Having a sample of observations, $x_i, i=1,..,n$, to test the null hypothesis stating that the data follows a specific distribution $F_0 (x)$, the Cram\'{e}r-von-Mises goodness-of-fit test statistic is,

\begin{equation}\label{e1} \omega^2_n=n\int_{-\infty}^{+\infty} \left ( {\hat {F}_n(x)} - F_0 (x)\right )^2 {\mathrm {d}}F_0(x), \end{equation} where $\hat {F}_n(x)$ is the empirical distribution function. If the test distribution $ F_0$ has unknown parameter(s), the statistic $\omega_n ^{2}$ is estimated as [e.g., \cite{Evans2008}], \begin{equation}\label{e2} \hat{\omega}_n ^{2}=\frac {1}{12n}+\sum _{i=1}^{n}\left[{\frac {2i-1}{2n}}-{\hat F}_0(x_{i:n})\right]^{2}, \end{equation} where $x_{1:n}$, $x_{2:n}$, $\cdots$, $x_{n:n}$ are sample order statistics. On the other hand, in some literature the statistic \cite{Cramer1928}, known as $L_2-$ distance,
\begin{equation}\label{e3} L_2=\int_{-\infty}^{+\infty} \left ( {\hat {F}_n(x)} - {\hat F}_0(x)\right)^2 {\mathrm {d}}x, \end{equation} has been used as a goodness-of-fit criteria to test the above mention null hypothesis, which measures the discrepancy between the empirical and theoretical distribution functions. This distance has the disadvantage that it is not distribution-free; thus if we want to apply this distance for testing goodness-of-fit, then the critical values depend on $F_0$.

My question now is how to calculate (using a program in R) this test statistics and even more difficult question how to calculate its corresponding p-value? To be more specific consider the following simulated sample

> set.seed(12345)
> x<-rexp(30,rate=2)
> hatLambda<-1/mean(x)
> hatLambda
 [1] 1.692009 #so we fit Exp. dist. with this rate to the observed sample in x      

Now I want to calculate the $L_2-$ distance, which show the distance between fitted exponential distribution and the empirical distribution, numerically. As asked before how to calculate the corresponding p-value (second priority)?

To more point please

1) I am already aware that calculating $\hat{\omega}_n ^{2}$ is so simple, but the reason for trying more difficult task when calculating $L_2-$ distance is comparing my results with some already available results.

2) when i hopefully get the answer, I will calculate $L_2-$ distance for testing available data against more complicated distributions like Pareto, lognormal, etc.

Many thanks in advance.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.