Statalist


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

Re: st: Question on a meta-regression problem


From   "G Livesey" <glivesey@inlogic.co.uk>
To   <statalist@hsphsun2.harvard.edu>
Subject   Re: st: Question on a meta-regression problem
Date   Mon, 27 Apr 2009 14:24:09 +0100

Firstly I would like to thank those who responded to my prior statlisters
email, and for the valuable references therein. I was advised to try GLST,
but for me this raises an important issue.
 
Previously I mentioned I was using metareg not glst to look at dose-response
data in observational studies, and this raised some query as to what problem
I had with glst, while at the same time it appeared that I was being firmly
recommended to use glst in preference to metareg.
 
I have had difficulty in application of the glst method to the analysis of
the slopes of observational studies I am working on. One of the problems I
have reproduced using "classical" data re-reported to illustrate the use of
glst in the Stata glst help file (from Greenland.1992 ) and the Stata book
Methods of Meta-analysis (from Wolk . 1999)
 
For me the results raise a central question, but first lets look at the
results I get from Wolk's data (Table 1). I can see that glst is far
superior to wls, but it is not clear to me that  this is owing largely to an
account being made of covariance gained from the case and none case data.
Rather it seems to me to be mostly an intercept problem that affects the
assessed slope reported by wls. 
 

Table 1 using Wolk..1999's data:        Coef          lci        uci
P              
Slope from wls                         -0.017       -0.046       0.012
0.246
Slope from glst                        -0.023       -0.048       0.001
0.063
Slope from metareg                     -0.022       -0.046       0.001
0.065         
 

By contrast, the glst and metareg approach give virtually identical results.
A common explanation is that the difference between wls and glst is an
account by the latter for raw case and non case data to gain access to a
pseudo-covariance, while wls does not. 
 
With the data from Wolk .1999 I find metareg to give the same result as
glst. For metareg  I first  re-associate the standard error value with the
referent, an se that is universally combined  with se values and reported
with the confidence intervals for quantiles >1. A similar level of identity
between metareg and glst as shown in Table 1 is obtained for the majority of
studies I am analysing as part of a current project, leading me to believe
that for these studies it is not necessary to apply glst. Such would be
useful because not all studies report full data, especially ones tending to
find lesser or unclear effects.
 
Now a different example. The Greenland.1999  data obtained via the glst help
file better represent the problem I see with a minority of studies I am
analyzing. Again the wls procedure has an intercept problem that affects the
reported slope.  But is there still a residual intercept problem affecting
the reported slope in the glst procedure or is this difference truly a
refection of the covariance effect alluded to? (There is certainly a problem
with the output from predict xb after glst, as xb does not passing among the
observations). 
 

Table 2 Greenland 1992                 Coef          lci         uci
P

Slope from wls                         0.033       -0.003       0.070
0.074
Slope from glst                        0.045        0.005       0.086
0.028
Slope from metareg                     0.050        0.010       0.090
0.013          
 
 

A characteristic of the minority of studies (like that analysed in Table 2)
seems to be that the data are scattered or at least apparently non-linear
near the intercept (when joining data points). Such may be real or more
often than not appears could be random. This leaves me believing that we may
still have an intercept problem affecting the slope in the glst approach.
If the latter is true we might expect meta-analysis of the slopes to be less
heterogeneous for slopes obtained by metareg than when obtained by glst,
which holds true at least for the studies I am analysing.  So is the glst
method as reliable as one may be wanting to believe? Can anyone explain
these data as robustly not due to an intercept problem?
 
The theory behind wanting to account for the 'lost covariance' when
analysing single studies may be sound, but I am worried that implementation
of the theoretical improvement has given rise to or leaves open another
important problem yet to be resolved. The question I raise is highly
important because results from glst analyses are often passed to national
and international organisations.

With many thanks for your help.

Geoff L.
Please consider the environment.
 
 
Geoffrey Livesey B.Sc., Ph.D., R.P.H.Nutr. 
Registered Public Health Nutritionist
 
Intelligent Nutrition and Health Research and Review
Commissioned by Industry, Government, and Academia
 
Member of SENSE, Professional Nutrition Consultants
www.sense-nutrition.org.uk
Member of VCG, Professional Bioscience Consultants   www.vcgllp.com 
 
 
 
INDEPENDENT NUTRITION LOGIC (INLogic) Ltd
NUTRITION RESEARCH MANAGEMENT AND CONSULTANCY
Limited Company registered in England and Wales Number 4991400
VAT Registration Number GB 731 9065 38
 
Tel:       +44-1953-606689
Mobile:  +44-7990-964609 
Fax:      +44-1953-600218
 
 
E:         glivesey@inlogic.co.uk
 
A:         Pealerswell House
            21 Bellrope Lane
            Wymondham
            Norfolk
            NR18 OQX
            United Kingdom
 
 
W:        www.inlogic.co.uk


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index