We strive to follow a Test Driven Development (TDD) pattern. All models or statistical functions that are added to the main code base are to have tests versus an existing statistical package, if possible.
Like many packages, statsmodels uses the Nose testing system and the convenient extensions in numpy.testing. Nose itself is an extension of Python's unittest. Nose will find any file, directory, function, or class name that matches the regular expression (?:^|[b_./-])[Tt]est. This is mainly functions that begin with test* and classes that begin with Test*.
You can run all the tests by:
>>> import statsmodels.api as sm >>> sm.test()
You can test submodules by:
NumPy provides a good introduction to unit testing with Nose and NumPy extensions here. It is worth a read for some more details. Here, we will document a few conventions we follow that are worth mentioning. Often we want to test a whole model at once rather than just one function, for example. The following is a pared down version test_discrete.py. In this case, several different models with different options need to be tested. The tests look something like
from numpy.testing import assert_almost_equal import statsmodels.api as sm from results.results_discrete import Spector class CheckDiscreteResults(object): """ res2 are the results. res1 are the values from statsmodels """ def test_params(self): assert_almost_equal(self.res1.params, self.res2.params, 4) decimal_tvalues = 4 def test_tvalues(self): assert_almost_equal(self.res1.params, self.res2.params, self.decimal_tvalues) # ... as many more tests as there are common results class TestProbitNewton(CheckDiscreteResults): """ Tests the Probit model using Newton's method for fitting. """ @classmethod def setupClass(cls): # set up model data = sm.datasets.spector.load() data.exog = sm.add_constant(data.exog) cls.res1 = sm.Probit(data.endog, data.exog).fit(method='newton', disp=0) # set up results res2 = Spector() res2.probit() cls.res2 = res2 # set up precision cls.decimal_tvalues = 3 def test_model_specifc(self): assert_almost_equal(self.res1.foo, self.res2.foo, 4) if __name__ == "__main__": import nose nose.runmodule(argv=[__file__, '-vvs', '-x', '--pdb'], exit=False)
The main workhorse is the CheckDiscreteResults class. Notice that we can set the level of precision for tvalues to be different than the default in the subclass TestProbitNewton. All of the test classes have a setupClass classmethod. Otherwise, Nose would reinstantiate the class before every single test method. If the fitting of the model is time consuming, then this is clearly undesirable. Finally, we have a script at the bottom so that we can run the tests should be running the Python file.
The test results are the final piece of the above example. For many tests, especially those for the models, there are many results against which you would like to test. It makes sense then to separate the hard-coded results from the actual tests to make the tests more readable. If there are only a few results it’s not necessary to separate the results. We often take results from some other statistical package. It is important to document where you got the results from and why they might differ from the results that we get. Each tests folder has a results subdirectory. Consider the folder structure for the discrete models:
tests/ __init__.py test_discrete.py results/ __init__.py results_discrete.py nbinom_resids.csv
It is up to you how best to structure the results. In the discrete model example, you will notice that there are result classes based around particular datasets with a method for loading different model results for that dataset. You can also include text files that hold results to be loaded by results classes if it is easier than putting them in the class itself.