[R] Re gression between adjacent columns - error with NAs
Krzysztof Sakrejda-Leavitt
krzysztof.sakrejda at gmail.com
Thu Jul 31 03:19:38 CEST 2008
A partial answer is...
After doing debug(lm.fit) and debug(lm), and waiting to see where the
messages come from, I can tell that when you do na.action=NULL, the R
goes to call the Fortran routine dqrls (through .Fortran) and when R
prepares the data to pass to the dqrls it triggers one of the error
checks. Upshot being that R decides dqrls ought not get NA's as values
to work with...
when you don't do na.action=NULL, the NA's get caught in the error
checking for lm.fit... which seems logical since the complaint describes
what you are doing: "0 (non-NA) cases".
If you're doing regression by adjacent columns, I think calling lm with
a column full of NA's as the independent variable is a mistake since lm
can't really do anything with that--your function should check that
there is a reason to call lm, maybe like so:
##Code start
SourceMat<-matrix(data=rnorm(100), ncol=10, nrow=10)
SourceMat[,3]<-c(NA)
tt<-time(SourceMat)
rownum=2
colnum=10
ResultMat<-matrix(NA, ncol=colnum, nrow=rownum)
#loop through each column in the source matrix:
for (i in 1:10)
{
sel_col<-SourceMat[col(SourceMat)==i] #selecting the correct
column in the matrix in turn
if(!all(is.na(sel_col))) {ResultMat[,i]<-coef(lm(tt~sel_col))}
}
##Code end
rcoder wrote:
> Hi Gabor,
>
> Thanks for your reply. I've written something that can be copied and pasted
> into your monitor to reproduce the error I am experiencing. Once the loop
> experiences a column full of NAs in SourceMat (column 3), it exits with
> errors, and ResultMat is only partially complete (up to column 2) with o/p
> intercept and slope results.
>
> When I include the 'na.action=NULL' statement, I get the following
> statement:
> Error in lm.fit(x, y, offset = offset, singular.ok = singular.ok, ...) :
> NA/NaN/Inf in foreign function call (arg 1)
>
> When I leave this statement out, I get the following:
> Error in lm.fit(x, y, offset = offset, singular.ok = singular.ok, ...) :
> 0 (non-NA) cases
>
> In either case, ResultMat is only filled up to column 2:
> [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10]
> [1,] 5.3611056 5.4099400 NA NA NA NA NA NA NA NA
> [2,] -0.8028985 -0.4078084 NA NA NA NA NA NA NA NA
>
>
> ##Code start
> SourceMat<-matrix(data=rnorm(100), ncol=10, nrow=10)
> SourceMat[,3]<-c(NA)
> tt<-time(SourceMat)
> rownum=2
> colnum=10
> ResultMat<-matrix(NA, ncol=colnum, nrow=rownum)
> #loop through each column in the source matrix:
> for (i in 1:10)
> {
> sel_col<-SourceMat[col(SourceMat)==i] #selecting the correct column
> in the matrix in turn
> ResultMat[,i]<-coef(lm(tt~sel_col, na.action=NULL))
> }
> ##Code end
>
> I would be grateful for any suggestions to avoid this problem.
>
> Thanks,
>
> rcoder
>
>
> rcoder wrote:
>
>> Well, in this case I don't think my original code would have helped
>> much...
>>
>> So, I've rewritten as below. I want to perform regression between one
>> column in a matrix and all other columns in the same matrix. I have a for
>> loop to achieve this, which succeeds in exporting intercept and slope
>> coefficients to a results matrix, except when a column that contains only
>> NAs is reached. Columns partially filled with NAs are handled, but the
>> code exits with errors when a single column is filled with NAs. I inserted
>> the 'na.action=NULL' statement within the lm() construct, but to no avail.
>> I would be very grateful for any advice.
>>
>>
>>> tt<-time(SourceMat)
>>> ResultMat<-matrix(NA, ncol=colnum, nrow=rownum) #creates an o/p
>>>
> template matrix
>
>> #loop through each column in the source matrix:
>>
>>> for (i in 1:5000)
>>>
>> {
>> sel_col<-[col(SourceMat)==i] #selecting the correct column in the
>> matrix in turn
>> SourceMat[,i]<-coef(lm(tt~sel_col), na.action=NULL)
>> }
>>
>> Thanks,
>>
>> rcoder
>>
>>
>> Gabor Grothendieck wrote:
>>
>>> Read the last line of every message to r-help.
>>>
>>> On Tue, Jul 29, 2008 at 6:15 PM, rcoder <mpdotbook at gmail.com> wrote:
>>>
>>>> Hi everyone,
>>>>
>>>> I am trying to apply linear regression to adjacent columns in a matrix
>>>> (i.e.
>>>> col1~col2; col3~col4; etc.). The columns in my matrix come with
>>>> identifiers
>>>> at the top of each column, but when I try to use these identifiers to
>>>> reference the columns in the regression function using rollapply(), the
>>>> columns are not recognised and the regression breaks down. Is there a
>>>> more
>>>> robust way to reference the columns I need, so that I can apply the
>>>> regression across the matrix; 'by.column', but every other column?
>>>>
>>>> Thanks,
>>>>
>>>> rcoder
>>>> --
>>>> View this message in context:
>>>> http://www.nabble.com/rolling-regression-between-adjacent-columns-tp18722392p18722392.html
>>>> Sent from the R help mailing list archive at Nabble.com.
>>>>
>>>> ______________________________________________
>>>> R-help at r-project.org mailing list
>>>> https://stat.ethz.ch/mailman/listinfo/r-help
>>>> PLEASE do read the posting guide
>>>> http://www.R-project.org/posting-guide.html
>>>> and provide commented, minimal, self-contained, reproducible code.
>>>>
>>>>
>>> ______________________________________________
>>> R-help at r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-help
>>> PLEASE do read the posting guide
>>> http://www.R-project.org/posting-guide.html
>>> and provide commented, minimal, self-contained, reproducible code.
>>>
>>>
>>>
>>
>
>
More information about the R-help
mailing list