trialdiff and the existing ecosystem

Complement, not duplicate

trialdiff does not attempt to replace low-level data frame comparison. It builds on the observation that detection is a solved problem, while clinical interpretation and downstream impact are not.

Tool / package	What it does	Where it stops	How trialdiff relates
diffdf	Detailed cell-level diff of two data frames, with keys	No clinical categorisation, no lineage, no impact	Optional low-level backend (backend = "diffdf"); trialdiff adds the clinical layers
waldo	General R object comparison, used by testthat	General purpose, not clinical	Optional low-level equality backend (backend = "waldo")
dplyr / base	Joins and set operations	Building blocks only	Used internally; no user-facing duplication
haven	Read SAS/SPSS/Stata	Transport only	Input reader for compare_cut()
admiral	ADaM derivation	Does not compare data cuts	Downstream of the comparison; lineage edges can point at admiral derivations
metacore / metatools	Metadata management and dataset checking	Metadata, not data-cut diffs	lineage_from_metadata() derives a lineage graph from a metacore object
cards	Analysis Results Data	Results, not change detection	Impacted “analysis” nodes can be cards/ARD objects
tern / rtables	TLG generation	Output generation	Outputs (TLFs) are lineage nodes flagged for review
SAS PROC COMPARE	Dataset comparison, value/label/length differences	No classification, lineage or impact	Conceptual ancestor; trialdiff is the R/pharmaverse equivalent plus context
Commercial platforms	End-to-end clinical data management and validation	Closed, licence-bound	trialdiff is open-source and scriptable

Integration points

• Input: any data frame, including haven-imported SAS datasets and admiral-derived ADaM.
• Metadata: variable labels are read from the label attribute, which haven, metacore and admiral all set.
• Lineage: edges can reference admiral derivations, cards ARDs and tern/rtables outputs.
• Output: JSON and list reports for automated QC pipelines; HTML/Quarto for human review.

Is this novel?

The novelty is not in comparison. It is in the combination of:

1. a clinical change taxonomy;
2. explicit, user-declared lineage between datasets, variables, analyses and outputs;
3. a transparent, policy-driven impact assessment that distinguishes “definitely”, “potentially” and “unlikely”, and refuses to claim statistical impact without a rerun;
4. reporting that lists exactly what a programmer or statistician must review.

No package in the pharmaverse currently occupies this position. See the project proposal in proposal/ for the full gap analysis and roadmap.