Search best set of co variable in epidemiological studies with DAG and 6-steps algoritms.

Script to build a DAG with bnlean with epidemiogogic data and apply 6 step algorithm to choice co variables with effectors and outcome descrided in Shrier and Platt

Why ?

Usually, in epidemiology, researcher presented outcome and one or more exposure/co - founder from a single logistic regression/ generalized linear models. Theses approaches may affect and bias estimates of exposure on outcome (see Westreich, D. & Greenland,). To avoid bias and computed risk factor with generalized linear models for each couple effectors/outcome with a specific set of confounder. we selected best set of co founder for each couple with algorithm 6-step of Shrier and Platt apply on Directed Acyclic Graphs (DAG).

Algoritms

Steps 1 build dag with epidemiological data and format data for 6 step algorims

• Script : build_dag.r

Description

Build a Bayesian network and inference, to obtain a acyclic and directed graphics we black list some relation

Requirement

• R-project
  • corrplot library
  • bnlearn library

Steps 2 defined covariable(s) for a specific Effector and Outcome

• Script : 6Step.version.py

• requirement

  • python, library itertools, sys, argparse used

• options :

  • --input_file
    • relation parents and child : example where A parents of B and C : A B,C

A B,C
B D
C D

• --out_file : output files contain list of covariable could be uses [need]

• --effector : effector to analyse [need]

• --outcome : outcome need to analyse in graphics [need]

• --excl : list of variable to exclude, should be separate by comma [optional]

• models of output :

• mc minimum covariable (Default)

• mmc : merged minimum covariable

• lc : maximum covariable

• algorithms for mc :

  • we defined one effector and outcome, all other variables of the DAG are considered as putative cofactors
  • we discarded all putative cofactors that did not validate step 1 of 6-step algorithm DAG: descent of effector in DAG
  • for each putative covariable we performed a 6-step DAG
  • if no covariable was identified in the 6-step DAG algorithms, we did all combinations of 2 covariables and ran them for each set of algorithms
  • if no set of 2 covariables were found in previous step of the 6 steps, we did this for all combinations of 3 covariables, followed by combination of 4, 5, 6 until we obtained one or set of covariables respecting 6-step DAG algorithm
  • we computed risk factors (relative risk) of effectors on the outcome using the covariable set defined previously, covariable set used for each couple of outcome/effector is reported in Supplementary Tables 7, 8 and 9.

Other script :

• simulate_phenotype.r: script to simulate a phenotype values with DAG know and relation between variables

Example

• script simulate_phenotype.r to build simulates values of phenotype files ressource/Pheno.sim.infocontains a relation father child and beta value, sd values

A B 0.2 0.01
B D 0.15 0.005
E C -0.3 0.01
A C -0.5 0.01
C F 0.25  0.1 
E F -0.3  0.01 
C D -0.25  0.01 

• we used beta values and sd values (random normal law) and DAG to computed a phenotype for 10 000 individuals see ressource/test_data.csv

• build DAG :
  • we used values generated in previous step set to build DAG with bnlearn

• we observed a relation A<=>B so we excluded B=>A to avoid to be cyclic

• we writed DAG in file

• search best covariable :

  • we used python script to generate putative covariable for effector and outcome F

python  6Step.v1.1.py --input_file ressource/dag_bnlearn.tab --out_file ressource/CovEffC_F --effector C --outcome F --model mmc

How to Cite?

articles in review : revalence and Risk Factors for Chronic Kidney Disease in Four Sub-Saharan African Countries: an AWI-Gen Cross-Sectional Population Study

(accepted in lancet global health) See bibliography section for DAG see library bnlearn and 6-step procedure Shrier et al, 2008

Bibliography

• Westreich, D. & Greenland, S. The table 2 fallacy: Presenting and interpreting confounder and modifier coefficients. Am. J. Epidemiol. 177, 292–298 (2013).
• Shrier, I. & Platt, R. W. Reducing bias through directed acyclic graphs. BMC Med. Res. Methodol. 8, 70 (2008).
• Evans, D., Chaix, B., Lobbedez, T., Verger, C. & Flahault, A. Combining directed acyclic graphs and the change-in-estimate procedure as a novel approach to adjustment-variable selection in epidemiology. BMC Med. Res. Methodol. 12, 156 (2012).
• Thornley, S., Marshall, R. j, Wells, S. & Rod, J. Using Directed Acyclic Graphs for Investigating Causal Paths for Cardiovascular Disease. J. Biom. Biostat. 04, (2013).
• Margaritis, D. Learning Bayesian Network Model Structure from Data. (Carnegie-Mellon University, Pittsburgh, PA, 2003). Scutari, M. Learning Bayesian Networks with the bnlearn R Package. J. Stat. Software, Artic. 35, 1–22 (2010).