Source code for statsmodels.distributions.discrete

from statsmodels.compat.pandas import deprecate_kwarg
from statsmodels.compat.scipy import apply_where

import numpy as np
from scipy.special import gammaln
from scipy.stats import nbinom, poisson, rv_discrete

from statsmodels.base.model import GenericLikelihoodModel
from statsmodels.tools.rng_qrng import check_random_state


class genpoisson_p_gen(rv_discrete):
    """Generalized Poisson distribution"""

    def _argcheck(self, mu, alpha, p):
        return (mu >= 0) & ~np.isnan(alpha) & (p > 0)

    def _logpmf(self, x, mu, alpha, p):
        mu_p = mu ** (p - 1.0)
        a1 = np.maximum(np.nextafter(0, 1), 1 + alpha * mu_p)
        a2 = np.maximum(np.nextafter(0, 1), mu + (a1 - 1.0) * x)
        logpmf_ = np.log(mu) + (x - 1.0) * np.log(a2)
        logpmf_ -= x * np.log(a1) + gammaln(x + 1.0) + a2 / a1
        return logpmf_

    def _pmf(self, x, mu, alpha, p):
        return np.exp(self._logpmf(x, mu, alpha, p))

    def mean(self, mu, alpha, p):
        return mu

    def var(self, mu, alpha, p):
        dispersion_factor = (1 + alpha * mu ** (p - 1)) ** 2
        var = dispersion_factor * mu
        return var


genpoisson_p = genpoisson_p_gen(name="genpoisson_p", longname="Generalized Poisson")


class zipoisson_gen(rv_discrete):
    """Zero Inflated Poisson distribution"""

    def _argcheck(self, mu, w):
        return (mu > 0) & (w >= 0) & (w <= 1)

    def _logpmf(self, x, mu, w):
        return apply_where(
            x != 0,
            (x, mu, w),
            (lambda x, mu, w: np.log(1.0 - w) + x * np.log(mu) - gammaln(x + 1.0) - mu),
            fill_value=np.log(w + (1.0 - w) * np.exp(-mu)),
        )

    def _pmf(self, x, mu, w):
        return np.exp(self._logpmf(x, mu, w))

    def _cdf(self, x, mu, w):
        # construct cdf from standard poisson's cdf and the w inflation of zero
        return w + poisson(mu=mu).cdf(x) * (1 - w)

    def _ppf(self, q, mu, w):
        # we just translated and stretched q to remove zi
        q_mod = (q - w) / (1 - w)
        x = poisson(mu=mu).ppf(q_mod)
        # set to zero if in the zi range
        if isinstance(x, np.ndarray):
            x[q < w] = 0
        elif np.isscalar(x) and q < w:
            return 0.0
        return x

    def mean(self, mu, w):
        return (1 - w) * mu

    def var(self, mu, w):
        dispersion_factor = 1 + w * mu
        var = dispersion_factor * self.mean(mu, w)
        return var

    def _moment(self, n, mu, w):
        return (1 - w) * poisson.moment(n, mu)


zipoisson = zipoisson_gen(name="zipoisson", longname="Zero Inflated Poisson")


class zigeneralizedpoisson_gen(rv_discrete):
    """Zero Inflated Generalized Poisson distribution"""

    def _argcheck(self, mu, alpha, p, w):
        return (mu > 0) & (w >= 0) & (w <= 1)

    def _logpmf(self, x, mu, alpha, p, w):
        return apply_where(
            x != 0,
            (x, mu, alpha, p, w),
            (
                lambda x, mu, alpha, p, w: np.log(1.0 - w)
                + genpoisson_p.logpmf(x, mu, alpha, p)
            ),
            fill_value=np.log(w + (1.0 - w) * genpoisson_p.pmf(x, mu, alpha, p)),
        )

    def _pmf(self, x, mu, alpha, p, w):
        return np.exp(self._logpmf(x, mu, alpha, p, w))

    def mean(self, mu, alpha, p, w):
        return (1 - w) * mu

    def var(self, mu, alpha, p, w):
        p = p - 1
        dispersion_factor = (1 + alpha * mu**p) ** 2 + w * mu
        var = dispersion_factor * self.mean(mu, alpha, p, w)
        return var


zigenpoisson = zigeneralizedpoisson_gen(
    name="zigenpoisson", longname="Zero Inflated Generalized Poisson"
)


class zinegativebinomial_gen(rv_discrete):
    """Zero Inflated Generalized Negative Binomial distribution"""

    def _argcheck(self, mu, alpha, p, w):
        return (mu > 0) & (w >= 0) & (w <= 1)

    def _logpmf(self, x, mu, alpha, p, w):
        s, p = self.convert_params(mu, alpha, p)
        return apply_where(
            x != 0,
            (x, s, p, w),
            (lambda x, s, p, w: np.log(1.0 - w) + nbinom.logpmf(x, s, p)),
            fill_value=np.log(w + (1.0 - w) * nbinom.pmf(x, s, p)),
        )

    def _pmf(self, x, mu, alpha, p, w):
        return np.exp(self._logpmf(x, mu, alpha, p, w))

    def _cdf(self, x, mu, alpha, p, w):
        s, p = self.convert_params(mu, alpha, p)
        # construct cdf from standard negative binomial cdf
        # and the w inflation of zero
        return w + nbinom.cdf(x, s, p) * (1 - w)

    def _ppf(self, q, mu, alpha, p, w):
        s, p = self.convert_params(mu, alpha, p)
        # we just translated and stretched q to remove zi
        q_mod = (q - w) / (1 - w)
        x = nbinom.ppf(q_mod, s, p)
        # set to zero if in the zi range
        if isinstance(x, np.ndarray):
            x[q < w] = 0
        elif np.isscalar(x) and q < w:
            return 0.0
        return x

    def mean(self, mu, alpha, p, w):
        return (1 - w) * mu

    def var(self, mu, alpha, p, w):
        dispersion_factor = 1 + alpha * mu ** (p - 1) + w * mu
        var = dispersion_factor * self.mean(mu, alpha, p, w)
        return var

    def _moment(self, n, mu, alpha, p, w):
        s, p = self.convert_params(mu, alpha, p)
        return (1 - w) * nbinom.moment(n, s, p)

    def convert_params(self, mu, alpha, p):
        size = 1.0 / alpha * mu ** (2 - p)
        prob = size / (size + mu)
        return (size, prob)


zinegbin = zinegativebinomial_gen(
    name="zinegbin", longname="Zero Inflated Generalized Negative Binomial"
)


class truncatedpoisson_gen(rv_discrete):
    """Truncated Poisson discrete random variable"""

    # TODO: need cdf, and rvs

    def _argcheck(self, mu, truncation):
        # this does not work
        # vector bound breaks some generic methods
        # self.a = truncation + 1 # max(truncation + 1, 0)
        return (mu >= 0) & (truncation >= -1)

    def _get_support(self, mu, truncation):
        return truncation + 1, self.b

    def _logpmf(self, x, mu, truncation):
        pmf = 0
        for i in range(int(np.max(truncation)) + 1):
            pmf += poisson.pmf(i, mu)

        # Skip pmf = 1 to avoid warnings
        log_1_m_pmf = np.full_like(pmf, -np.inf)
        loc = pmf > 1
        log_1_m_pmf[loc] = np.nan
        loc = pmf < 1
        log_1_m_pmf[loc] = np.log(1 - pmf[loc])
        logpmf_ = poisson.logpmf(x, mu) - log_1_m_pmf
        # logpmf_[x < truncation + 1] = - np.inf
        return logpmf_

    def _pmf(self, x, mu, truncation):
        return np.exp(self._logpmf(x, mu, truncation))


truncatedpoisson = truncatedpoisson_gen(
    name="truncatedpoisson", longname="Truncated Poisson"
)


class truncatednegbin_gen(rv_discrete):
    """Truncated Generalized Negative Binomial (NB-P) discrete random variable"""

    def _argcheck(self, mu, alpha, p, truncation):
        return (mu >= 0) & (truncation >= -1)

    def _get_support(self, mu, alpha, p, truncation):
        return truncation + 1, self.b

    def _logpmf(self, x, mu, alpha, p, truncation):
        size, prob = self.convert_params(mu, alpha, p)
        pmf = 0
        for i in range(int(np.max(truncation)) + 1):
            pmf += nbinom.pmf(i, size, prob)

        # Skip pmf = 1 to avoid warnings
        log_1_m_pmf = np.full_like(pmf, -np.inf)
        loc = pmf > 1
        log_1_m_pmf[loc] = np.nan
        loc = pmf < 1
        log_1_m_pmf[loc] = np.log(1 - pmf[loc])
        logpmf_ = nbinom.logpmf(x, size, prob) - log_1_m_pmf
        # logpmf_[x < truncation + 1] = - np.inf
        return logpmf_

    def _pmf(self, x, mu, alpha, p, truncation):
        return np.exp(self._logpmf(x, mu, alpha, p, truncation))

    def convert_params(self, mu, alpha, p):
        size = 1.0 / alpha * mu ** (2 - p)
        prob = size / (size + mu)
        return (size, prob)


truncatednegbin = truncatednegbin_gen(
    name="truncatednegbin", longname="Truncated Generalized Negative Binomial"
)


[docs] class DiscretizedCount(rv_discrete): """Count distribution based on discretized distribution Parameters ---------- distr : distribution instance d_offset : float, optional Offset for integer interval, default is zero. The discrete random variable is ``y = floor(x + offset)`` where x is the continuous random variable. Warning: not verified for all methods. add_scale : bool, optional If True (default), then the scale of the base distribution is added as parameter for the discrete distribution. The scale parameter is in the last position. kwds : keyword arguments The extra keyword arguments are used delegated to the ``__init__`` of the super class. Their usage has not been checked, e.g., currently the support of the distribution is assumed to be all non-negative integers. Notes ----- `loc` argument is currently not supported, scale is not available for discrete distributions in scipy. The scale parameter of the underlying continuous distribution is the last shape parameter in this DiscretizedCount distribution if ``add_scale`` is True. The implementation was based mainly on [1]_ and [2]_. However, many new discrete distributions have been developed based on the approach that we use here. Note, that in many cases authors reparameterize the distribution, while this class inherits the parameterization from the underlying continuous distribution. References ---------- .. [1] Chakraborty, Subrata, and Dhrubajyoti Chakravarty. "Discrete gamma distributions: Properties and parameter estimations." Communications in Statistics-Theory and Methods 41, no. 18 (2012): 3301-3324. .. [2] Alzaatreh, Ayman, Carl Lee, and Felix Famoye. 2012. “On the Discrete Analogues of Continuous Distributions.” Statistical Methodology 9 (6): 589-603. """ def __new__(cls, *args, **kwds): # rv_discrete.__new__ does not allow `kwds`, skip it # only does dispatch to multinomial return super(rv_discrete, cls).__new__(cls) def __init__(self, distr, d_offset=0, add_scale=True, **kwds): # kwds are extras in rv_discrete self.distr = distr self.d_offset = d_offset self._ctor_param = distr._ctor_param self.add_scale = add_scale if distr.shapes is not None: self.k_shapes = len(distr.shapes.split(",")) if add_scale: kwds.update({"shapes": distr.shapes + ", s"}) self.k_shapes += 1 # no shape parameters in underlying distribution elif add_scale: kwds.update({"shapes": "s"}) self.k_shapes = 1 else: self.k_shapes = 0 super().__init__(**kwds) def _updated_ctor_param(self): dic = super()._updated_ctor_param() dic["distr"] = self.distr return dic def _unpack_args(self, args): if self.add_scale: scale = args[-1] args = args[:-1] else: scale = 1 return args, scale @deprecate_kwarg("random_state", "rng") def _rvs(self, *args, size=None, rng=None): """Generate random variates. Parameters ---------- *args Shape parameters and scale forwarded to the underlying continuous distribution. size : int or tuple of ints, optional Number of random variates to generate. rng : int, array_like of int, numpy.random.Generator, or numpy.random.RandomState, optional If `rng` is None, a new ``Generator`` is created using fresh entropy from the operating system. If `rng` is an int or array of ints, a new ``Generator`` is created, seeded with `rng`. If `rng` is already a ``Generator`` or ``RandomState`` instance, that instance is used. random_state : int, array_like of int, numpy.random.Generator, or numpy.random.RandomState, optional .. deprecated:: 0.15 random_state has been deprecated. In-line with SPEC-007, use rng for passing a random number generator or seed. Returns ------- rv : ndarray Random variates from the discretized distribution. """ args, scale = self._unpack_args(args) if size is None: size = getattr(self, "_size", 1) rng = check_random_state(rng) rv = np.trunc( self.distr.rvs(*args, scale=scale, size=size, random_state=rng) + self.d_offset ) return rv def _pmf(self, x, *args): distr = self.distr if self.d_offset != 0: x = x + self.d_offset args, scale = self._unpack_args(args) p = distr.sf(x, *args, scale=scale) - distr.sf(x + 1, *args, scale=scale) return p def _cdf(self, x, *args): distr = self.distr args, scale = self._unpack_args(args) if self.d_offset != 0: x = x + self.d_offset p = distr.cdf(x + 1, *args, scale=scale) return p def _sf(self, x, *args): distr = self.distr args, scale = self._unpack_args(args) if self.d_offset != 0: x = x + self.d_offset p = distr.sf(x + 1, *args, scale=scale) return p def _ppf(self, p, *args): distr = self.distr args, scale = self._unpack_args(args) qc = distr.ppf(p, *args, scale=scale) if self.d_offset != 0: qc = qc + self.d_offset q = np.floor(qc * (1 - 1e-15)) return q def _isf(self, p, *args): distr = self.distr args, scale = self._unpack_args(args) qc = distr.isf(p, *args, scale=scale) if self.d_offset != 0: qc = qc + self.d_offset q = np.floor(qc * (1 - 1e-15)) return q
[docs] class DiscretizedModel(GenericLikelihoodModel): """experimental model to fit discretized distribution Count models based on discretized distributions can be used to model data that is under- or over-dispersed relative to Poisson or that has heavier tails. Parameters ---------- endog : array_like, 1-D Univariate data for fitting the distribution. exog : None Explanatory variables are not supported. The ``exog`` argument is only included for consistency in the signature across models. distr : DiscretizedCount instance (required) Instance of a DiscretizedCount distribution. See Also -------- DiscretizedCount Examples -------- >>> from scipy import stats >>> from statsmodels.distributions.discrete import ( ... DiscretizedCount, DiscretizedModel) >>> dd = DiscretizedCount(stats.gamma) >>> mod = DiscretizedModel(y, distr=dd) >>> res = mod.fit() >>> probs = res.predict(which="probs", k_max=5) """ def __init__(self, endog, exog=None, distr=None): if exog is not None: raise ValueError("exog is not supported") super().__init__(endog, exog, distr=distr) self._init_keys.append("distr") self.df_resid = len(endog) - distr.k_shapes self.df_model = 0 self.k_extra = distr.k_shapes # no constant subtracted self.k_constant = 0 self.nparams = distr.k_shapes # needed for start_params self.start_params = 0.5 * np.ones(self.nparams)
[docs] def loglike(self, params): """ Log-likelihood of the discretized distribution. Parameters ---------- params : array_like Shape (and scale) parameters of the underlying continuous distribution. Returns ------- float Sum of the log-probabilities of `endog` given `params`. """ # this does not allow exog yet, # model `params` are also distribution `args` # For regression model this needs to be replaced by a conversion method args = params ll = np.log(self.distr._pmf(self.endog, *args)) return ll.sum()
[docs] def predict(self, params, exog=None, which=None, k_max=20): """ Predict values for the discretized distribution. Parameters ---------- params : array_like Shape (and scale) parameters of the underlying continuous distribution. exog : None Explanatory variables are not supported. The ``exog`` argument is only included for consistency in the signature across models. which : str, optional Statistic to predict. The only currently implemented value is "probs", which returns the probability mass function evaluated at ``0, 1, ..., k_max - 1``. Any other value, including the default of None, raises a ``ValueError``. k_max : int, optional Number of support points, starting at zero, for which probabilities are returned. Returns ------- ndarray Probability mass function evaluated at ``0, 1, ..., k_max - 1``. """ if exog is not None: raise ValueError("exog is not supported") args = params if which == "probs": pr = self.distr.pmf(np.arange(k_max), *args) return pr else: raise ValueError('only which="probs" is currently implemented')
[docs] def get_distr(self, params): """ Frozen distribution instance of the discrete distribution. Parameters ---------- params : array_like Shape (and scale) parameters of the underlying continuous distribution. Returns ------- rv_discrete_frozen Frozen distribution instance of the discretized distribution with `params` substituted for the shape parameters. """ args = params distr = self.distr(*args) return distr