📄 Statsmodels: econometric and statistical modeling with Python

Summary

Economists have mostly done statistical work in proprietary tools such as Stata, SAS, GAUSS, MATLAB and NLOGIT, or in R. Python had no dedicated package for estimating and testing statistical and econometric models. The authors present statsmodels, a SciKit add-on to NumPy/SciPy. It began as Jonathan Taylor's models code, which moved from SciPy to NIPY and was then developed further during Google Summer of Code 2009 and 2010.

The design treats a model as an object that reduces data. A base Model class holds endog and exog data and provides placeholders for fit and predict. LikelihoodModel supports the regression models. Fitting returns Results objects that compute their statistics lazily, only when requested.

To ensure correctness, the developers follow a test-driven workflow. Outputs are checked against R, SAS, Stata or Monte Carlo simulations, and any mismatches are documented. Worked examples, including OLS on the highly collinear Longley data, show a uniform API across models. The paper matters because it laid out a transparent, replicable, open-source foundation for statistical modeling in Python.

Key points

• The main codebase has five model modules:
  • regression (OLS, WLS, GLS)
  • glm (six exponential-family distributions, at least ten link functions)
  • rlm (M-estimators with eight norms)
  • discretemod (Logit, Probit, MNLogit and Poisson, estimated by maximum likelihood)
  • contrast
• Supporting tools include:
  • tests for heteroskedasticity and autocorrelation
  • a framework for linear hypotheses on coefficients
  • descriptive statistics
  • SimpleTable output with LaTeX export
  • a reader for Stata .dta files
• The package ships fourteen datasets. The Longley 1967 data, with a condition number of 456,037, is used to check least-squares accuracy, and parameters are computed by default via the pseudoinverse.
• The development process uses distributed version control branches and a sandbox directory, with tests against reference software and code review on the mailing list. Releases are distributed via PyPI/easy_install.
• Planned additions include time-series models, kernel density estimation and nonparametric regression, panel data, systems of equations, and maximum entropy models.