4
$\begingroup$

I want to estimate the first-order transition matrix of a sequence in discrete time, e.g.

$$ s = 1,0,1,0,1,1,0,1,0,0, \dots$$

but states are not evenly spaced in time. So that even if $s_{t=1} = 1$ and $s_{t = 2} = 0$, and $s_{t=3} = 1$ and $s_{t = 4} = 0$, the time lag between $t_{1}$ and $t_{2}$ is one day, whereas the time lag between $t_{3}$ and $t_{4}$ is one week.

How can I estimate the correct transition matrix accounting for the time lag between states? References are well accepted.

$\endgroup$
  • 4
    $\begingroup$ If the time intervals are generally of one unit (day) and sometimes a multiple of the unit (e.g. 2 days, 7 days), then just replace the transition matrix, say $\mathbf{P}$, by a suitable power of it (e.g. $\mathbf{P}^2$, $\mathbf{P}^7$). It the time design is more complex with no clear unit, then use a Continuous Time Markov Chain model. $\endgroup$ – Yves May 13 '15 at 12:10
3
+50
$\begingroup$

Taking "one day" as your time unit, consider the complete sequence $x_1,\dots,x_n$ and sum over the "missing data" days for which you don't have an observation. More specifically, if you have only information about $k$ days $I=\{i_i,\dots,i_k\}\subset\{1,\dots,n\}$, then, defining $I^c=\{1,\dots,n\}\setminus\{i_i,\dots,i_k\}$, and using the marginal $$ p(x_{i_1},\dots,x_{i_k}) = \sum_{i\in I^c}\sum_{x_i\in\{0,1\}}\left(p(x_1)\prod_{j=1}^n p(x_j\mid x_{j-1})\right), $$ you have the likelihood of the transition matrix elements based only on the observed days. Maximization of this (in general intractable) likelihood looks like a job for the EM algorithm.

$\endgroup$

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.