_1, x_2, \ldots, x_n \}\), \(y = \{ y_1, y_2, \ldots, y_n \}\). We define the shifts of the medians as

\[\newcommand{\DSM}{\Delta_{\operatorname{SM}}} \DSM = \operatorname{median}(y) - \operatorname{median}(x). \]

and the Hodges-Lehmann location shift estimator as

\[\newcommand{\DHL}{\Delta_{\operatorname{HL}}} \DHL = \operatorname{median}(y_j - x_i). \]

We also consider the classic estimator that estimates the difference of means:

\[\newcommand{\Dbase}{\Delta_{\operatorname{0}}} \Dbase = \operatorname{mean}(y) - \operatorname{mean}(x). \]

The Gaussian efficiency of \(\DSM\) and \(DHL\) can be defined as follows:

\[e(\DSM) = \frac{\mathbb{V}[\Dbase]}{\mathbb{V}[\DSM]},\quad e(\DHL) = \frac{\mathbb{V}[\Dbase]}{\mathbb{V}[\DHL]}. \]

Numerical simulations

We conduct the following simulation:

  • Enumerate the sample size \(n\) from \(3\) to \(100\).
  • For each \(n\), generate \(100\,000\) pairs of random samples from \(\mathcal{N}(0, 1)\).
  • For each pair of samples, estimate the shift between them using \(\Dbase\), \(\DSM\), and \(\DHL\).
  • Calculate the Gaussian efficiency of \(\DSM\) and \(DHL\) using the above equations.

Here are the results:

As we can see, the Hodges-Lehmann location shift estimator is much more efficient than the shift of the medians.


  • [Hodges1963]
    Hodges, J. L., and E. L. Lehmann. 1963. Estimates of location based on rank tests. The Annals of Mathematical Statistics 34 (2):598–611.

The source code of this post and all the relevant files are available on GitHub.