On Thu, 15 Aug 2002 14:19:52 -0400 firstname.lastname@example.org wrote:
> I had a question concerning the use of pweights in general, and with respect to
> inequality an d poverty measures in particular. I think I found most of the
> answers on statalist, but I would like to be sure since the Yahoo search
> facility is not very good.
> Basically, I am using ineqdeco (and povdeco). The dataset (British Houshold
> Panel Survey) has pweights, but these are not allowed by any of the two
> programs. I have the choice of working with no weights at all, but I would
> prefer to use them. With ineqdeco, for instance, I get an error message when
> trying to use pweights, and if I try to "force" the weights, they are assumed to
> be aweights:
> ineqdeco income [w=pweight]
> (analytic weights assumed)
> >From what I could find on the server, after a point raised by Michael Smith,
> nick cox asked:
> >More seriously, can you compute aweights from your pweights?
> and Stephen Jenkins (implicitly?) replied:
> >1) The comment about a lack of estimators for inequality indices when
> >using complex survey data (i.e. with clustering and stratification) is
> >only relevant if one wants estimates of standard errors and confidence
> >intervals (in addition to the point estimates). For the point
> >estimates, just treat your pweights as aweights and use one of the many
> >Stata programs available for calculating inequality indices
> So my question is, like Nick's: is there a way to "translate" pweights into a or
> f weights, or using the p's as a's is ok as long as you just want point
> Does it make any sense to revise the ado files and adapt them to pweights (say,
> by replacing sum with svymean, etc., within the programs)? I can do this, but
> only if it makes sense of course!
> Finally, I am using the British Houshold Panel Survey (cross sections in
> principle), if it helps at all. Excuse me if this is not a Stata question: if
> using only observations that are present in every wave, but analysing only
> cross sections, is it better to use the weights, or just discard them?
> Longitudinal weights take attrition into account, but I think the cross section
> ones do not.
As my earlier message ("31181 Re: st: RE: inequality indexes and weighting Stephen P. Jenkins Thu
4/4/2002" said, things are much easier for poverty indices than for
inequality indices, as virtually all the former (including the Foster
et al family) are simple functions of moments of the data.
If you have p-weighted data, and want to estimate poverty indices:
Strategy 1: use -povdeco [aw= <p-weight varname>]: you'll get the
correct point estimates.
Strategy 2: generate suitable new variables and then use -svymean-.
This will allow p-weights (and also strata and cluster, if available).
As well as providing the same point estimates as Strategy 1, you will
also get the "correct" svy standard errors. -svymean- gives you
estimates based on exactly the same formulae as given in Howes and
Lanjouw (1998) [like you, also at the World Bank!], who in effect
reproduce formulae already well-known in the stats literature.
Because this is so straightforward to do, I have never bothered to can
it into a special routine.
What are the suitable new variables? Let y be income, and z the
poverty line. Consider the headcount ratio poverty measure (FGT(a),
a=0). Create new indicator variable I = 1 if poor, 0 otherwise. For
FGT(2), create new variable I*(1 - y/z); for FGT(2), create new
variable I*(1 - y/z)^2.
NB the BHPS has cluster/psu and strata variables in the AHHSAMP file
(AHHAC and ASTRATA).
Finally, it sounds as you are using a sample that is a balanced panel,
but then only analysing each cross-section separately -- that seems
a weird sample to me for repeated cross-section analysis. Also, be
sure to read the BHPS documentation about the different sorts of
weights and their properties (cross-sectional, longitudinal, and how to
handle the supplementary Welsh and Scottish samples (via wMEMORIG
variable)) Both longitudinal and cross-section weights take attrition
into account in senses that are too complicated (and unsuitable) to
[author of -ineqdeco- and -povdeco-, and researcher at the
place that produces the BHPS!]
Howes SR and Lanjouw JO (1998) "Poverty comparisons and
household survey design", Review of Income and Wealth, 44, 99-108
Professor Stephen P. Jenkins <email@example.com>
Institute for Social and Economic Research (ISER)
University of Essex, Colchester, CO4 3SQ, UK
Tel: +44 (0)1206 873374. Fax: +44 (0)1206 873151.
* For searches and help try: