This tutorial explains how to quickly implement new maximum likelihood models in statsmodels. We give two examples:
The GenericLikelihoodModel class eases the process by providing tools such as automatic numeric differentiation and a unified interface to scipy optimization functions. Using statsmodels, users can fit new MLE models simply by "plugging-in" a log-likelihood function.
In [ ]:
import numpy as np
from scipy import stats
import statsmodels.api as sm
from statsmodels.base.model import GenericLikelihoodModel
The Spector dataset is distributed with statsmodels. You can access a vector of values for the dependent variable (endog) and a matrix of regressors (exog) like this:
In [ ]:
data = sm.datasets.spector.load_pandas()
exog = data.exog
endog = data.endog
print(sm.datasets.spector.NOTE)
print(data.exog.head())
Them, we add a constant to the matrix of regressors:
In [ ]:
exog = sm.add_constant(exog, prepend=True)
To create your own Likelihood Model, you simply need to overwrite the loglike method.
In [ ]:
class MyProbit(GenericLikelihoodModel):
def loglike(self, params):
exog = self.exog
endog = self.endog
q = 2 * endog - 1
return stats.norm.logcdf(q*np.dot(exog, params)).sum()
Estimate the model and print a summary:
In [ ]:
sm_probit_manual = MyProbit(endog, exog).fit()
print(sm_probit_manual.summary())
Compare your Probit implementation to statsmodels' "canned" implementation:
In [ ]:
sm_probit_canned = sm.Probit(endog, exog).fit()
In [ ]:
print(sm_probit_canned.params)
print(sm_probit_manual.params)
In [ ]:
print(sm_probit_canned.cov_params())
print(sm_probit_manual.cov_params())
Notice that the GenericMaximumLikelihood class provides automatic differentiation, so we did not have to provide Hessian or Score functions in order to calculate the covariance estimates.
Consider a negative binomial regression model for count data with log-likelihood (type NB-2) function expressed as:
$$ \mathcal{L}(\beta_j; y, \alpha) = \sum_{i=1}^n y_i ln \left ( \frac{\alpha exp(X_i'\beta)}{1+\alpha exp(X_i'\beta)} \right ) - \frac{1}{\alpha} ln(1+\alpha exp(X_i'\beta)) + ln \Gamma (y_i + 1/\alpha) - ln \Gamma (y_i+1) - ln \Gamma (1/\alpha) $$with a matrix of regressors $X$, a vector of coefficients $\beta$, and the negative binomial heterogeneity parameter $\alpha$.
Using the nbinom distribution from scipy, we can write this likelihood
simply as:
In [ ]:
import numpy as np
from scipy.stats import nbinom
In [ ]:
def _ll_nb2(y, X, beta, alph):
mu = np.exp(np.dot(X, beta))
size = 1/alph
prob = size/(size+mu)
ll = nbinom.logpmf(y, size, prob)
return ll
In [ ]:
from statsmodels.base.model import GenericLikelihoodModel
In [ ]:
class NBin(GenericLikelihoodModel):
def __init__(self, endog, exog, **kwds):
super(NBin, self).__init__(endog, exog, **kwds)
def nloglikeobs(self, params):
alph = params[-1]
beta = params[:-1]
ll = _ll_nb2(self.endog, self.exog, beta, alph)
return -ll
def fit(self, start_params=None, maxiter=10000, maxfun=5000, **kwds):
# we have one additional parameter and we need to add it for summary
self.exog_names.append('alpha')
if start_params == None:
# Reasonable starting values
start_params = np.append(np.zeros(self.exog.shape[1]), .5)
# intercept
start_params[-2] = np.log(self.endog.mean())
return super(NBin, self).fit(start_params=start_params,
maxiter=maxiter, maxfun=maxfun,
**kwds)
Two important things to notice:
nloglikeobs: This function should return one evaluation of the negative log-likelihood function per observation in your dataset (i.e. rows of the endog/X matrix). start_params: A one-dimensional array of starting values needs to be provided. The size of this array determines the number of parameters that will be used in optimization.That's it! You're done!
The Medpar
dataset is hosted in CSV format at the Rdatasets repository. We use the read_csv
function from the Pandas library to load the data
in memory. We then print the first few columns:
In [ ]:
import statsmodels.api as sm
In [ ]:
medpar = sm.datasets.get_rdataset("medpar", "COUNT", cache=True).data
medpar.head()
The model we are interested in has a vector of non-negative integers as
dependent variable (los), and 5 regressors: Intercept, type2,
type3, hmo, white.
For estimation, we need to create two variables to hold our regressors and the outcome variable. These can be ndarrays or pandas objects.
In [ ]:
y = medpar.los
X = medpar[["type2", "type3", "hmo", "white"]].copy()
X["constant"] = 1
Then, we fit the model and extract some information:
In [ ]:
mod = NBin(y, X)
res = mod.fit()
Extract parameter estimates, standard errors, p-values, AIC, etc.:
In [ ]:
print('Parameters: ', res.params)
print('Standard errors: ', res.bse)
print('P-values: ', res.pvalues)
print('AIC: ', res.aic)
As usual, you can obtain a full list of available information by typing
dir(res).
We can also look at the summary of the estimation results.
In [ ]:
print(res.summary())
We can check the results by using the statsmodels implementation of the Negative Binomial model, which uses the analytic score function and Hessian.
In [ ]:
res_nbin = sm.NegativeBinomial(y, X).fit(disp=0)
print(res_nbin.summary())
In [ ]:
print(res_nbin.params)
In [ ]:
print(res_nbin.bse)
Or we could compare them to results obtained using the MASS implementation for R:
url = 'https://raw.githubusercontent.com/vincentarelbundock/Rdatasets/csv/COUNT/medpar.csv'
medpar = read.csv(url)
f = los~factor(type)+hmo+white
library(MASS)
mod = glm.nb(f, medpar)
coef(summary(mod))
Estimate Std. Error z value Pr(>|z|)
(Intercept) 2.31027893 0.06744676 34.253370 3.885556e-257
factor(type)2 0.22124898 0.05045746 4.384861 1.160597e-05
factor(type)3 0.70615882 0.07599849 9.291748 1.517751e-20
hmo -0.06795522 0.05321375 -1.277024 2.015939e-01
white -0.12906544 0.06836272 -1.887951 5.903257e-02
The statsmodels generic MLE and R parameter estimates agree up to the fourth decimal. The standard errors, however, agree only up to the second decimal. This discrepancy is the result of imprecision in our Hessian numerical estimates. In the current context, the difference between MASS and statsmodels standard error estimates is substantively irrelevant, but it highlights the fact that users who need very precise estimates may not always want to rely on default settings when using numerical derivatives. In such cases, it is better to use analytical derivatives with the LikelihoodModel class.