Math Inferential Statistics Article

T-Tests: Intro to Key Terms & One Sample t-test

A t-test is a type of statistical hypothesis test. This type of test can determine if 1) a sample mean is significantly different from a known population mean or 2) a sample mean is significantly different from another sample mean. Typically, we associate t-tests with the latter. This post will walk through a detailed example of just 1).

In order to fully grasp the concepts in this post, it'll help to have familiarity with the following concepts: z-tests, hypothesis testing, p-values, normal distribution, standard deviation, standard error and the central limit theorem.

There are three distinct types of t-tests: one sample t-tests, dependent sample t-tests and independent sample t-tests. Each type of test involves a comparison of two groups.

Hypothesis tests for comparing multiple groups are called analysis of variance (ANOVA) and will be covered in another post.

Each of the three types of tests involve the same variable tested in the two groups. For example, a t-test can compare scores on the same final exam for one teacher's students and another teacher's students. However, if one teacher provides an exam with different questions than another teacher, a simple hypothesis test can't adequately compare the students' performance against the other class because there are lurking variables. These are variables that are not explained in the single score of student's performance but would affect the interpretation of the t-tests results. Some examples of lurking variables are difficulty of questions, number of questions on the exam and time allotted to complete the exam.

Comparison: z-tests vs. t-tests

Both z-tests and t-tests are statistical hypothesis tests.

Comparison field z-test t-test
Purpose compares if the means of two datasets are significantly different from each other when population variance is known compares if the means of two sets of data are significantly different when population variance is unknown
distribution used normal student's t
population variance known unknown
sample size large (\(n>30\)) small (\(n<30\))

Overview of Three Types of t-tests

type of test group 1 group 2 purpose
one sample t-test population sample tests whether the mean of a population is significantly different than a sample mean
dependent sample t-test sample sample tests whether the difference in means of two depenendent samples is significantly different
independent sample t-test sample sample tests whether the difference in means of two independent samples is significantly different

Each of these tests is considered a parametric test because they assume that data comes from a population that meets specific probability distribution standards. These data requirements will be discussed individually for each test in separate posts.

One Sample t-test Descriptive Example

There's an academy that teaches students new to music how to play the guitar in just three months. At the end of the three months, students are assessed on their performance of the song Ziggy Stardust. For their performance, a proctor counts the number of errors performed by each student. There have been hundreds of instructors at this academy over the year and they all utilized the same methods for teaching. However, last year a new teacher, Mr. Brown, utilized a new method to teach a random group of 28 students in that 3 month period.

Each year, the academy updates their mean value for the number of errors on student's final performance. However, they don't record all the individual historical student scores for privacy purposes. Therefore, the population variance is unknown.

A calculation was also made for the mean number of errors by Mr. Brown's students. Does Mr. Brown's teaching significantly decrease the number of errors in student's performance compared to the population of students taught at the academy?

\(H_{o}: \mu_{\textrm{Mr. Brown}}>=\mu_{academy}\)

\(H_{a}: \mu_{\textrm{Mr. Brown}}<\mu_{academy}\)

In the hypotheses above, I use the symbol of \(\mu\) (a population mean) instead of \(\bar{x}\) (a sample mean) for Mr. Brown's students. This may seem confusing since we just have a sample of Mr. Brown's students.

These hypotheses are not about the exact provided data. Rather, they're inferences about the expected population means from which the sample comes from compared to the known population mean. A proper t-test analysis performs calculations to help infer what the expected population mean (that contains the sample) given just a sample mean. This description is part of the field of inferential statistics: use of a random sample of data taken from a population to describe and make inferences about the population. These inferences are important because they help to explain what would likely happen if we had access to a sufficiently large population of Mr. Brown's students.

In a one sample t-test like this, you could perform it if Mr. Brown taught at the academy or outside of the academy. However, there's other factors to consider in proper experiment design to ensure the results have merit. For this example, I used the more logical t-test design and included Mr. Brown as one of the many teachers at the academy.

Dependent Sample t-test Descriptive Example

A sample of 19 people were told to lie in bed ready to fall asleep at 10pm every night and would be woken up by an alarm at 6:30am for 15 days in a row. In that period, they tried to sleep as much as possible. With each night's sleep, sensors and a computer helped record a measurement of that person's total hours slept. This same test was later performed for the same group of people under the same circumstances, but with one difference: each night a person took a sleeping pill at 9:30pm.

After all the data was collected, I calculated the mean number of hours slept for people without using the pill and with using the pill. Does the sleeping pill help significantly increase the average hours slept for the group than without the pill?

\(H_{o}:\mu_{pill}>= \mu_{\textrm{no pill}}\)

\(H_{a}:\mu_{pill}>\mu_{\textrm{no pill}}\)

This is considered a pretest posttest type of t-test for dependent samples. The samples are considered dependent on one another because the same people who took the sleep test without the pills later took the test again with sleeping pills.

Typically, in a dependent t-test example such as this, the exact populations data with individual observations is unknown or one is unable to access the population data.

Independent Sample t-test Descriptive Example

There was a sample of 20 professional basketball players in 1940 that attempted to jump as high as they could and measurements were taken on their reach by the top of their hand. The same test was performed on a sample of 20 professional basketball players in 2019. After all this data was collected, I calculated the mean height in inches reached by the sample of players in 1940 and 2019. Do professional basketball players in 2019 jump significantly higher on average than basketball players in 1940?

Similar to the example with dependent t-tests above, these hypotheses are an inference on the expected true population means from which each sample comes from.

\(H_{o}\): \(\mu_{2019}<=\mu_{1940}\)

\(H_{a}\): \(\mu_{2019}>\mu_{1940}\)

Setup Code

Import Modules

from scipy.stats import t
import numpy as np
import scipy.stats as stats
import seaborn as sns
import matplotlib.pyplot as plt
% matplotlib inline

Set Visualization Style


Steps to Perform One Sample t-test

1) Setup the experiment to record measurements from a sample.

2) Set an alpha level for the test, a null hypothesis and alternative hypothesis.

3) Compute the sample mean, sample size and sample standard deviation. - These calculations will be used later in steps 6 and 7.

4) Obtain the population mean for the same variable measured.

5) Calculate the t-critical value.

6) Determine if the data meets the proper requirements to perform a one-sample t-test.

7) Calculate the t-statistic.

8) Compare the t-statistic to the t-critical value. Interpret results of the experiment based on the original hypotheses.

1) Setup Experiment

I'll continue with the descriptive example for the one-sample t-test mentioned above. The setup was designed so the count of errors for all student's performances of Ziggy Stardust by Mr. Brown's students were recorded.

Assign a python list of errors of students in the sample to the variable sample_student_errors.

sample_student_errors = [5, 5, 5, 5, 5, 5, 5, 4, 4, 4, 4, 4, 6, 6, 6, 6, 7, 7, 7, 3, 3, 3, 3, 2, 2, 8, 8, 9]

2) Set an Alpha Level and Original Hypotheses

The alpha level for this experiment will be \(0.05\).

alpha = 0.05

Does Mr. Brown's teaching significantly decrease the number of errors in student's performance compared to the population of students taught at the academy?

\(H_{o}: \mu_{\textrm{Mr. Brown}}>=\mu_{academy}\)

\(H_{a}: \mu_{\textrm{Mr. Brown}}<\mu_{academy}\)

This is a one-sided negative direction t-test.

3) Compute the Sample Mean, Sample Size and Sample Standard Deviation

Assign the variable n to be the size of the sample.

n = len(sample_student_errors)

Assign the variable sample_mean to be the mean of the sample.

sample_mean = np.mean(sample_student_errors)

Assign the variable sample_std_dev to be the standard deviation of the sample.

sample_std_dev = np.std(sample_student_errors, ddof=1)

4) Obtain the Population Mean

I use the Numpy package normal() method in the random class to create a Python list of floats to represent errors of students for the population. Assign this list to the variable population_student_errors since these are the values for the count of errors each student made performing Ziggy Stardust.

I cast all the float values to integer errors since teachers only count singular errors.

np.random.seed(3) # sets the random number generator to return the same results each time
population_student_errors = list(np.random.normal(7.1, 2.1, 1300))
population_student_errors = [int(error) for error in population_student_errors]

Assign the variable population_mean_errors to the mean of population_student_errors. It's ok if Mr. Brown's recent sample of students are utilized in this population.

population_mean_errors = round(np.mean(population_student_errors), 2)

5) Calculate the t-critical Value

Also, similar to the calculation of z-critical values for z-tests, a t-test requires calculation of a t-critical value which just uses the initial alpha level and \(n\) - the sample size.

The t-critical value provides a threshold of statistical unlikeliness on the t-distribution. The t-distribution is a probability distribution used to estimate population parameters in analysis for t-tests.

The math and logic used to infer the t-distribution is long and complex. For now, I'll cover some basics of the t-distribution and chug forward with this t-test analysis.

Visualization of z-distribution and t-distribution

This visualization is based on certain parameters such as the size of samples and/or population. They won't always appear in the exact shapes below but I still think the comparisons are a good general point of reference.

plt.figure(figsize=(11, 9))
z_distribution = np.random.normal(loc=0, size=6000)
t_distribution = np.random.standard_t(df=2, size=6000) # not a normal distribution
sns.distplot(z_distribution, hist=False, label='z-distribution')
sns.distplot(t_distribution, color='red', hist=False, label='t-distribution')
plt.ylabel("frequency of occurence", labelpad=14)
plt.xlabel("t-statistic or z-score", labelpad=14)
plt.title("Distribution of z-distribution and t-distribution", y=1.015, fontsize=20)
plt.xlim(-15, 15)
plt.ylim(-.01, .4)


The t-distribution has the following characteristics:

  • mean = 0
  • fairly symmetrical
  • typically not exactly normal given small \(n\) values; a very high \(n\) value will lead to a closer to normal distribution
  • as \(n\) increases, the sample standard deviation \(s\) approaches the value of the population standard deviation \(\sigma\)
  • \(variance = v/(v-2)\) in which \(v\) is the degrees of freedom and \(v>2\) (degrees of freedom explained below)

T-test errors are more prone to errors in results than a z-test. The reason for this is that compared to the z-distribution, the t-distribution has larger variance. Visually we see a wider curve and thicker tails.

Calculate degrees of freedom

We utilize degrees of freedom in t-test calculations. Degrees of freedom is simply the freedom to vary. We have 28 students in Mr. Brown's class. If I hypothetically picked out grades for 27 of the students, you must pick out the one remaining student's grade. For any 28 student scores I pick, you have \(28-1 = 27\) student's grades to choose from. In variable terms, this means if you have \(n\) numbers to choose from, you have \(n-1\) degrees of freedom.

In statistics, the calculation for degrees of freedom is the count of observations in the sample that are free to vary when estimating statistical parameters.

The t-distribution is defined by the degrees of freedom value and the alpha level.

Let's assign the variable degrees_of_freedom to be the degrees of freedom for students in Mr. Brown's class.

degrees_of_freedom = n-1

Calculate t-critical value

Similarities of a t-critical value to a z-score:

  • identifies a boundary for a critical region of its respective distribution (for a t-test, the t-distribution)
  • for a one-tailed test, there's either a positive or negative critical value with one tail probability
  • for a two-tailed test, there's a positive and negative critical value with their respective tail probabilities

For a t-test, we utilize knowledge of the alpha level, direction of the test and degrees of freedom.

I'm familiar with two ways to calculate the t-critical value. One way is using a table and another is done programatically using Python. I prefer the latter.

Udacity hosts a table to identify t-critical values at this page. The values inside the table represent the t-critical values. Each row shows these t-critical values for a degrees of freedom (df) value. The columns are labeled for a right tail probability. This table can be used for one-tail and two-tail tests.

In Python, as part of the Scipy package and t class, there's a method called ppf() that returns the t-critical value. The arguments of the method are:

  • q for the area under the curve of the t-distribution to the left of the t-critical value
  • df for the degrees of freedom (equivalent to \(n-1\) for a one sample t-test)

Below is the calculation for our guitar example in which we know the alpha value of \(0.05\) is the area under the t-distribution to the left of the t-critical value. We want to assign alpha to q becuase this is a one-sided negative direction test.

t_critical_value = round(stats.t.ppf(q=alpha, df=degrees_of_freedom), 3)

The t-critical value should be negative because this is a one-sided negative direction t-test.

6) Determine if Data Meets Requirements for One Sample t-test

One Sample t-tests Data Requirements

  • Each of the groups, the sample and population, have the same variable that is continous.

  • Population and sample each have an approximate normal distribution.

    • Non-normal distributions, especially those with large variance and/or skewness heavily reduce the power of the results of the t-test.
  • The sample should be randomly selected from a population.

  • Assume sample data is representative of the whole population and can therefore estimate population parameters.

  • Approximately equal variances of the population and sample.

Evaluate Data Requirements for Guitar Example

In this example, I'll assume we don't have access to each of the scores for the whole population of students. The population variance is also unknown.

Two requirements we can test with the data part of 1) sample has an approximate normal population and sample each have an approximate normal distribution.

Let's examine the distribution of the values in the sample with a histogram. I use the Matplotlib hist() method.

plt.figure(figsize=(11, 9))
plt.hist(sample_student_errors, bins=8, color='dodgerblue')
plt.title("Distribution of Errors of Students Taught by Private Instructor", y=1.015, fontsize=20)
plt.xlabel("count of errors", labelpad=14)
plt.ylabel("count of occurences", labelpad=14);


This distribution looks symmetrical, has no notable skew, and is centered around a central value.

sample_median_errors = np.median(sample_student_errors)
sample_mean_errors = np.mean(sample_student_errors)
print("Sample mean of {0} and median of {1}".format(sample_mean_errors, sample_median_errors))
Sample mean of 5.035714285714286 and median of 5.0

Above is the code that shows the sample mean and median are nearly the exact same too. Therefore, it's fairly safe to say that the sample has a normal distribution too.

Calculate the t-statistic

Similar to z-tests in the calculation of a z-score, a t-test requires calculation of a t-statistic. A one-sample t-test helps determine if the mean of a population is significantly different than a sample mean. Both the sample and population have the same variable measurement. Remember, in a one-sample t-test, we don't know the population standard deviation so we cannot use that in the following calculation. The equation for the t-statistic for a one-sample t-test is:

  • \(\bar{x}\) is sample mean and is a point estimate
  • \(\mu\) is our known current population mean
  • \(s\) is the sample standard deviation (remember to use bessel's correction with \(n-1\) in the denominator)
  • \(n\) is the sample size
  • \(\frac{s}{\sqrt{n}}\) is our expected standard deviation from a sampling distribution of means (term also called standard error of the means)

The t-statistic tells us how far our sample mean lies from the population mean in terms of standard error units on a sampling distribution of sample means. This sampling distribution comes from taking a sufficiently large number of size \(n\) samples from the population data. On the sampling distribution, both \(\bar{x}\) and \(\mu\) are on the same "scale" of \(n\) so it's OK to compare them. An anology that makes sense to me for this t-statistic is compared the quality of two apples of the same size from two different farms. This is an easy comparison because we're just comparing the quality. If we simply just looked at the difference between \(\bar{x}\) and \(\mu\) that would be comparing the quality of two apples of different sizes and different qualities from two different farms - a tough comparison to make sense of because we can't simply isolate just the quality in a bite of each.

The t-statstic equation above has a numerator and denominator that can each be made sense of.

Remember the t-critical value is \(-1.703\). The smaller the t-statistic, the more evidence we obtain to reject the null hypothesis and conclude Mr. Brown's students have significantly fewer errors than the general population.

The numerator measures the "strength" of the signal: the difference between the sample mean and population mean. In the guitar example, if Mr. Brown's students have far fewer errors than the general population, we'd expect a smaller numerator value and therefore more "strength". The difference:

numerator = sample_mean-population_mean_errors

This value of \(-1.52\) seems to indicate Mr. Brown's students do perform with fewer errors on average than the general population.

The denominator measures the variation or "noise" in the sample data. If hypothetically the errors of all \(4\) students in Mr. Brown's class were \(1\), \(5\), \(9\), and \(18\) - that would be a lot of variation. This would equate so a high \(s\) value. If hypothetically the errors of all \(4\) students in Mr. Brown's class were \(1\), \(2\), \(3\), and \(5\) - that would be less variation and result in a small \(s\) value.

The amount of variation (value in the denominator) is "noisier" (a larger value) - for a smaller \(n\) value. The demominator is also called the standard error.

standard_error = sample_std_dev/np.sqrt(n)
t_statistic = numerator/standard_error

In the Python Scipy package, there's a method called ttest_1samp() that calculates the t-statistic and p-value for a one sample two-sided t-test. However, the guitar errors example is a one-sided negative direction test. Let's use the returned p-value from this method and multiply it by \(2\) to get the p-value for a one-sided test.

results = stats.ttest_1samp(a=sample_student_errors, popmean=population_mean_errors)
Ttest_1sampResult(statistic=-4.443441113387382, pvalue=0.00013578924891219084)
two_sided_p_value = results[1]
one_sided_p_value = round(two_sided_p_value*2, 5)

This returned t-statistic from the manual calculation above is the same value returned from use of the ttest_1samp method.

Visualization of the t-distribution and a t-statistic

plt.figure(figsize=(10, 8))
t_distribution_df_28 = np.random.standard_t(n, size=3000)
sns.distplot(t_distribution_df_28, color='blue', hist=False, label='t-distribution of $n=28$')
plt.axvline(x=t_statistic, color='darkslateblue', linestyle='--', label='t-statistic of $-4.44$')
plt.xlabel("t-values", labelpad=14)
plt.ylabel("frequency of occurence", labelpad=14);
plt.title("t-distribution and location of t-statistic", y=1.015, fontsize=20)
plt.ylim(-.01, .4)
plt.xlim(-6, 6)


8) Interpret Results of Experiment

Conditions for Rejecting the Null Hypothesis

Similar to z-tests, with each t-test hypothesis, we want to determine if there is sufficient evidence to reject the null hypothesis. Below are details of hypothesis tests and means for evaluating one-sample t-tests.

null hypothesis alternative hypothesis alternative hypothesis interpretation number of tails on t-distribution tail(s) of t-distribution to reject the null
\(\bar{x}>=\mu\) \(\bar{x}<\mu\) the sample mean is less than the population mean one left
\(\bar{x}<=\mu\) \(\bar{x}>\mu\) the sample mean is greater than the population mean one right
\(\bar{x}=\mu\) \(\bar{x}\neq\mu\) no prediction on a direction for the difference between the sample mean and population mean two left or right

Here's what it means to reject the null hypothesis:

  • The t-statistic falls within the t-critical region (bounded by the t-critical value)
    • For a one-tailed test in the positive direction, the t-statistic is greater than the t-critical value
    • For a one-tailed test in the negative direction, the t-statistic is less than the t-critical value
    • For a two-tailed test, the t-statistic is greater than the positive t-critical value or the t-statistic is less than the negative t-critical value
  • The probability of obtaining the estimated population parameter due to random chance (the p-value) is less than the alpha level

Final Conclusion for Guitar Example

Given the table above, our guitar example is similar to the first row.

Ths t-statistic of \(-4.44\) is less than the t-critical value of \(-1.703\)

Also, the p-value of \(0.00027\) is less than the initial alpha level of \(0.05\). This p-value tells us the probability of getting this t-statistic if the null hypothesis is true.

Therefore, there is sufficient evidence to reject the null hypothesis. Mr. Brown's students have significantly fewer errors in playing Ziggy Stardust than the general population of students who learned at the academy.

Calculate a 95% Confidence Interval

This step is not needed for a t-test, but it offers additioanl valuable insights to our interpretation of the results.

A confidence interval provides a range of values in which there's a specified probability that the expected true population mean lies.

We performed a t-test with a sample of students from Mr. Brown and would like to estimate the population parameters for all students that would go through the same process of instruction by Mr. Brown.

Essentially, if the private teacher were to continually teach groups of 25 beginner students under the same circumstances as in the initial context, we'd expect 95% of those sample means to lie within what range?

We can express the formula for the confidence interval for the expected population of mean errors by students for their performance of Ziggy Stardust after instruction by Mr. Brown as:

$$(\bar{x}_{\textrm{Mr. Brown}}+\textrm{t-critical value}*SE,\:\bar{x}_{\textrm{Mr. Brown}}-\textrm{t-critical value}*SE)$$

\(\textrm{t-critical value}*SE\) is the margin of error and is 1/2 the width of our confidence interval. The margin of error is the number of standard errors away from the mean.

Previously, we wanted a tail probability in the t-distribution of 0.05 but since we now want a 95% confidence interval, we'd like 0.025 probability in the tails. Therefore, we need a new t-critical value.

# retrieve value <= probability
tail_probability = 0.025
t_critical_value_for_95_CI =round(t.ppf(q=tail_probability, df=degrees_of_freedom), 2)
print("t-critical value is {0} for tail probability of {1} and degrees of freedom of {2}".format(t_critical_value_for_95_CI, tail_probability, degrees_of_freedom))
t-critical value is -2.05 for tail probability of 0.025 and degrees of freedom of 27

Let's calculate the 95% confidence interval using the equation above.

mean_lower_95_CI_bound = round(sample_mean + t_critical_value_for_95_CI*standard_error, 3)
mean_upper_95_CI_bound = round(sample_mean - t_critical_value_for_95_CI*standard_error, 3)
print("95% CI: ({0}, {1})".format(mean_lower_95_CI_bound, mean_upper_95_CI_bound))
95% CI: (4.332, 5.739)

Let's visualize the bounds of the 95% confidence interval and indicate the two distributions at each bound.

sampling_distribution_theoretical_values_lower = np.random.normal(loc=mean_lower_95_CI_bound, scale=standard_error, size=5000)
sampling_distribution_theoretical_values_upper = np.random.normal(loc=mean_upper_95_CI_bound, scale=standard_error, size=5000)
plt.figure(figsize=(12, 9))
sns.distplot(sampling_distribution_theoretical_values_lower, color="g", hist=False)
sns.distplot(sampling_distribution_theoretical_values_upper, color="orange", hist=False)
plt.axvspan(mean_lower_95_CI_bound, mean_upper_95_CI_bound, alpha=0.5, color='lightsteelblue')
bbox_props = dict(boxstyle="round", fc="w", ec="0.2", alpha=0.9)
plt.text(sample_mean, 0.4, "Shaded area represents range\nof expected population\nmean ($\mu_{Brown}$) - equivalent\nto interval estimate of 95% CI", ha="center", 
         va="center", size=16, bbox=bbox_props)
plt.axvline(x=sample_mean, linestyle='--', linewidth=2.5, label="sample mean private instruction (point estimate)", c='purple')
plt.axvline(x=mean_lower_95_CI_bound, linestyle='--', linewidth=2, c='g', label='lower 95% CI bound')
plt.axvline(x=mean_upper_95_CI_bound, linestyle='--', linewidth=2, c='orange', label='upper 95% CI bound')
plt.xlabel("average errors", labelpad=14)
plt.ylabel("probability of occurence", labelpad=14)
plt.title("95% CI for Bounds of Population for Errors of Students Playing Song Taught by Private Instructor", y=1.015)
plt.legend(bbox_to_anchor=(1.01, 1), loc="upper left");


Visualization of Change in t-critical Value with Change in Degrees of Freedom

For an alpha level, as the degrees of freedom increases past \(22\), there is little change in the corresponding t-critical value.

Thus, if I wanted to run a one sample t-test, I would likely just get a sample of roughly size \(23\) which equates to a degree of freedom of \(22\). I'd only opt to retrive a larger sample if retrieving that data was easy.

Below is a visualization that shows the change in t-critical value as the sample size increases for a one sample t-test in the positive direction with an alpha level of \(0.05\).

t_critical_values = []
degree_of_freedom_values = list(range(4, 200))
for df_value in degree_of_freedom_values:
    t_critical_value = t.ppf(q=0.95, df=df_value)
plt.figure(figsize=(16, 11))
plt.plot(degree_of_freedom_values, t_critical_values, color='dodgerblue')
plt.axvline(x=22, label='df = 22', color='darkslategray', linestyle='--')
plt.text(x=70, y=1.85, horizontalalignment='center', fontsize=17, 
         s="Minimal change in t-critical values\nwith increasing degrees of freedom for df>22", 
         bbox=dict(facecolor='whitesmoke', boxstyle="round, pad=0.25"))
plt.xlabel("degrees of freedom [df]", labelpad=15)
plt.ylabel("t-critical value", labelpad=15)
plt.title("Change in T-Critical Values with Change in Sample Size for Alpha Level $0.05$", y=1.01, fontsize=21)
plt.ylim(0, 2.2);


The t-critical value barely decreases with increasing degrees of freedom past \(22\).