Intro to the stats module

from scipy import stats
import numpy as np
from xarray_einstats.stats import XrContinuousRV, rankdata, hmean, skew, median_abs_deviation
from xarray_einstats.tutorial import generate_mcmc_like_dataset
ds = generate_mcmc_like_dataset(11)
ds
<xarray.Dataset> Size: 6kB
Dimensions:  (plot_dim: 20, chain: 4, draw: 10, team: 6, match: 12)
Coordinates:
  * team     (team) <U1 24B 'a' 'b' 'c' 'd' 'e' 'f'
  * chain    (chain) int64 32B 0 1 2 3
  * draw     (draw) int64 80B 0 1 2 3 4 5 6 7 8 9
Dimensions without coordinates: plot_dim, match
Data variables:
    x_plot   (plot_dim) float64 160B 0.0 0.5263 1.053 1.579 ... 8.947 9.474 10.0
    mu       (chain, draw, team) float64 2kB 0.2296 0.5383 ... 0.4452 2.004
    sigma    (chain, draw) float64 320B 0.3703 0.00899 0.1398 ... 0.2246 0.2875
    score    (chain, draw, match) int64 4kB 1 0 0 1 4 1 1 0 ... 0 0 2 1 2 1 0 2

Probability distributions

Initialization

Initialization takes a class from an external library that defines distributions with array inputs and the positional/keyword arguments needed to initialize it. Tested distributions are:

  • SciPy distributions

  • PreliZ distributions

  • SciPy random variables (introduced in SciPy 1.14).

Note

Positional and keyword arguments are broadcasted with all other inputs, but they are then reconstructed and passed as is. For example: preliz.Normal(0, 1) is valid, so XrContinuousRV(pz.Normal, 0, 1) is also valid but scipy.stats.Normal(0, 1) is not valid only keyword arguments are accepted, so the same will apply when wrapping with with xarray-einstats.

Moreover, the use of XrContinuousRV and XrDiscreteRV only changes for some of the methods like pdf/pmf. However, PreliZ always uses pdf, even for discrete distributions so it needs to always be wrapped through XrContinuousRV

dist = stats.norm      # SciPy distribution
# import preliz as pz
# dist = pz.Normal     # PreliZ distribution
# dist = stats.Normal  # SciPy random variable

norm = XrContinuousRV(dist, ds["mu"], ds["sigma"])

Using its methods

Once initialized, you can use its methods exactly as you’d use them with frozen scipy distributions or preliz distributions. The only two differences are

  1. They now take scalars or DataArrays as inputs, arrays are only accepted as the arguments on which to evaluate the methods (in scipy docs they are represented by x, k or q depending on the method)

  2. size behaves differently in the rvs method; here size represents the shape of a single draw. This ensures that you don’t need to care about any broadcasting or alignment of arrays, xarray_einstats does this for you and will return a DataArray with combined shape (*size, *broadcasted_input_shape).

You can generate 100 random draws from the initialized distribution. As we have mentioned and unlike what would happen with scipy the output won’t have shape 100 but instead will have shape 100, *broadcasted_input_shape. xarray generates the broadcasted_input_shape and size is independent from it so you can relax and not care about broadcasting.

norm.rvs(size=100)
<xarray.DataArray (rv_dim0: 100, chain: 4, draw: 10, team: 6)> Size: 192kB
-0.1369 0.312 1.122 0.296 -0.3041 2.919 ... 0.2673 0.07624 0.02325 0.3478 2.179
Coordinates:
  * team     (team) <U1 24B 'a' 'b' 'c' 'd' 'e' 'f'
  * chain    (chain) int64 32B 0 1 2 3
  * draw     (draw) int64 80B 0 1 2 3 4 5 6 7 8 9
Dimensions without coordinates: rv_dim0

If the dimension names are not provided, xarray_einstats assings rv_dim# as dimension name as many times as necessary. To define the names manually you can use the dims argument:

norm.rvs(size=(5, 3), dims=["subject", "batch"])
<xarray.DataArray (subject: 5, batch: 3, chain: 4, draw: 10, team: 6)> Size: 29kB
0.1375 0.729 1.113 0.5688 0.1979 3.794 ... -0.1246 0.2988 0.3684 0.1294 1.877
Coordinates:
  * team     (team) <U1 24B 'a' 'b' 'c' 'd' 'e' 'f'
  * chain    (chain) int64 32B 0 1 2 3
  * draw     (draw) int64 80B 0 1 2 3 4 5 6 7 8 9
Dimensions without coordinates: subject, batch

In the output above we’ll often want to provide coordinate values to the new dimensions. .rvs takes a coords argument which can be used for that, but it isn’t much of an improvement over .rvs(...).assign_coords(coords). It is also possible however to skip size and dims altogether and use coords to inform both:

subject = [
    "Monstera deliciosa", "Monstera borsigiana", "Monstera siltepecana",
    "Monstera variegata", "Monstera pinnatipartita"
]
batch = ["March", "June", "October"]
norm.rvs(coords={"subject": subject, "batch": batch})
<xarray.DataArray (subject: 5, batch: 3, chain: 4, draw: 10, team: 6)> Size: 29kB
-0.1794 0.5826 1.101 -0.07705 0.1534 3.711 ... 0.7631 0.2766 0.4849 0.5801 2.014
Coordinates:
  * subject  (subject) <U23 460B 'Monstera deliciosa' ... 'Monstera pinnatipa...
  * batch    (batch) <U7 84B 'March' 'June' 'October'
  * team     (team) <U1 24B 'a' 'b' 'c' 'd' 'e' 'f'
  * chain    (chain) int64 32B 0 1 2 3
  * draw     (draw) int64 80B 0 1 2 3 4 5 6 7 8 9

The behaviour for other methods is similar:

norm.logpdf(ds["x_plot"])
<xarray.DataArray 'x_plot' (plot_dim: 20, chain: 4, draw: 10, team: 6)> Size: 38kB
-0.1177 -0.9821 -4.519 0.06682 0.02491 ... -594.3 -600.6 -561.9 -551.8 -386.4
Coordinates:
  * chain    (chain) int64 32B 0 1 2 3
  * draw     (draw) int64 80B 0 1 2 3 4 5 6 7 8 9
  * team     (team) <U1 24B 'a' 'b' 'c' 'd' 'e' 'f'
Dimensions without coordinates: plot_dim

For convenience, you can also use array_like input which is converted to a DataArray under the hood. In such cases, the dimension name is quantile for ppf and isf, point otherwise. In both cases, the values passed as input are preserved as coordinate values.

norm.ppf([.25, .5, .75])
<xarray.DataArray (quantile: 3, chain: 4, draw: 10, team: 6)> Size: 6kB
-0.02018 0.2885 0.8726 -0.204 -0.1332 ... 0.2786 0.2264 0.5523 0.6391 2.198
Coordinates:
  * quantile  (quantile) float64 24B 0.25 0.5 0.75
  * chain     (chain) int64 32B 0 1 2 3
  * draw      (draw) int64 80B 0 1 2 3 4 5 6 7 8 9
  * team      (team) <U1 24B 'a' 'b' 'c' 'd' 'e' 'f'
pdf = norm.pdf(np.linspace(-5, 5))
pdf
<xarray.DataArray (point: 50, chain: 4, draw: 10, team: 6)> Size: 96kB
5.321e-44 2.898e-49 4.753e-60 5.206e-41 ... 3.563e-57 4.449e-55 3.664e-24
Coordinates:
  * point    (point) float64 400B -5.0 -4.796 -4.592 -4.388 ... 4.592 4.796 5.0
  * chain    (chain) int64 32B 0 1 2 3
  * draw     (draw) int64 80B 0 1 2 3 4 5 6 7 8 9
  * team     (team) <U1 24B 'a' 'b' 'c' 'd' 'e' 'f'

Plot a subset of the pdf we just calculated with matplotlib.

import matplotlib.pyplot as plt
plt.rcParams["figure.facecolor"] = "white"

fig, ax = plt.subplots()
ax.plot(pdf.point, pdf.sel(team="d", chain=2), color="C0", alpha=.5)
ax.set(xlabel="x", ylabel="pdf of normal distribution", );
../_images/282dd206ef33a1272e380894e15d1a59996ddc4a66a1aa084fc9d91e5e65e746.png

Other functions

The rest of the functions in the module have a very similar API to their scipy counterparts, the only differences are:

  • They take dims instead of axis. Moreover, dims can be str or a sequence of str instead of a single integer only as supported by axis.

  • Arguments that take array_like as values take DataArray inputs instead. For example the scale argument in median_abs_deviation

  • They accept extra arbitrary kwargs, that are passed to xarray.apply_ufunc.

Here are some examples of using functions in the stats module of xarray_einstats with dims argument instead of axis.

hmean(ds["mu"], dims="team")
<xarray.DataArray 'mu' (chain: 4, draw: 10)> Size: 320B
0.1588 0.2123 0.5543 0.7826 0.1913 0.6035 ... 0.1269 0.712 0.3044 0.1936 0.1223
Coordinates:
  * chain    (chain) int64 32B 0 1 2 3
  * draw     (draw) int64 80B 0 1 2 3 4 5 6 7 8 9
rankdata(ds["score"], dims=("chain", "draw"), method="min")
<xarray.DataArray 'score' (match: 12, chain: 4, draw: 10)> Size: 4kB
14 14 14 14 14 31 14 1 31 14 31 1 14 1 ... 15 15 15 15 15 1 34 15 15 1 34 34 34
Dimensions without coordinates: match, chain, draw

Important

The statistical summaries and other statistical functions can take both DataArray and Dataset. Methods in probability functions and functions in linear algebra module are tested only on DataArrays.

When using Dataset inputs, you must make sure that all the dimensions in dims are present in all the DataArrays within the Dataset.

skew(ds[["score", "mu", "sigma"]], dims=("chain", "draw"))
<xarray.Dataset> Size: 176B
Dimensions:  (match: 12, team: 6)
Coordinates:
  * team     (team) <U1 24B 'a' 'b' 'c' 'd' 'e' 'f'
Dimensions without coordinates: match
Data variables:
    score    (match) float64 96B 1.466 0.2149 0.6788 1.361 ... 1.099 1.156 1.265
    mu       (team) float64 48B 0.8152 1.84 2.102 1.806 1.091 0.9678
    sigma    float64 8B 1.314
median_abs_deviation(ds)
<xarray.Dataset> Size: 32B
Dimensions:  ()
Data variables:
    x_plot   float64 8B 2.632
    mu       float64 8B 0.4878
    sigma    float64 8B 0.39
    score    float64 8B 1.0
%load_ext watermark
%watermark -n -u -v -iv -w -p xarray_einstats,xarray
Last updated: Thu May 22 2025

Python implementation: CPython
Python version       : 3.12.7
IPython version      : 8.29.0

xarray_einstats: 0.9.0
xarray         : 2025.4.0

matplotlib     : 3.10.1
numpy          : 2.2.6
xarray_einstats: 0.9.0
scipy          : 1.15.2

Watermark: 2.5.0