Analysis of variance (ANOVA)
This is one of the most fundamental tools in Statistics.
It appears very simple. Indeed, it is just
a way of representing Pythagorean theorem.
Yet it has very far reaching implications both in real data application
and in concept/theory
development. ANOVA ideas reappear in such seemingly unrelated
areas like
power spectrum analysis in time series or the study of
large sample properties of
U-statistics (Efron-Stein identity) and bootstrapping,.
ANOVA when used creatively can lead to a powerful system for exploring large dimensional data.
However, unfortunately from standard textbook, one gets the feeling
that
" What ANOVA
does is mainly for using F-test of signficance "
In this course , you shall see how this is such a missleading
conclusion.
+++++++++++++++++++++++++++++++++++++++++++
(1) One-way (Program in xlis-stat, oneway.lsp
: usage (oneway-model y) ; Show
me)
variation measure : sum of squares (SS) (Tell me more in class)
Partitioning
of SS :
Total = Between
group(class, cluster) + Within group(class, cluster)
Degree of freedom
(2). Random variable version of ANOVA identity (Tell me more)
Var (Y) = Var( E(Y|X)) + E ( Var (Y|X))
(3) Two-way (Download two-way program : usage ; ( twoway-model-additive y); show me)
Additive model : response = grand mean + row effect + column effect (Tell me more in class)
Interaction model: response = grand mean + row effect + column effect + interaction
Degree of freedoms (use or absue)
Single replicate : interaction cannot be seperated
from error
no degrees of freedom left for interactions. What to
do ?
(4) Creative ways of using ANOVA :
4.1. Tukey's one degree's of freedom
test :
response = grand mean + row effect + column effect + b ( row-effect times
column-effect) (Tell me more in class)
It is very easy to fit this model. Tukey shows that
to test b=0 or not, all you have to do is to pretend that grand mean,
row and column effects
are known; so you have a simple linear regression problem of testing if
the slope is zero.
4.2. Many rows or columns : More Fun for exploring the residual patterns.
4.2a. More to plot
: for example, plot the sum of squares of residuals in the additive
model for each row ( each column) against
the corresponding column effect (the row effect, respectively).
4.2b Combined use with PCA-model
for dimension reduction:
Since the residuals are naturally arranged in the matrix form(
let's call it
the residual matrix) , we can treat it as a data matrix,
and apply PCA to reduce dimension.
Rubber's Data - revisited : (Show
me )
It turns out that the residual matrix is nearly degenerated to a
rank one matrix.
Automatic Basis curve finding (to be discussed later)
(5) Singular value decomposition of data matrix
(6) Nested structure and crossing structure (To be discussed more later)