I'm running a HH severity model in Emblem and one of the factors (dwelling type) is significant and time-consistent after doing some level of grouping. However, after fitting the factor the fitted average curve is actually further away to the observed average than the fitted average from the reference model (before factor was included). Can anyone explain this? Thanks in advance!
Hi Sonnyshook, No complex aliases and none of the standard errors are red. I had a look at the Cook's Pearson for potential outliers and the highest value is 0.18. I read that only residuals with a Cook's distance of > 1 should be considered for exclusion/capping.
Is the dwelling type factor weighted differently to other factors such that its pulling a lot of weight. What criteria do you use for picking a significant factor. A new factor is admissible for the following reasons: You have certainty that the factor you have just fitted and all the factor levels within the factor are not based on suspect data You can not easily draw a straight line to the model line points without crossing standard error bars It has at least one or two betas with green errors in the Beta list (less than or equal to 50%) It fits the CA better to the OBS line There is good level of exposure in the factor levels with the narrowest error bars It makes the new model have a lower akaike coefficient, BIC, etc When fitted it does not have a signifcance greater than 5% for chi squared for likeihood of equality of the reference model to the newaly fitted one If we can agree on the above then the dwelling factor is indeed significant. And we can proceed to investigate further
My guess would be you have incorrectly specified the variance function. Do any initial tests of the claim size distribution show any discontinuities or are there any anomolies in your data? You should always be able to closely model a one way effect with a simple factor in a GLM - Except ofcourse when the selected variance function is not appropriate given the emperical variance - this always affects resulting parameter estimates.
