Skip to content

Warning message after "unscaling" predictor variable

3 messages · Rachael Mady, Ben Bolker, Dimitris Rizopoulos

Original
Rachael Mady
# 
Hello,

I have been successfully using the glmmTMB package, but have come across a
Warning today that I cannot solve. The data and the code are provided below
this posting.

I have been running mixed models with julian date scaled and centered (such
that it is a z-score; julian2 in the data). Today, I tried to run the same
models without julian date scaled and centered and received this warning:

*Warning message:*
*In nlminb(start = par, objective = fn, gradient = gr, control =
control$optCtrl) :*
*  NA/NaN function evaluation*

To recreate, see the code below. Mod1 is with julian not scaled/centered
and mod2 is with julian scaled.

*The data and code to reproduce situation: *

data <-
structure(list(sum.50 = c(2L, 1L, 2L, 0L, 0L, 7L, 0L, 6L, 1L,
0L, 3L, 8L, 1L, 0L, 2L, 7L, 0L, 0L, 1L, 3L, 2L, 0L, 8L, 9L, 6L,
1L, 8L, 8L, 0L, 5L, 0L, 0L, 5L, 3L, 1L, 5L, 2L, 0L, 0L, 2L, 7L,
0L, 0L, 7L, 1L, 0L, 5L, 8L, 5L, 3L, 0L, 4L, 8L, 2L, 7L, 0L, 2L,
7L, 0L, 1L, 12L, 5L, 0L, 14L, 0L, 5L, 5L, 2L, 6L, 0L, 3L, 1L,
0L, 4L, 5L, 1L, 0L, 3L, 9L, 1L, 13L, 0L, 5L, 7L, 8L, 5L, 0L,
9L, 11L, 0L, 0L, 4L, 3L, 0L, 4L, 7L, 7L, 7L, 0L, 1L, 5L, 1L),
    trtmt_simple = structure(c(3L, 1L, 2L, 3L, 1L, 2L, 1L, 2L,
    3L, 1L, 2L, 3L, 2L, 3L, 1L, 2L, 3L, 1L, 1L, 2L, 3L, 1L, 2L,
    3L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 2L, 3L,
    1L, 2L, 3L, 1L, 1L, 2L, 3L, 1L, 2L, 3L, 2L, 3L, 1L, 2L, 3L,
    1L, 3L, 1L, 2L, 2L, 3L, 1L, 2L, 3L, 1L, 3L, 1L, 2L, 3L, 1L,
    2L, 1L, 2L, 3L, 1L, 2L, 3L, 3L, 1L, 2L, 3L, 1L, 2L, 1L, 2L,
    3L, 2L, 3L, 1L, 2L, 3L, 1L, 1L, 2L, 3L, 1L, 2L, 3L, 2L, 3L,
    1L, 2L, 3L, 1L), .Label = c("control", "constant", "pulsed"
    ), class = "factor"), site = structure(c(3L, 4L, 5L, 3L,
    4L, 5L, 8L, 9L, 6L, 8L, 9L, 6L, 1L, 2L, 7L, 1L, 2L, 7L, 4L,
    5L, 3L, 4L, 5L, 3L, 6L, 8L, 9L, 6L, 8L, 9L, 2L, 7L, 1L, 2L,
    7L, 1L, 9L, 6L, 8L, 9L, 6L, 8L, 7L, 1L, 2L, 7L, 1L, 2L, 5L,
    3L, 4L, 5L, 3L, 4L, 6L, 8L, 9L, 1L, 2L, 7L, 1L, 2L, 7L, 3L,
    4L, 5L, 3L, 4L, 5L, 8L, 9L, 6L, 8L, 9L, 6L, 2L, 7L, 1L, 2L,
    7L, 1L, 4L, 5L, 3L, 9L, 6L, 8L, 9L, 6L, 8L, 7L, 1L, 2L, 7L,
    1L, 2L, 5L, 3L, 4L, 5L, 3L, 4L), .Label = c("bakh", "icel",
    "lid1", "lid2", "lid3", "mtpl", "nemi", "yb01", "yb02"), class =
"factor"),
    julian = c(12L, 12L, 12L, 14L, 14L, 14L, 19L, 19L, 19L, 21L,
    21L, 21L, 26L, 26L, 26L, 28L, 28L, 28L, 33L, 33L, 33L, 35L,
    35L, 35L, 40L, 40L, 40L, 42L, 42L, 42L, 47L, 47L, 47L, 49L,
    49L, 49L, 61L, 61L, 61L, 63L, 63L, 63L, 68L, 68L, 68L, 70L,
    70L, 70L, 75L, 75L, 75L, 77L, 77L, 77L, 84L, 84L, 84L, 89L,
    89L, 89L, 91L, 91L, 91L, 96L, 96L, 96L, 98L, 98L, 98L, 103L,
    103L, 103L, 105L, 105L, 105L, 110L, 110L, 110L, 112L, 112L,
    112L, 119L, 119L, 119L, 124L, 124L, 124L, 126L, 126L, 126L,
    131L, 131L, 131L, 133L, 133L, 133L, 138L, 138L, 138L, 140L,
    140L, 140L), julian2 = structure(c(-1.60484310158565,
-1.60484310158565,
    -1.60484310158565, -1.55457625178932, -1.55457625178932,
    -1.55457625178932, -1.42890912729851, -1.42890912729851,
    -1.42890912729851, -1.37864227750218, -1.37864227750218,
    -1.37864227750218, -1.25297515301137, -1.25297515301137,
    -1.25297515301137, -1.20270830321504, -1.20270830321504,
    -1.20270830321504, -1.07704117872422, -1.07704117872422,
    -1.07704117872422, -1.0267743289279, -1.0267743289279,
-1.0267743289279,
    -0.901107204437083, -0.901107204437083, -0.901107204437083,
    -0.850840354640757, -0.850840354640757, -0.850840354640757,
    -0.725173230149941, -0.725173230149941, -0.725173230149941,
    -0.674906380353615, -0.674906380353615, -0.674906380353615,
    -0.373305281575658, -0.373305281575658, -0.373305281575658,
    -0.323038431779332, -0.323038431779332, -0.323038431779332,
    -0.197371307288516, -0.197371307288516, -0.197371307288516,
    -0.14710445749219, -0.14710445749219, -0.14710445749219,
    -0.0214373330013746, -0.0214373330013746, -0.0214373330013746,
    0.0288295167949516, 0.0288295167949516, 0.0288295167949516,
    0.204763491082093, 0.204763491082093, 0.204763491082093,
    0.330430615572909, 0.330430615572909, 0.330430615572909,
    0.380697465369235, 0.380697465369235, 0.380697465369235,
    0.506364589860051, 0.506364589860051, 0.506364589860051,
    0.556631439656377, 0.556631439656377, 0.556631439656377,
    0.682298564147192, 0.682298564147192, 0.682298564147192,
    0.732565413943518, 0.732565413943518, 0.732565413943518,
    0.858232538434334, 0.858232538434334, 0.858232538434334,
    0.90849938823066, 0.90849938823066, 0.90849938823066, 1.0844333625178,
    1.0844333625178, 1.0844333625178, 1.21010048700862, 1.21010048700862,
    1.21010048700862, 1.26036733680494, 1.26036733680494, 1.26036733680494,
    1.38603446129576, 1.38603446129576, 1.38603446129576, 1.43630131109209,
    1.43630131109209, 1.43630131109209, 1.5619684355829, 1.5619684355829,
    1.5619684355829, 1.61223528537923, 1.61223528537923, 1.61223528537923
    ), .Dim = c(102L, 1L), "`\`scaled:center\``" = 75.8529411764706,
"`\`scaled:scale\``" = 39.7876534555816)), row.names = c(NA,
-102L), class = "data.frame")

# library
library(glmmTMB)

# model
mod1 <- glmmTMB(sum.50 ~ trtmt_simple + julian + (1|site), data = data,
ziformula=~1, family=nbinom1)
summary(mod1)

mod2 <- glmmTMB(sum.50 ~ trtmt_simple + julian2 + (1|site), data = data,
ziformula=~1, family=nbinom1)
summary(mod2)

Thank you very much in advance for your help!

Cheers,
Rachael
Ben Bolker
# 
Quick answer (1): if you get the same likelihood (or very similar,
e.g. difference < 1e-3) for both models, then it should be safe to
disregard the warning.

  (2) If you want to double-check that you're really getting equivalent
results, you can try unscaling the parameters "by hand": this is covered
in the following StackOverflow questions:

https://stackoverflow.com/questions/23642111/how-to-unscale-the-coefficients-from-an-lmer-model-fitted-with-a-scaled-respon/23643740#23643740

https://stackoverflow.com/questions/24268031/unscale-and-uncenter-glmer-parameters

  It would take a few minutes longer than I have right now to dig in and
see what causes the warning (i.e., what combination of
correlation/difference in parameter scales is actually causing the
problem).  My offhand guess would be that it's the centering, not the
scaling, that's important here, but I could be wrong.
On 2019-05-07 5:17 p.m., Rachael Mady wrote:
Dimitris Rizopoulos
# 
You could also give a try in GLMMadaptive 
(https://drizopoulos.github.io/GLMMadaptive/) that fits the same model 
using the adaptive Gaussian quadrature instead on the Laplace 
approximation. The equivalent code is:

fm <- mixed_model(sum.50 ~ trtmt_simple + julian, random = ~ 1 | site,
                   data = data, family = zi.negative.binomial(),
                   zi_fixed = ~ 1)
summary(fm)

For more examples check here:

https://drizopoulos.github.io/GLMMadaptive/articles/ZeroInflated_and_TwoPart_Models.html

https://drizopoulos.github.io/GLMMadaptive/articles/Goodness_of_Fit.html

Best,
Dimitris
On 5/7/2019 11:17 PM, Rachael Mady wrote: