[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: STATA and NHIS weight variables

From   Michael Drazer <>
Subject   Re: st: STATA and NHIS weight variables
Date   Mon, 20 Jul 2009 13:48:40 -0500

Steve, Stas, Joao, and Nick: thank you for the help.

Stas, your understanding of the design agrees with my own understanding, and the sample adult and sample adult cancer data do have weights (both wtfa and wtfa_sa, with wtfa ~= wtfa_sa ) for the individuals who completed the survey for the sample adult and cancer files. Adults not completing the cancer/sample adult surveys only have a wtfa weight.

I merged the three files because I created a new variable (referred to here as variable "D") that was defined using data in each of the three files - so variable A was only located in the person file, whereas variable B was only located in the Sample Adult file, and variable C was only located in the Cancer file. Variable D varies based upon the values of variables A, B, and C.

Re: the observations, some of the observations do overlap. For example, an observation for Adult "A" would be found in all three files (Person, Adult, Cancer) if they were randomly sampled, whereas an observation for Adult "B" would only be found in the Person file if they were surveyed for the Person file, but another adult (Adult "C", a spouse, relative, etc.) in the household was randomly sampled for the adult/cancer surveys. Even so, the merge was successful (I merged by household, family, and person record identifiers) and I have the correct # of total observations.

I suspect that the best way to address this issue, based on the helpful feedback of the Statalisters (thank you for the help, again) would be to redefine my newly created variable "D" for only those persons who completed all three surveys, and then to perform my analyses on that same subpopulation of persons who completed all three surveys.

In response to Steve's comment, I'm using Stata 10, so you were correct in identifying that there is no "psu()" option - my command was out-of-date (I was using this resource from NHIS in writing the command for my purposes. The new command reads:

svyset psu [pweight=wtfa_sa], strata(stratum) vce(linearized) singleunit(missing)

Please correct me if this does not sound correct.

Thanks again for the consideration and help,


On Jul 20, 2009, at 12:22 PM, Stas Kolenikov wrote:

That's terrible -- they have a 100k input file with 2200+ lines, but
did not bother to put just one more line for the appropriate design

Here's my understanding of NHIS: the household was conducted, and data
were collected for each person (by self or proxy reporting) producing
the person file. The weights in the person file would essentially be
the household weights. An adult was sampled to provide more detailed
information; their weight will be higher than the person weight by the
factor equal to (# of people in the household). All sampled adults
should have reported on the cancer supplement, so the sample adult and
the sample adult cancer data should have the same weights.

What was it that you tried to achieve by merging the three files? My
understanding is that you should've received almost a block-diagonal
structure of your data: the variables from the person data go with the
person(=household) weights, the variables from the sample adult and
cancer data go with the sample adult weights, and the
observations/variables do not overlap. Am I wrong? If you want to mix
information from them in your analysis, then you will have a lot of
missing data from the unsampled adults who did not provide the
information for the sample and cancer data sets. If only do the
analysis using the person data, you should use the wtfa weights from
the person data. If your analysis uses sample adult and cancer data,
you should use wtfa_sa weights. I expect that it will be impossible to
meaningfully analyze the data from all three files simultaneously, but
I will likely be wrong with this.

On Mon, Jul 20, 2009 at 11:07 AM, Michael
Drazer<> wrote:

I am new to STATA and am doing some work with a dataset that I created by merging together the 2005 sample adult, person, and sample adult cancer NHIS datasets. I'm trying to construct an appropriate survey design statement,
but am not sure how to properly define the pweights for my dataset.

According to the Sample Adult variable layout, the wtfa_sa weight should be used for most sample adult analyses - I noticed, however, that the Sample
Adult dataset also has values for the wtfa weight. All of the 98,649
observations in my merged dataset have wtfa weights, but only those 31,428
observations in the sample adult/sample adult cancer datasets contain
wtfa_sa weights. My question is: to properly weigh this data, should I use a
standard survey design statement, such as:

svyset [pweight=wtfa],strata(stratum)psu(psu)

even though it does not take into account the wtfa_sa weights, or is there a
better way to weigh this data?

For additional info re: variable layouts in the NHIS data, see here (the Variable layout PDFs are informative in terms of a description of the wtfa and wtfa_sa variables - see the person layout for wtfa and either the sample
adult or sample adult cancer layout for the wtfa_sa variable):

Thanks in advance,


*   For searches and help try:

Stas Kolenikov, also found at
Small print: I use this email account for mailing lists only.

*   For searches and help try:

*   For searches and help try:

© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index