bundles / scipy 1.17.1 / scipy / stats / _stats_py / quantile_test

function

`scipy.stats._stats_py:quantile_test`

source: /scipy/stats/_stats_py.py :9048

Signature

   def     quantile_test (    x    ,  * ,     q    =  0   ,    p    =  0.5   ,    alternative    =  two-sided     )

Summary

Perform a quantile test and compute a confidence interval of the quantile.

Extended Summary

This function tests the null hypothesis that q is the value of the quantile associated with probability p of the population underlying sample x. For example, with default parameters, it tests that the median of the population underlying x is zero. The function returns an object including the test statistic, a p-value, and a method for computing the confidence interval around the quantile.

Parameters

x : array_like

A one-dimensional sample.

q : float, default: 0

The hypothesized value of the quantile.

p : float, default: 0.5

The probability associated with the quantile; i.e. the proportion of the population less than q is p. Must be strictly between 0 and 1.

alternative : {'two-sided', 'less', 'greater'}, optional

Defines the alternative hypothesis. The following options are available (default is 'two-sided'):

'two-sided': the quantile associated with the probability p is not q.
'less': the quantile associated with the probability p is less than q.
'greater': the quantile associated with the probability p is greater than q.

Returns

result : QuantileTestResult

An object with the following attributes:

statistic: statistic
statistic_type: statistic_type
pvalue: pvalue

The object also has the following method:

confidence_interval(confidence_level=0.95): Computes a confidence interval around the the population quantile associated with the probability p. The confidence interval is returned in a namedtuple with fields low and high. Values are nan when there are not enough observations to compute the confidence interval at the desired confidence.

Notes

This test and its method for computing confidence intervals are non-parametric. They are valid if and only if the observations are i.i.d.

The implementation of the test follows Conover ^[1]. Two test statistics are considered.

T1: The number of observations in x less than or equal to q.

T1 = (x <= q).sum()

T2: The number of observations in x strictly less than q.

T2 = (x < q).sum()

The use of two test statistics is necessary to handle the possibility that x was generated from a discrete or mixed distribution.

The null hypothesis for the test is:

H0: The $p^{th}$ population quantile is q.

and the null distribution for each test statistic is $binom (n, p)$ . When alternative='less', the alternative hypothesis is:

H1: The $p^{th}$ population quantile is less than q.

and the p-value is the probability that the binomial random variable

Y \sim binom (n, p)

is greater than or equal to the observed value T2.

When alternative='greater', the alternative hypothesis is:

H1: The $p^{th}$ population quantile is greater than q

and the p-value is the probability that the binomial random variable Y is less than or equal to the observed value T1.

When alternative='two-sided', the alternative hypothesis is

H1: q is not the $p^{th}$ population quantile.

and the p-value is twice the smaller of the p-values for the 'less' and 'greater' cases. Both of these p-values can exceed 0.5 for the same data, so the value is clipped into the interval $[0, 1]$ .

The approach for confidence intervals is attributed to Thompson ^[2] and later proven to be applicable to any set of i.i.d. samples ^[3]. The computation is based on the observation that the probability of a quantile $q$ to be larger than any observations $x_{m} (1 \leq m \leq N)$ can be computed as

P (x_{m} \leq q) = 1 - k = 0 \sum m - 1 (k N) q^{k} (1 - q)^{N - k}

By default, confidence intervals are computed for a 95% confidence level. A common interpretation of a 95% confidence intervals is that if i.i.d. samples are drawn repeatedly from the same population and confidence intervals are formed each time, the confidence interval will contain the true value of the specified quantile in approximately 95% of trials.

A similar function is available in the QuantileNPCI R package ^[4]. The foundation is the same, but it computes the confidence interval bounds by doing interpolations between the sample values, whereas this function uses only sample values as bounds. Thus, quantile_test.confidence_interval returns more conservative intervals (i.e., larger).

The same computation of confidence intervals for quantiles is included in the confintr package ^[5].

Two-sided confidence intervals are not guaranteed to be optimal; i.e., there may exist a tighter interval that may contain the quantile of interest with probability larger than the confidence level. Without further assumption on the samples (e.g., the nature of the underlying distribution), the one-sided intervals are optimally tight.

Array API Standard Support

quantile_test has experimental support for Python Array API Standard compatible backends in addition to NumPy. Please consider testing these features by setting an environment variable SCIPY_ARRAY_API=1 and providing CuPy, PyTorch, JAX, or Dask arrays as array arguments. The following combinations of backend and device (or other capability) are supported.

====================  ====================  ====================
Library               CPU                   GPU
====================  ====================  ====================
NumPy                 ✅                     n/a                 
CuPy                  n/a                   ⛔                   
PyTorch               ⛔                     ⛔                   
JAX                   ⛔                     ⛔                   
Dask                  ⛔                     n/a                 
====================  ====================  ====================

See dev-arrayapi for more information.

Examples

Suppose we wish to test the null hypothesis that the median of a population is equal to 0.5. We choose a confidence level of 99%; that is, we will reject the null hypothesis in favor of the alternative if the p-value is less than 0.01. When testing random variates from the standard uniform distribution, which has a median of 0.5, we expect the data to be consistent with the null hypothesis most of the time.

import numpy as np
from scipy import stats
rng = np.random.default_rng(6981396440634228121)
rvs = stats.uniform.rvs(size=100, random_state=rng)

✓

stats.quantile_test(rvs, q=0.5, p=0.5)

✗

As expected, the p-value is not below our threshold of 0.01, so we cannot reject the null hypothesis. When testing data from the standard *normal* distribution, which has a median of 0, we would expect the null hypothesis to be rejected.

rvs = stats.norm.rvs(size=100, random_state=rng)

✓

stats.quantile_test(rvs, q=0.5, p=0.5)

✗

Indeed, the p-value is lower than our threshold of 0.01, so we reject the null hypothesis in favor of the default "two-sided" alternative: the median of the population is *not* equal to 0.5. However, suppose we were to test the null hypothesis against the one-sided alternative that the median of the population is *greater* than 0.5. Since the median of the standard normal is less than 0.5, we would not expect the null hypothesis to be rejected.

stats.quantile_test(rvs, q=0.5, p=0.5, alternative='greater')

✗

Unsurprisingly, with a p-value greater than our threshold, we would not reject the null hypothesis in favor of the chosen alternative. The quantile test can be used for any quantile, not only the median. For example, we can test whether the third quartile of the distribution underlying the sample is greater than 0.6.

rvs = stats.uniform.rvs(size=100, random_state=rng)

✓

stats.quantile_test(rvs, q=0.6, p=0.75, alternative='greater')

✗

The p-value is lower than the threshold. We reject the null hypothesis in favor of the alternative: the third quartile of the distribution underlying our sample is greater than 0.6. `quantile_test` can also compute confidence intervals for any quantile.

rvs = stats.norm.rvs(size=100, random_state=rng)
res = stats.quantile_test(rvs, q=0.6, p=0.75)
ci = res.confidence_interval(confidence_level=0.95)

✓

ci

✗

When testing a one-sided alternative, the confidence interval contains all observations such that if passed as `q`, the p-value of the test would be greater than 0.05, and therefore the null hypothesis would not be rejected. For example:

rvs.sort()
q, p, alpha = 0.6, 0.75, 0.95
res = stats.quantile_test(rvs, q=q, p=p, alternative='less')
ci = res.confidence_interval(confidence_level=alpha)
for x in rvs[rvs <= ci.high]:
    res = stats.quantile_test(rvs, q=x, p=p, alternative='less')
    assert res.pvalue > 1-alpha
for x in rvs[rvs > ci.high]:
    res = stats.quantile_test(rvs, q=x, p=p, alternative='less')
    assert res.pvalue < 1-alpha

✓

Also, if a 95% confidence interval is repeatedly generated for random samples, the confidence interval will contain the true quantile value in approximately 95% of replications.

dist = stats.rayleigh() # our "unknown" distribution
p = 0.2
true_stat = dist.ppf(p) # the true value of the statistic
n_trials = 1000
quantile_ci_contains_true_stat = 0
for i in range(n_trials):
    data = dist.rvs(size=100, random_state=rng)
    res = stats.quantile_test(data, p=p)
    ci = res.confidence_interval(0.95)
    if ci[0] < true_stat < ci[1]:
        quantile_ci_contains_true_stat += 1
quantile_ci_contains_true_stat >= 950

✓

This works with any distribution and any quantile, as long as the samples are i.i.d.

Aliases

scipy.stats.quantile_test