groceryheist / misclassificationmodels Goto Github PK

Right now we only support Gaussian and logistic models. It would be great to support more. Creating this issue to note and track this.

A partial list of models I want to support

• Count models (e.g., poisson; negative binomial; zero-inflated).
• Survival models (e.g., Cox models).
• Models for bounded continuous outcomes (e.g., gamma regression).

Uncertainty quantification can be a challenge in error correction models using MLE.
We currently use MLE with the quadratic approximation. #4 partially addresses this, but maybe Stan can do an even better job and also support a wider range of models more easily.

One thing to consider is if this package should focus on MLE methods and supporting Bayesian models should be developed in parallel, but in a separate package.

• Sargan's J Test
• Stratified diagnostic odds ratios and heterogeneity test (e.g. metafor)

In theory, profile likelihood provides better uncertainty quantification than the quadratic approximation does. This package currently only supports the quadradic approximation.

I already have some code that uses profile likelihood, but it isn't integrated in this package or tested.