In $\mathcal{F}$ test, the hypotheses involve multiple regression coefficients with the $\mathcal{F}_{0}$ test statistic: \begin{equation} z_{j} = \beta_{0} + \beta_{1} x_{1} + \beta_{2} x_{2} + ... + \beta_{n} x_{j} + \epsilon_{j} \end{equation} where Z is a dependent variable, $X$ the independent variable, and $\beta_{j} \in \mathbb{C}$ for $n \in \mathbb{N}$ \cite{ahlfors, ftests}. The null hypothesis is $H_{0}: \mathbf{L} \beta = \mathbf{c}$ without joint hypothesis.

I think the general form of the matrix $L$ is difficult one for computation. Consider a joint hypothesis $H_{0}: \beta_{1} = \beta_{2} = 0$. This leads to a term $X'X$ that is a diagonal matrix, see these lecture notes. So the corresponding matrix of $\mathbf{L} := X'X$ is a diagonal matrix and $\mathbf{c} := [[0 0 ... 0]^{-1}$ is a null vector.

Joint hypothesis

Here you can consider $\mathbf{L} \hat{\beta} - \mathbf{c} = 0$ for the regression of the observed value. For detailed steps, see the material of Zurich University, here and last part of it. The observed $\mathcal{F}$ test statistic is the following if the null hypothesis has a form $H_{0}: \beta_{1} = \beta_{2} = ... = \beta_{k} = 0$.

\begin{equation} F_{0} = \frac{ \left( SSR_{r} - SSR_{ur} \right) / q }{ SSR_{ur} / \left(n - (k+1) \right)}, \end{equation} where $SSR_{r}$ is the Sum of Squared Residuals of the restricted model, $SSR_{ur}$ is the same for the unrestricted model, and $n$ is the number of observations \cite{ftests} p. 4. This is similar to the one in Zurich University's materials.

\begin{equation} \mathbf{X'X} = \begin{bmatrix} \gamma_{1} & . & . & . & \gamma_{M+1} & 0 & . & . & . & 0 \\[0.3em] 0 & . & \gamma_{1} & . & . & \gamma_{M+1} & 0 & . & . & 0 \\[0.3em] 0 & 0 & 0 & . & . & 0 & \gamma_{1} & . & . & \gamma_{M+1} \end{bmatrix} \end{equation}

and the corresponding vector of $\mathbf{c} = [0 0 ... 0]^{-1}$ is a null vector here under $H_{0}$.

Real values vs Complex values

For me both are ok, but complex values are for useful for future applications. You can assume that the matrix $\mathbf{L}$ and $\mathbf{c}$ are real valued if it makes the analysis easier. Here however something about analysis with complex data.


In Matlab, you can do $\mathcal{F}$ test simply by

[h,p,ci,stats] = vartest2(J, model1)
[h2,p2,ci2,stats2] = vartest2(J, model2)

and compare the appropriate models. However, here, Matlab does not give you the values of the matrices. I would like to know in the context of Matlab, for instance, how it handles the matrices $\mathbf{L}$ and $\mathbf{c}$.

How can you present the matrices $\mathbf{L}$ and $\mathbf{c}$ for $\mathcal{F}$ test?

My sources

  • Complex analysis, Lars Ahlfors, 1979. (ahlfors)
  • Statistical Concepts: A Second Course, Lomax R, 2007. (ftests)
  • Some lecture materials from Zurich University Economic Department
  • Some lecture materials from Stanford University Political Sciences
  • $\begingroup$ i am vevy curious to know how Ahlfor's stands as a reference for this question. $\endgroup$
    – meh
    Commented Apr 5, 2015 at 16:24
  • $\begingroup$ @aginensky He started to use Z for dependent variables and X for independent variables and I use his convention. $\endgroup$ Commented Apr 5, 2015 at 16:25
  • $\begingroup$ truly not important, but z in Ahlfor's book means a complex variable- z = x+iy typically with x,y real. $\endgroup$
    – meh
    Commented Apr 5, 2015 at 16:27
  • $\begingroup$ @aginensky See stats.stackexchange.com/a/66268/3017 $\endgroup$ Commented Apr 5, 2015 at 16:29


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.