Transcript Slide Title - Princeton University

Linear Regression
Statistical Anecdotes:
 Do hospitals make you sick?
 Student’s story
Etymology of “regression”
Andy Jacobson
July 2006
Linear Regression
Statistical Anecdotes:
 Do hospitals make you sick?
 Student’s story
 Etymology of “regression”
Andy Jacobson
July 2006
Outline
1.
2.
3.
4.
5.
6.
Discussion of yesterday’s exercise
The mathematics of regression
Solution of the normal equations
Probability and likelihood
Sample exercise: Mauna Loa CO2
Sample exercise: TransCom3 inversion
http://www.aos.princeton.edu/WWWPUBLIC/
sara/statistics_course/andy/R/
corr_exer.r
18 July practical
mauna_loa.r
Today’s first example
transcom3.r
Today’s second example
dot-Rprofile
Rename to ~/.Rprofile (i.e., home dir)
hclimate.indices.r
cov2cor.r
ferret.palette.r
Get SOI, NAO, PDO, etc. from CDC
Convert covariance to correlation
Use nice ferret color palettes
geo.axes.r
Format degree symbols, etc., for maps
load.ncdf.r
Quickly load a whole netCDF file
svd.invert.r
mat4.r
svd_invert.m
atm0_m1.mat
R-intro.pdf
faraway_pra_book.pdf
Multiple linear regression using SVD
Read and write Matlab .mat files (v4 only)
Multiple linear regression using SVD (Matlab)
Data for the TransCom3 example
Basic R documentation
Julian Faraway’s “Practical Regression and ANOVA in R”
book
Multiple Linear Regression
Basis Set
Data
Parameters
Basis Functions
“Design matrix” A gives values of each basis
function at each observation location.
Basis Functions
Observations
Note that one column of (e.g., ai1) may be all ones, to represent the “intercept”.
From the Cost Function
to the Normal Equations
“Least squares” optimization minimizes sum of squared residuals (misfits
to data). For the time being, we assume that the residuals are IID:
Expanding terms:
Cost is minimized when
derivative w.r.t. x vanishes:
Rearranging:
Optimal parameter values
(note that ATA must be
invertible):
x-hat is BLUE
BLUE = Best Linear Unbiased Estimate
(not shown here: “best”)
Practical Solution of Normal Equations using SVD
If we could pre-multiply the forward equation by A-1, the
“pseudo-inverse” of A, we could get our answer directly:
For every M x N matrix A,
there exists a singular value
decomposition (SVD):
S is diagonal and contains
the Singular Values
The columns of U and V are
orthogonal to one another:
The pseudoinverse is thus:
U is M x M
S is N x N
V is N x N
Practical Solution of Normal Equations using SVD
If we could pre-multiply the forward equation by A-1, the
“pseudo-inverse” of A, we could get our answer directly:
The pseudoinverse is:
where
Practical Solution of Normal Equations using SVD
If we could pre-multiply the forward equation by A-1, the
“pseudo-inverse” of A, we could get our answer directly:
The pseudoinverse is:
And the parameter
uncertainty covariance
matrix is:
with
Gaussian Probability and Least Squares
Residuals vector:
Observations
Predictions
Probability of ri :
Likelihood of r :
N.B.: Only true if residuals are uncorrelated (independent).
Maximum Likelihood
Log-Likelihood of r :
Goodness-of-fit: 2 for N-M degrees of freedom has a known distribution, so
regression models such as this can be judged on the probability of getting a
given value of 2.
Probability and Least Squares
• Why should we expect Gaussian
residuals?
Random Processes
z1 <- runif(5000)
Random Processes
hist(z1)
Random Processes
z1 <- runif(5000)
z2 <- runif(5000)
What is the distribution of (z1 + z2) ?
Triangular Distribution
hist(z1+z2)
Central Limit Theorem
There are more ways to get a central value than an extreme one.
Probability and Least Squares
• Why should we expect Gaussian
residuals?
(1) Because the Central Limit Theorem is on
our side.
(2) Note that the LS solution is always a
minimum variance solution, which is useful by
itself. The “maximum-likelihood” interpretation is
more of a goal than a reality.
Weighted Least Squares:
More General “Data” Errors
Minimizing the 2 is equivalent to minimizing a cost function
containing a covariance matrix C of data errors:
The data error covariance matrix is often taken to be diagonal. This
means that you put different levels of confidence on different
observations (confidence assigned by assessing both measurement error
and amount of trust in your basis functions and linear model). Note that
this structure still assumes independence between the residuals.
Covariate Data Errors
Recall cost function:
Now allow off-diagonal
covariances in C.
N.B.
ij = ji
and
ii = i2.
Multivariate normal
PDF: J propagates
without trouble into
the likelihood expression.
Minimizing J still maximizes the likelihood
Fundamental Trick for
Weighted and Generalized Least Squares
Transform system (A,b,C) with data covariance matrix C
into system (A’,b’,C’), where C’ is the identity matrix:
The Cholesky decomposition
computes a “matrix square root”
such that if R=chol(C), then C=RR.
You can then solve the Ordinary Least Squares
problem A’x = b’, using for instance the SVD method.
Note that x remains in regular, untransformed space.