Statalist


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

RE: st: svyset when looking at children of respondents


From   "Ergo, Alex" <aergo@jhsph.edu>
To   "statalist@hsphsun2.harvard.edu" <statalist@hsphsun2.harvard.edu>
Subject   RE: st: svyset when looking at children of respondents
Date   Mon, 27 Oct 2008 11:37:21 -0400

Steve -- Thanks a lot for the clarifications and the additional suggestions!
Alex

________________________________________
From: owner-statalist@hsphsun2.harvard.edu [owner-statalist@hsphsun2.harvard.edu] On Behalf Of Steven Samuels [sjhsamuels@earthlink.net]
Sent: Saturday, October 25, 2008 6:26 PM
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: svyset when looking at children of respondents

On Oct 25, 2008, at 5:11 PM, Ergo, Alex wrote:

> Thanks a lot for your advice, Steve! It's extremely helpful.
> There's just one point in your response I'm not sure I understand.
> What do you mean when you write that each child will inherit the
> probability weight of his/her mother? Do I not need to adjust the
> weights to account for the fact that the number of children
> included in the analysis will vary across women?
No.  The survey programs will automatically take the number of
children into account.  If you were selecting a child at random from
each woman's birth history you might want to re-weight.
> Also, would you agree that, in your first suggested approach, I
> should keep women without children in the dataset and use the
> 'subpop' option with the svyset command?

I disagree. I would keep only the women with children in the data
set, and  I would not use -subpop-.  I  -subpop- only for descriptive
studies; your study has an analytic purpose. (See WE Deming, Some
Theory of Sampling", Dover Reprint, 1984). In essence, you are
conditioning on the population of women with births and, also, on
their covariate values.

> Once again, many thanks for your help.
> Alex
You are welcome. You haven't told us much about your study question.
However, if I were doing your analysis, I would first ignore the
survey design and fit multi-level models with -xtmelogit-.  I would
then use results from these analyses as initial parameter estimates
in -gllamm-.  With -svy: logistic- approach, you will get no
information about between-woman variation in the probability of an
infant death. (This might be available from other sources).  You run
the risk that the number of births is related to the probability of
infant death.  Sample sizes should be uninformative, but  pregnancy
outcomes  may not be.

I also withdraw my suggestion to include prior birth outcomes as
covariates.  Higher numbers may be predictive of greater risk solely
because they reflect the mother's inherently greater risk.  Also, the
effects of exposures may be may be changed. If an exposure has
already increased the risk of a prior infant death, its *additional*
effect on a current pregnancy will be reduced.  This is a tricky area
to interpret:  Is risk related to birth order? To the number of
children already in the house? Perhaps. But,these predictors may
themselves be outcomes related to infant death.  Luckily, JHSPH has
fine reproductive epidemiologists and statisticians to advise you.

-Steve
> ________________________________________
> From: owner-statalist@hsphsun2.harvard.edu [owner-
> statalist@hsphsun2.harvard.edu] On Behalf Of Steven Samuels
> [sjhsamuels@earthlink.net]
> Sent: Friday, October 24, 2008 2:21 PM
> To: statalist@hsphsun2.harvard.edu
> Subject: Re: st: svyset when looking at children of respondents
>
> Alex:
>
> Too start off,  each child will inherit the probability weight,
> stratum, and cluster assignments of its mother.
>
> I see two ways of accounting for clustering by mother:
>
> 1. Add mother as a last stage of "sampling" in your -svyset- command,
> but with no finite population correction. See the -svyset- help or
> the Stata Survey manual.  Use Stata's -svy: logit- command.
>
> 2. Fit a multilevel logistic model (level 1: child, level 2: mother)
> with -gllamm- (downloadable from SSC, manual http:// www.bepress.com/
> ucbbiostat/paper160). -gllamm- accepts probability weights and a
> cluster unit above than the highest level in your model. -gllamm- is
> not a -svy- enabled program, so you cannot use the stratum
> information in the survey design. However you can use stratum-level
> covariates. Stas Kolenikov has a demo at http://www.unc.edu/~skolenik/
> stata/gllamm-demo.html
>
> If you are interested in conditional, rather than marginal,
> predictions, you might choose to ignore the survey sampling weights
> altogether.
>
> I recommend the -glamm- option. With -glamm- you will be able to
> model woman-level effects as fixed and random.
>
> You should be aware of a potential bias in selecting the births for
> your study data. Women may prefer to end their pregnancies with a
> successful one (in some places, perhaps, with a successful male
> birth). If this is the case, you should exclude a woman's last birth
> from your data.   To guard against this problem, you may also include
> as a covariate the outcomes of prior pregnancies and births. I would
> not recommend this if you are interested in marginal, rather than
> conditional, prediction. (-gllamm- will do both kinds.)
>
> If you want to use Stata's -svy- commands, and you are combining
> multiple surveys, there are other issues. I suggest that you create
> "super-strata" which cross countries or survey periods with the
> within-survey strata.
>
> Good luck!
>
> -Steve
>
>
> On Oct 24, 2008, at 12:44 PM, Ergo, Alex wrote:
>
>> Dear A
>> I'm working with large population surveys. The individuals
>> interviewed are women of
>> reproductive age. Among many other things, the respondents provide
>> information relating
>> to their children. All this information is stored in the
>> respondent's record.
>> I would like to run some logistic regressions with infant mortality
>> as dependent
>> variable (1 if child died within the month following birth; 0
>> otherwise). In order
>> to create this dependent variable, I need to reshape the dataset
>> from wide to long
>> so as to have one live birth per record. I do not consider all the
>> children for
>> which information is available, but only those born up to 10 years
>> before the
>> mother's interview date.
>> In this situation, what is the most appropriate approach to account
>> for the complex
>> survey design? I thought of using the svyset command, but I'm not
>> sure how.  More
>> particularly, should I account for the clustering of live births at
>> the level of the
>> respondent and for the fact that respondents who did not have any
>> live birth in the
>> last 10 years are omitted from the regression analysis? If so, how?
>> Should I adjust
>> the sample weights when more than one child is from the same mother?
>> I'm using STATA 9.2.
>> I hope someone can help me with this.  Thanks in advance!
>> Alex
>>
>> *
>> *   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/



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