A reproducible layered directed acyclic graph protocol for secondary-data studies

Abstract

Introduction: Secondary-data studies (clinical, administrative or sensor records) embed selection, measurement and missing-data processes that are often treated as “preprocessing”, leaving key causal assumptions implicit and hard to audit.

Objective: To propose a reproducible protocol that makes causal assumptions explicit in secondary-data studies using layered directed acyclic graphs.

Methodology: The protocol separates a material causal system (exposure→outcome) and three additional layers: (i) selection/observability, (ii) operationalisation and measurement (constructs vs. recorded proxies) and (iii) missing-data mechanisms. It includes estimand and time alignment, extraction rules and cohort construction, plus quality control based on graph source code and figure traceability.

Results: Outputs include a reporting checklist (Table 1), a decision→threat→graph-signature mapping with conditional mitigations (Table 2), a related-work gap synthesis (Table 3), layered DAG examples (Figs. 1–9) and reproducible evaluations quantifying typical biases and sensitivity to missingness (Tables 4–7).

Conclusions: The layered approach makes selection, measurement and missingness assumptions explicit; improves reproducibility through code–figure–text traceability; and operationalises alignment between design/analysis choices and the estimand, limiting claims that exceed the available evidence. A minimal applied case shows that observation/missingness decisions can materially shift the estimand, motivating explicit sensitivity reporting (Δ) and denominators.