[R] Bootstrap and Jackknife Bias using Survey Package

[Thomas Lumley]

On Tue, 11 Apr 2006, Carlos Creva Singano (M2004078) wrote:

> Dear R users,
>
> I´m student of Master in Statistic and Data analysis, in New University 
> of Lisbon. And now i´m writting my dissertation in variance 
> estimation.So i´m using Survey Package to compute the principal 
> estimators and theirs variances.
>
> My data is from Incoming and Expendire Survey. This is stratified 
> Multi-stage Survey care out by National Statistic Institute of 
> Mozambique. My domain of analysis is Maputo City, the Capital of 
> Mozambique. I just compute the sampling errors using Survey Package, but 
> i have some droubles:


> 1. How to compute Bootstrap and Jackknife Bias of estimates, like mean?

I don't know how to do this for survey estimates, but if you do, you can 
compute it from the replicates.

> 2. How to see each replicate estimate of the parameter?

All the survey functions for replicate designs will return the replicates 
with the return.replicates=TRUE option

R> a <- svymean(~api00,rclus1,return.replicates=TRUE)
R> summary(a$replicates)
    Min. 1st Qu.  Median    Mean 3rd Qu.    Max.
   636.1   640.4   642.9   644.3   645.7   667.7
R> str(a)
List of 2
  $ mean      : atomic [1:1] 644
   ..- attr(*, "var")= num 693
   ..- attr(*, "statistic")= chr "mean"
  $ replicates: num [1:15] 643 648 646 643 645 ...
  - attr(*, "class")= chr "svrepstat"


> 3. Is it possible to use the Sitter algoritm in Bootstrap for stratified 
> and multi-stage sampling in Survey Package? Or is possible to use the 
> pseudo-population (obtained by replication the sample) in estimation 
> using Bootstrap?
>

as.svrepdesign( ,type="bootstrap")
does create a pseudo-population in designs with a finite population 
correction. This is as described in the reference by Davison and Canty 
given on the help page.

I don't know what you mean by the "Sitter algorithm".  Prof Sitter has 
written several papers on bootstrapping survey data -- can you give a more 
precise reference?

 	-thomas

Thomas Lumley			Assoc. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle