Add support for non-linear mixed models

[paul-buerkner]

As brms generates Stan code on the fly, it should be possible to translate non-linear model formulae into Stan code, given that Stan knows the occuring functions (such as exp or log; possibly under another name) or that I have defined these functions in the 'functions' block of the Stan code. Definitely more of a long term goal, though.

[paul-buerkner]

Non-linear mixed models are more or less working now. Still, there are quite a few things to do:

• document new functionality
• add more unit tests
• possibly improve efficiency of the Stan code (will stay as it is for now)
• improve comments in the Stan code
• add more input options for argument nonlinear
• write a vignette
• improve documentation
• allow random effects of different non-linear parameters to be correlated

Also, I am almost sure I will find some more bugs in the process.

[EmmanuelCharpentier]

Thank you for this very interesting and very handy package. The "nonlinear" option is a very big selling point !

A vignette, including several examples and at least one detailed example, is indeed urgently needed : the current documentation makes it hard for a newcomer to figure out what should go in the "nonlinear" part of the model, and how to specify a model with both linear and nonlinear parts.

[paul-buerkner]

I agree. The vignette should be ready within the next few weeks.

[aforren1]

If it helps in the meanwhile, I've worked a few simple examples using the Orange dataset at http://rpubs.com/aforren1/orange-nonlinear.

[paul-buerkner]

@aforren1 This is great, thanks! :-)

[paul-buerkner]

All of the main features of non-linear models have been implemented in brms version 0.8 and 0.9. The remaining stuff has been moved to other issues (see #91 and #90). Accordingly, I think this issue can be closed now.