🚀 go-pugleaf

   For the second time in less than a year, I am going through what is
the closest thing to Hell on Earth that I know of: arguing over sample
weighting with a French statistician :-)

   The last time, the matter was trivial -even though thoroughly disputed-
and reference to classics like Kisch's _Survey sampling_ finally proved enough.
In the meanwhile, I became almost crazy and had to be put back on track by
a few people on the list. This time however, the matter is more tricky and
I definitely could use some help.

   The data come from a two-stage sample design (households then
individuals) with sampling proportions varying according to age in the
second stage. The tricky thing is that the data are life histories that are
to be used in event history analysis (a.k.a survival regression or failure
time analysis). There is little litterature about the weighting of such
samples. Conventional sampling (and weighting) theory and techniques have
been developped for "ordinary" data and do not apply to longitudinal
data in a straightforward manner.

   The only reference on the weighting of life histories collected through
sampling procedures we know of is Hoem, Jan H., (1985) "Weighting,
misclassification and other issues in the analysis of survey samples of
life histories" in Heckman, J. J. and Singer, B, Eds. _Longitudinal
analysis of labor market data_, Cambridge University Press. Hoem first
considers the case where the sampling design is "non-informative", that is
when the life histories are selected in a such a way that each one of them
can be considered as independently selected of any other. In this case, he
proves (using a mathematical demonstration that I can hardly follow, less
criticize) that weighting is irrelevant. In cases where the sampling design
is informative, that is when the selection of the individuals who will
provide their life history is dependent on other individuals' life
histories, weighting is justified because the probability of their
selection is the outcome of a stochastic process (at least this is what I
understand of the paper).

   I therefore suggested that weights should be used so that the sampling
probability of all life histories be equal, using the weights that would
be appropriate for ordinary data. This view is supported by other people
around me who weighted similar samples when they used survival regression
models even though none of them can pretend to have a full understanding of
the probability and statistical problems involved. The other party -that I
strongly suspect of having little understanding of sampling if any and of
resenting the use of a software they are not used to- considers that there
is too little litterature on the matter to use any form of weighting.

   Here come the questions.

   Does anyone have enough knowledge of this matter to either confirm or
dismiss my understanding of Hoem's paper and my appreciation of our
problem?

   Does anyone know of other references dealing with this problem?

   Did anyone have to get his way through a similar problem?

   As usual, any help will be greatly appreciated.

Benoit Laplante
INRS/Urbanisation
Montreal, Quebec
LAPLANTE@INRS-URB.UQUEBEC.CA
LAPLANTE@UQUEBEC.BITNET