Skip to content

memory limit in aov

3 messages · Peter Dalgaard, Lucy Crooks

Lucy Crooks
# 
I want to do an unbalanced anova on 272,992 observations with 405  
factors including 2-way interactions between 1 of these factors and  
the other 404. After fitting only 11 factors and their interactions I  
get error messages like:

Error: cannot allocate vector of size 1433066 Kb
R(365,0xa000ed68) malloc: *** vm_allocate(size=1467461632) failed  
(error code=3)
R(365,0xa000ed68) malloc: *** error: can't allocate region
R(365,0xa000ed68) malloc: *** set a breakpoint in szone_error to debug

I think that the anova involves a matrix of 272,992 rows by 29025  
columns (using dummy variables)=7,900 million elements. I realise  
this is a lot! Could I solve this if I had more RAM or is it just too  
big?

Another possibility is to do 16 separate analyses on 17,062  
observations with 404 factors (although statistically I think the  
first approach is preferable). I get similar error messages then:

Error: cannot allocate vector of size 175685 Kb
R(365,0xa000ed68) malloc: *** vm_allocate(size=179904512) failed  
(error code=3)

I think this analysis requires a 31 million element matrix.

I am using R version 2.2.1 on a Mac G5 with 1 GB RAM running OS  
10.4.4. Can somebody tell me what the limitations of my machine (or  
R) are likely to be? Whether this smaller analysis is feasible? and  
if so how much more memory I might require?

The data is in R in a data frame of 272,992 rows by 406 columns. I  
would really appreciate any helpful input.

Lucy Crooks
Theoretical Biology
ETH Zurich
Peter Dalgaard
# 
Lucy Crooks <Lucy.Crooks at env.ethz.ch> writes:
You do not want to use aov() on unbalanced data, and especially not on
large data sets if random effects are involved. Rather, you need to
look at lmer() or just lm() if no random effects are present. 

However, even so, if you really have 29025 parameters to estimate, I
think you're out of luck. 8 billion (US) elements is 64G and R is not
able to handle objects of that size - the limit is that the size must
fit in a 32 bit integer (about 2 billion elements).

A quick calculation suggests that your factors have around 8 levels
each. Is that really necessary, or can you perhaps collapse some
levels?
Lucy Crooks
# 
Thanks for your reply.

Thanks for info on aov-hadn't been able to tell which to use from  
help pages. There are no random effects so will switch to lm().

The data are amino acid sequences, with factor being position and  
level which amino acid is present. There are indeed an average of  
around 8 per position (from 2 to 20). I don't think I can collapse  
the levels at least to start with as I don't know in advance which  
effect fitness (the y variable).

 From what you say R should be able to do the smaller analysis. So  
have increased the RAM and will try this again.

Lucy Crooks
On Feb 1, 2006, at 3:45 PM, Peter Dalgaard wrote: