metest: Mediation analysis

In RomanHornung/survmediation: Mediation analysis for (high-dimensional molecular) survival data

Mediation analysis

Description

Univariately test for mediation effect with respect to one or several numerical variables in a survival setting.

Usage

metest(times, cens, expos, M, z = NULL, simul = TRUE, npvalsim = 1e+5, riskincr, 
  exposinflc = TRUE, minflc = TRUE, returnmodels = FALSE)

Arguments

times	numeric. Vector of survival/censoring times.
cens	numeric. Survival status vector. Each entry must take the value "1" if the corresponding entry in times is a survival time and "0" if this entry is a censoring time.
expos	numeric. Exposure vector. Each value must take the value "1" if the exposure is present in the corresponding case and the value "0" if the exposure is absent for that case.
M	matrix or vector. If M is matrix than there are more than one (potential) mediators. In this case, each column contains the values of one of the (potential) mediators. If M is a vector than this vector simply contains the values of the (potential) mediator to be studied.
z	list. Named list, where each entry contains the values of one of the covariates to be adjusted for.
simul	logical. TRUE if the simulation approach should be used. FALSE if the delta approximation should be used.
npvalsim	numeric. Number of simulation iterations to be used in the simulation approach. Default is 1e+5. To obtain more precise p values use 1e+6 or 1e+7 iterations if computationally feasible.
riskincr	logical. TRUE if it is assumed that the presence of the exposure leads to an increase in risk. FALSE if presence of the exposure is assumed to lead to a decreased risk.
exposinflc	logical. TRUE if the influence of the exposure should be modelled as constant over time. FALSE if this influence should be modelled as time-varying.
minflc	logical. TRUE if the influence of the mediator should be modelled as constant over time. FALSE if this influence should be modelled as time-varying.
returnmodels	logical. TRUE if the Aalen model(s) (equation (1) in Lange & Hansen (2011)) and the mediator model(s) (equation (2) in Lange & Hansen (2011)) should be returned. Default is FALSE.

Details

For details see Lange & Hansen (2011).

Value

metest returns an object of class metest. An object of class "metest" is a list containing the following components:

pvalues

vector of length equal to the number of (potental) mediators tested, where the j-th entry corresponds to the p-value for the j-th (potential) mediator.

params

If returnmodels = FALSE (default), this will be NULL. If returnmodels = TRUE this will contain the Aalen model(s) (equation (1) in Lange & Hansen (2011)) and the mediator model(s) (equation (2) in Lange & Hansen (2011)). In the case of more than one tested mediator this will be a list of lists, each of which contains two elements "aalenfit" and "ols_m". In the case of only one tested mediator this will be a list with two elements "aalenfit" and "ols_m".

Author(s)

Roman Hornung

References

Lange, T., Hansen, J.V. (2011) Direct and indirect effects in a survival context. Epidemiology 22:575–581.

Examples

# Simulate a dataset:
#####################

# Simulation parameters:

# Number of observations:
n <- 100
# Number of (potential) mediators:
p <- 10

# Indices of true mediators:
mediatorind <- 1:3

# Baseline hazard rate:
lambda <- 0.05

# Effect of exposure on the three mediators:
exposmediatoreff <- 0.7

# Effect of the mediators on survival:
surveff <- 0.8

# Direct effect of the mediator on survival:
exposeff <- 0.6

# Hazard rate of exponentially distributed censoring times:
lambdacens <- 0.02

# Coefficient of linear predictor in the Cox regression
# model relating the (potential) mediators and the
# exposure to the outcome:

beta <- rep(0, p+1)
beta[mediatorind] <- surveff
beta[p+1] <- exposeff


# Simulate data:

# Set seed:
set.seed(123)

# Exposure:
expos <- rep(c(1,0), each=n/2)

# (Potential) mediators:
X <- matrix(nrow=n, ncol=p, data=rnorm(n*p))
X[expos==1, mediatorind] <- X[expos==1, mediatorind] + exposmediatoreff

# Uncensored survival times:
linpred <- cbind(X, expos)%*%beta
indhazards <- as.vector(lambda*(exp(linpred)))
timesuncens <- -log(runif(n))/indhazards

# Censoring times:
censtimes <- rexp(n, rate=lambdacens)

# Censoring variable:
cens <- ifelse(timesuncens < censtimes, 1, 0)

# Observed survival times:
times <- timesuncens
times[cens==0] <- censtimes[cens==0]

# data.frame:
simdata <- data.frame(X)
simdata$expos <- expos
simdata$times <- times
simdata$cens <- cens




# Mediation tests:
##################

# Time-constant influence of exposure and time-constant influences of mediators:

# First mediator only:
metest(times=simdata$times, cens=simdata$cens, expos=simdata$expos, M=simdata[,1], 
       simul = TRUE, npvalsim = 10000, riskincr=TRUE, exposinflc = TRUE, minflc = TRUE)
	   
# First mediator only, where Aalen model and mediator model is returned:
metest(times=simdata$times, cens=simdata$cens, expos=simdata$expos, M=simdata[,1], 
       simul = TRUE, npvalsim = 10000, riskincr=TRUE, exposinflc = TRUE, minflc = TRUE, returnmodels = TRUE)

# All (potential) mediators:
metest(times=simdata$times, cens=simdata$cens, expos=simdata$expos, 
       M=as.matrix(simdata[,1:10]), simul = TRUE, npvalsim = 10000, riskincr=TRUE, 
       exposinflc = TRUE, minflc = TRUE)


# Time-varying influence of exposure and time-constant influences of mediators:

metest(times=simdata$times, cens=simdata$cens, expos=simdata$expos, 
       M=as.matrix(simdata[,1:10]), simul = TRUE, npvalsim = 10000, 
       riskincr=TRUE, exposinflc = FALSE, minflc = TRUE)


# Time-constant influence of exposure and time-varying influences of mediators:

metest(times=simdata$times, cens=simdata$cens, expos=simdata$expos, 
       M=as.matrix(simdata[,1:10]), simul = TRUE, npvalsim = 10000, riskincr=TRUE, 
       exposinflc = TRUE, minflc = FALSE)


# Time-varying influence of exposure and time-varying influences of mediators:

metest(times=simdata$times, cens=simdata$cens, expos=simdata$expos, 
       M=as.matrix(simdata[,1:10]), simul = TRUE, npvalsim = 10000, riskincr=TRUE, 
       exposinflc = FALSE, minflc = FALSE)

RomanHornung/survmediation documentation built on July 3, 2023, 4:25 a.m.