scikit-survival 0.26.0__cp314-cp314-macosx_10_15_x86_64.whl

sksurv/nonparametric.py

@@ -0,0 +1,911 @@
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with this program.  If not, see <http://www.gnu.org/licenses/>.
+import numbers
+
+import numpy as np
+from scipy import stats
+from sklearn.base import BaseEstimator
+from sklearn.utils._param_validation import Interval, StrOptions
+from sklearn.utils.validation import check_array, check_consistent_length, check_is_fitted
+
+from .util import check_y_survival
+
+__all__ = [
+    "CensoringDistributionEstimator",
+    "kaplan_meier_estimator",
+    "nelson_aalen_estimator",
+    "ipc_weights",
+    "SurvivalFunctionEstimator",
+    "cumulative_incidence_competing_risks",
+]
+
+
+def _compute_counts(event, time, order=None):
+    """Count right-censored and uncensored samples at each unique time point.
+
+    Parameters
+    ----------
+    event : ndarray
+        Boolean event indicator.
+        Integer in the case of multiple risks.
+        Zero means right-censored event.
+        Positive values for each of the possible risk events.
+
+    time : ndarray
+        Survival time or time of censoring.
+
+    order : ndarray or None
+        Indices to order time in ascending order.
+        If None, order will be computed.
+
+    Returns
+    -------
+    times : ndarray
+        Unique time points.
+
+    n_events : ndarray
+        Number of events at each time point.
+        2D array with shape `(n_unique_time_points, n_risks + 1)` in the case of competing risks.
+
+    n_at_risk : ndarray
+        Number of samples that have not been censored or have not had an event at each time point.
+
+    n_censored : ndarray
+        Number of censored samples at each time point.
+    """
+    n_samples = event.shape[0]
+    n_risks = event.max() if (np.issubdtype(event.dtype, np.integer) and event.max() > 1) else 0
+
+    if order is None:
+        order = np.argsort(time, kind="mergesort")
+
+    uniq_times = np.empty(n_samples, dtype=time.dtype)
+    uniq_events = np.empty((n_samples, n_risks + 1), dtype=int)
+    uniq_counts = np.empty(n_samples, dtype=int)
+
+    i = 0
+    prev_val = time[order[0]]
+    j = 0
+    while True:
+        count_event = np.zeros(n_risks + 1, dtype=int)
+        count = 0
+        while i < n_samples and prev_val == time[order[i]]:
+            event_type = event[order[i]]
+            if event_type:
+                count_event[0] += 1
+                if n_risks:
+                    count_event[event_type] += 1
+
+            count += 1
+            i += 1
+
+        uniq_times[j] = prev_val
+        uniq_events[j] = count_event
+        uniq_counts[j] = count
+        j += 1
+
+        if i == n_samples:
+            break
+
+        prev_val = time[order[i]]
+
+    times = np.resize(uniq_times, j)
+    total_count = np.resize(uniq_counts, j)
+    if n_risks:
+        n_events = np.resize(uniq_events, (j, n_risks + 1))
+        n_censored = total_count - n_events[:, 0]
+    else:
+        n_events = np.resize(uniq_events, j)
+        n_censored = total_count - n_events
+
+    # offset cumulative sum by one
+    total_count = np.r_[0, total_count]
+    n_at_risk = n_samples - np.cumsum(total_count)
+
+    return times, n_events, n_at_risk[:-1], n_censored
+
+
+def _compute_counts_truncated(event, time_enter, time_exit):
+    """Compute counts for left truncated and right-censored survival data.
+
+    Parameters
+    ----------
+    event : ndarray
+        Boolean event indicator.
+
+    time_enter : ndarray
+        Time when a subject entered the study.
+
+    time_exit : ndarray
+        Time when a subject left the study due to an
+        event or censoring.
+
+    Returns
+    -------
+    times : ndarray
+        Unique time points.
+
+    n_events : ndarray
+        Number of events at each time point.
+
+    n_at_risk : ndarray
+        Number of samples that are censored or have an event at each time point.
+    """
+    if (time_enter > time_exit).any():
+        raise ValueError("exit time must be larger start time for all samples")
+
+    n_samples = event.shape[0]
+
+    uniq_times = np.sort(np.unique(np.r_[time_enter, time_exit]), kind="mergesort")
+    total_counts = np.empty(len(uniq_times), dtype=int)
+    event_counts = np.empty(len(uniq_times), dtype=int)
+
+    order_enter = np.argsort(time_enter, kind="mergesort")
+    order_exit = np.argsort(time_exit, kind="mergesort")
+    s_time_enter = time_enter[order_enter]
+    s_time_exit = time_exit[order_exit]
+
+    t0 = uniq_times[0]
+    # everything larger is included
+    idx_enter = np.searchsorted(s_time_enter, t0, side="right")
+    # everything smaller is excluded
+    idx_exit = np.searchsorted(s_time_exit, t0, side="left")
+
+    total_counts[0] = idx_enter
+    # except people die on the day they enter
+    event_counts[0] = 0
+
+    for i in range(1, len(uniq_times)):
+        ti = uniq_times[i]
+
+        while idx_enter < n_samples and s_time_enter[idx_enter] < ti:
+            idx_enter += 1
+
+        while idx_exit < n_samples and s_time_exit[idx_exit] < ti:
+            idx_exit += 1
+
+        risk_set = np.setdiff1d(order_enter[:idx_enter], order_exit[:idx_exit], assume_unique=True)
+        total_counts[i] = len(risk_set)
+
+        count_event = 0
+        k = idx_exit
+        while k < n_samples and s_time_exit[k] == ti:
+            if event[order_exit[k]]:
+                count_event += 1
+            k += 1
+        event_counts[i] = count_event
+
+    return uniq_times, event_counts, total_counts
+
+
+def _ci_logmlog(s, sigma_t, conf_level):
+    r"""Compute the pointwise log-minus-log transformed confidence intervals.
+    s refers to the prob_survival or the cum_inc (for the competing risks case).
+    sigma_t is the square root of the variance of the log of the estimator of s.
+
+    .. math::
+
+        \sigma_t = \mathrm{Var}(\log(\hat{S}(t)))
+    """
+    eps = np.finfo(s.dtype).eps
+    mask = s > eps
+    log_p = np.zeros_like(s)
+    np.log(s, where=mask, out=log_p)
+    theta = np.zeros_like(s)
+    np.true_divide(sigma_t, log_p, where=log_p < -eps, out=theta)
+
+    z = stats.norm.isf((1.0 - conf_level) / 2.0)
+    theta = z * np.multiply.outer([-1, 1], theta)
+    ci = np.exp(np.exp(theta) * log_p)
+    ci[:, ~mask] = 0.0
+    return ci
+
+
+def _km_ci_estimator(prob_survival, ratio_var, conf_level, conf_type):
+    """Helper to compute confidence intervals for the Kaplan-Meier estimate.
+
+    Parameters
+    ----------
+    prob_survival : ndarray, shape = (n_times,)
+        Survival probability at each unique time point.
+
+    ratio_var : ndarray, shape = (n_times,)
+        The variance ratio term for each unique time point.
+
+    conf_level : float
+        The level for a two-sided confidence interval.
+
+    conf_type : {'log-log'}
+        The type of confidence intervals to estimate.
+
+    Returns
+    -------
+    ci : ndarray, shape = (2, n_times)
+        Pointwise confidence interval.
+    """
+    if conf_type not in {"log-log"}:
+        raise ValueError(f"conf_type must be None or a str among {{'log-log'}}, but was {conf_type!r}")
+
+    if not isinstance(conf_level, numbers.Real) or not np.isfinite(conf_level) or conf_level <= 0 or conf_level >= 1.0:
+        raise ValueError(f"conf_level must be a float in the range (0.0, 1.0), but was {conf_level!r}")
+
+    sigma = np.sqrt(np.cumsum(ratio_var))
+    ci = _ci_logmlog(prob_survival, sigma, conf_level)
+    return ci
+
+
+def kaplan_meier_estimator(
+    event,
+    time_exit,
+    time_enter=None,
+    time_min=None,
+    reverse=False,
+    conf_level=0.95,
+    conf_type=None,
+):
+    """Computes the Kaplan-Meier estimate of the survival function.
+
+    See [1]_ for further description.
+
+    Parameters
+    ----------
+    event : array-like, shape = (n_samples,)
+        A boolean array where ``True`` indicates an event and ``False`` indicates
+        right-censoring.
+
+    time_exit : array-like, shape = (n_samples,)
+        Time of event or censoring.
+
+    time_enter : array-like, shape = (n_samples,), optional
+        Contains time when each individual entered the study for
+        left truncated survival data.
+
+    time_min : float, optional
+        Compute estimator conditional on survival at least up to
+        the specified time.
+
+    reverse : bool, optional, default: False
+        Whether to estimate the censoring distribution.
+        When there are ties between times at which events are observed,
+        then events come first and are subtracted from the denominator.
+        Only available for right-censored data, i.e. `time_enter` must
+        be None.
+
+    conf_level : float, optional, default: 0.95
+        The level for a two-sided confidence interval on the survival curves.
+
+    conf_type : None or {'log-log'}, optional, default: None.
+        The type of confidence intervals to estimate.
+        If `None`, no confidence intervals are estimated.
+        If "log-log", estimate confidence intervals using
+        the log hazard or :math:`log(-log(S(t)))` as described in [2]_.
+
+    Returns
+    -------
+    time : ndarray, shape = (n_times,)
+        Unique times.
+
+    prob_survival : ndarray, shape = (n_times,)
+        Survival probability at each unique time point.
+        If `time_enter` is provided, estimates are conditional probabilities.
+
+    conf_int : ndarray, shape = (2, n_times)
+        Pointwise confidence interval of the Kaplan-Meier estimator
+        at each unique time point.
+        Only provided if `conf_type` is not None.
+
+    Examples
+    --------
+    Creating a Kaplan-Meier curve:
+
+    .. plot::
+
+        >>> import matplotlib.pyplot as plt
+        >>> from sksurv.datasets import load_veterans_lung_cancer
+        >>> from sksurv.nonparametric import kaplan_meier_estimator
+        >>>
+        >>> _, y = load_veterans_lung_cancer()
+        >>> time, prob_surv, conf_int = kaplan_meier_estimator(
+        ...     y["Status"], y["Survival_in_days"], conf_type="log-log"
+        ... )
+        >>> plt.step(time, prob_surv, where="post")
+        [...]
+        >>> plt.fill_between(time, conf_int[0], conf_int[1], alpha=0.25, step="post")
+        <matplotlib.collections.FillBetweenPolyCollection object at 0x...>
+        >>> plt.ylim(0, 1)
+        (0.0, 1.0)
+        >>> plt.show()  # doctest: +SKIP
+
+    See also
+    --------
+    sksurv.nonparametric.SurvivalFunctionEstimator
+        Estimator API of the Kaplan-Meier estimator.
+
+    References
+    ----------
+    .. [1] Kaplan, E. L. and Meier, P., "Nonparametric estimation from incomplete observations",
+           Journal of The American Statistical Association, vol. 53, pp. 457-481, 1958.
+    .. [2] Borgan Ø. and Liestøl K., "A Note on Confidence Intervals and Bands for the
+           Survival Function Based on Transformations", Scandinavian Journal of
+           Statistics. 1990;17(1):35–41.
+    """
+    event, time_enter, time_exit = check_y_survival(event, time_enter, time_exit, allow_all_censored=True)
+    check_consistent_length(event, time_enter, time_exit)
+
+    if conf_type is not None and reverse:
+        raise NotImplementedError("Confidence intervals of the censoring distribution is not implemented.")
+
+    if time_enter is None:
+        uniq_times, n_events, n_at_risk, n_censored = _compute_counts(event, time_exit)
+
+        if reverse:
+            n_at_risk -= n_events
+            n_events = n_censored
+    else:
+        if reverse:
+            raise ValueError("The censoring distribution cannot be estimated from left truncated data")
+
+        uniq_times, n_events, n_at_risk = _compute_counts_truncated(event, time_enter, time_exit)
+
+    # account for 0/0 = nan
+    ratio = np.divide(
+        n_events,
+        n_at_risk,
+        out=np.zeros(uniq_times.shape[0], dtype=float),
+        where=n_events != 0,
+    )
+    values = 1.0 - ratio
+
+    if conf_type is not None:
+        ratio_var = np.divide(
+            n_events,
+            n_at_risk * (n_at_risk - n_events),
+            out=np.zeros(uniq_times.shape[0], dtype=float),
+            where=(n_events != 0) & (n_at_risk != n_events),
+        )
+
+    if time_min is not None:
+        mask = uniq_times >= time_min
+        uniq_times = np.compress(mask, uniq_times)
+        values = np.compress(mask, values)
+
+    prob_survival = np.cumprod(values)
+
+    if conf_type is None:
+        return uniq_times, prob_survival
+
+    if time_min is not None:
+        ratio_var = np.compress(mask, ratio_var)
+
+    ci = _km_ci_estimator(prob_survival, ratio_var, conf_level, conf_type)
+
+    return uniq_times, prob_survival, ci
+
+
+def nelson_aalen_estimator(event, time):
+    """Computes the Nelson-Aalen estimate of the cumulative hazard function.
+
+    See [1]_, [2]_ for further description.
+
+    Parameters
+    ----------
+    event : array-like, shape = (n_samples,)
+        A boolean array where ``True`` indicates an event and ``False`` indicates
+        right-censoring.
+
+    time : array-like, shape = (n_samples,)
+        Time of event or censoring.
+
+    Returns
+    -------
+    time : ndarray, shape = (n_times,)
+        Unique times.
+
+    cum_hazard : ndarray, shape = (n_times,)
+        Cumulative hazard at each unique time point.
+
+    Examples
+    --------
+    Creating a cumulative hazard curve:
+
+    .. plot::
+
+        >>> import matplotlib.pyplot as plt
+        >>> from sksurv.datasets import load_aids
+        >>> from sksurv.nonparametric import nelson_aalen_estimator
+        >>>
+        >>> _, y = load_aids(endpoint="death")
+        >>> time, cum_hazard = nelson_aalen_estimator(y["censor_d"], y["time_d"])
+        >>>
+        >>> plt.step(time, cum_hazard, where="post")
+        [...]
+        >>> plt.show()  # doctest: +SKIP
+
+    References
+    ----------
+    .. [1] Nelson, W., "Theory and applications of hazard plotting for censored failure data",
+           Technometrics, vol. 14, pp. 945-965, 1972.
+
+    .. [2] Aalen, O. O., "Nonparametric inference for a family of counting processes",
+           Annals of Statistics, vol. 6, pp. 701–726, 1978.
+    """
+    event, time = check_y_survival(event, time)
+    check_consistent_length(event, time)
+    uniq_times, n_events, n_at_risk, _ = _compute_counts(event, time)
+
+    y = np.cumsum(n_events / n_at_risk)
+
+    return uniq_times, y
+
+
+def ipc_weights(event, time):
+    """Compute inverse probability of censoring weights
+
+    Parameters
+    ----------
+    event : array-like, shape = (n_samples,)
+        A boolean array where ``True`` indicates an event and ``False`` indicates
+        right-censoring.
+
+    time : array-like, shape = (n_samples,)
+        Time when a subject experienced an event or was censored.
+
+    Returns
+    -------
+    weights : ndarray, shape = (n_samples,)
+        inverse probability of censoring weights
+
+    See also
+    --------
+    CensoringDistributionEstimator
+        An estimator interface for estimating inverse probability
+        of censoring weights for unseen time points.
+    """
+    if event.all():
+        return np.ones(time.shape[0])
+
+    unique_time, p = kaplan_meier_estimator(event, time, reverse=True)
+
+    idx = np.searchsorted(unique_time, time[event])
+    Ghat = p[idx]
+
+    assert (Ghat > 0).all()
+
+    weights = np.zeros(time.shape[0])
+    weights[event] = 1.0 / Ghat
+
+    return weights
+
+
+class SurvivalFunctionEstimator(BaseEstimator):
+    """Kaplan–Meier estimate of the survival function.
+
+    Parameters
+    ----------
+    conf_level : float, optional, default: 0.95
+        The level for a two-sided confidence interval on the survival curves.
+
+    conf_type : None or {'log-log'}, optional, default: None.
+        The type of confidence intervals to estimate.
+        If `None`, no confidence intervals are estimated.
+        If "log-log", estimate confidence intervals using
+        the log hazard or :math:`log(-log(S(t)))`.
+
+    See also
+    --------
+    sksurv.nonparametric.kaplan_meier_estimator
+        Functional API of the Kaplan-Meier estimator.
+    """
+
+    _parameter_constraints = {
+        "conf_level": [Interval(numbers.Real, 0.0, 1.0, closed="neither")],
+        "conf_type": [None, StrOptions({"log-log"})],
+    }
+
+    def __init__(self, conf_level=0.95, conf_type=None):
+        self.conf_level = conf_level
+        self.conf_type = conf_type
+
+    def fit(self, y):
+        """Estimate survival distribution from training data.
+
+        Parameters
+        ----------
+        y : structured array, shape = (n_samples,)
+            A structured array with two fields. The first field is a boolean
+            where ``True`` indicates an event and ``False`` indicates right-censoring.
+            The second field is a float with the time of event or time of censoring.
+
+        Returns
+        -------
+        self
+        """
+        self._validate_params()
+        event, time = check_y_survival(y, allow_all_censored=True)
+
+        values = kaplan_meier_estimator(event, time, conf_level=self.conf_level, conf_type=self.conf_type)
+        if self.conf_type is None:
+            unique_time, prob = values
+        else:
+            unique_time, prob, conf_int = values
+            self.conf_int_ = np.column_stack((np.ones((2, 1)), conf_int))
+
+        self.unique_time_ = np.r_[-np.inf, unique_time]
+        self.prob_ = np.r_[1.0, prob]
+
+        return self
+
+    def predict_proba(self, time, return_conf_int=False):
+        r"""Return probability of remaining event-free at given time points.
+
+        :math:`\hat{S}(t) = P(T > t)`
+
+        Parameters
+        ----------
+        time : array-like, shape = (n_samples,)
+            Time to estimate probability at.
+
+        return_conf_int : bool, optional, default: False
+            Whether to return the pointwise confidence interval
+            of the survival function.
+            Only available if :meth:`fit()` has been called
+            with the `conf_type` parameter set.
+
+        Returns
+        -------
+        prob : ndarray, shape = (n_samples,)
+            Probability of remaining event-free at the given time points.
+
+        conf_int : ndarray, shape = (2, n_samples)
+            Pointwise confidence interval at the passed time points.
+            Only provided if `return_conf_int` is True.
+        """
+        check_is_fitted(self, "unique_time_")
+        if return_conf_int and not hasattr(self, "conf_int_"):
+            raise ValueError(
+                "If return_conf_int is True, SurvivalFunctionEstimator must be fitted with conf_int != None"
+            )
+
+        time = check_array(time, ensure_2d=False, estimator=self, input_name="time")
+
+        # K-M is undefined if estimate at last time point is non-zero
+        extends = time > self.unique_time_[-1]
+        if self.prob_[-1] > 0 and extends.any():
+            raise ValueError(f"time must be smaller than largest observed time point: {self.unique_time_[-1]}")
+
+        # beyond last time point is zero probability
+        Shat = np.empty(time.shape, dtype=float)
+        Shat[extends] = 0.0
+
+        valid = ~extends
+        time = time[valid]
+        idx = np.searchsorted(self.unique_time_, time)
+        # for non-exact matches, we need to shift the index to left
+        eps = np.finfo(self.unique_time_.dtype).eps
+        exact = np.absolute(self.unique_time_[idx] - time) < eps
+        idx[~exact] -= 1
+        Shat[valid] = self.prob_[idx]
+
+        if not return_conf_int:
+            return Shat
+
+        ci = np.empty((2, time.shape[0]), dtype=float)
+        ci[:, extends] = np.nan
+        ci[:, valid] = self.conf_int_[:, idx]
+        return Shat, ci
+
+
+class CensoringDistributionEstimator(SurvivalFunctionEstimator):
+    """Kaplan–Meier estimator for the censoring distribution."""
+
+    def fit(self, y):
+        """Estimate censoring distribution from training data.
+
+        Parameters
+        ----------
+        y : structured array, shape = (n_samples,)
+            A structured array with two fields. The first field is a boolean
+            where ``True`` indicates an event and ``False`` indicates right-censoring.
+            The second field is a float with the time of event or time of censoring.
+
+        Returns
+        -------
+        self
+        """
+        event, time = check_y_survival(y)
+        if event.all():
+            self.unique_time_ = np.unique(time)
+            self.prob_ = np.ones(self.unique_time_.shape[0])
+        else:
+            unique_time, prob = kaplan_meier_estimator(event, time, reverse=True)
+            self.unique_time_ = np.r_[-np.inf, unique_time]
+            self.prob_ = np.r_[1.0, prob]
+
+        return self
+
+    def predict_ipcw(self, y):
+        r"""Return inverse probability of censoring weights at given time points.
+
+        :math:`\omega_i = \delta_i / \hat{G}(y_i)`
+
+        Parameters
+        ----------
+        y : structured array, shape = (n_samples,)
+            A structured array with two fields. The first field is a boolean
+            where ``True`` indicates an event and ``False`` indicates right-censoring.
+            The second field is a float with the time of event or time of censoring.
+
+        Returns
+        -------
+        ipcw : ndarray, shape = (n_samples,)
+            Inverse probability of censoring weights.
+        """
+        event, time = check_y_survival(y)
+        Ghat = self.predict_proba(time[event])
+
+        if (Ghat == 0.0).any():
+            raise ValueError("censoring survival function is zero at one or more time points")
+
+        weights = np.zeros(time.shape[0])
+        weights[event] = 1.0 / Ghat
+
+        return weights
+
+
+def _cum_inc_cr_ci_estimator(cum_inc, var, conf_level, conf_type):
+    if conf_type not in {"log-log"}:
+        raise ValueError(f"conf_type must be None or a str among {{'log-log'}}, but was {conf_type!r}")
+
+    if not isinstance(conf_level, numbers.Real) or not np.isfinite(conf_level) or conf_level <= 0 or conf_level >= 1.0:
+        raise ValueError(f"conf_level must be a float in the range (0.0, 1.0), but was {conf_level!r}")
+    eps = np.finfo(var.dtype).eps
+    sigma = np.zeros_like(var)
+    np.divide(np.sqrt(var), cum_inc, where=var > eps, out=sigma)
+    ci = _ci_logmlog(cum_inc, sigma, conf_level)
+    # make first axis the competing risks, the second axis the lower and upper confidence interval
+    ci = np.swapaxes(ci, 0, 1)
+    return ci
+
+
+def cumulative_incidence_competing_risks(
+    event,
+    time_exit,
+    time_min=None,
+    conf_level=0.95,
+    conf_type=None,
+    var_type="Aalen",
+):
+    """Non-parametric estimator of Cumulative Incidence function in the case of competing risks.
+
+    See the :ref:`User Guide </user_guide/competing-risks.ipynb>` and [1]_ for further details.
+
+    Parameters
+    ----------
+    event : array-like, shape = (n_samples,), dtype = int
+        Contains event indicators. A value of 0 indicates right-censoring,
+        while a positive integer from 1 to `n_risks` corresponds to a specific risk.
+        `n_risks` is the total number of different risks.
+        It assumes there are events for all possible risks.
+
+    time_exit : array-like, shape = (n_samples,)
+        Contains event or censoring times.
+
+    time_min : float, optional, default: None
+        Compute estimator conditional on survival at least up to
+        the specified time.
+
+    conf_level : float, optional, default: 0.95
+        The level for a two-sided confidence interval on the cumulative incidence curves.
+
+    conf_type : None or {'log-log'}, optional, default: None.
+        The type of confidence intervals to estimate.
+        If `None`, no confidence intervals are estimated.
+        If "log-log", estimate confidence intervals using
+        the log hazard or :math:`log(-log(S(t)))`.
+
+    var_type : {'Aalen', 'Dinse', 'Dinse_Approx'}, optional, default: 'Aalen'
+        The method for estimating the variance of the estimator.
+        See [2]_, [3]_ and [4]_ for each of the methods.
+        Only used if `conf_type` is not None.
+
+    Returns
+    -------
+    time : ndarray, shape = (n_times,)
+        Unique times.
+
+    cum_incidence : ndarray, shape = (n_risks + 1, n_times)
+        Cumulative incidence for each risk. The first row (``cum_incidence[0]``)
+        is the cumulative incidence of any risk (total risk). The remaining
+        rows (``cum_incidence[1:]``) are the cumulative incidences for each
+        competing risk.
+
+    conf_int : ndarray, shape = (n_risks + 1, 2, n_times)
+        Pointwise confidence interval (second axis) of the cumulative incidence function
+        at each unique time point (last axis)
+        for all possible risks (first axis), including overall risk (``conf_int[0]``).
+        Only provided if `conf_type` is not None.
+
+    Examples
+    --------
+    Creating cumulative incidence curves:
+
+    .. plot::
+
+        >>> import matplotlib.pyplot as plt
+        >>> from sksurv.datasets import load_bmt
+        >>> from sksurv.nonparametric import cumulative_incidence_competing_risks
+        >>>
+        >>> dis, bmt_df = load_bmt()
+        >>> event = bmt_df["status"]
+        >>> time = bmt_df["ftime"]
+        >>> n_risks = event.max()
+        >>>
+        >>> x, y, conf_int = cumulative_incidence_competing_risks(
+        ...     event, time, conf_type="log-log"
+        ... )
+        >>>
+        >>> plt.step(x, y[0], where="post", label="Total risk")
+        [...]
+        >>> plt.fill_between(x, conf_int[0, 0], conf_int[0, 1], alpha=0.25, step="post")
+        <matplotlib.collections.FillBetweenPolyCollection object at 0x...>
+        >>> for i in range(1, n_risks + 1):
+        ...     plt.step(x, y[i], where="post", label=f"{i}-risk")
+        ...     plt.fill_between(x, conf_int[i, 0], conf_int[i, 1], alpha=0.25, step="post")
+        [...]
+        <matplotlib.collections.FillBetweenPolyCollection object at 0x...>
+        >>> plt.ylim(0, 1)
+        (0.0, 1.0)
+        >>> plt.legend()
+        <matplotlib.legend.Legend object at 0x...>
+        >>> plt.show()  # doctest: +SKIP
+
+    References
+    ----------
+    .. [1] Kalbfleisch, J.D. and Prentice, R.L. (2002)
+           The Statistical Analysis of Failure Time Data. 2nd Edition, John Wiley and Sons, New York.
+    .. [2] Aalen, O. (1978a). Annals of Statistics, 6, 534–545.
+           We implement the formula in M. Pintilie: "Competing Risks: A Practical Perspective".
+           John Wiley & Sons, 2006, Eq. 4.5
+    .. [3] Dinse and Larson, Biometrika (1986), 379. Sect. 4, Eqs. 4 and 5.
+    .. [4] Dinse and Larson, Biometrika (1986), 379. Sect. 4, Eq. 6.
+    """
+    event, time_exit = check_y_survival(event, time_exit, allow_all_censored=True, competing_risks=True)
+    check_consistent_length(event, time_exit)
+
+    n_risks = event.max()
+    uniq_times, n_events_cr, n_at_risk, _n_censored = _compute_counts(event, time_exit)
+
+    # account for 0/0 = nan
+    n_t = uniq_times.shape[0]
+    ratio = np.divide(
+        n_events_cr,
+        n_at_risk[..., np.newaxis],
+        out=np.zeros((n_t, n_risks + 1), dtype=float),
+        where=n_events_cr != 0,
+    )
+
+    if time_min is not None:
+        mask = uniq_times >= time_min
+        uniq_times = np.compress(mask, uniq_times)
+        ratio = np.compress(mask, ratio, axis=0)
+
+    kpe = np.cumprod(1.0 - ratio[:, 0])
+    kpe_prime = np.r_[1.0, kpe[:-1]]
+    cum_inc = np.empty((n_risks + 1, n_t), dtype=float)
+    cum_inc[0] = 1.0 - kpe
+    cum_inc[1:] = np.cumsum((ratio[:, 1:].T * kpe_prime), axis=1)
+
+    if conf_type is None:
+        return uniq_times, cum_inc
+
+    if var_type == "Aalen":
+        var = _var_aalen(n_events_cr, kpe_prime, n_at_risk, cum_inc)
+    elif var_type == "Dinse_Approx":
+        var = _var_dinse_approx(n_events_cr, kpe_prime, n_at_risk, cum_inc)
+    elif var_type == "Dinse":
+        var = _var_dinse(n_events_cr, kpe_prime, n_at_risk)
+    else:
+        raise ValueError(f"{var_type=} must be one of 'Aalen', 'Dinse', or 'Dinse_Approx'.")
+
+    _x, _y, conf_int_km = kaplan_meier_estimator(event > 0, time_exit, conf_type="log-log")
+    ci = np.empty(shape=(n_risks + 1, 2, n_t), dtype=conf_int_km.dtype)
+    ci[0, :, :] = 1 - conf_int_km
+    ci[1:, :, :] = _cum_inc_cr_ci_estimator(cum_inc[1:], var, conf_level, conf_type)
+
+    return uniq_times, cum_inc, ci
+
+
+def _var_dinse_approx(n_events_cr, kpe_prime, n_at_risk, cum_inc):
+    """
+    Variance estimator from Dinse and Larson, Biometrika (1986), 379
+    See Section 4, Eqs. 6.
+    This is an approximation from the _var_dinse, so that one should be preferred.
+    However, this seems to be more common in the literature.
+    """
+    dr = n_events_cr[:, 0]
+    dr_cr = n_events_cr[:, 1:].T
+    irt = cum_inc[1:, :, np.newaxis] - cum_inc[1:, np.newaxis, :]
+    mask = np.tril(np.ones_like(irt[0]))
+
+    # var_a = np.sum(irt**2 * mask * (dr / (n_at_risk * (n_at_risk - dr))), axis=2)
+    var_a = np.einsum("rjk,jk,k->rj", irt**2, mask, dr / (n_at_risk * (n_at_risk - dr)))
+    var_b = np.cumsum(((n_at_risk - dr_cr) / n_at_risk) * (dr_cr / n_at_risk**2) * kpe_prime**2, axis=1)
+    # var_c = -2 * np.sum(irt * mask * dr_cr[:, np.newaxis, :] * (kpe_prime / n_at_risk**2), axis=2)
+    var_c = -2 * np.einsum("rjk,jk,rk,k->rj", irt, mask, dr_cr, kpe_prime / n_at_risk**2)
+
+    var = var_a + var_b + var_c
+    return var
+
+
+def _var_dinse(n_events_cr, kpe_prime, n_at_risk):
+    """
+    Variance estimator from Dinse and Larson, Biometrika (1986), 379
+    See Section 4, Eqs. 4 and 5
+    """
+    dr = n_events_cr[:, 0]
+    dr_cr = n_events_cr[:, 1:].T
+    theta = dr_cr * kpe_prime / n_at_risk
+    x = dr / (n_at_risk * (n_at_risk - dr))
+    cprod = np.cumprod(1 + x) / (1 + x)
+
+    nt_range = np.arange(dr.size)
+    i_idx = nt_range[:, None, None]
+    j_idx = nt_range[None, :, None]
+    k_idx = nt_range[None, None, :]
+    mask = ((j_idx < i_idx) & (k_idx > j_idx) & (k_idx <= i_idx)).astype(int)
+
+    _v1 = np.zeros_like(theta)
+    np.divide((n_at_risk - dr_cr), n_at_risk * dr_cr, out=_v1, where=dr_cr > 0)
+    v1 = np.cumsum(theta**2 * ((1 + _v1) * cprod - 1), axis=1)
+
+    corr = (1 - 1 / n_at_risk) * cprod - 1
+    v2 = 2 * np.einsum("rj,rk,ijk->ri", theta * corr, theta, mask)
+    var = v1 + v2
+
+    return var
+
+
+def _var_aalen(n_events_cr, kpe_prime, n_at_risk, cum_inc):
+    """
+    Variance estimator from Aalen
+    Aalen, O. (1978a). Nonparametric estimation of partial transition
+    probabilities in multiple decrement models. Annals of Statistics, 6, 534–545.
+    We implement it as shown in
+    M. Pintilie: "Competing Risks: A Practical Perspective". John Wiley & Sons, 2006, Eq. 4.5
+    This seems to be the estimator used in cmprsk, but there are some numerical differences with our implementation.
+    """
+    dr = n_events_cr[:, 0]
+    dr_cr = n_events_cr[:, 1:].T
+    irt = cum_inc[1:, :, np.newaxis] - cum_inc[1:, np.newaxis, :]
+    mask = np.tril(np.ones_like(irt[0]))
+
+    _va = np.zeros_like(kpe_prime)
+    den_a = (n_at_risk - 1) * (n_at_risk - dr)
+    np.divide(dr, den_a, out=_va, where=den_a > 0)
+    # var_a = np.sum(irt**2 * mask * _va, axis=2)
+    var_a = np.einsum("rjk,jk,k->rj", irt**2, mask, _va)
+
+    _vb = np.zeros_like(kpe_prime)
+    den_b = (n_at_risk - 1) * n_at_risk**2
+    np.divide(1.0, den_b, out=_vb, where=den_b > 0)
+    var_b = np.cumsum((n_at_risk - dr_cr) * dr_cr * _vb * kpe_prime**2, axis=1)
+
+    _vca = dr_cr * (n_at_risk - dr_cr)
+    _vcb = np.zeros_like(kpe_prime)
+    den_c = n_at_risk * (n_at_risk - dr) * (n_at_risk - 1)
+    np.divide(kpe_prime, den_c, out=_vcb, where=den_c > 0)
+    # var_c = -2 * np.sum(irt * mask * _vca[:, np.newaxis, :] * _vcb, axis=2)
+    var_c = -2 * np.einsum("rjk,jk,rk,k->rj", irt, mask, _vca, _vcb)
+
+    var = var_a + var_b + var_c
+    return var