1. Patients launch a $1.27 million crowdfunding campaign for ME/CFS gut microbiome study.
    Check out the website, Facebook and Twitter. Join in donate and spread the word!
9th Invest in ME International ME Conference, 2014 - Part 1: Autoimmunity and ME
Mark Berry begins a series of articles on the 9th Invest in ME International ME Conference in London, with a look at three presentations on autoimmunity
Discuss the article on the Forums.

(PACE Trial) Does the Heterogeneity of Chronic Fatigue Syndrome Moderate the Response

Discussion in 'Latest ME/CFS Research' started by Dolphin, Aug 7, 2011.

  1. Dolphin

    Dolphin Senior Member

    Messages:
    6,440
    Likes:
    4,667
    Does the Heterogeneity of Chronic Fatigue Syndrome Moderate the Response

    [Apologies - I can't change the heading - this isn't to do with PACE. If a moderator or admin could correct the subject line, it'd be great.]

    Vol. 80, No. 6, 2011
    --------------------------------------------------------------------------------

    http://content.karger.com/ProdukteD...ikelNr=327582&Ausgabe=255497&ProduktNr=223864

    * I've given each sentence its own paragraph
  2. oceanblue

    oceanblue Senior Member

    Messages:
    1,174
    Likes:
    342
    UK
    Thanks, Dolphin, that look interesting. My guess is they were expecting to find factors predicting response, particularly given that PACE was explicity based on a psychological model of the illness. And this is remarkably reasonable of them:
    Will try to source a copy of the paper.
  3. Dolphin

    Dolphin Senior Member

    Messages:
    6,440
    Likes:
    4,667
    But they say just before it:
    which rather undoes any admission.

    I'm not sure I'm going to know enough about LCA or LCR to know which is "better" but we'll see how things go, I certainly plan to read it.
  4. oceanblue

    oceanblue Senior Member

    Messages:
    1,174
    Likes:
    342
    UK
    I took that as supporting it - they are saying they found nothing that differentiates patients so have no grounds to rule any patients in or out.
  5. Dolphin

    Dolphin Senior Member

    Messages:
    6,440
    Likes:
    4,667
    I took that to refer to the LCA where all were positive.
    I'm hopeful I'll have a better idea when I see the full paper (and perhaps do a little reading); it looks like it might be something a drug company would have trouble saying i.e. if they found a group that had a negative response using a certain analysis.
  6. Dolphin

    Dolphin Senior Member

    Messages:
    6,440
    Likes:
    4,667
    Anyone have the Cognitive Behavioural Response Questionnaire?

    I've read the paper now. First a preliminary comment/question:

    Has anyone got this questionnaire? I often find the concept of symptom focusing in CFS is not well-defined. For example, if one has a symptom which is not an "approved" CFS symptom or say that such a symptom bothers you, that is called symptom focusing.

    It could be that what they really found was people who had more symptoms had a poor response.
  7. Snow Leopard

    Snow Leopard Senior Member

    Messages:
    2,203
    Likes:
    1,536
    Australia
    Seriously, how hard do you think they tried to find poorly responding subgroups, given the method? (I haven't read the paper).
  8. Dolphin

    Dolphin Senior Member

    Messages:
    6,440
    Likes:
    4,667
    This is really two separate studies. I'm afraid my suspicious mind makes me wonder whether they have deliberately published like this to confuse peopl as most people are not going to feel they are that knowledgeable about latent class analysis and latent class regression.

    The first analysis is about latent class analysis. This is about breaking up a group into subgroups based on the data one has collected. This was just done on the baseline data so doesn't have really anything to do with CBT.

    In CFS, it is very important to measure interesting CFS data. What I have seen happen with CDC data on the broad empiric criteria (the current study uses the Oxford criteria) is that when one measures body mass index and various measures that could be connected to body mass index/being overweight, what one ends up at are classes like obese and non-obese. It may be true that these are classes but these are not necessarily CFS groups with different causes for their illness: they are not really interesting subgroups of CFS.

    The second analysis looked at the response to CBT on one outcome measure, the Chalder Fatigue Questionnaire (Likert scoring i.e. 0-33). This led to 4 separate classes.

    I find it disappointing that they do not give data on the 38 predictors. There could be other interesting information there e.g. they measured 27 symptoms.

    There conclusion is fair:
    I don't know how they can justify this conclusion in the abstract:
    If one found people who had an adverse drug reaction, the relevant data is requires use of the outcome data: if there is some group that looks like it can be defined that has a bad reaction, that is relevant; one doesn't go on general subgroups that are independent of the treatment (i.e. subgroups one can find whether they have the treatment or not).
  9. Dolphin

    Dolphin Senior Member

    Messages:
    6,440
    Likes:
    4,667
    We don't get much data but it looks like a lot of the individuals in class 4 (of the latent class regression analysis) scored less than 11 on the Chalder Fatigue Questionnaire as the median value is 11 (obtained by measuring the figure which is pretty spread out so any small measurement errors wouldn't alter this much).

    They could be said to be characterised by higher depression scores, lower anxiety scores and low "symptom focusing" from Table 2. I view scores of under 11 and especially under 10 as artificial scores (remember there will be a spread around any median score). Actually I think scores of 0 on any item (question) is generally likely to be artificial - one doesn't need to score less than 11 to get a 0 on one item: if one scores 2 or more on other items, it is possible (and it seems quite likely that people with a CFS diagnosis would score 2+ on some items).
  10. oceanblue

    oceanblue Senior Member

    Messages:
    1,174
    Likes:
    342
    UK
    Thanks for all the analysis.

    Harms
    Their 'let 'em all have it' conclusion might be because the poor outcome group had a median change of -1 on CFQ, which they probably classified as neutral rather than harmful.

    However, using the PACE formula for a clinically meaningful difference of 0.5SD, we could say that a change in CFQ of -4 or more represents harm (0.5 SD=3.6). On this basis, well over 25% of the 'poor outcome' group (Group 1) were 'harmed' by the treatment (since the IQR for change is -5 to +3). This probably equates to around 10% of all patients in the study.

    Unreliable findings
    Another reason for their conclusion could be their observation that their findings need replication (how refreshingly honest). As the authors have mentioned, lots of people have used techniques like LCA to identify sub-groups in CFS patients and none of them find the same groups.

    Adding to my suspicion of the Latent Class Regression grouping is the finding that the most important paramater in defining different groups is 'weight fluctuations' - far more important than factors like symptom focusing or avoidance/resting. Huh? When has anyone ever mentioned weight fluctuations as such an important factor in CFS? The authors don't even comment on this strange finding.

    I'm assuming weight fluctuations is a symptom, since BMI wasn't measured, and maybe weight fluctuations correlates strongly with BMI but we don't know. If it is just a proxy for BMI then, as you suggest, they may have just picked out a sub-group who are chronically fatigued but don't have ME, as seems to have hapened in studies using the empirical criteria.
  11. Dolphin

    Dolphin Senior Member

    Messages:
    6,440
    Likes:
    4,667
    Ok, but they don't give any explicit reason in the text. Abstracts are supposed to summarise what the text says and comments there should be backed up. Of course, if they give a reason in the text people might have seen through it.

    Well spotted. Good analysis.

    I'm still not 100% sure we have definitely seen all the factors relevant for class 1 in the LCR - we only get to see 10 out of the 38: the others perhaps might not come up if they were not that different for the some of the classes (I'm afraid my statistics knowledge is not sufficient currently to know exactly what could be at play - I wish they had just given all 38 they had looked at).

    I must not have been clear if I gave the impression I thought weight fluctuations was to do with obesity - I didn't intend in that post to talk about weight fluctuations at all as I hadn't thought it through (I'm also not sure how they defined it).

    My point about obesity was from looking at the CDC studies on their broadly defined cohort (empiric criteria, Reeves et al. 2005 i.e. I think most don't have anything close to ME/CFS). They used some objective measures - I think they were from when they were doing the exclusions. So there were some groups classed by obesity - but that was really because the measures they were using were factors associated with obesity (can't remember exact ones - but probably things like blood sugar, cholesterol, etc.). My point is that they don't give interesting information on ME/CFS heterogeneity in my view - the sort that might persuade a doctor to give one patient an antiviral (say), another a drug to suppress the immune system and a third some drug that worked on the central nervous system (or whatever).

    Also, to clarify: while looking at patients at baseline (LCA) has been used before and because of what they the measures they used, I'm not sure it has that much value doesn't mean I necessarily think LCR has no value. For example, one might find that people for whom a particular drug/treatment isn't suitable e.g. who are over 60 and overweight (or have kidney problems or liver problems or whatever - with a huge amount of the variables having no value): not necessarily a proper CFS subgroup (or something that would show up on a LCA) but still useful in knowing who not to give a treatment to. And with CFS, such information I think could give useful information e.g. I'd imagine if one looked at who responded very well to GET and became a lot better objectively and those where there wasn't a huge change and indeed with some getting worse, that might be interesting - but it might not show up if the analysis had been restricted to LCA where other factors might "overpower" the differences.

    Basically, I think one has to be careful with a LCA as it could be drowned out with a lot of "noise" - irrelevant variables. LCR analysis looks like it could be more interesting esp. if all the data is presented, ideally objective outcome measures are used and the treatments could give interesting information
    (Aside: I think a LCR on GET might give better information than CBT in CFS as CBT can be a mish-mash of different treatments and people may improve for other reasons - although that might be more important if subjective measures are used).

    Just to repeat, as I mention LCR might not give a proper CFS subgroup e.g. if people with kidney problems (for example) couldn't cope with a drug/did badly on a drug, it might not give information about subgroups with the condition. However, with broadly defined CFS, LCR on GET-type treatments would seem like it has the potential to give some information.

    Anyway, that's possibly the first paper on LCR I've seen anywhere so I'm no expert.
  12. oceanblue

    oceanblue Senior Member

    Messages:
    1,174
    Likes:
    342
    UK
    Yes, the abstract didn't tally with the paper in a few places.

    As I understand it LCR is a close cousin of LCA, adding the depenedent variable (outcome) alongside all the things that go into LCA. As Latent Class analysis gives such unreliable results, i don't trust those for LCR either, though the principle is good. Or to put it another way, I don't think the underlying 'sub-groups' are robust/meaningfull enough, at least not based on the parameters they are using and so are unlikely to give useful predictions. If they could replicate the same 4 LCR groups on an independent sample of patients, then that would then be compelling.

    I also can't work out why they bother with groups at all. If you are using many parameters (38 in this case) to predict outcome, why not do it at an indivdual level? They are effectively using:
    1. Individual patient Paramaters-> determine which sub-group
    2. Base prediction on subgroup (all patients in group get same prediction)

    You made an interesting point about why obesity has come out as a factor in previous LCA studies.

    Surely paramaters->predict outcome would be a simpler way of doing it.

    Interesting point you made about why
  13. Dolphin

    Dolphin Senior Member

    Messages:
    6,440
    Likes:
    4,667
    Thanks for your thoughts and reading through my piece above which was longer than I intended and could probably have been more concise.

    I agree that testing results would be useful. Indeed with big enough samples (this was quite big) as you probably know, some researchers divide up the sample and use (say) the last third to test the model.

    Interesting point you make (although I'm not sure I fully understand it and of course in general I'm out of my depth talking about this). What a lot of papers I see (not on LCR) do is report on individual factors/correlations initially with a table, which wasn't done here. This would be a correlation with change score here. Then they would use reduction techniques to see if there are a few overall independent factors that could be used for everyone which I think is what you are saying. Sometimes I think researchers dichotomise the scores (top half of scores and bottom half of scores) and calculate a discriminant function with weightings for variables. I don't know enough to know if there are advantages in what they did e.g. why would somebody want to know whether a person was in class 2 or 3. It would seem to make more sense if there was more than one outcome measure e.g. class 2 did well on outcome measure A, class 3 didn't do as well on outcome measure A but did well on outcome measure B.

    One interesting overall point is that LCR class 1 had the worst initial CFQ scores.

    I have a feeling given when it was published and that Peter White doesn't usually publish CBT papers that this may be referred to in the PACE Trial analysis and it may have been advantageous to publish in this order with the particular contents of this paper (just as it was to publish the normative Chalder Fatigue Questionnaire scores last year) - but maybe I'm getting too suspicious.
  14. oceanblue

    oceanblue Senior Member

    Messages:
    1,174
    Likes:
    342
    UK
    Think some bits were missing from my original post for some reason, though it may not have made sense even in it's full form :D.

    An exploratory/confirmatory approach in this study, as you suggest, could have helped demonstrate the LCR sub-groups are real, and not yet another unreproducible quirk.

    I found the results difficult too. Apparently LCR is a form of Structural Equation Modelling and I haven't had the strength to tackle that section of my Biostatistics textbook yet... What struch me, though, was the apparent strength of some many factors as predictors. Even pain, ranked 25/38 as a predictor accoriding to the Wald statistic, had p<0.001. If this is anything to go by, then when PACE data is analysed they will find many powerful predictors of outcome. Here they will be looking at the ability of individual parameters to predict outcome (rather than group membership), which should make interpretation easier.
  15. Dolphin

    Dolphin Senior Member

    Messages:
    6,440
    Likes:
    4,667
    I missed the rankings - I must read up more on statistics (it is in my to do list and I have bought some books). Well spotted:
    and the change in fatigue score was the dependent variable.

    Also, earlier I said dichotomise (used in some other analyses) would mean divide half/half but of course, it doesn't have to be done like that - it could be, for example, dividing it up so any decreases are one group with the rest being the other.
  16. oceanblue

    oceanblue Senior Member

    Messages:
    1,174
    Likes:
    342
    UK
    Not sure if they dichotomised here: apparently LCR can work with continous variables to predict a categorical outcome (group membership) so who knows what's going on. How is it that such a soft discipline as psychology uses such impenetrably difficult statistics?
  17. Dolphin

    Dolphin Senior Member

    Messages:
    6,440
    Likes:
    4,667
    No, I don't believe they dichotomised, it was just mentioned earlier when I was mentioning some of the ways outcomes could be analysed - if that had been done, one would have had two groups and one could have looked at the predictors more clearly for disimprovement. That would not have been the same analysis as looking at correlations with change.
  18. oceanblue

    oceanblue Senior Member

    Messages:
    1,174
    Likes:
    342
    UK
    Ah, I see what you mean.

See more popular forum discussions.

Share This Page