Re: st: combining multiple surveys - adjustment of weights

[Steven Samuels <sjsamuels@gmail.com>]

I should clarify point 1. I meant, there is no statistical problem if  
all primary sampling units are defined in the same way.  Example where  
this was the case: In some of our surveys of  businesses, some were  
listed  at each location where they operated; others had one listing  
that covered all locations. If we drew the sample only from the  
original list, there was a problem whether or not our intended  
sampling units were locations or the parent businesses.

Steve

On Jan 28, 2011, at 9:04 AM, Steven Samuels wrote:


On Jan 27, 2011, at 11:45 PM, Dana Shills wrote:


> I am trying to combine multiple country survey datasets (10  
surveys) to
> estimate ols regressions of firm performance across the 10  
countries. 8 of
> the surveys are stratified random samples of firms in the 8  
countries (so
> the strata identifier in each dataset goes from 1 to some number N,  
that
> varies) and 2 others are simple random samples. The survey  
questionnaire
> is the same across countries.

> 1> Statistically are there any issues with combining multiple  
surveys when
> the sampling methods are different?

Not if the primary sampling units were the same in all surveys.

> 2> To declare the combined dataset in Stata as a survey, I create a  
new
> strata variable that doesn't overlap between the surveys (as  
suggested in
> the previous thread). But how should I deal with the probability  
weights
> provided in each dataset?

See below.

> Suppose I assume that the sum of the weights in each country  
(countrysum)
> represents the population, then can I just rescale each weight by  
(sum of
> weighs across all countries)/countrysum?

Don't assume: Check! If the weight sum in each country is  
approximately or exactly equal to the population size (number of  
firms), don't rescale.  Use the weights as given.

(Modification: if you know the population sizes exactly and had <100%  
response, you might apply non-response adjustments.  With SRS, the  
simplest  is to create a new weight=  N/n' where N is the known  
population size in the stratum and n' is the achieved sample size.   
For other methods, see section 8.5 of Sharon Lohr, Sampling: Design  
and Analysis, 1999 or 2010 editions, Brooks/Cole, Boston.)


> 3> I am running weighted regressions using the svy: regress command  
on the
> combined dataset. If the sample size varies greatly between the  
surveys,
> should I rescale each firm level variable in the regression by the  
sample
> size in that country or does svy take care of this?

Stata's survey commands will take care of it.



Steve

Steven J. Samuels
sjsamuels@gmail.com
18 Cantine's Island
Saugerties NY 12477
USA
Voice: 845-246-0774
Fax:    206-202-4783



*
*   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/