-------------------------

THIS ALL NEEDS TO BE EXAMPLE DRIVEN.... (grab an exercise...)

MAYBE START WITH NORMAL MODEL AND LIKELIHOOD!!
DERIVE/ASSERT MOST THINGS FROM THAT

A bit of a better account of the mean+distrib stuff is needed here...

linreg - n=20 production run data
E[Y|X=x] = b-0 + b1x

Y = E[] + eps

var y = sig2

names yhat fitted predicted residual

"line of best fit"

LS

estim eeq's
(p 18 bottom)

final form...

std fls (mybe worth mentioning but get from matrix form)
[he doesn't derive them immed either]

CI & PI's

ANOVA table  SS's decomposition

F statistic

----

Dummy variable regression...

(a single binary dummy isn't very interesting!)

------------------------------


Chapter 3: diagnostics and transforms


anscomb data example is worked out...

Lesson: look at the data!  Look at the source!

(blather about fitting a linear model to quadratic data)

meatball derivation of h_{ij} at bottom of p 57

hii = leverage

bottom of p 56 has nice leverage discussion
  "good" and "Bad" leverage - may be a bit much!
  "good" = (lev)& (not outlier)
  "bad" = (lev) & (outlier)

(and again a discussion of quadratic fit... blwah..)

high lev = (hii> 4/n)  (avg h00 ius 2/n)
ouliers = (std resid > 2)


3.2.5 has a nice discussion of normality of errors vs residuals

sqrt(std resid plots) kindf ofg in 3.2.6 -- takes a very long time!

3.3 transfromations

nonconst variance
% effects
nonlinearity

box-cox and why to stick to common fractions....
  - can use to suggest common fractions, but might as well just try
    common fractions!

inverse response plots pp 85ff