
140
We can make predictions from our line of best fit in two places - the
HOME
view and the
PLOT
view. The hp 38g was able to do this only from the
HOME
view.
Predicting using
PREDY
In the
HOME
view we use the functions
PREDY
and
PREDX
from the Stat-Two section of the
MATH
menu. The functions
PREDX
and
PREDY
use whatever was the last line of best
fit calculated. It is up to you to ensure that the
one you want used was the one last graphed.
If I want to predict a y value for x = 3, then I simply type
PREDY(3)
into the
HOME
view as shown right.
Many people choose to simply type ‘
PREDY
’ using the
ALPHA
button instead
of going through the
MATH
menu.
Predicting using the
PLOT
view
Using the
PLOT
view is the probably the more
visually appealing method of obtaining
predicted values. When you have plotted a set
of data and its fit curve then pressing up arrow
will change the focus of the values at the bottom of the screen from the data
points to the
PREDY
values. In the screen
snapshots shown right the focus changes from
data point 3 (2,2), to the
PREDY
value for x = 2
of 2.806.
If there is more than one data set (and fit lines) graphed then the up arrow
will move progressively from one to another and finally back to the first.
Pressing left and right arrows will move along the fit line but only on pixel
positions, which may not be suitable if the scale is not chosen carefully. A
better way is to use the
key to obtain
PREDY
values for any required
value of
x
, including values which would normally be off-screen.
Another aspect of bivariate stats needs to be remembered when the fit
chosen is not linear. Mathematically, the correlation coefficient is a strictly
linear measure of the goodness of fit and this means that the correlation
coefficient quoted in the
view is
always
for the linear model even when
the model chosen is not linear.