mn2 is software package to fit linear mixed models to experimental data. It allows to maximize the restricted/residual likelihood criteria and obtain estimatives of fixed effects and variance components. Prediction of random effects parameters and coefficients can be made over different structures of variance covariance.
The REML criteria in LMM can be calculated by a feed-forward neural network. Thus, the parameters of the model are obtained as those that maximumize that criteria. The framework of the neural networks allows the seamless calculation of functions and their derivatives. The gradients obtained by backpropagation over the computational graph are used to control the optimization process.
Requires python at least version 3.11. In addition to that, the following packages are dependencies for the use of mn2: pandas, patsy, scipy and torch. Matrix algebra is performed using tensors from torch. The function evaluations and the backpropagation of its gradients is also implemented in torch. Optimization is peformed through the interface implemented in scipy. Data can be be managed with pandas. Model design matrices can be generated by patsy.
pip3 install mn2