A Python implementation of Klein and Vella's semiparametric double-index estimator (JoAE, 2009)

This is a Python implementation of "A semiparametric model for binary response and continuous outcomes under index heteroscedasticity" by Klein and Vella (2009, Journal of Applied Econometrics). The code is a simplified version of the code used in Drerup, Enke, and von Gaudecker (2017, Journal of Econometrics). An older version of the code embedded in von Gaudecker's Templates for Reproducible Research Projects in Economics can be found here.

Current features.

• Model estimation
• Tables:
  • coefficients
  • average partial effects of the KV model
  • average partial effects of a similar probit
  • average partial effects of a similar probit alongside OLS results
• Figures:
  • a 3d plot with predicted probabilities
  • a 2d contour plot with joint densities of the indices
  • plots of the average structural functions along the indices
  • plots of the average structural functions alongside a plot of a comparable probit

All of the code below has been tested on a Mac. Let me know if you are interested in a Windows version. The required changes should be minimal.

Note: The code contains some changes that have not fully been tested yet. Any feedback is welcome: tdrerup@uni-bonn.de

Getting started.

While the code may work with your current Python (≥ 3.4) installation, it is helpful to work in a preset environment. To do so, first install either Anaconda or Miniconda. Then type

  source set-env.sh

to install all required packages in the version specified in environment.yml. You may need to use bash set-env.sh instead.

Activate the environment using

  source activate binary_choice_double_index

Whenever you work with the code, be sure to activate the environment beforehand. For newer versions of conda you may need to activate the environment using

  conda activate binary_choice_double_index

Usage Example.

We will show the basic usage of the commands using an example dataset (./data/example.csv). Run this example to make sure that the code works.

The specification we are going to estimate is saved in ./models/example.json. The layout of this file is straightforward. We will discuss the individual components below. To fit the model, type

  python kv_fit.py example

You should see output like the following:

  Starting pilot fit.
  Optimization terminated successfully.
           Current function value: 0.511914
           Iterations: 45
           Function evaluations: 62
           Gradient evaluations: 62
  Starting final fit.
  Optimization terminated successfully.
           Current function value: 0.397809
           Iterations: 46
           Function evaluations: 49
           Gradient evaluations: 49

Fitting the model takes about 1 minute on a 2015 MBP. More complicated models or models with more data can take substantially longer. Results (coefficients, final indices, etc.) will be packaged in a pickle and saved in ./bld/results/.

To obtain tables with

• coefficients
• average partial effects of the KV model
• average partial effects of a similar probit
• average partial effects of a similar probit alongside OLS results

run

  python kv_tables.py example all

Tables will be saved as .tex files inin ./bld/tables. As you can see from the tables, the example model contained one index with 5 variables and one with only 1.

To visualize the results, type

  python kv_figures.py example all

This will create the following figures:

• a 3d plot with predicted probabilities
• a 2d contour plot with joint densities of the indices
• plots of the average structural functions along the indices
• plots of the average structural functions alongside a plot of a comparable probit

All figures will be saved in ./bld/figures.

Details.

For a detailed description of the tables and figures produced by kv_tables.py and kv_figures.py, check The precision of subjective data and the explanatory power of economic models (2017) in the Journal of Econometrics) or the working paper version.

Data.

Store the data as a comma-separated file (.csv) in ./data/. The estimator expects the outcome to be binary (0/1, no categoricals). Explanatory variables are expected to be numerical, so dummies for categorical variables need to be created beforehand.

Model Setup.

The folder ./models/ contains json files with information about the model(s) to estimate. Each json contains the following entries:

• data: The name of the file that contains the data for this model.
• y_name: The name of the outcome variable (as stored in data).
• labels: A dictionary with the names / labels for the outcome variable and indices (mainly used for plotting)
• index_colnames: The names of the variables to include in the two indices:
  • index_1: A list of the variables for index 1
  • index_2: A list of the variables for index 2
• pilot: Settings for the KV pilot estimator
  • coeffs_start: Starting values for the coefficients for each index
  • trim: Percentiles for trimming of the densities
  • n_smoothing_stages: Number of smoothing stages (3 is currently not implemented)
  • maxiter: Maximum number of iterations
• final: Settings for the final KV estimator
  • trim: Percentiles for trimming of the densities
  • maxiter: Maximum number of iterations

Use the json for the example as a template for your own specification.