\
|
> = >
o
o o
c
c
i i
i i
B
B Pr ) Pr(
=
5)
|
.
|
\
|
+ u =
(
|
.
|
\
|
u
o
o o
o
i i
B B
1
) ( ) | Pr(
i i
B B yes + u = | o
| o
/ =
27
while the standard deviation of WTP is estimated as:
8)
The same formulae produce estimates of mean/median WTP and of
the scale parameter of WTP from the logit coefficient if WTP is
assumed to be a logistic variate.
A standard probit routine will automatically produce standard errors
for and , but not for and .
To obtain the covariance matrix of and , you can use the delta
method (Cameron, 1991).
First calculate the covariance matrix of and produced by the
probit routine V:
9)
where , and , with |(-) the
standard normal probability density function (pdf).
Next, compute the matrix G :
10) G =
| o
/ 1 =
o
|
o
o
|
V =
`
)
=
w z
B
B B
i
i
i i i
n
( )
1
2
1
1
z B
i i
= +
o | w z z z z
i i i i
( ) ( ) { ( )[ ( )]} = |
2
1 u u
(
(
1 0
1
2 2
/
/
/
|
o | |
28
Finally calculate the matrix product V
1
=GVG, with V
1
the covariance
matrix of and . Nb This can be done in LIMDEP.
If WTP is assumed to be a logistic variate, the steps required for the
delta method are the same, except that w(z) in expression (9) is
equal to {exp(z
i
) / [1 + exp(z
i
)]}.
in some studies, depending on the frequencies of the yes and no
responses to the payment questions, formula (7) produces a
negative mean/median WTP figure.
Perhaps a better way to avoid this problem is to work with a WTP
distribution that is defined only over the positive semi-axis. The
Weibull and the lognormal are examples of such distributions.
The cdf for a Weibull with parameters u (u>0) and o is
Mean WTP is where I(-) is the gamma function
Median WTP is
The log likelihood function becomes:
11)
where F is the cdf of the Weibull
o
F y e
y
( )
( / )
=
1
o
u
|
.
|
\
|
+ I 1
1
u
o
u
o
/ 1
)] 5 . 0 ln( [
| |
=
+
n
i
i i i i
B F WTP B F WTP
1
) , ; ( log ) 1 ( )) , ; ( 1 log( o u o u
29
median WTP is generally regarded as a robust, and conservative,
welfare statistic associated with the good or proposed policy. It is
usually estimated more precisely than mean WTP, and is interpreted
as the value at which 50% of the respondents would vote in support
of the program, and hence the cost at which the majority of the
population would be in support of it.
30
The Double Bounded Dichotomous Choice model
Double bounded models increase efficiency in three ways:
YN and NY answers bound WTP
NN and YY answers further constrain WTP
The number of observation is increased
The log likelihood function becomes:
where WTP
H
and WTP
L
are the lower and upper bound of the interval
around WTP defined above, F() is the cdf of WTP, and denotes
the vector of parameters that index the distribution of WTP. (Notice
that for respondents who give two yes responses, the upper bound
of WTP may be infinity, or the respondents income; for respondents
who give two no responses, the lower bound is either zero (if the
distribution of WTP admits only non-negative values) or negative
infinity (if the distribution of WTP is a normal or a logistic.))
| |
log log ( ; ) ( ; ) L F WTP F WTP
H L
i
n
=
=
u u
1
31
Non parametric models for contingent
valuation: the Turnbull estimator
1. Consider only the first bid answer
2. For bids indexed j=1,,M, calculate F
j
=N
j
/(N
j
+Y
j
) where N
j
is the
number of No responses to t
j
and Y
j
is the number of Yes
responses to the same bid, Tj= N
j
+Y
j
3. Beginning with j=1, compare F
j
and F
j+1
. Intuitively, % of Nos
should increase with the increase in the bid
4. If F
j+1
>F
j
then continue
5. If F
j+1
F
j
then pool cells j and j+1 into one cell with boundaries
(t
j
,t
j+2
], and calculate F
j
*=(N
j
+N
j+1
)/(T
j
+T
j+1
)=N
j
*/T
j
*. That is,
eliminate bid tj+1 and pool responses to bid t
j+1
with responses to
bid t
j
.
6. Continue until cells are pooled sufficiently to allow for a
monotonically increasing CDF
7. Set F
M+1
*=1, F
0
*=0
8. Calculate the PDF as the step difference in the final CDF:
f
j+1
*=F
j+1
*-F
J
* for each offered price. These represent consistent
estimates of the probability that WTP falls between price j and
price j+1.
32
9. Multiply each offered price (t
j
) by the probability that WTP falls
between it and the next highest price (t
j+1
)
10. Sum the quantities from step (9) over all prices to get an estimate
of the lower bound on WTP:
11. Calculate the variance of the lower bound as:
) ( ) (
*
0
*
1 j
M
j
j j LB
F F t WTP E
=
+
=
=
*
1
2
1
*
* *
) (
) 1 (
)) ( (
M
j
j j
j
j j
LB
t t
T
F F
WTP E V
33
Unrestricted Turnbull
tj N
j
T
j
F
j
F
j
* f
j
*
5 20 54 0.370 0.343 0.343
10 15 48 0.313 Pooled back Pooled back
25 46 81 0.568 0.568 0.225
50 55 95 0.579 0.579 0.011
100 106 133 0.797 0.797 0.218
200 82 94 0.872 0.872 0.075
300 72 81 0.889 0.889 0.017
300+ - - 1 1 0.111
E
LB
(WTP)=0*0.343+5*0.225+25*0.011+50*0.218+100*0.075+
+200*0.017+300*0.111=$56.50
V(E
LB
(WTP))=(0.343*0.657/102)*(5-0)
2
+(0.568*0.432/81)*(25-5)
2
+
+(0.579*0.421/95)*(50-25)
2
+(0.797-0.203/133)*(100-50)
2
+
+(0.872*0.128/94)*(200-100)
2
+(0.889*0.111/81)*(300-200)
2
=$29.52
34
Literature for this lecture
Haab-McConnell Valuing environmental and natural resources
chapters 1-5
Perman et al. Chapter 12
A good book for the CV: Mitchell-Carson Using surveys to value
public goods: the contingent valuation method Resources for the
Future, Washington, DC, 1989.
Read the paper Alberini, Rosato, Longo, Zanatta Information and
Willingness to Pay in a Contingent Valuation Study: The Value of S.
Erasmo in the Lagoon of Venice.
Ill post the slides and the paper on my website:
http://people.bath.ac.uk/al224/