Transcript B632_06lect13

Logistic Regression Analysis
Estimation and Interpretation
• Logit Estimation
– Log likelihood estimation
• Hypothesis Tests
• Interpretation
• Reversing Logits
– Averages
– Types
April 18, 2006
Lecture 13
Slide #1
Logit Estimation
• Logit uses maximum likelihood estimation
– Counterpart to minimizing least squares
• MLE identifies the probability of obtaining the
sample as a function of model parameters
(X’s):
– What values for b’s make the sample most likely?
April 18, 2006
Lecture 13
Slide #2
Logit Estimation, Continued
The probability that Y i = 1 is :
1
K 1
Pi =
where Li  0  k 1 k X ik
- Li
(1 + e )
Each observation i (Yi plus the Xi’s) contributes to the MLF by
Pi if Yi=1, and by 1-Pi if Yi=0. The contribution is:
PiYi (1 Pi )1Yi and the ML function is
L  {Pi (1  Pi )
Yi
1 Yi
}
In other words, the MLF is largest for the model that best
predicts when Y=1 or Y=0; when the predicted value of Y is
correct and close to 1 or 0, the MLF is maximized.
April 18, 2006
Lecture 13
Slide #3
Logit Estimation, Continued
The MLF identifies values of the b’s that maximize the
log likelihood:
log e L  {Yi log e Pi  (1  Yi ) log e (1 Pi )}
The solution involves taking the first derivative of the log
likelihood with respect to each of the b’s, setting them to
zero, and solving the simultaneous equation. The solution
of the equation isn’t linear, so it can’t be solved directly.
Instead, it’s solved through a sequential estimation process
that looks for successively better “fits” of the model.
April 18, 2006
Lecture 13
Slide #4
Example of Logit Estimation:
Radiation Protection Standards
3. Supra-linear relationship
High
Cancer
Incidence
1. Linear relationship
2. Sub-linear relationship
0
0
April 18, 2006
Radiation
Dose
Lecture 13
High
Slide #5
Measurement
• “Given your own knowledge of radiation effects on
humans and other organisms, which of the above
hypothesized relationships do you think is most likely
correct?”
• “On a scale where zero means not at all certain, and ten
means completely certain, how certain are you that the
relationship you identified is correct?”
• “Which of these three hypothesized relationships do you
think should be assumed for purposes of setting public
safety standards for managing radioactive materials?”
April 18, 2006
Lecture 13
Slide #6
Example of Logit Estimation,
Continued
• We need a binary dependent variable
– Focus only on those (the majority) who believe the threshold
model (quad-linear) is correct
– Predict the choice between the threshold (Quad=0) and the
linear (=1) model as the basis for their preferred safety standard
–
–
–
generate DR_correct = c4_28_ra
recode DR_correct (2=0) (1=1) (3=.) (4=.)
tabulate DR_correct
– Now recode the D-R function that is preferred for standard setting
–
–
–
generate DR_standard = c4_30_pb
recode DR_standard (2=0) (1=1) (3=.) (4=.)
tabulate DR_standard
April 18, 2006
Lecture 13
Slide #7
Example of Logit Estimation,
Continued
. logit DR_s tandard DR_cert ide ology sex if DR_correct==0
Iteration
Iteration
Iteration
Iteration
0:
1:
2:
3:
log
log
log
log
likelihood
likelihood
likelihood
likelihood
=
=
=
=
-607.99677
-584.66068
-584.57097
-584.57097
Logit estimates
Number of obs
LR chi2(3)
Prob > chi2
Pseudo R2
Log likelihood = -584.57097
=
=
=
=
943
46.85
0.0000
0.0385
-----------------------------------------------------------------------------DR_standard |
Coef.
Std. Err.
z
P>|z|
[95% Conf. Interval]
-------------+---------------------------------------------------------------DR_cert | -.1700262
.0299573
-5.68
0.000
-.2287414
-.111311
ideology | -.1041382
.0512416
-2.03
0.042
-.20457
-.0037065
sex | -.5071621
.1980048
-2.56
0.010
-.8952443
-.1190799
_cons |
1.229425
.3058896
4.02
0.000
.6298925
1.828958
------------------------------------------------------------------------------
April 18, 2006
Lecture 13
Slide #8
Example of Logit Estimation,
Continued
• Model “goodness of fit”:
. lstat
Logistic model for DR_standard
-------- True --- ----Classified |
D
~D |
Total
-----------+--------------------------+----------+
|
59
53 |
112
|
267
564 |
831
-----------+--------------------------+----------Total
|
326
617 |
943
Classified + if predicted Pr(D) >= .5
True D defined as DR_standard != 0
-------------------------------------------------Sensitivity
Pr( +| D)
18.10%
Specificity
Pr( -|~D)
91.41%
Positive predictive value
Pr( D| +)
52.68%
Negative predictive value
Pr(~D| -)
67.87%
-------------------------------------------------False + rate for true ~D
Pr( +|~D)
8.59%
False - rate for true D
Pr( -| D)
81.90%
False + rate for classified +
Pr(~D| +)
47.32%
False - rate for classified Pr( D| -)
32.13%
-------------------------------------------------Correctly classified
66.07%
--------------------------------------------------
April 18, 2006
Lecture 13
Slide #9
Logit Assumptions and Qualifiers
• The model is correctly specified
– True conditional probabilities are logistic function of
the X’s
– No important X’s omitted; no extraneous X’s included
– No significant measurement error
• The cases are independent
• No X is a linear function of other X’s
– Multicollinearity leads to imprecision
• Influential cases can bias estimates
• Sample size: n-K should exceed 100
– Independent covariation is critical
April 18, 2006
Lecture 13
Slide #10
Logit Hypothesis Tests
• Nested Model Tests (like F-Tests in OLS)
– Is a more complex model a better fit?
• Test to see if parameters for omitted variables are
statistically indistinguishable from zero:
 H2  2(log e LK  H  log e LK )
• Where the Chi-square table uses K degrees of
freedom.
• If p < 0.05, the complex model fits significantly
better
April 18, 2006
Lecture 13
Slide #11
More Logit Hypothesis Tests
• To test for the overall hypothesis that all b’s
are equal to zero (like an overall F-test):
– Compare the final log-likelihood with the initial
one, using the same formula:
Initial log likelihood = -607.997
Final log likelihood = -584.571
Difference =
-23.426

2
K1
 2(log e Li  log e L f )
= 46.85, df=K-1; p-value > 0.001
(see Hamilton p. 354)

April 18, 2006
Lecture 13
Slide #12
Still More Logit Hypothesis Tests
• z-statistic:
– Similar to the t-stat in OLS
– Compares the estimated coefficient to the
estimated standard error
– P-value is derived from the Chi-Square
distribution
• Attached to each estimated coefficient
– The p-value shows probability that the null
hypothesis is correct, given the data
April 18, 2006
Lecture 13
Slide #13
Interpreting Logits
• Logits can be used to directly calculate
odds:
anti log  e
Lˆ
• Logits can be reversed to obtain the
predicted probabilities:
Pˆ 
April 18, 2006
1
1 e
 Lˆ
Lecture 13
Slide #14
Interpreting Logits, Continued
How would you calculate the effect of a particular
independent variable, Xi, on the probability of Y = 1?
•
Set all Xj’s at their mean, then calculate
Pˆ 
•
1
1 e
 Lˆ
With Xi at it’s minimum and maximum.
Then calculate the difference.
April 18, 2006
Lecture 13
Slide #15
Estimated Probability Effects
.6
.5
.4
M edi
an s p
line
.3
.2
0
2
4
6
8
10
DR_ce rt
April 18, 2006
Lecture 13
Slide #16
Interpreting Logits, Continued
•Another method: “Typing”
•Calculate the logit for distinct types of
observations:
•Conservative, certain, male
•Liberal, uncertain female
•(or any permutation you like)
•Homework:
•Predict the choice between Quad and
LD-HR (will require recodes)
•Plot the effect of the risk index and
ideology on probability of a shift
April 18, 2006
Lecture 13
Slide #17
Coming Up...
• Chapter 7
– All
• Statistical Problems with Logit
– Effects of assumption failures
– Diagnostics
• Begin discussion of Factor Analysis
April 18, 2006
Lecture 13
Slide #18