Confidence intervals

We discussed the attached paper “Points of Significance: Error Bars”. This is important reading for just about everyone. Overlapping 95% CIs do not tell you that two coefficients are not significantly different from each other.

“In general, a gap between bars does not ensure significance, nor does overlap rule it out – it depends on the type of bar. Chances are you were surprised to learn this unintuitive result.”

GAMs

Pascale presented the work she’s done modelling data from an experimental study of tree growth. We discussed model diagnostics for GAMs and what could be considered a random effect.

Standardizing predictors

We spent a ton of time discussing the why, when, and how of centering and scaling predictors and/or response variables. Seems like an important topic to just about everyone. Maybe we should do a little workshop on it at some point.

Again, yes, you can centre binary predictors and factor predictors and this can help interpreting some models, especially if there are interactions:

• Schielzeth 2010: http://onlinelibrary.wiley.com/doi/10.1111/j.2041-210X.2010.00012.x/abstract
• http://seananderson.ca/2014/07/30/centering-interactions.html

Interactions

Natascia presented some work she’s doing where she’s fitting regression models with and without interactions. We talked about why centering predictors can be critical when interpreting models with interactions. We also talked about why you usually want to include a main effect if you include an interaction.

The beginnings of an ADMB replacement

If anyone out there is using ADMB, there’s a new R package on the block called TMB that is making waves. It has similar functionality to ADMB: https://github.com/kaskr/adcomp

I played with it a bit here: http://seananderson.ca/2014/10/17/tmb.html