# Why is t-distribution independent of true variance?

Why is the t-distribution independent of the true variance of the underlying function? Intuitively I would not expect this to be true. Can anyone explain why this is true, using words rather than mathematical formulas?

Here is an example of exactly what I mean:

Consider 2 functions F and G.

F returns values which are normally distributed around mean 0 with variance 1.

G returns values which are normally distributed around mean 0 with variance 5.

Consider taking a sample of size 20 from F or G.

If we take multiple samples like this from F, the sampling distribution will be Student's t-distribution with 20-1 degrees of freedom.

If we take multiple samples like this from G, the sampling distribution will be Student's t-distribution with 20-1 degrees of freedom.

Intuitively I would expect to get a different distribution.

• It seems like you are confusing things. What can be said is that if you have a sample from a normal distribution the t statistic will not dependent on the population variance. – Michael R. Chernick May 5 '17 at 19:06
• @MichaelChernick can you please point out the things that are wrong in my post? – Atte Juvonen May 5 '17 at 19:47
• Far from being any different, $F$ and $G$ are identical: you can view them as producing exactly the same quantities, but merely expressed in different units of measurement. ($F$ could measure a signed distance in Tuscan canne while $G$ could measure the same distance in bracci, for instance.) The $t$ statistic is one of many possible quantities that does not change when the units of measurement are changed. – whuber May 5 '17 at 20:01
• The T-distribution you refer to is the distribution of the test statistic $T$.The test-statistic $T$ is independent of the population variance because the population variance is a constant not a random variable. The test-statistic $T$ is asymptotically independent of the sample variances as well because of the central limit theorem. I think the test-statistic is also independent of the sample variance when the data are normal. They will not be independent when the data are non-normal. – AdamO Mar 28 '18 at 15:50

The first thing to note is that the sample mean and sample variance are independent. The proof of this is a bit long to include here but can be found online pretty easily. The sample mean has distribution $$\bar{X} -\mu \sim N(0, \sigma^{2} /n)$$ and the sample variance has dist $$(n-1)S^{2} / \sigma^{2} \sim \chi^{2}_{n-1}$$ With abuse of notation $$\frac{\bar{X}-\mu}{S / \sqrt{n}}= \frac{\sigma Z / \sqrt{n}}{\sigma \sqrt{\chi^{2}_{n-1}/(n-1)}/\sqrt{n}}=\frac{Z}{\sqrt{\chi^{2}_{n-1}/(n-1)}}=t_{n-1}$$ Here the Z is a standard normal, and is independent of the Chi squared term, and the final equality comes from the definition of a t random variable.

• Hi @AdamO, from what I understand he's not asking about a two sample t-test at all. The two distributions are brought up just to contrast the fact that the a t statistic computed from either one on the null H0: mu=0 would have the same distribution. – MHankin Mar 29 '18 at 18:00
• You're right. Somehow I read into F and G as being two pops from which one would want to compare variances. Disregard the previous comment. – AdamO Mar 29 '18 at 19:56

Keep in mind that the statistic is divided by the sample standard deviation so in a way you "use" this value as a (roughly speaking) standardisation tool.

Thus, if you take 20 sample values from either f then both the nominator and the denominator will be lower (most probably) compared to if you do the same from g. This will make both statistics to follow the t-distribution.

By definition (at least my definition), the Student's T distribution is the distribution of the variable

$$T=\sqrt n \frac {\overline{X_n} - \mu} {\sqrt{\frac 1 {n-1}\sum(X_i-\overline{X_n})^2}}$$

Where the $X_i$ are independent $N(\mu, \sigma^2)$ variables. The question is: why is this definition actually a definition? Why does the distribution of $T$ not depend on $\mu$ and $\sigma^2$?

A simple way to prove that the distribution is well-defined is to do the following algebraic manipulations. First, rewrite the sample variance by adding and removing $\mu$:

$$\sum(X_i-\overline{X_n})^2 = \sum((X_i-\mu)-(\overline{X_n}-\mu))^2$$

And now divide both parts of the fraction by $\sigma$:

$$T=\sqrt n \frac {(\overline{X_n} - \mu)/\sigma} {\sqrt{\frac 1 {n-1}\sum(\frac{{X_i-\mu}}\sigma-\frac{\overline{X_n}-\mu}\sigma)^2}}$$

The point is, we've gotten an expression of the form:

$$T=f\left(\frac{X_1-\mu}{\sigma}, ..., \frac{X_n-\mu}{\sigma}, \frac{\overline{X_n}-\mu}\sigma\right)$$

That is, we have expressed $T$ as a function of the variables $Y_i=\frac{X_i-\mu}{\sigma}$ and of $\frac{\overline{X_n}-\mu}\sigma$, which are all normally distributed variables. In fact, since $\frac{\overline{X_n}-\mu}\sigma$ can itself be expressed in terms of the $Y_i$, we can even write, for some complicated function $f$:

$$T=f(Y_1, ..., Y_n)$$

Since the joint distribution of the vector $(Y_1, ..., Y_n)$ doesn't depend on $\mu$ or $\sigma^2$ (it is $N(0, I_n)$), neither does the distribution of $T$.

• Wouldn't there be $n+1$ elements in your 3rd display of $T$? That $n+1$-th element is not $N(0,1)$ but $N(0, 1/n)$ approximately, and it's not immediately claimed or (to me) clear the $Y_{n+1}$would be independent of any of the previous $n$ observations. – AdamO Mar 28 '18 at 15:56
• @AdamO I removed the statement that $Y_{n+1}$ has unit variance. I'm not sure if $Y_{n+1}$ is independent from the $Y_i$, but I don't claim it is. Since $Y_{n+1}$ can be expressed as a function of the $Y_i$ ($i<n+1$), we only need independence of the $Y_i$ ($i<n+1$). – Jack M Mar 28 '18 at 16:29