\documentclass[mathserif]{beamer} % Get Computer Modern math font. \usefonttheme{serif} % Looks like Computer Modern for non-math text -- nice! \setbeamertemplate{navigation symbols}{} % Supress navigation symbols \usetheme{Berlin} % Diplays sections on top \usepackage[english]{babel} \setbeamertemplate{footline}[frame number] \mode % \mode{\setbeamercolor{background canvas}{bg=black!5}} \title{Within Cases} \subtitle {The Humble $t$-test} % (optional) \date{} % To suppress date \begin{document} \begin{frame} \titlepage \end{frame} \begin{frame} \frametitle{Overview} \tableofcontents \end{frame} \section{The Issue} \begin{frame}{Independent Observations}%{Subtitles are optional.} % - A title should summarize the slide in an understandable fashion % for anyone how does not follow everything on the slide itself. \begin{itemize} \item Most statistical models assume independent observations. \item Sometimes the assumption of independence is unreasonable. \item For example, times series and within cases designs. \end{itemize} \end{frame} \begin{frame}{Within Cases} \begin{itemize} \item A case contributes a value of the response variable for every value of a categorical explanatory variable. \item As opposed to explanatory variables that are \emph{Between Cases}: Explanatory variables partition the sample. \item It is natural to expect data from the same case to be correlated, \emph{not} independent. \item For example, the same subject appears in several treatment conditions \item Hearing study: How does pitch affect our ability to hear faint sounds? Subjects are presented with tones at a variety of different pitch and volume levels (in a random order). They press a key when they think they hear something. \item A study can have both within and between cases factors. \end{itemize} \end{frame} \begin{frame}{You may hear terms like} \begin{itemize} \item \textbf{Longitudinal}: The same variables are measured repeatedly over time. Usually lots of variables, including categorical ones, and large samples. If there's an experimental treatment, itŐs usually once at the beginning, like a surgery. Basically itŐs \emph{tracking} what happens over time. \item \textbf{Repeated measures}: Usually, same subjects experience two or more experimental treatments. Usually quantitative explanatory variables and small samples. \end{itemize} \end{frame} \section{Univariate} \begin{frame}{Student's Sleep Study (\emph{Biometrika}, 1908)} {First Published Example of a $t$-test} \begin{itemize} \item Patients take two sleeping medicines several days apart. \item Half get $A$ first, half get $B$ first. \item Reported hours of sleep are recorded. \item It's natural to subtract, and test whether the mean \emph{difference} equals zero. \item That's what Gossett did. \item But some might do an independent $t$-test with $n_1=n_2$. \item Is it harmful? \end{itemize} \end{frame} \begin{frame}{Conclusions from an earlier discussion} \begin{itemize} \item When covariance is positive, matched $t$-test has better power \item Each case serves as its own control. \item A huge number of unknown influences are removed by subtraction. \item This makes the analysis more precise. \end{itemize} \end{frame} \section{Multivariate} \begin{frame}{Hotelling's $t^2$}{Multivariate Matched $t$-test} \begin{itemize} \item $\mathbf{X}_1, \ldots, \mathbf{X}_n \stackrel{i.i.d.}{\sim} N_k(\boldsymbol{\mu},\boldsymbol{\Sigma})$ \item $\overline{\mathbf{X}}_n = \frac{1}{n} \sum_{i=1}^n \mathbf{X}_i$ and $\mathbf{S} = \frac{1}{n-1} \sum_{i=1}^n \left(\mathbf{X}_i-\overline{\mathbf{X}}_n\right) \left(\mathbf{X}_i-\overline{\mathbf{X}}_n\right)^\prime$ \item $t^2 = n\left(\overline{\mathbf{X}}_n - \boldsymbol{\mu}\right)^\prime \mathbf{S}^{-1} \left(\overline{\mathbf{X}}_n - \boldsymbol{\mu}\right) \sim T^2(k,n-1)$ \item That is, $\frac{n-k}{k(n-1)}t^2 \sim F(k,n-k)$ \item When $k=1$, reduces to the familiar $t^2=F(1,n-1)$ \item Test $H_0: \boldsymbol{\mu} = \boldsymbol{\mu}_0$ \end{itemize} \end{frame} \begin{frame}{Test \emph{Collections} of Contrasts} {$H_0: \mathbf{L}\boldsymbol{\mu} = \mathbf{h}$, where $ \mathbf{L}$ is $r \times k$} \begin{itemize} \item $t^2 = n\left(\overline{\mathbf{X}}_n - \boldsymbol{\mu}\right)^\prime \mathbf{S}^{-1} \left(\overline{\mathbf{X}}_n - \boldsymbol{\mu}\right) \sim T^2(k,n-1)$, \\ so if $H_0$ is true \item $t^2 = n\left(\mathbf{L}\overline{\mathbf{X}}_n - \mathbf{h}\right)^\prime \left(\mathbf{LSL}^\prime\right)^{-1} \left(\mathbf{L}\overline{\mathbf{X}}_n - \mathbf{h}\right) \sim T^2(r,n-1)$ \item Could also calculate contrast variables, like differences. \begin{itemize} \item Expected value of the contrast is the contrast of expected values. \item Just test (simultaneously) whether the means of the contrast variables are zero, using the first formula. \end{itemize} \item For 2 or more within-cases factors, use contrasts to test for main effects, interactions. \end{itemize} \end{frame} \begin{frame}{Compare Wald-like tests} Recall \begin{itemize} \item If $\mathbf{Y}_n = \sqrt{n}(\mathbf{T}_n-\boldsymbol{\theta}) \stackrel{d}{\rightarrow} \mathbf{Y} \sim N_k(\mathbf{0},\boldsymbol{\Sigma})$, then \begin{eqnarray*} W_n & = & n (\mathbf{LT}_n-\mathbf{h})^\prime \left( \mathbf{L}\widehat{\boldsymbol{\Sigma}}_n\mathbf{L}^\prime \right)^{-1} (\mathbf{LT}_n-\mathbf{h}) \stackrel{d}{\rightarrow} W \sim \chi^2(r) \\ t^2 & = & n\left(\mathbf{L}\overline{\mathbf{X}}_n - \mathbf{h}\right)^\prime \left(\mathbf{LSL}^\prime\right)^{-1} \left(\mathbf{L}\overline{\mathbf{X}}_n - \mathbf{h}\right) \sim T^2(r,n-1) \end{eqnarray*} \item And $F = \frac{n-r}{r(n-1)}t^2 \sim F(r,n-r) \Rightarrow t^2 = \frac{n-1}{n-r}\,rF \stackrel{d}{\rightarrow} Y \sim \chi^2(r) $ \item So the Hotelling $t$-squared test is robust with respect to normality. \end{itemize} \end{frame} \end{document}