Conservation program enrollment
mechanisms using auctions: what can
laboratory experiments tell us about the
use of imprecise cost information?
Daniel Hellerstein (ERS) and Sean Sylvia (AREC/UMD)
AERE Summer Workshop, Seattle WA, June 8-10, 2011.
The views expressed are the authors and should not be attributed to the Economic Research Service or
the USDA
Motivation
Conservation programs need some means of
choosing which applicants to accept …
to wit…
an enrollment mechanism
Goals of an enrollment mechanism:
•
•
•
•
Minimizing program expenditures/ maximizing benefits
Encouraging broad participation
Inducing adoption of enhanced environmental practices
Minimizing impacts on production
Example: the CRP’s EBI
The Conservation Reserve Program (CRP) is a ~31 million acre, $1.5
billion/year program established in 1986. Objectives include erosion
control, water quality protection, and providing wildlife habitat
The CRP’s enrollment mechanism
• Offers are ranked using an Environmental Benefits
Index (EBI) that incorporates environmental
impacts and the bid.
• Each parcel’s bid can not exceed a bid cap (a
maximum bid).
Cost heterogeneity
• Landowner costs are heterogeneous.
–If a single price were paid to all offers, owners of
low cost parcels could earn substantial rents
– A precise bid cap (equal to a parcel’s opportunity
cost) could deliver substantial savings to program
administrators.
•However, a poorly chosen bid cap
can increase total expenditures
Example: unbiased bid caps can stink
10 parcels with heterogeneous cost (but otherwise the same)
Goal: accept 5 of 10
parcels, whose cost range
from 1 to 10
Two bid cap measures:
1. accurate/unbiased
2. less accurate/upwardly
biased
Type
Total cost
Actual
15
Single price (6th
highest cost)
30
Less accurate
cap
18.7
Accurate &
unbiased cap
31.5
Prior findings: Stringent bid caps lead to
higher acquisition costs
Max bids were varied in stringency, from 80% (of a tickets
maximum possible cost) to 120%
• 80% yielded the highest acquisition costs
• 120% yielded acquisition costs similar to 80%
• Costs were minimized at 90%
Goals of this study
In auction setting with asymmetric bidders and noisy
assessments:
• What are the performance characteristics of several
different auction mechanisms?
We examined:
• Quotas
• Target bids
• Endogenous target bids
Ranking
Choices
Instrument
Design
Experimental Design
Twelve 1.5 hour sessions
On average, 10 participants per
session
5 or 6 treatments per session
8 auctions per treatment
Participant’s receive two
“tickets” per auction
Tickets have a randomly generated
cost, and a fairly large bid cap
Participant enters bids (BID) on zero, one or both tickets
Participant can also purchase “points” (q) on zero, one or both tickets
Simple case:
SCORE=BID- q
Accept the 12 lowest scoring
tickets
Earnings:
EARN= BID – ticketCost - (0.5 x q)
Asymmetric costs
Tickets belong to one of 4 types
Description of type
Cost range
Bid cap
Target bid
(a) Low cost with
low variance
30 -45
90
39
(b) Low cost with
medium variance
20-65
130
45
(c) Medium cost
with medium
variance
35 – 95
190
71
(d) High cost with
high variance
40- 150
300
94
Each participant receives:
1. an (a) or a (b) ticket, and
2. a (c ) or a (d) ticket
Three treatments
(that complement the “basic” treatment)
Description
Score calculation
(lower scores are better)
Quota
A two stage acceptance procedure:
1. Within each type, the highest scoring ticket is
dropped
2. The survivors are pooled, with A (A=12
usually) lowest scoring tickets accepted
BID- q
(q=quality points
purchased)
target
Bid
A target bid is assigned to each ticket type.
• Bids below the targetBid: score is reduced
• Bids above the targetBid: score is increased
BID - q +
((BID-0.5q)–targetBid) )
Endog
target
Bid
• Target is announced after all bids are received
• Target is the lowest bid received for this ticket type
• Thus, scores will be increased for all tickets
BID - q +
((BID-0.5q)–lowest_t) )
lowest_t= lowest bid
received for type t
tickets
Sample screen: basic treatment
Sample screen: targetBid treatment
Sample screen: quota treatment
Sample screen: endog targetBid
Example: session 5, round 6
(standard treatment)
$ and points
150
offer
90
cost
30
points
0
5
Type A
10
Type B
15
Type C
20
Type D
-30
Tickets, sorted by ticket-type (small to
large bid cap)
Example: session 5, round 28
(targetBid treatment)
Analysis
We focus on aggregate results
Several linear regressions are used to discern the
impacts of the several treatments.
Prior findings: Quota auctions can reduce
acquisition costs
• Using two ticket types, imposing a quota reduced acquisition costs by an average
of 8%.
• Cost reduction due to reduced bids by “low costs” tickets was greater than cost
increases due to accepting higher cost tickets
Standard
targetBid
quota
Endog
target
Profit
rate
Some aggregate dependent variables…
Mean (sd)
optCost
“optimal” cost of
acquisition
A
 sCost
 0.5 *  qt Acceptt
i
i
expend
Expenditures
 Accept * ( Bid
t
tickets)
sCost
t
T
(bids of accepted
524.3 (93.4)
T
t
875.7 (151.9)
 0.5qt )
t
Social cost (cost
T
 Accept  Cost
of accepted tickets)
t
t
569.1 (98.4)
 0.5qt 
t
expEff
socEff
Efficiency
avgProfit
Average profit
(of accepted
offers)
expenditures, or
forgone production,
over optimal cost:
expend/optCost
sCost/optCost
T
 Accept
t
1.68 (0.17)
1.09 (0.06)
21.3 (6.0)
* ( Bid t  (Costt  0.5qt ))
t
A
)
A = # tickets accepted,
T = # tickets offered
sCost= sorted ticket costs ,low to high (t=1..T)
Accept: 0/1 dummy, 1 if accepted (t=1..T)
Basic regressions
N=206, (t-stat)
expEff
avgProfit
socEff
intercept
1.19
5.7 (4.8)
1.1 (13.5)
target
-0.18
(-8.9)
-10.5
0.030 (2.5)
endogTarget
-0.13
(-6.0)
-7.3 (-9.3)
0.039 (3.1)
quota
-0.14
(-5.81)
-7.7 (- 8.3)
0.023
maxPrior
0.0054 (6.8)
0.21 (7.1)
0.00064 (1.4)
vickCost
-0.018 (-13.6)
-0.33 (-6.2)
-0.0023
vickRatio
0.73
12.5
0.09
Exper
-0.0041 (-0.9)
0.034 (.21)
-0.004 (-1.9)
qSmart
0.32
7.5 (3.1)
0.016 (0.4)
R-square [ f-stat ]
0.66 [< 0.001]
0.58 [< 0.001]
0.18 [< 0.001]
(9.9)
(7.3)
(5.2)
(-12.9)
(3.2)
(1.7)
(-2.7)
(1.4)
Panel regressions (panel=round number)
N=206
avgProfit (haus prob<0.81)
expEff (haus prob<0.01)
RE
FE
RE
FE
Intercept
5.7 (1.2)
-2.1 (-0.3)
1.19 (9.9)
1.20 (6.4)
Target
-10.5 (-12.9)
-8.2 (-5.9)
-0.18 (-8.9)
-.014 (-4.1)
endogTarget -7.9 (-9.3)
-5.4 (-3.9)
-0.12 (-6.0)
-0.083 (-2.3)
Quota
-7.7 (-8.4)
-5.2
-0.13 (-5.8)
-0.088 (-2.4)
maxPrior
0.21 (7.1)
0.21 (6.5)
0.0053 (6.8)
0.0051 (6.3)
vickCost
-0.33 (-6.2)
-0.37 (-2.4)
-0.18 (-13.6)
-0.015 (-3.8)
vickRatio
12.1 (3.2)
18.3 (2.1)
0.72 (7.33)
0.59 (2.7)
Exper
0.04 (0.2)
0.052 (-0.3)
-0.004 (-0.9)
-0.0069 (-1.6)
qSmart
7.5 (3.1)
9.9
0.32 (5.2)
0.37 (5.9)
R-sq [f-stat]
0.47 [<0.0001]
0.55
[<0.0001]
0.66 [<0.0001]
0.64 [<0.0001)
(-3.7)
(4.0)
Difference models
Y= avgProfit
intercept
Difference
Difference in difference
1.43 (109)
1.33 (7.8)
target
Change in targetBid
treatment
-9.7 (-37.2)
-6.9 (-10.1)
endogTarget
Change in endogTarg
treatment
-7.49 (-27.3)
-7.2 (-10.7)
Quota
Change in quota
treatment
-7.08 (-23.0)
-5.5 (-8.1)
maxPrior
Max accepted bid prior
round
0.093 (8.1)
0.02 (1.9)
vickCost
Total cost in a vickery
auction cost
-0.39 (-25.7)
n.a.
vickRatio
vickCost/optcost
15.8 (14.1)
n.a.
qSmart
1 if perfect q point
useage
3.3 (3.1)
-0.027 (-0.03)
R-square [f-stat]
0.54 [< 0.001]
0.06 [< 0.001]
N
1822
2565
Conclusions
• Use of an alternative auction mechanism can decrease
program expenditures
• Bid targets seem to be somewhat more effective than
endogenous bid targets or quotas
• Cost savings vary around 10%
• There is an increase in social costs (as more expensive
“lands” are enrolled), that range around 3% (and that are
highest in endogenous bid treatments)
…. Other conclusions
• Quality improvements, even when unambiguously
beneficial, are often not undertaken
• Failure to utilize quality improvements seems to be
related to
• treatment type (endogenous targetBid
treatments had worse results),
• bidder competence (accepted offers use quality
points more effectively)
Future work
• Examine other cost range distributions, and other
allocations of quality points
• Devise a structural model that uses individual
observations (ticket within an auction), to replace the
convenient aggregate models.
• Farmers & agricultural landowners as experimental
subjects
• etc etc etc
References:
Higgins, Nathaniel*, Michael Roberts*, and Daniel Hellerstein, 2011, Using Quotas to
Enhance Competition in Asymmetric Auctions: A Comparison of Theoretical and
Experimental Outcomes, for submission to Games and Economic Behavior
Hellerstein, Daniel* and Nathaniel Higgins*, 2010, “The Effective Use of Limited
Information: Do Bid Maximums Reduce Procurement Costs in Asymmetric Auctions?”
Agricultural and Resource Economics Review, 39 (2, April):288-304
Higgins, Nathaniel. "Computational and Experimental Market Design." PhD
Dissertation, University of Maryland, College Park, 2010.
Hellerstein, Daniel* and Nathaniel Higgins, 2009, “The Effective Use of Limited
Information: Do Bid Maximums Reduce Procurement Costs in Asymmetric
Auctions?”, Presentation at Northeastern Agricultural and Resource Economics
Association conference, Burlington VT, June
Higgins, Nathaniel*, Michael Roberts*, and Daniel Hellerstein, 2008, “Cost Saving
Procurement Auctions for Environmental Services”, Poster presentation at AAEA
Summer Meetings, Denver CO, July
Appendix: miscellaneous
tidbits
The CRP
Current enrollment (April 2011): 31.2 million acres
40
2.5
35
Million Acres
25
1.5
20
1
15
10
Billion Dollars
2
30
0.5
5
20
08
20
06
20
04
20
02
20
00
19
96
19
98
19
94
19
92
19
90
0
19
88
19
86
0
Year
CRP Acres
Source: ERS using FSA CRP contract data as of
October 2009
Continuous CRP Acres
Yearly $ Outlay
• Current acreage is a 5.6 million acre drop from
the 2007 peak (36.8 million)
• Current acreage includes 5.0 million acres of
continuous signup
• Average cost per acre: $46 for general, $102 for
continuous
Optimal: offer actual cost to everyone
Total Expenditure: 15
Offered,
and rejected
Offered, and
accepted
parcel
Examples of too stringent maximums
Offer 6 to everyone
Total Expenditure: 30
Expenditures, when using a single price, is twice actual cost
Offered &
accepted
Not offered
parcel
Offer the less accurate, and upwardly biased, cost
Total Expenditure: 18.7
A moderately upwardly biased cap is almost as good as the optimal
Offered &
rejected
Offered &
accepted
parcel
Offer the more accurate, and unbiased, cost
Total Expenditure: 31.5
An unbiased, and accurate, bid cap can be significantly worse than a
biased/inaccurate cap, and can be worse than single price
Not
offered
Offered &
accepted
How well are quality points used?
What influences efficient use of q points?
N=6470
Coefficient
(tstat)
Mean (stddev)
Min / Max
Intercept
0.51 (12.4)
Round
-0.0041 (-5.6)
18 (9.7)
2 / 38
cheapTalk
0.27 (0.42)
0.02
0 / 1 dummy
compQuests
0.07 (9.0)
4.6 (0.7)
2/5
Cost
-0.002 (-11.3)
60.4 (30.3)
20 / 150
Target
-0.15 (-8.9)
0.32
0 / 1 dummy
endogTarget
-0.16 (-8.6)
0.21
0 / 1 dummy
Quota
-0.085 (-4.2)
0.17
0 / 1 dummy
R-square
[f-test prob]
0.071
[<0.0001]
0.57 (0.46)
0 / 1
qSmart
(dependent
variable)
Models
Basic:
Ysr  Tsr T  X sr  X  Cr c  Z s  z   s   r   rs
Where:
s,r = session, and round within session
T = vector of treatment dummies (target, endogTarget, quota), only one of which is non-zero
X = round specific variables (such as maxPrior and qSmart)
Z = session descriptor (such as exper)
C = round specific costs (such as vickCost and vickRatio) – these have the same (or nearly the same)
values in round r, regardless of session,
eps = error components: session specific (s), round specific ( r), and observation specific (rs)
Y = average profit, ratio of actual expenditures to optimal (full information) cost, or ratio of actual costs
(of enrolled parcels) expenditures to optimal cost.
Panel (within round):
Y r  T r T  X r  X  C r c  Z s   s   rs
Notes:
Variables are demeaned , using round-specific means.
The eps_r “round specific” error component is conditioned out by the FE or RE estimator. (the
demeaning or quasi-demeaning).
The C variables, due to changes in # of participants, are not completely removed. However, their
variance is reduced, hence one expects they will have reduced influence in the regression.
Difference (within session):
Ys , r12  Ts , r12 T  X s , r12 X  Cs , r12  c   r12   s , r12
Notes:
This model uses all “pairs” or rounds that share a session, and that have at least one of the T
variables differ.
•Thus, it is a panel model, where each panel use observations within a single session.
• Note that the first differencing uses all “interesting” pairs within a session (it is not a simple
“adjacent round” first difference).
Definition of “first differencing” for a pair of observations in panel s:
x_s,r12 = x_s,r2 – x_s,r1
(where r1 and r2 are rounds).
The eps_s “session specific” error component is conditioned out by the first differencing.
The Z variables are also conditioned out.
Note that delta T can be negative, which means a treatment was no longer used.
Difference of differences:
2Ys12, r12  2Ts12, r12T  2 X s12, r12 X  2 s12, r12
Notes:
This model compares “pairs” of rounds between sessions s1 and s2.
•Each comparison uses pairs (S2, S2) that have the same first (r1) and second round (r2).
• S1 must have a round that is identical (in terms of T) to a round in S2
• S1 must have a round that is different than a round in S2
Thus, the difference within a pair is compared to a difference within another pair.
Definition of “difference in differencing” for a pair of observations spanning rounds r1 and r2, in
sessions s1 and s2:
x_s12r1r2 = (x_s1,r2 – x_s1,r1) - (x_s2,r2 – x_s2,r1)
Where the r2 treatments are the same, and the r1 treatments are different.
Note that the difference in differencing :
•Controls for Z (the first difference removes session specific variables)
•Controls for changes in eps_r (the within pair changes are the same)
•Controls for changes in cost structure (since r1 and r2 have the very similar cost structure
across all sessions), hence C is essentially conditioned out.