Last updated: Nov-06-2023

LSVCMM

Installation

You can install the development version of LSVCMM from GitHub with:

# install.packages("devtools")
devtools::install_github("fontaine618/LSVCMM")

Reference

For details on the LSVCMM method, refer to the Draft manuscript.

Example

This is a basic example which shows you how to solve a common problem:

library(LSVCMM)
instance = generate_synthetic_data()
fit = lsvcmm(
  data=instance$data, 
  response="response",
  subject="subject_id",
  time="time",
  vc_covariates="group",
  kernel=list(scale=c(0.1, 0.2, 0.3)),
  penalty=list(adaptive=0.5, penalize_intercept=T)
)

library(ggplot2)
ggplot() + 
  geom_line(
    data=fit$results,
    mapping=aes(x=penalty.lambda, y=ebich, color=as.factor(kernel.scale), group=kernel.scale)
  ) + scale_x_log10() + 
  theme_minimal() + labs(color="Kernel scale", x="Reg. parameter", y="EBIC")

i = which.min(fit$results$ebich)
t(fit$vc_path[,,i])
#>              [,1]       [,2]
#>  [1,]  0.00000000 -0.1905021
#>  [2,]  0.00000000  0.0000000
#>  [3,]  0.00000000  0.0000000
#>  [4,] -0.16977513  0.0000000
#>  [5,]  0.00000000  0.0000000
#>  [6,] -0.04898868  0.1558412
#>  [7,] -0.48785789  1.2715141
#>  [8,] -0.31260794  1.2169534
#>  [9,]  0.00000000  0.7558545
#> [10,] -0.22549099  1.0335752
#> [11,]  0.00000000  0.9595684
instance$true_values
#>    time b0        b1
#> 1   0.0  0 0.0000000
#> 2   0.1  0 0.0000000
#> 3   0.2  0 0.0000000
#> 4   0.3  0 0.0000000
#> 5   0.4  0 0.0000000
#> 6   0.5  0 0.1192029
#> 7   0.6  0 0.5000000
#> 8   0.7  0 0.8807971
#> 9   0.8  0 0.9820138
#> 10  0.9  0 0.9975274
#> 11  1.0  0 0.9996646
fit$results[i, ]
#>           llk      rss family.dispersion family.name
#> 290 -1219.911 956.1185          1.276527    gaussian
#>                    penalty.name penalty.alpha penalty.lambda penalty.adaptive
#> 290 adaptive_sparse_group_lasso             1   0.0005697523              0.5
#>     penalty.penalize_intercept link_function.name working_covariance.estimate
#> 290                       TRUE           identity                       FALSE
#>     working_covariance.ratio working_covariance.name kernel.name kernel.scale
#> 290                0.3703248       compound_symmetry    gaussian          0.3
#>     control.max_iter control.max_rounds control.rel_tol control.verbose
#> 290             1000                 50           1e-06               1
#>     control.update_method control.backtracking_fraction
#> 290                   PGD                           0.9
#>     control.two_step_estimation control.stepsize_factor      penalty df
#> 290                        TRUE                       1 2.869372e-06 12
#>     df_kernel  df_logn df_logn_kernel df_max      aic     aich      bic
#> 290  2.051598 78.72887          13.46     22 2463.822 2443.925 2518.551
#>         bich     ebic    ebich
#> 290 2453.282 2555.643 2459.623