Trondheim mini-symposium

Notes for a workshop here.

Schedule

• introductions (problem descriptions?): $\approx$30 min
• data manipulation and visualization overview: 30 min
• data manip/vis lab (??), R troubleshooting: 1 hour
• break
• GLMM outline lecture: 1 hour
• GLMM lab: the rest of the morning
• lunch
• work on projects!
• present projects (??)

Data manipulation and visualization

Data manipulation

• Basic ideas: metadata and formats
• (Big data, spatial data: specialized tools (RDBMS, GIS): not covered but see Spatial task view, High-performance computing task view, R Data manual)
• Wide vs long vs relational table format (multitable package)
• Data frames
• Data manipulation in R
  • Alternative paradigms
    • Programming paradigm: [, [[, indexing
    • Logical (high-level) paradigm: subset, transform, with, merge (not attach)
  • apply operations
    • general-purpose vs special purpose (colMeans etc.); performance, readability benefit vs. remembering more specific tools
  • plyr package: especially ddply
  • Dates and times: not covered but see chron, lubridate, Datetime … always use the least complex date/time format possible

Data visualization

• Three goals:
  • exploratory: understand patterns in data (efficiently, nonparametrically)
  • diagnostic: understand departures from model (nonparametrically)
  • presentation: tell a story (elegantly but honestly: trying to match the statistical model used, but not always)
• Tools
  • for graphics generally
    • Base R (programming/canvas paradigm): mish-mosh of functions at different levels of abstraction (plot; lines; segments, rect, arrows, text). Many, many, many extensions (plotrix package).
    • grid graphics: a low-level toolbox which can be used on its own. Object- rather than canvas-oriented. Murrell book.
    • lattice: built on grid, much higher-level
    • ggplot(2): also built on grid, similar to lattice but even higher-level (and weird) (Wilkinson, Wickham books).
    • weird/advanced (not covered): dynamic/3D/alternative: rgl, rggobi, playwith, animation
  • for diagnostics:
    • model methods (e.g. plot.lm), more flexible versions in lme (xyplot.lme
    • or write your own
    • fortify in ggplot
  • Mixed model-focused approaches:
    • Pooled and fixed-effect (grouped) models

GLMMs

Reminder about GLMs

• families (i.e. distributions): Poisson, binomial, neg binom (1 and 2), Gamma, lognormal …
• overdispersion

Reminder about mixed models

• "fixed" vs "random"
• nested vs crossed

Estimation

• PQL: glmmPQL
• Laplace/AGHQ: lme4, glmmADMB
• MCMC: MCMCglmm

Inference

• Wald
• LRTs (profiles)
• Posterior densities
• IC-based approaches
  • conditional AIC
• Inference on RE variances
• Parametric bootstrap

Challenges

• Estimation glitches
  • Convergence problems
  • Zero variance estimates/Perfect correlations
• $p$ values: finite-size corrections
• Spatial structure
• Temporal structure
• Phylogenetic/pedigree structure
• Nonlinearity (GAMs)
• Ordinal data
• Zero-inflation

Data files etc.

• Banta_trondheim.pdf, Banta_trondheim.R, Banta_trondheim.Rnw
• Banta_TotalFruits.csv
• glmm_funs.R
• Banta_glmmADMB_fits.RData
• Banta_MCMCglmm_fit.RData
• glmmADMB_0.6.4.tar.gz

Raw materials

• Harvard Forest, UWO, Concordia talks
• Banta examples
• other worked examples on this site (Owls, glycera, Culcita)
• sparrows & moose examples
• Vonesh seed predation example (lab 2 from the Book)
• examples at NCEAS site: wildflowers??