www.cmm.bristol.ac.uk/forum

Hi,

I'm running a 3-level cross-classified binary logistic regression model in MLwiN 2.30 following the Leckie & Bell (2013) Module 12 practical using IGLS for starting values then MCMC (cross-classified). I have only kept variables in the model that have made a significant difference to the DIC of 5 or more(overall model fit).

The coefficients and their standard errors make sense. HOWEVER, when I run the "intervals and tests" as per p.48-49 of the Leckie and Bell practical, the "p values" for 1 d.f. that go with the chi-squared statistics don't make sense. For example: Gender (male) results: coeff. 0.168 (S.E. 0.035) chi-squared 23.6 but +/- 95% sep. p= 0.068; +/- 95% jt. 0.148. Also: Ethnicity - Non-white results: coeff. -0.713 (S.E. 0.058) chi-sq 151.6 but +/- 95% sep. p = 0.113; +/- 95% jt. 0.248.

Also the variable distance from home to school reduces the DIC by over 4000, has a chi-squared value of 2454, but the p value is shown as 0.05!!! Something's not right.

This is the third time I've tried re-sorting and re-running the model and same problem occurs each time.

The data looks fine to me. Anyone got any ideas?

Hi SCEinShef,
This is a misunderstanding of the Interval and Tests window. It will only give you a chi-squared test statistic and not a P value which you will need the tail areas window to calculate. The number +/- 95% sep is simply 1/2 the interval width i.e. in
the case of Gender you have coeff 0.168 and 95% sep interval of 0.168 +/- 0.068 (similarly for joint interval). Generally the sep interval should be ~2*the se.
I hope this helps,
Bill.

Hi,

Yes I've just run the "Intervals & Tests" video and realised it's been too many years since I did the course (2008) and used MLwiN in my PhD (2009)!

Thanks for replying anyway.

Yours somewhat embarrassed but relieved that my results are ok!!

Sue