I am trying to come up with the transition intensities of a parameterized Markov model, i.e, the elements of the transition matrix weighted by the covariates through the inverse Logit relation:

$$\Theta= \begin{bmatrix} \theta_{11}&\theta_{12}\\ \theta_{21}&\theta_{22} \end{bmatrix}$$

where as usual, the matrix sums row-wise should equal 1, and the $\theta_{ij}$ are calculated thus (let's say we have as explanatory variables an intercept, Income, and a Location dummy taking values 0 or 1):

$$\theta_{ij}=\frac{e^{\beta^{ij}_{0}+\beta^{ij}_{1}Income+\beta^{ij}_{2}LocationDummy}}{1+\sum_\limits{i\neq j}e^{\beta^{ij}_{0}+\beta^{ij}_{1}Income+\beta^{ij}_{2}LocationDummy}}$$

I've already estimated the relevant parameters through Maximum Likelihood and now need to find the $\theta_{ij}$, denoting the transition probability between states $i$ and $j$ as per the formula above to fill in the transition matrix.

The trouble is that for most values of $\beta$'s times the covariates that I include, the expression goes to either $0$ or $+\infty$. If, say, income is 500, then the expression can very easily go to one of these bounds, whereas it should be a number between 0 and 1.

I have estimated the betas, and am now plugging them in the expression for $\theta_{ij}$ to get the matrix $\Theta$.

If it helps, I am using R, and know the invlogit function, which does essentially this, but for the standard logit. Is there some function in R that does this, or does this issue have some canonical fix, like normalizing income values?


Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Browse other questions tagged or ask your own question.