Discussion: View Thread

  • 1.  Hypothesis Testing of RSA

    Posted 09-26-2012 11:40
    Hello everyone

    Has anyone in the group familiar with looking at Respiratory Sinus Arryhthmia as an outcome?
    In the output that I have, this is the log transformed variance of the heart rate.
    One of the papers that I looked at, used the log transformed variance as the dependent variable in a linear regression model. I have never dealth with this type of data before and so was a little taken aback  at fitting a regression model to a variance.
     Any thought on the appropriate analytical model to look at RSA across time would be helpful.

    Thanks

    Sree

    -------------------------------------------
    Sreelatha Meleth
    Senior Research Statistician
    RTI International
    -------------------------------------------


  • 2.  RE:Hypothesis Testing of RSA

    Posted 09-26-2012 14:27


    -------------------------------------------
    Raymond Hoffmann
    Professor
    Medical College of Wisconsin
    -------------------------------------------

    While I haven't looked at RSA as an outcome, heart rhythms should be stable (variance near 0) with high values indicationg considerable risk.  So this is a good measure of an arhythmia severity.  Thus a gamma GLM with a lo would be my first choice (make sure you look at the goodness of fit, it could be log normal) and the corresponding GEE, generalized estimating equation, or a GLMM, generalized linear mixed model for the repeated observations.  Usually a log link is chosen as the default, but you might use an identity ling since it is already log transformed or untransform the data.  Look carefully at the actual transform; there probably won't be any zero's, but of course the log of zero or near zero is not well behaved (- infinity!).  The GEE is a marginal model while the GLMM is a individual level model, so it depends on which is more appropriate to what you are modeling.

    Again it also depends on the number of subjects and the number of follow-ups (numerical stability of the algorithms), so it would be very helpful if you described the problem in more detail.  Clearly there are a lot of choices along the way, e.g. the variance-covariance model, whether the time points are regular or irregular, etc.that have substantial effects on the model choice. 

    Among the better References for this are:
    SAS's book on Mixed models by Littell et al,
    Stata's book by Hardin and Hilbe or the one (I haven't had a chance to review the new two volume version) by Rabe-Hesketh.
    Fitmaurice and Laird's book on Longitudinal Models and
    Applied Mixed Models by Brown and Prescott

    All of these cover the range of problems with good examples and explanations and enough, but not too much, theory.

    Ray






  • 3.  RE:Hypothesis Testing of RSA

    Posted 09-26-2012 15:32
    Actually, in the fetal/maternal health literature, stable fetal heart rates (variance near 0) are considered a sign of sickness, and variable fetal heart rates are considered a sign of health...although it does depend in part on what behavioral state the fetus is in.  About the variance of heart rate in adults, I can only conjecture, but I conjecture it's pretty much the same story.

    -------------------------------------------
    Eric Siegel
    Biostatistician
    Univ of Arkansas for Medical Sciences
    -------------------------------------------





  • 4.  RE:Hypothesis Testing of RSA

    Posted 09-26-2012 15:38
    Yes - In adults - higher heart rate variability is considered a sign of good vagal tone and a good sign.

    -------------------------------------------
    Sreelatha Meleth
    RTI International
    -------------------------------------------








  • 5.  RE:Hypothesis Testing of RSA

    Posted 09-26-2012 17:35
    Dear Sree,

    The proposal is similar in concept to the Levene's Test (which has a certain simplicity and the flexibility to test not only the omnibus F-test but contrasts for specific hypotheses such as dose-related trends).  Also, for your particular problem, if you have Holter data (which is very dense) then you can study entire distributions.  In my experience, the potential of the available Holter data is rarely explored.  Lastly, if you are interested in the temporal aspect within a day (rather than a fixed testing window across days) then you will also want to consider modeling or controlling the circadian rhythm.

    Best regards,
    David

    -------------------------------------------
    David Reasner
    Albemarle Scientific Consulting LLC
    -------------------------------------------