www.cmm.bristol.ac.uk/forum

Dear colleagues,

I have a question:

I am using negative binomial regression, and i am looking to compare models, as per their goodness of fit. From reading, it seems that log likelihood would not work for neg. binomial/poisson regressions. And MCMC (dic criterion) does not work for this either on mlwin. I am supposed to use a wald's test.

I have dug around and found some examples of how this is done on Mlwin, but the combination of (a) the different conditions of those examples, (b) the un-intuitive interface for the "intervals and tests windows" for me, and (c) my lack of experience doing this are leaving me a bit stumped about what concretely to do.

Basically, I will end up with maybe 5 to 7 different models predicting my DV, and I want to find a straight-forward way to compare the models as wholes to one another (not just with or without one individual IV).

How do I use the intervals and test window to do this?

thanks in advance for any assistance you might offer,

I am attaching a screenshot of my example (my equation at the moment and intervals and tests windows).

thank you!

Dear csswader2 ,

Suppose you want to perform a Wald test of the current model against a simpler version where you omit the last two covariates in your model, fatalism and informult. This is equivalent to testing the null hypothesis that both coefficients are jointly equal to zero

H0: \beta_9 = 0 & \beta_10 = 0

against the alternative hypothesis that at least one of the two coefficients is significantly different from zero

Ha: \beta_9 != 0 | \beta_10 != 0

The null hypothesis consists of two restrictions to the model \beta_9 = 0 and \beta_10 = 0.

What you therefore need to do is two select two functions at the bottom of the Intervals and tests window (one for each restriction) and then enter a 1 in the first column in the row labelled fixed: fatalism (to indicate that you want to test whether \beta_9 = 0) and a 1 in the second column in the row labelled fixed: informult (to indicate that you want to test whether \beta_10 = 0). When you click Calc, the window will display the Chi-squared statistic and p-value associated with the joint test at the bottom of the window.

I hope that helps

Best wishes

George

Dear George,

thanks for the helpful reply! I will try it out. And does this apply here: "The degrees of freedom for the test is the difference in the number of parameters between the two models. " So that would be two?

very best,
Chris