[R] Survey package

[Thomas Lumley]

On Sun, 9 Sep 2007, eugen pircalabelu wrote:
> A short example:
>
> stratum id weight nh Nh  y sex
>     1      1      3     5 15 23   1
>     1      2      3     5 15 25   1
>     1      3      3     5 15 27   2
>     1      4      3     5 15 21   2
>     1      5      3     5 15 22   1
>     2      6      4     3 12 33   1
>     2      7      4     3 12 27   1
>     2      8      4     3 12 29   2
>
> where nh is size of sample stratum and Nh the corresponding  population value, and  y is  metric variable.
>
> Now if i let
>
> design <- svydesign( id=~1, data=age, strata=~stratum, fpc=~Nh)
> then weights(design)  gives me 3,3,3,3,3,4,4,4.
>
> If i then let
>
> x<- postStratify( design, strata=~sex, data.frame(sex=c("1","2"), freq=c(10,15)))
> the weights become
>
> 1                    2            3            4            5            6            7            8
> 2.17               2.17       5.35       5.35        2.17    1.73        1.73        4.28
>
> If i define
>
> design <- svydesign( id=~1, data=age )
> x<- postStratify( design, strata=~sex, data.frame(sex=c("1","2"), freq=c(10,15)))
> weights become  2 2 5 5 2 2 2 5
>
> The question: does poststratify recognize that i have already stratified 
> in the first design by stratum and then it post stratifies by sex? and 
> why is that? (because i don't have the full joint distribution, the 
> sex*stratum crossing, in order to apply correctly the post stratify 
> function) I see that Mr Lumley uses the postStratify function when the 
> design does not include strata (eg from ?poststratify:
>

This gives you a design stratified by stratum and post-stratified by sex, 
which is not the same as stratifying by stratum*sex or post-stratifying by 
stratum*sex.

In this case you should probably rake() on stratum and sex rather than 
just post-stratifying. Post-stratifying on sex is equivalent to one iteration 
of the iterative proportional fitting algorithm used in raking.

 	-thomas