Code to generate synthetic data experiments and associated results in "Domain constraints improve healthcare risk prediction when outcome data is missing".
- Conda environment. The conda environment we use is specified in domain_constraints.yml.
- Generating synthetic data. We provide code to generate data using three different data generating processes: 1) Heckman model (generate_synthetic_data_heckman.py), 2) Bernoulli-sigmoid model with uniform unobservables (generate_synthetic_data_bin_Y.py), and 3) Bernoulli-sigmoid model with normal unobservables (generate_synthetic_data_bin_Y.py). In all cases we provide cdoe to generate multiple synthetic datasets for which parameters are drawn from a normal distribution with user specified means and standard deviations.
-
Fitting model. We provide code to fit three different models in stan: 1) Heckman model (fit_model_heckman.py), 2) Bernoulli-sigmoid model with uniform unobservables (fit_model_bin_Y.py), and 3) Bernoulli-sigmoid model with normal unobservables (fit_model_bin_Y.py). The Heckman model estimates parameters
$(\beta_Y, \beta_T, \rho, \sigma^2)$ , the Bernoulli-sigmoid model with uniform unobservables estimates parameters$(\beta_Y, \beta_\Delta, \sigma^2)$ and uses a fixed$\alpha=1$ , the Bernoulli-sigmoid model with normal unobservables can either estimate parameters$(\beta_Y, \beta_\Delta, \sigma^2)$ with a fixed$\alpha$ or estimate parameters$(\beta_Y, \beta_\Delta, \alpha)$ with a fixed$\sigma^2$ . - Analyzing model fit. Once models have been fit, figures and results in the paper can be reproduced by running make_figures.ipynb. We provide code to reproduce Figure 2 in the main text of the paper, however this code can be slightly modified to create the other figures in Appendix C. This code relies on a compiled results file called results.csv containing paths to the data and stan samples. More details are included in make_figures.ipynb.