[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: when your sample is the entire population

From   "Austin Nichols" <[email protected]>
To   [email protected]
Subject   Re: st: when your sample is the entire population
Date   Fri, 18 Jan 2008 21:55:17 -0500

Paolo Pamini <[email protected]>:
You are describing a different problem--one of a confounding variable.
 Potential confounders in a non-experimental setting introduce a host
of problems, and techniques to deal with those problems are discussed
in a couple of papers in Stata Journal 7(4) due out any day now.

The problem of inference given population data can be combined with
the problem of confounders, but that would merely obscure the central

I intended my comment on a simple comparison of means or counts or
ratios across two categories as a kind of reductio ad absurdum for
those who believe no standard errors are necessary when dealing with
population data, but at least one Statalister is unconvinced.

Nick Cox and Michael Blasnik raise additional points, none (even
Nick's #5!) of which I see as incompatible with the notion that you
would nearly always want to compute a nonzero standard error for any
"estimate" drawn from population data.

On Jan 18, 2008 9:31 PM, Paolo Pamini <[email protected]> wrote:
> Hi all
> <<
> Suppose you have data on every kid in the school, not a survey, and every
> school in the district, and you want to test for some form of sex
> discrimination in assignment to a program.  Well, just see if more girls
> than boys (or vice versa) are assigned, and there is your evidence of
> discrimination, right?
> >>
> I think the problem doesn't disappear when working with explanatory
> variables: assume that assignment to a programme is determined by low income
> and that girls in your whole population are generally richer than boys, then
> you have an over-proportion of assigned boys, although not sex but income is
> the true explanatory variable. The whole story is even more complicated
> assuming relationships with further explanatory variables.
> As soon as you do a multiple regression you need significance tests even if
> working with the whole population, am I wrong?
> Regards
> Paolo
*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index