mediate_spcma {hdmed}

Sparse Principal Component Mediation Analysis for High-Dimensional Mediators

Description

mediate_spcma applies sparse principal component mediation analysis to mediation settings in which the mediators are high-dimensional.

Usage

mediate_spcma(
  A,
  M,
  Y,
  var_per = 0.8,
  n_pc = NULL,
  sims = 1000,
  boot_ci_type = "bca",
  ci_level = 0.95,
  fused = FALSE,
  gamma = 0,
  per_jump = 0.7,
  eps = 1e-04,
  maxsteps = 2000,
  seed = 1
)

Arguments

A

length n numeric vector containing exposure variable

M

n x p numeric matrix of high-dimensional mediators.

Y

length n numeric vector containing continuous outcome variable.

var_per

a numeric variable with the desired proportion of variance explained. Default is 0.8.

n_pc

optional numeric variable with the desired number of PCs, in which case var_per is ignored. Default is NULL and the number of PCs is determined based on the desired proportion of variance explained.

sims

number of Monte Carlo draws for nonparametric bootstrap or quasi-Bayesian approximation (see mediation::mediate()). Default is 1000.

boot_ci_type

character string indicating the type of bootstrap confidence intervals for when boot = TRUE. If "bca", bias-corrected and accelerated (BCa) confidence intervals will be estimated. If "perc", percentile confidence intervals will be estimated (see mediation::mediate()). Default is "bca".

ci_level

the designated confidence level. Default 0.95.

fused

logical variable for whether the fused LASSO should be used instead of the ordinary LASSO. Default is FALSE.

gamma

numeric variable >=0 indicating the ratio of the standard LASSO penalty to the fusion penalty (see genlasso::genlasso()). Ignored if fused = FALSE. Default is 0, meaning there is no standard penalty. Larger values result in more shrinkage and sparser PCs.

per_jump

numeric value used for tuning parameter selection - the quantile cut-off for total variance change under different lambda values in the LASSO. Default is 0.7. Larger values result in more shrinkage and sparser PCs

eps

numeric variable indicating the multiplier for the ridge penalty in case X is rank deficient (see genlasso::genlasso()). Default is 1e-4.

maxsteps

an integer specifying the maximum number of steps for the algorithm before termination (see genlasso::genlasso()). Default is 2000.

seed

seed used for fitting single-mediator models after PCA

Details

mediate_spcma performs principal component mediation analysis, comparable to mediate_pcma, with the modification that the PC loadings are sparsified by a flexible LASSO penalty. This has the potential make the PCs more interpretable, since, unlike in PCA, they are only linear combinations of a subset of mediators rather than all of them. The choice of LASSO penalties is determined by the fused argument - which, when set to TRUE, deploys a fused LASSO penalty that encourages the model to give consecutive mediators similar loadings. The default is fused = FALSE, and the standard LASSO penalty is used instead of the fusion penalty. Once the sparse PCs are computed, inference proceeds exactly like in PCMA, and the PC-mediators are evaluated with methods from the mediate package.

Value

A list containing:

• loadings: a matrix of the PC loadings.

• pcs: a matrix of the PCs.

• var_explained: the cumulative proportion of variance explained by the PCs.

• contributions: a data frame containing the estimates, confidence intervals, and p-values of the mediation contributions.

• effects: a data frame containing the estimated direct, global mediation, and total effects.

Source

https://rdrr.io/github/zhaoyi1026/spcma

References

Zhao, Y., Lindquist, M. A. & Caffo, B. S. Sparse principal component based high-dimensional mediation analysis. Comput. Stat. Data Anal. 142, 106835 (2020).

Examples

A <- med_dat$A
M <- med_dat$M
Y <- med_dat$Y

# Fit SPCMA with the fused LASSO penalty while choosing the number of PCs based
# on the variance they explain. In practice, var_per and sims should be higher.
out <- mediate_spcma(A, M, Y, var_per = 0.25, fused = TRUE, gamma = 2, sims = 10)
out$effects