mediate_lvma {hdmed}

Latent Variable Mediation Analysis

Description

mediate_lvma fits a high-dimensional mediation model described by Derkach et al. (2019), in which a small number of latent, unmeasured mediators replace the original mediators in the model.

Usage

mediate_lvma(A, M, Y, q, rhoLM, rhoEL, rhoLY, scale = TRUE, imax = 5000)

Arguments

A

length n numeric vector representing the exposure variable

M

n x p numeric matrix of high-dimensional mediators.

Y

length n numeric vector representing the continuous outcome variable.

q

number of latent mediators

rhoLM

numeric vector of candidate penalty parameters for the latent mediator-mediator associations in the joint likelihood. Default is a short toy sequence.

rhoEL

numeric vector of candidate penalty parameters for the exposure-latent mediator associations in the joint likelihood. Default is a short toy sequence.

rhoLY

numeric vector of candidate penalty parameters for the latent mediator-outcome associations in the joint likelihood. Default is a short toy sequence.

scale

logical flag for whether the inputted mediators should be standardized prior to the analysis. Default is FALSE, but TRUE may be worth attempting in case of errors.

imax

integer specifying the maximum number of iterations allowed. Default is 5000.

Details

LVMA is a latent variable mediation model which assumes, contrary to standard assumptions, that the inputted set of candidate mediators do not affect the outcome through the exposure on their own, but rather, occur as result of latent, unmeasured mediators which themselves transmit effects from the exposure to outcome. The required parameters for fitting this model are rhoLE, a regularization parameter for effects of the latent mediators on the inputted mediators; rhoEL, a regularization parameter for the effects of the exposure on the latent mediators; and rhoLY, a regularization parameter for the effects of the latent mediators on the exposure. These parameters should ideally be supplied by the user as vectors, so that each combination of the three parameters can be attempted in the estimation. However, this can be intensely computation costly, and for simplicity our default values are vectors of length 4, corresponding to a 64 by 64 parameter grid. In practice, Derkach et al. use a much larger grid with 5 values of rhoLM (ranging from 6 to 8.5), 40 values of rhoEY (ranging from 0 to 40), and 40 values of rhoLY (ranging from 0 to 75). Supplying longer parameter vectors makes the fit more flexible, but more computationally costly, and to reliably implement LVMA on real data one should use a larger parameter grid with parallel computation on a remote computing cluster, as did the authors. For more information on the likelihood, parameters, and mediation model, see the referenced article and/or its supplement files.

Value

A list containing the selected models based on AIC, BIC, and EBIC (recommended) as three sub-lists. The sub-lists include objects indicating the penalty set that was used (penalty), the values of the chosen parameters (e.g., EBIC), the exposure-latent mediator effects (AL_effects), the latent mediator-mediator effects (LM_effects, a data frame), the direct effect of the exposure on the outcome (AY_direct_effect), the the latent mediator-outcome effects (LY_effects), and binary vector indicating whether each mediator was determined to be active. Here, active mediators are those which are associated with a latent mediator that itself is associated with both A and Y.

Source

https://pubmed.ncbi.nlm.nih.gov/30859548/

References

Derkach, A., Pfeiffer, R. M., Chen, T.-H. & Sampson, J. N. High dimensional mediation analysis with latent variables. Biometrics 75, 745-756 (2019).

Examples

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

# Perform latent variable mediation analsis with 4 latent mediators and print
# whether the original 20 mediators are "actively" related to mediation
out <- mediate_lvma(A, M, Y, q = 4, rhoLM = 2, rhoEL = 2, rhoLY = 2, imax = 50)
table(out$EBIC_out$mediator_active)