Studentized range distribution

Applications

Critical values of the studentized range distribution are used in Tukey's range test.

The studentized range is used to calculate significance levels for results obtained by data mining, where one selectively seeks extreme differences in sample data, rather than only sampling randomly.

The Studentized range distribution has applications to hypothesis testing and multiple comparisons procedures. For example, Tukey's range test and Duncan's new multiple range test (MRT), in which the sample x₁, ..., x_n is a sample of means and q is the basic test-statistic, can be used as post-hoc analysis to test between which two groups means there is a significant difference (pairwise comparisons) after rejecting the null hypothesis that all groups are from the same population (i.e. all means are equal) by the standard analysis of variance.

Related distributions

When only the equality of the two groups means is in question (i.e. whether μ₁ = μ₂), the studentized range distribution is similar to the Student's t distribution, differing only in that the first takes into account the number of means under consideration, and the critical value is adjusted accordingly. The more means under consideration, the larger the critical value is. This makes sense since the more means there are, the greater the probability that at least some differences between pairs of means will be significantly large due to chance alone.

Derivation

The studentized range distribution function arises from re-scaling the sample range R by the sample standard deviation s, since the studentized range is customarily tabulated in units of standard deviations, with the variable q = R⁄s . The derivation begins with a perfectly general form of the distribution function of the sample range, which applies to any sample data distribution.

In order to obtain the distribution in terms of the "studentized" range q, we will change variable from R to s and q. Assuming the sample data is normally distributed, the standard deviation s will be χ distributed. By further integrating over s we can remove s as a parameter and obtain the re-scaled distribution in terms of q alone.

General form

For any probability density function f_X, the range probability density f_R is:

${\displaystyle f_R(r;k)=k(k-1){\int }_{-\mathrm{\infty }}^{\mathrm{\infty }}{f}_X\left(t+\frac{1}{2}r\right)f_X\left(t-\frac{1}{2}r\right){\left[{\int }_{t-\frac{1}{2}r}^{t+\frac{1}{2}r}{f}_X(x)\mathrm{d}x\right]}^{k-2}\mathrm{d}t}$

What this means is that we are adding up the probabilities that, given k draws from a distribution, two of them differ by r, and the remaining k − 2 draws all fall between the two extreme values. If we change variables to u where ${\displaystyle u=t-\frac{1}{2}r}$ is the low-end of the range, and define F_X as the cumulative distribution function of f_X, then the equation can be simplified:

${\displaystyle f_R(r;k)=k(k-1){\int }_{-\mathrm{\infty }}^{\mathrm{\infty }}{f}_X(u+r)f_X(u){\left[F_X(u+r)-F_X(u)\right]}^{k-2}\mathrm{d}u}$

We introduce a similar integral, and notice that differentiating under the integral-sign gives

${\displaystyle \begin{array}{cc}\frac{\mathrm{\partial }}{\mathrm{\partial }r}& \left[k{\int }_{-\mathrm{\infty }}^{\mathrm{\infty }}{f}_X(u)[F_X(u+r)-F_X(u){]}^{k-1}\mathrm{d}u\right]\\ =& k(k-1){\int }_{-\mathrm{\infty }}^{\mathrm{\infty }}{f}_X(u+r)f_X(u)[F_X(u+r)-F_X(u){]}^{k-2}\mathrm{d}u\end{array}}$

which recovers the integral above, so that last relation confirms

${\displaystyle \begin{array}{cc}{F}_R(r;k)& =k{\int }_{-\mathrm{\infty }}^{\mathrm{\infty }}{f}_X(u)[F_X(u+r)-F_X(u){]}^{k-1}\mathrm{d}u\\ & =k{\int }_{-\mathrm{\infty }}^{\mathrm{\infty }}[F_X(u+r)-F_X(u){]}^{k-1}\mathrm{d}{F}_X(u)\end{array}}$

because for any continuous cdf

${\displaystyle \frac{\mathrm{\partial }{F}_R(r;k)}{\mathrm{\partial }r}=f_R(r;k)}$

Special form for normal data

The range distribution is most often used for confidence intervals around sample averages, which are asymptotically normally distributed by the central limit theorem.

In order to create the studentized range distribution for normal data, we first switch from the generic f_X and F_X to the distribution functions φ and Φ for the standard normal distribution, and change the variable r to s·q, where q is a fixed factor that re-scales r by scaling factor s:

${\displaystyle f_R(q;k)=sk(k-1){\int }_{-\mathrm{\infty }}^{\mathrm{\infty }}\phi (u+sq)\phi (u){\left[\mathrm{\Phi }(u+sq)-\mathrm{\Phi }(u)\right]}^{k-2}\mathrm{d}u}$

Choose the scaling factor s to be the sample standard deviation, so that q becomes the number of standard deviations wide that the range is. For normal data s is chi distributed and the distribution function f_S of the chi distribution is given by:

${\displaystyle f_S(s;\nu )\mathrm{d}s=\{\begin{array}{ll}{\displaystyle \frac{{\nu }^{\nu /2}{s}^{\nu -1}{e}^{-\nu s^2/2}}{2^{\left(\nu /2-1\right)}\mathrm{\Gamma }(\nu /2)}}\mathrm{d}s& \text{for }0<s<\mathrm{\infty },\\ 0& \text{otherwise}.\end{array}}$

Multiplying the distributions f_R and f_S and integrating to remove the dependence on the standard deviation s gives the studentized range distribution function for normal data:

${\displaystyle f_R(q;k,\nu )=\frac{{\nu }^{\nu /2}k(k-1)}{2^{\left(\nu /2-1\right)}\mathrm{\Gamma }(\nu /2)}{\int }_0^{\mathrm{\infty }}{s}^{\nu }{e}^{-\nu s^2/2}{\int }_{-\mathrm{\infty }}^{\mathrm{\infty }}\phi (u+sq)\phi (u){\left[\mathrm{\Phi }(u+sq)-\mathrm{\Phi }(u)\right]}^{k-2}\mathrm{d}u\mathrm{d}s}$

where

q is the width of the data range measured in standard deviations,

ν is the number of degrees of freedom for determining the sample standard deviation, and

k is the number of separate averages that form the points within the range.

The equation for the pdf shown in the sections above comes from using

${\displaystyle e^{-\nu s^2/2}=\sqrt{2\pi }\phi (\sqrt{\nu }s)}$

to replace the exponential expression in the outer integral.