Class Notes for Logistic Regression –15^th and 20^th February

Reminders:

(1) Homework 3 on Design due in two weeks (Monday);

(2) First Exam is Wednesday 1^st March.

Wednesday 15^th Class

· GdLM’s: distribution and link function – for logistic, binomial distribution and logit link – gives equation (4.3)

· E.g. 4.1 – Polish girls – odds ratio, LD50, scale, residuals

· E.g. 4.2 – ECMO and respiratory illness in children – odds ratio, LD50 makes no sense here, model fits data exactly (no LOF test)

· E.g. 4.3 – Tobacco budworms – get the X scale right first, then compare two groups (M vs F), like ANOCOV, test of one common line for both genders – not a Full-and-Reduced-F-test, but a -2DLL test.

Monday 20^th Class

· E.g. 4.4 – proteins – when we use only p1, residuals tell us others may be important; stepwise tells us that p1, p2 and p4 are important predictors of MD, whereas age and p3 add nothing additional (and so need not be included). Even though 3 women are much older than the others, get very similar results when they are excluded. Focus here is on stepwise and on odds ratio interpretations.

· E.g. 4.5 – lung disease – similar to last example in that the focus here is on stepwise and on odds ratio interpretations (tougher here).

· E.g. 4.6 – pups – we overstate the case (lie) when we ignore the pup-to-pub variability; treatments might not differ here; illustrates overdispersion and random effects.

Wednesday 22^nd class: log-linear models

Monday 27^th class: proportional odds model for ordinal logits