Re: st: RE: Probability and non-response weights; how can I create a composite weight?

[Steve Samuels <sjsamuels@gmail.com>]

Try:

http://www.statcan.gc.ca/pub/12-001-x/2010001/article/11246-eng.pdf


Steve

On Nov 23, 2011, at 7:46 AM, Ángel Rodríguez Laso wrote:

Dear Cam,

The Haziza et al paper looks valuable, but the link you have
incorporated is the same than that for the Little and Vartivarian
paper. Can you provide the proper link?

Thank you very much.

Angel Rodriguez-Laso

2011/11/21 Cameron McIntosh <cnm100@hotmail.com>:
> Although a very popular adjustment method, there are some reasons not to simply multiple sampling weights by the inverse of the response rates:
> 
> Little, R.J., & Vartivarian, S. (2003). On weighting the rates in non-response weights. Statistics in Medicine, 22(9), 1589–1599.http://deepblue.lib.umich.edu/bitstream/2027.42/34860/1/1513_ftp.pdf
> 
> I would also suggest having a look at the variance estimation side of things:
> Haziza, D., Thompson, K.J., & Yung, W. (2010). The effect of nonresponse adjustments on variance estimation. Survey Methodology, 36(1), 35-43. http://deepblue.lib.umich.edu/bitstream/2027.42/34860/1/1513_ftp.pdf
> The Haziza et al. (2010) paper also has a decent overview of and a number of excellent references on non-response weighting.
> 
> HTH,
> 
> Cam
> 
>> From: s.scholes@ucl.ac.uk
>> To: statalist@hsphsun2.harvard.edu
>> Subject: st: RE: Probability and non-response weights; how can I create a composite weight?
>> Date: Mon, 21 Nov 2011 11:20:43 +0000
>> 
>> Arturo, here are my thoughts.
>> 
>> 1) Any adjustments to sampling weights (whether through direct standardisation or poststratification) will change both the estimates and standard errors.
>> 2) I cannot see any objection to the composite weight you suggest. I would be very interested to know why such a weight is not recommended.
>> 3) If using individual level data then the sum of the composite weight over all individuals should equal the population total of individuals (if all women in responding households are surveyed then
>> the responding households need to be weighted up to compensate for the non-responding HHs in each sampled village; and the sampled number of villages need to be weighted up to the total number of villages).
>> This you seem to do in your composite weight.
>> 
>> Hope this helps
>> Shaun
>> 
>> 
>> 
>> 
>> 
>> 
>> 
>> 
>> 
>> -----Original Message-----
>> From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Arturo Rodriguez
>> Sent: 18 November 2011 21:14
>> To: statalist@hsphsun2.harvard.edu
>> Subject: st: Probability and non-response weights; how can I create a composite weight?
>> 
>> Hi everybody,
>> 
>> I am working with survey data which was collected through a one-stage cluster sample with stratification (i.e. health clinics in a region were selected to be part of the study first and then villages surrounding those clinics were randomly sampled, all households in the selected villages and all women in those households were then surveyed).
>> 
>> My question is regarding the proper use of a composite design weight.
>> Here are the details:
>> 
>> (1) DESIGN WEIGHTS
>> I have calculated probability weights for each h_th stratum ... (i.e.
>> pw = total_villages_h / sampled_villages_h) I have also calculated unit non-response weights for each h_th stratum ... (i.e. unrw = total_eligible_households_h / sampled_households_h)
>> 
>> I then created a composite weights (sw) by multiplying pw * unrw = sw and used the following command to define my survey data:
>> svyset village_number [pweight = sw], strata(clinic) fpc(total_villages_h)
>> 
>> I have read that using a composite weight (sw = pw*unrw) is not recommended and that I should weight my sample first by pw and then weight again by unrw.
>> My first question is: Can STATA do this automatically? How? If not, do I have to multiply each of my 300+ variables by pw and then by unrw?
>> What is the most efficient way
>> to do this?
>> 
>> I asked STATA to calculate estimated population totals for each of the areas surrounding the clinics but I am getting figures that are very low compared to what is known to be true. Is there anything I can do to correct this statistically? Or is it just that my sample is not representative?
>> 
>> (2) POST-STRATIFICATION WEIGHTS
>> Will adding poststrata and postweight to my svyset make any difference in obtaining better "population around clinic" point estimates? Or will it only change my STD ERRORS?
>> 
>> Thanks in advance for all the wisdom.
>> *
>> *   For searches and help try:
>> *   http://www.stata.com/help.cgi?search
>> *   http://www.stata.com/support/statalist/faq
>> *   http://www.ats.ucla.edu/stat/stata/
>> 
>> 
>> 
>> *
>> *   For searches and help try:
>> *   http://www.stata.com/help.cgi?search
>> *   http://www.stata.com/support/statalist/faq
>> *   http://www.ats.ucla.edu/stat/stata/
> 
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
> 

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/