You are on page 1of 9

NONLINEAR REGRESSION Exponential Models

In many cases the form of the functional relationship between the predictor variable
and the response variable is not linear. However, in some cases the form of the func-
tional relationship may be known either through previous experience or on the basis of
assumptions made the variables.

In population growth models, the rate of growth of a population is often proportional


to the population size and so if the population size at time t is denoted by N (t) the
relationship between time and population size can be obtained by solving the differential
equation

dN
= kN k constant.
dt

The solution of this differential equation gives

N (t) = N0 ekt

where the population size at time zero is N0 and so the response variable N and the
predictor variable t are related by an exponential function.

The parameterization in a nonlinear model is not unique as the model

y = N0 ekt

can also be represented as

y = β0 β1t

where the second model is obtained from the first by defining β0 = N0 and β1 = ek .

9.1
NONLINEAR REGRESSION Exponential Models

Example

Catalase is a common enzyme which is lost with exposure to sunlight. The concentration
of catalase (µg/10 mL) as a function of time (minutes) over a period of eighty minutes
is noted by Mitchell and Anderson (1965) with the following results.

Time (t) Concentration (y)

0 121.0
10 74.0
30 30.0
50 12.0
60 6.7
70 3.7
80 2.0

A plot of concentration against time shows that the relationship between concentration
and time is not linear and it has been suggested that the relationship may be of the
form

y = aekt a, k constants

Taking logarithms gives

ln y = ln a + kt

and so a linear relationship between ln y and t.

9.2
NONLINEAR REGRESSION Exponential Models

Fitting the linear model using Minitab gives

The regression equation is


ln(y) = 4.86 - 0.0503 t

Predictor Coef SE Coef T P


Constant 4.85625 0.07139 68.03 0.000
t -0.050347 0.001392 -36.16 0.000

S = 0.103666 R-Sq = 99.6% R-Sq(adj) = 99.5%

Analysis of Variance

Source DF SS MS F P
Regression 1 14.050 14.050 1307.43 0.000
Residual Error 5 0.054 0.011
Total 6 14.104

with estimates of k as -0.05, of ln a as 4.856 and so of a as 128.5.

This gives the functional relationship between concentration and time as

y = 128.5e−0.05t .

9.3
NONLINEAR REGRESSION Power Function Models

The relationship between the response variable y and the predictor variable x can also
be in the form a power function where

y = β0 xβ1

where taking logarithms gives

ln y = ln β0 + β1 ln x

and so a linear relationship between ln y and ln x.

Example

The data below give the length (L) in centimetres and the weight (W ) in pounds for
yellow fin tuna in the Central Pacific region.

L W

70 14.3
80 21.5
90 30.8
100 42.5
110 56.8
120 74.1
130 94.7
140 119.0
160 179.0
180 256.0

A plot of weight against length indicates that the relationship is not linear and it is
thought that the relationship between length and weight is a power function of the form

W = β0 L β1 .

Taking logarithms gives

ln W = ln β0 + β1 ln L

and so a linear relationship between ln W and ln L.

9.4
NONLINEAR REGRESSION Power Function Models

Fitting the linear model using Minitab gives

The regression equation is


ln(W) = - 10.3 + 3.06 ln(L)

Predictor Coef SE Coef T P


Constant -10.3233 0.0055 -1880.55 0.000
ln(L) 3.05588 0.00116 2637.96 0.000

S = 0.00105083 R-Sq = 100.0% R-Sq(adj) = 100.0%

Analysis of Variance

Source DF SS MS F P
Regression 1 7.6842 7.6842 6958826.13 0.000
Residual Error 8 0.0000 0.0000
Total 9 7.6842

with estimates of β1 as 3.06, of ln β0 as -10.32 and so of β0 as 0.000033. This gives the


functional relationship between length and weight as

W = 0.000033L3.06 .

9.5
NONLINEAR REGRESSION Logistic Growth Curves

The exponential model assumes that the population grows in size indefinitely. Most
populations do not grow in size indefinitely with the size being restricted by limited
food supply, lack of space, increasing predator activity or other control mechanisms.
Assuming an upper bound to the population size of K individuals, called the carrying
capacity of the environment, can lead to a population growth model of the form

N (t) = K 1 − e−kt
which was first proposed by E. A. Mitscherlich in 1939 and fits growth in agricultural
populations very well. A model which fits the growth of biological populations better is
one which assumes that the growth rate is related to both the current population size,
N (t) and the remaining carrying capacity of the environment, K − N (t).

In this model, the population grows slowly at first, reaches its maximum growth rate at
a population size of K/2 individuals and then grows more slowly towards the maximum
size of K individuals. A graph of this type of population growth is an S-shaped or
sigmoid curve. The simplest function having a sigmoid curve graph is the logistic
function which was introduced into population dynamics by the Belgian mathematician
P. F. Verhulst in 1838. The equation specifying the logistic growth curve is

K
N (t) = b, k, K positive constants
1 + be−kt

and is obtained by solving the differential equation

dN
= kN (K − N ).
dt

A possible approach to fitting a regression of population size on time is to linearize the


problem. This involves considering transformations of the predictor or response so that
the relationship between the transformed values is linear. An algebraic transformation
of Mitscherlich’s growth model gives ln[(K − N )/K] = −kt and so there is a linear
relationship between the variables ln[(K − N )/K] and t. Similarly, a linearization of
the logistic growth curve gives ln[(K − N )/N ] = ln b − kt which is a linear equation of
ln[(K − N )/N ] against t.

As the equation of the logistic curve involves three constants, namely b, k and K, any
attempt to fit the curve to a set of observed points by a linearization requires that one
of the constants is estimated before the other two are fitted in the linearized model.
A graph of the observed points usually allows a reasonable estimate of the carrying
capacity K and so the other two constants can be estimated from a suitable linear plot.

9.6
NONLINEAR REGRESSION Logistic Growth Curves

In 1913, Tor Carlson grew Sacharomyces, a common type of yeast, in laboratory cul-
tures. He began with a few yeast cells in a suitable nutritive solution and recorded
the population size for the next eighteen hours. For the results given below, assuming
logistic growth, find the constants b, k and K of the appropriate logistic growth curve.

Hours (t) Amt of yeast (N)

0 9.6
1 18.3
2 29.0
3 47.2
4 71.1
5 119.1
6 174.6
7 257.3
8 350.7
9 441.0
10 513.3
11 559.7
12 594.8
13 629.4
14 640.8
15 651.1
16 655.9
17 659.6
18 661.8

A plot of the above observations suggests that the observations do follow a logistic
growth curve, and the carrying capacity can be estimated from the graph to be 665.

With K = 665, the following values of ln[(K − N )/N ] are obtained and a plot of these
values against t produces a graph in which the points do appear to lie on a straight line
and so the assumption that the observations follow a logistic growth curve is justified.

9.7
NONLINEAR REGRESSION Logistic Growth Curves

Hours (t) ln[(K-N)/N]

0 4.22
1 3.56
2 3.09
3 2.57
4 2.12
5 1.52
6 1.03
7 0.46
8 -0.11
9 -0.68
10 -1.22
11 -1.67
12 -2.14
13 -2.87
14 -3.28
15 -3.85
16 -4.28
17 -4.81
18 -5.33

Fitting a linear regression to these values gives

The regression equation is


ln((K-N)/N) = 4.16 - 0.531 Time

Analysis of Variance

Source DF SS MS F P
Regression 1 160.52 160.52 43027.89 0.000
Residual Error 17 0.06 0.00
Total 18 160.58

with the line of best fit having intercept of 4.16 and slope -0.531, so that the appropriate
estimates of the constants are b = e4.16 = 64 and k = 0.5 to give the appropriate
sigmoid curve as
665
N (t) =
1 + 64e−0.5t

9.8
NONLINEAR REGRESSION Logistic Growth Curves

In SPSS the nonlinear regression can be fitted directly using the Analyse menu where
we specify the dependent variable, the nonlinear model and initial estimates of the
parameters as follows

Analyse>Regression>Nonlinear
Dependent [y]
Model Expression [b1/(1 + b2*EXP(-b3*x))]
Parameters
b1(665)
b2(50)
b3(1)

to get the following estimates of the parameters.

Parameter Estimates
Parameter Estimate Std. Error 95% Confidence Interval
Lower Bound Upper Bound
b1 663.022 1.703 659.412 666.632
b2 71.576 2.927 65.371 77.782
b3 .547 .006 .535 .559

9.9

You might also like