bartz 0.5.0__py3-none-any.whl → 0.7.0__py3-none-any.whl

bartz/BART.py

@@ -22,17 +22,73 @@
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 # SOFTWARE.
 
-import functools
+"""Implement a class `gbart` that mimics the R BART package."""
+
+import math
+from collections.abc import Sequence
+from functools import cached_property
+from typing import Any, Literal, Protocol
 
 import jax
 import jax.numpy as jnp
+from equinox import Module, field
+from jax.scipy.special import ndtr
+from jaxtyping import (
+    Array,
+    Bool,
+    Float,
+    Float32,
+    Int32,
+    Integer,
+    Key,
+    Real,
+    Shaped,
+    UInt,
+)
+from numpy import ndarray
+
+from bartz import mcmcloop, mcmcstep, prepcovars
+from bartz.jaxext.scipy.special import ndtri
+from bartz.jaxext.scipy.stats import invgamma
+
+FloatLike = float | Float[Any, '']
+
+
+class DataFrame(Protocol):
+    """DataFrame duck-type for `gbart`.
+
+    Attributes
+    ----------
+    columns : Sequence[str]
+        The names of the columns.
+    """
+
+    columns: Sequence[str]
+
+    def to_numpy(self) -> ndarray:
+        """Convert the dataframe to a 2d numpy array with columns on the second axis."""
+        ...
 
-from . import grove, jaxext, mcmcloop, mcmcstep, prepcovars
 
+class Series(Protocol):
+    """Series duck-type for `gbart`.
 
-class gbart:
+    Attributes
+    ----------
+    name : str | None
+        The name of the series.
     """
-    Nonparametric regression with Bayesian Additive Regression Trees (BART).
+
+    name: str | None
+
+    def to_numpy(self) -> ndarray:
+        """Convert the series to a 1d numpy array."""
+        ...
+
+
+class gbart(Module):
+    R"""
+    Nonparametric regression with Bayesian Additive Regression Trees (BART) [2]_.
 
     Regress `y_train` on `x_train` with a latent mean function represented as
     a sum of decision trees. The inference is carried out by sampling the
@@ -40,55 +96,108 @@ class gbart:
 
     Parameters
     ----------
-    x_train : array (p, n) or DataFrame
+    x_train
         The training predictors.
-    y_train : array (n,) or Series
+    y_train
         The training responses.
-    x_test : array (p, m) or DataFrame, optional
+    x_test
         The test predictors.
-    usequants : bool, default False
+    type
+        The type of regression. 'wbart' for continuous regression, 'pbart' for
+        binary regression with probit link.
+    sparse
+        Whether to activate variable selection on the predictors as done in
+        [1]_.
+    theta
+    a
+    b
+    rho
+        Hyperparameters of the sparsity prior used for variable selection.
+
+        The prior distribution on the choice of predictor for each decision rule
+        is
+
+        .. math::
+            (s_1, \ldots, s_p) \sim
+            \operatorname{Dirichlet}(\mathtt{theta}/p, \ldots, \mathtt{theta}/p).
+
+        If `theta` is not specified, it's a priori distributed according to
+
+        .. math::
+            \frac{\mathtt{theta}}{\mathtt{theta} + \mathtt{rho}} \sim
+            \operatorname{Beta}(\mathtt{a}, \mathtt{b}).
+
+        If not specified, `rho` is set to the number of predictors p. To tune
+        the prior, consider setting a lower `rho` to prefer more sparsity.
+        If setting `theta` directly, it should be in the ballpark of p or lower
+        as well.
+    xinfo
+        A matrix with the cutpoins to use to bin each predictor. If not
+        specified, it is generated automatically according to `usequants` and
+        `numcut`.
+
+        Each row shall contain a sorted list of cutpoints for a predictor. If
+        there are less cutpoints than the number of columns in the matrix,
+        fill the remaining cells with NaN.
+
+        `xinfo` shall be a matrix even if `x_train` is a dataframe.
+    usequants
         Whether to use predictors quantiles instead of a uniform grid to bin
-        predictors.
-    sigest : float, optional
+        predictors. Ignored if `xinfo` is specified.
+    rm_const
+        How to treat predictors with no associated decision rules (i.e., there
+        are no available cutpoints for that predictor). If `True` (default),
+        they are ignored. If `False`, an error is raised if there are any. If
+        `None`, no check is performed, and the output of the MCMC may not make
+        sense if there are predictors without cutpoints. The option `None` is
+        provided only to allow jax tracing.
+    sigest
         An estimate of the residual standard deviation on `y_train`, used to set
         `lamda`. If not specified, it is estimated by linear regression (with
         intercept, and without taking into account `w`). If `y_train` has less
         than two elements, it is set to 1. If n <= p, it is set to the standard
         deviation of `y_train`. Ignored if `lamda` is specified.
-    sigdf : int, default 3
+    sigdf
         The degrees of freedom of the scaled inverse-chisquared prior on the
         noise variance.
-    sigquant : float, default 0.9
+    sigquant
         The quantile of the prior on the noise variance that shall match
         `sigest` to set the scale of the prior. Ignored if `lamda` is specified.
-    k : float, default 2
+    k
         The inverse scale of the prior standard deviation on the latent mean
         function, relative to half the observed range of `y_train`. If `y_train`
         has less than two elements, `k` is ignored and the scale is set to 1.
-    power : float, default 2
-    base : float, default 0.95
+    power
+    base
         Parameters of the prior on tree node generation. The probability that a
         node at depth `d` (0-based) is non-terminal is ``base / (1 + d) **
         power``.
-    maxdepth : int, default 6
-        The maximum depth of the trees. This is 1-based, so with the default
-        ``maxdepth=6``, the depths of the levels range from 0 to 5.
-    lamda : float, optional
-        The scale of the prior on the noise variance. If ``lamda==1``, the
-        prior is an inverse chi-squared scaled to have harmonic mean 1. If
-        not specified, it is set based on `sigest` and `sigquant`.
-    offset : float, optional
+    lamda
+        The prior harmonic mean of the error variance. (The harmonic mean of x
+        is 1/mean(1/x).) If not specified, it is set based on `sigest` and
+        `sigquant`.
+    tau_num
+        The numerator in the expression that determines the prior standard
+        deviation of leaves. If not specified, default to ``(max(y_train) -
+        min(y_train)) / 2`` (or 1 if `y_train` has less than two elements) for
+        continuous regression, and 3 for binary regression.
+    offset
         The prior mean of the latent mean function. If not specified, it is set
-        to the mean of `y_train`. If `y_train` is empty, it is set to 0.
-    w : array (n,), optional
+        to the mean of `y_train` for continuous regression, and to
+        ``Phi^-1(mean(y_train))`` for binary regression. If `y_train` is empty,
+        `offset` is set to 0. With binary regression, if `y_train` is all
+        `False` or `True`, it is set to ``Phi^-1(1/(n+1))`` or
+        ``Phi^-1(n/(n+1))``, respectively.
+    w
         Coefficients that rescale the error standard deviation on each
         datapoint. Not specifying `w` is equivalent to setting it to 1 for all
         datapoints. Note: `w` is ignored in the automatic determination of
         `sigest`, so either the weights should be O(1), or `sigest` should be
         specified by the user.
-    ntree : int, default 200
-        The number of trees used to represent the latent mean function.
-    numcut : int, default 255
+    ntree
+        The number of trees used to represent the latent mean function. By
+        default 200 for continuous regression and 50 for binary regression.
+    numcut
         If `usequants` is `False`: the exact number of cutpoints used to bin the
         predictors, ranging between the minimum and maximum observed values
         (excluded).
@@ -101,50 +210,43 @@ class gbart:
 
         Before running the algorithm, the predictors are compressed to the
         smallest integer type that fits the bin indices, so `numcut` is best set
-        to the maximum value of an unsigned integer type.
-    ndpost : int, default 1000
+        to the maximum value of an unsigned integer type, like 255.
+
+        Ignored if `xinfo` is specified.
+    ndpost
         The number of MCMC samples to save, after burn-in.
-    nskip : int, default 100
+    nskip
         The number of initial MCMC samples to discard as burn-in.
-    keepevery : int, default 1
-        The thinning factor for the MCMC samples, after burn-in.
-    printevery : int, default 100
-        The number of iterations (including skipped ones) between each log.
-    seed : int or jax random key, default 0
+    keepevery
+        The thinning factor for the MCMC samples, after burn-in. By default, 1
+        for continuous regression and 10 for binary regression.
+    printevery
+        The number of iterations (including thinned-away ones) between each log
+        line. Set to `None` to disable logging.
+
+        `printevery` has a few unexpected side effects. On cpu, interrupting
+        with ^C halts the MCMC only on the next log. And the total number of
+        iterations is a multiple of `printevery`, so if ``nskip + keepevery *
+        ndpost`` is not a multiple of `printevery`, some of the last iterations
+        will not be saved.
+    seed
         The seed for the random number generator.
-    initkw : dict
-        Additional arguments passed to `mcmcstep.init`.
+    maxdepth
+        The maximum depth of the trees. This is 1-based, so with the default
+        ``maxdepth=6``, the depths of the levels range from 0 to 5.
+    init_kw
+        Additional arguments passed to `bartz.mcmcstep.init`.
+    run_mcmc_kw
+        Additional arguments passed to `bartz.mcmcloop.run_mcmc`.
 
     Attributes
     ----------
-    yhat_train : array (ndpost, n)
-        The conditional posterior mean at `x_train` for each MCMC iteration.
-    yhat_train_mean : array (n,)
-        The marginal posterior mean at `x_train`.
-    yhat_test : array (ndpost, m)
-        The conditional posterior mean at `x_test` for each MCMC iteration.
-    yhat_test_mean : array (m,)
-        The marginal posterior mean at `x_test`.
-    sigma : array (ndpost,)
-        The standard deviation of the error.
-    first_sigma : array (nskip,)
-        The standard deviation of the error in the burn-in phase.
-    offset : float
+    offset : Float32[Array, '']
         The prior mean of the latent mean function.
-    scale : float
-        The prior standard deviation of the latent mean function.
-    lamda : float
-        The prior harmonic mean of the error variance.
-    sigest : float or None
+    sigest : Float32[Array, ''] | None
         The estimated standard deviation of the error used to set `lamda`.
-    ntree : int
-        The number of trees.
-    maxdepth : int
-        The maximum depth of the trees.
-
-    Methods
-    -------
-    predict
+    yhat_test : Float32[Array, 'ndpost m'] | None
+        The conditional posterior mean at `x_test` for each MCMC iteration.
 
     Notes
     -----
@@ -156,128 +258,293 @@ class gbart:
     - If ``usequants=False``, R BART switches to quantiles anyway if there are
       less predictor values than the required number of bins, while bartz
       always follows the specification.
+    - Some functionality is missing.
     - The error variance parameter is called `lamda` instead of `lambda`.
-    - `rm_const` is always `False`.
-    - The default `numcut` is 255 instead of 100.
-    - A lot of functionality is missing (variable selection, discrete response).
     - There are some additional attributes, and some missing.
+    - The trees have a maximum depth.
+    - `rm_const` refers to predictors without decision rules instead of
+      predictors that are constant in `x_train`.
+    - If `rm_const=True` and some variables are dropped, the predictors
+      matrix/dataframe passed to `predict` should still include them.
 
+    References
+    ----------
+    .. [1] Linero, Antonio R. (2018). “Bayesian Regression Trees for
+       High-Dimensional Prediction and Variable Selection”. In: Journal of the
+       American Statistical Association 113.522, pp. 626-636.
+    .. [2] Hugh A. Chipman, Edward I. George, Robert E. McCulloch "BART:
+       Bayesian additive regression trees," The Annals of Applied Statistics,
+       Ann. Appl. Stat. 4(1), 266-298, (March 2010).
     """
 
+    _main_trace: mcmcloop.MainTrace
+    _burnin_trace: mcmcloop.BurninTrace
+    _mcmc_state: mcmcstep.State
+    _splits: Real[Array, 'p max_num_splits']
+    _x_train_fmt: Any = field(static=True)
+
+    ndpost: int = field(static=True)
+    offset: Float32[Array, '']
+    sigest: Float32[Array, ''] | None = None
+    yhat_test: Float32[Array, 'ndpost m'] | None = None
+
     def __init__(
         self,
-        x_train,
-        y_train,
+        x_train: Real[Array, 'p n'] | DataFrame,
+        y_train: Bool[Array, ' n'] | Float32[Array, ' n'] | Series,
         *,
-        x_test=None,
-        usequants=False,
-        sigest=None,
-        sigdf=3,
-        sigquant=0.9,
-        k=2,
-        power=2,
-        base=0.95,
-        maxdepth=6,
-        lamda=None,
-        offset=None,
-        w=None,
-        ntree=200,
-        numcut=255,
-        ndpost=1000,
-        nskip=100,
-        keepevery=1,
-        printevery=100,
-        seed=0,
-        initkw=None,
+        x_test: Real[Array, 'p m'] | DataFrame | None = None,
+        type: Literal['wbart', 'pbart'] = 'wbart',  # noqa: A002
+        sparse: bool = False,
+        theta: FloatLike | None = None,
+        a: FloatLike = 0.5,
+        b: FloatLike = 1.0,
+        rho: FloatLike | None = None,
+        xinfo: Float[Array, 'p n'] | None = None,
+        usequants: bool = False,
+        rm_const: bool | None = True,
+        sigest: FloatLike | None = None,
+        sigdf: FloatLike = 3.0,
+        sigquant: FloatLike = 0.9,
+        k: FloatLike = 2.0,
+        power: FloatLike = 2.0,
+        base: FloatLike = 0.95,
+        lamda: FloatLike | None = None,
+        tau_num: FloatLike | None = None,
+        offset: FloatLike | None = None,
+        w: Float[Array, ' n'] | None = None,
+        ntree: int | None = None,
+        numcut: int = 100,
+        ndpost: int = 1000,
+        nskip: int = 100,
+        keepevery: int | None = None,
+        printevery: int | None = 100,
+        seed: int | Key[Array, ''] = 0,
+        maxdepth: int = 6,
+        init_kw: dict | None = None,
+        run_mcmc_kw: dict | None = None,
     ):
+        # check data and put it in the right format
         x_train, x_train_fmt = self._process_predictor_input(x_train)
-        y_train, _ = self._process_response_input(y_train)
+        y_train = self._process_response_input(y_train)
         self._check_same_length(x_train, y_train)
         if w is not None:
-            w, _ = self._process_response_input(w)
+            w = self._process_response_input(w)
             self._check_same_length(x_train, w)
 
+        # check data types are correct for continuous/binary regression
+        self._check_type_settings(y_train, type, w)
+        # from here onwards, the type is determined by y_train.dtype == bool
+
+        # set defaults that depend on type of regression
+        if ntree is None:
+            ntree = 50 if y_train.dtype == bool else 200
+        if keepevery is None:
+            keepevery = 10 if y_train.dtype == bool else 1
+
+        # process sparsity settings
+        theta, a, b, rho = self._process_sparsity_settings(
+            x_train, sparse, theta, a, b, rho
+        )
+
+        # process "standardization" settings
         offset = self._process_offset_settings(y_train, offset)
-        scale = self._process_scale_settings(y_train, k)
-        lamda, sigest = self._process_noise_variance_settings(
-            x_train, y_train, sigest, sigdf, sigquant, lamda, offset
+        sigma_mu = self._process_leaf_sdev_settings(y_train, k, ntree, tau_num)
+        lamda, sigest = self._process_error_variance_settings(
+            x_train, y_train, sigest, sigdf, sigquant, lamda
         )
 
-        splits, max_split = self._determine_splits(x_train, usequants, numcut)
+        # determine splits
+        splits, max_split = self._determine_splits(x_train, usequants, numcut, xinfo)
         x_train = self._bin_predictors(x_train, splits)
-        y_train, lamda_scaled = self._transform_input(y_train, lamda, offset, scale)
 
-        mcmc_state = self._setup_mcmc(
+        # setup and run mcmc
+        initial_state = self._setup_mcmc(
             x_train,
             y_train,
+            offset,
             w,
             max_split,
-            lamda_scaled,
+            lamda,
+            sigma_mu,
             sigdf,
             power,
             base,
             maxdepth,
             ntree,
-            initkw,
+            init_kw,
+            rm_const,
+            theta,
+            a,
+            b,
+            rho,
         )
         final_state, burnin_trace, main_trace = self._run_mcmc(
-            mcmc_state, ndpost, nskip, keepevery, printevery, seed
+            initial_state,
+            ndpost,
+            nskip,
+            keepevery,
+            printevery,
+            seed,
+            run_mcmc_kw,
+            sparse,
         )
 
-        sigma = self._extract_sigma(main_trace, scale)
-        first_sigma = self._extract_sigma(burnin_trace, scale)
-
-        self.offset = offset
-        self.scale = scale
-        self.lamda = lamda
+        # set public attributes
+        self.offset = final_state.offset  # from the state because of buffer donation
+        self.ndpost = ndpost
         self.sigest = sigest
-        self.ntree = ntree
-        self.maxdepth = maxdepth
-        self.sigma = sigma
-        self.first_sigma = first_sigma
 
-        self._x_train_fmt = x_train_fmt
-        self._splits = splits
+        # set private attributes
         self._main_trace = main_trace
+        self._burnin_trace = burnin_trace
         self._mcmc_state = final_state
+        self._splits = splits
+        self._x_train_fmt = x_train_fmt
 
+        # predict at test points
         if x_test is not None:
-            yhat_test = self.predict(x_test)
-            self.yhat_test = yhat_test
-            self.yhat_test_mean = yhat_test.mean(axis=0)
+            self.yhat_test = self.predict(x_test)
+
+    @cached_property
+    def prob_test(self) -> Float32[Array, 'ndpost m'] | None:
+        """The posterior probability of y being True at `x_test` for each MCMC iteration."""
+        if self.yhat_test is None or self._mcmc_state.y.dtype != bool:
+            return None
+        else:
+            return ndtr(self.yhat_test)
+
+    @cached_property
+    def prob_test_mean(self) -> Float32[Array, ' m'] | None:
+        """The marginal posterior probability of y being True at `x_test`."""
+        if self.prob_test is None:
+            return None
+        else:
+            return self.prob_test.mean(axis=0)
+
+    @cached_property
+    def prob_train(self) -> Float32[Array, 'ndpost n'] | None:
+        """The posterior probability of y being True at `x_train` for each MCMC iteration."""
+        if self._mcmc_state.y.dtype == bool:
+            return ndtr(self.yhat_train)
+        else:
+            return None
+
+    @cached_property
+    def prob_train_mean(self) -> Float32[Array, ' n'] | None:
+        """The marginal posterior probability of y being True at `x_train`."""
+        if self.prob_train is None:
+            return None
+        else:
+            return self.prob_train.mean(axis=0)
+
+    @cached_property
+    def sigma(self) -> Float32[Array, ' nskip+ndpost'] | None:
+        """The standard deviation of the error, including burn-in samples."""
+        if self._burnin_trace.sigma2 is None:
+            return None
+        else:
+            assert self._main_trace.sigma2 is not None
+            return jnp.sqrt(
+                jnp.concatenate([self._burnin_trace.sigma2, self._main_trace.sigma2])
+            )
+
+    @cached_property
+    def sigma_mean(self) -> Float32[Array, ''] | None:
+        """The mean of `sigma`, only over the post-burnin samples."""
+        if self.sigma is None:
+            return None
+        else:
+            return self.sigma[len(self.sigma) - self.ndpost :].mean(axis=0)
+
+    @cached_property
+    def varcount(self) -> Int32[Array, 'ndpost p']:
+        """Histogram of predictor usage for decision rules in the trees."""
+        return mcmcloop.compute_varcount(
+            self._mcmc_state.forest.max_split.size, self._main_trace
+        )
+
+    @cached_property
+    def varcount_mean(self) -> Float32[Array, ' p']:
+        """Average of `varcount` across MCMC iterations."""
+        return self.varcount.mean(axis=0)
+
+    @cached_property
+    def varprob(self) -> Float32[Array, 'ndpost p']:
+        """Posterior samples of the probability of choosing each predictor for a decision rule."""
+        varprob = self._main_trace.varprob
+        if varprob is None:
+            max_split = self._mcmc_state.forest.max_split
+            p = max_split.size
+            peff = jnp.count_nonzero(max_split)
+            varprob = jnp.where(max_split, 1 / peff, 0)
+            varprob = jnp.broadcast_to(varprob, (self.ndpost, p))
+        return varprob
+
+    @cached_property
+    def varprob_mean(self) -> Float32[Array, ' p']:
+        """The marginal posterior probability of each predictor being chosen for a decision rule."""
+        return self.varprob.mean(axis=0)
+
+    @cached_property
+    def yhat_test_mean(self) -> Float32[Array, ' m'] | None:
+        """The marginal posterior mean at `x_test`.
+
+        Not defined with binary regression because it's error-prone, typically
+        the right thing to consider would be `prob_test_mean`.
+        """
+        if self.yhat_test is None or self._mcmc_state.y.dtype == bool:
+            return None
+        else:
+            return self.yhat_test.mean(axis=0)
+
+    @cached_property
+    def yhat_train(self) -> Float32[Array, 'ndpost n']:
+        """The conditional posterior mean at `x_train` for each MCMC iteration."""
+        x_train = self._mcmc_state.X
+        return self._predict(x_train)
 
-    @functools.cached_property
-    def yhat_train(self):
-        x_train = self._mcmc_state['X']
-        yhat_train = self._predict(self._main_trace, x_train)
-        return self._transform_output(yhat_train, self.offset, self.scale)
+    @cached_property
+    def yhat_train_mean(self) -> Float32[Array, ' n'] | None:
+        """The marginal posterior mean at `x_train`.
 
-    @functools.cached_property
-    def yhat_train_mean(self):
-        return self.yhat_train.mean(axis=0)
+        Not defined with binary regression because it's error-prone, typically
+        the right thing to consider would be `prob_train_mean`.
+        """
+        if self._mcmc_state.y.dtype == bool:
+            return None
+        else:
+            return self.yhat_train.mean(axis=0)
 
-    def predict(self, x_test):
+    def predict(
+        self, x_test: Real[Array, 'p m'] | DataFrame
+    ) -> Float32[Array, 'ndpost m']:
         """
         Compute the posterior mean at `x_test` for each MCMC iteration.
 
         Parameters
         ----------
-        x_test : array (p, m) or DataFrame
+        x_test
             The test predictors.
 
         Returns
         -------
-        yhat_test : array (ndpost, m)
-            The conditional posterior mean at `x_test` for each MCMC iteration.
+        The conditional posterior mean at `x_test` for each MCMC iteration.
+
+        Raises
+        ------
+        ValueError
+            If `x_test` has a different format than `x_train`.
         """
         x_test, x_test_fmt = self._process_predictor_input(x_test)
-        self._check_compatible_formats(x_test_fmt, self._x_train_fmt)
+        if x_test_fmt != self._x_train_fmt:
+            msg = f'Input format mismatch: {x_test_fmt=} != x_train_fmt={self._x_train_fmt!r}'
+            raise ValueError(msg)
         x_test = self._bin_predictors(x_test, self._splits)
-        yhat_test = self._predict(self._main_trace, x_test)
-        return self._transform_output(yhat_test, self.offset, self.scale)
+        return self._predict(x_test)
 
     @staticmethod
-    def _process_predictor_input(x):
+    def _process_predictor_input(x) -> tuple[Shaped[Array, 'p n'], Any]:
         if hasattr(x, 'columns'):
             fmt = dict(kind='dataframe', columns=x.columns)
             x = x.to_numpy().T
@@ -288,19 +555,12 @@ class gbart:
         return x, fmt
 
     @staticmethod
-    def _check_compatible_formats(fmt1, fmt2):
-        assert fmt1 == fmt2
-
-    @staticmethod
-    def _process_response_input(y):
+    def _process_response_input(y) -> Shaped[Array, ' n']:
         if hasattr(y, 'to_numpy'):
-            fmt = dict(kind='series', name=y.name)
             y = y.to_numpy()
-        else:
-            fmt = dict(kind='array')
         y = jnp.asarray(y)
         assert y.ndim == 1
-        return y, fmt
+        return y
 
     @staticmethod
     def _check_same_length(x1, x2):
@@ -308,18 +568,29 @@ class gbart:
         assert get_length(x1) == get_length(x2)
 
     @staticmethod
-    def _process_noise_variance_settings(
-        x_train, y_train, sigest, sigdf, sigquant, lamda, offset
-    ):
-        if lamda is not None:
+    def _process_error_variance_settings(
+        x_train, y_train, sigest, sigdf, sigquant, lamda
+    ) -> tuple[Float32[Array, ''] | None, ...]:
+        if y_train.dtype == bool:
+            if sigest is not None:
+                msg = 'Let `sigest=None` for binary regression'
+                raise ValueError(msg)
+            if lamda is not None:
+                msg = 'Let `lamda=None` for binary regression'
+                raise ValueError(msg)
+            return None, None
+        elif lamda is not None:
+            if sigest is not None:
+                msg = 'Let `sigest=None` if `lamda` is specified'
+                raise ValueError(msg)
             return lamda, None
         else:
             if sigest is not None:
-                sigest2 = sigest * sigest
+                sigest2 = jnp.square(sigest)
             elif y_train.size < 2:
                 sigest2 = 1
             elif y_train.size <= x_train.shape[0]:
-                sigest2 = jnp.var(y_train - offset)
+                sigest2 = jnp.var(y_train)
             else:
                 x_centered = x_train.T - x_train.mean(axis=1)
                 y_centered = y_train - y_train.mean()
@@ -329,182 +600,214 @@ class gbart:
                 dof = len(y_train) - rank
                 sigest2 = chisq / dof
             alpha = sigdf / 2
-            invchi2 = jaxext.scipy.stats.invgamma.ppf(sigquant, alpha) / 2
+            invchi2 = invgamma.ppf(sigquant, alpha) / 2
             invchi2rid = invchi2 * sigdf
             return sigest2 / invchi2rid, jnp.sqrt(sigest2)
 
     @staticmethod
-    def _process_offset_settings(y_train, offset):
+    def _check_type_settings(y_train, type, w):  # noqa: A002
+        match type:
+            case 'wbart':
+                if y_train.dtype != jnp.float32:
+                    msg = (
+                        'Continuous regression requires y_train.dtype=float32,'
+                        f' got {y_train.dtype=} instead.'
+                    )
+                    raise TypeError(msg)
+            case 'pbart':
+                if w is not None:
+                    msg = 'Binary regression does not support weights, set `w=None`'
+                    raise ValueError(msg)
+                if y_train.dtype != bool:
+                    msg = (
+                        'Binary regression requires y_train.dtype=bool,'
+                        f' got {y_train.dtype=} instead.'
+                    )
+                    raise TypeError(msg)
+            case _:
+                msg = f'Invalid {type=}'
+                raise ValueError(msg)
+
+    @staticmethod
+    def _process_sparsity_settings(
+        x_train: Real[Array, 'p n'],
+        sparse: bool,
+        theta: FloatLike | None,
+        a: FloatLike,
+        b: FloatLike,
+        rho: FloatLike | None,
+    ) -> (
+        tuple[None, None, None, None]
+        | tuple[FloatLike, None, None, None]
+        | tuple[None, FloatLike, FloatLike, FloatLike]
+    ):
+        if not sparse:
+            return None, None, None, None
+        elif theta is not None:
+            return theta, None, None, None
+        else:
+            if rho is None:
+                p, _ = x_train.shape
+                rho = float(p)
+            return None, a, b, rho
+
+    @staticmethod
+    def _process_offset_settings(
+        y_train: Float32[Array, ' n'] | Bool[Array, ' n'],
+        offset: float | Float32[Any, ''] | None,
+    ) -> Float32[Array, '']:
         if offset is not None:
-            return offset
+            return jnp.asarray(offset)
         elif y_train.size < 1:
-            return 0
+            return jnp.array(0.0)
         else:
-            return y_train.mean()
+            mean = y_train.mean()
 
-    @staticmethod
-    def _process_scale_settings(y_train, k):
-        if y_train.size < 2:
-            return 1
+        if y_train.dtype == bool:
+            bound = 1 / (1 + y_train.size)
+            mean = jnp.clip(mean, bound, 1 - bound)
+            return ndtri(mean)
         else:
-            return (y_train.max() - y_train.min()) / (2 * k)
+            return mean
 
     @staticmethod
-    def _determine_splits(x_train, usequants, numcut):
-        if usequants:
+    def _process_leaf_sdev_settings(
+        y_train: Float32[Array, ' n'] | Bool[Array, ' n'],
+        k: float,
+        ntree: int,
+        tau_num: FloatLike | None,
+    ):
+        if tau_num is None:
+            if y_train.dtype == bool:
+                tau_num = 3.0
+            elif y_train.size < 2:
+                tau_num = 1.0
+            else:
+                tau_num = (y_train.max() - y_train.min()) / 2
+
+        return tau_num / (k * math.sqrt(ntree))
+
+    @staticmethod
+    def _determine_splits(
+        x_train: Real[Array, 'p n'],
+        usequants: bool,
+        numcut: int,
+        xinfo: Float[Array, 'p n'] | None,
+    ) -> tuple[Real[Array, 'p m'], UInt[Array, ' p']]:
+        if xinfo is not None:
+            if xinfo.ndim != 2 or xinfo.shape[0] != x_train.shape[0]:
+                msg = f'{xinfo.shape=} different from expected ({x_train.shape[0]}, *)'
+                raise ValueError(msg)
+            return prepcovars.parse_xinfo(xinfo)
+        elif usequants:
             return prepcovars.quantilized_splits_from_matrix(x_train, numcut + 1)
         else:
             return prepcovars.uniform_splits_from_matrix(x_train, numcut + 1)
 
     @staticmethod
-    def _bin_predictors(x, splits):
+    def _bin_predictors(x, splits) -> UInt[Array, 'p n']:
         return prepcovars.bin_predictors(x, splits)
 
-    @staticmethod
-    def _transform_input(y, lamda, offset, scale):
-        y = (y - offset) / scale
-        lamda = lamda / (scale * scale)
-        return y, lamda
-
     @staticmethod
     def _setup_mcmc(
-        x_train,
-        y_train,
-        w,
-        max_split,
-        lamda,
-        sigdf,
-        power,
-        base,
-        maxdepth,
-        ntree,
-        initkw,
+        x_train: Real[Array, 'p n'],
+        y_train: Float32[Array, ' n'] | Bool[Array, ' n'],
+        offset: Float32[Array, ''],
+        w: Float[Array, ' n'] | None,
+        max_split: UInt[Array, ' p'],
+        lamda: Float32[Array, ''] | None,
+        sigma_mu: FloatLike,
+        sigdf: FloatLike,
+        power: FloatLike,
+        base: FloatLike,
+        maxdepth: int,
+        ntree: int,
+        init_kw: dict[str, Any] | None,
+        rm_const: bool | None,
+        theta: FloatLike | None,
+        a: FloatLike | None,
+        b: FloatLike | None,
+        rho: FloatLike | None,
     ):
         depth = jnp.arange(maxdepth - 1)
         p_nonterminal = base / (1 + depth).astype(float) ** power
-        sigma2_alpha = sigdf / 2
-        sigma2_beta = lamda * sigma2_alpha
+
+        if y_train.dtype == bool:
+            sigma2_alpha = None
+            sigma2_beta = None
+        else:
+            sigma2_alpha = sigdf / 2
+            sigma2_beta = lamda * sigma2_alpha
+
         kw = dict(
             X=x_train,
-            y=y_train,
+            # copy y_train because it's going to be donated in the mcmc loop
+            y=jnp.array(y_train),
+            offset=offset,
             error_scale=w,
             max_split=max_split,
             num_trees=ntree,
             p_nonterminal=p_nonterminal,
+            sigma_mu2=jnp.square(sigma_mu),
             sigma2_alpha=sigma2_alpha,
             sigma2_beta=sigma2_beta,
+            min_points_per_decision_node=10,
             min_points_per_leaf=5,
+            theta=theta,
+            a=a,
+            b=b,
+            rho=rho,
         )
-        if initkw is not None:
-            kw.update(initkw)
+
+        if rm_const is None:
+            kw.update(filter_splitless_vars=False)
+        elif rm_const:
+            kw.update(filter_splitless_vars=True)
+        else:
+            n_empty = jnp.count_nonzero(max_split == 0)
+            if n_empty:
+                msg = f'There are {n_empty}/{max_split.size} predictors without decision rules'
+                raise ValueError(msg)
+            kw.update(filter_splitless_vars=False)
+
+        if init_kw is not None:
+            kw.update(init_kw)
+
         return mcmcstep.init(**kw)
 
     @staticmethod
-    def _run_mcmc(mcmc_state, ndpost, nskip, keepevery, printevery, seed):
+    def _run_mcmc(
+        mcmc_state: mcmcstep.State,
+        ndpost: int,
+        nskip: int,
+        keepevery: int,
+        printevery: int | None,
+        seed: int | Integer[Array, ''] | Key[Array, ''],
+        run_mcmc_kw: dict | None,
+        sparse: bool,
+    ):
+        # prepare random generator seed
         if isinstance(seed, jax.Array) and jnp.issubdtype(
             seed.dtype, jax.dtypes.prng_key
         ):
-            key = seed
+            key = seed.copy()
+            # copy because the inner loop in run_mcmc will donate the buffer
         else:
             key = jax.random.key(seed)
-        callback = mcmcloop.make_simple_print_callback(printevery)
-        return mcmcloop.run_mcmc(key, mcmc_state, nskip, ndpost, keepevery, callback)
 
-    @staticmethod
-    def _predict(trace, x):
-        return mcmcloop.evaluate_trace(trace, x)
-
-    @staticmethod
-    def _transform_output(y, offset, scale):
-        return offset + scale * y
-
-    @staticmethod
-    def _extract_sigma(trace, scale):
-        return scale * jnp.sqrt(trace['sigma2'])
-
-    def _show_tree(self, i_sample, i_tree, print_all=False):
-        from . import debug
-
-        trace = self._main_trace
-        leaf_tree = trace['leaf_trees'][i_sample, i_tree]
-        var_tree = trace['var_trees'][i_sample, i_tree]
-        split_tree = trace['split_trees'][i_sample, i_tree]
-        debug.print_tree(leaf_tree, var_tree, split_tree, print_all)
-
-    def _sigma_harmonic_mean(self, prior=False):
-        bart = self._mcmc_state
-        if prior:
-            alpha = bart['sigma2_alpha']
-            beta = bart['sigma2_beta']
-        else:
-            resid = bart['resid']
-            alpha = bart['sigma2_alpha'] + resid.size / 2
-            norm2 = jnp.dot(
-                resid, resid, preferred_element_type=bart['sigma2_beta'].dtype
+        # prepare arguments
+        kw = dict(n_burn=nskip, n_skip=keepevery, inner_loop_length=printevery)
+        kw.update(
+            mcmcloop.make_default_callback(
+                dot_every=None if printevery is None or printevery == 1 else 1,
+                report_every=printevery,
+                sparse_on_at=nskip // 2 if sparse else None,
             )
-            beta = bart['sigma2_beta'] + norm2 / 2
-        sigma2 = beta / alpha
-        return jnp.sqrt(sigma2) * self.scale
-
-    def _compare_resid(self):
-        bart = self._mcmc_state
-        resid1 = bart['resid']
-        yhat = grove.evaluate_forest(
-            bart['X'],
-            bart['leaf_trees'],
-            bart['var_trees'],
-            bart['split_trees'],
-            jnp.float32,
         )
-        resid2 = bart['y'] - yhat
-        return resid1, resid2
-
-    def _avg_acc(self):
-        trace = self._main_trace
-
-        def acc(prefix):
-            acc = trace[f'{prefix}_acc_count']
-            prop = trace[f'{prefix}_prop_count']
-            return acc.sum() / prop.sum()
-
-        return acc('grow'), acc('prune')
-
-    def _avg_prop(self):
-        trace = self._main_trace
-
-        def prop(prefix):
-            return trace[f'{prefix}_prop_count'].sum()
-
-        pgrow = prop('grow')
-        pprune = prop('prune')
-        total = pgrow + pprune
-        return pgrow / total, pprune / total
-
-    def _avg_move(self):
-        agrow, aprune = self._avg_acc()
-        pgrow, pprune = self._avg_prop()
-        return agrow * pgrow, aprune * pprune
-
-    def _depth_distr(self):
-        from . import debug
-
-        trace = self._main_trace
-        split_trees = trace['split_trees']
-        return debug.trace_depth_distr(split_trees)
-
-    def _points_per_leaf_distr(self):
-        from . import debug
-
-        return debug.trace_points_per_leaf_distr(
-            self._main_trace, self._mcmc_state['X']
-        )
-
-    def _check_trees(self):
-        from . import debug
+        if run_mcmc_kw is not None:
+            kw.update(run_mcmc_kw)
 
-        return debug.check_trace(self._main_trace, self._mcmc_state)
+        return mcmcloop.run_mcmc(key, mcmc_state, ndpost, **kw)
 
-    def _tree_goes_bad(self):
-        bad = self._check_trees().astype(bool)
-        bad_before = jnp.pad(bad[:-1], [(1, 0), (0, 0)])
-        return bad & ~bad_before
+    def _predict(self, x):
+        return mcmcloop.evaluate_trace(self._main_trace, x)