MOTIVATION FOR THE \(t\) DISTRIBUTION:


[What follows was first written in a post in SE Mathematics]


The \(t\)-Student distribution is a “natural” \(pdf\) to define.

If \(X_1, \ldots, X_n\) are iid observations \(\sim N(\mu,\sigma^2)\),



\(\Large\frac{\bar{X}\,-\,\mu}{\sigma/\sqrt{n}} \sim N(0,1)\).

And \(\Large\frac{\bar{X}\,-\,\mu}{\sigma/\sqrt{n}}\) is the formula of the Z-statistic.

If the \(X_1,\ldots,X_n\) are not normally distributed, the expression will tend to \(\sim N(0,1)\) as \(n\mapsto\infty\), the basis for the central limit theorem (CLT).

If the standard deviation of the population, \(\sigma\), is unknown we can replace it by the estimation based on a sample, \(S\), but then the expression (one-sample t-test statistic) will follow a \(t\)-distribution:

\(\Large \frac{\bar{X}\,-\,\mu}{S/\sqrt{n}}\sim t_{n-1}\)

with \(\large S=\sqrt{\frac{\sum(X_i-\bar X)^2}{n-1}}\) (Bessel’s correction).

Why?

\[\Large\frac{\bar{X}\,-\,\mu}{S/\sqrt{n}} = \frac{\bar{X}\,-\,\mu}{\frac{\sigma}{\sqrt{n}}} \frac{1}{\frac{S}{\sigma}}= Z\,\frac{1}{\frac{S}{\sigma}} = \frac{Z}{\sqrt{\frac{\sum(X_i-\bar X)^2}{(n-1)\,\sigma^2}}} \sim\frac{Z}{\sqrt{\frac{\chi_{n - 1}^2}{n-1}}} \sim t_{n-1}\small \tag 1\]

How did the chi square (\(\chi^2\)) made its entry into the \(pdf\)? It’s the distribution that models \(X^2\) with \(X\sim N(0,1)\).

In general, if \(X_1, \ldots, X_n\) are \(\sim \chi^2_{1\,df}\), then \(X_1 +\cdots+X_n \sim \chi^2\,\small(n\text{ df})\) here.

And in the case of the \(t\)-distribution the chi-square is suitable to model the sum of squared normals \(\large \sum(X_i-\bar X)^2\) in (1)), a well known property derived here, typically with \(n\) degrees of freedom, but why is it \(n\,-\,1\) here? In other words, why…

\(\Large \frac{\sum(X_i - \bar X)^2}{\sigma^2}\) in equation (1) becomes \(\Large \chi_{n-1}^2\) and not \(\Large \chi_{n}^2\)?

The explanation is in the answer here.

In a nutshell, there aren’t \(n\) \(X_i\) independently distributed random variables corresponding to the different observations, each one following a \(\chi^2\) distribution when squared (\(\small(X-\bar X)^2\)). In fact there are only \(n-1\) because of the insertion of the \(\bar X\) in the mix.



DERIVATION OF THE \(pdf\) OF THE \(t\) DISTRIBUTION:



First let’s recapute equation (1) by multiplying by \(\frac{\sigma}{\sigma}\) the Z-statistic \(\Large \frac{\bar X - \mu}{\frac{\sigma}{\sqrt{n}}}\):

\(\Large t= \frac{\bar{X}-\mu}{S/\sqrt{n}}=\frac{\frac{\bar{X}-\mu}{\sigma/\sqrt{n}}}{\sqrt{S^2/\sigma^2}} = \frac{Z}{\sqrt{\frac{\sum(X_i-\bar X)^2}{(n-1)\,\sigma^2}}} \small \tag1\)

In the above expression, \((\bar{X}-\mu)/(\mu/\sqrt{n})\,\,\sim \,\,N(0,1)\)



and \(\sqrt{S^2/\sigma^2} =\sqrt{ \frac{\sum_{x=1}^n(X - \bar{X})^2}{n-1}/\sigma^2} \,\,\sim \,\,\sqrt{\chi_{(n-1)}^2/(n-1)}\).



Assigning \(U = (\bar{X}-\mu)/(\mu/\sqrt{n})\) and \(V=\sqrt{S^2/\sigma^2}\), characterized as \(U \sim N(0,1)\) and \(V \sim \chi_k^2\), expression 1 becomes \(U/\sqrt{V/k}\).



With the premise of independence, the joint density is:



\(\Large f_{U,V}(u,v) = \frac{1}{(2\pi)^{1/2}} e^{-u^2/2} \frac{1}{\Gamma(\frac{k}{2})\,2^{k/2}}\,v^{(k/2)-1}\, e^{-v/2}\) with \(-\infty<u<\infty\) and \(0<v<\infty\).



Making the transformation \(\Large t=\frac{u}{\sqrt{v/k}}\) and \(\Large w=v\), (hence, \(\Large u=t\,(\frac{w}{k})^{1/2}\)), and with \(\Large (w/k)^{1/2}\) as the Jacobian, the marginal pdf will be:



\(\large f_T(t) = \int_0^\infty \,f_{U,V}\bigg(t\,(\frac{w}{k})^{1/2},w\bigg)(w/k)^{1/2}dw\)



\(=\Large \frac{1}{(2\pi)^{1/2}}\frac{1}{\Gamma(\frac{k}{2})2^{k/2}}\, \int_0^\infty\, e^{-\frac{\bigg(t(\frac{w}{k})^{1/2}\bigg)^2}{2}} w^{(k/2)-1} e^{-(\frac{w}{2})} \frac{w^{1/2}}{k^{1/2}}\,dw\)



\(=\Large \frac{1}{(2\pi)^{1/2}}\frac{1}{\Gamma(\frac{k}{2})2^{k/2}k^{1/2}}\, \Large \int_0^\infty\, w^{((k+1)/2)-1}\, e^{-(1/2)(1 + t^2/k)w} \,dw\)



The next step entails identifying in the previous equation the kernel of a gamma distribution pdf:



\(\Large x^{\alpha-1}\,e^{x\,\lambda}\)



with parameters \(\large (\alpha=(k+1)/2,\,\lambda=(1/2)(1+t^2/k))\).

The generic pdf for the gamma distribution is,

\(\Large \frac{\lambda^\alpha}{\Gamma(\alpha)}\,x^{\alpha-1}\,e^{x\,\lambda}\)


The strategy is then to synthesize the entire gamma pdf within the improper integral in our \(f_T(t)\) pdf in progress, so that we can simplify it as just \(1\), as we know to be true of all pdf’s. To get away with it we need to multiply numerator and denominator by the same coefficient:

\(\Large \frac{\Gamma(\alpha)\,\lambda^\alpha}{\Gamma(\alpha)\,\lambda^\alpha}\). And since neither \(\alpha\) nor \(\lambda\) include the integrating factor \(w\) we can include them inside the integral, or leave them out. Naturally, we want to leave within the integral \(\Large \frac{\lambda^\alpha}{\Gamma(\alpha)}\), and keep \(\Large \frac{\Gamma(\alpha)}{\lambda^\alpha}\) outside the integral. Now \(f_T(t)\) will look hideous for just one second:



\(\Large= \frac{1}{(2\pi)^{1/2}}\frac{1}{\Gamma(\frac{k}{2})2^{k/2}k^{1/2}}\, \int_0^\infty\frac{((1/2)(1+t^2/k))^{(k+1)/2}}{\Gamma((k+1)/2)}\) \(w^{((k+1)/2)-1} e^{-(1/2)(1 + t^2/k)w} dw\, \frac{\Gamma((k+1)/2)}{((1/2)(1+t^2/k))^{(k+1)/2}}\)



… because everything between \(\int\) and \(dw\) is just the gamma \(pdf\) integrated over its entire support, so it becomes \(1\), and we are left with:



\(=\Large \frac{1}{(2\pi)^{1/2}}\frac{1}{\Gamma(\frac{k}{2})2^{k/2}k^{1/2}}\, \frac{\Gamma((k+1)/2)}{((1/2)(1+t^2/k))^{(k+1)/2}}\)



\(=\Large \frac{1}{(2\pi)^{1/2}}\frac{1}{\Gamma(\frac{k}{2})2^{k/2}k^{1/2}}\,\Gamma((k+1)/2)\, \Big[\frac{2}{(1+t^2/k)}\Big]^{(k+1)/2}\)



\(=\Large \frac{\Gamma(\frac{k+1}{2})}{\Gamma(\frac{k}{2})}\, \frac{1}{(2\pi)^{1/2}2^{k/2}k^{1/2}}\, \Big[\frac{2}{(1+t^2/k)}\Big]^{(k+1)/2}\)



\(=\Large \frac{\Gamma(\frac{k+1}{2})}{\Gamma(\frac{k}{2})}\, \frac{1}{(2\pi)^{1/2}2^{k/2}k^{1/2}}\, \frac{2^{(k+1)/2}}{(1+t^2/k)^{(k+1)/2}}\)



\(=\Large \frac{\Gamma(\frac{k+1}{2})}{\Gamma(\frac{k}{2})}\, \frac{1}{(\pi)^{1/2}k^{1/2}}\, \frac{1}{(1+t^2/k)^{(k+1)/2}}\)





\(\Large f_T(t)= \frac{\Gamma(\frac{k+1}{2})}{\Gamma(\frac{k}{2})}\, \frac{1}{\sqrt{k\,\pi}}\, \Big(1+\frac{t^2}{k}\Big)^{-\frac{k+1}{2}}\)



which is the \(pdf\) of the \(t\)-Student or Gosset distribution with \(k\) degrees of freedom.


Home Page