On Mar 27, 5:18=A0am, Bob <frott...@[EMAIL PROTECTED]
> wrote:
> Hi,
>
> I am trying to do a wage regression (ln(wage) dependent on several
> individual characteristics like age, education, region, etc.) based on
> household survey data and now I don't know if =A0and why sample weights
> (here: sample inflation factors, multipliers to inflate the sample to
> the total population) should be used in the regression and if so, how
> this is done.
> I saw some references where they discussed the issue but I didn't
> really understand why the one way or the other is preferred.
>
> Theoretically, I think one could clone the individual observations
> (single household) to equal the respective sample inflation factor and
> adding an error term from the distribution of the subgroup sample to
> each clone. But practically, the size of the data would not be
> manageable.
>
> Can anyone point me to a gentle introduction reference regarding this
> issues or give me some clues?
>
> Many thanks,
> =A0 Bob
The design of the survey has to be taken into account beyond the
weights, since most household surveys use quite complex designs with
multiple levels of stratification, sampling, etc. If the data come
from one of the big national surveys, they usually have some do***ent
describing the recommended method of analysis. Usually you have to use
regression methods designed specifically for survey data to get
correct standard errors of the estimates (weighted regression will
give the right estimates, but the standard errors will be too low). I
know that Stata, SAS, R do have such facilities. In those programs you
just specify the strata, primary sampling units and weights and they
will adjust for those in the regression models.
Aniko
|
