Independent T-test

The indepentent T-test is a parametric test used to test for a statistically significant difference in the means between 2 groups. As with all parametric tests, there are certain conditions that need to be met in order for the test results to be considered reliable.

Fear not if math is not your strong suit. All this is being calucated when using the methods of a statistical software or programming language. It's good to know what is going on behind the scences. References for this section are provided at the end of the page.

T-test with Python

Don't forget to check the assumptions before interpreting the results! First to load the libraries needed. This demonstration will include 2 ways to conduct an independent sample t-test in Python. One with Researchpy and the other with Scipy.stats.

import pandas as pd
import researchpy as rp
import scipy.stats as stats

Now to load the data set and take a high level look at the variables.

df = pd.read_csv("https://raw.githubusercontent.com/researchpy/Data-sets/master/blood_pressure.csv")
df.info()

Indepdent t-test using Researchpy

The method returns 2 data frames, one that contains the summary statistical information and the other that contains the statistical test information. If the returned data frames are not stored as a Python object then the output will be less clean than it can be since it will be displayed as a tuple - see below.

rp.ttest(group1= df['bp_after'][df['sex'] == 'Male'], group1_name= "Male",
         group2= df['bp_after'][df['sex'] == 'Female'], group2_name= "Female")

If stored as Python objects, they will be stored as Pandas data frames. This means that anything that can be done to a traditional Pandas data frame can be done to these results.

summary, results = rp.ttest(group1= df['bp_after'][df['sex'] == 'Male'], group1_name= "Male",
                            group2= df['bp_after'][df['sex'] == 'Female'], group2_name= "Female")
print(summary)

print(results)

Before the results should be interpreted, the assumptions of the test should be checked. For example purposes, the results will be interpreted before checking the assumptions.

Interpretation

Indepdent t-test using Scipy.stats

This method conducts the independent sample t-test and returns only the t test statistic and it's associated p-value. For more information about this method, please refer to the official documentation page.

stats.ttest_ind(df['bp_after'][df['sex'] == 'Male'],
                df['bp_after'][df['sex'] == 'Female'])

Interpretation

Assumption check

The assumptions in this section need to be met in order for the test results to be considered valid. A more in-depth look at parametric assumptions is provided here, which includes some potential remedies.

The two samples are independent

This assumption is tested when the study is designed. What this means is that no individual has data in group A and B.

Population distributions are normal

One of the assumptions is that the sampling distribution is normally distributed. This test of normality applies to the difference in values between the groups. One method for testing this assumption is the Shapiro-Wilk test. This can be completed using the shapiro() method from Scipy.stats.

sampling_difference = df['bp_after'][df['sex'] == 'Male'].values - \
                      df['bp_after'][df['sex'] == 'Female'].values

stats.shapiro(sampling_difference)

Unfortunately, the output is not labelled but is in the format of (W test statistic, p-value). The test is not significant which indicates the sampling distribution is normally distributed.

Another way to test the assumption is through a visual check- this is helpful when the sample is large. The reason this is true is that as the sample size increases, the statistical test's ability to reject the null hypothesis increases, i.e. it gains power to detect smaller differences as the sample size n increases.

One method of visually checking the distribution is to use a probability plot with or without the correlation value, $R^2$, to assess the observed values correlation with the theoretical distribution in question - in the current case it would be the Gaussian (a.k.a the normal) distribution. This can be completed by using the probplot() method from Scipy.stats.

fig = plt.figure(figsize= (20, 10))
ax = fig.add_subplot(111)

normality_plot, stat = stats.probplot(sampling_difference, plot= plt, rvalue= True)
ax.set_title("Probability plot of sampling difference", fontsize= 20)
ax.set

plt.show()

Both methods support the same finding which is that the sampling distribution is normally distributed.

Homogeneity of variance

One of the assumptions is that both groups have equal variances. One method for testing this assumption is the Levene's test of homogeneity of variances. This can be completed using the levene() method from Scipy.stats.

stats.levene(df['bp_after'][df['sex'] == 'Male'],
             df['bp_after'][df['sex'] == 'Female'],
             center= 'mean')

The test is significant which indicates the groups have a different amount of variation and that the t-test may not be the best statistical method to be used. Again, it may be worthwhile to check this assumption visually as well.

fig = plt.figure(figsize= (20, 10))
ax = fig.add_subplot(111)


p_bp_male = plt.hist(df['bp_after'][df['sex'] == 'Male'], label= "Male",
                     density= True,
                     alpha=0.75)
p_bp_female = plt.hist(df['bp_after'][df['sex'] == 'Female'], label= "Female",
                       density= True,
                       alpha=0.75)


plt.suptitle("Distribution of Post-Blood Pressure \n between Males and Females", fontsize= 20)
plt.xlabel("Blood Pressure", fontsize= 16)
plt.ylabel("Probability density", fontsize= 16)

plt.text(133, .025,
         f"$\mu= {df['bp_after'][df['sex'] == 'Female'].mean(): .1f}, \ \sigma= {df['bp_after'][df['sex'] == 'Female'].std(): .1f}$")
plt.text(160, .025,
         f"$\mu= {df['bp_after'][df['sex'] == 'Male'].mean(): .1f}, \ \sigma= {df['bp_after'][df['sex'] == 'Male'].std(): .1f}$")


plt.show()

There are different ways to handle heteroskedasticity (unequal variance) and a decision needs to be made. Some options include, but is not limited to, transformming the dependent variable (outcome), could use trimmed means, robust standard errors, or use a parametric test suchs as the Welch's t-test.