negative-response-conventions

A simulation study of what the common ways of handling negative response values do to reported toxicity estimates, for responses that can legitimately fall below zero — specific growth rate above all.

Six data-handling conventions are compared on simulated data where the true ErC10, ErC50 and no-effect concentration are known by construction, so each can be scored rather than merely compared. Every arm is fitted the way bayesnec recommends: model-averaged over the declining candidate set, with package defaults throughout.

This compendium is the citable record behind the Modelling growth data and other potentially negative response values vignette (example7) in bayesnec.

The conventions

arm	what the analyst does	family
measured	use the values as recorded	gaussian
floored	replace negatives with zero	gaussian
deleted	delete the negative rows, keep the rest of their group	gaussian
censored	declare negatives left-censored at zero	gaussian
beta	floor, scale to (0, 1), fit a Beta	Beta
gamma	floor, fit a Gamma	Gamma

No arm is handed a prior built by another, so a contrast between arms is the contrast a practitioner would obtain rather than a likelihood-only contrast.

Running it

Everything runs from the repository root.

Rscript analysis/run_block.R 1 20      # iterations 1-50 in every cell, 20 workers
Rscript analysis/collate.R             # aggregate whatever exists -> results/metrics.csv
Rscript analysis/run_block.R 2 20      # iterations 51-100
Rscript analysis/collate.R

Blocks are balanced: block 1 runs iterations 1–50 in every cell, not cell 1 to completion. That is what makes an early collation meaningful, and it is what lets the vignette be drafted against 50 iterations and refreshed later without any prose changing.

The run is resumable. Each (cell, arm, iteration) writes its own file under results/, and a unit whose file already exists is skipped, so a crashed block is restarted with the same command.

Two things that are not optional

backend = "cmdstanr". brms defaults to rstan, which recompiles every model in every session. A single arm ran for ten minutes under rstan without producing a fit. fit_arm() refuses to run unless cmdstanr_write_stan_file_dir is set, because the cost of discovering this late is days rather than minutes.

The project library. lib/ holds the bayesnec build the study is about. Every script puts it first on .libPaths(). Do not run against a different installed version: the study exists to measure behaviour that changed in 2.2.0, and an older library silently measures the old behaviour.

Provenance

• bayesnec built from open-AIMS/bayesnec at 0976caca (dev), version 2.1.3.33, installed into lib/.
• R/simulate.R is copied verbatim from open-AIMS/negative-sgr at 0181d66e, so the generating model is demonstrably the same one the earlier study used.
• Nothing else is carried over. That study remains the frozen record of the results the vignette quoted before this one replaced them, and it is not modified.

Layout

R/simulate.R     generating model (carried over, do not edit)
R/cells.R        the design cells and their truths
R/arms.R         the six conventions, and the fitting call
R/estimates.R    endpoint extraction, model weights, per-unit runner
analysis/        run_block.R, collate.R
results/         one .rds per (cell, arm, iteration); metrics.csv
lib/             project-local bayesnec
cmdstan_cache/   compiled Stan programs, keyed by code hash

REDESIGN-claude.md is the implementation specification: what the study must establish, what changed from the previous design and why, the cell table, the staging plan and its Monte Carlo standard errors, the metrics schema, and what is out of scope. The collaborator-facing plan is open-AIMS/bayesnec#296.