[6544] in s-news-athena
GEE s.e. generally smaller than naive s.e., why?
daemon@ATHENA.MIT.EDU (Ottar N. Bjornstad)
Wed Mar 1 15:43:07 1995
Date: Wed, 01 Mar 1995 14:56:30 +0100
To: s-news@utstat.toronto.edu
From: bjornsto@ERE.UMontreal.CA (Ottar N. Bjornstad)
Hi all,
At school I was taught that the problem with positively autocorrelated data
is that they lead to too anti-conservative statistics ("Type I error"; Too
narrow C.I.s). This is because we tend to underestimate the magnitude of
the error-component when the autocorrelation is positive.
I have some spatially autocorrelated data (autocorrelation is positive).
And I have a matrix representing the degree of pairwise autocorrelation
between each sample (calculated from the spatial correlogram of the
residuals of the model).
The data are counts so I wanted to use the the function in the gee library
(I know I have promised to summarize the topic of analysis of autocorrelated
count data; but I have to understand this myself first). My problem is as
follows: When I compare the S.E. of the naive model (the one assuming
independence, which as far is I understand is supposed to be
anti-conservative) with those of the GEE analysis (which is supposed to take
the autucorrelation into account) the former is nearly always wider. This
is exactly the opposite of what I expect. I would have had no problem with
the standard errors of the most important variables becoming more narrow
(perhaps due to increased presission). But when this happens to nearly all
variables I become very suspicious - What is wrong with me when I think
that the naive S.E. are to narrow, and expect that the corrected S.E. should
be wider?
In desperation,
Ottar,
--------------------------------------------
Ottar N. Bjornstad
Until 10th of may 1995:
bjornsto@ere.umontreal.ca
Univ. Montreal., Dept. Sci. Biol., CP 6128, Montreal, H3C 3J7, Canada.
Tel. (514) 343 6111 ext 1233, Fax. (514) 343-2293
After:
ottarnb@bio.uio.no
Div. Zool., Dept. Biol., Univ. of Oslo, Box 1050 Blindern, 0316 Oslo, Norway
Tel. + 47 22857295, Fax. + 47 22854605
