Hotelling's T-squared distribution
Template:Redirect Template:Probability distribution
In statistics, particularly in hypothesis testing, the Hotelling's T-squared distribution (T2), proposed by Harold Hotelling,<ref name=H1931 /> is a multivariate probability distribution that is tightly related to the F-distribution and is most notable for arising as the distribution of a set of sample statistics that are natural generalizations of the statistics underlying the Student's t-distribution. The Hotelling's t-squared statistic (t2) is a generalization of Student's t-statistic that is used in multivariate hypothesis testing.<ref name='jonhson'>Template:Cite book</ref>
Motivation
The distribution arises in multivariate statistics in undertaking tests of the differences between the (multivariate) means of different populations, where tests for univariate problems would make use of a t-test. The distribution is named for Harold Hotelling, who developed it as a generalization of Student's t-distribution.<ref name=H1931>Template:Cite journal</ref>
Definition
If the vector <math>d</math> is Gaussian multivariate-distributed with zero mean and unit covariance matrix <math>N(\mathbf{0}_{p}, \mathbf{I}_{p, p})</math> and <math>M</math> is a <math>p \times p</math> random matrix with a Wishart distribution <math>W(\mathbf{I}_{p, p}, m)</math> with unit scale matrix and m degrees of freedom, and d and M are independent of each other, then the quadratic form <math>X</math> has a Hotelling distribution (with parameters <math>p</math> and <math>m</math>):<ref>Eric W. Weisstein, MathWorld</ref> <math display="block">X = m d^T M^{-1} d \sim T^2(p, m).</math>
It can be shown that if a random variable X has Hotelling's T-squared distribution, <math>X \sim T^2_{p,m}</math>, then:<ref name=H1931/> <math display="block"> \frac{m-p+1}{pm} X\sim F_{p,m-p+1} </math> where <math>F_{p,m-p+1}</math> is the F-distribution with parameters p and m − p + 1.
Hotelling t-squared statistic
Let <math>\hat{\mathbf \Sigma}</math> be the sample covariance:
<math display="block"> \hat{\mathbf \Sigma} = \frac 1 {n-1} \sum_{i=1}^n \left(\mathbf{x}_i -\overline{\mathbf{x}}\right) \left(\mathbf{x}_i - \overline{\mathbf{x}}\right)' </math>
where we denote transpose by an apostrophe. It can be shown that <math>\hat{\mathbf \Sigma}</math> is a positive (semi) definite matrix and <math>(n-1)\hat{\mathbf \Sigma}</math> follows a p-variate Wishart distribution with n − 1 degrees of freedom.<ref name="MKB">Template:Cite book</ref> The sample covariance matrix of the mean reads <math>\hat{\mathbf \Sigma}_\overline{\mathbf x}=\hat{\mathbf \Sigma}/n</math>.<ref name="Fogelmark2018">Template:Cite journal</ref>
The Hotelling's t-squared statistic is then defined as:<ref>Template:Cite web</ref>
<math display="block"> t^2=(\overline{\mathbf x}-\boldsymbol{\mu})'\hat{\mathbf \Sigma}_\overline{\mathbf x}^{-1} (\overline{\mathbf x}-\boldsymbol{\mathbf\mu})=n(\overline{\mathbf x}-\boldsymbol{\mu})'\hat{\mathbf \Sigma}^{-1} (\overline{\mathbf x}-\boldsymbol{\mathbf\mu}), </math>
which is proportional to the Mahalanobis distance between the sample mean and <math>\boldsymbol{\mu}</math>. Because of this, one should expect the statistic to assume low values if <math>\overline{\mathbf x} \approx \boldsymbol{\mu}</math>, and high values if they are different.
From the distribution,
<math display="block">t^2 \sim T^2_{p,n-1}=\frac{p(n-1)}{n-p} F_{p,n-p} ,</math>
where <math>F_{p,n-p}</math> is the F-distribution with parameters p and n − p.
In order to calculate a p-value (unrelated to p variable here), note that the distribution of <math>t^2</math> equivalently implies that
<math display="block"> \frac{n-p} {p(n-1)} t^2 \sim F_{p,n-p} .</math>
Then, use the quantity on the left hand side to evaluate the p-value corresponding to the sample, which comes from the F-distribution. A confidence region may also be determined using similar logic.
Motivation
Let <math>\mathcal{N}_p(\boldsymbol{\mu},{\mathbf \Sigma})</math> denote a p-variate normal distribution with location <math>\boldsymbol{\mu}</math> and known covariance <math>{\mathbf \Sigma}</math>. Let
<math display="block">{\mathbf x}_1,\dots,{\mathbf x}_n\sim \mathcal{N}_p(\boldsymbol{\mu},{\mathbf \Sigma})</math>
be n independent identically distributed (iid) random variables, which may be represented as <math>p\times1</math> column vectors of real numbers. Define
<math display="block">\overline{\mathbf x}=\frac{\mathbf{x}_1+\cdots+\mathbf{x}_n}{n}</math>
to be the sample mean with covariance <math>{\mathbf \Sigma}_\overline{\mathbf x}={\mathbf \Sigma}/ n</math>. It can be shown that
<math display="block">(\overline{\mathbf x}-\boldsymbol{\mu})'{\mathbf \Sigma}_\overline{\mathbf x}^{-1}(\overline{\mathbf x}-\boldsymbol{\mathbf\mu})\sim\chi^2_p ,</math>
where <math>\chi^2_p</math> is the chi-squared distribution with p degrees of freedom.<ref>End of chapter 4.2 of Template:Harvp</ref>
Template:Math proof Template:Collapse bottom Alternatively, one can argue using density functions and characteristic functions, as follows. Template:Collapse top Template:Math proof \cdot \cancel{n} \cdot \boldsymbol\Sigma^{-1} \right|^{1/2} \\ & = \left| \left[ (\cancel{\boldsymbol\Sigma^{-1}} -2i \theta \cancel{\boldsymbol\Sigma^{-1}} ) \cancel{\boldsymbol\Sigma} \right]^{-1} \right|^{1/2} \\ & = |\mathbf I_p-2 i \theta \mathbf I_p|^{-1/2} \end{align} </math> where <math>I_p</math> is an identity matrix of dimension <math>p</math>. Finally, calculating the determinant, we obtain: <math display="block"> \begin{align} & = (1-2 i \theta)^{-p/2} \end{align} </math> which is the characteristic function for a chi-square distribution with <math>p</math> degrees of freedom. <math>\;\;\;\blacksquare</math> }} Template:Collapse bottom
Two-sample statistic
If <math>{\mathbf x}_1,\dots,{\mathbf x}_{n_x}\sim N_p(\boldsymbol{\mu},{\mathbf \Sigma})</math> and <math>{\mathbf y}_1,\dots,{\mathbf y}_{n_y}\sim N_p(\boldsymbol{\mu},{\mathbf \Sigma})</math>, with the samples independently drawn from two independent multivariate normal distributions with the same mean and covariance, and we define
<math display="block">\overline{\mathbf x}=\frac{1}{n_x}\sum_{i=1}^{n_x} \mathbf{x}_i \qquad \overline{\mathbf y}=\frac{1}{n_y}\sum_{i=1}^{n_y} \mathbf{y}_i</math>
as the sample means, and
<math display="block">\begin{align} \hat{\mathbf \Sigma}_{\mathbf x} &= \frac{1}{n_x-1}\sum_{i=1}^{n_x} \left(\mathbf{x}_i-\overline{\mathbf x}\right) \left(\mathbf{x}_i-\overline{\mathbf x}\right)' \\ \hat{\mathbf \Sigma}_{\mathbf y}&= \frac{1}{n_y-1}\sum_{i=1}^{n_{y}} \left(\mathbf{y}_i-\overline{\mathbf y}\right) \left(\mathbf{y}_i-\overline{\mathbf y}\right)' \end{align}</math>
as the respective sample covariance matrices. Then
<math display="block">\hat{\mathbf \Sigma}= \frac{(n_x - 1) \hat{\mathbf \Sigma}_{\mathbf x} + (n_y - 1) \hat{\mathbf \Sigma}_{\mathbf y}}{n_x+n_y-2}</math>
is the unbiased pooled covariance matrix estimate (an extension of pooled variance).Template:Anchor
Finally, the Hotelling's two-sample t-squared statistic is
<math display="block">t^2 = \frac{n_x n_y}{n_x+n_y}(\overline{\mathbf x}-\overline{\mathbf y})'\hat{\mathbf \Sigma}^{-1}(\overline{\mathbf x}-\overline{\mathbf y}) \sim T^2(p, n_x+n_y-2)</math>
Related concepts
It can be related to the F-distribution by<ref name="MKB"/>
<math display="block">\frac{n_x+n_y-p-1}{(n_x+n_y-2)p}t^2 \sim F(p,n_x+n_y-1-p).</math>
The non-null distribution of this statistic is the noncentral F-distribution (the ratio of a non-central Chi-squared random variable and an independent central Chi-squared random variable) <math display="block">\frac{n_x+n_y-p-1}{(n_x+n_y-2)p}t^2 \sim F(p,n_x+n_y-1-p;\delta),</math> with <math display="block">\delta = \frac{n_x n_y}{n_x+n_y}\boldsymbol{d}'\mathbf{\Sigma}^{-1}\boldsymbol{d},</math> where <math>\boldsymbol{d}=\mathbf{\overline{x} - \overline{y}}</math> is the difference vector between the population means.
In the two-variable case, the formula simplifies nicely allowing appreciation of how the correlation, <math>\rho</math>, between the variables affects <math>t^2</math>. If we define <math display="block">d_{1} = \overline{x}_{1}-\overline{y}_{1}, \qquad d_{2} = \overline{x}_{2}-\overline{y}_{2}</math> and <math display="block">s_1 = \sqrt{\Sigma_{11}} \qquad s_2 = \sqrt{\Sigma_{22}} \qquad \rho = \Sigma_{12}/(s_1 s_2) = \Sigma_{21}/(s_1 s_2)</math> then <math display="block">t^2 = \frac{n_x n_y}{(n_x+n_y)(1-\rho ^2)} \left [ \left ( \frac{d_1}{s_1} \right )^2+\left ( \frac{d_2}{s_2} \right )^2-2\rho \left ( \frac{d_1}{s_1} \right )\left ( \frac{d_2}{s_2} \right ) \right ] </math> Thus, if the differences in the two rows of the vector <math>\mathbf d = \overline{\mathbf x}-\overline{\mathbf y}</math> are of the same sign, in general, <math>t^2</math> becomes smaller as <math>\rho</math> becomes more positive. If the differences are of opposite sign <math>t^2</math> becomes larger as <math>\rho</math> becomes more positive.
A univariate special case can be found in Welch's t-test.
More robust and powerful tests than Hotelling's two-sample test have been proposed in the literature, see for example the interpoint distance based tests which can be applied also when the number of variables is comparable with, or even larger than, the number of subjects.<ref>Template:Cite journal</ref><ref>Template:Cite journal</ref>
See also
- Student's t-test in univariate statistics
- Student's t-distribution in univariate probability theory
- Multivariate Student distribution
- F-distribution (commonly tabulated or available in software libraries, and hence used for testing the T-squared statistic using the relationship given above)
- Wilks's lambda distribution (in multivariate statistics, Wilks's Λ is to Hotelling's T2 as Snedecor's F is to Student's t in univariate statistics)