Statalist


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

Re: st: Trend test in meta analysis


From   "Austin Nichols" <austinnichols@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: Trend test in meta analysis
Date   Fri, 21 Dec 2007 21:23:09 -0500

Vijay, Tom, and Ben:
I have read neither of the references cited by Ben and Tom, but it
seems to me a more fundamental problem is that the meta-analysis
includes observational studies (I assume). The principle of
meta-analysis is to combine several low-precision unbiased estimates
to get a high-precision unbiased estimate; if every study exhibits a
positive bias, can one even consider a meta-analysis?  In that case,
some kind of sensitivity testing  la Rosenbaum's _Observational
Studies_ seems in order.

That said, I am wondering if some kind of sign test (see -help
signtest-) might not be a good way forward, e.g. count the number of
studies with s1>0 (where s1 is the coef on category 1, low dose
smoker, with the reference category nonsmoker represented by the
zero--could also omit comparisons with zero) and s1<0, those with
s2>s1 and with s2<s1, those with s2>0 and s2<0, etc. Under the null of
no trend, all of these events have probability one half. A similar
approach would be to use -nptrend oddsrat, by(category)-, I think.

Does that make sense in this context?  This is very far afield from my
own field...

On Dec 21, 2007 4:59 PM, Tom Trikalinos <ttrikalin@gmail.com> wrote:
> Hi vijay
>
> I'm not clear what form your data have.
>
> Let me say right off the bat that such a question (dose-response)
> could be evaluated on a qualitative basis (with a plot) without giving
> a p-value. This should not be discarded as an option, given the strong
> assumptions that are inherent in the data-abstraction process during
> meta-analysis.
>
> That being said see if the paper by Jesse Berlin et al.
> "Meta-analysis of epidemiologic dose-response data." Epidemiology.
> 1993 May;4(3):218-28. PMID: 8512986 helps.
>
> Two general comments
> A.  first summing up the ORs 1 to 10 and then seeing for a trend (e.g.
> with a meta-regression) is subject to Simpson's paradox. This is why
> you need an approach like the on described in the cited paper.
>
> B.  I'm not sure how metap could help you.  -metap- does a
> meta-analysis of significance levels, a whole family of non-parametric
> meta-analysis methods (metap implements 3 of numerous approaches).
> This is essentially an omnibus test and therefore does not have the
> interpretation you would wish. A meta-analysis of p-values asks is
> there evidence of significant deviation from the null in AT LEAST ONE
> of the studies?  A typical misinterpretaion of metap results is that
> this is the p-value for the overall summary effect.
>
>
> hope these thoughts help
>
> tom
>
>
>
> On Dec 21, 2007 3:52 PM, Jayaprakash, Vijay
>
> <Vijay.Jayaprakash@roswellpark.org> wrote:
> > Hi Stata users,
> >        I'm trying to do a meta analysis of data from 10 different studies. The smoking variable in each study is stratified into categories (Category 1:1-10 cigarettes, Category 2: 10-20 cigs, Category3: 20-30 cigs etc.). I calculated the OR for each category (of smoking) by study.  I then did a meta analysis and calculated the OR for each category by random effects model using the meta command:
> > meta  or1 lcl1 ucl1, ci eform

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/



© Copyright 1996–2021 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index