Re: st: sampling weight

[Stas Kolenikov <skolenik@gmail.com>]

On Thu, Sep 27, 2012 at 9:20 AM, Nick Cox <njcoxstata@gmail.com> wrote:
> I think there is some misunderstanding here. Stas wants you to
> describe the sampling design that was used to produce your dataset,
> not to design a survey yourself. This means exactly how the data were
> collected.
>
> More broadly, neither Stas nor anybody else can give good advice to
> you on how to analyse your dataset without an idea of how that dataset
> was generated. (I  am guessing you did not visit the cities and select
> the people yourself.)  Perhaps this is not even documented clearly,
> but the point remains. Using any kind of weights is dubious unless you
> know from documentation of the survey that those weights make sense.

If you did go select the people yourself, you had to have known
something about sampling to have done this properly. If a survey
organization collected the data, they should have provided you a
methods report describing how the sample was drawn, and how the data
were collected. If they have not, it should have been their
responsibility, and you are in a position to ask them. Of course, I
can imagine a number of worst case scenarios when the data were
collected, but no proper report was written, and the person who
oversaw the data collection left the organization, etc. But usually
there are ways to find out about how the sample was drawn.

-- 
-- 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



>
> Nick
>
> On Thu, Sep 27, 2012 at 3:07 PM, Lynn Lee <lynn09v@gmail.com> wrote:
>
>> I have no idea about sampling design ( I never learned before.). The below
>> are just my idea about choice of simple weights. I generate a new variable,
>> which is total number of individuals in each city in the data set. And I
>> choose this new variable as weights, type in -[pweight=total]- , looks like
>> Stata11 can do this weighted regression for me. But I can not figure out
>> how Stata11 do weighting. Could you please give me suggestion about basics
>> of sampling design (or some web link)?  I am new to sampling design, I do
>> not know how to describe in full detail.
>
> Stas Kolenikov
>
>> These are steps in the right direction. Please describe your sampling design
>> in full detail, so that we could brainstorm and see what the right
>> specifications should be.
>
> Lynn Lee
>
>>> I just want to do simple sampling.
>>>
>>> Take "webuse total" for example. I am wondering how was "swgt" generated?
>> I
>>> guess: obs 1 has her corresponding sampling weight, swgt=25964, which is
>> the
>>> total population in her group; obs 4 has his corresponding sampling
>> weight,
>>> swgt=4312, which is the total population in his group;etc.  Is that right?
>>>
>>> So, if I use this logic in my downloaded survey data sets, I can group all
>>> the obs into different sampling weight over residence place and gender.
>>> Like: I calculate total number of individuals who were in the dataset
>>> according to their resident city , say, total number of individuals in
>> city
>>> 1 is 1000 in dataset, total number of individuals in city n is 400 in the
>>> data set, then, I generate this city-total-individuals as a new variable
>>> (weight). (Or I can even be more detailed, total number of people in the
>>> data set over city, gender, age.) In regression, I simply use command "reg
>> y
>>> x1 x2 x3 [pweight=total]". Can this way correct in part for unweighted
>> data
>>> set?
>>>
>>> Suppose the mean of total(weights) is 500, min is 100 and max is 800.Then,
>>> weighted analysis will give at most 800/100 times the weights to
>> potentially
>>> under-sampled observations. Do I understand correctly?
>
> Stas Kolenikov
>
>>> If Lynn obtained her sample in a rigorous way by enumerating the
>> dwellings,
>>> she should have all the inputs into the probability of selection, and the
>>> baseline sampling weight is the inverse of that.
>>> Then she would want to correct for non-response, which would be the
>> fraction
>>> of those responding to the survey among those sampled.
>>>
>>> If Lynn is interested in a specific population (females of reproductive
>> age,
>>> say), and that's who the survey collected the data on, then she would need
>>> to get the total population counts for that specific population (which may
>>> prove even more difficult).
>>>
>>> If she does not have these figures, then I don't really know what to do.
>> As
>>> they say, when you approach a statistician with collected data in hand,
>> they
>>> can only tell you what killed your study.
>
>>> On Wed, Sep 26, 2012 at 8:15 AM, JVerkuilen (Gmail)
>
>>>> On Wed, Sep 26, 2012 at 2:49 AM, Lynn Lee <lynn09v@gmail.com> wrote:
>>>>
>>>>> Any suggestion to suggest which weight is better? Or, other types of
>>> weights
>>>>> may be better than population weights?
>>>>
>>>> Do you have a few accurately observed variables such as the population
>>>> age and gender breakdown? If so you can often create
>>>> post-stratification weights (through a process called "raking") that
>>>> make your samples align with the associations observed in those
>>>> tables.
>>>>
>>>> A quick -findit raking- turned up a program -ipfraking- written by
>>>> Stas Kolenikov and available from his website. Hopefully he'll chime
>>>> in.
> *
> *   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/