paired_ttest_resample: Resampled paired t test

Resampled paired t test procedure to compare the performance of two models

from mlxtend.evaluate import paired_ttest_resample


Resampled paired t test procedure (also called k-hold-out paired t test) is a popular method for comparing the performance of two models (classifiers or regressors); however, this method has many drawbacks and is not recommended to be used in practice [1], and techniques such as the paired_ttest_5x2cv should be used instead.

To explain how this method works, let's consider to estimator (e.g., classifiers) A and B. Further, we have a labeled dataset D. In the common hold-out method, we typically split the dataset into 2 parts: a training and a test set. In the resampled paired t test procedure, we repeat this splitting procedure (with typically 2/3 training data and 1/3 test data) k times (usually 30). In each iteration, we train A and B on the training set and evaluate it on the test set. Then, we compute the difference in performance between A and B in each iteration so that we obtain k difference measures. Now, by making the assumption that these k differences were independently drawn and follow an approximately normal distribution, we can compute the following t statistic with k-1 degrees of freedom according to Student's t test, under the null hypothesis that the models A and B have equal performance:

Here, computes the difference between the model performances in the th iteration, , and represents the average difference between the classifier performances, .

Once we computed the t statistic we can compute the p value and compare it to our chosen significance level, e.g., . If the p value is smaller than , we reject the null hypothesis and accept that there is a significant difference in the two models.

To summarize the procedure:

  1. i := 0
  2. while i < k:
  3. split dataset into training and test subsets
  4. fit models A and B to the training set
  5. compute the performances of A and B on the test set
  6. record the performance difference between A and B
  7. i := i + 1
  8. compute t-statistic
  9. compute p value from t-statistic with k-1 degrees of freedom
  10. compare p value to chosen significance threshold

The problem with this method, and the reason why it is not recommended to be used in practice, is that it violates the assumptions of Student's t test [1]:


Example 1 - Resampled paired t test

Assume we want to compare two classification algorithms, logistic regression and a decision tree algorithm:

from sklearn.linear_model import LogisticRegression
from sklearn.tree import DecisionTreeClassifier
from import iris_data
from sklearn.model_selection import train_test_split

X, y = iris_data()
clf1 = LogisticRegression(random_state=1)
clf2 = DecisionTreeClassifier(random_state=1)

X_train, X_test, y_train, y_test = \
    train_test_split(X, y, test_size=0.25,

score1 =, y_train).score(X_test, y_test)
score2 =, y_train).score(X_test, y_test)

print('Logistic regression accuracy: %.2f%%' % (score1*100))
print('Decision tree accuracy: %.2f%%' % (score2*100))
Logistic regression accuracy: 97.37%
Decision tree accuracy: 94.74%

Note that these accuracy values are not used in the paired t test procedure as new test/train splits are generated during the resampling procedure, the values above are just serving the purpose of intuition.

Now, let's assume a significance threshold of for rejecting the null hypothesis that both algorithms perform equally well on the dataset and conduct the paired sample t test:

from mlxtend.evaluate import paired_ttest_resampled

t, p = paired_ttest_resampled(estimator1=clf1,
                              X=X, y=y,

print('t statistic: %.3f' % t)
print('p value: %.3f' % p)
t statistic: -1.809
p value: 0.081

Since , we cannot reject the null hypothesis and may conclude that the performance of the two algorithms is not significantly different.

While it is generally not recommended to apply statistical tests multiple times without correction for multiple hypothesis testing, let us take a look at an example where the decision tree algorithm is limited to producing a very simple decision boundary that would result in a relatively bad performance:

clf2 = DecisionTreeClassifier(random_state=1, max_depth=1)

score2 =, y_train).score(X_test, y_test)
print('Decision tree accuracy: %.2f%%' % (score2*100))

t, p = paired_ttest_resampled(estimator1=clf1,
                              X=X, y=y,

print('t statistic: %.3f' % t)
print('p value: %.3f' % p)
Decision tree accuracy: 63.16%
t statistic: 39.214
p value: 0.000

Assuming that we conducted this test also with a significance level of , we can reject the null-hypothesis that both models perform equally well on this dataset, since the p-value () is smaller than .


paired_ttest_resampled(estimator1, estimator2, X, y, num_rounds=30, test_size=0.3, scoring=None, random_seed=None)

Implements the resampled paired t test procedure to compare the performance of two models (also called k-hold-out paired t test).




For usage examples, please see