- Upgrade dependency to pymc 2.3.3
- Remove LBA model as likelihood seems broken
- Compatibility with pandas > 0.13.
- Fix problem that causes stats to not be generated when loading model.
- Update packages to work with anaconda 1.9.
- Refactored posterior predictive plots and added tutorial: http://ski.clps.brown.edu/hddm_docs/tutorial_post_pred.html
- Smaller bugfixes.
- Works with PyMC 2.3.
- Experimental features:
- Updated HLBA model but currently has bad recovery.
- Added sample_emcee() to use the emcee parallel sampler. Seems to work but requires some tuning and does not seem to beat slice sampling.
- New and improved HDDM model with the following changes:
- Priors: by default model will use informative priors
(see http://ski.clps.brown.edu/hddm_docs/methods.html#hierarchical-drift-diffusion-models-used-in-hddm)
If you want uninformative priors, set
informative=False
. - Sampling: This model uses slice sampling which leads to faster convergence while being slower to generate an individual sample. In our experiments, burnin of 20 is often good enough.
- Inter-trial variablity parameters are only estimated at the group level, not for individual subjects.
- The old model has been renamed to
HDDMTransformed
. - HDDMRegression and HDDMStimCoding are also using this model.
- Priors: by default model will use informative priors
(see http://ski.clps.brown.edu/hddm_docs/methods.html#hierarchical-drift-diffusion-models-used-in-hddm)
If you want uninformative priors, set
- HDDMRegression takes patsy model specification strings. See http://ski.clps.brown.edu/hddm_docs/howto.html#estimate-a-regression-model and http://ski.clps.brown.edu/hddm_docs/tutorial_regression_stimcoding.html#chap-tutorial-hddm-regression
- Improved online documentation at http://ski.clps.brown.edu/hddm_docs
- A new HDDM demo at http://ski.clps.brown.edu/hddm_docs/demo.html
- Ratcliff's quantile optimization method for single subjects and
groups using the
.optimize()
method - Maximum likelihood optimization.
- Many bugfixes and better test coverage.
- hddm_fit.py command line utility is depracated.
- Models are now pickable. (This means they can be loaded and saved. Critically, it is now also trivial to run multiple models in parallel that way.)
HDDM 0.4 is now distributed under the simplified BSD license (see the LICENSE file) instead of GPLv3.
- Handling of outliers via mixture model. http://ski.clps.brown.edu/hddm_docs/howto.html#deal-with-outliers
- New model HDDMRegression to allow estimation of trial-by-trial regressions with a covariate. http://ski.clps.brown.edu/hddm_docs/howto.html#estimate-a-regression-model
- New model HDDMStimulusCoding. http://ski.clps.brown.edu/hddm_docs/howto.html#code-subject-responses
- New model HLBA -- a hierarchical Linear Ballistic Accumulator model (hddm.HLBA).
- Posterior predictive quantile plots (see model.plot_posterior_quantiles()).
- model.load_db() is working again.
- Fixed annoying bug that broke plotting of posterior predictive.
- Complete rewrite of the underlying model creation engine (kabuki) to allow for more flexible model creation including transforms. This enabled development of a new HDDM default model without explicit parameter bounds.
- Group mean distributions are now Gibbs sampled and group variability distributions are now slice sampled leading to much improved convergence and mixing.
- MAP approximation of hierarchical models for better initialization.
- Improved documentation (check out the `How-to`_ section).
- Chi-square fitting using the Ratcliff quantile method.
- Posterior predictive checks.
- Better model initialization that shouldn't fail.
- Many bugfixes.
- Major internal overhaul.
- Flexible HDDM model class to fit group and subject models.
- Heavily optimized cython likelihoods.