An overview of the model, examples, references, and other documentation can be found on Read the Docs.
PyBLP is a Python 3 implementation of routines for estimating the demand for differentiated products with BLP-type random coefficients logit models. This package was created by Jeff Gortmaker in collaboration with Chris Conlon.
Development of the package has been guided by the work of many researchers and practitioners. For a full list of references, including the original work of Berry, Levinsohn, and Pakes (1995), refer to the references section of the documentation.
If you use PyBLP in your research, we ask that you also cite Conlon and Gortmaker (2020), which describes the advances implemented in the package.
@article{PyBLP,
author = {Conlon, Christopher and Gortmaker, Jeff},
title = {Best practices for differentiated products demand estimation with {PyBLP}},
journal = {The RAND Journal of Economics},
volume = {51},
number = {4},
pages = {1108-1161},
doi = {https://doi.org/10.1111/1756-2171.12352},
url = {https://onlinelibrary.wiley.com/doi/abs/10.1111/1756-2171.12352},
eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1111/1756-2171.12352},
year = {2020}
}
The PyBLP package has been tested on Python versions 3.6 through 3.9. The SciPy instructions for installing related packages is a good guide for how to install a scientific Python environment. A good choice is the Anaconda Distribution, since it comes packaged with the following PyBLP dependencies: NumPy, SciPy, SymPy, and Patsy. For absorption of high dimension fixed effects, PyBLP also depends on its companion package PyHDFE, which will be installed when PyBLP is installed.
However, PyBLP may not work with old versions of its dependencies. You can update PyBLP's Anaconda dependencies with:
conda update numpy scipy sympy patsy
You can update PyHDFE with:
pip install --upgrade pyhdfe
You can install the current release of PyBLP with pip:
pip install pyblp
You can upgrade to a newer release with the --upgrade flag:
pip install --upgrade pyblp
If you lack permissions, you can install PyBLP in your user directory with the --user flag:
pip install --user pyblp
Alternatively, you can download a wheel or source archive from PyPI. You can find the latest development code on GitHub and the latest development documentation here.
Once installed, PyBLP can be incorporated into projects written in many other languages with the help of various tools that enable interoperability with Python.
For example, the reticulate package makes interacting with PyBLP in R straightforward (when supported, Python objects can be converted to their R counterparts with the py_to_r function):
library(reticulate)
pyblp <- import("pyblp", convert=FALSE)
pyblp$options$flush_output <- TRUE
Similarly, PyCall can be used to incorporate PyBLP into a Julia workflow:
using PyCall
pyblp = pyimport("pyblp")
The py command serves a similar purpose in MATLAB:
py.pyblp
- R-style formula interface
- Bertrand-Nash supply-side moments
- Multiple equation GMM
- Demographic interactions
- Product-specific demographics
- Flexible micro moments that can match statistics based on survey data
- Support for micro moments based on second choice data
- Fixed effect absorption
- Nonlinear functions of product characteristics
- Concentrating out linear parameters
- Flexible random coefficient distributions
- Parameter bounds and constraints
- Random coefficients nested logit (RCNL)
- Approximation to the pure characteristics model
- Varying nesting parameters across groups
- Logit and nested logit benchmarks
- Classic BLP instruments
- Differentiation instruments
- Optimal instruments
- Adjustments for simulation error
- Tests of overidentifying and model restrictions
- Parametric boostrapping post-estimation outputs
- Elasticities and diversion ratios
- Marginal costs and markups
- Passthrough calculations
- Profits and consumer surplus
- Newton and fixed point methods for computing pricing equilibria
- Merger simulation
- Custom counterfactual simulation
- Synthetic data construction
- SciPy or Artleys Knitro optimization
- Fixed point acceleration
- Monte Carlo, quasi-random sequences, quadrature, and sparse grids
- Importance sampling
- Custom optimization and iteration routines
- Robust and clustered errors
- Linear or log-linear marginal costs
- Partial ownership matrices
- Analytic gradients
- Finite difference Hessians
- Market-by-market parallelization
- Extended floating point precision
- Robust error handling
- Fast, "Robust," and Approximately Correct (FRAC) estimation
- Analytic Hessians
- Mathematical Program with Equilibrium Constraints (MPEC)
- Generalized Empirical Likelihood (GEL)
- Discrete types
Please use the GitHub issue tracker to submit bugs or to request features.