Comments on: Tues Apr 3

By: Rex Tien

Rex Tien — Tue, 03 Apr 2012 13:24:24 +0000

Could you explain a bit more about the concept of deviance and null deviance to get a more intuitive understanding of these measures?

By: Rich Truncellito

Rich Truncellito — Tue, 03 Apr 2012 13:22:41 +0000

In modern regression models, the representation of noise or the error term isn’t explicitly stated as it is in linear regression models. To be clear, is this implied status of the error term in modern regression models so that the noise may take on other, nonlinear relationships to the deterministic parts of the models?

By: Noah

Noah — Tue, 03 Apr 2012 13:09:51 +0000

What is it about the link functions that make them typically concave which makes loglikelihood maximization easy?

By: Rob Rasmussen

Rob Rasmussen — Tue, 03 Apr 2012 13:07:05 +0000

Could you go over more in class the method for calculating the Poisson regression?
Also, why is the typical link function for Poisson regression the log function?

By: Yijuan Du

Yijuan Du — Tue, 03 Apr 2012 11:46:57 +0000

I think it may be better to give an example showing in the picture if the data follow other distribution, using the ordinary linear model may result in a big departure.

By: Ben Dichter

Ben Dichter — Tue, 03 Apr 2012 06:51:26 +0000

Why is the condition of independence between variables so important for regression?

By: Matt Panico

Matt Panico — Tue, 03 Apr 2012 06:37:30 +0000

Is there any situation in which the GLM link function would not be in the exponential family?

By: Shubham Debnath

Shubham Debnath — Tue, 03 Apr 2012 03:50:33 +0000

As I was reading I was wondering about reparameterization, so I’m glad that was the last section. How much does the initial value of the parameters affect the iterative procedure? How often will it not converge to a desired answer?

By: David Zhou

David Zhou — Tue, 03 Apr 2012 03:20:41 +0000

Can you give some very broad guidelines for choosing starting values for nonlinear least-squares problems?

By: Sharlene Flesher

Sharlene Flesher — Tue, 03 Apr 2012 03:07:38 +0000

In 14.2.2 you mention how it is helpful to have good starting values and reparameterize in solving nonlinear least squares problems- can this always be done or are there often cases that don’t have intuitive starting points and can’t easily be reparameterized?