History

0.3.0 (2026-06-10)

Warning

Breaking changes since 0.2.0 — read before re-running old analyses:

  • Group-level SD prior default changed halfcauchy → halfnormal (BaseModel.group_sd_dist). This silently changes the posterior of every hierarchical regression fit. HalfCauchy’s infinite-variance tail let poorly-identified group SDs run away (funnels, divergences); HalfNormal’s light tail tames them. To reproduce pre-0.3.0 numbers, set model.group_sd_dist = 'halfcauchy' per instance.

  • The regressors= keyword is deprecated in favour of the explicit fixed_regressors= (population-mean design, no per-subject offset) and random_regressors= (per-subject random effect) split. regressors= still works bit-for-bit (it maps to a random slope on every column) but now emits a DeprecationWarning.

  • Default lapse group prior changed to Beta “1/x” (p_lapse / p_outlier transform 'beta'); the old logit-Normal prior is opt-in via model.lapse_group = 'logit_normal'.

Sampling robustness and convergence (what makes the DDM/RDM models above actually mix on real, multi-subject data):

  • RT outlier contaminant (p_outlier) baked into every DDM and race-diffusion model (HSSM-style mixture). On a fraction p_outlier of trials the response is a flat-over-RT lapse rather than the diffusion process. Fixed p_outlier = 0.05 by default (set 0.0 for pure WFPT, 'hierarchical' to estimate it). This is what cracks the 66-subject DDM: the slow-RT tail otherwise breaks convergence, inflates Pareto-k (unreliable LOO), and shows up as a PPC tail misfit — one cause, three symptoms, all fixed (r̂ 1.59 → 1.00, min ESS 7 → 3200+).

  • Configurable group-SD prior group_sd_dist ('halfnormal' | 'halfcauchy'); see the breaking note above.

  • Opt-in multipath Pathfinder init (find_init='pathfinder', MAP-seeded) for genuinely multimodal posteriors; 'mapjitter' remains the default and is usually sufficient once the contaminant is in.

  • New guide notes/fitting_ddm_without_divergences.md in the source tree — the full recipe and a symptom → diagnosis → fix catalogue for DDM/RDM convergence.

Fixed vs random effects API

  • fixed_regressors / random_regressors on the regression models (static choice and the DDM magnitude-comparison regression model), letting you put a population-mean (fixed) effect on a between-subjects covariate while keeping a per-subject random intercept — the correct parameterisation for group contrasts (a random slope on a between-subjects covariate is non-identified; bauer now warns). New tutorial: Fixed vs random effects.

Other modelling additions

  • memory_as_sv DDM variant: routes the frozen memory noise into Ratcliff across-trial drift variability sv (perceptual noise stays within-trial), making the two noise sources separately identifiable. Composes with the contaminant.

  • Race-diffusion fit_w_d / fit_w_s toggles to fix the discriminative gain (w_d ≡ 1) or ablate the overall-magnitude (sum) drift term.

  • consistent_choice_noise for static choice models: normalise the choice likelihood by the SD of the noisy posterior mean (KLW/DDM-consistent), matching how the accumulator models treat decision noise.

The original 0.3.0 feature set (unreleased until now):

  • DDM and race-diffusion models for choice + RT data, sharing the same Bayesian observer cognitive front-end as the static-choice models:

    • DDMMagnitudeComparisonModel, DDMFlexibleNoiseComparisonModel, DDMRiskModel, DDMFlexibleNoiseRiskModel (+ regression variant) — Wiener WFPT likelihood via hssm.likelihoods.logp_ddm.

    • RaceDiffusionMagnitudeComparisonModel, RaceDiffusionFlexibleNoiseComparisonModel, RaceDiffusionRiskModel, RaceDiffusionFlexibleNoiseRiskModel (+ regression variant) — analytical Wald-race likelihood with advantage=True decomposition by default (van Ravenzwaaij 2020 style).

  • JAX backend support via --backend {numpyro,blackjax} in the CLI fit scripts. JAX-NUTS is 1.5–3× faster on CPU and 5–30× faster on GPU (NVIDIA L4, with chain_method='vectorized' so chains run in parallel on a single device).

  • CLI fit scripts under bauer/scripts/: fit_garcia.py, fit_dehollander2024.py, fit_dehollander_tms.py.

  • SLURM job templates under bauer/scripts/slurm_jobs/ for cluster fitting, including a CUDA-env build job, a generic run_fit.sh wrapper, a JAX backend benchmark, and a full per-dataset production submit script.

  • Bundled datasets added (loaders in bauer.utils.data): load_dehollander2024_risk (dotcloud, N=30), load_dehollander2024_symbolic (Arabic-numeral risk, N=58), load_dehollander_tms_risk (TMS risky choice, N=35 sessions 2/3), load_bedi2026 (Bedi 2026 abstract-value estimation pilot, N=13).

  • Unified PPC API: BaseModel.ppc now returns the same long-format DataFrame as the DDM/RDM PPCs — index = paradigm levels + ppc_sample, single simulated_choice column.

  • Renamed polynomial_order → spline_order throughout (no backwards-compat alias).

  • Cleanup: removed SafeVsRisky* family, RNPModel, several deprecated prior_estimate options in RiskModel, and the incorporate_probability parameter. 26 stale dev notebooks deleted; bundled CSVs trimmed of derived columns.

0.2.0 (2026-04-03)

  • Refactored model classes into a dedicated bauer/models/ package (psychophysics, magnitude, risky_choice submodules).

  • Added pyproject.toml with fully declared runtime and optional dependencies.

  • Added environment.yml for reproducible conda environment setup.

  • Added tutorial notebooks covering psychophysics / magnitude comparison, risky choice with the KLW model, and de Hollander et al. (2024) stake effects.

  • load_dehollander2024() data loader added to bauer.utils.data.

  • Documentation expanded with API reference, concepts page, and tutorial.

0.1.0 (2022-11-17)

  • First release on PyPI.