I was reading a paper in which it was assumed that $\varepsilon_1,\cdots,\varepsilon_n$ conditional on $X$ possess serial (non-linear) dependence, such that

\begin{equation} P[\varepsilon_t\geq0\mid\varepsilon_1,\cdots,\varepsilon_{t-1},X]=P[\varepsilon_t<0\mid\varepsilon_1,\cdots,\varepsilon_{t-1},X]=\frac{1}{2} \end{equation} Then the signs $s(\varepsilon_1),\cdots,s(\varepsilon_n)$ are i.i.d and distributed as $Bi(1,0.5)$. Proof: We can write the likelihood function of the signs conditional on X as \begin{eqnarray} l(s(\varepsilon_1),\cdots,s(\varepsilon_n)\mid X)&=&\prod\limits_{t=1}^{n}P[\varepsilon_t\geq0\mid\varepsilon_1,\cdots,\varepsilon_{t-1},X]^{s(\varepsilon_t)}P[\varepsilon_t<0\mid\varepsilon_1,\cdots,\varepsilon_{t-1},X]^{1-s(\varepsilon_t)}\\ &=&\left(\frac{1}{2}\right)^{s(\varepsilon_t)}\left(\frac{1}{2}\right)^{1-s(\varepsilon_t)}=\left(\frac{1}{2}\right)^n \end{eqnarray} This holds for any combination of $t=1,\cdots,n$, if there is a permutation $\pi:i\rightarrow j$ such that the earlier assumption on the conditional median holds. Now instead lets assume we are interested in the signs $s(\varepsilon_1+\beta x_1),\cdots,s(\varepsilon_n+\beta x_n)$. Intuitively, conditional on $X$, as $\beta x_1,\cdots,\beta x_n$ are constant, and since $s(\varepsilon_1),\cdots,s(\varepsilon_n)$ are independent, then the signs $s(\varepsilon_1+\beta x_1),\cdots,s(\varepsilon_n+\beta x_n)$ should also be independent. However, if we write the likelihood function, we would not observe this \begin{equation} l(s(\varepsilon_1+\beta x_1),\cdots,s(\varepsilon_n+\beta x_n)\mid X)=\\ \prod\limits_{t=1}^{n}P[\varepsilon_t\geq-\beta x_{t}\mid\varepsilon_1,\cdots,\varepsilon_{t-1},X]^{s(\varepsilon_t+\beta x_t)}P[\varepsilon_t<-\beta x_t\mid\varepsilon_1,\cdots,\varepsilon_{t-1},X]^{1-s(\varepsilon_t+\beta x_t)} \end{equation} and as before since no assumptions exists on the median of $\varepsilon_t+\beta x_t$ conditional on its own past and $X$, then the joint probabilities $P[\varepsilon_t\geq-\beta x_{t}\mid\varepsilon_1,\cdots,\varepsilon_{t-1},X]$ vary across time. Is my conclusion correct that thus, the signs $s(\varepsilon_1+\beta x_1),\cdots,s(\varepsilon_n+\beta x_n)$ cannot be concluded to be independent? Or am I missing something here?

Thanks in advance.


Your conclusion is correct. If signs $s(\varepsilon_1),\cdots,s(\varepsilon_n)$ are i.i.d and distributed as $Bi(1,0.5)$, you cannot conclude that signs $s(\varepsilon_1+\beta x_1),\cdots,s(\varepsilon_n+\beta x_n)$ are independend (or identically distributed). For example, if $\beta x_i > 0$, it is possible that $\varepsilon_n+\beta x_i$ is always positive, whatever $s(\varepsilon_i)$ is. If all elements of $\beta x$ are positive, you can randomly select signs of $s(\varepsilon_1),\cdots,s(\varepsilon_n)$, and then select these variables in such way that if $\varepsilon_{i-1}+\beta x_{i-1}<0$ then $\varepsilon_i+\beta x_i>0$.

| cite | improve this answer | |
  • $\begingroup$ I couldn't follow the relevance of your final statement to the question. Could there be some typos in it? $\endgroup$ – whuber May 26 at 13:15
  • $\begingroup$ @whuber - yes, fixed. $\endgroup$ – user31264 May 26 at 17:14

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.