📊

SAS → R Transition

Hands-on Seminar · Statistics Denmark · 2026

Tools, patterns, and institutional knowledge for a durable open-source migration

Agenda

What We Will Cover

🗺️

Migration Philosophy

Replace the pipeline, not just the syntax

🔧

Conversion Tools

AI translators, open-source accelerators

💊

Pharmaverse Lessons

Pipeline packages & AGENTS.md pattern

⚠️

Watch-Outs

Silent bugs, numerical discrepancies

Foundation

Two Approaches to Migration

❌ Syntax Translation

Line-by-line conversion

• Converts SAS DATA steps → R code mechanically
• Produces R that "thinks in SAS"
• Fragile — SAS idioms don't map cleanly
• Macros encoding business rules are lost

✅ Pipeline Replacement

Rebuild each stage in idiomatic R

• Map SAS roles to purpose-built R packages
• Tidyverse + domain packages (survey, gt, haven)
• Maintainable, reviewable, submission-ready
• Pharmaverse proved this at scale

Layer 1

AI-Powered Conversion Tools

Tool	Type	Best For	License
CSO Ireland Transcompiler	Shiny app + OpenAI API	NSO context — built by stats office for stats office	MIT / FOSS
codeconvert.ai	Browser-based, no signup	Live seminar demos — paste SAS, get R instantly	Free tier
Claude 3.7 Sonnet / GPT-4o	Direct API / chat	Module-by-module macro decomposition	Commercial API
Brillio / AWS Bedrock	Human-in-the-loop pipeline	Large-scale batch migration with audit trail	Commercial
HMS Code2X	Agentic migration tool	Claims ~75% automation; regulated environments	Commercial

⚠️ All tools handle DATA steps and PROC SQL well — all fail silently on complex SAS macros

Layer 2

SAS → R Package Crosswalk

SAS Concept	R Equivalent
DATA step	`dplyr` verbs — mutate, filter, group_by, summarise
PROC SQL	`dbplyr`, `dplyr`, `sqldf`
PROC MEANS / SUMMARY	`summarise()`, `janitor`, `gtsummary`
PROC FREQ	`table()`, `janitor::tabyl()`
PROC SGPLOT	`ggplot2`
ODS / RTF output	`gt`, `flextable`, `Quarto`
SAS macros	R functions + `purrr`
.sas7bdat files	`haven::read_sas()`
SAS formats / labels	`forcats` (factor labels)

💊

Lessons from Pharmaverse

The pharmaceutical industry has run the largest regulated SAS→R migrations in the world.
Their patterns are directly transferable to national statistics.

Pharmaverse

The Pharmaverse Package Stack

SAS Role	Pharmaverse R Package	Equivalent
PROC IMPORT / .sas7bdat	`{haven}`	Read SAS datasets; handles date epoch shift
SDTM derivation macros	`{sdtm.oak}`	Map raw data to SDTM domains (DM, AE …)
ADaM derivation macros	`{admiral}`	Build ADSL, ADAE, ADTTE with auditable derivations
PROC EXPORT / XPT files	`{xportr}`	Submission-compliant XPT with define.xml metadata
RTF/PDF output macros	`{r2rtf}`, `{rtables}`	Clinical-quality table output
SAS log files	`{logrx}`	R execution logs mirroring SAS .log format
Interactive data review	`{teal}`	Shiny ADaM explorer

Each package replaces a process role — not individual lines of code

Merck / PharmaSUG 2025

GenAI-Assisted Conversion Workflow

1

Import first — haven::read_sas() to ingest .sas7bdat without rewriting ETL

2

Decompose macros — split into modules (data prep → model → output) and convert each separately; feeding the full macro to an LLM fails

3

Manual parameter verification — replace PROC MIXED → lmer() + emmeans(); check degrees-of-freedom method explicitly

4

Numerical equivalence check — run SAS and R outputs side-by-side; Merck confirmed identical PK summary results

5

Output: use templates — GPT-4o is weak at RTF macros; use existing {r2rtf} templates and adapt

Key Pattern

AGENTS.md — Encoding Institutional Knowledge

The {admiral} package publishes a 1,300-line auto-generated AGENTS.md — a README for AI agents, not humans

What admiral's AGENTS.md contains

ADaM conventions LLMs can't infer — e.g. flag vars take "Y" or NA, never "N"

Full ecosystem pipeline context: sdtm.oak → admiral → xportr → XPT

Unit testing rules, documentation workflow, compliance check comment

DST equivalent would encode

StatBank table format conventions and px-file structure

Eurostat reporting requirements and metadata standards

Survey weighting conventions that an LLM cannot guess

rOpenSci now requires AGENTS.md as a peer-review artifact — this norm is arriving regardless

⚠️ Watch Out

Silent Failure Modes

SAS Macros encode business rules, not syntax

LLM translators handle DATA steps and PROC SQL well but fail silently on macros — macro logic must be interpreted by a human, not translated character by character

Missing values behave differently

In SAS, IF x > 5 where x = . evaluates to FALSE. In R, x > 5 where x = NA returns NA — propagating silently through pipelines. CAMIS working group documents all such discrepancies.

PROC MIXED → lmer() is the hardest step

Kenward-Roger vs Satterthwaite degrees-of-freedom methods differ by default between SAS and R — producing different p-values for the same model. Always verify manually.

haven::read_sas() drops some format labels silently

Always run str() on imported data and cross-check variable types before any analysis step

Resources

Key References for the Seminar

admiral AGENTS.md (live)

github.com/pharmaverse/admiral

1,300-line production AGENTS.md — best real-world example of encoding domain knowledge for AI

CSO Ireland Transcompiler

github.com/CSOIreland/Code-Transcompiler-using-LLMs-public

MIT-licensed Shiny app built by a national statistics office for SAS→R; deployable on Posit Connect

pharmaverse/examples

pharmaverse.github.io/examples

End-to-end SDTM→ADaM→TLG pipelines in R; runnable on Posit Cloud

CAMIS Working Group

IAIMS / R Consortium

Documents numerical discrepancies between SAS, R, and Python for standard statistical procedures

EpiR Handbook — Transition to R

epirhandbook.com

Public health / statistics-office flavored crosswalk guide; free and openly licensed

PharmaSUG 2025 AI-136

pharmasug.org/proceedings/2025

Merck's validated GenAI-assisted PK table conversion — numerical equivalence confirmed

Hands-On Seminar

Suggested Seminar Structure

🎬 Demo

Live translation — paste a PROC MEANS + DATA step into Claude 3.7; inspect R output; fix one edge case manually to teach critical evaluation, not blind trust

📥 Import

haven::read_sas() — import an existing .sas7bdat file; run str(); identify any label-drop issues before touching analysis

🔬 CAMIS

Equivalence check — run one real procedure in both SAS and R; dramatise the "bitwise reproducibility" problem for official statistics

🤖 AGENTS

Write a DST AGENTS.md stub — encode one DST output convention (e.g. StatBank px-format rule) as a reusable AI instruction file

🚀 Deploy

CSO Ireland transcompiler — explore self-hosting the MIT-licensed app on Posit Connect as a DST internal tool

🇩🇰

The Key Insight

Automated tools handle syntax.
AGENTS.md handles institutional knowledge.
The latter determines whether migrated code is actually correct.

FOSS-first

haven + tidyverse + pharmaverse patterns

Validate always

CAMIS equivalence checks for official statistics

Encode conventions

AGENTS.md before AI touches DST pipelines