eribean / girth_mcmc Goto Github PK

Add a function to support the GPCM model in both Bayesian and Variational Inference.

Just did a quick comparison for a 2PL model on a dichotomous dataset between the R package mirt and girth_mcmc. Are there any specific reasons that the scale for the difficulty values are inverted (for lack of a better term)?

Need to support datasets that have missing values.

The graded unfolding model is an ideal point IRT model and an extension of the PCM #9. The framework for this model reuses a lot from the partial credit model and should be straightforward to implement. Not so straightforward is obtaining good initial estimates for item locations. The polarity ambiguity of item locations can cause posterior methods to diverge since to stable points will exist....

Add the other two often used polytomous models.

Future PYMC3 is going to incorporate JAX and will bring with it better GPU support, keep an eye on this.

Polytomous estimation can take some time, look into variational estimation to speed some things up.

Add support for estimating multidimensional models.

Return the 95% confidence interval from the trace.