Difference between Sampling a population Vs Bootstrapping

I am finding it difficult to understand the concept of Bootstrapping in statistics . I know what sampling is , that is , taking a 'sample_size' number of observations from a population to estimate some of that population statistics like mean , SD etc . I thought bootstrapping was doing that same process of sampling multiple times , but it doesn't look like that's a proper way to put it . Some sources say bootstrapping takes a number of samples with size equal to the original dataset while some others say it takes samples of desired sample size from within a bigger sample of a dataset. All of these definitions got me confused .

Could someone please explain the difference between the two in simple and intuitive manner ? i.e , what exactly is each one of them doing ?

• Could you quote where and what exactly did you read? In general, bootstrap takes sample with replacement from the data of size the same as the size of the data.
– Tim
Jul 5 '20 at 19:56
• One obtains the usual sample by sampling from the population. A bootstrapping sample is different because one samples with replacement from the sample itself. But, Efron showed that the relationship between the usual sample and the population is the same as the relationship between the bootstrap sample and the sample under certain conditions. ( mainly that the statistic one is constructing converges, I think ). This obviously doesn't do the concept justice but John Fox has a decent intro to bootstrapping. Google for "John Fox Bootstrapping" and it will come up. Jul 5 '20 at 20:41
• Tim and mlofton , I believe you both have different say regarding whether bootstrapping is done from the original dataset or a sample of the dataset . Are both possible ? Jul 6 '20 at 6:42

When you take a sample from a population, you are gathering information about the population, which you might use for making a confidence interval or for testing a hypothesis about population parameters (perhaps the population mean $$\mu).$$

When you 're-sample' as in bootstrapping, you are analyzing data already taken from a population. Re-sampling does not provide any new information about the population. (But it might help you better to understand the data you already have.)

You refer to two kinds of 're-sampling' that are in common use.

• In making a nonparametric bootstrap confidence interval, based on $$n$$ existing observations, you might take a large number $$B$$ of re-samples from your data. You would sample with replacement and re-samples would be of size $$n.$$ Nonparametric bootstrapping is often used when you do not know the distribution 'family' of the population. (In particular, you would rarely use a nonparametric bootstrap for data known to have been sampled from a normal or an exponential population.)
• In making a parametric bootstrap, you typically know the distribution family of the population, but perhaps not the values of particular parameters (such as $$\mu$$ or $$\sigma.)$$. In that case, you use the data to estimate parameters(s) and then use bootstrapping to get a confidence interval to go with each estimate. Then you use the estimated parameter value(s) to simulate a re-sample of size $$n$$ from the population. So you are not re-sampling directly from the data but from a population suggested by the data.

Nonparametric bootstrap CI for population mean. Suppose I have a vector y that contains $$n$$ observations from a population of unknown distribution. I want to make a 95% nonparametric bootstrap confidence interval for the population mean $$\mu.$$ Here are summary statistics and a histogram:

summary(y); sd(y)
Min. 1st Qu.  Median    Mean 3rd Qu.    Max.
5.83   11.85   16.16   16.64   19.27   53.99
 7.774901   # sample SD
hist(y, prob=T, col="skyblue2");  rug(y) Here is R code to make a simple quantile bootsrap CI for $$\mu.$$ I take a large number $$B = 3000$$ re-samples of size $$n = 50$$ from y and find the average $$\bar X^*$$ of each re-sample [denoted as a.re in the R code.].

set.seed(2020)
a.re = replicate(3000, mean(sample(y, 50, rep=T)))
quantile(a.re, c(.025, .975))
2.5%    97.5%
14.65229 18.95220

So I could say that a 95% nonparametric bootstrap CI for $$\mu$$ is $$(14.7, 19.0).$$ The interval contains the sample mean $$\bar Y = 16.64,$$ but $$\bar Y$$ is not necessarily at the center of the CI. [Confession: Because I can see form the histogram that the sample is skewed, it might have been better to do a bias-corrected CI, but I'm trying to illustrate re-sampling, but get into nuances of various types of bootstrap CIs.]

Below is a histogram of the re-sampled averages a with vertical lines indicating the bootstrap CI for $$\mu.$$

hist(a.re, prob=T, col="wheat")
abline(v=q, col="blue") Parametric bootstrap CI. Now suppose I know that that the sample y came from a gamma distribution with shape parameter $$\alpha = 5$$ and unknown rate parameter $$\lambda.$$ A reasonable estimator of $$\lambda$$ is $$\hat \lambda = \alpha/\bar Y = 5/16.64 = 0.30.$$

Now I simulate a large number $$B = 3000$$ samples of size $$n = 50$$ from the distribution $$\mathsf{Gamma}(\alpha=5, \lambda = .3).$$ Then I find $$\hat\lambda^*$$ from each sample. At the end, I can use $$(0.27, 0.34)$$ as a 95% parametric bootstrap CI for $$\lambda.$$

set.seed(2020)
lam.re = replicate(3000, 5/mean(rgamma(50, 5, .3)))
q = quantile(lam.re, c(.025,.975));  q
2.5%     97.5%
0.2668468 0.3416872

Here is a histogram of the bootstrap distribution of lam.re along with vertical bars showing the 95% parametric bootstrap CI for $$\lambda.$$ [There are ways to use the gamma distribution to make a CI for $$\lambda$$ without bootstrapping. But in this case the bootstrap method works very well.]

hist(lam.re, prob=T, col="wheat")
abline(v = q, col="blue") Remark. For both of these bootstrap CIs I did re-sampling. But in these procedures I sample no additional data from the population.

Note: In this case the 'population' was R's function to generate gamma data. Data y for this demo were sampled as follows;

set.seed(2020)
y = round(rgamma(50, 5, .3), 2)