blme optimizer warnings
A couple of guesses here in addition to what Ben mentioned, but you
likely don't want a prior on the covariance of the random effects and
the weights should be on the scale of inverse variances. The following
replicates the numbers for the CCREM column from table 1:
blmer(g ~ 1 + (1 | Study) + (1 | Subscale) + (1 | Outcome:Study:Subscale),
data = meta, weights = Precision, control =
lmerControl(optimizer = "bobyqa"),
resid.prior = point(1), cov.prior = NULL)
On Wed, May 13, 2020 at 10:04 PM Sijia Huang <huangsjcc at gmail.com> wrote:
Here it is. Thanks! A demonstration and evaluation of the use of cross-classified random-effects models for meta-analysis On Wed, May 13, 2020 at 6:57 PM Ben Bolker <bbolker at gmail.com> wrote:
Can you give a more specific reference? I can't immediately guess from Fern?ndez-Castilla's google scholar page which article it is ... On 5/13/20 9:36 PM, Sijia Huang wrote: Thanks for the quick reply, Ben! I am replicating the Fern?ndez-Castilla et al. (2018) article. Below are the data they have in the article. Anything I can do to resolve the issue? Thanks!
meta
Study Outcome Subscale g Variance Precision 1 1 1 1 -0.251 0.024 41.455 2 2 1 1 -0.069 0.001 1361.067 3 3 1 5 0.138 0.001 957.620 4 4 1 1 -0.754 0.085 11.809 5 5 1 1 -0.228 0.020 49.598 6 6 1 6 -0.212 0.004 246.180 7 6 2 7 0.219 0.004 246.095 8 7 1 1 0.000 0.012 83.367 9 8 1 2 -0.103 0.006 162.778 10 8 2 3 0.138 0.006 162.612 11 8 3 4 -0.387 0.006 160.133 12 9 1 1 -0.032 0.023 44.415 13 10 1 5 -0.020 0.058 17.110 14 11 1 1 0.128 0.017 59.999 15 12 1 1 -0.262 0.032 31.505 16 13 1 1 -0.046 0.071 14.080 17 14 1 6 -0.324 0.003 381.620 18 14 2 6 -0.409 0.003 378.611 19 14 3 7 0.080 0.003 385.319 20 14 4 7 -0.140 0.003 385.542 21 15 1 1 0.311 0.005 185.364 22 16 1 1 0.036 0.005 205.063 23 17 1 6 -0.259 0.001 925.643 24 17 2 7 0.196 0.001 928.897 25 18 1 1 0.157 0.013 74.094 26 19 1 1 0.000 0.056 17.985 27 20 1 1 0.000 0.074 13.600 28 21 1 6 -0.013 0.039 25.425 29 21 2 7 -0.004 0.039 25.426 30 22 1 1 -0.202 0.001 1487.992 31 23 1 1 0.000 0.086 11.628 32 24 1 1 -0.221 0.001 713.110 33 25 1 1 -0.099 0.001 749.964 34 26 1 5 -0.165 0.000 6505.024 35 27 1 1 -0.523 0.063 15.856 36 28 1 1 0.000 0.001 1611.801 37 29 1 6 0.377 0.045 22.045 38 29 2 7 0.575 0.046 21.677 39 30 1 1 0.590 0.074 13.477 40 31 1 1 0.020 0.001 1335.991 41 32 1 1 0.121 0.043 23.489 42 33 1 1 -0.101 0.003 363.163 43 34 1 1 -0.101 0.003 369.507 44 35 1 1 -0.104 0.004 255.507 45 36 1 1 -0.270 0.003 340.761 46 37 1 1 0.179 0.150 6.645 47 38 1 2 0.468 0.020 51.255 48 38 2 4 -0.479 0.020 51.193 49 39 1 5 -0.081 0.024 42.536 50 40 1 1 -0.071 0.043 23.519 51 41 1 1 0.201 0.077 13.036 52 42 1 6 -0.070 0.006 180.844 53 42 2 7 0.190 0.006 180.168 54 43 1 1 0.277 0.013 79.220 55 44 1 5 -0.086 0.001 903.924 56 45 1 5 -0.338 0.002 469.260 57 46 1 1 0.262 0.003 290.330 58 47 1 5 0.000 0.003 304.959 59 48 1 1 -0.645 0.055 18.192 60 49 1 5 -0.120 0.002 461.802 61 50 1 5 -0.286 0.009 106.189 62 51 1 1 -0.124 0.006 172.261 63 52 1 1 0.023 0.028 35.941 64 53 1 5 -0.064 0.001 944.600 65 54 1 1 0.000 0.043 23.010 66 55 1 1 0.000 0.014 72.723 67 56 1 5 0.000 0.012 85.832 68 57 1 1 0.000 0.012 85.832 On Wed, May 13, 2020 at 6:00 PM Ben Bolker <bbolker at gmail.com> wrote:
Without looking very carefully at this: * unless your response variable is somehow already centered at zero by design, a model with no intercept at all is going to be weird/problematic (random effects are always zero-centered by definition). * is it really OK to have an infinite scale in your wishart prior? (It may be fine, I'm not immediately familiar with the blme parameterizations, it just looks weird) * the fact that your standard devs are all exactly 1 suggests that the optimizer bailed out before actually doing anything (these are the default starting values). Can you provide a reproducible example? On 5/13/20 8:53 PM, Sijia Huang wrote:
Hi everyone, I am fitting a cross-classified model with blme, but getting 1 optimizer warning. The code and output are shown below. Any suggestions regarding fixing the estimation issue? Thanks!
meta.example <- blmer(g~0+(1|Study)+(1|Subscale)+
1|Outcome:Study:Subscale), + data=meta, weights = Variance, + resid.prior = point(1), + control = lmerControl(optimizer="bobyqa"))
meta.example
Cov prior : Outcome:Study:Subscale ~ wishart(df = 3.5, scale = Inf,
posterior.scale = cov, common.scale = TRUE)
: Study ~ wishart(df = 3.5, scale = Inf, posterior.scale =
cov,
common.scale = TRUE)
: Subscale ~ wishart(df = 3.5, scale = Inf, posterior.scale
=
cov, common.scale = TRUE) Resid prior: point(value = 1) Prior dev : NaN Linear mixed model fit by maximum likelihood ['blmerMod'] Formula: g ~ 0 + (1 | Study) + (1 | Subscale) + (1 |
Outcome:Study:Subscale)
Data: meta
Weights: Variance
AIC BIC logLik deviance df.resid
Inf Inf -Inf Inf 64
Random effects:
Groups Name Std.Dev.
Outcome:Study:Subscale (Intercept) 1
Study (Intercept) 1
Subscale (Intercept) 1
Residual 1
Number of obs: 68, groups: Outcome:Study:Subscale, 68; Study, 57;
Subscale, 7
No fixed effect coefficients
convergence code 0; 1 optimizer warnings; 0 lme4 warnings
Best,
Sijia
[[alternative HTML version deleted]]
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
[[alternative HTML version deleted]]
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models