If single race/ethnicity in cohort
This stage of the analysis is specific to the chosen exposure/outcome and the specified adjustment variables. Below is the code for all of the analyses to run for the birth size project. Please be sure to update the cohort and date information in the below code for your analysis, as well as the destination path. Finally, be sure to update the column names of the exposure/outcome(s) of interest, the adjustment variables, and the table 1 variables. These should correspond to column names in the dataframe specified in the phenofinal
argument of the dataAnalysis
function.
A few important specifications to note in the dataAnalysis
function:
- By specifying
vartype="ExposureCont"
, the birth size characteristic is modeled as a continuous exposure and assumed to be numeric - By specifying
robust=TRUE
, analyses are run via robust regression using iterated re-weighted least squares (Huber weights), and White’s estimator for the variance - The
Table1vars
argument is used to specify the variables to include in our descriptive Table 1 - The
adjustmentvariables
argument specifies the variables we will adjust for in our models RunUnadjusted=TRUE
, indicates to run unadjusted modelsRunAdjusted=TRUE
, indicates to also run adjusted modelsRunCellTypeAdjusted=TRUE
, indicates to also run adjusted models, further adjusting for estimated cell compositionRunSexSpecific=TRUE
, indicates to also run the models stratified by infant sexRunCellTypeInteract=TRUE
, indicates to evaluate interactions with cell compositions
Additional details regarding the dataAnalysis
function can be found by running ?dataAnalysis
Quick check to make sure the function runs in your cohort
Given the modeling approaches used, the dataAnalysis
function requires a good deal of time to run. We recommend first checking whether the function runs on a relatively small subset of sites (i.e. 100 CpG loci). If you encounter any issues, please let us know. If not, proceed to the next step.
If you plan to run your analysis in parallel
Running your site-specific analysis in parallel (argument runparallel=TRUE
) can greatly reduce the computation time. To run your analysis in parallel, you will need to specify the number of cores to use (argument number_cores
). You can use the detectCores()
function in the parallel
library to check the number of cores available; depending on your computing resources, you may need to specify less cores than the total number available. If you are not planning to run your analysis in parallel, specify runparallel=FALSE
.
## if running in parallel, checking the number of available cores
library(parallel)
detectCores() # should probably choose at least one less than the number available
allvarsofinterest=c("BWT_Zscore","BirthLength_Zscore","HeadCircum_Zscore","wlr_Zscore")
## You can reduce this dataframe to whatever variables you have.
## For example, if you only have birthweight, you would specify:
## allvarsofinterest=c("BWT_Zscore")
for (i in 1:length(allvarsofinterest)){
cat("Exposure:",allvarsofinterest[i],"\n")
tempresults<-dataAnalysis(phenofinal=phenodataframe,
betafinal=Betasnooutliers[1:100,],
array="450K",
maxit=100,
robust=TRUE,
Omega=processedOut$Omega,
vartype="ExposureCont",
varofinterest=allvarsofinterest[i],
Table1vars=c("BWT","Gestage","Sex","Age","Parity","MaternalEd",
"Smoke","preBMI","Ethnic","Meanlog2oddsContamination"),
StratifyTable1=FALSE,
StratifyTable1var=NULL,
adjustmentvariables=c("Sex","Age","Parity","MaternalEd",
"Smoke","preBMI","Ethnic","Meanlog2oddsContamination"),
RunUnadjusted=TRUE,
RunAdjusted=TRUE,
RunCellTypeAdjusted=TRUE,
RunSexSpecific=TRUE,
RunCellTypeInteract=TRUE,
RestrictToSubset=FALSE,
RestrictionVar=NULL,
RestrictToIndicator=NULL,
number_cores=8,
runparallel=TRUE,
destinationfolder="H:\\UCLA\\PACE\\Birthweight-placenta",
savelog=TRUE,
cohort="HEBC",analysisdate="20220709",
analysisname="main")
}
Running the final models
Now running the models for all CpG loci
for (i in 1:length(allvarsofinterest)){
cat("Exposure:",allvarsofinterest[i],"\n")
tempresults<-dataAnalysis(phenofinal=phenodataframe,
betafinal=Betasnooutliers,
array="450K",
maxit=100,
robust=TRUE,
Omega=processedOut$Omega,
vartype="ExposureCont",
varofinterest=allvarsofinterest[i],
Table1vars=c("BWT","Gestage","Sex","Age","Parity","MaternalEd",
"Smoke","preBMI","Ethnic","Meanlog2oddsContamination"),
StratifyTable1=FALSE,
StratifyTable1var=NULL,
adjustmentvariables=c("Sex","Age","Parity","MaternalEd",
"Smoke","preBMI","Ethnic","Meanlog2oddsContamination"),
RunUnadjusted=TRUE,
RunAdjusted=TRUE,
RunCellTypeAdjusted=TRUE,
RunSexSpecific=TRUE,
RunCellTypeInteract=TRUE,
RestrictToSubset=FALSE,
RestrictionVar=NULL,
RestrictToIndicator=NULL,
number_cores=8,
runparallel=TRUE,
destinationfolder="H:\\UCLA\\PACE\\Birthweight-placenta",
savelog=TRUE,
cohort="HEBC",analysisdate="20220709",
analysisname="main")
}
Check model convergence
The function dataAnalysis includes an indicator of whether each site-specific model converged. If the models are not converging, you can increase the number of specified iterations for the robust regression models using the argument maxit in the dataAnalysis function. The current default number of iterations is 100; increasing this number will make the function slower.
baseoutputdirectory<-"H:/UCLA/PACE/Birthweight-placenta/HEBC_20220709_Output"
listchecking<-as.list(rep(NA,length(allvarsofinterest)))
names(listchecking)<-allvarsofinterest
for (i in 1:length(allvarsofinterest)){
tempvarofinterest<-allvarsofinterest[i]
cat("Exposure:",tempvarofinterest,"\n")
tempdirectory<-paste(baseoutputdirectory,"/",tempvarofinterest,"_main",sep="")
setwd(tempdirectory)
tempfilename<-paste("HEBC_20220709_",tempvarofinterest,"_main_allanalyses.RData",sep="")
load(tempfilename)
if("CellInteraction" %in% names(alldataout)) alldataout$CellInteraction<-NULL
if("warnings" %in% colnames(alldataout[[1]])) listchecking[[i]]<-lapply(alldataout,function(x) if(length(x)>1) table(x$warnings))
}
## Check them out
listchecking
Data upload
The PACEanalysis functions will create an “Output” folder in your specified destination directory, which will contain everything to be shared for the subsequent meta-analysis. To submit your results, zip this Output folder, and send it to Alexandra (ambinder@hawaii.edu).