scipy.stats.ttest_ind¶

scipy.stats.
ttest_ind
(a, b, axis=0, equal_var=True, nan_policy='propagate', permutations=None, random_state=None, alternative='twosided', trim=0)[source]¶ Calculate the Ttest for the means of two independent samples of scores.
This is a twosided test for the null hypothesis that 2 independent samples have identical average (expected) values. This test assumes that the populations have identical variances by default.
 Parameters
 a, barray_like
The arrays must have the same shape, except in the dimension corresponding to axis (the first, by default).
 axisint or None, optional
Axis along which to compute test. If None, compute over the whole arrays, a, and b.
 equal_varbool, optional
If True (default), perform a standard independent 2 sample test that assumes equal population variances [1]. If False, perform Welch’s ttest, which does not assume equal population variance [2].
New in version 0.11.0.
 nan_policy{‘propagate’, ‘raise’, ‘omit’}, optional
Defines how to handle when input contains nan. The following options are available (default is ‘propagate’):
‘propagate’: returns nan
‘raise’: throws an error
‘omit’: performs the calculations ignoring nan values
The ‘omit’ option is not currently available for permutation tests or onesided asympyotic tests.
 permutationsnonnegative int, np.inf, or None (default), optional
If 0 or None (default), use the tdistribution to calculate pvalues. Otherwise, permutations is the number of random permutations that will be used to estimate pvalues using a permutation test. If permutations equals or exceeds the number of distinct partitions of the pooled data, an exact test is performed instead (i.e. each distinct partition is used exactly once). See Notes for details.
New in version 1.7.0.
 random_state{None, int,
numpy.random.Generator
, numpy.random.RandomState
}, optionalIf seed is None (or np.random), the
numpy.random.RandomState
singleton is used. If seed is an int, a newRandomState
instance is used, seeded with seed. If seed is already aGenerator
orRandomState
instance then that instance is used.Pseudorandom number generator state used to generate permutations (used only when permutations is not None).
New in version 1.7.0.
 alternative{‘twosided’, ‘less’, ‘greater’}, optional
Defines the alternative hypothesis. The following options are available (default is ‘twosided’):
‘twosided’
‘less’: onesided
‘greater’: onesided
New in version 1.6.0.
 trimfloat, optional
If nonzero, performs a trimmed (Yuen’s) ttest. Defines the fraction of elements to be trimmed from each end of the input samples. If 0 (default), no elements will be trimmed from either side. The number of trimmed elements from each tail is the floor of the trim times the number of elements. Valid range is [0, .5).
New in version 1.7.
 Returns
 statisticfloat or array
The calculated tstatistic.
 pvaluefloat or array
The twotailed pvalue.
Notes
Suppose we observe two independent samples, e.g. flower petal lengths, and we are considering whether the two samples were drawn from the same population (e.g. the same species of flower or two species with similar petal characteristics) or two different populations.
The ttest quantifies the difference between the arithmetic means of the two samples. The pvalue quantifies the probability of observing as or more extreme values assuming the null hypothesis, that the samples are drawn from populations with the same population means, is true. A pvalue larger than a chosen threshold (e.g. 5% or 1%) indicates that our observation is not so unlikely to have occurred by chance. Therefore, we do not reject the null hypothesis of equal population means. If the pvalue is smaller than our threshold, then we have evidence against the null hypothesis of equal population means.
By default, the pvalue is determined by comparing the tstatistic of the observed data against a theoretical tdistribution. When
1 < permutations < binom(n, k)
, wherek
is the number of observations in a,n
is the total number of observations in a and b, andbinom(n, k)
is the binomial coefficient (n
choosek
),
the data are pooled (concatenated), randomly assigned to either group a or b, and the tstatistic is calculated. This process is performed repeatedly (permutation times), generating a distribution of the tstatistic under the null hypothesis, and the tstatistic of the observed data is compared to this distribution to determine the pvalue. When
permutations >= binom(n, k)
, an exact test is performed: the data are partitioned between the groups in each distinct way exactly once.The permutation test can be computationally expensive and not necessarily more accurate than the analytical test, but it does not make strong assumptions about the shape of the underlying distribution.
Use of trimming is commonly referred to as the trimmed ttest. At times called Yuen’s ttest, this is an extension of Welch’s ttest, with the difference being the use of winsorized means in calculation of the variance and the trimmed sample size in calculation of the statistic. Trimming is reccomended if the underlying distribution is longtailed or contaminated with outliers [4].
References
 1
https://en.wikipedia.org/wiki/Ttest#Independent_twosample_ttest
 2
 3
 4
Yuen, Karen K. “The TwoSample Trimmed t for Unequal Population Variances.” Biometrika, vol. 61, no. 1, 1974, pp. 165170. JSTOR, www.jstor.org/stable/2334299. Accessed 30 Mar. 2021.
 5
Yuen, Karen K., and W. J. Dixon. “The Approximate Behaviour and Performance of the TwoSample Trimmed t.” Biometrika, vol. 60, no. 2, 1973, pp. 369374. JSTOR, www.jstor.org/stable/2334550. Accessed 30 Mar. 2021.
Examples
>>> from scipy import stats >>> rng = np.random.default_rng()
Test with sample with identical means:
>>> rvs1 = stats.norm.rvs(loc=5, scale=10, size=500, random_state=rng) >>> rvs2 = stats.norm.rvs(loc=5, scale=10, size=500, random_state=rng) >>> stats.ttest_ind(rvs1, rvs2) Ttest_indResult(statistic=0.4390847099199348, pvalue=0.6606952038870015) >>> stats.ttest_ind(rvs1, rvs2, equal_var=False) Ttest_indResult(statistic=0.4390847099199348, pvalue=0.6606952553131064)
ttest_ind
underestimates p for unequal variances:>>> rvs3 = stats.norm.rvs(loc=5, scale=20, size=500, random_state=rng) >>> stats.ttest_ind(rvs1, rvs3) Ttest_indResult(statistic=1.6370984482905417, pvalue=0.1019251574705033) >>> stats.ttest_ind(rvs1, rvs3, equal_var=False) Ttest_indResult(statistic=1.637098448290542, pvalue=0.10202110497954867)
When
n1 != n2
, the equal variance tstatistic is no longer equal to the unequal variance tstatistic:>>> rvs4 = stats.norm.rvs(loc=5, scale=20, size=100, random_state=rng) >>> stats.ttest_ind(rvs1, rvs4) Ttest_indResult(statistic=1.9481646859513422, pvalue=0.05186270935842703) >>> stats.ttest_ind(rvs1, rvs4, equal_var=False) Ttest_indResult(statistic=1.3146566100751664, pvalue=0.1913495266513811)
Ttest with different means, variance, and n:
>>> rvs5 = stats.norm.rvs(loc=8, scale=20, size=100, random_state=rng) >>> stats.ttest_ind(rvs1, rvs5) Ttest_indResult(statistic=2.8415950600298774, pvalue=0.0046418707568707885) >>> stats.ttest_ind(rvs1, rvs5, equal_var=False) Ttest_indResult(statistic=1.8686598649188084, pvalue=0.06434714193919686)
When performing a permutation test, more permutations typically yields more accurate results. Use a
np.random.Generator
to ensure reproducibility:>>> stats.ttest_ind(rvs1, rvs5, permutations=10000, ... random_state=rng) Ttest_indResult(statistic=2.8415950600298774, pvalue=0.0052)
Take these two samples, one of which has an extreme tail.
>>> a = (56, 128.6, 12, 123.8, 64.34, 78, 763.3) >>> b = (1.1, 2.9, 4.2)
Use the trim keyword to perform a trimmed (Yuen) ttest. For example, using 20% trimming,
trim=.2
, the test will reduce the impact of one (np.floor(trim*len(a))
) element from each tail of sample a. It will have no effect on sample b becausenp.floor(trim*len(b))
is 0.>>> stats.ttest_ind(a, b, trim=.2) Ttest_indResult(statistic=3.4463884028073513, pvalue=0.01369338726499547)