Regression Analysis With Scilab

Regression Analysis with SCILAB
By
Gilberto E. Urroz, Ph.D., P.E.
Distributed by
i nfoClearinghouse.com
2001 Gilberto E. Urroz

All Rights Reserved
A "zip" file containing all of the programs in this document (and other
SCILAB documents at InfoClearinghouse.com) can be downloaded at the
following site:
http://www.engineering.usu.edu/cee/faculty/gurro/Software_Calculators/Scil
ab_Docs/ScilabBookFunctions.zip
The author's SCILAB web page can be accessed at:

http://www.engineering.usu.edu/cee/faculty/gurro/Scilab.html
Please report any errors in this document to: [email protected]
REGRESSION ANALYSIS
Simple linear regression

Covariance and Correlation
Additional equations and definitions
Standard error of the estimate
A function for calculating linear regression of two variables
Confidence intervals and hypothesis testing in linear regression
A function for a comprehensive linear regression analysis
An example for linear regression analysis using function linregtable
SCILAB function reglin
2
6
6
7
8
9
11
11
13
Graphical display of multivariate data
13
Multiple linear regression

Example of multiple linear regression using matrices
Covariance in multiple linear regression
Confidence intervals and hypotheses testing in multiple linear regression
Coefficient of multiple determination
A function for multiple linear regression analysis
Application of function multiplelinear
A function for predicting values from a multiple regression
16
17
18
20
22
22
25
28
Simple linear regression using function multiplelinear
29
Analysis of residuals
Scaling residuals
Influential observations
A function for residual analysis
Applications of function residuals
32
34
35
35
36
Multiple linear regression with function datafit
40
Polynomial data fitting
42
Exercises
46
Download at InfoClearinghouse.com
Regression Analysis
The idea behind regression analysis is to verify that a function
y = f(x)
fits a given data set
{(x1,y1), (x2,y2),,( xn,yn)}
after obtaining the parameters that identify function f(x). The value x represents one or more
independent variables. The function f(x) can be, for example, a linear function, i.e.,
y = mx+b, or y = b0 + b1x1 + + bkxk),

a polynomial function, i.e.,
y = b0 + b1x1 + + bpxp,
or other non-linear functions. The procedure consists in postulating a form of the function to
be fitted, y = f(x), which will depend, in general, of a number of parameters, say {b0, b1, ,
bk}. Then we choose a criteria to determine the values of those parameters. The most
commonly used is the least-square criteria, by which the sum of the squares of the errors (SSE)
involved in the data fitting is minimized. The error involved in fitting point i in the data set is
given by
ei = yi- y i ,
thus, the quantity to be minimized is
n
SSE = e = ( y i y i ) 2 .
i =1
2
i
i =1
Minimization of SSE is accomplished by taking the derivatives of SSE with respect to each of the
parameters, b0, b1, , bk, and setting these results to zero, i.e., (SSE)/b0 = 0, (SSE)/b0 = 0,
, (SSE)/bk = 0. The resulting set of equations is then solved for the values b0, b1, , bk.
After finding the parameters by minimization of the sum of square errors (SSE), we can test
hypotheses about those parameters under certain confidence levels to complete the regression
analysis. In the following section we present the regression analysis of a simple linear
regression.
Simple linear regression

Consider the data set represented in the figure below and represented by the set {(x1,y1),
(x2,y2),,( xn,yn)}. Suppose that the equation
y = mx+b,
representing a straight line in the x-y plane is used to represent the relationship between the
values x and y from the set. The fitted value of y corresponding to point xi is
y i = mxi+b,
and the corresponding error in the fitting is
ei = yi- y i = yi-(mxi+b) = yi-mxi-b.
The sum of square errors (SSE) to be minimized is

n
SSE (m, b) = ( y i mxi b) 2 .

i =1
To determine the values of m and b that minimize the sum of square errors, we use the
conditions
( SSE ) = 0
a
( SSE ) = 0
b
from which we get the so-called normal equations:

n
y
i =1
= b n + m xi
i =1
i =1
i =1
i =1
xi yi = b xi + m xi2
This is a system of linear equations with m and b as the unknowns. In matricial form, these
equations are written as
n
xi
in=1
x2
i
i =1
n
yi
m
= ni =1
.
n
b
xi
xi yi
i =1
i =1
For example, consider the data set given in the following table
x
y
1.2
6.05
2.5
11.6
4.3
15.8
8.3
21.8
11.6
36.8
The following SCILAB commands will calculate the values of m and b to minimize SSE. A plot of
the original data and the straight line fitting is also produced. The column vector p stores the
values of m and b. Thus, for this case m=2.6827095 and b=3.4404811.
-->x=[1.2,2.5,4.3,8.3,11.6];y=[6.05,11.6,15.8,21.8,36.8];
-->Sx=sum(x);Sx2=sum(x^2);Sy=sum(y);Sxy=sum(x.*y);n=length(x);
-->A=[Sx,n;Sx2,Sx];B=[Sy;Sxy];p=A\B
p
!
!
=
2.6827095 !
3.4404811 !
-->deff('[y]=yh(x)','y=p(1).*x+p(2)')
-->plot2d(xf,yf,1,'011',' ',rect)
-->plot2d(x,y,-1,'011',' ',rect)
-->xtitle('Simple linear regression','x','y')
The value of the sum of square errors for this fitting is:
-->yhat=yh(x);err=y-yhat;SSE=sum(err^2)
SSE = 23.443412
To illustrate graphically the behavior of SSE(m,b) we use the following function

SSEPlot(mrange,brange,x,y), where mrange and brange are vectors with ranges of values of m
and b, respectively, and x and y are the vectors with the original data.
function [] = SSEPlot(mrange,brange,x,y)
n=length(mrange); m=length(brange);
SSE = zeros(n,m);
deff('[y]=f(x)','y=slope*x+intercept')
for i = 1:n
for j = 1:m
slope = mrange(i);intercept=brange(j);
yhat = f(x);err=y-yhat;SSE(i,j)=sum(err^2);
end;
end;
xset('window',1);plot3d(mrange,brange,SSE,45,45,'m@b@SSE');
xtitle('Sum of square errors')
xset('window',2);contour(mrange,brange,SSE,10);
xtitle('Sum of square errors','m','b');
The function produces a three-dimensional plot of SSE(m,b) as well as a contour plot of the
function. To produce the plots we use the following SCILAB commands:
-->mr = [2.6:0.01:2.8];br=[3.3:0.01:3.5];
-->getf('SSEPlot')
-->SSEPlot(mr,br,x,y)
The following two lines modify the contour plot.

-->plot2d([p(1)],[p(2)],-9,'011',' ',[2.600 3.30 2.790 3.50])
-->xstring(p(1)+0.002,p(2)+0.002,'minimum SSE')
Covariance and Correlation

The concepts of covariance and correlation were introduced in Chapter 14 in relation to
bivariate random variables. For a sample of data points (x,y), such as the one used for the
linear regression in the previous section, we define covariance as
s xy =
1 n
( xi x)( yi y)
n 1 i =1
The sample correlation coefficient for x,y is defined as
r xy =
s xy
sx sy
where sx, sy are the standard deviations of x and y, respectively, i.e.
s x2 =
1
n 1
s y2 =
n
1
( yi y )2
n 1 i =1
( xi x ) 2
i =1
The correlation coefficient is a measure of how well the fitting equation, i.e., ^y = mx+b, fits
the given data. The values of rxy are constrained in the interval (-1,1). The closer the value of
rxy is to +1 or -1, the better the linear fitting for the given data.
Additional equations and definitions

Let's define the following quantities:
1 n
= ( xi x ) = (n 1) s = xi xi
n i =1
i =1
i =1
n
S xx
2
x
1 n
S yy = ( yi y) = (n 1) s = yi yi
n i=1
i =1
i =1
n
2
y
1 n n
Sxy = (xi x)(yi y) = (n 1) sxy = xi yi xi yi
n i=1 i=1
i=1
i=1
n
From which it follows that the standard deviations of x and y, and the covariance of x,y are
given, respectively, by
S xx
n 1
sx =
S yy
sy =
n 1
S xy
s xy =
n 1
Also, the sample correlation coefficient is
S xy
rxy =
S xx S yy
In terms ofx,y, Sxx, Syy, and Sxy, the solution to the normal equations is:
m=
S xy
S xx
s xy
s x2
b = y mx
Standard error of the estimate

The function y i = mxi+b in a linear fitting is an approximation to the regression curve of a
random variable Y on a random variable X,
Y = Mx + B + ,
where is a random error.
If we have a set of n data points (xi, yi), then we can write
Yi = Mxi + B + i,
i = 1,2,,n, where Yi are independent, normally distributed random variables with mean
( + xi) and common variance 2, and i are independent, normally distributed random
variables with mean zero and the common variance 2.
Let yi = actual data value, ^yi = mxi + b = least-square prediction of the data. Then, the
prediction error is:
ei = yi - ^yi = yi - (mxi +b).
The prediction error being an estimate of the regression error , an estimate of 2 is the socalled standard error of the estimate,
2
1 n
SSE S yy ( S xy ) / S xx n 1 2
2
s =
[ y i (mxi + b)] =
=
=
s y (1 rxy2 )
n 2 i =1
n2
n2
n2
2
e
A function for calculating linear regression of two variables

Using the definitions provided in this section, we can use the following user-defined function,
linreg, to calculate the different parameters of a simple linear regression.
The function
returns the slope, m, and intercept, b, of the linear function, the covariance, sxy, the
correlation coefficient, rxy, the mean and standard deviations of x and y (x,sx,y,sy), and the
standard error of the estimate, se. The function also produces a plot of the original data and of
the fitted equation. A listing of the function follows:
function [rxy,sxy,slope,intercept]=linreg(x,y)
n=length(x);m=length(y);
if m<>n then
error('linreg - Vectors x and y are not of the same length.');
abort;
end;
Sxx
Syy
Sxy
sx
sy
sxy
rxy
xbar
ybar
slope
intercept
se
=
=
=
=
=
=
=
=
=
=
=
=
sum(x^2)-sum(x)^2/n;
sum(y^2)-sum(y)^2/n;
sum(x.*y)-sum(x)*sum(y)/n;
sqrt(Sxx/(n-1));
sqrt(Syy/(n-1));
Sxy/(n-1);
Sxy/sqrt(Sxx*Syy);
mean(x);
mean(y);
Sxy/Sxx;
ybar - slope*xbar;
sqrt((n-1)*sy^2*(1-rxy^2)/(n-2));
xmin
= min(x);
xmax
= max(x);
xrange = xmax-xmin;
xmin
= xmin - xrange/10;
xmax
= xmax + xrange/10;
xx
= [xmin:(xmax-xmin)/100:xmax];
deff('[y]=yhat(x)','y=slope*x+intercept');
yy
= yhat(xx);
ymin
= min(y);
ymax
= max(y);
yrange = ymax - ymin;
ymin
= ymin - yrange/10;
ymax
= ymax + yrange/10;
rect
= [xmin ymin xmax ymax];
plot2d(xx,yy,1,'011',' ',rect);
xset('mark',-9,1);
plot2d( x, y,-9,'011',' ',rect);
xtitle('Linear regression','x','y');
As an example, we will use the following data set:

x
y
4.5
113
5.6
114
7.2
109
11.2
96.5
15
91.9
20
82.5
First we enter the data into vectors x and y, and then call function linreg.
-->x=[4.5,5.6,7.2,11.2,15,20];y=[113,114,109,96.5,91.9,82.5];
-->[rxy,sxy,slope,intercept] = linreg(x,y)
xbar = 10.583333
ybar = 101.15
sx = 6.0307269
sy = 12.823221
intercept = 123.38335
slope = - 2.1007891
sxy = - 76.405
rxy = - .9879955
The correlation coefficient rxy = -0.9879955 corresponds to a decreasing linear function. The
fact that the value of the correlation coefficient is close to -1 suggest a good linear fitting.
Confidence intervals and hypothesis testing in linear regression

The values m and b in the linear fitting
B in the regression curve
y i = mxi+b are approximations to the parameters M and

Y = Mx + B + .
Therefore, we can produce confidence intervals for the parameters M and B for a confidence
level . We can also perform hypotheses testing on specific values of the parameters.
Confidence limits for regression coefficients:
For the slope ():

m (t
n-2,/2)se/Sxx
< < m + (t
n-2,/2)se/Sxx,
For the intercept ():

b (t
n-2,/2)se[(1/n)+x
/Sxx]1/2 < < b + (t
n-2,/2)se[(1/n)+x
/Sxx]1/2,
where t follows the Students t distribution with = n 2, degrees of freedom, and n

represents the number of points in the sample.
Confidence interval for the mean value of Y at x = x0, i.e., mx0 + b:

[mx0+b (t
; mx0+b +(t
2
1/2
]
n-2, /2)se[(1/n)+(x0-x) /Sxx]
Limits of prediction: confidence interval for the predicted value Y0=Y(x0):
[mx0+b (t
2
1/2
n-2,/2)se[(1/n)+(x0-x) /Sxx]
2
1/2
n-2,/2)se[1+(1/n)+(x0-x) /Sxx]
; mx0+b+(t
2
1/2
]
n-2, /2)se[1+(1/n)+(x0-x) /Sxx]
Hypothesis testing on the slope, :
The null hypothesis, H0: = 0, is tested against the alternative hypothesis, H1: 0.
The test statistic is
t0 = (m-0)/(se/Sxx),
represents the number of points in the sample. The test is carried out as that of a mean
value hypothesis testing, i.e., given the level of significance, , determine the critical
value of t, t/2, then, reject H0 if t0 > t/2 or if t0 < - t/2.
If you test for the value 0= 0, and it turns out you do not reject the null hypothesis, H0:
= 0, then, the validity of a linear regression is in doubt. In other words, the sample data
does not support the assertion that 0. Therefore, this is a test of the significance of
the regression model.
Hypothesis testing on the intercept , :
The null hypothesis, H0: = 0, is tested against the alternative hypothesis, H1: 0.
t0 = (b-0)/[(1/n)+x2/Sxx]1/2,
represents the number of points in the sample. The test is carried out as that of a mean
10
value hypothesis testing, i.e., given the level of significance, , determine the critical
value of t, t/2, then, reject H0 if t0 > t/2 or if t0 < - t/2.
A function for a comprehensive linear regression analysis

The following function, linregtable, produces a comprehensive analysis of linear regression
returning not only the basic information produced by function linreg, but also including a table
of data including the fitted values, the errors, and the confidence intervals for the mean value
and the predicted values of the regression line. The function also returns estimated errors of
the slope and intercept, and performs hypotheses testing for the cases M = 0 and B = 0.
function [se,rxy,sxy,slope,intercept,sy,sx,ybar,xbar]=linreg(x,y)
n=length(x);m=length(y);
if m<>n then
error('linreg - Vectors x and y are not of the same length.');
abort;
end;
Sxx
Syy
Sxy
sx
sy
sxy
rxy
xbar
ybar
slope
intercept
se
=
=
=
=
=
=
=
=
=
=
=
=
sum(x^2)-sum(x)^2/n;
sum(y^2)-sum(y)^2/n;
sum(x.*y)-sum(x)*sum(y)/n;
sqrt(Sxx/(n-1));
sqrt(Syy/(n-1));
Sxy/(n-1);
Sxy/sqrt(Sxx*Syy);
mean(x);
mean(y);
Sxy/Sxx;
ybar - slope*xbar;
sqrt((n-1)*sy^2*(1-rxy^2)/(n-2));
xmin
= min(x);
xmax
= max(x);
xrange = xmax-xmin;
xmin
= xmin - xrange/10;
xmax
= xmax + xrange/10;
xx
= [xmin:(xmax-xmin)/100:xmax];
deff('[y]=yhat(x)','y=slope*x+intercept');
yy
= yhat(xx);
ymin
= min(y);
ymax
= max(y);
yrange = ymax - ymin;
ymin
= ymin - yrange/10;
ymax
= ymax + yrange/10;
rect
= [xmin ymin xmax ymax];
plot2d(xx,yy,1,'011',' ',rect);
xset('mark',-9,1);
plot2d( x, y,-9,'011',' ',rect);
xtitle('Linear regression','x','y');
An example for linear regression analysis using function linregtable

For an application of function linregtable consider the following (x,y) data.
significance level of 0.05.
11
We use a
x
y
2.0
5.5
2.5
7.2
3.0
9.4
3.5
10.0
4.0
12.2
The following SCILAB commands are used to load the data and perform the regression analysis:
-->getf('linregtable')
-->x=[2.0,2.5,3.0,3.5,4.0];y=[5.5,7.2,9.4,10.0,12.2];
-->linregtable(x,y,0.05)
Regression line:
y
Significance level
Value of t_alpha/2
Confidence interval for slope
Confidence interval for intercept
Covariance of x and y
Correlation coefficient
Standard error of estimate
Standard error of slope
Standard error of intercept
Mean values of x and y
Standard deviations of x and y
Error sum of squares
=
=
=
=
=
=
=
=
=
=
=
=
=
3.24*x + -.86
.05
3.18245
[2.37976;4.10024]
[-3.51144;1.79144]
2.025
.98972
.42740
.27031
.83315
3 8.86
.79057 2.58805
.548
--------------------------------------------------------------------------------------x
y
^y
error
C.I. mean
C.I. predicted
--------------------------------------------------------------------------------------2
5.5
5.62
-.12
4.56642
6.67358
3.89952
7.34048
2.5
7.2
7.24
-.04
6.49501
7.98499
5.68918
8.79082
3
9.4
8.86
.54
8.25172
9.46828
7.37002
10.35
3.5
10
10.48
-.48
9.73501
11.225
8.92918
12.0308
4
12.2
12.1
.1
11.0464
13.1536
10.3795
13.8205
---------------------------------------------------------------------------------------
Reject the null hypothesis H0:Slope = 0.

Test parameter for hypothesis testing on the slope (t) = 11.9863
Do not reject the null hypothesis H0:Intercept = 0.
Test parameter for hypothesis testing on the intercept (t) =
-.44117
The plot of the original data and the fitted data, also produced by function linregtable, is
shown next:
12
The graph shows a good linear fitting of the data confirmed by a correlation coefficient
(0.98972) very close to 1.0. The hypotheses testing indicate that the null hypothesis H0:b = 0
cannot be rejected, i.e., a zero intercept may be substituted for the intercept of -0.86 with a
95% confidence level. On the other hand, the null hypothesis H0:m=0 is rejected, indicating a
proper linear relationship.
SCILAB function reglin

SCILAB provides function reglin, with call: [m,b,sig] = reglin(x,y), which returns the values of
the slope, m, the intercept, b, and the standard deviation of the residual, , for the linear
fitting y =mx+b. For the data of the example above, using reglin we obtain:
--> [m,b,sig] = reglin(x,y)
sig = 0.3310589, b = -0.85, m = 3.24
Graphical display of multivariate data

In the next section we will present techniques of analysis for multiple linear regression in
which we use fittings of the form y = b0 + b1x1 + b2x2 + b3x3 + + bnxn. Before we present
the details of the analysis, however, we want to introduce a simple way to visualize
relationships between pairs of variables in a multivariate data set. The proposed graph is an
array of plots representing the relationships between independent variables xi and xj, for ij,
as well as the relationship of the dependent variable y and each of the independent variables
xi. Function multiplot, which takes as input a matrix X whose columns are values of the
independent variables, and a (row) vector y, which represents the dependent variable,
produces such array of plots. A listing of the function multiplot follows next.
function [] = multiplot(X,y)
//Produces a matrix of plots:
//
//
//
//
---x1---x2-vs-x1
.
y-vs-x1
x1-vs-x2
---x2--.
y-vs-x2
x1-vs-x3 ...
x2-vs-x3 ...
.
y-vs-x3 ...
x1-vs-y
x2-vs-y
.
---y---
[m n] = size(X);
nr = n+1; nc = nr;
XX = [X y'];
xset('window',1);
xset('default');
xbasc();
xset('mark',-1,1);
for i = 1:nr
for j = 1:nc
mtlb_subplot(nr,nc,(i-1)*nr+j);
if i <> j then
rect= [min(XX(:,j)) min(XX(:,i)) max(XX(:,j)) max(XX(:,i))];
plot2d(XX(:,j),XX(:,i),-1);
13
if i==nr & j == nc then

xtitle(' ','y','y');
elseif i==nr then
xtitle(' ','x'+string(j),'y');
elseif j==nc then
xtitle(' ','y','x'+string(i));
else
xtitle(' ','x'+string(j),'x'+string(i))
end;
end;
end;
end;
xset('font',2,5);
for i = 1:nr
for j = 1:nc
mtlb_subplot(nr,nc,(i-1)*nr+j);
if i==j then
plot2d([0],[0],1,'010',' ',[0 0 10 10]);
if i==nr & j==nc then
xstring(3,5,'y');
else
xstring(3,5,'x'+string(i));
end;
end;
end;
end;
To sub-divide the plot window into subplots, function multiplot uses function mtlb_subplot, a
function that emulates Matlabs function subplot (which explains the prefix mtlb_). Details of
this, and other functions with the mtlb_ prefix are presented in more detail in Chapter 20.
To illustrate the use of function multiplot we will use the following data set:
____________________
x1
x2
y
____________________
2.3 21.5 147.47
3.2 23.2 165.42
4.5 24.5 170.60
5.1 26.2 184.84
6.2 27.1 198.05
7.5 28.3 209.96
____________________
The SCILAB commands to produce the plot array are shown next.
-->x1 = [2.3 3.2 4.5 5.1 6.2 7.5];
-->x2 = [21.5 23.2 24.5 26.2 27.1 28.3];
-->y
= [147.47 165.42 170.60 184.84 198.05 209.96];
-->X=[x1' x2'];
-->getf('multiplot')
-->multiplot(X,y)
14
__________________________________________________________________________________
__________________________________________________________________________________
A result like this array of plots is useful in determining some preliminary trends among the
variables. For example, the plots above show strong dependency between x1 and x2, besides
the expected dependency of y on x1 or y on x2. In that sense, variables x1 and x2 are not
independent of each other. When we refer to them as the independent variables, the meaning
is that of variables that explain y, which is, in turn, referred to as the dependent variable.
15

The subject of multiple linear regression was first introduced in Chapter 5 as an example of
applications of matrix operations. For multiple linear regression fitting consider a data set of
the form
x1
x11
x2
x21
x3
x31
x4
x5
xn1
y
y1
x12
x22
x32
xn2
y2
x13
x32
x33
xn3
y3
x1,m-1
x1,m
.
x
2,m-1
2,m
3,m-1
3,m
n,m-1
ym-1
n,m
ym
to which we fit an equation of the form
y = b0 + b1x1 + b2x2 + b3x3 + + bnxn.

If we write the independent variables x1, x2, , xn into a row vector, i.e., xi = [x1i x2i xni], and
the coefficients b0 b1 b2 bn into a column vector b = [b0 b1 b2 b3 bn]T, we can write
y i = xib.
If we put together the matrix, X = [x1 x2
_
1
x11
x21
1
x12
x22
1
x13
x32
.
.
.
.
.
.
1
x1,m
x 2,m
_
and the vector,
xn]T, i.e.,
x31
x32
x33
.
.
x 3,m
xn1
xn2
xn3
.
.
x n,m
_
y = [ y1 y 2 .... y n ]T , we can summarize the original table into the expression

y = X b.
The error vector is

e=y-
y ,
and the sum of squared erros, SSE, is

SSE = eT e = (y -
y )T(y - y ) = (y-Xb)T(y-Xb).
To minimize SSE we write (SSE)/b = 0. It can be shown that this results in the expression
XT X b = XTy, from which it follows that
b = (XT X)-1 XTy
16
An example for calculating the vector of coefficients b for a multiple linear regression was
presented in Chapter 5. The example is repeated here to facilitate understanding of the
procedure.
Example of multiple linear regression using matrices

use the following data to obtain the multiple linear fitting
y = b0 + b1x1 + b2x2 + b3x3,
x1
1.20
2.50
3.50
4.00
6.00
x2
3.10
3.10
4.50
4.50
5.00
x3
2.00
2.50
2.50
3.00
3.50
y
5.70
8.20
5.00
8.20
9.50
With SCILAB you can proceed as follows:

First, enter the vectors x1, x2, x3, and y, as row vectors:
->x1 = [1.2,2.5,3.5,4.0,6.0]
x1 = !
1.2
2.5
3.5
4.
-->x2 = [3.1,3.1,4.5,4.5,5.0]
x2 = !
3.1
3.1
4.5
4.5
-->x3 = [2.0,2.5,2.5,3.0,3.5]
x3 = !
2.
2.5
2.5
3.
-->y = [5.7,8.2,5.0,8.2,9.5]
y = !
5.7
8.2
5.
8.2
6. !
5. !
3.5 !
9.5 !
Next, we form matrix X:

-->X = [ones(5,1) x1' x2' x3']
X =
!
!
!
!
!
1.
1.
1.
1.
1.
1.2
2.5
3.5
4.
6.
3.1
3.1
4.5
4.5
5.
2.
2.5
2.5
3.
3.5
!
!
!
!
!
The vector of coefficients for the multiple linear equation is calculated as:
-->b =inv(X'*X)*X'*y
b =
! - 2.1649851 !
! - .7144632 !
! - 1.7850398 !
!
7.0941849 !
17
Thus, the multiple-linear regression equation is:

y^ = -2.1649851-0.7144632x1 -1.7850398x2 +7.0941849x3.
This function can be used to evaluate y for values of x given as [x1,x2,x3]. For example, for
[x1,x2,x3] = [3,4,2], construct a vector xx = [1,3,4,2], and multiply xx times b, to obtain y(xx):
-->xx = [1,3,4,2]
xx = !
1.
3.
4.
2. !
-->xx*b
ans = 2.739836
The fitted values of y corresponding to the values of x1, x2, and x3 from the table are obtained
from y = Xb:
-->X*b
ans =
!
5.6324056
!
8.2506958
!
5.0371769
!
8.2270378
!
9.4526839
!
!
!
!
!
Compare these fitted values with the original data as shown in the table below:
x1
1.20
2.50
3.50
4.00
6.00
x2
3.10
3.10
4.50
4.50
5.00
x3
2.00
2.50
2.50
3.00
3.50
y
5.70
8.20
5.00
8.20
9.50
y-fitted
5.63
8.25
5.04
8.23
9.45
This procedure will be coded into a user-defined SCILAB function in an upcoming section
incorporating some of the calculations for the regression analysis as shown next.
An array of plots showing the dependency of the different variables involved in the multiple
linear fitting is shown in the following page. It was produced by using function multiplot.
-->multiplot(X,y);
(See plot in next page).
Covariance in multiple linear regression

In simple linear regression we defined a standard error of the estimate, se, as an approximation
to the variance of the distribution of errors. The standard error of the estimate, also known
as the mean square error, for multiple linear regression is defined as
18
se = MSE = SSE/(m-n-1),
where m is the number of data points available and n is the number of coefficients required for
the multiple linear fitting.
The matrix C = se2(XTX)-1 is a symmetric matrix known as the covariance matrix. The diagonal
elements cii are the variances associated with each of the coefficients bi, i.e.,
Var(bi) = cii,
while elements off the diagonal, cij, ij, are the covariances of bi and bj, i.e.,
Cov(bi,bj) = cij, ij.
The square root of the variances Var(bi) are referred to as the standard error of the estimate
for each coefficient, i.e.,
se(bi) = cii = [Var(bi)]1/2.
__________________________________________________________________________________
__________________________________________________________________________________
19
Confidence intervals and hypotheses testing in multiple linear

regression
The multiple linear relationship defined earlier, namely,
y = b0 + b1x1 + b2x2 + b3x3 + + bnxn,

is an approximation to the multiple linear model
Y = 0 + 1x1 + 2x2 + 3x3 + + nxn + ,
where Y is a normally-distributed random variable with mean y , is also normally-distributed
with mean zero. The standard deviation of the distributions of both Y and is .
An
approximation to the value of is the standard error of the estimate for the regression, se,
defined earlier.
Confidence intervals for the coefficients
Using a level of confidence , we can write confidence intervals for each of the
coefficients i in the linear model for Y, as
bi (t
m-n,/2)se(bi)
< i < bi + (t
m-n,/2)
se(bi) ,
for i=1,2,,n, where bi is the i-th coefficient in the linear fitting, tm-n,/2 is the value of the
Students t variable for = m-n degrees of freedom corresponding to a cumulative
probability of 1-/2, and se(bi) is the standard error of the estimate for bi.
Confidence interval for the mean value of Y at x = x0, i.e., x0Tb:

[x0Tb (t
[x0 T C x0]1/2 ; x0Tb + (t
m-n,/2)
m-n,/2)[x0
C x0]1/2 2]
where C is the covariance matrix.
Limits of prediction: confidence interval for the predicted value Y0=Y(x0):

[x0Tb (t
m-n,/2)se[1+x0
(C/se) x0]1/2 ; x0Tb + (t m-n,/2)se[1+x0 T (C/se) x0]1/2 2]
Hypothesis testing on the coefficients, i:
The null hypothesis, H0: i = 0, is tested against the alternative hypothesis, H1: i 0.
t0 = (bi-0)/se(bi)
where t follows the Students t distribution with = m-n, degrees of freedom. The test is
carried out as that of a mean value hypothesis testing, i.e., given the level of significance,
, determine the critical value of t, t/2, then, reject H0 if t0 > t/2 or if t0 < - t/2. Of
interest, many times, is the test that a particular coefficient bi be zero, i.e., H0: i = 0.
20
Hypothesis testing for significance of regression

This test is aimed at determining if a linear regression indeed exists between y and x. The
null hypothesis to be tested is H0: 1=2=3 ==n=0, against the alternative hypothesis,
H1:i 0 for at least one value of i. The appropriate test to be conducted is an F test
based on the test parameter
F0 =
SSR / n
MSR
=
,
SSE /(m n 1) MSE
where MSR = SSR/n is known as the regression mean square, MSE = SSE/(m-n-1) = se2 is the
mean square error, SSE is the sum of squared errors (defined earlier), and SSR is the
regression sum of squares defined as
m
SSR = ( y i y ) 2 ,
i =1
wherey is the mean value of y.

The parameter F0 follows the F distribution with n
degrees of freedom in the numerator and m-n-1 degrees of freedom in the denominator.
For a given confidence level , we will determine the value F from the appropriate F
distribution, and reject the null hypothesis H0 if F0 > F.
Analysis of variance for testing significance of regression

Analysis of variance is the name to the general method that produces the parameters
described above for the testing of significance of regression. The method is based on the
so-called analysis of variance identity,
SST = SSR + SSE,
where SSE and SSR were described above and SST, the total corrected sum of squares, is
described as
m
SST = ( y i y ) 2 .
i =1
The term SSR accounts for the variability in yi due to the regression line, while the terms
SSE accounts for the residual variability not incorporated in SSR. The term SSR has n
degrees of freedom, i.e., the same number of coefficients in the multiple linear fitting.
The term SSE has m-n-1 degrees of freedom, while SST has n-1 degrees of freedom.
21
Analysis of variance for the test of significance of regression is typically reported in a table
that includes the following information:
Variation source Sum of squares Degrees of freedom Mean square

F0
SSR
n
MSR
MSR/MSE
Regression
SSE
m-k-1
MSE
Residual/error
SST
m-1
Total
Coefficient of multiple determination

The coefficient of multiple determination R2 is defined in terms of SST, SSR, and SSE, as
R2 =
SSR
SSE
= 1
,
SST
SST
while the positive square root of this value is referred to as the multiple correlation
coefficient, R. This multiple correlation coefficient, for the case of a simple linear regression,
is the same as the correlation coefficient rxy. Values of R2 are restricted to the range [0,1].
Unlike the simple correlation coefficient, rxy, the coefficient of multiple determination, R2, is
not a good indicator of linearity. A better indicator is the adjusted coefficient of multiple
determination,
2
Radj
= 1
SSE /(m n 1)
.
SST /( m 1)
A function for multiple linear regression analysis

The following function, multiplelinear, includes the calculation of the coefficients for a
multiple linear regression equation, their standard error of estimates, their confidence
interval, and the recommendation for rejection or not rejection for the null hypotheses H0:i =
0. The function also produces a table of the values of y, the fitted values y , the errors, and
the confidence intervals for the mean linear regression Y, and for the predicted linear
regression. The analysis-of-variance table is also produced by the function. Finally, the
function prints the values of the standard error of the estimate, se, the coefficient of multiple
determination, the multiple correlation coefficient, the adjusted coefficient of multiple
determination, and the covariance matrix. The function returns the vector of coefficients, b,
the covariance matrix, cov(xi,xj), and the standard error of the estimate, se. The arguments
of the function are a matrix XA whose columns include the column data vectors x1, x2, , xn, a
column vector y containing the dependent variable data, and a level of confidence, alpha
(typical values are 0.01, 0.05, 0.10).
22
A listing of the function follows:

function [b,C] = multiplelinear(XA,y,alpha)
[m n] = size(XA);
X = [ones(m,1) XA];
b=inv(X'*X)*X'*y;
yh = X*b;
e=y-yh;
SSE=e'*e;
MSE = SSE/(m-n-1);
se = sqrt(MSE);
C = MSE*inv(X'*X);
[nC mC]=size(C);
seb = [];
//Size of original X matrix

//Augmenting matrix X
//Coefficients of function
//Fitted value of y
//Errors or residuals
//Sum of squared errors
//Mean square error
//Standard error of estimate
//Covariance matrix
//Standard errors for coefficients
for i = 1:nC
seb = [seb; sqrt(C(i,i))];
end;
ta2 = cdft('T',m-n,1-alpha/2,alpha/2); //t_alpha/2
sY = []; sYp = [];
//Terms involved in C.I. for Y, Ypred
for i=1:m
sY = [sY; sqrt(X(i,:)*C*X(i,:)')];
sYp = [sYp; se*sqrt(1+X(i,:)*(C/se)*X(i,:)')];
end;
CIYL = yh-sY;
//Lower limit for C.I. for mean Y
CIYU = yh+sY;
//Upper limit for C.I. for mean Y
CIYpL = yh-sYp;
//Lower limit for C.I. for predicted Y
CIYpU = yh+sYp;
//Upper limit for C.I. for predicted Y
CIbL = b-ta2*seb;
//Lower limit for C.I. for coefficients
CIbU = b+ta2*seb;
//Upper limit for C.I. for coefficients
t0b = b./seb;
//t parameter for testing H0:b(i)=0
decision = [];
//Hypothesis testing for H0:b(i)=0
for i = 1:n+1
if t0b(i)>ta2 | t0b(i)<-ta2 then
decision = [decision; ' reject
'];
else
decision = [decision; ' do not reject'];
end;
end;
ybar = mean(y);
SST = sum((y-ybar)^2);
SSR = sum((yh-ybar)^2);
MSR
MSE
F0
Fa
=
=
=
=
//Mean value of y
//Total sum of squares
//Residual sum of squares
SSR/n;
//Regression mean square
SSE/(m-n-1);
//Error mean square
MSR/MSE;
//F parameter for significance of regression
cdff('F',n,m-n-1,1-alpha,alpha);
//F_alpha
R2 = 1-SSE/SST; R = sqrt(R2);
R2a = 1-(SSE/(m-n-1))/(SST/(m-1));
//Coeff. of multiple regression

//Adj. Coeff. of multiple regression
//Printing of results
printf(' ');
printf('Multiple linear regression');
printf('==========================');
printf(' ');
printf('Table of coefficients');
printf('------------------------------------------------------------------------');
23
printf('
i
b(i)
se(b(i))
Lower
Upper
t0
H0:b(i)=0');
printf('------------------------------------------------------------------------');
for i = 1:n+1
printf('%4.0f %10g %10g %10g %10g %10g '+decision(i),...
i-1,b(i),seb(i),CIbL(i),CIbU(i),t0b(i));
end;
printf('------------------------------------------------------------------------');
printf('
t_alpha/2 = %g',ta2);
printf('------------------------------------------------------------------------');
printf(' ');printf(' ');
printf('Table of fitted values and errors');
printf('--------------------------------------------------------------------------------');
printf('
i
y(i)
yh(i)
e(i)
C.I. for Y
C.I.
for Ypred');
printf('--------------------------------------------------------------------------------');
for i = 1:m
printf('%4.0f %10.6g %10.6g %10.6g %10.6g %10.6g %10.6g %10.6g',...
i,y(i),yh(i),e(i),CIYL(i),CIYU(i),CIYpL(i),CIYpU(i));
end;
printf('--------------------------------------------------------------------------------');
printf(' ');printf(' ');
printf('Analysis of variance');
printf('--------------------------------------------------------');
printf('Source of
Sum of
Degrees of
Mean')
printf('variation
squares
freedom
square
F0');
printf('--------------------------------------------------------');
printf('Regression
%10.6g %10.0f %10.6g %10.6g',SSR,n,MSR,F0');
printf('Residual
%10.6g %10.0f %10.6g
',SSE,m-n-1,MSE);
printf('Total
%10.6g %10.0f
',SST,m-1);
printf('--------------------------------------------------------');
printf('With F0 = %g and F_alpha = %g,',F0,Fa);
if F0>Fa then
printf('reject the null hypothesis H0:beta1=beta2=...=betan=0.');
else
printf('do not reject the null hypothesis H0:beta1=beta2=...=betan=0.');
end;
printf('--------------------------------------------------------');
disp(' ');
printf('Additional information');
printf('---------------------------------------------------------');
printf('Standard error of estimate (se)
= %g',se);
printf('Coefficient of multiple determination (R^2)
= %g',R2);
printf('Multiple correlation coefficient (R)
= %g',R);
printf('Adjusted coefficient of multiple determination = %g',R2a);
printf('---------------------------------------------------------');
printf(' ');
printf('Covariance matrix:');
disp(C);
printf(' ');
printf('---------------------------------------------------------');
//Plots of residuals - several options
24
for j = 1:n
xset('window',j);xset('mark',-9,2);xbasc(j);
plot2d(XA(:,j),e,-9)
xtitle('Residual plot - error vs. x'+string(j),'x'+string(j),'error');
end;
xset('window',n+1);xset('mark',-9,2);
plot2d(y,e,-9);
xtitle('Residual plot - error vs. y','y','error');
xset('window',n+2);xset('mark',-9,2);
plot2d(yh,e,-9);
xtitle('Residual plot - error vs. yh','yh','error');
Application of function multiplelinear

Consider the multiple linear regression of the form y = b0+b1x1+b2x2 for the data shown in the
following table:
x2
x3
y
x1
1.20
3.10
2.00
5.70
2.50
3.10
2.50
8.20
3.50
4.50
2.50
5.00
4.00
4.50
3.00
8.20
6.00
5.00
3.50
9.50
First, we load the data:
-->x1=[1.2,2.5,3.5,4.0,6.0];x2=[3.1,3.1,4.5,4.5,5.0];
-->x3=[2.0,2.5,2.5,3.0,3.5];y =[5.7,8.2,5.0,8.2,9.5];
-->X=[x1' x2' x3']
X =
!
1.2
3.1
2. !
!
2.5
3.1
2.5 !
!
3.5
4.5
2.5 !
!
4.
4.5
3. !
!
6.
5.
3.5 !
-->y=y'
y =
!
5.7 !
!
8.2 !
!
5. !
!
8.2 !
!
9.5 !
Then, we call function multiplelinear to obtain information on the multiple linear regression:
-->[b,C,se]=multiplelinear(X,y,0.1);
==========================
Table of coefficients
------------------------------------------------------------------------i
b(i)
se(b(i))
Lower
Upper
t0 H0:b(i)=0
------------------------------------------------------------------------0
-2.16499
1.14458
-5.50713
1.17716
-1.89152 do not reject
1
-.71446
.21459
-1.34105 -.0878760
-3.3295 reject
2
-1.78504
.18141
-2.31477
-1.25531
-9.83958 reject
3
7.09418
.49595
5.64603
8.54234
14.3043 reject
------------------------------------------------------------------------t_alpha/2 = 2.91999
-------------------------------------------------------------------------
25
Table of fitted values and errors

------------------------------------------------------------------------------i
y(i)
yh(i)
e(i)
C.I. for Y
C.I. for Ypred
------------------------------------------------------------------------------1
5.7
5.63241
.0675944
5.54921
5.7156
5.5218
5.74301
2
8.2
8.2507 -.0506958
8.15624
8.34515
8.13913
8.36226
3
5
5.03718 -.0371769
4.93663
5.13772
4.92504
5.14931
4
8.2
8.22704 -.0270378
8.12331
8.33077
8.11459
8.33949
5
9.5
9.45268
.0473161
9.3565
9.54887
9.34096
9.56441
------------------------------------------------------------------------------Analysis of variance
-------------------------------------------------------Source of
Sum of
Degrees of
Mean
variation
squares
freedom
square
F0
-------------------------------------------------------Regression
14.2965
3
4.7655
414.714
Residual
.0114911
1
.0114911
Total
14.308
4
-------------------------------------------------------With F0 = 414.714 and F_alpha = 53.5932,
reject the null hypothesis H0:beta1=beta2=...=betan=0.
--------------------------------------------------------
Additional information
--------------------------------------------------------Standard error of estimate (se)
= .10720
Coefficient of multiple determination (R^2)
= .99920
Multiple correlation coefficient (R)
= .99960
Adjusted coefficient of multiple determination = .99679
---------------------------------------------------------
Covariance matrix:
!
1.3100522
!
.2388806
! - .1694459
! - .5351636
.2388806
.0460470
.0313405
.1002469
.1694459
.0313405
.0329111
.0534431
.5351636
.1002469
.0534431
.2459639
!
!
!
!
The results show that, for a confidence level = 0.1, the hypothesis H0:0 = 0 may not be
rejected, meaning that you could eliminate that term from the multiple linear fitting. On the
other hand, the test for significance of regression indicates that we cannot reject the
hypothesis that a linear regression does exist. Plots of the residuals against variables x1, x2, x3,
y, and ^y are shown next.
26
27
A function for predicting values from a multiple regression

The following function, mlpredict, takes as arguments a row vector x containing values of the
independent variables [x1 x2 xn] in a multiple linear regression, b is a column vector
containing the coefficients of the linear regression, C is the corresponding covariance matrix,
se is the standard error of the estimate, and alpha is the level of confidence. The function
returns the predicted value y and prints the confidence intervals for the mean value Y and for
the predicted value of Y.
function [y] = mlpredict(x,b,C,se,alpha)
nb = length(b); nx = length(x);
if nb<>nx+1 then
error('mlpredict - Vectors x and b are of incompatible length.');
abort;
else
n = nx;
end;
[nC mC] = size(C);
if nC<>mC then
error('mlpredict - Covariance matrix C must be a square matrix.');
28
abort;
elseif nC<>n+1 then
error('mlpredict - Dimensions of covariance matrix incompatible with vector
b.');
abort;
end;
xx = [1 x];
y = xx*b;
CIYL = y CIYU = y +
CIYpL = y CIYpU = y +
sqrt(xx*C*xx');
sqrt(xx*C*xx');
se*sqrt(1+xx*(C/se)*xx');
se*sqrt(1+xx*(C/se)*xx');
//augment vector x
//calculate y
//Lower limit C.I.
//Upper limit C.I.
//Lower limit C.I.
//Upper limit C.I.
for
for
for
for
mean Y
mean Y
predicted Y
predicted Y
//Print results
printf(' ');
disp(' ',x,'For x = ');
printf('Multiple linear regression prediction is y = %g',y);
printf('Confidence interval for mean value Y
= [%g,%g]',CIYL,CIYU);
printf('Confidence interval for predicted value Y = [%g,%g]',CIYpL,CIYpU);
An application of function mlpredict, using the values of b, C, and se obtained from function
multiplelinear as shown above, is presented next:
-->y=mlpredict(x,b,C,se,0.1);
For x =
!
2.
3.5
2.8 !
Multiple linear regression prediction is y = 10.0222

Confidence interval for mean value Y
= [9.732,10.3123]
Confidence interval for predicted value Y = [9.87893,10.1654]
Simple linear regression using function multiplelinear

Function multiplelinear can be used to produce the regression analysis for a simple linear
regression as illustrated in the following example. The data to be fitted is given in the
following table:
x
1.02
1.05
1.25
1.34
1.49
1.44
0.94
1.30
1.59
1.40
y
90.02
89.14
91.48
93.81
96.77
94.49
87.62
91.78
99.43
94.69
29
A plot of the data is produced with:

-->plot2d(x,y,-1);xtitle(Simple linear regression example, x,y);
The regression analysis using function multiplelinear follows:
-->[b,C,se] = multiplelinear(X,Y,0.05);
==========================
------------------------------------------------------------------------i
b(i)
se(b(i))
Lower
Upper
t0 H0:b(i)=0
------------------------------------------------------------------------0
72.2564
2.03849
67.645
76.8678
35.446 reject
1
16.1206
1.5701
12.5688
19.6724
10.2672 reject
------------------------------------------------------------------------t_alpha/2 = 2.26216
-------------------------------------------------------------------------
Table of fitted values and errors

------------------------------------------------------------------------------i
y(i)
yh(i)
e(i)
C.I. for Y
C.I. for Ypred
------------------------------------------------------------------------------1
90.02
88.6994
1.32059
88.1769
89.2219
87.552
89.8468
2
89.14
89.183 -.0430274
88.6967
89.6694
88.052
90.3141
3
91.48
92.4071
-.92714
92.081
92.7333
91.3363
93.4779
4
93.81
93.858 -.0479932
93.5232
94.1928
92.7845
94.9315
5
96.77
96.2761
.49392
95.8173
96.7349
95.1568
97.3953
6
94.49
95.4701
-.98005
95.0634
95.8767
94.3715
96.5686
7
87.62
87.4098
.21024
86.7835
88.036
86.2107
88.6089
8
91.78
93.2132
-1.43317
92.8897
93.5366
92.1432
94.2831
9
99.43
97.8881
1.54186
97.307
98.4692
96.7124
99.0639
10
94.69
94.8252
-.13523
94.4535
95.1969
93.7394
95.9111
-------------------------------------------------------------------------------
30
Analysis of variance
-------------------------------------------------------Source of
Sum of
Degrees of
Mean
variation
squares
freedom
square
F0
-------------------------------------------------------Regression
109.448
1
109.448
105.416
Residual
8.30597
8
1.03825
Total
117.754
9
-------------------------------------------------------With F0 = 105.416 and F_alpha = 5.31766,
-------------------------------------------------------Additional information
= 1.01894
= .92946
= .96409
--------------------------------------------------------Covariance matrix:
!
4.1554489
! - 3.1603934
- 3.1603934 !
2.4652054 !
---------------------------------------------------------
Compare the results with those obtained using function linreg:

-->[se,rxy,sxy,m,b,sy,sx,ybar,xbar]=linreg(x,y);
-->[m,b]
ans = !
16.120572
72.256427 !
-->se
se = 1.0189435
-->rxy
rxy = .9640868
The data fitting produced with linreg is shown in the next figure:
31
Analysis of residuals
Residuals are simply the errors between the original data values y and the fitted values
i.e., the values
e = y-
y ,
y .
Plots of the residuals against each of the independent variables, x1, x2, , xn, or against the
original data values y or the fitted values y can be used to identify trends in the residuals. If
the residuals are randomly distributed about zero, the plots will show no specific pattern for
the residuals. Thus, residual analysis can be used to check the assumption of normal
distribution of errors about a zero mean. If the assumption of normal distribution of residuals
about zero does not hold, the plots of residuals may show specific trends.
For example, consider the data from the following table:
__________________________________________________________
x2
p
q
r
x1
__________________________________________________________
1.1
22.1
1524.7407
3585.7418
1558.505
2.1
23.2
1600.8101
3863.4938
1588.175
3.4
24.5
1630.6414
4205.4344
1630.79
4.2
20.4
1416.1757
3216.1131
1371.75
5.5
25.2
1681.7725
4409.4029
1658.565
6.1
23.1
1554.5774
3876.8094
1541.455
7.4
19.2
1317.4763
2975.0054
1315.83
8.1
18.2
1324.6139
2764.6509
1296.275
9.9
20.5
1446.5163
3289.158
1481.265
11.
19.1
1328.9309
2983.4153
1458.17
__________________________________________________________
The table shows two independent variables, x1 and x2, and three dependent variables, i.e., p =
p(x1,x2), q = q(x1,x2), and r = r(x1,x2). We will try to fit a multiple linear function to the three
functions, e.g., p = b0 + b1x1 + b2x2, using function multiplelinear, with the specific purpose of
checking the distribution of the errors. Thus, we will not include in this section the output for
the function calls. Only the plots of residuals (or errors) against the fitted data will be
presented. The SCILAB command required to produce the plots are shown next.
-->x1=[1.1 2.1 3.4 4.2 5.5 6.1 7.4 8.1 9.9 11.0];
-->x2 = [22.1 23.2 24.5 20.4 25.2 23.1 19.2 18.2 20.5 19.1];
-->p = [ 1524.7407
1600.8101
1630.6414
1416.1757
1681.7725 ...
-->
1554.5774
1317.4763
1324.6139
1446.5163
1328.9309 ];
-->q = [ 3585.7418
3863.4938
4205.4344
3216.1131
4409.4029 ...
-->
3876.8094
2975.0054
2764.6509
3289.158
2983.4153 ];
-->r = [ 1558.505
1588.175
1630.79
1371.75
1658.565
1541.455 ...
-->
1315.83
1296.275
1481.265
1458.17 ];
-->X = [x1' x2'];
-->[b,C,se] = multiplelinear(X,p',0.01);
-->[b,C,se] = multiplelinear(X,q',0.01);
-->[b,C,se] = multiplelinear(X,r',0.01);
32
The following table summarizes information regarding the fittings:

Fitting
2
adj
Significance
p=p(x1,x2)
0.97658
0.98822
0.96989
reject H0
q=q(x1,x2)
0.99926
0.99963
0.99905
reject H0
r=r(x1,x2)
0.8732
0.93445
0.83697
reject H0
The three fitting show good values of the coefficients R2, R, and R2adj, and the F test for the
significance of the regression indicates rejecting the null hypothesis H0:1 = 2 = = n = 0 in
all cases. In other words, a multiple linear fitting is acceptable for all cases. The residual
plots, depicted below, however, indicate that only the fitting of p=p(x1,x2) shows a random
distribution about zero. The fitting of q=q(x1,x2) shows a clear non-linear trend, while the
fitting of r=r(x1,x2) shows a funnel shape.
33
Scaling residuals
The errors ei may be standardized by using the values
zei = ei/se,
where se is the standard error of the estimate from the multiple linear fitting. If the errors
are normally distributed then approximately 95% of them will fall within the interval (-2,+2).
Any residual located far outside of this interval may indicate an outlier point. Outlier points
are points that do not follow the general trend of the data. They may indicate an error in the
data collection or simply an extremely large or low value in the data. If sufficient justification
exists, outlier points may be discarded and the regression repeated with the remaining data
points.
Another way to scale residuals is to use the so-called studentized residuals defined as
ri =
ei
s e 1 hii
i = 1,2,,m, where hii are the diagonal elements in the hat matrix, H, defined as
H = X(XTX)-1X.
This matrix relates the observed values y to the fitted values y , i.e., y = Hy. Thus, the
name hat matrix. Using the values hii we can write the studentized standard error of the i-th
residual as
se(ei) = se(1-hii)1/2.
34
Influential observations
Sometimes in a simple or multiple linear fitting there will be one or more points whose effects
on the regression are unusually influential. Typically, these so-called influential observations
correspond to outlier points that tend to drag the fitting in one or other direction. To
determine whether or not a point is influential we computed the Cooks distance defined as
di =
ei2 hii
zei2 hii
ri 2 hii
=
=
,
(n + 1) s e2 (1 + hii ) 2 (n + 1)(1 + hii ) 2 (n + 1)(1 hii )
i=1,2,,m. A value of di > 1 may indicate an influential point.
A function for residual analysis

The following function, residuals, produces a comprehensive residual analysis including the
standardized and studentized residuals, studentized standard error of estimates for the
residuals, and the corresponding Cooks distances. The function takes as input a matrix X
whose columns represent the independent variables, i.e., X = [x1 x2 xn], and the column
vector y containing the dependent variable. The function produces a table of results, plots of
the residuals and Cook distances, and returns the values se(ei) , zei, ri, and di. A listing of the
function follows:
function [e,stderr,ze,r,d] = residuals(XA,y)

//Multiple
//e
=
//stderr =
//ze
=
//r
=
//d
=
linear regression - residual analysis

residuals
standard errors of residuals
standardized residuals
studentized residuals
Cook's distances
[m n] = size(XA);
X = [ones(m,1) XA];
H = X*inv(X'*X)*X';
yh = H*y;
e=y-yh;
SSE=e'*e;
MSE = SSE/(m-n-1);
se = sqrt(MSE);
[nh mh] = size(H);
h=[];
for i=1:nh
h=[h;H(i,i)];
end;
see = se*(1-h).^2;
ze
= e/se;
r
= e./see;
d
= r.*r.*h./(1-h)/(n+1);
//Size of original X matrix

//Augmenting matrix X
//H ('hat') matrix
//Fitted value of y
//Errors or residuals
//Sum of squared errors
//Mean square error
//Standard error of estimate
//Size of matrix H
//Vector h
//standard errors for residuals

//standardized residuals
//studentized residuals
//Cook's distances
//Printing of results
printf(' ');
35
printf('Residual Analysis for Multiple Linear Regression');

printf('--------------------------------------------------------------------------------');
printf('
i
y(i)
yh(i)
e(i)
se(e(i))
ze(i)
r(i)
d(i)');
printf('--------------------------------------------------------------------------------');
for i = 1:m
printf('%4.0f %10.6g %10.6g %10.6g %10.6g %10.6g %10.6g %10.6g',...
i,y(i),yh(i),e(i),see(i),ze(i),r(i),d(i));
end;
printf('--------------------------------------------------------------------------------');
printf('Standard error of estimate (se)
= %g',se);
printf('--------------------------------------------------------------------------------');
printf(' ');
//Plots of residuals - several options
xset('window',1);xset('mark',-9,2);
plot2d(yh,e,-9);
xtitle('Residual plot - residual vs. y','y','e');
plot2d(yh,ze,-9);
xtitle('Residual plot - standardized residual vs. yh','yh','ze');
plot2d(yh,r,-9);
xtitle('Residual plot - studentized residual vs. yh','yh','r');
plot2d(yh,d,-9);
xtitle('Cook distance plot','yh','d');
Applications of function residuals

Applications of function residuals to the multiple linear fittings p = p(x1,x2), q = q(x1,x2), and r
= r(x1,x2), follow. First, we load function residuals:
-->getf(residuals)
Residual analysis for p = p(x1,x2)

-->residuals(X,p');
36
Residual Analysis for Multiple Linear Regression

------------------------------------------------------------------------------i
y(i)
yh(i)
e(i)
se(e(i))
ze(i)
r(i)
d(i)
------------------------------------------------------------------------------1
1524.74
1521.69
3.04682
8.02336
.13006
.37974
.0340687
2
1600.81
1578.51
22.3035
13.2116
.95204
1.68817
.31503
3
1630.64
1645.41
-14.7688
12.6407
-.63042 -1.16835
.16442
4
1416.18
1424.51
-8.33603
13.589
-.35583
-.61344
.0392612
5
1681.77
1678.61
3.16424
6.70004
.13507
.47227
.0646745
6
1554.58
1565.08
-10.5036
15.6687
-.44835
-.67035
.0333676
7
1317.48
1353.87
-36.3932
14.7089 -1.55347 -2.47422
.53468
8
1324.61
1298.97
25.6415
10.9047
1.09453
2.35141
.85834
9
1446.52
1418.35
28.1663
11.9558
1.2023
2.35587
.73966
10
1328.93
1341.25
-12.3207
9.18077
-.52592 -1.34202
.35865
------------------------------------------------------------------------------Standard error of estimate (se)
= 23.427
-------------------------------------------------------------------------------
Notice that most of the standardized residuals are within the interval (-2,2), and all of the
values di are less than one. This residual analysis, thus, does not reveal any outliers. Similar
results are obtained from the following table corresponding to the fitting q = q(x1,x2).
Residual analysis for q = q(x1,x2)
-->residuals(X,q');

------------------------------------------------------------------------------i
y(i)
yh(i)
e(i)
se(e(i))
ze(i)
r(i)
d(i)
------------------------------------------------------------------------------1
3585.74
3594.19
-8.45215
5.89407
-.49112
-1.43401
.48583
2
3863.49
3869.77
-6.27968
9.70543
-.36489
-.64703
.0462771
3
4205.43
4196.81
8.62144
9.28605
.50096
.92843
.10383
4
3216.11
3221.54
-5.43054
9.98268
-.31555
-.54400
.0308755
5
4409.4
4388.96
20.4452
4.92194
1.188
4.15388
5.00333
6
3876.81
3891.61
-14.8006
11.5105
-.86001
-1.28584
.12277
7
2975.01
2970.09
4.91247
10.8054
.28545
.45463
.0180525
8
2764.65
2738.01
26.6417
8.01076
1.54806
3.32574
1.71704
9
3289.16
3310.89
-21.7285
8.78291
-1.26257
-2.47396
.81567
10
2983.42
2987.34 -3.92933
6.74432
-.22832
-.58261 .0675956

= 17.2098
-------------------------------------------------------------------------------
37
Residual analysis for r = r(x1,x2)

-->residuals(X,r');
------------------------------------------------------------------------------i
y(i)
yh(i)
e(i)
se(e(i))
ze(i)
r(i)
d(i)
------------------------------------------------------------------------------1
1558.51
1493.88
64.627
17.7424
1.2475
3.64252
3.13458
2
1588.18
1558.26
29.9105
29.2154
.57737
1.02379
.11586
3
1630.79
1634.99
-4.20322
27.953 -.0811352
-.15037 .00272348
4
1371.75
1419.35
-47.6025
30.05
-.91888
-1.58411
.26181
5
1658.57
1683.84
-25.2722
14.8161
-.48783
-1.70573
.84366
6
1541.46
1574.41
-32.9551
34.6489
-.63614
-.95112
.0671713
7
1315.83
1372.19
-56.3571
32.5265
-1.08787
-1.73265
.26221
8
1296.27
1322.31
-26.0318
24.1141
-.50249
-1.07953
.18091
9
1481.27
1455.37
25.8962
26.4384
.49988
.97949
.12786
10
1458.17
1386.18
71.9883
20.3018
1.3896
3.5459
2.50387
= 51.8051
-------------------------------------------------------------------------------
The table corresponding to the fitting q=q(x1,x2) shows two residuals, e1 and e10 whose Cooks
distance is larger than one. These two points, even if their standardized residuals are in the
interval (-2,2), may be considered outliers. The residual analysis eliminating these two
suspected outliers is shown next.
Residual analysis for r = r(x1,x2) eliminating suspected outliers

To eliminate the outliers we modify matrix X and vector r as follows:
-->XX = X(2:9,:)
XX =
!
!
!
!
!
!
!
!
2.1
3.4
4.2
5.5
6.1
7.4
8.1
9.9
23.2
24.5
20.4
25.2
23.1
19.2
18.2
20.5
!
!
!
!
!
!
!
!
-->rr = r(2:9)
rr =
!
1588.175
1630.79
1296.275
1481.265 !
1371.75
1658.565
1541.455
1315.83
-->residuals(XX,rr');
38

------------------------------------------------------------------------------i
y(i)
yh(i)
e(i)
se(e(i))
ze(i)
r(i)
d(i)
------------------------------------------------------------------------------1
1588.18
1542.72
45.459
10.9925
1.27186
4.13546
4.57875
2
1630.79
1627.12
3.67448
17.1966
.10280
.21367 .00672191
3
1371.75
1393.76
-22.008
14.0957
-.61574
-1.56133
.48136
4
1658.57
1681.94
-23.3752
9.70312
-.65399
-2.40904
1.77832
5
1541.46
1563.69
-22.2351
22.8787
-.62210
-.97187
.0786800
6
1315.83
1345.33
-29.4969
18.9967
-.82527
-1.55274
.29870
7
1296.27
1291.79
4.48228
12.618
.12541
.35523
.0287305
8
1481.27
1437.77
43.4995
8.21833
1.21703
5.29299
10.1365
= 35.7423
-------------------------------------------------------------------------------
Even after eliminating the two influential observations we find that the remaining e1 and e8 are
influential in the reduced data set. We can check the analysis of residuals eliminating these
two influential observations as shown next:
-->XXX = XX(2:7), rrr = rr(2:7)

XXX =
!
!
!
!
!
!
3.4
4.2
5.5
6.1
7.4
8.1
rrr
!
!
!
!
!
!
!
=
1630.79
1371.75
1658.565
1541.455
1315.83
1296.275 !
-->residuals(XXX,rrr');
------------------------------------------------------------------------------i
y(i)
yh(i)
e(i)
se(e(i))
ze(i)
r(i)
d(i)
-------------------------------------------------------------------------------
1630.79
1601.8
28.9923
33.2701
.20573
.87142
.40176
2
1371.75
1557.26
-185.509
65.1625
-1.31635
-2.84688
1.9071
3
1658.57
1484.88
173.68
96.7163
1.23241
1.79577
.33395
4
1541.46
1451.48
89.9739
96.4308
.63844
.93304
.0909297
5
1315.83
1379.11
-63.2765
63.918
-.44900
-.98996
.23759
6
1296.27
1340.14
-43.8605
35.9466
-.31123
-1.22016
.72952
= 140.927
-------------------------------------------------------------------------------
Even after eliminating the two points e1 and e8 from the reduced data set, another influential
point is identified, e2. We may continue eliminating influential points, at the risk of running
out of data, or try a different data fitting.
39
Multiple linear regression with function datafit

SCILAB provides function datafit, introduced in Chapter 8, which can be used to determine the
coefficients of a fitting function through a least-square criteria. Function datafit is used for
fitting data to a model by defining an error function e = G(b,x) where b is a column vector of m
rows representing the parameters of the model, and x is a column vector of n rows
representing the variables involved in the model. Function datafit finds a solution to the set
of k equations ei = G(b,xi) = 0.
The simplest call to function datafit is
[p,err]=datafit(G,X,b0)
where G is the name of the error function G(b,x), X is a matrix whose rows consists of the
different vectors of variables, i.e., X = [x1; x2; ;xn], and b0 is a column vector representing
initial guesses of the parameters b sought.
Function datafit can be used to determine the coefficients of a multiple linear fitting as
illustrated in the example below. The data to be fit is given in the following table:
x1
25
31
45
60
65
72
80
84
75
60
50
38
x2
24
21
24
25
25
26
25
25
24
25
25
23
x3
91
90
88
87
91
94
87
86
88
91
90
89
x4
100
95
110
88
94
99
97
96
110
105
100
98
y
240
236
290
274
301
316
300
296
267
276
288
261
First, we load the data and prepare the matrix XX for the application of function datafit.
-->XY=...
-->[25 24
-->31 21
-->45 24
-->60 25
-->65 25
-->72 26
-->80 25
-->84 25
-->75 24
-->60 25
-->50 25
-->38 23
91
90
88
87
91
94
87
86
88
91
90
89
100
95
110
88
94
99
97
96
110
105
100
98
240
236
290
274
301
316
300
296
267
276
288
261];
-->XX=XY';
40
Next, we define the error function to be minimized and call function datafit:
-->deff('[e]=G(b,z)',...
--> 'e=b(1)+b(2)*z(1)+b(3)*z(2)+b(4)*z(3)+b(5)*z(4)-z(5)')
-->[b,er]=datafit(G,XX,b0)
er = 1699.0093
b =
! - 102.71289 !
!
.6053697 !
!
8.9236567 !
!
1.4374508 !
!
.0136086 !
You can check these results using function multiplelinear as follows:

-->size(XY)
ans =
!
12.
5. !
-->X=XY(1:12,1:4);y=XY(1:12,5);
//Prepare matrix X and vector y
-->[bb,C,se]=multiplelinear(X,y,0.10); //Call function multiplelinear
==========================
------------------------------------------------------------------------i
b(i)
se(b(i))
Lower
Upper
t0 H0:b(i)=0
------------------------------------------------------------------------0
-102.713
207.859
-489.237
283.81
-.49415 do not reject
1
.60537
.36890 -.0806111
1.29135
1.64103 do not reject
2
8.92364
5.30052
-.93293
18.7802
3
1.43746
2.39162
-3.00988
5.88479
.60104 do not reject
4
.0136093
.73382
-1.35097
1.37819
------------------------------------------------------------------------t_alpha/2 = 1.85955
------------------------------------------------------------------------Table of fitted values and errors
------------------------------------------------------------------------------i
y(i)
yh(i)
e(i)
C.I. for Y
C.I. for Ypred
------------------------------------------------------------------------------1
240
258.758
-18.758
248.31
269.206
214.676
302.84
2
236
234.114
1.88623
219.86
248.367
175.738
292.49
3
290
266.689
23.3109
255.836
277.542
221.107
312.271
4
274
282.956
-8.95646
271.601
294.312
235.506
330.407
5
301
291.815
9.18521
284.131
299.498
257.72
325.909
6
316
309.356
6.64355
296.747
321.966
257.204
361.509
7
300
295.186
4.81365
287.17
303.203
259.916
330.457
8
296
296.157
-.15677
286.462
305.851
254.842
337.472
9
267
284.85
-17.8502
273.464
296.236
237.285
332.415
10
276
288.938
-12.9376
281.73
296.145
256.503
321.372
11
288
281.378
6.62157
274.517
288.24
250.134
312.622
12
261
254.802
6.19798
247.76
261.844
222.94
286.664
-------------------------------------------------------------------------------
41
-------------------------------------------------------Source of
Sum of
Degrees of
Mean
variation
squares
freedom
square
F0
-------------------------------------------------------Regression
4957.24
4
1239.31
5.10602
Residual
1699.01
7
242.716
Total
6656.25
11
-------------------------------------------------------With F0 = 5.10602 and F_alpha = 2.96053,
--------------------------------------------------------
= 15.5793
= .74475
= .86299
!
!
!
!
!
43205.302
10.472107
142.32333
389.41606
43.653591
- 10.472107
.1360850
- 1.4244286
.4289197
- .0095819
- 142.32333
- 1.4244286
28.095536
- 5.3633513
.1923068
- 389.41606
.4289197
- 5.3633513
5.7198487
- .1563728
- 43.653591 !
- .0095819 !
.1923068 !
- .1563728 !
.5384939 !
---------------------------------------------------------
Notice that the error, e, returned by datafit is the sum of squared errors, SSE, returned by
function multiplelinear. The detailed regression analysis provided by multiplelinear indicates
that the hypothesis for significance of regression is to be rejected, i.e., the linear model is not
necessarily the best for this data set. Also, the coefficient of multiple regression and its
adjusted value are relatively small.
Note: Function datafit can be used to fit linear and non-linear functions.
application of function datafit were presented in Chapter 8.
Details of the
Polynomial data fitting

A function for polynomial data fitting was developed in Chapter 5 to illustrate the use of matrix
operations. In this section, we present the analysis of polynomial data fitting as a special case
of a multilinear regression. A polynomial fitting is provided by a function of the form
y = b0 + b1z + b2z2 + + bnzn.

This is equivalent to a multiple linear fitting if we take x1 = z, x2 = z2, , xn = zn. Thus, given a
data set {(z1,y1), (z2,y2), , (zm,ym)}, we can use function multiplelinear to obtain the
coefficients [b0,b1,b2,,bn] of the polynomial fitting. As an example, consider the fitting of a
cubic polynomial to the data given in the following table:
42
z
y
2.10
13.41
3.20
46.48
4.50
95.39
6.80
13.50
18.40
21.00
380.88 2451.55 5120.46 8619.14
The SCILAB instructions for producing the fitting are shown next. First, we load the data:
-->z=[2.10,3.20,4.50,6.80,13.50,18.40,21.00];
-->y=[13.41,46.48,95.39,380.88,2451.55,5120.46,8619.14];
Next, we prepare the vectors for the multiple linear regression and call the appropriate
function:
-->x1 = z; x2 = z^2; x3 = z^3; X = [x1' x2' x3']; yy = y';
-->[b,C,se]=multiplelinear(X,yy,0.01)
==========================
------------------------------------------------------------------------i
b(i)
se(b(i))
Lower
Upper
t0 H0:b(i)=0
------------------------------------------------------------------------0
-467.699
664.835
-3528.66
2593.26
1
223.301
262.938
-987.289
1433.89
2
-23.3898
26.1662
-143.861
97.0817
3
1.56949
.74616
-1.86592
5.00491
------------------------------------------------------------------------t_alpha/2 = 4.60409
------------------------------------------------------------------------Table of fitted values and errors
------------------------------------------------------------------------------i
y(i)
yh(i)
e(i)
C.I. for Y
C.I. for Ypred
------------------------------------------------------------------------------1
13.41
-87.3819
100.792
-351.859
177.095
-4793.08
4618.32
2
46.48
58.78
-12.3
-115.377
232.937
-3048.97
3166.53
3
95.39
206.529
-111.139
25.4059
387.653
-3024.28
3437.34
4
380.88
462.697
-81.8173
221.167
704.228
-3836.65
4762.04
5 2451.55
2145.6
305.95
1889.3
2401.9
-2415.28
6706.48
6 5120.46
5499.33
-378.865
5272.7
5725.95
1463.9
9534.75
7 8619.14
8441.76
177.38
8143.74
8739.78
3141.72
13741.8
-------------------------------------------------------------------------------
-------------------------------------------------------Source of
Sum of
Degrees of
Mean
variation
squares
freedom
square
F0
-------------------------------------------------------Regression
6.64055e+07
3 2.21352e+07
222.864
Residual
297964
3
99321.4
Total
6.67034e+07
6
-------------------------------------------------------With F0 = 222.864 and F_alpha = 29.4567,
--------------------------------------------------------
43
= 315.153
= .99553
= .99776
!
442005.47
! - 166781.09
!
15518.351
! - 412.44179
- 166781.09
69136.156
- 6728.9534
183.73635
15518.351
- 6728.9534
684.66847
- 19.297326
- 412.44179
183.73635
- 19.297326
.5567618
!
!
!
!
---------------------------------------------------------
With the vector of coefficients found above, we define the cubic function:
-->deff('[y]=yf(z)','y=b(1)+b(2)*z+b(3)*z^2+b(4)*z^3')
Then, we produce the fitted data and plot it together with the original data:
-->zz=[0:0.1:25];yz=yf(zz);
-->rect = [0 -500 25 15000]; //Based on min. & max. values of z and y
-->xset('mark',-1,2)
-->plot2d(zz,yz,1,'011',' ',rect)
-->plot2d(z,y,-1,'011',' ',rect)
-->xtitle('Fitting of a cubic function','z','y')
44
Alternatively, we can use function datafit to obtain the coefficients of the fitting as follows:
-->XX = [z;y]
XX
!
!
=
2.1
13.41
3.2
46.48
4.5
95.39
6.8
380.88
13.5
2451.55
18.4
5120.46
21.
!
8619.14 !
-->b0=ones(4,1)
b0 =
!
!
!
!
1.
1.
1.
1.
!
!
!
!
-->[b,er]=datafit(G,XX,b0)
er =
297964.09
=
! - 467.6971 !
!
223.30004 !
! - 23.389786 !
!
1.5694905 !
45
Exercises
[1]. To analyze the dependency of the mean annual flood, Q(cfs), on the drainage area, A(mi2)
for a given region, data from six experimental watersheds is collected. The data is summarized
in the table below:
2
A(mi ) 16.58 3.23 16.8 42.91 8.35 6.04
105 465 1000 290 157
Q(cfs) 455
(a) Use function linreg to perform a simple linear regression analysis on these data. The
purpose is to obtain a relationship of the form Q = mA + b. (b) Determine the covariance of A
and B, (c)the correlation coefficient, and (d) the standard error of the estimate.
[2] For the data of problem [1] use function linregtable to perform the linear fitting. (a) What
is the decision regarding the hypothesis that the slope of the linear fitting may be zero at a
level of confidence of 0.05? (b) What is the decision regarding the hypothesis that the
intercept of the linear fitting may be zero at a level of confidence of 0.05?
[3] For the data of problem [1] use function multiplelinear to produce the data fitting. (a)
What is the decision regarding the hypothesis that the slope of the linear fitting may be zero at
a level of confidence of 0.05? (b) What is the decision regarding the hypothesis that the
intercept of the linear fitting may be zero at a level of confidence of 0.05? (b) What is the
decision regarding the hypothesis that the linear fitting may not apply at all?
[4]. The data shown in the table below represents the monthly precipitation, P(in), in a
particular month, and the corresponding runoff, R(in), out of a specific hydrological basin for
the period 1960-1969.
Year
P (in)
R (in)
1960 1961 1962 1963 1964 1965 1966 1967 1968 1969
1.95 10.8 3.22 4.51 6.71 1.18 4.82 6.38 5.97 4.64
0.46 2.85 0.99
1.4 1.98 0.45 1.31 2.22 1.36 1.21
(a) Use function linreg to perform a simple linear regression analysis on these data. The
purpose is to obtain a relationship of the form R = mP + b. (b) Determine the covariance of P
and R, (c)the correlation coefficient, and (d) the standard error of the estimate.
[5] For the data of problem [1] use function linregtable to perform the linear fitting. (a) What
is the decision regarding the hypothesis that the slope of the linear fitting may be zero at a
level of confidence of 0.05? (b) What is the decision regarding the hypothesis that the
intercept of the linear fitting may be zero at a level of confidence of 0.05?
[6] For the data of problem [1] use function multiplelinear to produce the data fitting. (a)
What is the decision regarding the hypothesis that the slope of the linear fitting may be zero at
a level of confidence of 0.05? (b) What is the decision regarding the hypothesis that the
intercept of the linear fitting may be zero at a level of confidence of 0.05? (b) What is the
decision regarding the hypothesis that the linear fitting may not apply at all?
[7]. The following table shows data indicating the monthly precipitation during the month of
February, Pf, and during the month of March, Pm, as well as the runoff during the month of
March, Rm, for a specific watershed during the period 1935-1958.
46
Year Pm Pf
Rm
1935 9.74 4.11 6.15
1936 6.01 3.33 4.93
1937 1.30 5.08 1.42
1938 4.80 2.41 3.60
1939 4.15 9.64 3.54
1940 5.94 4.04 2.26
1941 2.99 0.73 0.81
1942 5.11 3.41 2.68
1943 7.06 3.89 4.68
1944 6.38 8.68 5.18
1945 1.92 6.83 2.91
1946 2.82 5.21 2.84
1947 2.51 1.78 2.02
1948 5.07 8.39 3.27
1949 4.63 3.25 3.05
1950 4.24 5.62 2.59
1951 6.38 8.56 4.66
1952 7.01 1.96 5.40
1953 4.15 5.57 2.60
1954 4.91 2.48 2.52
1955 8.18 5.72 6.09
1956 5.85 10.19 4.58
1957 2.14 5.66 2.02
1958 3.06 3.04 2.59
a. Use function multiplot to show the dependence of the many variables.
b. Use function multiplelinear to check the multiple linear fitting Rm = b0 + b1Pm +
b2Pf.
c. For a level of confidence of 0.05, what are the decisions regarding the hypotheses
that each of the coefficients may be zero?
d. What is the decision regarding the hypothesis that the linear fitting may not apply
at all for the same level of confidence?
e. What value of runoff for the month of March is predicted if the precipitation in
the month of March is 6.2 in, and that of the month of February is 3.2 in?
f. What are the confidence intervals for the mean value and the prediction for the
data of question (e) at a confidence level 0.05?
[8]. In the analysis of runoff produced by precipitation into a watershed, often we are
required to estimate a parameter known as the time of concentration (tc) which determines
the time to the peak of the hydrograph produced by the watershed. It is assumed that the
time of concentration is a function of a characteristic watershed length (L), of a characteristic
watershed slope (S), and of a parameter known as the runoff curve number (CN). Runoff curve
numbers are numbers used by the U.S. Soil Conservation Service in the estimation of runoff
from watersheds. Runoff curve numbers are typically functions of the location of the
watershed and of its soil and vegetation covers. The following table shows values of the time
of concentration, tc(hr), the watershed length, L(ft), the watershed slope, S(%), and the runoff
curve number (CN) for 5 experimental watersheds.
47
tc(hr)
L (ft)
S (%)
CN
0.2
800
2
75
0.2
0.2
0.3
0.3
1200 2100 2000 1500
3
4
6
1
84
88
70
85
(a) Use function multiplot to show the interdependence of the various variables in the table.
(b) Assuming that a multiple-linear fitting can be used to explain the dependence of tc on L, S,
and CN, use function multiplelinear to determine the coefficients of the fitting. (c) For a level
of confidence of 0.01, what are the decisions regarding the hypotheses that each of the
coefficients may be zero? (d) What is the decision regarding the hypothesis that the linear
fitting may not apply at all for the same level of confidence? (e) What value of the time of
concentration is predicted for L = 1750 ft, S = 5%, and CN = 80. (f) What are the confidence
intervals for the mean value and the prediction for the data of question (e) at a confidence
level 0.05?
[9]. The data in the table below shows the peak discharge, qp(cfs), the rainfall intensity,
i(in/hr), and the drainage area, A(acres), for rainfall events in six different watersheds.
qp(cfs)
i(in/hr)
A(acres)
23
3.2
12
45
4.6
21
44
5.1
18
64
3.8
32
68
6.1
24
62
7.4
16
(b) Assuming that a multiple-linear fitting can be used to explain the dependence of qp on i,
and A, use function multiplelinear to determine the coefficients of the fitting. (c) For a level
of confidence of 0.1, what are the decisions regarding the hypotheses that each of the
concentration is predicted for i = 5.6 in/hr and A = 25 acres. (f) What are the confidence
level 0.10?
[10]. Measurements performed across a pipeline diameter produce the following table of
velocities, v(fps), as function of the radial distance, r(in), measured from the pipe centerline.
48
r(in)
2.7
2.6
2.5
2.4
2.3
2.2
2.1
2.0
1.9
1.8
1.7
1.6
1.5
1.4
V(fps)
57.73
58.30
59.25
59.70
59.80
60.60
61.20
61.50
62.20
62.70
62.90
63.05
63.65
64.10
r(in)
1.3
1.2
1.1
1.0
0.9
0.8
0.7
0.5
0.3
0.1
0.1
0.3
0.5
0.7
V(fps)
64.40
64.80
64.80
65.20
65.50
65.50
65.90
66.30
66.40
66.50
66.50
66.30
66.00
65.70
r(in)
0.9
1.1
1.3
1.4
1.5
1.6
1.7
1.8
1.9
2.0
2.1
2.2
2.3
2.4
V(fps)
65.50
64.80
64.10
63.70
63.40
63.00
62.70
62.50
62.10
61.25
61.20
60.55
60.00
59.40
r(in)
2.5
2.6
2.7
2.8
2.9
3.0
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
V(fps)
58.90
58.40
57.50
57.00
56.60
55.90
54.80
54.20
53.20
52.35
50.80
49.50
47.70
44.45
(a) Plot the velocity profile indicated in the table. Notice that the values of r start at 2.7 in
and decrease to 0.1in, just to increase again from 0.1 in to 3.8 in. What these values of r
represent are velocity measurements at both sides of the centerline along a diameter of the
pipe. To produce an accurate plot of the velocity distribution, take the values of r from 2.7 in
down to 0.1 in as negative, and the remaining values as positive. (b) Using SCILAB function
datafit, fit a logarithmic function of the form v = b0 + b1 ln(y) to the data, where y = R-r, and R
is the pipe radius measured as R = 3.958 in.
[11]. The tables below show measurements of stagnation pressure on an air jet in a test set
up. The values of x represent distances from the jet outlet, where the values of r represent
distances from the jet centerline measured radially outwards. The stagnation pressure values
are used to calculate air velocities at the locations indicated by using v = (2gh)1/2, where g is
the acceleration of gravity. To make the units consistent, we recommend that you transform
the data to feet by using 1 in = (1/12) ft, and 1 cm = 0.0328 ft, and calculate the velocities
using g = 32.2 ft/s2.
Along centerline
x(in)
h(cm)
-0.25
20.58
0.00
20.38
0.50
20.16
1.00
20.16
3.00
20.16
5.00
19.6
6.00
18.25
6.50
17.11
Across jet at x = 12 in
r(in)
h(cm)
8.50
0.00
5.00
0.00
4.50
0.15
4.20
0.20
3.70
0.47
3.50
0.62
3.09
0.94
2.83
1.19
49
r(in)
-0.50
-1.00
-1.50
-2.00
-2.50
-3.00
-3.50
-4.20
h(cm)
6.08
5.06
3.97
2.73
1.78
1.11
0.70
0.25
8.00
10.00
15.00
20.00
13.52
9.28
4.14
2.23
2.55
2.00
1.50
1.00
0.50
0.00
1.62
2.62
3.91
5.12
6.07
6.51
-4.50
-4.80
-5.00
-5.30
-5.50
0.17
0.12
0.07
0.02
0.00
(a) Convert the data columns to feet and calculate the velocities corresponding to the different
values of h. (b) Plot the centerline velocities against the distance x and fit an equation of the
form
v(x) = b0/(b1 + b2x)
to the data resulting from the first table. (c) Plot the velocity v at section x = 12 in against the
radial distance |r|, and fit an equation of the form
v(r) = b0exp(-b1x2)
to the data resulting from the second table.
[12]. For relatively low pressures, Henrys law relates the vapor pressure of a gas, P, to the
molar fraction of the gas in mixture, x. The law is stated as P = kx, where k is known as
Henrys constant. To determine Henrys constant in practice we use data of P against x and
fit a straight line, i.e., P = b0 + b1x. If the value of b0 can be taken as zero, then, b1 = k.
Given the pressure-molar fraction data shown in the next table, use function linreg to
determine Henrys constant for the solubility of the following elements or compouns in water
at temperature indicated:
Sulfur dioxide, 23 C
3
P(mmHg) x(10 )
5
0.3263
10
0.5709
50
2.329
100
4.213
200
7.448
300
10.2
Carbon monoxide, 19 C
3
P(mmHg)
x(10 )
900
2.417
1000
2.685
2000
5.354
3000
8.000
4000
10.630
5000
13.230
6000
15.800
7000
18.280
8000
20.670
50
Hydrogen, 23 C
3
P(mmHg) x(10 )
1100
1.861
2000
3.382
3000
5.067
4000
6.729
6000
9.841
8000
12.560
[13]. In the analysis of liquid mixtures it is often necessary to determine parameters known as
activity coefficients. For the mixture of two liquids the van Laar equations can be used to
determine the activity coefficients, 1 and 2, in terms of the molecular fractions x1 and x2:
ln 1 =
A x1
1 +
B x2
, ln 2 =
B x1
1 +
A x2
The molecular fractions are related by

x1 + x2 = 1.
The table below shows the activity coefficients for a liquid mixture as functions of the
molecular fraction x1. Use these data and SCILAB function datafit to obtain the values of the
van Laar coefficients A and B for the data.
x1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
4.90
2.90
1.95
1.52
1.30
1.20
1.10
1.04
1.01
2
1.05
1.20
1.30
1.50
1.70
2.00
2.25
2.60
2.95
[14]. An alternative relationship between the activity coefficients, 1 and 2, in terms of the
molecular fractions x1 and x2 are the Margules equations:
ln 1 = (2B-A)x22 + 2(A-B)x23
ln 2 = (2A-B)x12 + 2(B-A)x23.
Using the data of problem [13] and SCILAB function datafit, determine the coefficients A and B
of the Margules equations.
[15]. Infiltration into soil is typically modeled using Hortons equation

f = fc + (f0-fc)e-kt,
where f is the infiltration rate, fc is the infiltration rate at saturation, f0 is the initial
infiltration rate, t is time, and k is a constant that depends primarily on the type of soil and
vegetation of the area of interest.
51
The following table shows measurements of the infiltration rate, f, as function of time, t, for a
specific storm in a watershed.
t(hr)
1
2
3
4
5
6
8
10
12
f(cm/hr)
3.9
3.4
3.1
2.7
2.5
2.3
2
1.8
1.54
t(hr)
14
16
18
20
22
24
26
28
30
f(cm/hr)
1.43
1.36
1.31
1.28
1.25
1.23
1.22
1.2
1.2
(a) Use SCILABs function datafit to obtain the parameters f0, fc, and k for the Hortons
equation.
(b) Plot the original data and the fitted data in the same set of axes.
[16]. The following data represents different properties of granite samples taken at the
locations indicated by the coordinates x(mi) and y(mi) on a specific site. The properties listed
in the table are as follows: x1 = percentage of quartz in the sample, x2 = color index (a
percentage), x3 = percentage of total feldspar, w = specific gravity (water = 1.0).
x1
x2
x3
21.3
38.9
26.1
29.3
24.5
30.9
27.9
22.8
20.1
16.4
15.0
0.6
18.4
19.5
34.4
26.9
28.7
28.5
38.4
28.1
37.4
5.5
2.7
11.1
6.0
6.6
3.3
1.9
1.2
5.6
21.3
18.9
35.9
16.6
14.2
4.6
8.6
5.5
3.9
3.0
12.9
3.5
73.0
57.4
62.6
63.6
69.1
65.1
69.1
76.0
74.1
61.7
65.6
62.5
64.9
65.4
60.7
63.6
65.8
67.8
57.6
59.0
57.6
2.63
2.64
2.64
2.63
2.64
2.61
2.63
2.63
2.65
2.69
2.67
2.83
2.70
2.68
2.62
2.63
2.61
2.62
2.61
2.63
2.63
0.920
1.150
1.160
1.300
1.400
1.590
1.750
1.820
1.830
1.855
2.010
2.040
2.050
2.210
2.270
2.530
2.620
3.025
3.060
3.070
3.120
6.090
3.625
6.750
3.010
7.405
8.630
4.220
2.420
8.840
10.920
14.225
10.605
8.320
8.060
2.730
3.500
7.445
5.060
5.420
12.550
12.130
52
0.9
8.8
16.2
2.2
29.1
24.9
39.6
17.1
0.0
19.9
1.2
13.2
13.7
26.1
19.9
4.9
15.5
0.0
4.5
0.0
4.0
23.4
29.5
22.9
34.9
5.5
28.4
5.1
6.9
3.6
11.3
47.8
11.6
34.8
18.8
21.2
2.3
4.1
18.8
12.2
39.7
30.5
63.8
24.1
12.4
9.8
74.4
55.4
77.6
69.3
65.7
67.8
56.6
70.9
52.2
67.2
64.0
67.4
64.0
71.2
76.0
74.30
69.70
60.20
63.90
35.20
71.80
63.10
60.40
2.78
2.76
2.63
2.74
2.64
2.70
2.63
2.71
2.84
2.68
2.84
2.74
2.74
2.61
2.63
2.77
2.72
2.83
2.77
2.92
2.77
2.79
2.69
3.400
3.520
3.610
4.220
4.250
4.940
5.040
5.060
5.090
5.240
5.320
5.320
5.330
5.350
5.610
5.850
6.460
6.590
7.260
7.420
7.910
8.470
8.740
15.400
9.910
11.520
16.400
11.430
5.910
1.840
11.760
16.430
11.330
8.780
13.730
12.450
1.430
4.150
13.840
11.660
14.640
12.810
16.610
14.650
13.330
15.770
(b) Assuming that a multiple-linear fitting can be used to explain the dependence of w on x1,
x2, and x3, use function multiplelinear to determine the coefficients of the fitting. (c) For a
level of confidence of 0.1, what are the decisions regarding the hypotheses that each of the
concentration is predicted for x1 =17, x2 = 25, and x3 = 55. (f) What are the confidence
level 0.10?
REFERENCES (for all SCILAB documents at InfoClearinghouse.com)

Abramowitz, M. and I.A. Stegun (editors), 1965,"Handbook of Mathematical Functions with Formulas, Graphs, and
Mathematical Tables," Dover Publications, Inc., New York.
Arora, J.S., 1985, "Introduction to Optimum Design," Class notes, The University of Iowa, Iowa City, Iowa.
Asian Institute of Technology, 1969, "Hydraulic Laboratory Manual," AIT - Bangkok, Thailand.
Berge, P., Y. Pomeau, and C. Vidal, 1984,"Order within chaos - Towards a deterministic approach to turbulence," John
Wiley & Sons, New York.
Bras, R.L. and I. Rodriguez-Iturbe, 1985,"Random Functions and Hydrology," Addison-Wesley Publishing Company,
Reading, Massachussetts.
Brogan, W.L., 1974,"Modern Control Theory," QPI series, Quantum Publisher Incorporated, New York.
Browne, M., 1999, "Schaum's Outline of Theory and Problems of Physics for Engineering and Science," Schaum's
outlines, McGraw-Hill, New York.
Farlow, Stanley J., 1982, "Partial Differential Equations for Scientists and Engineers," Dover Publications Inc., New
York.
53
Friedman, B., 1956 (reissued 1990), "Principles and Techniques of Applied Mathematics," Dover Publications Inc., New
York.
Gomez, C. (editor), 1999, Engineering and Scientific Computing with Scilab, Birkhuser, Boston.
Gullberg, J., 1997, "Mathematics - From the Birth of Numbers," W. W. Norton & Company, New York.
Harman, T.L., J. Dabney, and N. Richert, 2000, "Advanced Engineering Mathematics with MATLAB - Second edition,"
Brooks/Cole - Thompson Learning, Australia.
Harris, J.W., and H. Stocker, 1998, "Handbook of Mathematics and Computational Science," Springer, New York.
Hsu, H.P., 1984, "Applied Fourier Analysis," Harcourt Brace Jovanovich College Outline Series, Harcourt Brace
Jovanovich, Publishers, San Diego.
Journel, A.G., 1989, "Fundamentals of Geostatistics in Five Lessons," Short Course Presented at the 28th International
Geological Congress, Washington, D.C., American Geophysical Union, Washington, D.C.
Julien, P.Y., 1998,Erosion and Sedimentation, Cambridge University Press, Cambridge CB2 2RU, U.K.
Keener, J.P., 1988, "Principles of Applied Mathematics - Transformation and Approximation," Addison-Wesley
Publishing Company, Redwood City, California.
Kitanidis, P.K., 1997,Introduction to Geostatistics - Applications in Hydogeology, Cambridge University Press,
Cambridge CB2 2RU, U.K.
Koch, G.S., Jr., and R. F. Link, 1971, "Statistical Analysis of Geological Data - Volumes I and II," Dover Publications,
Inc., New York.
Korn, G.A. and T.M. Korn, 1968, "Mathematical Handbook for Scientists and Engineers," Dover Publications, Inc., New
York.
Kottegoda, N. T., and R. Rosso, 1997, "Probability, Statistics, and Reliability for Civil and Environmental Engineers,"
The Mc-Graw Hill Companies, Inc., New York.
Kreysig, E., 1983, "Advanced Engineering Mathematics - Fifth Edition," John Wiley & Sons, New York.
Lindfield, G. and J. Penny, 2000, "Numerical Methods Using Matlab," Prentice Hall, Upper Saddle River, New Jersey.
Magrab, E.B., S. Azarm, B. Balachandran, J. Duncan, K. Herold, and G. Walsh, 2000, "An Engineer's Guide to
MATLAB", Prentice Hall, Upper Saddle River, N.J., U.S.A.
McCuen, R.H., 1989,Hydrologic Analysis and Design - second edition, Prentice Hall, Upper Saddle River, New Jersey.
Middleton, G.V., 2000, "Data Analysis in the Earth Sciences Using Matlab," Prentice Hall, Upper Saddle River, New
Jersey.
Montgomery, D.C., G.C. Runger, and N.F. Hubele, 1998, "Engineering Statistics," John Wiley & Sons, Inc.
Newland, D.E., 1993, "An Introduction to Random Vibrations, Spectral & Wavelet Analysis - Third Edition," Longman
Scientific and Technical, New York.
Nicols, G., 1995, Introduction to Nonlinear Science, Cambridge University Press, Cambridge CB2 2RU, U.K.
Parker, T.S. and L.O. Chua, , "Practical Numerical Algorithms for Chaotic Systems, 1989, Springer-Verlag, New York.
Peitgen, H-O. and D. Saupe (editors), 1988, "The Science of Fractal Images," Springer-Verlag, New York.
Peitgen, H-O., H. Jrgens, and D. Saupe, 1992, "Chaos and Fractals - New Frontiers of Science," Springer-Verlag, New
York.
Press, W.H., B.P. Flannery, S.A. Teukolsky, and W.T. Vetterling, 1989, Numerical Recipes - The Art of Scientific
Computing (FORTRAN version), Cambridge University Press, Cambridge CB2 2RU, U.K.
Raghunath, H.M., 1985, "Hydrology - Principles, Analysis and Design," Wiley Eastern Limited, New Delhi, India.
Recktenwald, G., 2000, "Numerical Methods with Matlab - Implementation and Application," Prentice Hall, Upper
Saddle River, N.J., U.S.A.
54
Rothenberg, R.I., 1991, "Probability and Statistics," Harcourt Brace Jovanovich College Outline Series, Harcourt Brace
Jovanovich, Publishers, San Diego, CA.
Sagan, H., 1961,"Boundary and Eigenvalue Problems in Mathematical Physics," Dover Publications, Inc., New York.
Spanos, A., 1999,"Probability Theory and Statistical Inference - Econometric Modeling with Observational Data,"
Cambridge University Press, Cambridge CB2 2RU, U.K.
Spiegel, M. R., 1971 (second printing, 1999), "Schaum's Outline of Theory and Problems of Advanced Mathematics for
Engineers and Scientists," Schaum's Outline Series, McGraw-Hill, New York.
Tanis, E.A., 1987, "Statistics II - Estimation and Tests of Hypotheses," Harcourt Brace Jovanovich College Outline
Series, Harcourt Brace Jovanovich, Publishers, Fort Worth, TX.
Tinker, M. and R. Lambourne, 2000, "Further Mathematics for the Physical Sciences," John Wiley & Sons, LTD.,
Chichester, U.K.
Tolstov, G.P., 1962, "Fourier Series," (Translated from the Russian by R. A. Silverman), Dover Publications, New York.
Tveito, A. and R. Winther, 1998, "Introduction to Partial Differential Equations - A Computational Approach," Texts in
Applied Mathematics 29, Springer, New York.
Urroz, G., 2000, "Science and Engineering Mathematics with the HP 49 G - Volumes I & II", www.greatunpublished.com,
Charleston, S.C.
Urroz, G., 2001, "Applied Engineering Mathematics with Maple", www.greatunpublished.com, Charleston, S.C.
Winnick, J., , "Chemical Engineering Thermodynamics - An Introduction to Thermodynamics for Undergraduate
Engineering Students," John Wiley & Sons, Inc., New York.
55

Regression Analysis With Scilab

Uploaded by

Copyright:

Available Formats

Regression Analysis With Scilab

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Regression Analysis With Scilab

Uploaded by

Copyright:

Available Formats

Regression Analysis with SCILAB

2001 Gilberto E. Urroz

The author's SCILAB web page can be accessed at:

Please report any errors in this document to: [email protected]

Simple linear regression

Graphical display of multivariate data

Multiple linear regression

Simple linear regression using function multiplelinear

Multiple linear regression with function datafit

Polynomial data fitting

2001 Gilberto E. Urroz

y = mx+b, or y = b0 + b1x1 + + bkxk),

Simple linear regression

2001 Gilberto E. Urroz

The sum of square errors (SSE) to be minimized is

SSE (m, b) = ( y i mxi b) 2 .

from which we get the so-called normal equations:

2001 Gilberto E. Urroz

To illustrate graphically the behavior of SSE(m,b) we use the following function

2001 Gilberto E. Urroz

The following two lines modify the contour plot.

2001 Gilberto E. Urroz

Covariance and Correlation

The sample correlation coefficient for x,y is defined as

where sx, sy are the standard deviations of x and y, respectively, i.e.

Additional equations and definitions

2001 Gilberto E. Urroz

Also, the sample correlation coefficient is

Standard error of the estimate

If we have a set of n data points (xi, yi), then we can write

2001 Gilberto E. Urroz

A function for calculating linear regression of two variables

2001 Gilberto E. Urroz

As an example, we will use the following data set:

Confidence intervals and hypothesis testing in linear regression

y i = mxi+b are approximations to the parameters M and

2001 Gilberto E. Urroz

Confidence limits for regression coefficients:

For the slope ():

For the intercept ():

/Sxx]1/2 < < b + (t

where t follows the Students t distribution with = n 2, degrees of freedom, and n

Confidence interval for the mean value of Y at x = x0, i.e., mx0 + b:

Limits of prediction: confidence interval for the predicted value Y0=Y(x0):

Hypothesis testing on the slope, :

Hypothesis testing on the intercept , :

2001 Gilberto E. Urroz

A function for a comprehensive linear regression analysis

An example for linear regression analysis using function linregtable

2001 Gilberto E. Urroz

Reject the null hypothesis H0:Slope = 0.

2001 Gilberto E. Urroz

SCILAB function reglin

Graphical display of multivariate data

2001 Gilberto E. Urroz

if i==nr & j == nc then

= [147.47 165.42 170.60 184.84 198.05 209.96];

2001 Gilberto E. Urroz

2001 Gilberto E. Urroz