Chapter 06.

05
Adequacy of Models for Regression

After reading this chapter, you should be able to

1. determine if a linear regression model is adequate
2. determine how well the linear regression model predicts the response variable.

Quality of Fitted Model

In the application of regression models, one objective is to obtain an equation
y  f (x ) that best describes then response data points ( x1 , y1 ), ( x 2 , y 2 ),......., ( x n , y n ) .
Consequently, we are faced with answering two basic questions.
1. Does the model y  f (x) describe the data adequately, that is, is there an adequate
fit?
2. How well does the model predict the response variable (predictability)?
To answer these questions, let us limit our discussion to straight line models as
nonlinear models require a different approach. Some authors [1] claim that nonlinear model
parameters are not unbiased.
To exemplify our discussion, we will take example data to go through the process of
model evaluation. Given below is the data for the coefficient of thermal expansion vs.
temperature for steel. We assume a linear relationship between the data as
 (T )  a 0  a1T

Table 1 Values of coefficient of thermal expansion vs. temperature.

T (  F)  (μin/in/  F)
-340 2.45
-260 3.58
-180 4.52
-100 5.28
-20 5.86
60 6.36
Following the procedure for conducting linear regression as given in Chapter 06.03, we get
 (T )  6.0325  0.0096964T
Let us now look at how we can evaluate the adequacy of a linear regression model.

1. Plot the data and the regression model.

Figure 1 shows the data and the regression model. From a visual check, it looks like the
model explains the data adequately.

06.05.1
06.05.2 Chapter 06.05

Figure 1 Plot of coefficient of thermal expansion vs. temperature data points and regression
line.

2. Calculate the standard error of estimate.

The standard error of estimate is defined as
Sr
s / T 
n2
where
n
S r   ( i  a0  a1Ti ) 2
i 1

Table 2 Residuals for data.

Ti i a 0  a1Ti  i  a 0  a1Ti
-340 2.45 2.7357 -0.28571
-260 3.58 3.5114 0.068571
-180 4.52 4.2871 0.23286
-100 5.28 5.0629 0.21714
-20 5.86 5.8386 0.021429
60 6.36 6.6143 -0.25429
Adequacy of Regression Model 06.05.3

Table 2 shows the residuals of the data to calculate the sum of the square of residuals as
S r  ( 0.28571) 2  (0.068571) 2  (0.23286) 2  (0.21714) 2
 (0.021429) 2  ( 0.25429) 2
 0.25283
The standard error of estimate
Sr
s / T 
n2
0.25283

62
 0.25141
The units of s / T are same as the units of  . How is the value of the standard error of
estimate interpreted? We may say that on average the difference between the observed and
predicted values is 0.25141 μin/in/  F . Also, we can look at the value as follows. About
95% of the observed  values are between  2 s / T of the predicted value (see Figure 2).
This would lead us to believe that the value of  in the example is expected to be accurate
within  2 s / T =  2  0.25141 =  0.50282 μin/in/  F .

Figure 2 Plotting the linear regression line and showing the regression standard error.

One can also look at this criterion as finding if 95% of the scaled residuals for the model are
in the domain [-2,2], that is
06.05.4 Chapter 06.05

 i  a0  a1Ti
Scaled residual 
s / T
For the example,
s / T  0.25141
Table 4 Residuals and scaled residuals for data.
Ti i  i  a 0  a1Ti Scaled Residuals
-340 2.45 -0.28571 -1.1364
-260 3.58 0.068571 0.27275
-180 4.52 0.23286 0.92622
-100 5.28 0.21714 0.86369
-20 5.86 0.021429 0.085235
60 6.36 -0.25429 -1.0115
and the scaled residuals are calculated in Table 4. All the scaled residuals are in the [-2,2]
domain.

3. Calculate the coefficient of determination.

Denoted by r 2 , the coefficient of determination is another criterion to use for checking the
adequacy of the model.
To answer the above questions, let us start from the examination of some measures of
discrepancies between the whole data and some key central tendency. Look at the two
equations given below.
n
S r     i  ˆ i 
2

i 1
n
(1)
    i  a 0  a1Ti 
2

i 1
n
S t    i   
2
(2)
i 1
where
n

 i
  i 1

n
For the example data
6

 i
  i 1

6
2.45  3.58  4.52  5.28  5.86  6.36

6
 4.6750 μin/in/  F
n
S t    i   
2

i 1
Adequacy of Regression Model 06.05.5

 ( 2.2250) 2  ( 1.0950) 2  ( 0.15500) 2  (0.60500) 2  (1.1850) 2  (1.6850) 2

 10.783

Table 5 Difference between observed and average value.

Ti i i  
-340 2.45 -2.2250
-260 3.58 -1.0950
-180 4.52 -0.15500
-100 5.28 0.60500
-20 5.86 1.1850
60 6.36 1.6850
where S r is the sum of the square of the residuals (residual is the difference between the
observed value and the predicted value), and S t is the sum of the square of the difference
between the observed value and the average value.
What inferences can we make about the two equations? Equation (2) measures the
discrepancy between the data and the mean. Recall that the mean of the data is a measure of
a single point that measures the central tendency of the whole data. Equation (2) contrasts
with Equation (1) as Equation (1) measures the discrepancy between the vertical distance of
the point from the regression line (another measure of central tendency). This line obtained
by the least squares method gives the best estimate of a line with least sum of deviation. S r
as calculated quantifies the spread around the regression line.
The objective of least squares method is to obtain a compact equation that best
describes all the data points. The mean can also be used to describe all the data points. The
magnitude of the sum of squares of deviation from the mean or from the least squares line is
therefore a good indicator of how well the mean or least squares characterizes the
whole data. We can liken the sum of squares deviation around the mean, S t as the error or
variability in y without considering the regression variable x , while S r , the sum of
squares deviation around the least square regression line is error or variability in y
remaining after the dependent variable x has been considered.
The difference between these two parameters measures the error due to describing or
characterizing the data in one form instead of the other. A relative comparison of this
difference  S t  S r  , with the sum of squares deviation associated with the mean S t
describes a quantity called coefficient of determination, r 2
S  Sr
r2  t (5)
St
10.783  0.25283

10.783
 0.97655
Based on the value obtained above, we can claim that 97.7% of the original uncertainty in the
value of  can be explained by the straight-line regression model of
 (T )  6.0325  0.0096964T .
06.05.6 Chapter 06.05

Going back to the definition of the coefficient of determination, one can see that S t is the
variation without any relationship of y vs. x , while S r is the variation with the straight-
line relationship.
The limits of the values of r 2 are between 0 and 1. What do these limiting values of
r 2 mean? If r 2  0 , then S t  S r , which means that regressing the data to a straight line
does nothing to explain the data any further. If r 2  1 , then S r  0 , which means that the
straight line is passing through all the data points and is a perfect fit.

Caution in the use of r 2

a) The coefficient of determination, r 2 can be made larger (assumes no collinear
points) by adding more terms to the model. For instance, n  1 terms in a regression
equation for which n data points are used will give an r 2 value of 1 if there are no
collinear points.
b) The magnitude of r 2 also depends on the range of variability of the regressor ( x )
variable. Increase in the spread of x increases r 2 while a decrease in the spread of
x decreases r 2 .
c) Large regression slope will also yield artificially high r 2 .
d) The coefficient of determination, r 2 does not measure the appropriateness of the
linear model. r 2 may be large for nonlinearly related x and y values.
e) Large coefficient of determination r 2 value does not necessarily imply the regression
will predict accurately.
f) The coefficient of determination , r 2 does not measure the magnitude of the
regression slope.
g) These statements above imply that one should not choose a regression model solely
based on consideration of r 2 .

4. Find if the model meets the assumptions of random errors.

These assumptions include that the residuals are negative as well as positive to give a mean
of zero, the variation of the residuals as a function of the independent variable is random, the
residuals follow a normal distribution, and that there is no auto correlation between the data
points.
To illustrate this better, we have an extended data set for the example that we took.
Instead of 6 data points, this set has 22 data points (Table 6). Drawing conclusions from
small or large data sets for checking assumption of random error is not recommended.

Table 6 Instantaneous thermal expansion coefficient as a function of temperature.

Adequacy of Regression Model 06.05.7

Instantaneous
Temperature
Thermal Expansion
F μin/in/F
80 6.47
60 6.36
40 6.24
20 6.12
0 6.00
-20 5.86
-40 5.72
-60 5.58
-80 5.43
-100 5.28
-120 5.09
-140 4.91
-160 4.72
-180 4.52
-200 4.30
-220 4.08
-240 3.83
-260 3.58
-280 3.33
-300 3.07
-320 2.76
-340 2.45
06.05.8 Chapter 06.05

Figure 3 Plot of thermal expansion coefficient vs. temperature data points and regression line
for more data points.
Regressing the data from Table 2 to the straight line regression line
 (T )  a 0  a1T
and following the procedure for conducting linear regression as given in Chapter 06.03, we
get (Figure 3)
  6.0248  0.0093868T
Adequacy of Regression Model 06.05.9

Figure 4 Plot of residuals.

Figure 4 shows the residuals for the example as a function of temperature. Although
the residuals seem to average to zero, but within a range, they do not exhibit this zero mean.
For an initial value of T , the averages are below zero. For the middle values of T , the
averages are below zero, and again for the final values of T , the averages are below zero.
This may be considered a violation of the model assumption.
Figure 4 also shows the residuals for the example are following a nonlinear variance.
This is a clear violation of the model assumption of constant variance.
06.05.10 Chapter 06.05

Figure 5 Histogram of residuals.

Figure 5 shows the histogram of the residuals. Clearly, the histogram is not showing
a normal distribution, and hence violates the model assumption of normality.
To check that there is no autocorrelation between observed values, the following rule
of thumb can be used. If n is the number of data points, and q is the number of times the
sign of the residual changes, then if
(n  1) n 1
 n 1  q   n 1 ,
2 2
you most likely do not have an autocorrelation. For the example, n  22 , then
(22  1) 22  1
 22  1  q   22  1
2 2
5.9174  q  15.083
is not satisfied as q  2 . So this model assumption is violated.

References

ADEQUACY OF REGRESSION MODELS

Topic Adequacy of Regression Models
Summary Textbook notes of Adequacy of Regression Models
Major General Engineering
Adequacy of Regression Model 06.05.11

Authors Autar Kaw, Egwu Kalu

Date May 16, 2020
Web Site http://numericalmethods.eng.usf.edu