knitr::opts_chunk$set( collapse = TRUE, comment = "#>", message = FALSE, warning = FALSE )
In this vignette, we work through a one-sample T-test. We'll use the same data we used in vignette("one-sample-z-test")
, so you compare the result of the tests.
Let's suppose that a student is interesting in estimating how many memes their professors know and love. So they go to class, and every time a professor uses a new meme, they write it down. After a year of classes, the student has recorded the following meme counts, where each count corresponds to a single class they took:
[ 3, 7, 11, 0, 7, 0, 4, 5, 6, 2 ]
Note: For a Z-test, we need to know the population standard deviation $\sigma$. With T-tests, this is unnecessary, and we estimate the standard deviation from the data. This results in some additional uncertainty in our test statistic, which is reflected in the heavier tails of the T distribution compared to the normal distribution.
Before we can do a T-test, we need to make check if we can reasonably treat the mean of this sample as normally distributed. This happens is the case of either of following hold:
Since we have a small sample, we let's check if the data comes from a normal distribution using a normal quantile-quantile plot.
# read in the data x <- c(3, 7, 11, 0, 7, 0, 4, 5, 6, 2) # make the qqplot qqnorm(x) qqline(x)
Since the data lies close the line $y = x$, and has no notable systematic deviations from line, it's safe to treat the sample as coming from a normal distribution. We can proceed with our hypothesis test.
Let's test the null hypothesis that, on average, professors know 3 memes. That is
[ H_0: \mu = 3 \qquad H_A: \mu \neq 3 ]
First we need to calculate our T-statistic. Let's use do this with R. Remember that the T-statistic is defined as
[ T = \frac{\bar x - \mu_0}{s / \sqrt{n}} \sim t_{n-1} ]
where $\bar x = \frac 1n \sum_{i=1}^n x_i$ is the sample mean, $\mu_0$ is our proposed value for the population mean, $s = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (x_i - \bar x)^2}$ is the sample standard deviation, and $n$ is the sample size. This test statistic then has a T distribution with $n-1$ degrees of freedom.
In R this looks like:
n <- length(x) # calculate the z-statistic t_stat <- (mean(x) - 3) / (sd(x) / sqrt(n)) t_stat
Since the sample size is 10, to calculate a two-sided p-value, we need to find
[ \begin{align} P(|t_9| \ge |1.38|) &= P(t_9 \ge 1.38) + P(t_9 \le -1.38) \ &= 1 - P(t_9 \le 1.38) + P(t_9 \le -1.38) \ \end{align} ]
To do this we need to c.d.f. of a $t_9$ distribution.
library(distributions3) T_9 <- StudentsT(df = 9) # make a T_9 distribution 1 - cdf(T_9, 1.38) + cdf(T_9, -1.38)
Note that we saved t_stat
above so we could have also done
1 - cdf(T_9, abs(t_stat)) + cdf(T_9, -abs(t_stat))
which is slightly more accurate since there is no rounding error.
So our p-value is about 0.20. You should verify this with a T-table.
Note that you should get the same value from cdf(T_9, 1.38)
and looking
up 1.38
on a T-table.
You may also have seen a different formula for the p-value of a two-sided T-test, which makes use of the fact that the T distribution is symmetric:
[ \begin{align} P(|t_9| \ge |1.38|) &= 2 \cdot P(t_9 \le -|1.38|) \end{align} ]
Using this formula we get the same result:
2 * cdf(T_9, -1.38)
Finally, sometimes we are interest in one sided T-tests. For the test
[ \begin{align} H_0: \mu \le 3 \qquad H_A: \mu > 3 \end{align} ]
the p-value is given by
[ P(t_9 > 1.38) ]
which we calculate with
1 - cdf(T_9, 1.38)
For the test
[ H_0: \mu \ge 3 \qquad H_A: \mu < 3 ]
the p-value is given by
[ P(t_9 < 1.38) ]
which we calculate with
cdf(T_9, 1.38)
If you want to verify that your calculation is correct, R has a function t.test()
that performs T-tests and calculates T confidence intervals for means.
To get a T statistic, degrees of freedom of the sampling distribution, and the p-value we pass t.test()
a vector of data. We tell t.test()
our null hypothesis by passing the mu
argument. In our case, we want to test mu = 3
.
t.test(x, mu = 3)
If we don't specify mu
, t.test()
assumes that we want to test the hypothesis $H_0 : \mu = 0$. This looks like
t.test(x)
We can also get one-sided p-values from t.test()
by specifying the alternative hypothesis. For the test
[ \begin{align} H_0: \mu \le 3 \qquad H_A: \mu > 3 \end{align} ]
we would use
t.test(x, mu = 3, alternative = "greater")
For the test
[ H_0: \mu \ge 3 \qquad H_A: \mu < 3 ]
we would use
t.test(x, mu = 3, alternative = "less")
Both of these results agree with our hand calculations from earlier.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.