Metida.jl: experimental Julia LinearMixedModel software [Software]

posted by PharmCat  – Russia, 2020-10-03 21:59 (1272 d 12:24 ago) – Posting: # 21968
Views: 3,845

Hello, dear colleagues!

Some time ago I released the mixed model solution to bioequivalence for Julia (ReplicateBE.jl) and it works fine but this package is not flexible and can calculate only one kind of task. Now I try to make a general instrument for linear mixed models with some approaches that were used in ReplicateBE.jl.

Metida.jl is an experimental Julia package for fitting linear mixed model with flexible covariance structure. Main goal - to implement the ability to develop custom covariance structures and easy way for contributors to do it. Top goal - to realize the most used PROC MIXED abilities from SAS.

Now Metida.jl in the early development stage and checked on one small dataset (SAS code). Covariance type implemented: Scaled identity, Variance components, Heterogeneous Compound Symmetry.

If Julia is already installed Metida.jl can be installed very simple way:

import Pkg; Pkg.add("Metida")

Version 0.1.3 on this moment available.

To check you can use this example:

using Metida, StatsBase, StatsModels, CSV, DataFrames
df = CSV.File(dirname(pathof(Metida))*"\\..\\test\\csv\\df0.csv") |> DataFrame
categorical!(df, :subject);
categorical!(df, :period);
categorical!(df, :sequence);
categorical!(df, :formulation);

lmm = LMM(@formula(var ~ sequence + period + formulation), df;
random = VarEffect(@covstr(formulation), CSH),
repeated = VarEffect(@covstr(formulation), VC),
subject = :subject)

fit!(lmm)


There are many work to do: check other validation datasets, add other covariance types, performance things and other... So I will be very appreciate for any issues on github or here, advices ets.
It will be very cool if you guided me for any validation aspects.

Note: One of the most notable problems is the long initial start-up time. The second run of the same command is much faster. Worst of all, at the moment, changing the covariance structure also requires recompilation and takes a certain amount of time.

Some comparations with PROC MIXED SAS:

using Metida, StatsBase, StatsModels, CSV, DataFrames
df = CSV.File(dirname(pathof(Metida))*"\\..\\test\\csv\\df0.csv") |> DataFrame

# EXAMPLE 1
################################################################################
################################################################################
################################################################################
#=
PROC MIXED data=df0;
CLASSES subject sequence period formulation;
MODEL  var = sequence period formulation/ DDFM=SATTERTH s;
RANDOM  formulation/TYPE=CSH SUB=subject G V;
REPEATED/GRP=formulation SUB=subject R;
RUN;

REML: 10.06523862
=#

lmm = LMM(@formula(var ~ sequence + period + formulation), df;
random   = VarEffect(@covstr(formulation), CSH),
repeated = VarEffect(@covstr(formulation), VC),
subject  = :subject)

fit!(lmm)
#=
Linear Mixed Model: var ~ sequence + period + formulation
Random 1:
   Model: formulation
   Type: HeterogeneousCompoundSymmetry (3)
   Coefnames: ["formulation: 1", "formulation: 2"] Repeated:
   Model: formulation
   Type: VarianceComponents (2)
   Coefnames: ["formulation: 1", "formulation: 2"]
Status: converged

   -2 logREML: 10.0652

   Fixed effects:

Name             Estimate     SE
(Intercept)      1.57749      0.334543
sequence: 2      -0.170833    0.384381
period: 2        0.195984     0.117228
period: 3        0.145014     0.109171
period: 4        0.157363     0.117228
formulation: 2   -0.0791667   0.0903709

Random effects:

   θ vector: [0.455584, 0.367656, 1.0, 0.143682, 0.205657]
Random 1   formulation: 1   var   0.207557
Random 1   formulation: 2   var   0.135171
Random 1   Rho              rho   1.0
Repeated   formulation: 1   var   0.0206445
Repeated   formulation: 2   var   0.0422948
=#

# EXAMPLE 2
################################################################################
################################################################################
################################################################################
#=
PROC MIXED data=df0;
CLASSES subject sequence period formulation;
MODEL  var = sequence period formulation/ DDFM=SATTERTH s;
RANDOM  formulation/TYPE=VC SUB=subject G V;
REPEATED/GRP=formulation SUB=subject R;
RUN;

REML: 16.06148160
=#

lmm = LMM(
    @formula(var ~ sequence + period + formulation), df;
    random   = VarEffect(@covstr(formulation), SI),
    repeated = VarEffect(@covstr(formulation), VC),
    subject  = :subject,
)
fit!(lmm)
#=
Linear Mixed Model: var ~ sequence + period + formulation
Random 1:
   Model: formulation
   Type: ScaledIdentity (1)
   Coefnames: ["formulation: 1", "formulation: 2"] Repeated:
   Model: formulation
   Type: VarianceComponents (2)
   Coefnames: ["formulation: 1", "formulation: 2"]
Status: converged

   -2 logREML: 16.0615

   Fixed effects:

Name             Estimate     SE
(Intercept)      1.57212      0.305807
sequence: 2      -0.170833    0.279555
period: 2        0.204087     0.289957
period: 3        0.155769     0.11308
period: 4        0.160015     0.289957
formulation: 2   -0.0791667   0.279555

Random effects:

   θ vector: [0.412436, 0.145184, 0.220819]
Random 1   Var              var   0.170103
Repeated   formulation: 1   var   0.0210784
Repeated   formulation: 2   var   0.048761
=#

#EXAMPLE 3
################################################################################
################################################################################
################################################################################
#=
PROC MIXED data=df0;
CLASSES subject sequence period formulation;
MODEL  var = sequence period formulation/ DDFM=SATTERTH s;
RANDOM  subject/TYPE=VC G V;
RUN;

REML: 10.86212458
=#

lmm = LMM(@formula(var ~ sequence + period + formulation), df;
    random = VarEffect(@covstr(subject), SI)
    )
fit!(lmm)
#=
Linear Mixed Model: var ~ sequence + period + formulation
Random 1:
   Model: subject
   Type: ScaledIdentity (1)
   Coefnames: ["subject: 1", "subject: 2", "subject: 3", "subject: 4", "subject: 5"] Repeated:
   Model: nothing
   Type: ScaledIdentity (1)
   Coefnames: -

Status: converged

   -2 logREML: 10.8621

   Fixed effects:

Name             Estimate     SE
(Intercept)      1.61         0.309774
sequence: 2      -0.170833    0.383959
period: 2        0.144167     0.116706
period: 3        0.08         0.115509
period: 4        0.144167     0.116706
formulation: 2   -0.0791667   0.0833617

Random effects:

   θ vector: [0.410574, 0.182636]
Random 1   Var   var   0.168571
Repeated   Var   var   0.0333559
=#

#EXAMPLE 4
################################################################################
################################################################################
################################################################################
#=
PROC MIXED data=df0;
CLASSES subject sequence period formulation;
MODEL  var = sequence period formulation/ DDFM=SATTERTH s;
RANDOM  period/TYPE=VC G V;
RANDOM  formulation/TYPE=VC G V;
RUN;

REML: 25.12948063
=#
lmm = LMM(
    @formula(var ~ sequence + period + formulation), df;
    random   = [VarEffect(@covstr(period), VC), VarEffect(@covstr(formulation), VC)] )
fit!(lmm)

#=
Linear Mixed Model: var ~ sequence + period + formulation
Random 1:
   Model: period
   Type: VarianceComponents (4)
   Coefnames: ["period: 1", "period: 2", "period: 3", "period: 4"] Random 2:
   Model: formulation
   Type: VarianceComponents (2)
   Coefnames: ["formulation: 1", "formulation: 2"] Repeated:
   Model: nothing
   Type: ScaledIdentity (1)
   Coefnames: -

Status: converged

   -2 logREML: 25.1295

   Fixed effects:

Name             Estimate     SE
(Intercept)      1.61         0.249491
sequence: 2      -0.170833    0.192487
period: 2        0.144167     0.269481
period: 3        0.08         0.266717
period: 4        0.144167     0.269481
formulation: 2   -0.0791667   0.192487

Random effects:

   θ vector: [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.421717]
Random 1   period: 1        var   0.0
Random 1   period: 2        var   0.0
Random 1   period: 3        var   0.0
Random 1   period: 4        var   0.0
Random 2   formulation: 1   var   0.0
Random 2   formulation: 2   var   0.0
Repeated   Var              var   0.177845
=#

#EXAMPLE 5
################################################################################
################################################################################
################################################################################
#=
PROC MIXED data=df0;
CLASSES subject sequence period formulation;
MODEL  var = sequence period formulation/ DDFM=SATTERTH s;
RANDOM  formulation/TYPE=UN(1) SUB=subject G V;
RUN;

REML: 16.24111264
=#

lmm = LMM(@formula(var ~ sequence + period + formulation), df;
    random  = VarEffect(@covstr(formulation), VC),
    subject = :subject)
fit!(lmm)

#=
Linear Mixed Model: var ~ sequence + period + formulation
Random 1:
   Model: formulation
   Type: VarianceComponents (2)
   Coefnames: ["formulation: 1", "formulation: 2"] Repeated:
   Model: nothing
   Type: ScaledIdentity (1)
   Coefnames: -

Status: converged

   -2 logREML: 16.2411

   Fixed effects:

Name             Estimate     SE
(Intercept)      1.61         0.334718
sequence: 2      -0.170833    0.277378
period: 2        0.144167     0.289463
period: 3        0.08         0.117047
period: 4        0.144167     0.289463
formulation: 2   -0.0791667   0.277378

Random effects:

   θ vector: [0.447322, 0.367367, 0.185068]
Random 1   formulation: 1   var   0.200097
Random 1   formulation: 2   var   0.134959
Repeated   Var              var   0.0342502
=#
################################################################################


Best regards,

Complete thread:

UA Flag
Activity
 Admin contact
22,957 posts in 4,819 threads, 1,636 registered users;
94 visitors (0 registered, 94 guests [including 7 identified bots]).
Forum time: 09:23 CET (Europe/Vienna)

With four parameters I can fit an elephant,
and with five I can make him wiggle his trunk.    John von Neumann

The Bioequivalence and Bioavailability Forum is hosted by
BEBAC Ing. Helmut Schütz
HTML5