Adama Science Technology University @ 2005 E.

Nonlinear Models for Regression

After reading this chapter, you should be able to
1. derive constants of nonlinear regression models,
2. use in examples, the derived formula for the constants of the nonlinear regression
model, and
3. linearize (transform) data to find constants of some nonlinear regression models.

From fundamental theories, we may know the relationship between two variables.
An example in chemical engineering is the Clausius-Clapeyron equation that relates vapor
pressure P of a vapor to its absolute temperature, T .
log P   A 
B
(1)
T
where A and B are the unknown parameters to be determined. The above equation is not
linear in the unknown parameters. Any model that is not linear in the unknown parameters is
described as a nonlinear regression model.

Nonlinear models using least squares

The development of the least squares estimation for nonlinear models does not
generally yield equations that are linear and hence easy to solve. An example of a nonlinear
regression model is the exponential model.

Exponential model
Given x1 ,y1  , x2 ,y 2  , . . . xn ,y n  , best fit y  ae bx to the data. The variables a and
b are the constants of the exponential model. The residual at each data point xi is
Ei  yi  ae bxi (2)
The sum of the square of the residuals is
n
S r   Ei2
i 1

 
n
  yi  ae bxi
2
(3)
i 1

To find the constants a and b of the exponential model, we minimize S r by differentiating

with respect to a and b and equating the resulting equations to zero.

S r
  
n
  2 yi  ae bxi  e bxi  0
a i 1

S r
  
n
  2 yi  ae bxi  axi e bxi  0 (4a,b)
b i 1
or
n n
  yi e bxi  a e 2bxi  0
i 1 i 1

11
Adama Science and Technology University @ 2005

n n

 yi xi e bxi  a xi e 2bxi  0
i 1 i 1
(5a,b)

Equations (5a) and (5b) are nonlinear in a and b and thus not in a closed form to be
solved as was the case for linear regression. In general, iterative methods (such as
Gauss-Newton iteration method, method of steepest descent, Marquardt's method, direct
search, etc) must be used to find values of a and b .
However, in this case, from Equation (5a), a can be written explicitly in terms of b as
n

ye i
bxi

a i 1
n
(6)
e
i 1
2 bxi

Substituting Equation (6) in (5b) gives

n

n ye i
bxi
n

 yi xi e bxi  i 1
n x e i
2 bxi
0 (7)
i 1
e
i 1
2 bxi i 1

This equation is still a nonlinear equation in b and can be solved best by numerical methods
such as the bisection method or the secant method.

Example 1
Many patients get concerned when a test involves injection of a radioactive material. For
example for scanning a gallbladder, a few drops of Technetium-99m isotope is used. Half of
the technetium-99m would be gone in about 6 hours. It, however, takes about 24 hours for
the radiation levels to reach what we are exposed to in day-to-day activities. Below is given
the relative intensity of radiation as a function of time.

Table 1 Relative intensity of radiation as a function of time

t (hrs) 0 1 3 5 7 9
 1.000 0.891 0.708 0.562 0.447 0.355

If the level of the relative intensity of radiation is related to time via an exponential formula
  Ae t , find
a) the value of the regression constants A and  ,
b) the half-life of Technium-99m, and
c) the radiation intensity after 24 hours.

Solution
a) The value of  is given by solving the nonlinear Equation (7),
Nonlinear Regression

n  e i
ti
n
f      i t i e ti  i 1
n t e  i
2 ti
0 (8)
i 1
e i 1
2 t i i 1

and then the value of A from Equation (6),

n

 e  i
ti

A i 1
n
(9)
e
i 1
2 t i

Equation (8) can be solved for  using bisection method. To estimate the initial
guesses, we assume λ  0.120 and λ  0.110 . We need to check whether these values
first bracket the root of f    0 . At λ  0.120 , the table below shows the evaluation of
f  0.120 .

Table 2 Summation value for calculation of constants of model

i ti  i  i t i e ti  i e ti e 2ti t i e 2ti
1 0 1 0.00000 1.00000 1.00000 0.00000
2 1 0.891 0.79205 0.79205 0.78663 0.78663
3 3 0.708 1.4819 0.49395 0.48675 1.4603
4 5 0.562 1.5422 0.30843 0.30119 1.5060
5 7 0.447 1.3508 0.19297 0.18637 1.3046
6 9 0.355 1.0850 0.12056 0.11533 1.0379
6

i 1
6.2501 2.9062 2.8763 6.0954
From Table 2
n6
6

 t e
i 1
i i
0.120ti
 6.2501
6

 e
i 1
i
0.120ti
 2.9062
6

i 1
e2 0.120ti  2.8763
6

t e 
i 1
i
2 0.120ti
 6.0954

f  0.120  6.2501  6.0954

2.9062
2.8763
 0.091357
Similarly
f  0.110  0.10099

7
Adama Science and Technology University @ 2005

Since
f  0.120  f  0.110  0 ,
the value of λ falls in the bracket of  0.120,0.110 . The next guess of the root then is
 0.120   0.110

2
 0.115
Continuing with the bisection method, the root of f    0 is found as   0.11508 . This
value of the root was obtained after 20 iterations with an absolute relative approximate error
of less than 0.000008%.
From Equation (9), A can be calculated as
6

  e i
ti

A i 1
6

e 
i 1
2 ti

1  e 0.11508( 0 )  0.891  e 0.11508(1)  0.708  e 0.11508( 3) 

0.562  e 0.11508( 5)  0.447  e 0.11508( 7 )  0.355  e 0.11508( 9 )

e 2 ( 0.11508)( 0 )  e 2 ( 0.11508)(1)  e 2 ( 0.11508)( 3) 
e 2 ( 0.11508)( 5)  e 2( 0.11508)( 7 )  e 2 ( 0.11508)( 9 )
2.9373

2.9378
 0.99983
The regression formula is hence given by
  0.99983 e 0.11508t

1
b) Half life of Technetium-99m is when   
2 t 0

0.99983  e 0.11508t 
1
0.99983e 0.11508(0)
2
e 0.11508t  0.5
 0.11508t  ln(0.5)
t  6.0232 hours
c) The relative intensity of the radiation after 24 hrs is
  0.99983  e 0.1150824
 6.3160 102
6.3160  10 2
This implies that only  100  6.3171% of the initial radioactive intensity is left
0.99983
after 24 hrs.
Nonlinear Regression

Figure 1 Relative intensity of radiation as a function of temperature using an

exponential regression model.

Growth model
Growth models common in scientific fields have been developed and used
successfully for specific situations. The growth models are used to describe how something
grows with changes in the regressor variable (often the time). Examples in this category
include growth of thin films or population with time. Growth models include
a
y (10)
1  be cx
a
where a, b and c are the constants of the model. At x  0 , y  and as x   ,
1 b
y a.
The residuals at each data point xi , are
a
Ei  y i  (11)
1  be cxi
The sum of the square of the residuals is
n
S r   Ei2
i 1
2
n
 a 
   yi   cxi 
(12)
i 1  1  be 

7
Adama Science and Technology University @ 2005

To find the constants a , b and c we minimize S r by differentiating with respect to a ,

b and c , and equating the resulting equations to zero.

S r
 
n 

2e cxi ae c xi  yi e cxi  b    0 ,
a 
i 1  e cxi  b
2
 
 
S r
 
n  cx

2ae i byi  e i  yi  a  
cx
 0,

b 
i 1  e cxi  b 
3 
 
S r
 
n 

 2abxi e cxi byi  e cxi  yi  a  
 0.
 (13a,b,c)
c 
i 1  e cxi  b
3
 
 
One can use the Newton-Raphson method to solve the above set of simultaneous nonlinear
equations for a , b and c .

Example 2
The height of a child is measured at different ages as follows.

Table 3 Height of the child at different ages.

t ( yrs) 0 5.0 8 12 16 18
H (in ) 20 36.2 52 60 69.2 70

Estimate the height of the child as an adult of 30 years of age using the growth model,
a
H
1  be ct
Solution
The saturation growth model of height, H vs. age, t is given as
a
H
1  be ct
where the constants a , b and c are the roots of the simultaneous nonlinear equation system

6
 
 2e cti ae cti  H i e cti  b  
  0

i 1  
e b
cti 2
 

6 

2ae cti bH i  e cti H i  a   
  0

i 1  
e cti  b
3
 

6  ct

  2abt i e bH i  e H i  a    0
ct  

i i

(14a,b,c)

i 1  
e cti  b
3 
 
Nonlinear Regression

We need initial guesses of the roots to get the iterative process started to find the root of
those equations. Suppose we use three of the given data points such as (0, 20), (12, 60) and
(18, 70) to find the initial guesses of roots; we have
a
20 
1  be c ( 0)
a
60 
1  be c (12)
a
70 
1  be c (18)
One can solve three unknowns a , b and c from the three equations as
a  7.5534  101
b  2.7767
c  1.9772  101
Applying the Newton-Raphson method for simultaneous nonlinear equations, one can get the
roots
a  7.4321101
b  2.8233
c  2.1715 101
The saturation growth model of the height of the child then is
7.4321  101
H 1
1  2.8233e 2.171510 t
The height of the child as an adult of 30 years of age is
7.4321  101
H 1
1  2.8233e 2.171510 (30)
 74 "
Polynomial Models
Given n data points ( x1 , y1 ), ( x2 , y 2 ),......, ( xn , y n ) use least squares method to regress the
data to an m th order polynomial.
y  a0  a1 x  a2 x 2    am x m , m  n (15)
The residual at each data point is given by
Ei  yi  a0  a1 xi  . . .  am xim (16)
The sum of the square of the residuals is given by
n
S r   Ei2
i 1
(17)
 
n
  y i  a0  a1 xi  . . .  a m x m 2
i
i 1
To find the constants of the polynomial regression model, we put the derivatives with respect
to a i to zero, that is,

7
Adama Science and Technology University @ 2005

Figure 2 Height of child as a function of age saturation growth model.

S r
 
n
  2 y i  a 0  a1 xi  . . .  a m xim (1)  0
a 0 i 1
S r
 
n
  2 y i  a 0  a1 xi  . . .  a m xim ( xi )  0
a1 i 1
. . . . . . . . . . . . . . . . . . . . . (18)
. . . . . . . . . . . . . . . . . . . . .
S r
 
n
  2 y i  a 0  a1 xi  . . .  a m xim ( xim )  0
a m i 1
Setting those equations in matrix form gives
   
   n 
 n  n
  n

  xi  . . .  xim   a   yi 
  i 1   i 1    0   i 1 
 n  a   n 
   xi    xi2  . . .  xim1   1     xi yi
n n
 (19)
  i 1   i 1   i 1  . . .  i 1 
. . . . . . . . . . .  a  . . . 
  m   n 
 n m   n m1   n 2m    xim yi 
  xi    xi  . . .  xi    i 1 
 i 1   i 1   i 1 
The above are solved for a0 , a1 ,..., am
Nonlinear Regression

Example 3
To find contraction of a steel cylinder, one needs to regress the thermal expansion coefficient
data to temperature

Table 4 The thermal expansion coefficient at given different temperatures

Temperature, T Coefficient of thermal

( F) expansion,  (in/in/  F)
80 6.47  10 6
40 6.24  10 6
-40 5.72  10 6
-120 5.09  10 6
-200 4.30  10 6
-280 3.33  10 6
-340 2.45  10 6

Fit the above data to α  a0  a1T  a2T 2

Solution
Since α  a0  a1T  a2T 2 is the quadratic relationship between the thermal expansion
coefficient and the temperature, the coefficients a0 , a1 , a2 are found as follows
  n   n 2   n 
 n   Ti    Ti   i
  i 1   i 1   a   i 1 

0
 n   n 2  n 3     n 
  Ti    Ti    Ti   a1    Ti  i
 i n1   i 1   i 1      i 1 
n 
 n 4    2   T 2
a
 T 2   n 3
  i   Ti    Ti   i i
 i 1


 i 1   i 1   i 1  

Table 5 Summations for calculating constants of model

i T (  F)  (in/in/  F) T 2 T3
6.4700 106
1 80 6.4000  103 5.1200  105
6.2400  10 6
2 40 1.6000  10 3 6.4000  10 4
5.7200  10 6
3 -40 1.6000  10 3  6.4000  10 4
5.0900  10 6
4 -120 1.4400  10 4  1.7280  10 6
4.3000  10 6
5 -200 4.0000  10 4  8.0000  10 6

7
Adama Science and Technology University @ 2005

3.3300  10 6
6 -280 7.8400  10 4  2.1952  10 7
2.4500  10 6
7 -340 1.1560  10 5  3.9304  10 7
7

  8.6000  10 2 3.3600  10 5
i 1 2.5800  10 5  7.0472  10 7

Table 5 (cont)
i T 4
T  T 2 
1 4.0960  10 7
5.1760  10 4 4.1408  10 2
2 2.5600  10 6 2.4960  10 4 9.9840  10 3
 2.2880  10 4
3 2.5600  10 6 9.1520  10 3
 6.1080  10 4
4 2.0736  108 7.3296  10 2
 8.6000  10 4
5 1.6000  109 1.7200  10 1
 9.3240  10 4
6 6.1466  109 2.6107  10 1
1.3363  1010  8.3300  10 4
7 2.8322  10 1
7
2.1363  1010  2.6978  10 3

i 1
8.5013  10 1

n7
7

T
i 1
i   8.6000  10 2
7

T
i 1
i
2
 2.5580  10 5
7

T
i 1
i
3
  7.0472  10 7
7

T
i 1
i
4
 2.1363 1010
7


i 1
i  3.3600  10 5
7

T 
i 1
i i   2.6978  10 3
7

T
i 1
i
2
 i  8.5013  10 1
We have
Nonlinear Regression

 7.0000  8.6000  10 22.5800  10 5  a0   3.3600  10 5 

   
 8.600  10
2
2.5800  10 5  7.0472  10 7   a1    2.6978  10 3 
 2.5800  10 5  7.0472  10 7 2.1363  1010  a 2   8.5013  10 1 

Solving the above system of simultaneous linear equations, we get
a0   6.0217  10 
6

 a    6.2782  10 9 
 1  
 11 
a 2   1.2218  10 
The polynomial regression model is
  a0  a1T  a2T 2
 6.0217  10 6  6.2782  10 9 T  1.2218  10 11T 2

Transforming the data to use linear regression formulas

Examination of the nonlinear models above shows that in general iterative methods are
required to estimate the values of the model parameters. It is sometimes useful to use simple
linear regression formulas to estimate the parameters of a nonlinear model. This involves
first transforming the given data such as to regress it to a linear model. Following the
transformation of the data, the evaluation of model parameters lends itself to a direct solution
approach using the least squares method. Data for nonlinear models such as exponential,
power, and growth can be transformed.
Exponential Model
As given in Example 1, many physical and chemical processes are governed by the
exponential function.
  ae bx (20)
Taking natural log of both sides of Equation (20) gives
ln   ln a  bx (21)
Let
z  ln 
a0  ln a implying a  e ao
a1  b
then
z  a0  a1 x (22)

7
Adama Science and Technology University @ 2005

Figure 3 Second-order polynomial regression model for coefficient of thermal expansion

as a function of temperature.

The data z versus x is now a linear model. The constants a 0 and a1 can be found using the
equation for the linear model as
n n n
n xi z i   xi  z i
a1  i 1 i 1 i 1
2
n
  n
n xi2    xi  (23a,b)
i 1  i 1 
_ _
a 0  z  a1 x
Now since a 0 and a1 are found, the original constants with the model are found as
b  a1
(24a,b)
a  e a0

Example 4
Repeat Example 1 using linearization of data.

Solution
  Ae t
Nonlinear Regression

ln    ln  A  t
Assuming
y  ln 
a0  ln  A
a1  
We get
y  a0  a1t
This is a linear relationship between y and t .
n n n
n t i y i   t i  y i
a1  i 1 i 1 i 1
2
  n n
n t    t i  i
2

i 1  i 1 
a0  y  a1t (25a,b)

Table 6 Summations of data to calculate constants of model.

i ti i yi  ln  i t i yi t i2
1 0 1 0.00000 0.0000 0.0000
2 1 0.891 -0.11541 -0.11541 1.0000
3 3 0.708 -0.34531 -1.0359 9.0000
4 5 0.562 -0.57625 -2.8813 25.000
5 7 0.447 -0.80520 -5.6364 49.000
6 9 0.355 -1.0356 -9.3207 81.000
6

 i 1
25.000 -2.8778 -18.990 165.00

n6
6

t
i 1
i  25.000
6

y
i 1
i  2.8778
6

t y
i 1
i i  18.990
6

t
i 1
i
2
 165.00

From Equation (25a,b) we have

6 18.990  25 2.8778
a1 
6165.00  25
2

 0.11505

7
Adama Science and Technology University @ 2005

 2.8778
  0.11505
25
a0 
6 6
4
 2.6150  10
Since
a0  ln  A
A  e a0
4
 e 2.615010
 0.99974
  a1  0.11505
The regression formula then is
  0.99974  e 0.11505t
Compare the formula to the one obtained without data linearization,
  0.99983  e0.11508t
b) Half-life is when
1
  
2 t 0

0.99974  e 0.11505t  0.99974 e 0.11505( 0 )

1
2
 0.11508t
e  0.5
 0.11505t  ln( 0.5)
t  6.0248 hours
c) The relative intensity of radiation, after 24 hours is
  0.99974e 0.1150524
 0.063200
6.3200  10 2
This implies that only  100  6.3216% of the initial radioactivity is left after
0.99974
24 hours.
Logarithmic Functions
The form for the log regression models is
y   0  1 ln x  (26)
This is a linear function between y and ln x  and the usual least squares method applies in
which y is the response variable and ln x  is the regressor.
Nonlinear Regression

Figure 4 Exponential regression model with transformed data for relative intensity of
radiation as a function of temperature.

Example 5
Sodium borohydride is a potential fuel for fuel cell. The following overpotential   vs.
current i  data was obtained in a study conducted to evaluate its electrochemical kinetics.

Table 7 Electrochemical Kinetics of borohydride data.

 (V ) -0.29563 -0.24346 -0.19012 -0.18772 -0.13407 -0.0861
i ( A) 0.00226 0.00212 0.00206 0.00202 0.00199 0.00195

At the conditions of the study, it is known that the relationship that exists between the
overpotential   and current i  can be expressed as
  a  b ln i (27)
where a is an electrochemical kinetics parameter of borohydride on the electrode. Use the
data in Table 7 to evaluate the values of a and b .

Solution
Following the least squares method, Table 8 is tabulated where
x  ln i
y 
We obtain

7
Adama Science and Technology University @ 2005

y  a  bx (28)
This is a linear relationship between y and x , and the coefficients b and a are found as
follow
n n n
n  xi y i   xi  y i
b i 1 i 1 i 1
2
 
n n
n x    xi  2
1
i 1  i 1 
a  y  bx (29a,b)

Table 8 Summation values for calculating constants of model

# i y  x  ln(i) x 2 x y
1 0.00226 -0.29563 -6.0924 37.117 1.8011
2 0.00212 -0.24346 -6.1563 37.901 1.4988
3 0.00206 -0.19012 -6.1850 38.255 1.1759
4 0.00202 -0.18772 -6.2047 38.498 1.1647
5 0.00199 -0.13407 -6.2196 38.684 0.83386
6 0.00195 -0.08610 -6.2399 38.937 0.53726
6


i 1
0.012400 -1.1371 -37.098 229.39 7.0117

n6
6

x
i 1
i  37.098
6

y
i 1
i  1.1371
6

x y
i 1
i i  7.0117
6

x
i 1
2
i  229.39

67.0117    37.098 1.1371

b
6229.39   37.098
2

 1.3601

 1.1371  37.098
a   1.3601
6 6
 8.5990
Hence
  8.5990  1.3601 ln i
Nonlinear Regression

Figure 5 Overpotential as a function of current.  (V )

Power Functions
The power function equation describes many scientific and engineering phenomena. In
chemical engineering, the rate of chemical reaction is often written in power function form as
y  ax b (30)
The method of least squares is applied to the power function by first linearizing the data (the
assumption is that b is not known). If the only unknown is a , then a linear relation exists
between x b and y . The linearization of the data is as follows.
ln  y   ln a   b ln x  (31)
The resulting equation shows a linear relation between ln y  and ln x  .
Let
z  ln y
w  ln( x)
a0  ln a implying a  e a0
a1  b
we get
z  a0  a1 w (32)

7
Adama Science and Technology University @ 2005

n n n
n wi z i   wi  z i
a1  i 1 i 1 i 1
2
 n
 n
n w    wi  2
i
i 1  i 1  (33a,b)
n n

 zi w i
a0  i 1
 a1 i 1

n n
Since a 0 and a1 can be found, the original constants of the model are
b  a1
(34a,b)
a  e a0

Example 6
The progress of a homogeneous chemical reaction is followed and it is desired to evaluate the
rate constant and the order of the reaction. The rate law expression for the reaction is known
to follow the power function form
 r  kC n (35)
Use the data provided in the table to obtain n and k .

Table 9 Chemical kinetics.

C A (gmol/l) 4 2.25 1.45 1.0 0.65 0.25 0.006
 rA (gmol/l  s) 0.398 0.298 0.238 0.198 0.158 0.098 0.048

Solution
Taking the natural log of both sides of Equation (35), we obtain
ln  r   ln k   n ln C 
Let
z  ln  r 
w  ln C 
a0  ln(k ) implying that k  e a0 (36)
a1  n (37)
We get
z  a0  a1 w
This is a linear relation between z and w , where
n n n
n wi z i   wi  z i
a1  i 1 i 1 i 1
2
 n
 n
n w    wi  2
i
i 1  i 1 
Nonlinear Regression

 n   n 
  zi    wi 
a 0   i 1  a  i 1  (38a,b)
 n  1 n 
   
   

Table 10 Kinetics rate law using power function

i C r w z w z w2
1 4 0.398 1.3863 -0.92130 -1.2772 1.9218
2 2.25 0.298 0.8109 -1.2107 -0.9818 0.65761
3 1.45 0.238 0.3716 -1.4355 -0.5334 0.13806
4 1 0.198 0.0000 -1.6195 0.0000 0.00000
5 0.65 0.158 -0.4308 -1.8452 0.7949 0.18557
6 0.25 0.098 -1.3863 -2.3228 3.2201 1.9218
7 0.006 0.048 -5.1160 -3.0366 15.535 26.173
7


i 1
-4.3643 -12.391 16.758 30.998

n7
7

w
i 1
i  4.3643
7

z
i 1
i  12.391
7

w z
i 1
i i  16.758
7

w
i 1
2
i  30.998

From Equation (38a,b)

7  16.758   4.3643   12.391
a1 
7  30.998   4.3643
2

 0.31943

 12.391  4.3643
a0   .31943
7 7
 1.5711
From Equation (36) and (37), we obtain
k  e 1.5711
 0.20782

7
Adama Science and Technology University @ 2005

n  a1
 0.31941
Finally, the model of progress of that chemical reaction is
 r  0.20782  C 0.31941

Figure 6 Kinetic chemical reaction rate as a function of concentration.

Growth Model
Growth models common in scientific fields have been developed and used successfully for
specific situations. The growth models are used to describe how something grows with
changes in a regressor variable (often the time). Examples in this category include growth of
thin films or population with time. In the logistic growth model, an example of a growth
model in which a measurable quantity y varies with some quantity x is
ax
y (39)
bx
For x  0 , y  0 while as x   , y  a . To linearize the data for this method,
1 b x

y ax
(40)
b1 1
 
ax a
Let
Nonlinear Regression

1
z
y
1
w  ,
x
1 1
a 0  implying that a 
a a0
b a
a1  implying b  a1  a  1
a a0
Then
z  a0  a1 w (41)
The relationship between z and w is linear with the coefficients a 0 and found as follows.
n n n
n wi z i   wi  z i
a1  i 1 i 1 i 1
2

n
 n
n w    wi 
2
i
i 1  i 1 

 n   n 
  zi    wi 
a0   i 1  a1  i 1  (42a,b)
 n   n 
   
   
Finding a 0 and a1 , then gives the constants of the original growth model as
1
a
a0
a
b 1 (43a,b)
a0

NONLINEAR REGRESSION
Topic Nonlinear Regression