Transcript Chapter 8 Part A Power Point

Chapter 8: Linear Regression—
Part A
A.P. Statistics
Linear Model
• Making a scatterplot allows you to describe
the relationship between the two quantitative
variables.
• However, sometimes it is much more useful to
use that linear relationship to predict or
estimate information based on that real data
relationship.
• We use the Linear Model to make those
predictions and estimations.
Linear Model
Normal Model
Allows us to make predictions
and estimations about the
population and future
events.
Linear Model
Allow us to make predictions
and estimations about the
population and future
events.
It is a model of real data, as
long as that data has a
nearly symmetric
distribution.
It is a model of real data, as
long as that data has a
linear relationship between
two quantitative variables.
Linear Model and the Least Squared
Regression Line
• To make this model, we need to find a line of
best fit.
• This line of best fit is the “predictor line” and
will be the way we predict or estimate our
response variable, given our explanatory
variable.
• This line has to do with how well it minimizes
the residuals.
Residuals and the Least Squares
Regression Line
• The residual is the
difference between the
observed value and the
predicted value.
• It tells us how far off the
model’s prediction is at
that point
• Negative residual:
predicted value is too big
(overestimation)
• Positive residual:
predicted value is too
small (underestimation)
Residuals
Least Squares Regression Line
• The LSRL attempts to find a line where the
sum of the squared residuals are the smallest.
• Why not just find a line where the sum of the
residuals is the smallest?
– Sum of residuals will always be zero
– By squaring residuals, we get all positive values,
which can be added
– Emphasizes the large residuals—which have a big
impact on the correlation and the regression line
Scatterplot of Math and Verbal SAT
scores
Scatterplot of Math and Verbal SAT
scores with incorrect LSRL
Scatterplot of Math and Verbal SAT
scores with correct LSRL
Simple Regression
Model of Collection 1
Response attribute (numeric): Verbal_SAT
Predictor attribute (numeric): Math_SAT
Sample count: 6
Equation of least-squares regression line:
Verbal_SAT = 1.11024 Math_SAT 75.424
Correlation coefficient, r = 0.954082
r-squared = 0.91027, indicating that
91.027% of the variation in Verbal_SAT is
accounted for by Math_SAT.
The best estimate for the slope is 1.11024
+/- 0.4839 at a 95 % confidence level.
(The standard error of the slope is
0.174288.)
When Math_SAT = 0 , the predicted value
for a future observation of Verbal_SAT is 75.4244 +/- 288.073.
Correlation and the Line
(Standardized data)
• LSRL passes through z x
and z y
• LSRL equation is: zˆ y  rz x
“moving one standard
deviation from the mean in x,
we can expect to move about
r standard deviations from the
mean in y .”
Interpreting Standardized Slope of
LSRL
LSRL of scatterplot:
zˆ fat  0 . 83 z protein
For every standard deviation
above (below) the mean a
sandwich is in protein, we’ll
predict that that its fat
content is 0.83 standard
deviations above (below) the
mean.
LSRL that models data in real units
yˆ  b0  b1 x
b 0  y - intercept
b1  slope
b 0  y  b1 x
Protein
Fat
x  17 . 2 g
y  23 . 5 g
s x  14 . 0 g
s y  16 . 4 g
r  0 . 83
b1 
rs y
sx
LSRL
Equation:
Interpreting LSRL
Slope: One additional gram of
protein is associated with an
additional 0.97 grams of fat.
y-intercept: An item that has
zero grams of protein will have
6.8 grams of fat.
f aˆ t  6 . 8  0 . 97 protein
ALWAYS CHECK TO SEE IF yINTERCEPT MAKES SENSE IN THE
CONTEXT OF THE PROBLEM AND
DATA
Properties of the LSRL
The fact that the Sum of Squared Errors (SSE,
same as Least Squared Sum)is as small as
possible means that for this line:
• The sum and mean of the residuals is 0
• The variation in the residuals is as small as
possible
• The line contains the point of averages x , y 
Assumptions and Conditions for using
LSRL
Quantitative Variable
Condition
Straight Enough Condition
if not—re-express
Outlier Condition
with and without ?
Residuals and LSRL
• Residuals should be
used to see if a linear
model is appropriate
and in addition the LSRL
that was calculated
• Residuals are the part
of the data that has not
been modeled in our
linear model
Residuals and LSRL
What to Look for in a
Residual Plot to Satisfy
Straight Enough Condition:
Looking at a scatterplot of the residuals
vs. the x-value is a good way to check
the Straight Enough Condition, which
determines if a linear model is
appropriate.
No patterns, no interesting
features (like direction or
shape), should stretch
horizontally with about same
scatter throughout, no bends or
outliers.
The distribution of residuals
should be symmetric if the
original data is straight enough.
Residuals, again
When analyzing the relationship
between two variables (thus far)
ALWAYS:
• Plot the data and describe the relationship*
Quantitati ve Data
• Check Three Regression
Straight Enough
Assumptions/Conditions Outlier
• Compute correlation coefficient
• Compute Least Squared Regression Line
• Check Residual Plot (Again)
• Interpret relationship (intercept, slope, correlation
and general conclusion)
* Calculate mean and standard deviation for each variable, if possible
S1 = mean
S2 = s
Final
Collection 1
84.933333
6.540715
Collection 1
Midterm
83.466667
7.4437574
S1 = mean
S2 = s
Model of Collection 1Simple Regression
Response attribute (numeric): Final
Predictor attribute (numeric): Midterm
Sample count: 15
Equation of least-squares
regression line:
Final = 0.752149 Midterm +
22.154
Correlation coefficient, r =
0.855994
r-squared = 0.73273, indicating
that 73.273% of the variation in
Final is accounted for by
Midterm.
The best estimate for the slope
is 0.752149 +/- 0.272187 at a 95
% confidence level. (The
standard error of the slope is
0.125991.)
When Midterm = 0 , the
predicted value for a future
observation of Final is 22.154
+/- 24.0299.