Why Studentizing the Sample Mean Produces a t Variable

Why Replace the Unknown Standard Deviation?

For an independent sample from a normal population, the sample mean has a normal distribution. If the population standard deviation $\sigma$ is known, the mean can be standardized using $\sigma/\sqrt{n}$. In many applications, however, $\sigma$ is unknown. Replacing it with the sample standard deviation $S$ changes the distribution of the standardized statistic: it is Student t rather than standard normal.

The reason is not merely that $S$ is an estimate. For a normal sample, the sample mean and sample variance have a particular exact relationship: the standardized mean is independent of the appropriately scaled sample variance, and that scaled variance has a chi-square distribution. Combining those facts produces the t distribution.

Model and notation: Let $X_1,\ldots,X_n$ be an independent random sample from $N(\mu,\sigma^2)$, where $\sigma^2>0$ and $n\geq 2$. Define the sample mean and unbiased sample variance by $$ \overline{X}=\frac{1}{n}\sum_{i=1}^n X_i,\qquad S^2=\frac{1}{n-1}\sum_{i=1}^n(X_i-\overline{X})^2. $$ The population parameters $\mu$ and $\sigma$ are fixed; the sample statistics $\overline{X}$ and $S$ are random.

The exact result requires an independent sample from a normal population. It is not an exact finite-sample result for every population distribution. Approximation questions for other models require separate justification.

The Three Ingredients

The first ingredient is the normal sampling distribution of the mean. Since the observations are independent and each has mean $\mu$ and variance $\sigma^2$, their average has mean $\mu$ and variance $\sigma^2/n$:

\overline{X}\sim N\left(\mu,\frac{\sigma^2}{n}\right), \qquad Z=\frac{\overline{X}-\mu}{\sigma/\sqrt{n}}\sim N(0,1).

The second ingredient is the sample-variance result for a normal sample. With the denominator $n-1$, the scaled sample variance has a chi-square distribution with $n-1$ degrees of freedom:

U=\frac{(n-1)S^2}{\sigma^2}\sim\chi^2_{n-1}.

The third ingredient is essential: for a normal sample, $\overline{X}$ and $S^2$ are independent. Thus $Z$ and $U$, which are functions of those respective statistics, are independent as well. The independence is special to the normal model; knowing the two marginal distributions alone would not establish the t result.

Definition: If $Z\sim N(0,1)$ and $U\sim\chi^2_\nu$ are independent, then $$ T=\frac{Z}{\sqrt{U/\nu}} $$ has a Student t distribution with $\nu$ degrees of freedom, written $T\sim t_\nu$.

Why the Mean and Variance Are Independent

A short geometric argument explains the independence. Center and scale each observation, setting $W_i=(X_i-\mu)/\sigma$. The vector of these values has independent standard normal components. Apply an orthogonal transformation whose first coordinate is $$ Y_0=\frac{1}{\sqrt{n}}\sum_{i=1}^n W_i =\frac{\sqrt{n}(\overline{X}-\mu)}{\sigma}. $$ An orthogonal transformation preserves the sum of squares. It also preserves the standard multivariate normal density, which is proportional to the exponential of minus one-half the sum of squared coordinates. Consequently, the transformed coordinates $Y_0,Y_1,\ldots,Y_{n-1}$ are independent standard normal variables.

The first transformed coordinate contains the sample mean. The remaining coordinates describe deviations from that mean. Preservation of the sum of squares gives $$ \sum_{j=1}^{n-1}Y_j^2 =\sum_{i=1}^{n}W_i^2-Y_0^2 =\frac{1}{\sigma^2}\sum_{i=1}^{n}(X_i-\overline{X})^2 =\frac{(n-1)S^2}{\sigma^2}. $$ The left side is a sum of squares of $n-1$ independent standard normal variables, so it has a $\chi^2_{n-1}$ distribution. It is independent of $Y_0$, because $Y_0$ is independent of all the other transformed coordinates. This establishes both the sample-variance chi-square law and its independence from the sample mean.

Deriving the Studentized-Mean Distribution

When $\sigma$ is unknown, studentizing means dividing the difference between the sample mean and the population mean by its estimated standard error $S/\sqrt{n}$. Substitute the definitions of $Z$ and $U$ into this ratio:

\begin{aligned} \frac{\overline{X}-\mu}{S/\sqrt{n}} &=\frac{\sqrt{n}(\overline{X}-\mu)/\sigma}{S/\sigma}\\ &=\frac{Z}{\sqrt{S^2/\sigma^2}}\\ &=\frac{Z}{\sqrt{\{(n-1)S^2/\sigma^2\}/(n-1)}}\\ &=\frac{Z}{\sqrt{U/(n-1)}}. \end{aligned}

The numerator is standard normal, the quantity $U$ in the denominator is chi-square with $n-1$ degrees of freedom, and they are independent. By the definition of the Student t distribution, the ratio has a $t_{n-1}$ distribution.

Theorem: For an independent sample $X_1,\ldots,X_n\sim N(\mu,\sigma^2)$, using the sample variance with denominator $n-1$, $$ \frac{\overline{X}-\mu}{S/\sqrt{n}}\sim t_{n-1}. $$ The degrees of freedom are $n-1$ because the sample deviations from $\overline{X}$ satisfy $\sum_{i=1}^n(X_i-\overline{X})=0$, leaving $n-1$ independent normal components in their squared sum.

Worked Examples

Worked Example: Studentizing a Mean from Raw Data

A hypothetical laboratory records five independent turnaround times, in hours: $18,20,20,22,20$. Assume these times are a random sample from $N(\mu,\sigma^2)$. Calculate the studentized statistic for testing the reference value $\mu_0=19$, and identify its null distribution.

Method. Calculate the sample mean and unbiased sample standard deviation, then use $t_{\mathrm{obs}}=(\overline{x}-\mu_0)/(s/\sqrt{n})$. Under the null hypothesis and the stated normal-sample model, the statistic has $n-1$ degrees of freedom.

The sample mean is $$ \overline{x}=\frac{18+20+20+22+20}{5}=\frac{100}{5}=20. $$ The deviations from 20 are $-2,0,0,2,0$, so the sum of their squares is $4+0+0+4+0=8$. Thus $$ s^2=\frac{8}{5-1}=2,\qquad s=\sqrt{2}. $$ The estimated standard error is $s/\sqrt{5}=\sqrt{2/5}$. Therefore, $$ t_{\mathrm{obs}}=\frac{20-19}{\sqrt{2/5}} =\sqrt{\frac{5}{2}}\approx 1.5811. $$ Under $H_0:\mu=19$, the statistic has a $t_4$ distribution. The arithmetic checks: the observed mean is one hour above the reference value, so the statistic is positive.

Common mistake avoided: Dividing the sum of squared deviations by $n=5$ would give a variance of $8/5$, not the unbiased sample variance used in the t statistic. The required denominator here is $n-1=4$.

Worked Example: Using Sample Summaries

A hypothetical study of $n=10$ independent measurements reports $\overline{x}=52$ and $s=6$. Assume a normal population and consider the reference mean $\mu_0=50$. Calculate the studentized statistic and state its distribution if the reference mean is the true mean.

Method. Use the reported sample standard deviation directly in the estimated standard error $s/\sqrt{n}$. The statistic uses the difference from the reference mean, not the sample mean alone.

The estimated standard error is $$ \frac{s}{\sqrt{n}}=\frac{6}{\sqrt{10}}. $$ The observed difference is $52-50=2$, so $$ t_{\mathrm{obs}}=\frac{2}{6/\sqrt{10}} =\frac{\sqrt{10}}{3} \approx 1.0541. $$ Since $n=10$, the degrees of freedom are $10-1=9$. If $\mu=50$, the statistic has a $t_9$ distribution. The positive value is consistent with the observed sample mean being greater than 50.

Common mistake avoided: Using $s$ itself as the standard error would give $2/6$. A sample mean averages $n$ independent observations, so its estimated standard error is $s/\sqrt{n}$, not $s$.

Worked Example: A Negative Studentized Mean

A hypothetical sensor produces six independent readings: $10,11,12,12,13,14$. Assume the readings come from $N(\mu,\sigma^2)$. Calculate the statistic relative to $\mu_0=13$, and identify its null distribution.

Method. First compute the center and sample variance; then retain the sign of $\overline{x}-\mu_0$ when studentizing.

The sum of the readings is $72$, so $\overline{x}=72/6=12$. The deviations from 12 are $-2,-1,0,0,1,2$, with squared deviations summing to $4+1+0+0+1+4=10$. Therefore, $$ s^2=\frac{10}{6-1}=2,\qquad s=\sqrt{2}. $$ The estimated standard error is $\sqrt{2}/\sqrt{6}=\sqrt{1/3}$. The observed statistic is $$ t_{\mathrm{obs}}=\frac{12-13}{\sqrt{1/3}}=-\sqrt{3}\approx -1.7321. $$ If $\mu=13$, its null distribution is $t_5$. The result is negative, as it must be because the sample mean is below the reference mean.

Common mistake avoided: Reversing the subtraction would produce $+1.7321$. The defined statistic has $\overline{x}-\mu_0$ in its numerator, so its sign records whether the sample mean lies below or above the reference value.

Assumptions, Interpretation, and an Exam Trap

The theorem describes a sampling distribution, not a claim that every observed statistic equals a t-distributed value in any setting. Its exact finite-sample conclusion rests on independent, identically distributed normal observations and the usual sample variance with denominator $n-1$. Under a tested null value $\mu_0$, set $\mu=\mu_0$ in the theorem; then $(\overline{X}-\mu_0)/(S/\sqrt{n})\sim t_{n-1}$. If the null value is not the true population mean, this central t-distribution statement does not apply to that statistic as written.

The main exam trap is mixing up $S^2$ and a variance computed with denominator $n$. The chi-square result is $$ \frac{(n-1)S^2}{\sigma^2}\sim\chi^2_{n-1}, $$ where $S^2=\sum(X_i-\overline{X})^2/(n-1)$. Replacing $n-1$ by $n$ changes the scale and does not give the stated chi-square variable. A second trap is forgetting independence: a normal numerator and a chi-square denominator do not by themselves imply a t distribution unless they are independent.

Key takeaway: The t statistic arises because the normal sample mean gives a standard normal numerator, while the independent sample variance gives a chi-square denominator with $n-1$ degrees of freedom. Their ratio is exactly $t_{n-1}$ under the normal-sample model.

Check Your Understanding

For each problem, use the unbiased sample variance where raw data are given. The exact t result assumes the stated independent normal-sample model.

A random sample of four measurements is $3,5,5,7$, assumed normal. Relative to $\mu_0=4$, calculate $t_{\mathrm{obs}}$ and give its null degrees of freedom. (Answer: $t_{\mathrm{obs}}=\sqrt{3/2}\approx1.2247$, $3$ degrees of freedom.)
A normal sample has $n=8$, $\overline{x}=31.5$, and $s=4$. For $\mu_0=30$, calculate the studentized statistic and give its null distribution. (Answer: $t_{\mathrm{obs}}=3\sqrt{2}/4\approx1.0607$, $t_7$.)
Five independent normal observations are $6,8,10,12,14$. Calculate the statistic for $\mu_0=12$, showing the sample variance, and state its null degrees of freedom. (Answer: $s^2=10$, $t_{\mathrm{obs}}=-\sqrt{2}\approx-1.4142$, $4$ degrees of freedom.)
For an independent normal sample of size 12, define $Z=\sqrt{12}(\overline{X}-\mu)/\sigma$ and $U=11S^2/\sigma^2$. State the distributions and independence of $Z$ and $U$, then identify the distribution of the studentized mean. (Answer: $Z\sim N(0,1)$, $U\sim\chi^2_{11}$, independently, and $(\overline{X}-\mu)/(S/\sqrt{12})\sim t_{11}$.)
An analyst has $n=9$ independent observations from a population known to be uniform, not normal. The analyst computes $(\overline{X}-\mu)/(S/\sqrt{9})$. Is its exact distribution guaranteed to be $t_8$ by the normal-sample result? Identify the key issue. (Answer: No. The exact result requires a normal population; the uniform model does not meet that assumption.)

What You'll Learn