# Nonparametric bootstrap: Assumptions and number of bootstrap samples?

1. Jul 20, 2014

I've been figuring out the use of the nonparametric bootstrap and if I understand correctly, this is the procedure:

1. Take an original sample, a vector x = (x1, ..., xn)

2. Generate k vectors, each called a 'bootstrap sample', of the same length as x by random sampling (with replacement) of the original vector, x, e.g. I have b1 = (x3, xn, x1, ..., x13), b2 = (x2, x5, ..., xn) etc.

3. Now I calculate any statistic $\hat{\theta} = f(\bf{x})$ on each bootstrap sample and my bootstrapped statistic is the mean of the statistic across the distribution and my confidence intervals on that bootstrapped statistic can be found using the inverse cdf of a normal distribution.

If everything is correct above, I have two questions:

i. How do I determine the number of bootstrap samples to take, k? Is there a principled way to determine this? Without this, I would just have to keep repeating the same procedure with increasing k until there's some kind of convergence on the mean $\bar{\hat{\theta}}$? But this seems computationally taxing.

ii. What assumptions must be correct for this procedure to work? I'm guessing that $\hat{\theta}$ must have finite variance? What else?

2. Jul 21, 2014

### Stephen Tashi

I think your description of the procedure is correct.

That's a good question. A major task is define what it means to say it "works". Do you have a sophisticated understanding of the meaning of a confidence interval? In particular, do you understand that the usual sort of confidence interval does NOT let you make statements about a population parameter being in a specific numerical interval. (For example, you can't conclude things like "There is a 90% chance that the population mean is in the interval [ 2.3 - 0.62, 2.3 + 0.62]. )

If you mean $\bar{\hat{\theta}}$ to be an estimate of a parameter of the distribution from which the sample is taken, I don't think there is any guarantee that the results of the bootstrap "converge" to the value of that parameter as you generate more boostrap samples.

3. Jul 24, 2014

1. Unfortunately no, I don't have a very sophisticated understanding of the meaning of a confidence interval (I wouldn't be able to write a philosophical debate about it). But I do have a basic grasp of the pitfalls. What's one school of thought I could practice on "what works" without going too deep into the foundations?

2. Hm, that's problematic. How would I know what's a good bootstrap approximation if it doesn't converge?

I'm sorry if I sound like I'm just looking for more clues but I don't have a strong intuition on how to attack this.

4. Jul 25, 2014

### Stephen Tashi

A confidence interval isn't the same as a "credible interval" (http://en.wikipedia.org/wiki/Credible_interval).

Suppose we are trying to estimate a property of a population by bootstrapping. We have a large batch of samples and from it we repeatedly select smaller bathches. How close our estimate is to the actual value of the populaton parameter depends on 1)How well the large batch of samples matches the population distribution and 2) How we estimate the parameter from the bootstrap samples

Once you have the large batch of samples, you can usually produce smaller (frequentist) confidence intervals by doing more bootstrap sampling in 2). More boostrap samples can't improve the mis-estimation that may be introduced in 1). The total confidence interval size depends on both 1) and 2).

It would be easier to discuss bootstrapping if we discuss estimating a specific thing.