Why resample by centering vs. the sample estimate, not null parameter value?

Question

In a classic paper on conducting bootstrapped hypothesis tests, Hall & Wilson (1991) present the following guideline, which I am trying to understand:

On first read, this seemed obvious: $\widehat{\theta}^{*}$ are being resampled under $H_0$, so are consistent for $\theta_0$, whereas $\widehat{\theta}$ is consistent for $\theta$. Thus, $\widehat{\theta}^{*} - \widehat{\theta}$ is consistent for $\theta_0 - \theta$, which makes sense. In contrast, $\widehat{\theta}^{*} - \theta_0$ is consistent for $\theta_0 - \theta_0 = 0$, which is not useful.

But then I realized, on rereading the paragraph above the First Guideline and a subsequent example, that Hall & Wilson are computing $\widehat{\theta}^{*}$ by resampling directly from the sample, rather than under the null. In that case, I have no idea why they recommend resampling $\widehat{\theta}^{*} - \widehat{\theta}$, which is consistent for 0!

References

Hall, P., & Wilson, S. R. (1991). Two guidelines for bootstrap hypothesis testing. Biometrics, 757-762.

Answer 1

The idea is that irrespective of the choice of $\theta_0$ the bootstrap distribution of $\widehat\theta^*$ - $\widehat\theta$ resembles the sampling distribution of $\widehat\theta$ - $\theta$. If you instead use $\theta_0$ you will be pulling the estimate closer to $\theta_0$ and thus lose power under the alternative. I think this is essentially what Hall and Wilson are saying but possibly a little simpler and hopefully clearer for you.