1 lecture ten. 2 lecture part i: regression part ii: experimental method
TRANSCRIPT
1
Lecture Ten
2
Lecture
• Part I: Regression
• Part II: Experimental Method
3
Outline: Regression
• The Assumptions of Least Squares
• The Pathologies of Least Squares
• Diagnostics for Least Squares
Assumptions
• Expected value of the error is zero, E[e)t)]= 0
• The error is independent of the explanatory variable, E{e(t) [x(t)-Ex(t)]}=0
• The errors are independent of one another, E[e(i)e(j)] = 0 , i not equal to j.
• The variance is homoskedatic, E[e(i)]2=E[e(j)]2
• The error is normal with mean zero and variance
2
5
18.4 Error Variable: Required Conditions
• The error is a critical part of the regression model.• Four requirements involving the distribution of
must be satisfied.– The probability distribution of is normal.– The mean of is zero: E() = 0.
– The standard deviation of is for all values of x.
– The set of errors associated with different values of y are all independent.
The Normality of
From the first three assumptions we have:y is normally distributed with meanE(y) = 0 + 1x, and a constant standard deviation
From the first three assumptions we have:y is normally distributed with meanE(y) = 0 + 1x, and a constant standard deviation
0 + 1x1
0 + 1x2
0 + 1x3
E(y|x2)
E(y|x3)
x1 x2 x3
E(y|x1)
The standard deviation remains constant,
but the mean value changes with x
7
Pathologies
• Cross section data: error variance is heteroskedatic. Example, could vary with firm size. Consequence, all the information available is not used efficiently, and better estimates of the standard error of regression parameters is possible.
• Time series data: errors are serially correlated, i.e auto-correlated. Consequence, inefficiency.
8
Pathologies ( Cont. )• Explanatory variable is not independent of
the error. Consequence, inconsistency, i.e. larger sample sizes do not lead to lower standard errors for the parameters, and the parameter estimates (slope etc.) are biased.
• The error is not distributed normally. Example, there may be fat tails. Consequence, use of the normal may underestimate true 95 % confidence intervals.
9
Pathologies (Cont.)
• Multicollinearity: The independent variables may be highly correlated. As a consequence, they do not truly represent separate causal factors, but instead a common causal factor.
10
18.9 Regression Diagnostics - I
• The three conditions required for the validity of the regression analysis are:– the error variable is normally distributed.– the error variance is constant for all values of x.– The errors are independent of each other.
• How can we diagnose violations of these conditions?
11
Residual Analysis• Examining the residuals (or standardized
residuals), help detect violations of the required conditions.
• Example 18.2 – continued:– Nonnormality.
• Use Excel to obtain the standardized residual histogram.
• Examine the histogram and look for a bell shaped. diagram with a mean close to zero.
12
Diagnostics ( Cont. )
• Multicollinearity may be suspected if the t-statistics for the coefficients of the explanatory variables are not significant but the coefficient of determination is high. The correlation between the explanatory variable can then be calculated. To see if it is high.
13
Diagnostics
• Is the error normal? Using EViews, with the view menu in the regression window, a histogram of the distribution of the estimated error is available, along with the coefficients of skewness and kurtosis, and the Jarque-Bera statistic testing for normality.
14
Diagnostics (Cont.)
• To detect heteroskedasticity: if there are sufficient observations, plot the estimated errors against the fitted dependent variable
Heteroscedasticity• When the requirement of a constant variance is violated
we have a condition of heteroscedasticity.• Diagnose heteroscedasticity by plotting the residual
against the predicted y.
+ + ++
+ ++
++
+
+
+
+
+
+
+
+
+
+
++
+
+
+
The spread increases with y
y
Residualy
+
+++
+
++
+
++
+
+++
+
+
+
+
+
++
+
+
16
Homoscedasticity• When the requirement of a constant variance is not violated we
have a condition of homoscedasticity.• Example 18.2 - continued
-1000
-500
0
500
1000
13500 14000 14500 15000 15500 16000
Predicted Price
Re
sid
ua
ls
17
Diagnostics ( Cont.)
• Autocorrelation: The Durbin-Watson statistic is a scalar index of autocorrelation, with values near 2 indicating no autocorrelation and values near zero indicating autocorrelation. Examine the plot of the residuals in the view menu of the regression window in EViews.
18
Non Independence of Error Variables
– A time series is constituted if data were collected over time.
– Examining the residuals over time, no pattern should be observed if the errors are independent.
– When a pattern is detected, the errors are said to be autocorrelated.
– Autocorrelation can be detected by graphing the residuals against time.
19
Patterns in the appearance of the residuals over time indicates that autocorrelation exists.
+
+++ +
++
++
+ +
++ + +
+
++ +
+
+
+
+
+
+Time
Residual Residual
Time+
+
+
Note the runs of positive residuals,replaced by runs of negative residuals
Note the oscillating behavior of the residuals around zero.
0 0
Non Independence of Error Variables
20
Fix-Ups
• Error is not distributed normally. For example, regression of personal income on explanatory variables. Sometimes a transformation, such as regressing the natural logarithm of income on the explanatory variables may make the error closer to normal.
21
Fix-ups (Cont.)• If the explanatory variable is not
independent of the error, look for a substitute that is highly correlated with the dependent variable but is independent of the error. Such a variable is called an instrument.
22
Data Errors: May lead to outliers
• Typos may lead to outliers and looking for ouliers is a good way to check for serious typos
23
Outliers• An outlier is an observation that is unusually small or
large.• Several possibilities need to be investigated when an
outlier is observed:– There was an error in recording the value.
– The point does not belong in the sample.
– The observation is valid.
• Identify outliers from the scatter diagram.
• It is customary to suspect an observation is an outlier if its |standard residual| > 2
24
+
+
+
+
+ +
+ + ++
+
+
+
+
+
+
+
The outlier causes a shift in the regression line
… but, some outliers may be very influential
++++++++++
An outlier An influential observation
25
Procedure for Regression Diagnostics• Develop a model that has a theoretical basis.• Gather data for the two variables in the model.• Draw the scatter diagram to determine whether a linear
model appears to be appropriate.• Determine the regression equation.• Check the required conditions for the errors.• Check the existence of outliers and influential observations• Assess the model fit.
• If the model fits the data, use the regression equation.
26
Part II: Experimental Method
27
Outline
• Critique of Regression
28
Critique of Regression
• Samples of opportunity rather than random sample
• Uncontrolled Causal Variables– omitted variables– unmeasured variables
• Insufficient theory to properly specify regression equation
29
Experimental Method: # Examples
• Deterrence
• Aspirin
• Miles per Gallon
30
Deterrence and the Death Penalty
31
Isaac Ehrlich Study of the Death Penalty: 1933-1969
Isaac Ehrlich Study of the Death Penalty: 1933-1969
Homicide Rate Per CapitaControl Variables
probability of arrestprobability of conviction given charged Probability of execution given conviction
Causal Variableslabor force participation rateunemployment ratepercent population aged 14-24 yearspermanent incometrend
32
Long Swings in the Homicide Rate in the US: 1900-1980
Source: Report to the Nation on Crime and Justice
33
Ehrlich Results: Elasticities of Homicide with respect to Controls
Ehrlich Results: Elasticities of Homicide with respect to Controls
Control Elasticity Average Valueof Control
Prob. of Arrest -1.6 0.90
Prob. of ConvictionGiven Charged
-0.5 0.43
Prob. of ExecutionGiven Convicted
-0.04 0.026
Source: Isaac Ehrlich, “The Deterrent Effect of Capital Punishment
34
Critique of Ehrlich by Death Penalty Opponents
Critique of Ehrlich by Death Penalty Opponents
Time period used: 1933-1968period of declining probability of execution
Ehrlich did not include probability of imprisonment given conviction as a control variable
Causal variables included are unconvincing as causes of homicide
35
United States Bureau of Justice Statisticshttp://www.ojp.usdoj.gov/bjs/
36
Experimental Method
• Police intervention in family violence
37
http://www.ojp.usdoj.gov/bjs/
United States Bureau of Justice Statistics
38
United States Bureau of Justice Statisticshttp://www.ojp.usdoj.gov/bjs/
39
Police Intervention with Experimental Controls
Police Intervention with Experimental Controls
A 911 call from a family memberthe case is randomly assigned for “treatment”
A police patrol responds and visits the householdpolice calm down the family membersbased on the treatment randomly assigned, the
police carry out the sanctions
40
Why is Treatment Assigned Randomly?
Why is Treatment Assigned Randomly?
To control for unknown causal factorsassign known numbers of cases, for example
equal numbers, to each treatmentwith this procedure, there should be an even
distribution of difficult cases in each treatment group
41
911 call(characteristics of household Participants unknown)
Random Assignment
code blue code gold
patrol responds patrol responds
settles the household settles the household
verbally warn the husband take the husband to jail for the night
42
Experimental Method: Clinical Trials• Doctors Volunteer
• Randomly assigned to two groups
• treatment group takes an aspirin a day
• the control group takes a placebo (sugar pill) per day
• After 5 years, 11,037 experimentals have 139 heart attacks (fatal and non fatal) pE = 0.0126
• after 5 years, 11034 controls have 239 heart attacks, pc= 0.0217
Conclusions from the Clinical Trials• Hypotheses: H0 : pC = pE , or pC - pE = 0.; Ha : (pC -
pE ) 0.
• Statistic:Z = [ C - E ) – (pC - pE )]/( pC - pE )
• Var C - E ) = Var( C ) + Var ( E )
• recall, from the variance for a proportionSE SE( C -
E )={[ c (1- c )]/nc + [ E(1- E )]/nE }1/2
• { [0.={[0217 ( 1- 0.0217)/ 11,034] + [0.0126 ( 1 – 0.0126)/ 11,039}1/2
• = 0.00175, so z = (.2017-.0126)/.00175• z= 5.2•
pp
pp p p
p p p p p p
44
Experimental Method
• Experimental Design: Paired Comparisons
Cab Brand A Brand B Difference
1 27.01 26.95 0.06
2 20.00 20.44 -0.44
3 23.41 25.05 -1.64
4 25.22 26.32 -1.10
5 30.11 29.56 0.55
6 25.55 26.60 -1.05
7 22.23 22.93 -0.70
8 19.78 20.23 -0.45
9 33.45 33.95 -0.50
10 25.22 26.01 -0.79
Sample Mean 25.20 25.80 -0.60
Standard Deviation 4.27 4.10 0.61
Table 1: Miles Per Gallon for Brand A and Brand B