Skip to content

Choosing best approach for a crossover experiment

2 messages · Bradley Carlson, Thierry Onkelinx

#
Hello all,


I have data from a repeated measures, crossover experiment that goes a bit
beyond my experience. I wanted to be sure I was approaching the data in the
right way. The sample size is small as this is an independent study project
with an undergraduate student, but we see some interesting patterns and I
hope it holds up to a proper analysis. Pardon the length of the
explanation: I want to be clear about what the problems are.


I'm trying to test whether snakes respond differently to certain odor cues.
We have 5 individual snakes. The first week, each snake was tested in 3
sets of tests, with each set of tests on a different day. In each set, they
were exposed to all 4 cues in a randomized order (C B A D on day 1, then A
B D C on day 2, etc.) Don't worry about order effects - that's not where
I'm going with this. I expect the same snakes' behavior to be correlated
both across days (some snakes exhibit stronger responses reliably), and I
expect their behavior in response to the 4 different cues to be correlated
within days (e.g., if they happen to be warmer on Monday, they might have a
stronger response to all 4 cues). I don't necessarily expect this
day-to-day variation to be correlated among individuals (I doubt Monday
would be a high response day for all snakes). What I care about are the
within-subjects, within-date differences between the cues (do they reliably
respond more strongly to cue A than the other cues tested on the same day?)


***My first question is what would be the proper formatting for a repeated
measures analysis of this in lmer**.* I was torn between a few different
options:


Behavior ~ Cue + (1|Subject) + (1|Date)   <--- Seems to assume that date
effects are similar for all individuals

Behavior ~ Cue + (1|Subject/Date)

Behavior ~ Cue + (1|Subject:Date)


Any advice on which of these are more appropriate for the structure of my
data?


The second issue is that I retested these same snakes in a second battery
of tests. The structure is identical, except instead of the first 4 cues (A
B C D), two cues were switched out for new ones (A B E F). I could perform
a separate analysis of the data from this battery of tests, using the best
format from the above data. However, since it is the same individual
animals, and since the A vs. B contrast is present in both batteries of
tests, it would be nice to put all the data into a single analysis. This
would strengthen the A vs. B comparisons and seems more elegant than two
separate analyses. ***Are there any reasons I couldn't do this?*** The only
thing I could think of would be that some pairwise comparisons would have
never actually been performed on the same day (e.g., C was part of the
first battery of tests, and E was part of the second), but accounting for
date effects should control for this, I think.


Finally, I was considering bootstrap analysis to generate CIs for testing
null hypotheses, as my data are hard to normalize with
transformations. ***Would
that eliminate the ability to do pairwise, post-hoc comparisons
(using lsmeans or multcomp)?*** I'd like to know which cues are different
from which. I could do this by re-running the analysis with each cue type
as the reference level - then the effects reported would be pairwise with
respect to the reference cue, but this doesn't account for multiple
comparisons.


Thank you so much in advance for taking the time to wade through this and
offer me any thoughts. I'd also be happy to send the data and have someone
put together a script that seems reasonable to them, so I can learn the
nuances of LMM a little better.


Best,

Brad


Assistant Professor of Biology

Wabash College

Crawfordsville, IN
6 days later
#
Dear Bradley,

For your first question you could consider the combination of crossed
and interaction random effects. (1|Subject) + (1|Date) +
(1|Subject:Date) That might be overkill given that you have only 4
observations for each level of Subject:Date. Therefore I'd rather go
for (1|Subject) + (1|Date).

For the second one: though not the most efficient design to test 6
treatment, it seems reasonable to me to analyse them with a single
model.

For the last question: What is hard to normalise? The response or the
residuals? Note that only the residuals are assumed to be normal.

Best regards,


ir. Thierry Onkelinx
Statisticus / Statistician

Vlaamse Overheid / Government of Flanders
INSTITUUT VOOR NATUUR- EN BOSONDERZOEK / RESEARCH INSTITUTE FOR NATURE
AND FOREST
Team Biometrie & Kwaliteitszorg / Team Biometrics & Quality Assurance
thierry.onkelinx at inbo.be
Kliniekstraat 25, B-1070 Brussel
www.inbo.be

///////////////////////////////////////////////////////////////////////////////////////////
To call in the statistician after the experiment is done may be no
more than asking him to perform a post-mortem examination: he may be
able to say what the experiment died of. ~ Sir Ronald Aylmer Fisher
The plural of anecdote is not data. ~ Roger Brinner
The combination of some data and an aching desire for an answer does
not ensure that a reasonable answer can be extracted from a given body
of data. ~ John Tukey
///////////////////////////////////////////////////////////////////////////////////////////


Van 14 tot en met 19 december 2017 verhuizen we uit onze vestiging in
Brussel naar het Herman Teirlinckgebouw op de site Thurn & Taxis.
Vanaf dan ben je welkom op het nieuwe adres: Havenlaan 88 bus 73, 1000 Brussel.

///////////////////////////////////////////////////////////////////////////////////////////



2017-12-05 17:50 GMT+01:00 Bradley Carlson <carbrae at gmail.com>: