3
$\begingroup$

My understanding is the distinction between subsampling and the m out of n bootstrap is that subsampling draws without replacement but the m out of n bootstrap does not.

If we are not in a situation with m = n, is there a reason to prefer the m out of n bootstrap? I am struggling to think of an intuitive reason why subsampling would not dominate, but I realize there may be some math justifying m out of n.

$\endgroup$
1
5
$\begingroup$

You are correct.

The bootstrap samples with replacement while subsampling samples without replacement.

Theoretically, when $m \ll n$, it does not matter if you sample with or without replacement (if you sample with replacement and $m \ll n$ the probability of drawing the same observation twice can be made arbitrarily small as $\frac{m}{n} \to 0$).

In practice, the two procedures usually provide very similar results. The key assumption that makes subsampling valid is that $m \to \infty$ and $\frac{m}{n} \to 0$ as $n \to \infty$. (To be precise if $\theta_n$ is your test statistic and $\theta$ is the true value and if $\tau_n (\theta_n - \theta)$ converges to a limit law, then $m \to \infty$, $\frac{m}{n} \to 0$ and $\frac{\tau_m}{\tau_n} \to 0$ have to hold if $n \to \infty$).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.