Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

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

RE: st: weights

From   "Cheng, Hsu-Chih" <>
To   "" <>
Subject   RE: st: weights
Date   Fri, 17 May 2013 23:26:17 +0000

Thanks for the very instructive suggestions from Nick, Austin, David, and Steve. Best, Simon

From: [] on behalf of Steve Samuels []
Sent: Wednesday, May 15, 2013 6:44 PM
Subject: Re: st: weights

It's possible that the weights have been normalized to sum to sample
size or to another constant. Simon can check this hypothesis easily

. total weight_variable

I think that such normalization is regrettable. For one thing,
normalization makes it impossible to estimate population totals. Also,
the normalized weights lose much of their diagnostic value. For example,
I once discovered a defective design by comparing estimated totals to
external values. The original weights have a natural scale, the "number of
population members represented"; the normalized weights do not. Thus
with original weights, it is easy to both identify and characterize
extremes. With normalized weights, extremely small
weights (close to or < 1) cannot even be detected.


On May 15, 2013, at 3:17 PM, David Hoaglin wrote:

Dear Simon,

It would be helpful to hear from others who have more experience with
the various types of weights, but I doubt that you can determine, from
only the numerical values of the weights in your data, which type of
weight they are.  The person who created the dataset should have
documented the process that produced the weights and thus explained
the type of weight.

David Hoaglin

On Tue, May 14, 2013 at 6:15 PM, Cheng, Hsu-Chih <> wrote:
> Dear Statalist veterans,
> Stata offers four options of weights: frequency weights, analytic weights, sampling weights, and importance weights.  Winship and Radbill (1994) suggest that un-weighted regression is preferred because it is less biased and more consistent than the weighted analysis, but their discussion is applicable only to sampling weights.  If the values of weights in my data center around 1 (some values are smaller than 1; some are greater), is it possible that these are still sampling weights?  Or, if the weights in the data are analytic or importance weights, what are the properties of using these weights in regression analysis?  Any suggestion or direction to read more on this issue is highly appreciate.
> Best,
> Simon

*   For searches and help try:

*   For searches and help try:
*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index