# Evidence-based directed acyclic graphs for perinatal pharmacoepidemiologic studies in rheumatology: a structured approach for development and implementation in administrative health data

## TL;DR

This paper presents a structured method for creating causal diagrams to study the effects of rheumatology drugs on birth defects using health data.

## Contribution

A systematic approach for developing evidence-based DAGs in perinatal pharmacoepidemiology using administrative health data.

## Key findings

- A DAG with 21 nodes was developed to evaluate tsDMARDs and congenital anomalies.
- Three key confounders were identified for adjustment in multivariable models.
- The approach improves study design in perinatal pharmacoepidemiology research.

## Abstract

Evidence-based Directed Acyclic Graphs (DAGs) are effective tools to comprehensively visualize complex causal and biasing pathways in pharmacoepidemiologic research in rheumatology. This paper outlines the process of developing and implementing a DAG, using a cohort study evaluating the impact of targeted synthetic disease-modifying antirheumatic drugs (tsDMARDs) on congenital anomalies as a case example. We include a discussion of how factors would be operationalized into variables in administrative data within the case example.

DAG Development involved: 1) identifying exposure and outcome, 2) identifying factors affecting the exposure, 3) identifying factors affecting the outcome, 4) identifying factors affecting both the exposure and outcome, 5) ascertaining relationships between factors, and lastly, 6) finalizing the DAG in DAGitty v3.1.

The final DAG for our case example on evaluating the association between tsDMARDs and congenital anomalies consisted of 21 nodes (points in the diagram representing factors such as exposures, outcomes, confounders, or mediators): 1 affecting the exposure, 12 affecting the outcome, 7 on the biasing pathways, and 1 mediator (maternal infection) on the exposure-outcome pathway. One minimally sufficient adjustment set was identified to inform confounder adjustment in a multivariable model, consisting of: concomitant conventional synthetic DMARDs, rheumatic disease activity, and maternal demographics (i.e., age, place of residence, race/ethnicity). Implications for implementing this DAG in a study using administrative health data include comprehensively revealing confounders to be adjusted for.

Our systematic approach to developing a DAG is particularly valuable for improving study designs in the growing field of perinatal pharmacoepidemiology in rheumatology, where there is a critical need for robust perinatal data on novel arthritis medications.