agresti/franklin statistics, 1 of 88 section 11.4 what do we learn from how the data vary around...
TRANSCRIPT
Agresti/Franklin Statistics, 1 of 88
Section 11.4
What Do We Learn from How the Data Vary Around the
Regression Line?
Agresti/Franklin Statistics, 2 of 88
Residuals and Standardized Residuals
A residual is a prediction error – the difference between an observed outcome and its predicted value• The magnitude of these residuals depends
on the units of measurement for y
A standardized version of the residual does not depend on the units
Agresti/Franklin Statistics, 3 of 88
Standardized Residuals
Standardized residual:
The se formula is complex, so we rely on software to find it
A standardized residual indicates how many standard errors a residual falls from 0
Often, observations with standardized residuals larger than 3 in absolute value represent outliers
)ˆ(
)ˆ(
yyse
yy
Typo on Pg 553 of Text.
Corrected Version
Agresti/Franklin Statistics, 4 of 88
Example: Detecting an Underachieving College Student
Data was collected on a sample of 59 students at the University of Georgia
Two of the variables were:• CGPA: College Grade Point Average
• HSGPA: High School Grade Point Average
Example 13 in Text
Agresti/Franklin Statistics, 5 of 88
Example: Detecting an Underachieving College Student
A regression equation was created from the data:
• x: HSGPA
• y: CGPA
Equation: xy 64.019.1ˆ
Agresti/Franklin Statistics, 6 of 88
Example: Detecting an Underachieving College Student
MINITAB highlights observations that have standardized residuals with absolute value larger than 2:
Agresti/Franklin Statistics, 7 of 88
Example: Detecting an Underachieving College Student
Consider the reported standardized residual of -3.14
• This indicates that the residual is 3.14 standard errors below 0
• This student’s actual college GPA is quite far below what the regression line predicts
Agresti/Franklin Statistics, 8 of 88
Analyzing Large Standardized Residuals
Does it fall well away from the linear trend that the other points follow?
Does it have too much influence on the results?
Note: Some large standardized residuals may occur just because of ordinary random variability
Agresti/Franklin Statistics, 9 of 88
Histogram of Residuals
A histogram of residuals or standardized residuals is a good way of detecting unusual observations
A histogram is also a good way of checking the assumption that the conditional distribution of y at each x value is normal• Look for a bell-shaped histogram
Agresti/Franklin Statistics, 10 of 88
Histogram of Residuals
Suppose the histogram is not bell-shaped: • The distribution of the residuals is not
normal
However….
• Two-sided inferences about the slope parameter still work quite well
• The t- inferences are robust
Agresti/Franklin Statistics, 11 of 88
The Residual Standard Deviation
For statistical inference, the regression model assumes that the conditional distribution of y at a fixed value of x is normal, with the same standard deviation at each x
This standard deviation, denoted by σ, refers to the variability of y values for all subjects with the same x value
Agresti/Franklin Statistics, 12 of 88
The Residual Standard Deviation
The estimate of σ, obtained from the data, is:
2
)ˆ( 2
n
yys
Agresti/Franklin Statistics, 13 of 88
Example: How Variable are the Athletes’ Strengths?
From MINITAB output, we obtain s, the residual standard deviation of y:
For any given x value, we estimate the mean y value using the regression equation and we estimate the standard deviation using s: s = 8.0
0.855
8.3522 s
Agresti/Franklin Statistics, 14 of 88
Confidence Interval for µy
We estimate µy, the population mean of y
at a given value of x by:
We can construct a 95 %confidence interval for µy using:
bxay ˆ
)ˆ (ˆ025.
yofsety
Agresti/Franklin Statistics, 15 of 88
Prediction Interval for y
The estimate for the mean of y at a fixed value of x is also a prediction for an individual outcome y at the fixed value of x
Most regression software will form this interval within which an outcome y is likely to fall• This is called a prediction interval for y
bxay ˆ
(See Figure 11.10)
Agresti/Franklin Statistics, 16 of 88
The Residual Standard Deviation
Difference in limit of CI and “s”
2
)ˆ( 2
n
yys
Agresti/Franklin Statistics, 17 of 88
Prediction Interval for y vs Confidence Interval for µy
The prediction interval for y is an inference about where individual observations fall
• Use a prediction interval for y if you want to predict where a single observation on y will fall for a particular x value
Agresti/Franklin Statistics, 18 of 88
Prediction Interval for y vs Confidence Interval for µy
The confidence interval for µy is an
inference about where a population mean falls
• Use a confidence interval for µy if you want
to estimate the mean of y for all individuals having a particular x value
Agresti/Franklin Statistics, 19 of 88
Example: Predicting Maximum Bench Press and Estimating its Mean
Agresti/Franklin Statistics, 20 of 88
Example: Predicting Maximum Bench Press and Estimating its Mean
Use the MINITAB output to find and interpret a 95% CI for the population mean of the maximum bench press values for all female high school athletes who can do x = 11 sixty-pound bench presses
For all female high school athletes who can do 11 sixty-pound bench presses, we estimate the mean of their maximum bench press values falls between 78 and 82 pounds
Agresti/Franklin Statistics, 21 of 88
Example: Predicting Maximum Bench Press and Estimating its Mean
Use the MINITAB output to find and interpret a 95% Prediction Interval for a single new observation on the maximum bench press for a randomly chosen female high school athlete who can do x = 11 sixty-pound bench presses
For all female high school athletes who can do 11 sixty-pound bench presses, we predict that 95% of them have maximum bench press values between 64 and 96 pounds
Agresti/Franklin Statistics, 22 of 88
Decomposing the Error
OR
Regression SS + Residual SS= Total SS
Regression SS:=P
(yi ¡ ¹y)2 =P
(yi ¡ ¹y)2 ¡P
(yi ¡ yi )2
F=(MS Reg)/(MSE). More general the “t” test (in cases studied in this class it is effectively “t” squared)
However in more complicated models (more explanatory variables) the difference and utility of this becomes apparent
In software(e.g. ANOVA), a sumof squaresdivided by df iscalled theMeanSquare For example MSE stands for mean square error :=
P(yi ¡ yi )2)=df
Agresti/Franklin Statistics, 23 of 88
Section 11.5
Exponential Regression: A Model for Nonlinearity
Agresti/Franklin Statistics, 24 of 88
Nonlinear Regression Models
If a scatterplot indicates substantial curvature in a relationship, then equations that provide curvature are needed
• Occasionally a scatterplot has a parabolic appearance: as x increases, y increases then it goes back down
• More often, y tends to continually increase or continually decrease but the trend shows curvature
Agresti/Franklin Statistics, 25 of 88
Example: Exponential Growth in Population Size
Since 2000, the population of the U.S. has been growing at a rate of 2% a year
• The population size in 2000 was 280 million
• The population size in 2001 was 280 x 1.02
• The population size in 2002 was 280 x (1.02)2
• …
• The population size in 2010 is estimated to be
• 280 x (1.02)10
• This is called exponential growth
Agresti/Franklin Statistics, 26 of 88
Exponential Regression Model
An exponential regression model has the formula:
For the mean µy of y at a given value of x, where α and β are parameters
x
y
Agresti/Franklin Statistics, 27 of 88
Exponential Regression Model
In the exponential regression equation, the explanatory variable x appears as the exponent of a parameter
The mean µy and the parameter β can take only positive values
As x increases, the mean µy increases when β>1
It continually decreases when 0 < β<1
Agresti/Franklin Statistics, 28 of 88
Exponential Regression Model
For exponential regression, the logarithm of the mean is a linear function of x
When the exponential regression model holds, a plot of the log of the y values versus x should show an approximate straight-line relation with x
Agresti/Franklin Statistics, 29 of 88
Example: Explosion in Number of People Using the Internet
Agresti/Franklin Statistics, 30 of 88
Example: Explosion in Number of People Using the Internet
Agresti/Franklin Statistics, 31 of 88
Example: Explosion in Number of People Using the Internet
Agresti/Franklin Statistics, 32 of 88
Example: Explosion in Number of People Using the Internet
Using regression software, we can create the exponential regression equation:
x: the number of years since 1995. Start with x = 0 for 1995, then x=1 for 1996, etc
y: number of internet users
Equation: xy )7708.1(38.20ˆ
Agresti/Franklin Statistics, 33 of 88
Interpreting Exponential Regression Models
In the exponential regression model,
the parameter α represents the mean value of y when x = 0;
The parameter β represents the multiplicative effect on the mean of y for a one-unit increase in x
x
y
Agresti/Franklin Statistics, 34 of 88
Example: Explosion in Number of People Using the Internet
In this model:
The predicted number of Internet users in 1995 (for which x = 0) is 20.38 million
The predicted number of Internet users in 1996 is 20.38 times 1.7708
xy )7708.1(38.20ˆ