Transcript Lecture notes for 11/21/00

Sociology 601, Class17: October 27, 2009
• Linear relationships. A & F, chapter 9.1
• Least squares estimation. A & F 9.2
• The linear regression model (9.3)
1
Example of a linear Relationship
Disagree men are better politicians
70%
60%
50%
40%
1996
1993
1994
1988
1989
1990
1991
1985
1986
1982
1983
1977
1978
30%
1974
1975
Percent more egalitarian
80%
Year of GSS
2
Equation for a linear relationship
A linear relationship is a relationship between two variables Y
and X that can be defined by the equation:
Y =  + X
•Y is the value for the response variable
•X is the value for the explanatory variable
• is the Y-intercept
• is the slope
3
Example of a linear relationship
Change over time in attitudes about gender
Y = α +βX
•Y (the response variable) = % disagree that men make
better politicians than women
•X (the explanatory variable) = year of survey
•α(Y-intercept) = value of y when x=0
•β(the slope) = change in y per unit of x
4
Example of a linear relationship
Change over time in attitudes about gender
Yhat = a +bX = -25.62 + 0.013*year
year
observed
predicted
1974
49.2%
47.6%
1975
48.3%
49.0%
1977
47.2%
51.6%
1978
54.5%
52.9%
1982
59.2%
58.2%
1983
61.1%
59.5%
1985
58.6%
62.2%
1986
60.5%
63.5%
1988
65.3%
66.1%
1989
64.7%
67.5%
1990
69.4%
68.8%
1991
71.1%
70.1%
1993
74.4%
72.7%
1994
74.6%
74.1%
1996
74.9%
76.7%
5
The Dangers of Extrapolation
Change over time in attitudes about gender
Yhat = a +bX = -25.62 + 0.013*year
year
observed
predicted
1974
49.2%
47.6%
1996
74.9%
76.7%
2008
70.0%
92.6%
2020
?
108.4%
6
Example of a linear Relationship
Disagree women should take care of
home, not the country
80%
70%
y = -19.83 + 0.010 * x
60%
50%
1996
1993
1994
1988
1989
1990
1991
1985
1986
1982
1983
1977
1978
40%
1974
1975
Percent more egalitarian
90%
Year of GSS
7
Key terms for linear relationships
• Explanatory variable: a variable that we think of as
explaining or “causing” the value of another variable.
(also called the independent variable)
• We reserve X to denote the explanatory variable
• Response variable: a variable that we think of as being
explained or “caused” by the value of another variable.
(also called the dependent variable)
• We reserve Y to denote the response variable
(Q: what happens if both variables explain each other?)
8
More key terms for linear relationships
•  : the slope of a linear relationship
•  : the increment in y per one unit of x
o If  > 0, the relationship between the explanatory and
response variables is positive.
o If  < 0, the relationship between the explanatory and
response variables is negative.
o If  = 0, the explanatory and response variables are said
to be independent.
• if x is multiplied by 12 (e.g., months rather than years), then ’ = ?
• if x is divided by 10 (e.g., decades rather than years), then ’ = ?
• if y is multiplied by 100 (e.g., percentage points rather than
proportion), then ’ = ?
9
• if you subtract 1974 from x, then ’ = ?
More key terms for linear relationships
•
 : the y-intercept of a linear relationship
•
 is the value of y when x = 0.
o this is sometimes a meaningless value of x way beyond
its observed range.
•
 : determines the height of the line up or down on the
y-axis
•
•
•
•
if x is multiplied by 12 (e.g., months rather than years), then ’ = ?
if x is divided by 10 (e.g., decades rather than years), then ’ = ?
if y is multiplied by 100 (e.g., percentage points rather than proportion),
then ’ = ?
if you subtract 1974 from x, then ’ = ?
(note:  and  are both population parameters like )
10
More key terms for linear relationships
• model: a formula that provides a simple approximation for
the relationship between variables.
• The linear function is the simplest model for a
relationship between two interval scale variables.
• Regression analysis: using linear models to study…
o the form of a relationship between variables
o the strength of a relationship between variables
o whether a statistically significant relationship exists
between variables
11
Another Example of a linear Relationship
Disagree men are better politicians
70%
60%
50%
40%
1996
1993
1994
1988
1989
1990
1991
1985
1986
1982
1983
1977
1978
30%
1974
1975
Percent more egalitarian
80%
Year of GSS
12
9.2 Predicting Y-scores
using least squares regression
• Next, we study relationships between two variables where
• there are multiple cases of X, and
• Y scores do not always line up on a straight line.
• There is some scatter to the data points.
• The objective is still to predict a value of Y, given a value of
X.
13
Linear prediction: an example.
• Chaves, M. and D.E. Cann. 1992. “Regulation, Pluralism,
and Religious Market Structure.” Rationality and Society
4(3): 272-290.
• observations for 18 countries
• outcome var: weekly percent attending religious services
• variable name – “attend”
• explanatory var: level of state regulation of religion
• variable name – “regul”
• (not really interval scale), ordinal ranking 0-6
14
Plotting a linear relationship in STATA
. plot attend regul
a
t
t
e
n
d
82 +
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 +
*
*
*
*
*
*
*
*
*
*
*
*
*
*
+----------------------------------------------------------------+
0
regul
6
15
Solving a least squares regression, using STATA
. regress attend regul
Source |
SS
df
MS
-------------+-----------------------------Model | 2240.05128
1 2240.05128
Residual | 3715.94872
16 232.246795
-------------+-----------------------------Total |
5956
17 350.352941
Number of obs
F( 1,
16)
Prob > F
R-squared
Adj R-squared
Root MSE
=
=
=
=
=
=
18
9.65
0.0068
0.3761
0.3371
15.24
-----------------------------------------------------------------------------attend |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
-------------+---------------------------------------------------------------regul | -5.358974
1.72555
-3.11
0.007
-9.016977
-1.700972
_cons |
36.83761
5.395698
6.83
0.000
25.39924
48.27598
------------------------------------------------------------------------------
• b is the coefficient for “regul”.
• a is the coefficient for “_cons”.
(ignore all the other output for now)
%attend = 36.8 - 5.4 * regul
16
Finding the predicted values of religious attendance (yhat) for each
observed level of regulation (x)
%attend = 36.8 - 5.4 * regul
regul
calculation
yhat =
predicted
attendance
0
36.8 – 5.4*0
36.8%
1
36.8 – 5.4*1
31.5%
2
36.8 – 5.4*2
26.1%
3
36.8 – 5.4*3
20.8%
4
36.8 – 5.4*4
15.4%
5
36.8 – 5.4*5
10.0%
6
36.8 – 5.4*6
4.7%
17
Finding the predicted values
for each observed level of X, using STATA
. predict pattend
(option xb assumed; fitted values)
. tabulate pattend regul
Fitted |
regul
values |
0
1
2
3
5
6|
Total
-----------+-----------------------------------------------------+-------4.683761 |
0
0
0
0
0
2|
2
10.04274 |
0
0
0
0
2
0|
2
20.76068 |
0
0
0
5
0
0|
5
26.11966 |
0
0
2
0
0
0|
2
31.47863 |
0
1
0
0
0
0|
1
36.83761 |
6
0
0
0
0
0|
6
-----------+-----------------------------------------------------+-------Total |
6
1
2
5
2
|
18
you can use the predict command only after you have used
regress to estimate the regression function.
18
Plotting the predicted values
for each observed level of X - in STATA
. plot pattend regul
36.8376 +
| *
|
|
F
|
i
|
*
t
|
t
|
e
|
*
d
|
|
v
|
*
a
|
l
|
u
|
e
|
s
|
|
*
|
|
4.68376 +
*
+----------------------------------------------------------------+
0
regul
6
19
Interpreting a regression
• When the data are scattered, we need to ask two questions:
o Is the data suitable for a linear model?
o If so, how do we draw a line through it?
• Checking suitability (i.e,. assumptions)
• scattergrams
• crosstabs (including means and sd’s by x-levels)
• The assumptions of a linear regression are violated if
o the plot / crosstab suggests a nonlinear relationship
o there are severe outliers (extreme x or y scores)
o there is evidence of heteroskedasticity
(the amount of “scatter” of the dots depends on the 20xscore)
Possible prediction methods
Once you have decided that a linear model is appropriate, how
do you choose a linear equation with a scattered mess of
dots?
A.) Calculate a slope from any two points?
B.) Calculate the average slope of all the points (with the
least error)?
C.) Calculate the slope with the least squared error)?
All these solutions may be technically unbiased, but C. is
generally accepted as the most efficient. (C gives a slope that
is, on average, closest to the slope of the population.)
21
Least squares prediction: formal terms
• population equation for a linear model: Y =  +  X + ε
• equation for a given observation: Yi = a + bXi + ei
where Yi and Xi are observed values of Y and X,
and ei is the error in observation Yi .
• prediction for a given value of X, based on a sample:
• Yhat = a + bX, where Yhat is the predicted value of Y
• Note that Yi – Yhat = ei = residual for observation i
22
Least squares prediction: equation for b
• goal for a given sample: estimate b and a such that
 (Yi – Yhat )2 is as small as possible.
(To derive the solution: start with Q =  (Yi – a - bXi )2 , take
partial differentials of Q with respect to a and b, and solve
for relative minima. This will not be tested in class!)
• solution:
( X  X )(Y  Y )
b
,
2
( X  X )
a  Y  bX
23
Least squares prediction: more terms
•  (Yi – Yhat )2
is also called the sum of squared errors or SSE.
(Also called the residual sum of squares, the squared errors
in the response variable left over after you control for
variation due to the explanatory variable.)
• The method that calculates b and a to produce the smallest
possible SSE is the method of least squares.
• b and a are least squares estimates
• The prediction line Yhat = a + bX is the least squares line
24
Least squares prediction: still more terms
• For a given observation, the prediction error ei
(Yi – Yhat ) is called the residual.
• An atypical X or Y score or a large residual can be called an
outlier.
o outliers can bias an estimate of a slope
o outliers can increase the possibility of a type I error of
inference.
o outlier Y scores are especially troublesome when they
are associated with extreme values of X.
o outliers sometimes belong in the data, sometimes not.
o Q: DC homicide rates?
25
Calculating the residuals for each observation, using
STATA
. predict rattend, residuals
. summarize attend pattend rattend if country=="Ireland"
Variable |
Obs
Mean
Std. Dev.
Min
Max
-------------+-------------------------------------------------------attend |
1
82
.
82
82
pattend |
1
36.83761
.
36.83761
36.83761
rattend |
1
45.16239
.
45.16239
45.16239
• reminder: you can only use the predict command after you have
used regress to estimate the regression function.
26
Plotting the residuals for each observed level of X, using
STATA
. plot rattend regul
45.1624 +
|
|
|
|
|
R
|
e
|
s
|
i
|
d
|
u
|
a
|
l
|
s
|
|
|
|
|
|
-19.4786 +
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
+----------------------------------------------------------------+
0
regul
do you notice the residual that is an outlier?
27
More on Sums of Squares:
• Sum of Squares refers to the act of taking each ‘error’, squaring
it, and adding it to all the other errors in the sample.
This operation is analogous to calculating a variance, without dividing by n1.
• Sum of Squares Total (SST) refers to the difference between a
score yi and the overall mean Ybar.
 (Yi – Ybar )2
• Sum of Squares Error (SSE), also called Sum of Squares
Residual (SSR), refers to the difference between a score yi and
the corresponding prediction from the regression line Yhat.
 (Yi – Yhat )2
28
9.3 the linear regression model
The conceptual problem:
• The linear model Y =  +  X has limited use because it is
deterministic and cannot account for variability in Y-values
for observations with the same X-value.
The conceptual solution:
• The linear regression model E(Y) =  +  X is a probabilistic
model more suited to the variable data in social science
research.
• A regression function describes how the mean of the
response variable changes according to the value of an
explanatory variable.
• For example, we don’t expect qll college graduates to earn
more than all high school graduates, but we expect the
mean earnings of college graduates to be greater than29the
mean earnings of high school graduates.
A standard deviation
for the linear regression model
• A new problem:
How do we describe variation about the means of a
regression line?
• A solution:
The conditional standard deviation  refers to variability of
Y values about the conditional population mean
E(Y) =  +  X
for subjects with the same value of X.
ˆ 
Q: why n-2?
SSE

n2
ˆ )2
(Y  Y
n2
30
The linear regression model:
example of conditional standard deviation
• Church attendance and state control problem:
SSE (also called SSR) = 3715.9
n = 18,
n-2 = 16
ˆ 
SSE

n2
3715.9
 15.24
16
31
Solving a least squares regression, using STATA
. regress attend regul
Source |
SS
df
MS
-------------+-----------------------------Model | 2240.05128
1 2240.05128
Residual | 3715.94872
16 232.246795
-------------+-----------------------------Total |
5956
17 350.352941
Number of obs
F( 1,
16)
Prob > F
R-squared
Adj R-squared
Root MSE
=
=
=
=
=
=
18
9.65
0.0068
0.3761
0.3371
15.24
-----------------------------------------------------------------------------attend |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
-------------+---------------------------------------------------------------regul | -5.358974
1.72555
-3.11
0.007
-9.016977
-1.700972
_cons |
36.83761
5.395698
6.83
0.000
25.39924
48.27598
------------------------------------------------------------------------------
• b is the coefficient for “regul”.
• a is the coefficient for “_cons”.
(ignore all the other output for now)
%attend = 36.8 - 5.4 * regul
32
interpreting the conditional standard deviation
Church attendance and state control problem:
•
For every level of state control of religion, the standard
deviation for the predicted mean church attendance is 15.24
percentage points. (Draw chart on board)
• By assumptions of the regression model, this is true for
every level of state control.
• (Is that assumption valid in this case?)
33
Conditional standard deviation and
Marginal standard deviation
• Degrees of freedom are different
• E(Y) is different: Ybar versus Yhat
• Conditional s.d. is usually smaller than marginal.
sY 
ˆ 
(Y  Y ) 2
n 1
2
ˆ
(Y  Y )
n2
34