I'll post if I come up with anything good.

Maybe I have, at last:

**Fatal Flaw in Glandular Fever model?**
This study found that psychological factors at baseline (initial infection) predicted which patients went on to develop CF (Chronic Fatigue) or CFS, and which did not. Or rather which psychological factors had a rather modest impact on the risk of developing CF/CFS. But the authors' approach in this study might be technically flawed.

*This is a bit geeky so bear with me*
The used a modelling technique called Logistic Regression to work out which baseline factors were predictors of the binary outcome i.e. developing CF or CFS. For such models to work, they need a lot more subjects than predictors and their final model had 217 subjects and 8 predictors, which gives an apparently healthy ratio of 27 subjects per predictor.

Except, it turns out that things are different in Logistic Regression where you have binary outcomes (subjects either have CFS or they don't) as opposed to normal regression where the outcome is a continuous score (e.g. Depression score, 0-30 points). Specifically, in Logistic Regression, it turns out the relevant number of subjects is the number of events i.e CFS cases in this study.

**There were only 17 CFS cases and 8 predictors: a wholly inadequate ratio of 2.1 events per predictor.**
This issue with Logistic Regression comes from a

2004 paper by Babyak, who points out that simulation studies show that:

When only 2 events per predictor were available, 30% to 70% of the

estimates had bias greater than 100%! **At a minimum, these**

**results should make us very wary of an article that does not at**

**least meet the rough guideline of 10 to 15 events per predictorâ€”**

**an all too common feature of many published articles**.

Incidentally, this Babyak paper was cited by James Coyne, Psychology professor and

scourge of flaky psychological research.

So, assuming Babyak is correct (I'd be very grateful for any views on this) then the Moss-Morris paper finding that psychological factors at infection predict development of CFS is deeply flawed. In any case their results are of rather modest effects, and the predictors of CF at 3 months are completely different from the predictors of CFS at 6 months, suggesting the model isn't up to much. It may be because they have vastly overfitted data to produce their model.