Skip to content

Roadmap for selecting an approach to analyzing repeated measures data

2 messages · Frank E Harrell Jr, Doran, Harold

Frank E Harrell Jr
# 
Dear Group,

At http://biostat.mc.vanderbilt.edu/tmp/summary.pdf I have put a draft 
of a roadmap for choosing a method for analyzing serial (longitudinal) 
data.  If anyone has feedback about this, including adding criteria for 
judging methods that I may have missed, I would appreciate hearing from you.

Thanks
Frank
Doran, Harold
# 
Thanks for this, Frank. Quick comment (decided to put on list rather
than just send to you directly, hope that's OK). Under mixed models, you
might consider creating two smaller columns, one for the fixed effects
and another for the random effects. Under the random effects, you might
consider checking "biased in general" given that the BLUPs are biased,
but have smaller mean squared error.

Bias I think gives a slightly misleading perception about these
estimates, but maybe something else can be added as an additional row
that might counterbalance the "bias" issue should you choose to check
it. 

Under footnote 'h', I think you mean "unlike". However, why do you say
it doesn't use "standard maximum likelihood" methods? This led me to
ask, what is a standard ML method anyway and why would the methods used
in mixed models packages not be "standard" (e.g., is EM not a standard
method?). Since you are referring to the use of the LRT statistic, I
think maybe you're talking about REML, in which case the fixed effects
are integrated out using a uniform prior. So, should it read
"marginalize the fixed effects" or "marginalize the random effects" (I
don't know, just asking) since we have the marginal distribution of the
random effects after integrating out the fixed effects.

I don't know why you might say it's hard to get CL for mixed models.
Again, there are multiple "things' to get CLs for, fixed effects, random
effects, and some packages provide CLs for the marginal variance
components. Now, the latter are indeed hard to get when the distribution
is not symmetric, in which case MCMC methods like those using MCMCsamp
can illustrate. But, the first two I don't think are hard, are they? We
can use the asymptotic standard errors of the fixed effects to generate
the CLs and the conditional variances of the random effects for CLs of
the BLUPs.

Maybe add a row for "assumes constant variance over time" since some of
the methods can account for non-constant variance over time.

HTH,
Harold