1
$\begingroup$

I am trying to model a set of data I have physical reason to believe can be represented by a bimodal normal cumulative distribution function (Technically it is a bimodal log-normal CDF, but I think I can handle transforming the predictor by log(x) ). I can successfully find the best fit to the data using MLE, e.g. $$ \hat{y}(x;\mu_1,\sigma_1,\mu_2,\sigma_2) = \frac{1}{2} + \frac{w}{2}erf \left(\frac{x-\mu_1}{\sqrt{2}\sigma_1} \right) + \frac{(1-w)}{2}erf \left(\frac{x-\mu_2}{\sqrt{2}\sigma_2} \right) $$ However, I also want to establish a prediction interval on the data for theoretical future measurements of y at some x. The only way I have come up with to do this is to set up a GLM with what I would interpret as a sort of bimodal probit model, e.g. $$ \hat{y} = \Phi ( b_0 + b_1x ) * b_2 + \Phi ( b_3 + b_4x ) * (1-b_2) $$ However, I can't invert the CDF to get the right side to be some sort of linear function like I can with a mono-modal CDF, e.g. $$ G(\hat{y}) = \Phi^{-1}(\hat{y}) = b_0 + b_1x $$ Am I thinking about this incorrectly? Is there a way to do a bimodal CDF link function? Perhaps there is a better way of determining how well a dataset fits a given bimodal CDF and establish a prediction interval that does not use a GLM?

$\endgroup$
1
$\begingroup$

I think your chosen function may be inherently nonlinear, though it's partially linearizable (each of $(b_0,b_1)$, $(b_3,b_4)$ and $b_2$ may be made linear in a transformed model, given the others).

If you need a binomial model for your response, you might need to use generalized nonlinear models.

$\endgroup$
1
  • $\begingroup$ Thanks. I think you are right that I can't do this with a GLM. Perhaps I should just look for a better way to estimate variance of y and use that. $\endgroup$ – Thursdays Coming Sep 23 '14 at 16:58

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.