ó áp7]c@sÀdZddlmZddlZddlmZddlmZddl m Z d„Z d„Z d e fd „ƒYZd e fd „ƒYZd e fd„ƒYZde fd„ƒYZdS(s´ Created on Mon Jul 26 08:34:59 2010 Author: josef-pktd changes: added offset and zero-inflated version of Poisson - kind of ok, need better test cases, - a nan in ZIP bse, need to check hessian calculations - found error in ZIP loglike - all tests pass with Issues ------ * If true model is not zero-inflated then numerical Hessian for ZIP has zeros for the inflation probability and is not invertible. -> hessian inverts and bse look ok if row and column are dropped, pinv also works * GenericMLE: still get somewhere (where?) "CacheWriteWarning: The attribute 'bse' cannot be overwritten" * bfgs is too fragile, doesn't come back * `nm` is slow but seems to work * need good start_params and their use in genericmle needs to be checked for consistency, set as attribute or method (called as attribute) * numerical hessian needs better scaling * check taking parts out of the loop, e.g. factorial(endog) could be precalculated iÿÿÿÿ(tprint_functionN(tstats(t factorial(tGenericLikelihoodModelcCstjtj||ƒƒS(N(tnptmaxtabs(tarr1tarr2((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pytmaxabs&scCstjtj||dƒƒS(Ni(RRR(RR((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pyt maxabsrel)stNonlinearDeltaCovcBs>eZdZd„Zdd„Zd„Zd„Zd„ZRS(sRAsymptotic covariance by Deltamethod the function is designed for 2d array, with rows equal to the number of equations and columns equal to the number of parameters. 1d params work by chance ? fun: R^{m*k) -> R^{m} where m is number of equations and k is the number of parameters. equations follow Greene cCs||_||_||_dS(N(tfuntparamst cov_params(tselfR R R((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pyt__init__9s  cKsK|dkr|j}n|jddƒddlm}|||j|S(Ntepsilong-Cëâ6?iÿÿÿÿ(t approx_fprime(tNoneR t setdefaulttstatsmodels.tools.numdiffRR (RR tkwdsR((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pytgrad>s   cCs4|jƒ}tjtj||jƒ|jƒ}|S(N(RRtdotRtT(Rtgtcovar((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pytcovEs $cCs|j|jƒS(N(R R (R((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pytexpectedJscCsw|jƒ}|jƒ}tj|ƒ}||}tjtj|jtjj|ƒƒ|ƒ}|tj j ||ƒfS(N( RRRtsizeRRtlinalgtinvRtchi2tsf(Rtvaluetmtvtdftdifftlmstat((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pytwaldNs    -N( t__name__t __module__t__doc__RRRRRR)(((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pyR ,s      t PoissonGMLEcBs eZdZd„Zd„ZRS(sfMaximum Likelihood Estimation of Poisson Model This is an example for generic MLE which has the same statistical model as discretemod.Poisson. Except for defining the negative log-likelihood method, all methods and results are generic. Gradients and Hessian and all resulting statistics are based on numerical differentiation. cCsFtj|j|ƒ}|j}tj|ƒ||tjt|ƒƒS(s Loglikelihood of Poisson model Parameters ---------- params : array-like The parameters of the model. Returns ------- The log likelihood of the model evaluated at `params` Notes -------- .. math:: \ln L=\sum_{i=1}^{n}\left[-\lambda_{i}+y_{i}x_{i}^{\prime}\beta-\ln y_{i}!\right] (RRtexogtendogtexptlogR(RR tXBR/((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pyt nloglikeobsgs cCs\t|dƒst‚n@|j}|j}tjtj||ƒƒ}tj|ddƒSdS(sOreturn frozen scipy.stats distribution with mu at estimated prediction tresulttlociN( thasattrt ValueErrorR4R RR0RRtpoisson(RR.R4R tmu((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pytpredict_distribution|s    (R*R+R,R3R:(((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pyR-Ys  tPoissonOffsetGMLEcBs)eZdZdddd„Zd„ZRS(suMaximum Likelihood Estimation of Poisson Model This is an example for generic MLE which has the same statistical model as discretemod.Poisson but adds offset Except for defining the negative log-likelihood method, all methods and results are generic. Gradients and Hessian and all resulting statistics are based on numerical differentiation. tnonecKsu|dk rF|jdkr4|dd…df}n|jƒ|_n d|_tt|ƒj||d||dS(Nigtmissing(RtndimtraveltoffsettsuperR;R(RR/R.R@R=R((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pyR™s  cCsS|jtj|j|ƒ}|j}tj|ƒ||tjt|ƒƒ}|S(s Loglikelihood of Poisson model Parameters ---------- params : array-like The parameters of the model. Returns ------- The log likelihood of the model evaluated at `params` Notes -------- .. math:: \ln L=\sum_{i=1}^{n}\left[-\lambda_{i}+y_{i}x_{i}^{\prime}\beta-\ln y_{i}!\right] (R@RRR.R/R0R1R(RR R2R/tnloglik((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pyR3©s *N(R*R+R,RRR3(((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pyR;Œs t PoissonZiGMLEcBs)eZdZdddd„Zd„ZRS(sÊMaximum Likelihood Estimation of Poisson Model This is an example for generic MLE which has the same statistical model as discretemod.Poisson but adds offset and zero-inflation. Except for defining the negative log-likelihood method, all methods and results are generic. Gradients and Hessian and all resulting statistics are based on numerical differentiation. There are numerical problems if there is no zero-inflation. R<cKsòtt|ƒj||d|||dk rh|jdkrV|dd…df}n|jƒ|_n d|_|dkr›tj|j dfƒ|_ n|j j d|_ tj tj|j ƒdfƒ|_dg|_|jjdƒdS(NR=igit start_paramstzi(RARCRRR>R?R@RtonestnobsR.tshapetnparamsthstackRDt cloneattrt exog_namestappend(RR/R.R@R=R((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pyRÏs   $ cCs¼|d }ddtj|dƒ}|jtj|j|ƒ}|j}tjd|ƒ tj|ƒ||tjt|ƒƒ}tj|tj||dk ƒƒ ||dk<|S(s Loglikelihood of Poisson model Parameters ---------- params : array-like The parameters of the model. Returns ------- The log likelihood of the model evaluated at `params` Notes -------- .. math:: \ln L=\sum_{i=1}^{n}\left[-\lambda_{i}+y_{i}x_{i}^{\prime}\beta-\ln y_{i}!\right] iÿÿÿÿii(RR0R@RR.R/R1R(RR tbetatgammR2R/RB((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pyR3çs  <2N(R*R+R,RRR3(((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pyRCÀs (R,t __future__RtnumpyRtscipyRtstatsmodels.compat.scipyRtstatsmodels.base.modelRR R tobjectR R-R;RC(((s;lib/python2.7/site-packages/statsmodels/miscmodels/count.pyts   -34