3
$\begingroup$

Suppose you only have these information from a sample data: $X_i$ and $w_i$, $i=1,...,N$, where $w_i$'s are the respective sampling weights(not integers).

Is it possible to obtain a valid bootstrap estimate of say, variance of $X_i$?

I understand that I might have to readjust the sampling weights(for weighted data) for each bootstrap replicate to obtain a valid bootstrap estimate. But considering if I only have the above information, are there other ways to do this?

Improve this question
$\endgroup$
4
  • $\begingroup$ There are several different meanings of "sampling weights"; for instance, Stata natively supports three different ones. Could you therefore explain what your weights mean? $\endgroup$
    – whuber ♦
    Commented Feb 27, 2014 at 17:05
  • $\begingroup$ hi @whuber, in Stata it would be "pweights" $\endgroup$
    – stats_newb
    Commented Feb 28, 2014 at 1:50
  • $\begingroup$ My recollection is that these are probability weights: that is, you selected cases independently and randomly but with different chances of inclusion in the sample. Is that correct? $\endgroup$
    – whuber ♦
    Commented Feb 28, 2014 at 14:40
  • $\begingroup$ Hi @whuber, these are weights that are calculated (with post-stratification applied) when sampling population and population of target inference are disparate. $\endgroup$
    – stats_newb
    Commented Mar 3, 2014 at 8:24

1 Answer 1

Reset to default
4
$\begingroup$

Short answer: No.

Marginally longer answer: A single vector of weights tells you nothing about:

  1. Stratification
  2. Clustering
  3. Calibration variables

A proper bootstrap scheme would involve the following (I can refer to my paper on this).

  1. Take a bootstrap sample (with replacement) of clusters within each stratum, independently across strata.
  2. Re-calibrate the sample to the same population totals that were used for the main weights.

Since you don't have strata and clusters, you cannot do 1. Since you don't have the calibration variables, you cannot do 2.

Improve this answer
$\endgroup$
1
  • 1
    $\begingroup$ What if there were no strata or clusters? For example, if you have sampling weights calculated after the event to match a sample to a population? $\endgroup$
    – dash2
    Commented May 6, 2022 at 18:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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