Statistics, lies and "Meta-analysis"

[Michaelangelica]

"Meta-analysis"

This is a statistical analysis new to me.

What do the maths wizards think of it?

Junkfood Science

Last week’s report on the meta-analyses done by researchers at the Wolfson Institute of Preventive Medicine in London, led many to ask ‘Just what is a meta-analysis’?

 

Since there’s another example coming right up, let’s take a moment to understand these new types of studies. You might not come to think of these studies as studies at all, but many believe they are.

 

Meta-analysis is a statistical method first proposed in 1976 by an educational psychology statistician, Gene V. Glass, as a way to analyze findings from a bunch of individual studies.

 

A meta-analysis is an analysis of other analyses to create a new study.

 

This technique is frequently used when there are no large, high quality, randomized, double-blind, placebo-controlled clinical trials — the gold standard — to prove the validity of a treatment or theory.

So a meta-analysis lumps together whatever evidence is available: the good, bad and indifferent. Some studies may show a weak positive statistical association, others report none, and others may even report a negative correlation.

It can end up giving the same weight to well-designed studies as poor ones, and create mud. By pooling what are oftentimes weak studies together, it is hoped to create a statistically stronger estimation of an effect.

And therein lies the rub.

 

A favorite definition among critics is that of professor John Brignell, PhD, who authored The Epidemiologists: Have they got scares for you!--

 

Meta Analysis is making a strong chain by combining weak links.

 

When you’re reading about a study and see the word, “meta-analysis,” it’s a warning sign to proceed with extreme caution.

Junkfood Science

[InfiniteNow]

I see your Junkfood offering and counter with an NCBI PubMed presentation. Touche! :cup:

 

 

Entrez PubMed

The results of small meta-analyses should be regarded with caution, even if the p value shows extreme statistical significance. Larger meta-analyses (i.e., those with several hundred events) are likely to be more reliable and may be clinically useful. Well-conducted meta-analyses of large trials using individual patient data may provide the best estimates of treatment effects in the cohort overall and in clinically important subgroups.

 

 

Also, an online book for ye:

 

Methods of Meta-Analysis: correcting error and... - Google Book Search

[Michaelangelica]

So some agree its OK, some don't

What's the consensus???

[InfiniteNow]

That it works quite well when done correctly.

Cheers. :cup:

[Michaelangelica]

That it works quite well when done correctly.

Cheers. :cup:

OK Thanks

What's "correctly"?

[InfiniteNow]

In this context, "correctly" implies ensuring that the data being used is taken from a large population of sources, and that those sources each had a large subject population. There are other aspects to doing a meta-study "correctly", and the link to eBook I gave covers most.

 

 

Who's that Pearson guy? :cup:

[Michaelangelica]

Who's that Pearson guy? :cup:

Que?