Skip to content

aov() -> lme() conversion difficulty

1 message · Ben Bolker

#
[cc'ing back to r-sig-mixed: it's generally a good idea to keep the
list "in the loop"]
On 11-03-16 12:12 PM, Brian Edward wrote:
Binomial will exactly work here.
Unless there is something non-exchangeable about the individual tests
(e.g. you think the probability of a hit might depend on the sequence of
the test), then I don't see why you can't do the paired t-test: that is,
calculate the differences per individual in the number of two-eye and
one-eye successes and do a t test against a population mean of zero.
The t-test is reasonably robust to non-normality.  You could also do a
paired Wilcoxon test (?wilcox.test).
Hits ~ b_i + eps_person + eps_(person:numeyes) + eps_resid

  where the first term is the fixed effect of number of eyes,
the second is the among-subject effect, the third is the
among-tests-within-subject effect, and the the fourth is residual variation
(I'm still not sure if you're using total number of hits per person/eye
combination as your data (0-10), or whether you're using success on a
particular trial (0/1).  If the latter, you're violating normality
pretty badly.  If the former, then the third and fourth terms above are
completely confounded [as I suggested based on your zero residual error].)
This is actually a slightly more complicated model, because it allows
for different variances in the success with one vs two eyes.