Bugs item #1585977, was opened at 2006-10-27 16:06
Message generated for change (Settings changed) made by warnes
You can respond by visiting:
https://sourceforge.net/tracker/?func=detail&atid=453021&aid=1585977&group_id=48422
Please note that this message will contain a full copy of the comment thread,
including the initial issue submission, for this request,
not just the latest update.
>Category: User Education
Group: None
>Status: Closed
>Resolution: Invalid
Priority: 5
Private: No
Submitted By: Nobody/Anonymous (nobody)
>Assigned to: Gregory Warnes (warnes)
Summary: some floating point values are being altered
Initial Comment:
I am using
Python 2.3.5 (#2, Mar 6 2006, 10:12:24) [GCC 4.0.3 20060304
(prerelease) (Debian 4.0.2-10)] on linux2
I do not know which version of RPy I am using; how can I find
that information?
I also do not know which version of R I am controlling with RPy;
how can I find that?
The manually-controlled R I'm using is R version 1.0 (v2004-10
-14) on Mac OSX 10.3.9.
In a nutshell, it appears that some floating point values are being
altered, either on the way from Py to R or back.
I am calling the R function ks.test from RPy, and occasionally I get
negative p-values, which is not possible. When I perform the identical
command in R manually, I do not get the same result. Further
investigation showed other cases where the answers were different, but
I will stick with this example.
----
In R:
data844 =
c(0.98326834745,0.98371832020,0.98390990837,0.98410972477,0.
98448770184,0.98459213417,0.98838426859,0.98853557083,0.988
76263763,0.98905877724,0.98989187920)
ks.test(data844,"punif")
output:
One-sample Kolmogorov-Smirnov test
data: data844
D = 0.9833, p-value = 1.158e-09
alternative hypothesis: two.sided
----
In python:
from rpy import *
data844 =
[0.98326834745,0.98371832020,0.98390990837,0.98410972477,0.9
8448770184,0.98459213417,0.98838426859,0.98853557083,0.9887
6263763,0.98905877724,0.98989187920]
r.ks_test(data844,"punif")
output:
{'data.name': ['c(0.98326834745, 0.9837183202,
0.98390990837, 0.98410972477, ', '0.98448770184,
0.98459213417, 0.98838426859,
0.98853557083, 0.98876263763, ', '0.98905877724,
0.9898918792)'],
'alternative': 'two.sided',
'method': 'One-sample Kolmogorov-Smirnov test',
'p.value': -2.2204460492503131e-16, <-----
note difference from 1.158e-09
'statistic': {'D': 0.98326834745000002}}
----
Note that my input vector contains highly precise values. One guess is
that the data going into R is truncated to less precision. The problem
with that theory is that, mathematically, we should never expect ks.test
to produce a negative p-value. But I can't vouch for the fact that R will
never do that.
My second guess is that the result somehow gets mis-converted on the
way back to python. This would be acceptable if the sign didn't
change.
Are there any simple tests you can suggest I do that would help me
figure out where the error is introduced? I read the manual section
(manual version 0.3.3, not necessarily in sync with the installed version
here) on type conversion and I didn't see any mention of floats other
than with respect to NaN and list-vs-tuple.
I can live with some conversion error. The problem I have right now is
that it's difficult to trust the results I'm getting, so I'm trying to figure
out where the problem is, if it's something I can correct for, or detect,
or what have you.
Thanks for any help,
Bob H
----------------------------------------------------------------------
Comment By: Gregory Warnes (warnes)
Date: 2006-10-30 15:38
Message:
Logged In: YES
user_id=9316
Hi,
ks.test returns an *object* of class 'htest'. When you ask
R to print this object, it checks the value of the p.value
field before printing it. If it is negative it prints the
"<" symbol. For example
> tmp = ks.test(data844,"punif")
> tmp
One-sample Kolmogorov-Smirnov test
data: data844
D = 0.9833, p-value < 2.2e-16
alternative hypothesis: two.sided
But looking at the actual object itself you see:
> str(tmp)
List of 5
$ statistic : Named num 0.983
..- attr(*, "names")= chr "D"
$ p.value : num -2.22e-16
$ alternative: chr "two.sided"
$ method : chr "One-sample Kolmogorov-Smirnov test"
$ data.name : chr "data844"
- attr(*, "class")= chr "htest"
and asking for just the p-value component gives
> tmp$p.value
[1] -2.220446e-16
So, you need to check for negative values in your program
code and handle them appropriately.
-Greg
----------------------------------------------------------------------
Comment By: Peter (maubp)
Date: 2006-10-30 14:31
Message:
Logged In: YES
user_id=259020
Confirmed behaviour - but its not a bug in rpy!
I am using rpy 0.99.2 (if I recall correctly) with windows
XP, R 2.3.1, python 2.3
I tried several things including:
from rpy import *
data844 = [0.98326834745, 0.98371832020, 0.98390990837,
0.98410972477, 0.98448770184, 0.98459213417,
0.98838426859, 0.98853557083, 0.98876263763,
0.98905877724, 0.98989187920]
print result["p.value"]
Giving:
-2.22044604925e-016
In R, as you say:
> ks.test(data844,"punif")
One-sample Kolmogorov-Smirnov test
data: data844
D = 0.9833, p-value < 2.2e-16
alternative hypothesis: two.sided
However, try this at the R command line:
> ks.test(data844,"punif")$p.value
[1] -2.220446e-16
Clearly ks.test seems to have stored the p-value internally
as a negative number (I have no idea why), and rpy is
converting this faithfully.
i.e. You do have not found a bug in rpy. You might have
found a bug in R or perhaps an undocumented feature..
----------------------------------------------------------------------
Comment By: Nobody/Anonymous (nobody)
Date: 2006-10-27 22:35
Message:
Logged In: NO
I have some more info.
First, the R version controlled by RPy is 2.2.1. The manual version I was
using was R 2.0.1.
Second, part of the problem is that the result from ks.test is not always a
number. Some times it is what I
would call a relation. I now have my hands on a command-line version of R
2.2.1, and doing the
command ks.test(data844,"punif") gives me
...
D = 0.9833, p-value < 2.2e-16
...
RPy seems to be converting the "< 2.2e-16" result into -2.2e-16.
Bob H
----------------------------------------------------------------------
You can respond by visiting:
https://sourceforge.net/tracker/?func=detail&atid=453021&aid=1585977&group_id=48422
-------------------------------------------------------------------------
Using Tomcat but need to do more? Need to support web services, security?
Get stuff done quickly with pre-integrated technology to make your job easier
Download IBM WebSphere Application Server v.1.0.1 based on Apache Geronimo
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=120709&bid=263057&dat=121642
_______________________________________________
rpy-list mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/rpy-list
