[R-sig-ME] c++ exception (unknown reason) when using an offset of the slope with glmer

Ben Bolker bbolker at gmail.com
Thu Jan 29 16:00:57 CET 2015


-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

  The intercept seems awfully extreme, but I guess that's basically
because a dose of zero is actually unrealistic.

  I got a little carried away and explored this in

  http://rpubs.com/bbolker/glmer_offset

The bottom line is that I think you can work around the offset issues
if necessary (although I agree that it does technically constitute a
bug in lme4; I will post an issue at
https://github.com/lme4/lme4/issues when I get around to it, or
someone else would be welcome to), but that a GLMM actually seems like
overkill for this problem.

  cheers
    Ben Bolker

On 15-01-28 11:10 AM, xavier.paoletti at curie.fr wrote:
> Thank you very much. Probabilities of event range approximately
> between .05 to .60
> 
> You are right, If I choose an offset of the slope=0.1, I can obtain
>  estimates of the intercept.
> 
> The offset of 1.5 came from the expected increase in the risk of
> event when escalating the dose from 4 to 6.
> 
> If I fit the model without offset, I get the following etimates for
> the fixed effects intercept:: -12.7 dose     : 2.3
> 
> 
> Finally, the reason for choosing an offset is to reduce the
> dimensionality of the model due to the sampling matrix. I work on
> an extension of phase I dose escalation design in oncology, where
> the proportion of data that is sampled at one or 2 dose levels 
> increases with the overall sample size. Therefore after 30, 40, 50
>  patients, the contribution of this dose level to the likelihood is
>  massive. Esimating both the intercept and the slope of the
> dose-response relationship gets useless or even misleading.
> 
> 
> I am not sure to understand why offset of the dose = 1.5 is
> misleading for the intercept estimate, but I will dig in .
> 
> Thanks again for your help
> 
> Xavier
> 
> 
> 
> 
> 
> Ben Bolker <bbolker at gmail.com> 28/01/2015 15:45
> 
> A <xavier.paoletti at curie.fr> cc "r-sig-mixed-models at r-project.org"
> <r-sig-mixed-models at r-project.org> Objet Re: [R-sig-ME] c++
> exception (unknown reason) when using an offset of the slope with
> glmer
> 
> 
> 
> 
> 
> 
> Confirmed on
> 
> R Under development (unstable) (2015-01-26 r67627) Platform:
> i686-pc-linux-gnu (32-bit) lme4 1.1.8
> 
> I will see what I can figure out.  I suspect the main problem is
> that the doses range from 4 to 6, so with an offset of (1.5*dose),
> that says that the logit-probability or log-odds should range from
> 6 to 9, which corresponds to a baseline probability of 0.997 to
> 0.999.  Those are very high probabilities: they're going to make it
> very hard to make a sensible model.   Can you say a little bit more
> about what you're trying to do/why an offset of 1.5 makes sense?
> 
> Ben Bolker
> 
> 
> 
> On 15-01-27 08:24 AM, xavier.paoletti at curie.fr wrote:
>> Thanks for your attention.
> 
>> Here there are. 180 rows, 4 columns + obs number. Obs    patid 
>> cycle dose DLT 1       1     1  4.1   0 2       2     1  4.8   0
>> 3 3     1  5.3   0 4       4     1  5.7   0 5       5     1  6.0
>> 1 6       6     1  5.7   1 7       7     1  5.7   0 8       8
>> 1 5.3   0 9       9     1  5.3   0 10     10     1  5.7   0 11
>> 11 1  5.3   0 12     12     1  5.3   0 13     13     1  5.3   0
>> 14 14     1  5.3   0 15     15     1  5.7   0 16     16     1
>> 5.3 0 17     17     1  5.3   0 18     18     1  5.3   0 19     19
>> 1 5.3   0 20     20     1  5.7   0 21     21     1  5.7   1 22
>> 22 1  5.7   1 23     23     1  5.7   0 24     24     1  5.3   0
>> 25 25     1  5.3   0 26     26     1  5.3   0 27     27     1
>> 5.3 1 28     28     1  5.7   0 29     29     1  5.7   1 30     30
>> 1 5.7   0 31      1     2  4.1   0 32      2     2  4.8   0 33
>> 3 2  5.3   0 34      4     2  5.7   0 35      5     2  6.0   0
>> 36 6     2  5.7   0 37      7     2  5.7   1 38      8     2  5.3
>> 0 39      9     2  5.3   0 40     10     2  5.7   0 41     11
>> 2 5.3   0 42     12     2  5.3   0 43     13     2  5.3   0 44
>> 14 2  5.3   0 45     15     2  5.7   0 46     16     2  5.3   0
>> 47 17     2  5.3   0 48     18     2  5.3   0 49     19     2
>> 5.3 0 50     20     2  5.7   0 51     21     2  5.7   0 52     22
>> 2 5.7   0 53     23     2  5.7   1 54     24     2  5.3   0 55
>> 25 2  5.3   0 56     26     2  5.3   0 57     27     2  5.3   0
>> 58 28     2  5.7   0 59     29     2  5.7   0 60     30     2
>> 5.7 1 61      1     3  4.1   0 62      2     3  4.8   0 63      3
>> 3 5.3   0 64      4     3  5.7   0 65      5     3  6.0   1 66
>> 6 3  5.7   1 67      7     3  5.7   1 68      8     3  5.3   0
>> 69 9     3  5.3   0 70     10     3  5.7   0 71     11     3  5.3
>> 0 72     12     3  5.3   0 73     13     3  5.3   0 74     14
>> 3 5.3   0 75     15     3  5.7   0 76     16     3  5.3   0 77
>> 17 3  5.3   0 78     18     3  5.3   0 79     19     3  5.3   1
>> 80 20     3  5.7   0 81     21     3  5.7   0 82     22     3
>> 5.7 0 83     23     3  5.7   1 84     24     3  5.3   0 85     25
>> 3 5.3   0 86     26     3  5.3   1 87     27     3  5.3   0 88
>> 28 3  5.7   0 89     29     3  5.7   0 90     30     3  5.7   1
>> 91 1     4  4.1   0 92      2     4  4.8   0 93      3     4  5.3
>> 0 94      4     4  5.7   0 95      5     4  6.0   0 96      6
>> 4 5.7   0 97      7     4  5.7   0 98      8     4  5.3   0 99
>> 9 4  5.3   0 100    10     4  5.7   0 101    11     4  5.3   1
>> 102 12     4  5.3   1 103    13     4  5.3   0 104    14     4
>> 5.3 0 105    15     4  5.7   1 106    16     4  5.3   0 107    17
>> 4 5.3   0 108    18     4  5.3   0 109    19     4  5.3   0 110
>> 20 4  5.7   0 111    21     4  5.7   0 112    22     4  5.7   0
>> 113 23     4  5.7   0 114    24     4  5.3   0 115    25     4
>> 5.3 0 116    26     4  5.3   0 117    27     4  5.3   0 118    28
>> 4 5.7   0 119    29     4  5.7   0 120    30     4  5.7   0 121
>> 1 5  4.1   0 122     2     5  4.8   0 123     3     5  5.3   0
>> 124 4     5  5.7   0 125     5     5  6.0   0 126     6     5
>> 5.7   1 127     7     5  5.7   0 128     8     5  5.3   1 129
>> 9     5 5.3   0 130    10     5  5.7   0 131    11     5  5.3   0
>> 132    12 5  5.3   0 133    13     5  5.3   1 134    14     5
>> 5.3   0 135 15     5  5.7   0 136    16     5  5.3   0 137    17
>> 5  5.3 0 138    18     5  5.3   0 139    19     5  5.3   0 140
>> 20     5 5.7   0 141    21     5  5.7   1 142    22     5  5.7
>> 0 143    23 5  5.7   0 144    24     5  5.3   0 145    25     5
>> 5.3   0 146 26     5  5.3   0 147    27     5  5.3   0 148    28
>> 5  5.7 0 149    29     5  5.7   0 150    30     5  5.7   0 151
>> 1     6 4.1   0 152     2     6  4.8   0 153     3     6  5.3   0
>> 154     4 6  5.7   1 155     5     6  6.0   1 156     6     6
>> 5.7   0 157 7     6  5.7   0 158     8     6  5.3   0 159     9
>> 6  5.3   0 160    10     6  5.7   1 161    11     6  5.3   0 162
>> 12     6 5.3   0 163    13     6  5.3   0 164    14     6  5.3
>> 0 165    15 6  5.7   0 166    16     6  5.3   0 167    17     6
>> 5.3   0 168 18     6  5.3   0 169    19     6  5.3   0 170    20
>> 6  5.7 1 171    21     6  5.7   1 172    22     6  5.7   0 173
>> 23     6 5.7   0 174    24     6  5.3   0 175    25     6  5.3
>> 0 176    26 6  5.3   1 177    27     6  5.3   0 178    28     6
>> 5.7   0 179 29     6  5.7   1 180    30     6  5.7   0
> 
> 
> 
> 
>> Ben Bolker <bbolker at gmail.com> 27/01/2015 14:14
> 
>> A <xavier.paoletti at curie.fr> cc
>> "r-sig-mixed-models at r-project.org" 
>> <r-sig-mixed-models at r-project.org> Objet Re: [R-sig-ME] c++ 
>> exception (unknown reason) when using an offset of the slope
>> with glmer
> 
> 
> 
> 
> 
> 
>> Your posted data set got removed by the mailing list machinery. 
>> Can you post it somewhere publicly accessible?
> 
> 
>> On Tue, Jan 27, 2015 at 3:45 AM,  <xavier.paoletti at curie.fr> 
>> wrote:
>>> 
>>> Dear all,
>>> 
>>> I am new on this forum and I hope my request follows the right 
>>> format.
>>> 
>>> I use R 3.1.2 and lme1.7 on a mac OS X (snow leopard) or
>>> Windows OS.
>>> 
>>> I try to fit a longitudinal logistic mixed effect model on 
>>> dose-time response data where the response is measured several 
>>> times at the same dose. The probability of response increases 
>>> with the dose. There is a set of discrete doses (let's say 6)
>>> but most of the data are measured at 1 or 2 doses.
>>> 
>>> I use a very simple logstic model with a random intercept and
>>> the dose effect. In the simplest case, there is not time
>>> effect. Furthermore, I would like to set the slope of the dose
>>> to some value
>> using
>>> offset slope=1.5.
>>> 
>>> The command line, glmer(DLTb ~ offset(slope*dose) + (1 |
>>> patid), family=binomial,data=dataAllCRM,nAGQ=10) gives the
>>> following error: Under Mac OSX: (function (fr, X, reTrms,
>>> family, nAGQ = 1L, verbose =
>> 0L,
>>> control = glmerControl(),  : c++ exception (unknown reason)
>>> 
>>> Under Windows: Error: (maxstephalfit) PIRLS step-halvings
>>> failed to
>> reduce
>>> deviance in pwrssUpdate
>>> 
>>> Whatever the value of the offset, I get the same error. If I 
>>> remove the offset or if I remove the variable, some estimates 
>>> are obtained.
>>> 
>>> Please find attached an example of dataset as an illustration;
>>> I get the same error for all tested datasets.
>>> 
>>> 
>>> Thank you very much for your help.
>>> 
>>> Best regards,
>>> 
>>> Xavier _______________________________________________ 
>>> R-sig-mixed-models at r-project.org mailing list 
>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> 
> 
> 
> 
> 

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)

iQEcBAEBAgAGBQJUykspAAoJEOCV5YRblxUHfHMIAKYw44W2SDYxy8sy/VUHlaNE
V1EcmfYu8wS8JH1bfk4nyi71+kvG9DZMdQL2SDBZAHeJo4iu2k4Burqdz+y01KaO
8n2X0oFAzJQfyP3vQJfEf2a90BwFFaUx2kTXuDq64a3sm8hGk6mtwYtMzzvQFamS
Xl9IKdCJNrelf9d4WkP0l9ZYzx7R1QABU7/i22YUoGmWygnR6bzqS/m1oyxqijSq
stoEZVAinl3ADK0zlnByuxBiPVev+swrYbG68GNzlk/a9Oi8gB0lVPvxdh5tiKB4
Ulpu7cbMO8xlQYR+H+SaDyi9NSxPuQat14iFLcmX82Yzf0RlEq0Vri+c4HCikck=
=/Msf
-----END PGP SIGNATURE-----



More information about the R-sig-mixed-models mailing list