Deriving the One-Sample T Statistic

Why Standardize a Difference in Means?

A sample mean will usually differ from a hypothesized population mean, even when that hypothesized value is correct. The size of the difference alone does not tell us how surprising it is: a difference of 2 units may be substantial when sample means are precise, but modest when they vary considerably from sample to sample. A test statistic puts the observed difference on a scale that accounts for its sampling variability.

For a one-sample test, the quantity of interest is a population mean $\mu$. The null hypothesis specifies a value $\mu_0$, so the difference we observe is $\bar{x}-\mu_0$. The one-sample t statistic expresses that difference in estimated standard errors of the sample mean.

Definition: For a sample of $n$ quantitative observations with sample mean $\bar{x}$ and sample standard deviation $s$, the estimated standard error of the sample mean is $s/\sqrt{n}$. For a null hypothesis $H_0:\mu=\mu_0$, the one-sample t statistic is $$ t=\frac{\bar{x}-\mu_0}{s/\sqrt{n}}. $$ It is the observed difference between the sample mean and the null value divided by the estimated standard error of the sample mean.

From Sampling Variability to the Statistic

The derivation begins with the variability of a sample mean. Suppose $X_1,\ldots,X_n$ are independent observations from a population with mean $\mu$ and standard deviation $\sigma$. The sample mean is $ \bar{X}=(X_1+\cdots+X_n)/n$. Independence means the variance of a sum is the sum of the variances, so

\operatorname{Var}(\bar{X}) =\operatorname{Var}\left(\frac{X_1+\cdots+X_n}{n}\right) =\frac{1}{n^2}\left(n\sigma^2\right) =\frac{\sigma^2}{n}.

Taking the square root gives the standard deviation of the sampling distribution of $\bar{X}$, also called its standard error when the population parameters are specified:

SE(\bar{X})=\frac{\sigma}{\sqrt{n}}.

Under $H_0:\mu=\mu_0$, the sample mean is centered at $\mu_0$. If $\sigma$ were known, we could standardize the difference using the true standard error, $\sigma/\sqrt{n}$. In practice, the population standard deviation is generally unknown. We estimate it with the sample standard deviation $s$, calculated as

s=\sqrt{\frac{\sum_{i=1}^{n}(x_i-\bar{x})^2}{n-1}}.

Replacing $\sigma$ by $s$ in the standard error gives the estimated standard error $s/\sqrt{n}$. Dividing the observed difference by this estimate produces the one-sample t statistic:

\text{estimated standard error}=\frac{s}{\sqrt{n}}, \qquad t=\frac{\text{observed difference from the null value}} {\text{estimated standard error}} =\frac{\bar{x}-\mu_0}{s/\sqrt{n}}.

This is the key derivation: the numerator measures how far the sample mean is from the value specified by the null hypothesis, and the denominator measures the typical sampling variation in a sample mean, estimated from the data. The resulting statistic is dimensionless. For example, if the measurements are in kilograms, both the numerator and denominator are in kilograms, so the units cancel.

Model and Reference Distribution

The formula can be calculated for many samples, but its exact t reference distribution requires a model. For the exact result, assume the observations are independent random observations from a normal population with mean $\mu$ and standard deviation $\sigma$. Under the null hypothesis $H_0:\mu=\mu_0$, the one-sample statistic follows a Student t distribution with $n-1$ degrees of freedom. The t distribution accounts for the extra uncertainty from estimating $\sigma$ with $s$. Its precise degrees-of-freedom index is part of identifying the reference distribution; the statistic’s formula itself comes from standardizing the difference by the estimated standard error.

Theorem: If $X_1,\ldots,X_n$ are independent observations from a normal population and $H_0:\mu=\mu_0$ is true, then $$ T=\frac{\bar{X}-\mu_0}{S/\sqrt{n}} $$ has a Student t distribution with $n-1$ degrees of freedom. The observed statistic $t=(\bar{x}-\mu_0)/(s/\sqrt{n})$ is the value of $T$ calculated from the sample.

This distributional statement and the definition of the statistic are related but distinct. The formula tells us how to calculate the observed value. The model and null hypothesis tell us which distribution is used to assess that value as evidence. The examples below focus on deriving and calculating the statistic, not on finding a p-value or making a final test decision.

Worked Example: Calculating the Statistic From Raw Data

In a hypothetical exercise, five independent measurements, in centimeters, are 18, 20, 21, 22, and 24. Suppose the null hypothesis specifies $\mu_0=20$ cm. For this example, assume the observations are a random sample from a normal population.

First calculate the sample mean:

\bar{x}=\frac{18+20+21+22+24}{5} =\frac{105}{5}=21\text{ cm}.

The deviations from the mean are $-3,-1,0,1,3$ cm, whose squared deviations sum to $9+1+0+1+9=20\text{ cm}^2$. Thus,

s=\sqrt{\frac{20}{5-1}} =\sqrt{5} \approx 2.236\text{ cm}, \qquad \frac{s}{\sqrt{n}}=\frac{2.236}{\sqrt{5}}=1.000\text{ cm}.

The sample mean is $1$ cm above the hypothesized mean. Standardizing that difference gives

t=\frac{21-20}{2.236/\sqrt{5}} =\frac{1}{1}=1.00.

The observed sample mean is one estimated standard error above the null value. The independent random sampling and normal-population assumptions were stipulated in the hypothetical setup, so the exact t reference distribution applies.

How to Read the Components

Each part of the statistic has a specific role. The numerator $\bar{x}-\mu_0$ is a signed difference: it is positive when the sample mean is above the null value and negative when it is below. The denominator $s/\sqrt{n}$ is positive when the sample has nonzero variation. It estimates how much sample means would typically vary under repeated sampling, at the sample size used.

The statistic is therefore not the difference in the original measurement units. It is a difference measured in estimated standard-error units. A value $t=2$, for instance, means the sample mean is two estimated standard errors above $\mu_0$; $t=-2$ means it is two estimated standard errors below. Neither value, by itself, is a probability or a conclusion about the null hypothesis.

The sample size affects the denominator through $\sqrt{n}$. Holding $s$ and the observed difference fixed, increasing $n$ makes the estimated standard error smaller and the absolute value of $t$ larger. This reflects greater precision in the sample mean, not a change to the units or definition of the mean itself.

Worked Example: Using Summary Statistics

Suppose a hypothetical random sample of $n=16$ independent, normally distributed measurements of a concentration has $\bar{x}=74.5$ milligrams per liter and $s=8.0$ milligrams per liter. The null value is $\mu_0=70.0$ milligrams per liter.

The estimated standard error is

\frac{s}{\sqrt{n}}=\frac{8.0}{\sqrt{16}} =\frac{8.0}{4}=2.0\text{ mg/L}.

The sample mean is $74.5-70.0=4.5$ mg/L above the null value. Therefore,

t=\frac{74.5-70.0}{8.0/\sqrt{16}} =\frac{4.5}{2.0}=2.25.

The observed difference is 2.25 estimated standard errors above the null value. This calculation uses the stated summary statistics; it does not require the original 16 measurements. The assumptions for the exact t model are stipulated here: independent random observations and a normal population.

Sign, Scale, and Sample Size

The sign of $t$ preserves the direction of the difference. If $\bar{x}<\mu_0$, the numerator and statistic are negative; if $\bar{x}>\mu_0$, they are positive. The size $|t|$ describes the distance from the null value in estimated standard-error units. These features make the statistic useful for comparing a sample mean with a specified target, while retaining information about which side of the target the sample mean lies on.

Changing the measurement scale does not change the statistic, provided the data and null value are converted consistently. For example, multiplying every observation and $\mu_0$ by a positive conversion factor multiplies both the difference and $s/\sqrt{n}$ by that factor. The ratio is unchanged. This is another way to see that $t$ has no measurement units.

Worked Example: A Sample Mean Below the Null Value

In a hypothetical setting, a random sample of $25$ independent, normally distributed times has sample mean $\bar{x}=4.6$ minutes and sample standard deviation $s=1.5$ minutes. The null mean is $\mu_0=5.0$ minutes.

The estimated standard error is

\frac{s}{\sqrt{n}}=\frac{1.5}{\sqrt{25}} =\frac{1.5}{5}=0.30\text{ minutes}.

Since the sample mean is $4.6-5.0=-0.4$ minutes below the null value, the statistic is

t=\frac{4.6-5.0}{1.5/\sqrt{25}} =\frac{-0.4}{0.30} \approx -1.33.

The negative sign indicates that the sample mean is below the null value; its magnitude indicates a distance of about 1.33 estimated standard errors. The sample size, independence, and normal-population model are given as part of this hypothetical setup.

Why the Estimated Standard Error Matters

Using $s/\sqrt{n}$ rather than $s$ is essential. The sample standard deviation describes variation among individual observations. The standard error describes variation in the sample mean across repeated samples. These are different quantities: as $n$ increases, the spread among individual observations need not shrink, but the sample mean generally becomes more precise.

There is also a reason the statistic is called a t statistic rather than a z statistic. If $\sigma$ were known, standardizing by $\sigma/\sqrt{n}$ would use the true standard error. In the usual one-sample mean problem, $\sigma$ is unknown, so $s$ must be used instead. That substitution adds uncertainty, which is reflected in the t reference distribution. Treating the estimated standard error as though it were known exactly would overlook this feature.

Key takeaway: The one-sample t statistic is obtained by taking the sample mean’s signed difference from the null value and dividing by the estimated standard error $s/\sqrt{n}$. It reports the difference in estimated standard-error units; its reference distribution depends on the sampling model.

Common Mistake

A common mistake is to divide $\bar{x}-\mu_0$ by $s$ and call the result the one-sample t statistic. That uses the spread of individual observations as the denominator, rather than the estimated standard error of the sample mean. The correction is to divide $s$ by $\sqrt{n}$ first:

t=\frac{\bar{x}-\mu_0}{s/\sqrt{n}}, \quad\text{not}\quad \frac{\bar{x}-\mu_0}{s}.

The second expression is a different standardization and does not account for the sample size in the way required for the one-sample mean test. Another mistake is to interpret $t$ itself as a p-value. The statistic is calculated from the data; a p-value requires comparing it with an appropriate reference distribution under the null model.

Check Your Understanding

Use the definitions and derivation above to answer the following questions.

In one sentence, explain why the one-sample t statistic divides $\bar{x}-\mu_0$ by $s/\sqrt{n}$, rather than by $s$.
A sample has $n=9$, $\bar{x}=12$, $s=3$, and null value $\mu_0=10$. Calculate the estimated standard error and the t statistic.
What does a negative value of $t$ say about the sample mean relative to the hypothesized mean?
State the sampling assumptions for which the one-sample statistic has an exact Student t reference distribution.
If all observations and the null value are converted from one unit to another by multiplying by the same positive factor, what happens to $t$? Explain briefly.