ds_DistributedBoosting.jl

This Julia package implements algorithms for performing heuristic componentwise likelihood-based boosting on distributed data under data protection constraints for data distributed via DataSHIELD. The algorithm is solemnly based on aggregated data and no individual-level data are shared.

The data need to be provided on DataSHIELD server and a login (username / password) needs to be provided. The most important assumption is that the data are standardized beforehand (mean=0, sd=1). Please only provide already standardized data, otherwise the results will not be valid.

The package uses RCall, thus R needs to be available. You need to install the R-Packages opal, dsBaseClient, dsStatsClient, and dsModellingClient in your R-library.

Installation:

The ds_DistributedBoosting package is installed by:

using Pkg
Pkg.add(PackageSpec(url="https://github.com/danielazoeller/ds_DistributedBoosting.jl.git", 
    rev="master"))

Overview of functions

The following tables provide an overview of the functions of the package, together with a short description. You can find more detailed descriptions for each function using the Julia help mode (entered by typing ? at the beginning of the Julia command prompt). Note that ds_boosting is overloaded.

New mutable structures

Structure name	Short description
Unibeta	Containing pooled univariable estimates and corresponding labels.
Covarmat	Containing pooled covariance matrix and corresponding labels.
DSLogin	Containing login information for DataSHIELD-servers.
Bootsscratch	Containing all relevant information for the algorithm. Resulting structure.

Functions for starting and ending a session

Function name	Short description
ds_loadPkg	Load needed R-packages.
ds_login	Login into DataSHIELD server.
ds_check	Check if the loaded variables are standardized.
ds_start	Wrapper function for ds_loadPKg, ds_check, ds_start.
ds_logout	Logout of DataSHIELD server.

Functions for calculating pooled univariable effects estimates and pooled covariances using DataSHIELD.

Function name	Short description
calc_covarmat	Calculate pooled covariance matrix using DataSHIELD.
calc_covarmat!	Expand pooled covariance matrix using DataSHIELD.
calc_unibeta	Calculate pooled univariable effect estimate using DataSHIELD.

Functions for extracting relevant informations for next boosting step

Function name	Short description
selectionsofcovs	Get variable names of the ones with the next highest score function.
getselections	Get variables names that should be called next according to heuristic (with buffer).
getcovarfromtriangular	Extract covriance of all variables with selected covariate from pooled covariance matrix.

Functions to perform boosting steps

Function name	Short description
boost!	Perform a single boosting step.
reboost!	Re-do boosting steps for newly called covariances.

Functions to perform complete distributed heuristic boosting using DataSHIELD

Function name	Short description
ds_boosting	Wrap all relevant steps - Available as single call or as wrapper with login/logout functions.

Examples

Prerequisite for running the example code here is that the ds_DistributedBoosting package is installed:

Pkg.clone("https://github.com/danielazoeller/ds_DistributedBoosting.jl.git")

Additionally, the standardized data need to be provided on DataSHIELD server and a login needs to be provided. You can use the data stored here containing 500 simulated predictors for 500 individuals on 2 servers (250 each). You need to install the R-Packages opal, dsBaseClient, dsStatsClient, and dsModellingClient in your R-library.

The example code needs to be adjusted to your own settings.

Reference

The algorithm is explained and evaluated in:

Zöller, D., Lenz, S., Binder, H. (2018). Distributed multivariable modeling for signature development under data protection constraints. eprint arXiv:1803.00422. ArXiv.

You can find the article here.