Models for categorical data

• Poisson regression is directly useful for natural counts and rates
• However, we often create counts from nominal data: tables
• Loglinear modelling uses poisson regression to model the structure of tables in terms of cell counts
• The logit transformation gives us logistic regression with binary dependent variables, either
  • grouped (N successes out of M trials), or
  • individual (each individual fails or succeeds, 0 or 1 success out of one trial)
• By extension it gives us
  • multinomal logistic regression
  • ordinal logistic regression