8
$\begingroup$

I have never had problems with R crashing before.

I am using the mice package (mice 2.13) to perform multiple imputations. The code works fine on some subsets of the data, but when I run it on other subsets, R crashes (not immediately - after some time). From the output in R just before it crashes, I believe it is using the 2l.pan method of imputation (from the pan package) I have run update.packages() already.

How can I diagnose this problem ? 

Problem signature:
  Problem Event Name:   APPCRASH
  Application Name: Rgui.exe
  Application Version:  2.151.59607.0
  Application Timestamp:    4fe47a63
  Fault Module Name:    R.dll
  Fault Module Version: 2.151.59607.0
  Fault Module Timestamp:   4fe47a4e
  Exception Code:   c0000005
  Exception Offset: 0000000000032ec8
  OS Version:   6.1.7601.2.1.0.256.4
  Locale ID:    2057
  Additional Information 1: 7782
  Additional Information 2: 77823beb5887f451c3dd7ae4fe931995
  Additional Information 3: 4491
  Additional Information 4: 4491b41bf90894717964f5eef2cccd84

Update

I have managed to create a reproducible example, with data:

require(foreign)
require(mice)
require(pan)

dt.fail <- read.csv("http://goo.gl/pg8um")
dt.fail$X <- NULL

dt.fail$out <- as.factor(dt.fail$out )
dt.fail$grp<- as.factor(dt.fail$grp)
dt.fail$v1<- as.factor(dt.fail$v1)
dt.fail$v2<- as.factor(dt.fail$v2)
dt.fail$v3 <- as.factor(dt.fail$v3)
dt.fail$v7<- as.factor(dt.fail$v7)
dt.fail$v8 <- as.factor(dt.fail$v8)
dt.fail$v9 <- as.factor(dt.fail$v9)
dt.fail$v11 <- as.factor(dt.fail$v11)
dt.fail$v12 <- as.factor(dt.fail$v12)


PredMatrix <- quickpred(dt.fail)
PredMatrix['CTP',] <- c(1,-2,0,0,0,0,0,0,0,0,1,0,1,1,0,2)



impute = mice(
data=dt.fail, 
    m = 1, 
    maxit = 1,
    imputationMethod = c(
    "logreg",   # out
    "",      # grp   ----> cluster grouping factor
    "pmm",  # v1
    "polyreg",  # v2
    "logreg",   # v3
    "pmm",  # v4
    "logreg",   # v5
    "logreg",   # v6
    "polyreg",  # v7 ----> auxilliary
    "polyreg",  # v8 ----> auxilliary
    "polyreg",  # v9 ----> auxilliary
    "polyreg",  # v10 ----> auxilliary
    "",     # v11 ----> complete
    "",     # v12 ----> complete
    "2l.pan",   # CTP ----> multilevel imputation
    ""),        # const ----> needed for multilevel impuitation

predictorMatrix = PredMatrix, seed = 101
)

And for completeness, here is the predictor matrix I was using:

    .     out grp v1 v2 v3 v4 v5 v6 v7 v8 v9 v10 v11 v12 CTP const
out     0   0  0  0  0  0  1  1  0  0  0   0   0   0   0     0
grp     0   0  0  0  0  0  0  0  0  0  0   0   0   1   0     0
v1      0   0  0  0  0  0  0  0  0  0  0   0   1   1   0     0
v2      0   0  0  0  0  1  1  1  0  1  0   0   1   1   1     0
v3      0   0  0  0  0  1  1  1  0  1  1   0   1   1   1     0
v4      0   0  0  1  1  0  1  1  0  1  1   0   1   1   1     0
v5      1   1  0  0  0  0  0  1  0  1  0   0   1   0   0     0
v6      1   1  0  1  0  1  1  0  0  1  0   0   1   0   0     0
v7      0   0  0  0  0  0  1  1  0  1  0   0   0   1   0     0
v8      0   0  0  0  0  0  1  1  0  0  0   0   1   1   0     0
v9      0   0  0  0  1  1  1  1  0  1  0   0   1   1   1     0
v10     0   0  0  0  0  0  1  1  0  1  0   0   1   1   0     0
v11     0   0  0  0  0  0  0  0  0  0  0   0   0   0   0     0
v12     0   0  0  0  0  0  0  0  0  0  0   0   0   0   0     0
CTP     1  -2  0  0  0  0  0  0  0  0  1   0   1   1   0     2
const   0   0  0  0  0  0  0  0  0  0  0   0   0   0   0     0
Improve this question
$\endgroup$
6
  • $\begingroup$ The first thing I would do is see if setting options(error=dump.frames) gets you anything after the error. R might be able to write the callstack to file before bottoming out. $\endgroup$
    – Matthew Plourde
    Feb 1, 2013 at 14:50
  • $\begingroup$ I say 'might', but I would be pleasantly surprised if R did in fact manage to do so. Take a look at the 'Just-in-time debugging' section at stats.uwo.ca/faculty/murdoch/software/debuggingR . The mingw debugger it mentions might be what you need. I haven't used it, but apparently it will display a stack dump at the time of a crash. $\endgroup$
    – Matthew Plourde
    Feb 1, 2013 at 15:15
  • $\begingroup$ I'd recommend contacting the authors of the package - this sort of a crash is usually an indication something is wrong with their C code $\endgroup$
    – hadley
    Feb 1, 2013 at 17:04
  • $\begingroup$ If R reliably (heh) crashes on certain data subsets and just as reliably doesn't crash on other subsets, then you've got a good start on describing the bug situation. See if you can write any intermediate results to a file, in order to further zoom in on the data and the specific function (or sub-function) call that's blowing up. $\endgroup$
    – Carl Witthoft
    Feb 1, 2013 at 17:32
  • $\begingroup$ @hadley thanks, I left a message for Stef Van Buuren on CrossValidated in a comment to an answer he gave to me earlier asking him to check here... $\endgroup$
    – Robert Long
    Feb 1, 2013 at 18:42

2 Answers 2

Reset to default
5
$\begingroup$

I occasionally have problems with the 2l methods for large data, but have never seen R itself crash on it. My guess would be that they are related to sparse data (very small clusters). How many predictors do you have relative to cluster size?

Some suggestions:

In your data, you have several covariates that have incomplete data but that are not imputed. Please check whether mice removes them before imputation by setting maxit = 0 and inspects imp$log. If you want to use these as predictors, you should specify an imputation method for them.

The mice package does not use any own fortran or C code, but pan may (I don't know). If you are really determined to find the source of the problem, I suggest that you consult the book by Matloff, which contains chapter on advanced debugging techniques.

The obvious other route is to try to simplify the model. Remove superfluous predictors, use a flat file (e.g. pmm) with cluster allocation as a fixed factor, and check whether the intra-class correlations of the observed and impute data are similar.

The intercept term is automatically added by `mice.impute.2l.pan', so you do not need that.

Hope this helps.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thanks ! In the original data (~18000 obs) the minimum cluster size was 10 and there were 4 predictors for CTP in the predictor matrix. I am now imputing all variables and I have obtained a smaller dataset (2800 obs, 15 vars) which is still causing R to crash - I have updated the question with a reproducible example. $\endgroup$
    – Robert Long
    Feb 2, 2013 at 19:33
3
$\begingroup$

I found what is causing the crash - there was one missing value in grp (which was not being imputed). Still, it does not seem quite right that it crashes R ! After running 

dt.fail <- dt.fail[!is.na(dt.fail$grp),]

it no longer crashes, but instead generates the following error:

Error in order(dfr$group) : argument 1 is not a vector

I will post a seperate question about that.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thanks for finding out. I will address this is a future release. $\endgroup$
    – Stef van Buuren
    Apr 10, 2013 at 15:14
  • $\begingroup$ As Hadley suggested, the problem was in the compiled code, in this case the pan() function. I have added tighter error checking, so the issue is solved in mice 2.16. $\endgroup$
    – Stef van Buuren
    Apr 27, 2013 at 14:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.