
Regression of the second kind
When there are uncer tainties in both axes, there is no reason to emphasize the x
axis, and the same procedure can be followed commuting the roles of x and y.
Generally another regression line is obtained, which is given by:
y
0
¼ a
0
þ n
0
x ð7:22Þ
with another pair of coefficients:
n
0
¼
s
2
y
s
xy
a
0
¼
yy n
0
xx
ð7:23Þ
This regression line minimizes the sum of the square of the residual abscissae:
SSR
x
¼
X
N
i ¼1
x
i
y
0
a
0
n
0
2
ð7:24Þ
If the two lines are close to each other or, in other words, if:
n
0
n and therefore a
0
a ð7:25Þ
then it can be said that there is a good correlation between the two physical
quantities x and y. The correlation coefficient defined by:
R ¼
s
xy
s
x
s
y
hence R
2
¼
n
n
0
ð7:26Þ
is a measure of the interdependenc e of x and y. It is not a measure of the quality
of the fit or of the accuracy of the estimates of the coefficients, since jRj¼1 for
any fit based on only two sets of points. The estimates of the errors on a and n
are calculated in ‘Confi dence in the coefficients’, above.
Orthogonal regression
If the two regressions of the second kind are calculated and different results are
obtained, the problem is to choose the coefficients: which pair is the closest to
the reality? Since each pair of coefficients is obtained assuming that one variable
is exactly known, it is likely that the best set is neither of them and instead lies in
between, but where?
There are several answers to that question, none of them being really satis-
factory. One recipe is to take an average slope:
nn ¼
n þ n
0
2
ð7:27Þ
or a weighted average slope:
nn ¼
"
y
n þ "
x
n
0
2
ð7:28Þ
150 Ventilation and Airflow in Buildings
Kommentare zu diesen Handbüchern