The standard errors, z-values and p-values match more-or-less with the results I posted. Keep in mind that these are even more approximate for a GLMM than for a LMM so “matching results” should be done with a rather wide tolerance.
@dmbates
I must be really dense today or something is off in my install (Win10-64).
Using Julia 1.4.2
I updated MixedModels
and ran your initial example code:
which for me yields
and not your results
I’ve gotten the same results with both the Julia CL and Atom IDE.
Do you have any suggestions what the problem might be?
I was using MixedModels#master as in
(@v1.4) pkg> add MixedModels#master
It seems that the coefficients are properly estimated but the standard deviation of the random effects and the standard errors are incorrect. @palday Do you recall if there was an issue related to weights in the released version v2.3.0? You have been doing much more on GLMMs while I have been down the automatic differentiation rabbit hole.
I was unable to reproduce the incorrect fit with MixedModels 2.3.0 – neither with nAGQ=1 nor nAGQ=3.
I tried to rule out the user hardware error by using a JuliaBox but ran into issues with adding the package 
The standard error issue that I’m aware of never made it into a release – there have been some intermittent issues with weights but I don’t think they’re related.
One thing I saw when skimming the thread is that there was one fit at the top with nAGQ > 1 and fast=true. This corresponds to a configuration that simply isn’t possible in lme4, namely AGQ but while optimizing on only the random (and not the fixed) effects. In lme4, you get the Laplace approximation (nAGQ=1) when you optimize on only the RE (confusingly with the option nAGQ=0).
A cliche’d suggestion: can you run locally, but after restarting julia? So restart julia, do the updates, then do the fitting. Don’t update while a package is loaded.
I’ve been working on a RCall compatibility for MixedModels <=> lme4 in my package JellyMe4 (it’s on the Julia registry). The latest release won’t move glmer to GeneralizedMixedModel but should be able to move Bernoulli models in the other direction (and will preserve contrasts). The funny thing there is that I’m able to reproduce @laut 's fit that way:
julia> using RCall, JellyMe4
[ Info: Precompiling RCall [6f49c342-dc21-5d91-9882-a32aef131414]
[ Info: Precompiling JellyMe4 [19ac8677-9a15-4623-9afd-84acc6165ce7]
julia> model = (m, nonzero);
julia> @rput model;
[ Info: contrasts on interaction terms are assumed to decompose into products of contrasts on the first-order terms
[ Info: (if you don't know what that means, you're probably fine)
┌ Warning: Some accuracy is lost in translation for GLMMs. This is fine with plotting, but proceed with caution for inferences.
│ You can try letting the model "reconverge" in lme4 with
│ update(model, control=glmerControl(optimizer="nloptwrap", calc.derivs=FALSE)).
└ @ JellyMe4 ~/.julia/packages/JellyMe4/Hupt7/src/glmerMod.jl:144
[ Info: lme4 handles deviance for GLMMs differently than MixedModels.jl
[ Info: for the correct comparison, examine -2*logLik(RModel) and deviance(JuliaModel)
julia> R"summary(model)"
RObject{VecSxp}
Generalized linear mixed model fit by maximum likelihood (Laplace
Approximation) [glmerMod]
Family: binomial ( logit )
Formula: Survived ~ 1 + Age + Sex + Age:Sex + (1 | Class)
Data: jellyme4_data
Control: glmerControl(optimizer = "nloptwrap", optCtrl = list(maxeval = 1),
calc.derivs = FALSE)
AIC BIC logLik deviance df.resid
49.0 54.9 -19.5 39.0 19
Scaled residuals:
Min 1Q Median 3Q Max
-2.0631 -0.7030 0.5001 0.7373 2.1385
Random effects:
Groups Name Variance Std.Dev.
Class (Intercept) 0.4006 0.6329
Number of obs: 24, groups: Class, 4
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 1.2618 0.9218 1.369 0.1710
AgeChild -0.1235 1.4687 -0.084 0.9330
SexMale -2.6101 1.2214 -2.137 0.0326 *
AgeChild:SexMale 1.8948 1.9781 0.958 0.3381
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Correlation of Fixed Effects:
(Intr) AgChld SexMal
AgeChild -0.552
SexMale -0.666 0.416
AgChld:SxMl 0.411 -0.728 -0.617
julia> m
Generalized Linear Mixed Model fit by maximum likelihood (nAGQ = 1)
Survived ~ 1 + Age + Sex + Age & Sex + (1 | Class)
Distribution: Bernoulli{Float64}
Link: LogitLink()
Deviance: 2210.0511
Variance components:
Column Variance Std.Dev.
Class (Intercept) 37.72355 6.14195
Number of obs: 24; levels of grouping factors: 4
Fixed-effects parameters:
─────────────────────────────────────────────────────────────────
Estimate Std.Error z value P(>|z|)
─────────────────────────────────────────────────────────────────
(Intercept) 1.26179 3.29347 0.383119 0.7016
Age: Child -0.123528 3.25191 -0.0379863 0.9697
Sex: Male -2.61011 1.45291 -1.79647 0.0724
Age: Child & Sex: Male 1.89476 4.19518 0.451653 0.6515
─────────────────────────────────────────────────────────────────
One thing to note is that weights don’t work in the current JellyMe4 release, so those are getting lost in translation. This suggests that something is happening with the weights in your bad run…
Just copying and pasting @dmbates code in the first answer, I’m also getting a different result in variance, Std.Dev and Std.errors estimates.
using RDatasets, MixedModels
titanic = RDatasets.dataset("datasets", "Titanic");
m = fit(GeneralizedLinearMixedModel, @formula(Survived ~ 1 + Age + Sex + (1|Class)), titanic, Binomial(), wts=float.(titanic.Freq))
m
Generalized Linear Mixed Model fit by maximum likelihood (nAGQ = 1)
Survived ~ 1 + Age + Sex + (1 | Class)
Distribution: Binomial{Float64}
Link: LogitLink()
Deviance: 2227.9565
Variance components:
Column Variance Std.Dev.
Class (Intercept) 26.841378 5.1808665
Number of obs: 32; levels of grouping factors: 4
Fixed-effects parameters:
────────────────────────────────────────────────────
Estimate Std.Error z value P(>|z|)
────────────────────────────────────────────────────
(Intercept) 1.12409 2.76454 0.406611 0.6843
Age: Child 1.04558 2.03346 0.514188 0.6071
Sex: Male -2.41575 1.16525 -2.07317 0.0382
────────────────────────────────────────────────────
I’m on MixedModels v2.3.0
I’m reasonably sure that’s the correct answer (because that’s the only one that make sense in terms of other things we know are going on).
That said, there are only 4 levels of class, so we’re already really pushing the limits of what we should be fitting. After all, we know that variance estimates (outside of a mixed models context) based on 4 observations generally won’t be great.
@palday I’ll try that now, I just tested that MixedModels doesn’t load on JuliaBox’s VM’s on
1.0.5
or Julia 1.3.1
This is my quick dirty way of testing reproducibility.
I’m not sure what’s going on there, but MixedModels 2.3.0 does have Julia 1.3 listed in its compatibility requirements, so it’s not surprising that it doesn’t run on earlier versions (in part because 1.3 has some threading and linear algebra enhancements that we take advantage of).
@palday but lme4 says that the results are statistically significant.
I think because when weighted the properly sample size is 2,201
sum(as.data.frame(Titanic)$Freq)
My guess is that the optimum isn’t well defined and that we’re seeing two local optima. I don’t have time at the moment to check the profiles (the code for this isn’t finalized in MixedModels.jl and I – and I suspect @dmbates too – have spent too much time in Julia lately to remember how to produce a profile plot quickly there). Because looking at the random effects – which is where the magic happens – we’re seeing approximately the same two answers again and again. Given that there are only 4 groups, this isn’t too surprising, but I would have to double check.
If I de-tabulate the tabulated Titanic dataset so that each row represents an individual person the results change:
I don’t know why. I understood the weights parameter to be intended for just this purpose.
See:
and:
It produces the exact same [except residuals] results on a GLM but not GLMER.
If I use the de-tabulated data in Julia:
R
write.csv(x = tidyr::uncount(data = data.frame(Titanic), weights = data.frame(Titanic)$Freq)[,-5], file = "C://Users/laut/Desktop/titanic.tsv")
Julia
titanic2 = CSV.File("C:\\Users\\laut\\Desktop\\titanic.tsv"; delim = "\t", header = true) |> DataFrame;
MixedModels.fit(GeneralizedLinearMixedModel, @formula(Survived ~ 1 + Age * Sex + (1 | Class)), titanic2, Bernoulli(), nAGQ = 1)
The problem seems to be a difference between the released version of the MixedModels package (v2.3.0) and the master branch on the repository (v3.0.0-DEV). In v2.3.0 the estimated standard deviation of the random effects is over 6 in the model with the interaction term
$ julia-1.4.2 --project
_
_ _ _(_)_ | Documentation: https://docs.julialang.org
(_) | (_) (_) |
_ _ _| |_ __ _ | Type "?" for help, "]?" for Pkg help.
| | | | | | |/ _` | |
| | |_| | | | (_| | | Version 1.4.2 (2020-05-23)
_/ |\__'_|_|_|\__'_| | Official https://julialang.org/ release
|__/ |
(tmp) pkg> st
Status `/tmp/Project.toml`
[a93c6f00] DataFrames v0.21.3
[ff71e718] MixedModels v2.3.0
[6f49c342] RCall v0.13.7
julia> using DataFrames, MixedModels, RCall
julia> titanic = filter(:Freq => !iszero, rcopy(R"as.data.frame(datasets::Titanic)"));
julia> const GLMM = GeneralizedLinearMixedModel;
julia> form = @formula Survived ~ 1 + Age*Sex + (1|Class);
julia> m = fit(GLMM, form, titanic, Bernoulli(), wts=titanic.Freq, nAGQ=3)
Generalized Linear Mixed Model fit by maximum likelihood (nAGQ = 3)
Survived ~ 1 + Age + Sex + Age & Sex + (1 | Class)
Distribution: Bernoulli{Float64}
Link: LogitLink()
Deviance: 2210.0492
Variance components:
Column Variance Std.Dev.
Class (Intercept) 37.752020 6.144267
Number of obs: 24; levels of grouping factors: 4
Fixed-effects parameters:
──────────────────────────────────────────────────────────────────
Estimate Std.Error z value P(>|z|)
──────────────────────────────────────────────────────────────────
(Intercept) 0.423055 4.02214 0.105182 0.9162
Age: Adult -1.77117 2.73242 -0.648208 0.5169
Sex: Female 0.715531 3.93831 0.181685 0.8558
Age: Adult & Sex: Female 1.89466 4.19524 0.451621 0.6515
──────────────────────────────────────────────────────────────────
whereas in v3.0.0-DEV (which requires Julia-v1.4.0 or later)
$ julia-1.4.2 --project
_
_ _ _(_)_ | Documentation: https://docs.julialang.org
(_) | (_) (_) |
_ _ _| |_ __ _ | Type "?" for help, "]?" for Pkg help.
| | | | | | |/ _` | |
| | |_| | | | (_| | | Version 1.4.2 (2020-05-23)
_/ |\__'_|_|_|\__'_| | Official https://julialang.org/ release
|__/ |
(tmp) pkg> add MixedModels#master
Updating git-repo `https://github.com/JuliaStats/MixedModels.jl.git`
Updating registry at `~/.julia/registries/General`
Updating git-repo `https://github.com/JuliaRegistries/General.git`
Resolving package versions...
Updating `/tmp/Project.toml`
[ff71e718] ↑ MixedModels v2.3.0 ⇒ v3.0.0-DEV #master (https://github.com/JuliaStats/MixedModels.jl.git)
Updating `/tmp/Manifest.toml`
[69666777] + Arrow v0.2.4
[becb17da] + Feather v0.5.6
[53afe959] + FlatBuffers v0.5.4
[ff71e718] ↑ MixedModels v2.3.0 ⇒ v3.0.0-DEV #master (https://github.com/JuliaStats/MixedModels.jl.git)
(tmp) pkg> st
Status `/tmp/Project.toml`
[a93c6f00] DataFrames v0.21.3
[ff71e718] MixedModels v3.0.0-DEV #master (https://github.com/JuliaStats/MixedModels.jl.git)
[6f49c342] RCall v0.13.7
julia> using DataFrames, MixedModels, RCall
[ Info: Precompiling MixedModels [ff71e718-51f3-5ec2-a782-8ffcbfa3c316]
julia> titanic = filter(:Freq => !iszero, rcopy(R"as.data.frame(datasets::Titanic)"));
julia> const GLMM = GeneralizedLinearMixedModel;
julia> form = @formula Survived ~ 1 + Age*Sex + (1|Class);
julia> m = fit(GLMM, form, titanic, Bernoulli(), wts=titanic.Freq, nAGQ=3)
Generalized Linear Mixed Model fit by maximum likelihood (nAGQ = 3)
Survived ~ 1 + Age + Sex + Age & Sex + (1 | Class)
Distribution: Bernoulli{Float64}
Link: LogitLink()
Deviance: 2210.0492
Variance components:
Column Variance Std.Dev.
Class (Intercept) 0.40044704 0.63280885
Number of obs: 24; levels of grouping factors: 4
Fixed-effects parameters:
────────────────────────────────────────────────────────────────
Estimate Std.Error z value P(>|z|)
────────────────────────────────────────────────────────────────
(Intercept) 0.422887 0.414341 1.02 0.3074
Age: Adult -1.77117 0.281569 -6.29 <1e-9
Sex: Female 0.715344 0.405834 1.76 0.0780
Age: Adult & Sex: Female 1.89475 0.432309 4.38 <1e-4
────────────────────────────────────────────────────────────────
Sorry that this took so long to straighten out. I guess we better get v3.0.0 released soon.
How do you install that beta version from github?
add https://github.com/JuliaStats/MixedModels.jl/blob/master/src/MixedModels.jl
The reason that I used RCall instead of RDatasets to obtain the data set is to ensure that it is a more faithful copy of the R dataset. The factor level orders are the same as those in R. If you create a CSV file then read the CSV, which I believe is what happens in RDatasets, you can’t guarantee you will get the same ordering.
1 Like
One of the big changes between 2.3 and 3.0-DEV is that fast=false fits use a two-stage optimization processes in 2.3, but a single stage in 3.0-DEV. There are a few examples where this leads to very different fits…
You can install the dev version with ]add MixedModels#master
1 Like