mssm 0.1.26

Toolbox for estimating Generalized additive mixed models (GAMMs), semi-Markov-switching (sms) GAMMs, and Impulse response sms GAMMs.

mssm: Markov-switching Spline Models

Description

mssm is a toolbox to estimate Generalized Additive Mixed Models (GAMMs) semi Markov-switching GAMMs (sMs-GAMMs) and sMs Impulse Response GAMMs (sMs-IR-GAMMs). The main branch is updated frequently to reflect new developments. The stable branch should reflect the latest releases. If you don't need the newest functionality, you should install from the stable branch (see below for instructions).

Installation

The easiest option is to install from pypi via pip.

1. Setup a conda environment with python > 3.10
2. Install mssm via pip

The latest release of mssm can be installed from pypi. So to install mssm simply run:

conda create -n mssm_env python=3.10
conda activate mssm_env
pip install mssm
pip install matplotlib # Only needed for tutorials

The fourth line, installing matplotlib is only necessary if you want to run the tutorials. Note: pypi will only reflect releases (Basically, the state of the stable branch). If you urgently need a feature currently only available on the main branch, consider building from source.

Building from source

You can also build directly from source. This requires conda or an installation of eigen (setup.py then expects eigen in "usr/local/include/eigen3". This will probably not work on windows - the conda strategy should.). Once you have conda installed, install eigen from conda-forge. After cloning and navigating into the downloaded repository you can then install via:

pip install .

To get started

• With GAMMs: Take a look at tutorial 1 in the tutorial folder.
• With sms-IR-GAMMs: Take a look at tutorial 2.
• With sms-GAMMs: Take a look at tutorial 3.

Contributing

Contributions are welcome! Feel free to open issues or make pull-requests to main. Some problems that could use work are listed below.

• Tensor smooth term setup is not very efficient and takes quite a lot of time.
• Spline bases are always first evaluated as dense matrices followed by pruning zeros away afterwards. This is quite wasteful.
• Fellner-Schall update computes generalized inverse for every term. This works but is not necessary. For univariate terms with a single penalty this could be based on the pre-computed rank (see Wood, 2020: Inference and computation with generalized additive models and their extensions) of the penalty matrix (for B-splines this is super easy to obtain).
• Constraints. Sum-to-zero constraints (1.T @ f(x) = 0) are not so attractive from a sparse model perspective. However, that alternatives result in wider CIs is well documented (see Wood, 2017 for example). Finding a good trade-off is still high on my list, the alternatives implemented so far perform okayish.