Re: st: appending two survey data sets

[Steve Samuels <sjsamuels@gmail.com>]

Ameya, For an SRS design you don't need to get the population N in each
stratum, just the number of centers in each stratum and the number of
eligible respondents in each sampled center. The data will contain,
obviously, the number of selected centers and selected respondents in
each.

You have a potential bias problem if the design was SRS and population
"sizes" of the health centers were skewed, e.g. there were relative few
"large" centers and more "small" ones. In such a case, respondents from
smaller centers may be over-represented.. The only simple fix is
post-stratification by center "size". Additionally, consider adding
center "size" to the regression models (see example below.)

> appending adds observations and I want to compare
> trends across both years), how do I do that?

If you wish to compare means or proportions
(let csize be a grouping of center sizes)
***********************
svy: mean myvar, over(year)
xi: svy: reg myvar i.year

svy: mean myvar over(year csize) 
xi: svy: reg myvar i.year i.csize i.year*i.csize
**********************

For some sampling references, see: 
http://www.stata.com/statalist/archive/2012-09/msg01058.html.


Steve 
 
On Oct 31, 2012, at 6:41 PM, Stas Kolenikov wrote:

On 1, 2, 3, the short answers are "yes", "yes" and "yes". The longer
answers depend on what you have at hand. If you had a simple random
sample at each stage, then you simply muliply through the ratios (# of
units sampled)/(# of units in the population) to get the probability
of selection. A smarter survey statistician would design a PPS survey,
in which hospitals would be selected with probabilities proportional
to the measure of size (# of beds, # of hospitalized, etc.). You
obviously have to make the names of your survey design variables the
same in two data sets.

A short answer to 4 is to -generate int year=2009- in one data set and
-year=2011- in the other before appending. I am not sure as to what's
the best way to approach 5, as it really depends on the computing
capacity you may have at hand. 800 variables and 10,000 observations
would produce at most 64Mb data set, and one would really have to go
back to the hardward from late 1990s to have problems with a data set
of this size.

"


-- 
-- Stas Kolenikov, PhD, PStat (SSC)  ::  http://stas.kolenikov.name
-- Senior Survey Statistician, Abt SRBI  ::  work email kolenikovs at
srbi dot com
-- Opinions stated in this email are mine only, and do not reflect the
position of my employer

On Wed, Oct 31, 2012 at 5:13 PM, Ameya Bondre
<ameyabondre.jhsph@gmail.com> wrote:
> My name is Ameya Bondre and I am working on two survey data sets for a
> sustainability study, and had few questions.
> 
> The study design:
> 
> To give you a background - I have to compare a range of conditions
> (health behaviors, diseases and health services) in a region, at the
> end of a health program (year 2009 - endline survey), with similar
> conditions two years after the program stopped (year 2011 - evaluation
> survey, to measure sustainability of program activities). I have two
> data sets for the two cross-sectional surveys conducted in 2009 and
> 2011. The surveys are independent (as in, the sampling was done again
> in 2011). The populations surveyed each time, are different
> cross-sections of the same region. Both surveys involve the same
> sampling technique with "block" as the stratum, "health center" as the
> primary sampling unit and "respondents/mothers" as the secondary
> sampling unit (but the variable names for these design variables are
> different in 2009 and 2011 data sets). I am using STATA 10. No FPC
> correction has been applied as per the program reports.
> 
> Questions (sampling weights and svy command):
> 
> 1) I have probability weights already given in the 2009 data sets but
> I don't have those built in, for the 2011 data sets. I have been told
> that the entire sampling method was similar for both years. Am I
> understanding correctly that I first need to calculate weights for all
> observations for 2011, then append data sets, and then set up the
> combined data set as a "survey set"?
> 
> 2) Further, do I need to create the sampling weight variable by
> calculating probability weights for 2011 observations (which I already
> have for 2009) ? if yes, what's the method to get weights - would I
> require the region's population (N) in 2011?
> 
> 3) Do I need to create new design variables for the svyset command,
> after appending the two data sets? (like one variable for psu, strata,
> weight - taking both data sets into account)
> 
> Questions (appending data sets)
> 
> 4) In appending, I am not able to label the variables/observations for
> 2011 separately from 2009, to identify them as "2009" and "2011"
> variables  (as appending adds observations and I want to compare
> trends across both years), how do I do that?
> 
> 4) Since I am using STATA 10 with limited memory and my data sets are
> huge (800 odd variables and sample sizes in thousands); can I append
> few variables at a time (that I need to analyze, for certain
> regressions), instead of the entire data set - would that affect the
> survey design of the new combined data set, after appending?
> 
> 
> 
> Please do let me know if any question is not clear. Thanks for your time..
> 
> Best,
> Ameya Bondre
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/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/faqs/resources/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/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/