mediate_hima {hdmed}

High-Dimensional Mediation Analysis

Description

mediate_hima fits a high-dimensional mediation model with the minimax concave penalty as proposed by Zhang et al. (2016), estimating the mediation contributions of potential mediators.

Usage

mediate_hima(
  A,
  M,
  Y,
  C1 = NULL,
  C2 = NULL,
  binary_y = FALSE,
  n_include = NULL,
  ...
)

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 or binary outcome variable.

C1

optional numeric matrix of covariates to include in the outcome model.

C2

optional numeric matrix of covariates to include in the mediator model.

binary_y

logical flag for whether Y should be interpreted as a binary variable with 1/0 coding rather than as continuous. Default is FALSE.

n_include

integer specifying the number of top markers from sure independent screening to be included. Default is NULL, in which case n_include will be either ceiling(n/log(n)) if binary_Y = F, or ceiling(n/(2*log(n))) if binary_Y = T. If n_include >= p, all mediators are included with no screening. Note that if binary_y = F, screening is performed based on the single-mediator outcome model p-values, and if binary_y = F, screening is based on the the mediator model p-values.

...

other arguments passed to hdi.

Details

The first step in HIMA is to perform sure independence screening (SIS) to choose the n_include mediators that are most associated with the outcome (when Y is continuous) or the exposure (when Y is binary), based on p-values from linear regression. The second step is to fit the outcome model for the remaining mediators with the minimax concave penalty. HIMA then fits the mediator models using linear regression among those mediators that have both survived SIS (in step 1) and been selected by the MCP (in step 2), which enables estimation of the mediation contributions. The global indirect effect is estimated by summing these contributions, and the direct effect is estimated by subtracting the global indirect effect from an estimate of the total effect. We compute p-values for the mediation contributions by taking the maximum of the \alpha_a and \beta_m p-values, where the beta p-values are obtained via a second, unpenalized generalized linear model containing only the mediators selected by the MCP. We include this p-value computation so that our function replicates the behavior of the HIMA function from HIMA package, the function on which ours is based, but we caution that the beta p-values may be over-optimistic due to double-dipping, since the mediators tested in the unpenalized model are only those chosen by the penalized model. Note also that the HIMA authors apply Bonferroni correction to the final, maxed p-values to account for multiple testing, which we choose to leave up to the user. For more information, see the "HIMA" R package along with the provided reference.

Value

A list containing:

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

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

References

Zhang, H. et al. Estimating and testing high-dimensional mediation effects in epigenetic studies. Bioinformatics 32, 3150-3154 (2016).

Examples

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

# Fit hima with continuous outcome
out <- mediate_hima(A, M, Y)
head(out$contributions)
out$effects

# Fit hima with binary outcome
Y1 <- as.numeric(Y > mean(Y))
out1 <- mediate_hima(A, M, Y1, binary_y = TRUE)
head(out1$contributions)
out1$effects