# Maximum Likelihood Estimation (Generic models)ΒΆ

This tutorial explains how to quickly implement new maximum likelihood models in statsmodels. We give two examples:

1. Probit model for binary dependent variables
2. Negative binomial model for count data

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.

## Example 1: Probit model¶

In [ ]:
from __future__ import print_function
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)


Them, we add a constant to the matrix of regressors:

In [ ]:
exog = sm.add_constant(exog, prepend=True)

::

Number of Observations - 32

Number of Variables - 4

Variable name definitions::

improved.  1 indicates an improvement.
TUCE  - Test score on economics test
PSI   - participation in program
GPA   - Student's grade point average

GPA  TUCE  PSI
0  2.66    20    0
1  2.89    22    0
2  3.28    24    0
3  2.92    12    0
4  4.00    21    0


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()

Optimization terminated successfully.
Current function value: 0.400588
Iterations: 292
Function evaluations: 494
MyProbit Results
==============================================================================
Model:                       MyProbit   AIC:                             33.64
Method:            Maximum Likelihood   BIC:                             39.50
Date:                Mon, 20 Jul 2015
Time:                        17:43:29
No. Observations:                  32
Df Residuals:                      28
Df Model:                           3
==============================================================================
coef    std err          z      P>|z|      [95.0% Conf. Int.]
------------------------------------------------------------------------------
const         -7.4523      2.542     -2.931      0.003       -12.435    -2.469
GPA            1.6258      0.694      2.343      0.019         0.266     2.986
TUCE           0.0517      0.084      0.617      0.537        -0.113     0.216
PSI            1.4263      0.595      2.397      0.017         0.260     2.593
==============================================================================

In [ ]:
print(sm_probit_canned.params)
print(sm_probit_manual.params)

Optimization terminated successfully.
Current function value: 0.400588
Iterations 6

In [ ]:
print(sm_probit_canned.cov_params())
print(sm_probit_manual.cov_params())

const   -7.452320
GPA      1.625810
TUCE     0.051729
PSI      1.426332
dtype: float64
[-7.45233176  1.62580888  0.05172971  1.42631954]


Notice that the GenericMaximumLikelihood class provides automatic differentiation, so we didn't have to provide Hessian or Score functions in order to calculate the covariance estimates.

## Example 2: Negative Binomial Regression for Count Data¶

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

          const       GPA      TUCE       PSI
const  6.464166 -1.169668 -0.101173 -0.594792
GPA   -1.169668  0.481473 -0.018914  0.105439
TUCE  -0.101173 -0.018914  0.007038  0.002472
PSI   -0.594792  0.105439  0.002472  0.354070
[[  6.46416770e+00  -1.16966617e+00  -1.01173180e-01  -5.94788993e-01]
[ -1.16966617e+00   4.81472116e-01  -1.89134586e-02   1.05438227e-01]
[ -1.01173180e-01  -1.89134586e-02   7.03758392e-03   2.47189191e-03]
[ -5.94788993e-01   1.05438227e-01   2.47189191e-03   3.54069512e-01]]

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


### New Model Class¶

We create a new model class which inherits from GenericLikelihoodModel:

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!

### Usage Example¶

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



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"]]
X["constant"] = 1


Then, we fit the model and extract some information:

In [ ]:
mod = NBin(y, X)
res = mod.fit()

/Users/tom.augspurger/Envs/py3/lib/python3.4/site-packages/IPython/kernel/__main__.py:3: SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead

See the the caveats in the documentation: http://pandas.pydata.org/pandas-docs/stable/indexing.html#indexing-view-versus-copy
app.launch_new_instance()


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)

Optimization terminated successfully.
Current function value: 3.209014
Iterations: 805
Function evaluations: 1238


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())

Parameters:  [ 0.2212642   0.70613942 -0.06798155 -0.12903932  2.31026565  0.44575147]
Standard errors:  [ 0.05059259  0.07613047  0.05326097  0.06854141  0.06794692  0.01981542]
P-values:  [  1.22297920e-005   1.76978024e-020   2.01819161e-001   5.97481600e-002
2.15079626e-253   4.62688439e-112]
AIC:  9604.95320583


### Testing¶

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())

                                 NBin Results
==============================================================================
Dep. Variable:                    los   Log-Likelihood:                -4797.5
Model:                           NBin   AIC:                             9605.
Method:            Maximum Likelihood   BIC:                             9632.
Date:                Mon, 20 Jul 2015
Time:                        17:43:31
No. Observations:                1495
Df Residuals:                    1490
Df Model:                           4
==============================================================================
coef    std err          z      P>|z|      [95.0% Conf. Int.]
------------------------------------------------------------------------------
type2          0.2213      0.051      4.373      0.000         0.122     0.320
type3          0.7061      0.076      9.275      0.000         0.557     0.855
hmo           -0.0680      0.053     -1.276      0.202        -0.172     0.036
white         -0.1290      0.069     -1.883      0.060        -0.263     0.005
constant       2.3103      0.068     34.001      0.000         2.177     2.443
alpha          0.4458      0.020     22.495      0.000         0.407     0.485
==============================================================================

In [ ]:
print(res_nbin.params)

                     NegativeBinomial Regression Results
==============================================================================
Dep. Variable:                    los   No. Observations:                 1495
Model:               NegativeBinomial   Df Residuals:                     1490
Method:                           MLE   Df Model:                            4
Date:                Mon, 20 Jul 2015   Pseudo R-squ.:                 0.01215
Time:                        17:43:31   Log-Likelihood:                -4797.5
converged:                       True   LL-Null:                       -4856.5
LLR p-value:                 1.404e-24
==============================================================================
coef    std err          z      P>|z|      [95.0% Conf. Int.]
------------------------------------------------------------------------------
type2          0.2212      0.051      4.373      0.000         0.122     0.320
type3          0.7062      0.076      9.276      0.000         0.557     0.855
hmo           -0.0680      0.053     -1.277      0.202        -0.172     0.036
white         -0.1291      0.069     -1.883      0.060        -0.263     0.005
constant       2.3103      0.068     34.001      0.000         2.177     2.443
alpha          0.4458      0.020     22.495      0.000         0.407     0.485
==============================================================================

In [ ]:
print(res_nbin.bse)

type2       0.221231
type3       0.706175
hmo        -0.067990
white      -0.129065
constant    2.310288
alpha       0.445758
dtype: float64


Or we could compare them to results obtained using the MASS implementation for R:

url = 'http://vincentarelbundock.github.com/Rdatasets/csv/COUNT/medpar.csv'
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



### Numerical precision¶

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.