Code for Galichon-Salanie's "Estimating Separable Matching Models"

Usage

Create a virtual environment, e.g. with python3 -m venv env. Activate it with source env/bin/activate.

Install the requirements with pip install -r requirements.txt. Among the packages it downloads are two created by Bernard Salanié: bs_python_utils and cupid_matching. The former is just a set of utility programs. The latter contains code to solve for the stable matching and estimate the parameters of separable matching models with MDE and Poisson-GLM. The code in this folder relies heavily on these two packages, which are documented on Salanie's website: bs_python_utils and cupid_matching.

Choose the parameters in config.py and run the code with python main.py. The program will create a folder Results and save a plot and a pickled file with the estimates for the sample sample_size defined in config.py.

Each simulation sample (that is, n_sim=1) takes a few seconds (4 seconds on a Mac M2 Max 2023) to estimate the Choo and Siow model by the two methods in the paper --- minimum distance and Poisson GLM. The code is parallelized over samples, unless you choose use_multiprocessing=False. By default, it uses all except 2 of your CPUs.

Structure of the code

The master program main.py reads the parameters in config.py.

1. If do_create_samples is True it uses create_samples.py to read the Choo and Siow datasets in the data_dir directory and to create two samples in samples_dir (both directories are specified in config,.py). The two samples correspond to the small and large samples described in Section 6 of the paper. The files created have the marriage patterns by age (*muxy.txt), the margins (*nx.txt and *my.txt), and the variance-covariance matrix of these estimates (*varmus.pkl).
2. It calls read_data.py, which reads the sample defined by sample_size in config.py and prepares it for the simulation. read_data.py also has code to add a small positive number (see zero_guard in config.py) for zero cells; this is used in the MDE simulation.
3. specification.py creates the basis functions according to the specification given by degrees in config.py.
4. Then main.py runs the simulation via simulate.py as defined by config.py.

Configuration

All parameters of the simulation are in config.py:

• do_create_samples, do_simuls, plot_simuls, do_simuls_mde, do_simuls_poisson define what the program does;
• n_sim is the number of simulated samples;
• use_multiprocessing and nb_cpus define the parallelization;
• zero_guard is the small positive number added to zero cells in the sample for MDE estimation;
• degrees is a list of tuples that define the degrees of the polynomials for the basis functions; e.g. an (a,b) tuple means that the basis function is $L_a(x)L_b(y)$, where $x$ and $y$ are the ages of the partners and $L_a$ is the Legendre polynomial of degree $a$. In addition to these terms, the basis functions also include the constant term; $\mathbf{1}(x>y)$, and a term proportional to $\max(x-y,0)$. The function generate_bases in specification.py creates the basis functions.

Questions

Please direct all questions to Bernard Salanie.