Re: Repeated measures in large data set

Posted by Rich Ulrich on Nov 19, 2013; 11:54pm
URL: http://spssx-discussion.165.s1.nabble.com/Repeated-measures-in-large-data-set-tp5723115p5723177.html

[many prior attempts.]
There is pretty general disparagement for using the very-high-d.f.
tests that you can have, but I don't know if the survey people or
data-miners have come up with the proper replacements.  (Someone?)
When they do, one recommended solution will be something like this.

Given your set of data, I think that for Within comparisons (which
you seem to be talking about) I would consider using a conservative
error term constructed as follows: 

For each of the 6 samples (2 regions x 3 groups), find the 11 d.f. error
term for the linear trend across 12 months; pool these, resulting in a
66 d.f. error.   Sixty-six gives pretty good robustness.  "Deviation from
linear trend" should give a fairly practical basis for being meaningful.

One piece of pragmatic advice for large N, which has been around for a
very long time, is that you should simple ignore all tests; focus on the
effect sizes that are meaningful in some other sense.

 - When you have dozens or hundreds of tests, you can always sort them
from largest to smallest, and talk about the largest.  That gets you the
right set to talk about, anyway.  (I have seen the error where some PI
spends far-too-much time on over-interpreting some diddly 0.05 test
while he ignores various  < 0.001  results.)

--
Rich Ulrich

> Date: Sat, 16 Nov 2013 15:00:56 -0500

> From: [hidden email]
> Subject: Repeated measures in large data set
> To: [hidden email]
>
> I have a date set with approximately 1,000,000 people. Medication usage was
> recorded monthly throughout one calendar year (i.e. each person has 12 time
> points). The variables are numeric and refer to dosage.
> I'm interested in comparing use across time, between two different regions and
> three different groups. I've run Repeated Measures models with factors and
> interactions. Everything is significant because the n is so large. Is there a
> better way to do this? The differences between months are very small but all
> pairwise comparisons are significant. How do I know which are meaningful?
> (I'm particularly interested in comparing one month to the preceding and
> following months).
>
> Thanks!
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD