Skip to content

aov and missing values

3 messages · Pedro J. Aphalo, Peter Dalgaard

Pedro J. Aphalo
# 
I am learning R, and although I have looked in the documentation, I may
be asking something obvious. Sorry, if that is the case. 

In a split-plot design if there is a missing subunit summary gives me a
table with two rows for the same factor, one in the error within section
and one in the section using error between units. With no data missing
the table is "normal". How does one interpret the table when data is
missing?, or is it that aov cannot cope with missing values in this
case?

Thanks in advance for any help.

Pedro.

The model I am using is of the form "y ~ A * B * C + block +
Error(A/B)", I tried the simpler model "y ~ A * B + block + Error(A)"
with the same result.

The behaviour is the same in Version 1.1.1 and Version 1.2.0 Under
development (unstable) (2000-10-11). If it matters, I am using Windows.
Peter Dalgaard
# 
"Pedro J. Aphalo" <pedro.aphalo at joensuu.fi> writes:
As a general matter, the theory for these designs is based on having
balanced data. The analysis with aov is based on an orthogonal
decompostion according to the error terms, which in the balanced case
means splitting into averages and differences of the values within a plot.
If each plot has received the same treatments, then there is no information
in the averages of the effect of treatment (and therefore it does not
appear in the between-unit stratum), but there will be if the
plots have received different sets of treatment. This will be the case
if some plots have missing values, but not only that: The error model
also becomes dubious (it essentially implies that averages have the same
variance no matter how many observations they are averages of).
Peter Dalgaard
# 
Peter Dalgaard BSA <p.dalgaard at biostat.ku.dk> writes:
Oops, sorry. It's the other way around: It implies that the variance
of single observations is different in order to get the same ratio for
the variance of averages. I.e. the average of two observations is
always half that of one observation. If the within-plot observation is
small, then this becomes clearly unrealistic.