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: Testing whether two estimated survey means are statistically different

From   Nick Winter <>
Subject   Re: st: Testing whether two estimated survey means are statistically different
Date   Thu, 20 Oct 2011 10:00:08 -0400

There is a subtle difference between -svy: reg- and -svy: mean- followed by -lincom-. The former corresponds to a t-test with the assumption that the variances in the two populations are equal; the latter corresponds (almost exactly) to a t-test without that assumption. The "almost exactly" is due to a difference in calculation of degrees of freedom.

Leaving aside -svy-, we can see the slightly different standard errors:

sysuse auto
reg mpg foreign
ttest mpg, over(foreign)

mean: mpg, over(foreign)
lincom _b[Domestic]-_b[Foreign]
ttest mpg, over(foreign) unequal

- Nick Winter

On 10/20/2011 4:58 AM, Maarten Buis wrote:
On Wed, Oct 19, 2011 at 11:33 PM, Andrew Wade<>  wrote:
Presumably the regression approach, where you have a single dummy for the sub-population, will be identical (or near identical) to:
- estimating svy: mean with over command (ie. mean for sub-population, and not in sub-population).
- lincom after svy: mean.

Or could the regression approach be applied, whereby you suppress the constant term?
Wouldn't this alternate regression approach be closer to what I'm after?

In terms of coefficients, yes, but in terms of test, no. These two
test are truely equivalent, meaning they contain exactly the same
information and nothing is gained by chosing one over the other. Think
of it this way: You can say Adam has 3 and Berta has 2 more (constant
is 3 and the coefficient of Berta is 2) or you can say Adam has 3 and
Berta has 5 (suppress the constant and add dummies for Adam and
Berta). The two statements contain exactly the same information.

  The only reason for prefering one over the other is ease of
explanation. Sometimes we have hypotheses or traditions within a
(sub-)discipline which would make it easier for your audience to
understand one parameterization over the other.

Hope this helps,

Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
*   For searches and help try:

Nicholas Winter                                 434.924.6994 t
Associate Professor                             434.924.3359 f
Department of Politics         e
University of Virginia w
S385 Gibson Hall, South Lawn
*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index