That doesn't really make much sense. Think about how you would use this model to get a prediction. If you're like "ok so x will be 5 and then since I used variance as a predictor if I want a prediction I need to plug in the variance... Uh... guys what's the variance of the point we're trying to predict?"
Well it would be more like this: measure vibrations from a beam being bent in the wind, the higher the wind speed the greater the fluctuations in bending and so the greater the variance in the data is... therefore once wind speed vs variance is plotted, could the measurement (or recording) of variance be used to estimate the wind speed hitting the beam?
PS It doesn't sound right to me and it's not my idea! But I can't quite explain why it's not right.
PS It doesn't sound right to me and it's not my idea! But I can't quite explain why it's not right.
Well, I guess anything is legitimate that makes a practical sense. In this case, physically, you are interested in the amplitude of the vibrations and the variance is a poor proxy for this. I mean, theoretically, you could have a large variance even for a smaller amplitude - so if you really care about the amplitude, that is what you would need to model.
regatds
regatds
Yes, the output variable i.e. the dependent variable, can be the variance or the standard deviation.
Have a look at gamlss and these notes (look at slide 26). Read their education papers.
Also have a look at DHGLM (double hierarchical glm) with both mean and dispersion as dependent variable by Lee Nelder and Pawitan. (e.g. at this)
These methods will make a breakthrough in the future.
Have a look at gamlss and these notes (look at slide 26). Read their education papers.
Also have a look at DHGLM (double hierarchical glm) with both mean and dispersion as dependent variable by Lee Nelder and Pawitan. (e.g. at this)
These methods will make a breakthrough in the future.