Almost just as soon as I finished the NIRS workshop in Rochester, I was back on the road! This past week, I was in Princeton for a statistics workshop.
The invited speaker was Dr. Stefan Th. Gries from UC Santa Barbara. He’s a corpus linguist, so thankfully most of his examples were somewhat familiar to me. Variables like “givenness of the subject” can be a little tricky if you’re not familiar! Additionally, he came prepared with all the R code we’d need, and several sample datasets.
During the workshop, we used R to explore the sample datasets with models and plots. The main focus was the mixed-effects model selection process. This has always been confusing for me, and I sincerely appreciated the clarity of the overview. It was a great confidence-booster!
Here’s Dr. Gries’ recommended strategy:
- Formulate a model with the most complete fixed effects and random effects structure, using REML.
- Run model selection process for the random effects (i.e., simplify the random effects structure).
- Refit that model using ML.
- Run model selection process for the fixed effects (i.e., simplify the fixed effects structure).
- Refit the final model with REML again.
- Interpret the output of the final model, and run diagnostics (e.g., categorization accuracy).
- Plot fixed and random effects.
And, of course, here are some photos from the trip!