# RECORD 2ce98014780d8a025c9acf3e943a33fd71e8d5b6
# TITLE: Forest pests and home values: The importance of accuracy in damage assessment and geocoding of properties
# YEAR: 2017.0  VENUE: nan  DOI: 10.1016/j.jfe.2017.02.002
# AUTHORS: K. Moeltner; C. Blinn; T. Holmes
# PAGES: 10  CANDIDATE TABLE PAGES (1-based): [6, 8]
# ABSTRACT: We examine the impact of measurement errors in geocoding of property locations and in the assessment of Mountain Pine Beetle-induced tree damage within the proximity of a given residence on estimated losses in home values. For our sample of homes in the wildland-urban interface of the Colorado front range and using a novel matching estimator with Bayesian regression adjustment we find that both types of errors can lead to substantial biases in estimated losses. Our results confirm that the Forest Service's Aerial Detection Survey is generally too coarse to be informative for property valuation that depends on highly localized spatial data.


===== PAGE 1 (score 11, numeric-paren 0, table-words 0, var-words 11) =====

Journal
 of
 Forest
 Economics
 26
 (2017)
 46–55
Contents
 lists
 available
 at
 ScienceDirect
Journal
 of
 Forest
 Economics
jo
 ur
 nal
 ho
 me  p
 age:
 www.elsevier.com/locate/jfe
Forest
 pests
 and
 home
 values:
 The
 importance
 of
 accuracy
 in
 damage
assessment
 and
 geocoding
 of
 properties
Klaus
 Moeltner a,∗,  Christine
 E.  Blinn b, Thomas
 P.
 Holmes c
a Department
 of
 Agricultural
 and
 Applied
 Economics,
 Virginia
 Tech,
 Blacksburg,
 VA,
 USA
b Department
 of
 Forest
 Resources
 and
 Environmental
 Conservation,
 Virginia
 Tech,
 Blacksburg,
 VA,
 USA
c USDA
 Forest
 Service,
 Southern
 Research
 Station,
 Research
 Triangle
 Park,
 NC,
 USA
a
 r  t
 i c  l e  
i  n  f  o
Article
 history:
Received
 6
 December
 2016
Accepted
 2
 February
 2017
Keywords:
Mountain
 Pine
 Beetle
Matching
 estimator
Measurement
 error
Aerial
 Detection
 Survey
a
 b
 s
 t  r  a
 c  t
We  examine
 the  impact  of measurement
 errors  in  geocoding
 of  property
 locations
 and  in  the  assessment
of
 Mountain
 Pine  Beetle-induced
 tree  damage
 within  the  proximity
 of a given
 residence
 on  estimated
losses
 in  home  values.  For  our  sample
 of  homes
 in  the  wildland-urban
 interface
 of  the  Colorado
 front
range
 and  using  a  novel
 matching
 estimator
 with  Bayesian
 regression
 adjustment
 we  find  that  both
 types
of errors  can lead  to substantial
 biases
 in  estimated
 losses.
 Our results  confirm
 that  the  Forest  Service’s
Aerial
 Detection
 Survey  is generally
 too  coarse  to be  informative
 for  property
 valuation
 that  depends
 on
highly
 localized
 spatial
 data.
© 2017  Department
 of  Forest  Economics,
 Swedish
 University
 of Agricultural
 Sciences,  Ume ˚a.
Published
 by  Elsevier
 GmbH.  All  rights
 reserved.
Introduction
The
 accuracy
 of
 address-based
 geocoding
 (ABG),
 as
 used
 in
 many
geographic
 information
 system
 (GIS)
 software
 packages,
 has
 been
subject
 to
 scrutiny
 in
 recent
 years
 by
 spatially
 focused
 disciplines
such
 as
 geography
 and
 epidemiology.
 The
 emerging
 consensus
 in
these
 areas
 is
 that
 ABG
 can
 lead
 to
 measurable
 positional
 errors,
when
 compared
 to
 ground-truthing
 or
 interpretation
 via
 aerial
imagery.
 Furthermore,
 these
 errors
 tend
 to
 be
 more
 pronounced
in
 rural
 areas
 with
 lower
 population
 density.
For
 example,
 Cayo
 and
 Talbot
 (2003)
 examine
 3000
 addresses
of
 residential
 properties
 in
 upstate
 New
 York
 and
 find
 a
 median
error,
 in
 terms
 of
 the
 Euclidean
 distance
 between
 the
 true
 and
 the
presumed
 location
 of
 the
 actual
 housing
 structure,
 of
 38
 meters
(m),
 78
 m,
 and
 201
 m
 in
 urban,
 suburban,
 and
 rural
 settings,
 respec-
tively.
 Schootman
 et
 al.
 (2007), in
 turn,
 report
 a
 mean
 error
 of
 42
 m
and
 a
 median
 deviation
 of
 31
 m
 for
 299
 addresses
 in
 urban
 St.
 Louis,
Missouri.
 Zandbergen
 (2007), considering
 over
 100,000
 addresses
in
 urban/suburban
 Orange
 County,
 Florida,
 discovers
 a
 mean
 error
of
 66
 m,
 with
 a
 standard
 deviation
 of
 over
 430
 m.
As
 mentioned
 in
 Burra
 et
 al.
 (2002), and
 discussed
 in
 detail
 in
Zandbergen
 (2007)
 these
 positional
 errors
 can
 introduce
 serious
bias
 into
 an
 underlying
 statistical
 analysis
 if
 they
 are
 systematic
 in
nature,
 which
 is
 often
 the
 case
 in
 a
 given
 empirical
 application. For
∗Corresponding
 author.
example,
 in
 Zandbergen
 (2007)’s
 context
 ABG
 generally
 “moves”
residential
 locations
 closer
 to
 major
 traffic
 arteries
 than
 they
 really
are.
 This
 leads
 to
 an
 over-estimation
 of
 the
 number
 of
 children
exposed
 to
 traffic-related
 air
 pollution.
 Similar
 biases
 are
 reported
in
 Harada
 and
 Shimada
 (2006)
 and
 in
 Hay
 et
 al.
 (2009)
 in
 the
 context
of
 spatial
 densities
 and
 hot
 spots
 for
 crime
 locations.
Somewhat
 surprisingly
 given
 the
 recent
 proliferation
 in
 spa-
tially
 explicit
 property
 valuation
 studies,
 the
 potential
 inaccuracy
of
 ABG
 has
 largely
 been
 ignored
 in
 the
 applied
 economics
 literature.
As
 we  illustrate
 in
 this
 study,
 pinpointing
 the
 accurate
 location
 of
a
 private
 residence
 can
 be
 very
 important
 if
 the
 effect
 of
 environ-
mental
 changes
 in
 the
 immediate
 vicinity
 of
 a
 home
 constitutes
 the
central
 focus
 of
 a
 given
 analysis.
In
 our
 application
 we
 consider
 tree
 mortality
 induced
 by
 the
Mountain
 Pine
 Beetle
 (MPB)
 within
 a
 close
 perimeter
 of
 a
 given
residence
 in
 the
 wildland-urban
 interface
 (WUI)
 of
 the
 Colorado
front
 range.
 The
 MPB
 (dentroctonus
 ponderosae) is
 a
 native
 forest
pest
 that
 has
 shown
 explosive
 population
 growth
 in
 recent
 years,
with
 annual
 forest
 destruction
 rates
 in
 North
 America
 and
 Canada
comparable
 to
 that
 of
 all
 wildfires
 combined
 (Carroll
 et
 al.,
 2004;
Bentz
 et
 al.,
 2010;
 Sims
 et
 al.,
 2010;
 Man,
 2012).
This
 poses
 a
 second
 spatially
 explicit
 risk
 of
 bias
 in
 estimated
 real
estate
 capitalization
 rates
 if
 tree
 mortality
 itself
 is
 assessed
 with
systematic
 error.
 Specifically,
 we  examine
 the
 home-level
 accu-
racy
 of
 MPB
 damage
 as
 determined
 by
 the
 U.S.
 Forest
 Service’s
Aerial
 Detection
 Survey
 (ADS)
 relative
 to
 case-by-case
 photo-
interpretation
 by
 a
 remote
 sensing
 expert.
http://dx.doi.org/10.1016/j.jfe.2017.02.002
1104-6899/©
 2017
 Department
 of
 Forest
 Economics,
 Swedish
 University
 of
 Agricultural
 Sciences,
 Ume ˚a.
 Published
 by
 Elsevier
 GmbH.
 All
 rights
 reserved.


===== PAGE 6 (score 44, numeric-paren 0, table-words 12, var-words 9) =====

K.
 Moeltner
 et
 al.
 / Journal
 of
 Forest
 Economics
 26
 (2017)
 46–55
 
51
Table  2
Sample
 statistics.
Treated
 observations
 
Eligible
 control
 observations
Variable
 
Mean
 / pct.
 
std
 
min
 
max
 
mean
 
std
 
min
 
max
Full
 sample
 (treated
 =
 304,
 eligible
 controls
 = 604)
Price
 (000s)
 
$355
 
$191
 
$77
 
$1,300
 
$460
 
$364
 
$55
 
$4,313
Age
 
30.67
 
16.61
 
0
 
92
 
30.87
 
23.53
 
0
 
151
sqft
 (00s)
18.80
 
8.46
 
3.68
 
66.24
 
21.29
 
9.57
 
4.71
 
71.47
Lot
 sqft.
 (000s)
 
114.96
 
163.19
 
0.87
 
1590.30
 
101.68
 
259.06
 
0.00
 
2132.70
Bedrooms
 
3.02
 
0.94
 
0
 
7
 
3.24
 
0.96
 
1
 
9
Bathrooms
 
2.37
 
0.97
 
0
 
6
 
2.75
 
1.18
 
0
 
9
Near
 fire
 
11.84%
 
23.84%
Sold
 1st
 quarter
 
21.71%
 
25.00%
Sold
 2nd
 quarter
31.25%
 
25.17%
Sold
 3rd
 quarter
29.93%
 
34.11%
Sold
 4th
 quarter
 
8.22%
 
8.61%
Golden
 (treated
 =
 161,
 eligible
 controls
 = 349)
Price
 (000s)
 
$413
 
$208
 
$78
 
$1,300
 
$469
 
$255
 
$81
 
$2,200
Age
 
30.92
 
16.61
 
3
 
89
 
28.93
 
21.56
 
0
 
136
sqft
 (00s)
 
20.77
 
9.11
 
5.76
 
66.24
 
23.10
 
9.56
 
4.71
 
71.47
Lot
 sqft.
 (000s)
 
100.53
 
154.02
 
0.87
 
1590.30
 
66.24
 
188.45
 
0.00
 
2132.70
Bedrooms
 
3.27
 
0.93
 
1
 
7
 
3.42
 
0.96
 
1
 
9
Bathrooms
 
2.68
 
0.92
 
1
 
6
 
3.04
 
1.16
 
1
 
9
Near
 fire
6.83%
 
22.64%
Sold
 1st
 quarter
 
19.88%
 
23.50%
Sold
 2nd
 quarter
 
32.92%
 
27.22%
Sold
 3rd
 quarter
 
27.95%
 
34.38%
Sold
 4th
 quarter
 
8.07%
 
8.60%
Boulder
 (treated
 =
 20,
 eligible
 controls
 =
 86)
Price
 (000s)
$315
 
$108
 
$159
 
$589
 
$737
 
$705
 
$108
 
$4,313
Age
 
40.30
 
15.19
 
9
 
70
 
38.13
 
25.85
 
3
 
131
sqft
 (00s)
 
15.55
 
5.22
 
7.38
 
24.57
 
21.11
 
11.61
 
5.86
 
61.27
Lot
 sqft.
 (000s)
 
154.82
 
232.83
 
17.09
 
916.50
 
158.97
 
393.20
 
0.71
 
1742.40
Bedrooms
 
3.05
 
1.05
 
1
 
5
 
3.21
 
0.93
 
1
 
5
Bathrooms
 
2.00
 
0.86
 
1
 
3
 
2.77
 
1.22
 
1
 
7
Near
 fire
 
45.00%
 
66.28%
Sold
 1st
 quarter
30.00%
 
23.26%
Sold
 2nd
 quarter
 
15.00%
 
27.91%
Sold
 3rd
 quarter
 
40.00%
 
34.88%
Sold
 4th
 quarter
 
15.00%
 
9.30%
std:
 standard
 deviation;
 min
 (max):
 minimum
 (maximum);
 pct.:
 percent.
than
 treated
 properties,
 and
 control
 homes
 in
 Boulder
 have
 higher
average
 square
 footage
 than
 treated
 residences.
 We  use
 the
 basic
housing
 attributes
 captured
 in
 the
 table
 in
 our
 balanced
 matching
procedure,
 and
 add
 to
 these
 the
 fire
 and
 sales
 quarter
 indicators
 in
the
 auxiliary
 regression
 model.
Table
 3
 shows
 sample
 counts
 before
 and
 after
 matching,
 as
 well
as
 post-matching
 counts
 for
 the
 three
 types
 of
 errors
 that
 can
 occur
using
 one
 of
 the
 flawed
 interpretation
 methods.
 The
 first
 block
 of
rows
 refers
 to
 the
 full
 sample.
 The
 first
 two
 rows,
 which
 give
 the
number
 of
 identified
 treated
 and
 eligible
 control
 properties,
 are
repeated
 from
 Table
 1. The
 third
 row
 shows
 the
 remaining
 control
observations
 after
 matching,
 that
 is
 the
 set
 of
 controls
 that
 actually
feed
 into
 the
 econometric
 analysis
 and
 the
 estimated
 treatment
effect.
 In
 all
 cases
 the
 number
 of
 these
 matched
 controls
 is
 smaller
than
 the
 number
 of
 treated.
 This
 is
 as
 expected,
 since
 the
 balancing
and
 matching
 algorithms
 allow
 for
 replacement
 in
 the
 selection
 of
controls,
 following
 best
 practices
 in
 the
 matching
 literature
 (e.g.
 Ho
et
 al.,
 2007;
 Abadie
 and
 Imbens,
 2011).
Rows
 four
 through
 six
 provide
 a
 detailed
 look
 at
 omission
and
 swapping
 errors.
 As
 is
 evident
 from
 the
 table,
 all
 three
 non-
benchmark
 approaches
 introduce
 all
 three
 types
 of
 errors,
 although
error
 counts
 are
 substantially
 smaller
 for
 the
 “PI,
 ABG”
 version
compared
 to
 the
 ADS-based
 approaches.
 Specifically,
 the
 latter
drop
 over
 50%
 of
 actually
 treated
 from
 the
 analysis.
 As
 discussed
above
 this
 not
 only
 hurts
 efficiency
 via
 reduced
 sample
 sizes,
 but
can
 introduce
 selection
 bias
 if
 this
 attrition
 is
 non-random.
 Post-
matching
 swapping
 errors
 are
 also
 considerable
 for
 the
 ADS-based
strategies,
 with
 approximately
 17%
 of
 actually
 treated
 moved
 to
the
 bin
 of
 matched
 controls,
 and
 approximately
 10%
 of
 controls
interpreted
 as
 treated.
Analogous
 results
 for
 Golden
 are
 captured
 in
 the
 second
 block
of
 rows.
 As
 can
 be
 seen
 from
 the
 table,
 the
 benchmark
 remote
sensing
 approach
 identifies
 161
 treated
 observations,
 which
 leaves
349
 eligible
 control
 homes.
 After
 balanced
 matching
 113
 of
 these
are
 actually
 selected
 as
 matched
 controls,
 with
 some
 serving
 as
matches
 for
 multiple
 treated.
 In
 comparison,
 the
 “PI,
 ABG”
 inter-
pretation
 completely
 omits
 five
 treated
 homes
 from
 the
 analysis,
erroneously
 sorts
 three
 treated
 homes
 into
 the
 bin
 of
 matched
 con-
trols,
 and
 misinterprets
 two
 control
 homes
 as
 treated.
 As
 for
 the
sample
 at
 large,
 errors
 introduced
 by
 the
 ADS-based
 methods
 are
more
 staggering,
 leading
 to
 the
 omission
 of
 close
 to
 80%
 of
 actually
treated
 in
 both
 cases.
 In
 addition,
 several
 treated
 homes
 (12
 and
15,
 respectively)
 are
 misplaced
 into
 the
 matched
 control
 bin,
 and
 a
comparable
 number
 of
 control
 cases
 (11
 and
 10,
 respectively)
 are
falsely
 used
 as
 treated
 properties.
Results
 for
 Boulder
 are
 given
 in
 the
 bottom
 block
 of
 rows
 of
Table
 3.
 The
 benchmark
 interpretation
 produces
 20
 treated
 homes,
86
 eligible
 controls,
 and
 15
 matched
 controls.
 Working
 with
 flawed
geocodes
 in
 the
 “PI,
 ABG”
 interpretation
 drops
 four
 treated
 alto-
gether,
 and
 misplaces
 another
 treated
 and
 three
 controls.
 While
these
 figures
 may
 seem
 diminutive,
 they
 still
 constitute
 a
 sizable
share
 of
 all
 available
 observations
 for
 this
 small
 sample.
 As
 is
 the
case
 for
 the
 full
 sample
 and
 Golden,
 the
 ADS-based
 approaches
 drop
even
 more
 treated
 (eight
 in
 both
 cases)
 and
 induce
 some
 erroneous
swapping
 of
 actually
 treated
 (one
 in
 each
 case)
 and
 actual
 controls
(10
 in
 each
 case).


===== PAGE 7 (score 21, numeric-paren 0, table-words 5, var-words 11) =====

52
 
K.
 Moeltner
 et
 al.
 / Journal
 of
 Forest
 Economics
 26
 (2017)
 46–55
Table
 3
Sorting
 effects
 based
 on
 remote
 sensing
 approaches.
PI,
 vis.
 
PI,
 ABG
 
ADS,
 vis.
 
ADS,
 ABG
Full
 sample
Treated
 
304
 
293
 
146
 
139
Eligible
 controls
 
604
 
615
 
762
 
769
Matched
 controls
 
206
 
200
 
120
 
114
Unused
 actually
 treated
 
–
 
10
 
173
 
183
Actually
 treated
 used
 as
 matched
 controls
 
–
 
11
 
50
 
45
Actual
 controls
 used
 as
 treated
 
–
 
10
 
65
 
63
Golden
Treated
 
161
 
155
 
32
 
28
Eligible
 controls
 
349
 
355
 
478
 
482
Matched
 controls
 
113
 
114
 
30
 
28
Unused
 actually
 treated
 
–
 
5
 
128
 
128
Actually
 treated
 used
 as
 matched
 controls
 
–
 
3
 
12
 
15
Actual
 controls
 used
 as
 treated
 
–
 
2
 
11
 
10
Boulder
Treated
 
20
 
18
 
21
 
21
Eligible
 controls
 
86
 
88
 
85
 
85
Matched
 controls
15
 
15
 
18
 
18
Unused
 actually
 treated
 
–
 
4
 
8
 
8
Actually
 treated
 used
 as
 matched
 controls
 
–
 
1
 
1
 
1
Actual
 controls
 used
 as
 treated
 
–
 
3
 
10
 
10
PI:
 damage
 assessment
 via
 expert
 photo-interpretation;
 ADS:
 damage
 assessment
 via
 Aerial
 Detection
 Survey;
 ABG:
 address-based
 geocoding
 using
 generic
 GIS
 software;
vis.:
 expert
 geocoding
 using
 photo-imagery.
As
 discussed
 previously,
 swapping
 effects,
 by
 themselves,
should
 unambiguously
 bias
 the
 ATT
 towards
 zero.
 In
 theory,
 the
omission
 effect,
 by
 itself,
 can
 be
 bias-neutral
 if
 omissions
 are
 com-
pletely
 random.
 However,
 the
 bias
 can
 go
 towards
 either
 direction
if
 omitted
 treated
 homes
 differ
 from
 the
 sample
 at
 large
 in
 relevant
dimensions.
 In
 our
 case,
 such
 a
 dimension
 could
 be
 the
 intensity
 of
infestation,
 that
 is
 the
 actual
 count
 or
 proportion
 of
 affected
 trees
surrounding
 a
 property.
 For
 example,
 ADS
 fly-overs
 may
 focus
 pri-
marily
 on
 the
 most
 heavily
 affected
 areas
 and
 miss
 low-intensity
impacts.
 This
 would
 bias
 the
 ATT
 away
 from
 zero
 and
 lead
 to
 an
over-estimation
 of
 losses.
 Similarly,
 if
 homes
 are
 “moved”
 from
their
 woody
 surroundings
 to
 the
 center
 of
 a
 highway
 by
 ABG,
 it
is
 feasible
 that
 the
 erroneous
 new
 location
 will
 “pick
 up”
 some
 tree
damage
 only
 in
 the
 most
 heavily
 affected
 areas.
 Our
 expert
 assess-
ment
 did
 not
 capture
 actual
 counts
 of
 diseased
 trees
 due
 to
 budget
and
 time
 limitations,
 which
 preempts
 a
 rigorous
 quantitative
 ver-
ification
 of
 intensity-based
 selection.
Econometric
 analysis
As
 mentioned
 above,
 we
 match
 homes
 on
 age
 (in
 years
 since
construction),
 square
 footage
 (in
 units
 of
 100
 square
 feet),
 lot
 size
(in
 1000s
 of
 square
 feet),
 number
 of
 bedrooms,
 and
 number
 of
 bath-
rooms.
 To
 check
 on
 balance
 within
 the
 GenMatch
 algorithm,
 we
also
 include
 all
 possible
 squared
 and
 interaction
 terms
 for
 these
variables,
 as
 suggested
 in
 Diamond
 and
 Sekhon
 (2013). Example
balancing
 results
 for
 the
 benchmark
 case
 are
 given
 in
 Appendix
 B.
The
 original
 set
 of
 matching
 variables
 are
 also
 included
 in
 the
 aux-
iliary
 regression
 that
 underlies
 the
 regression-adjusted
 ATT,
 with
the
 addition
 of
 binary
 indicators
 for
 the
 occurrence
 of
 a
 wildfire
within
 five
 kilometers
 and
 the
 past
 five
 years,
 as
 well
 as
 the
 annual
quarter
 in
 which
 a
 give
 sale
 occurred.
 The
 full-sample
 regression
also
 contains
 sub-regional
 fixed
 effects.
As
 described
 above,
 the
 auxiliary
 regression
 is
 estimated
 in
 a
Bayesian
 framework,
 with
 diffuse
 normal
 priors
 for
 all
 coefficients
(mean
 =
 0,
 variance
 =
 109),
 and
 a
 flat
 inverse-gamma
 prior
 for
 the
variance
 (shape
 =
 scale
 =
 1/2).
 We  let
 our
 Gibbs
 Sampler
 burn
 in
 for
5000
 draws
 to
 reduce
 the
 effect
 of
 starting
 values,
 and
 retain
 10,000
additional
 draws
 of
 parameters
 for
 inference.
 Specifically,
 we  use
each
 draw
 of
 regression
 coefficients
 to
 generate
 an
 estimate
 of
 pre-
dicted
 prices
 for
 a
 given
 treated
 home
 and
 its
 matched
 control,
that
 is
 the
 terms
 ˆ(xi)
 and
 ˆ(xj) in
 Eq.
 (3). These
 predicted
 prices
then
 feed
 into
 an
 estimate
 of
 the
 overall
 ATT
 via
 Eq.
 (2). Ultimately,
therefore,
 our
 Bayesian
 model
 de
 facto
 produces
 10,000
 draws
 to
characterize
 the
 distribution
 of
 the
 sought
 ATT.
Estimation
 results
 are
 captured
 in
 Table
 4. As
 before,
 the
 table
exhibits
 three
 blocks
 of
 rows
 for
 the
 full
 sample,
 Golden,
 and
 Boul-
der,
 respectively.
 The
 columns
 show
 the
 posterior
 mean,
 standard
deviation,
 proportion
 of
 posterior
 distribution
 to
 the
 right
 of
 zero
(p
 >
 0),
 and
 the
 lower
 and
 upper
 bound
 of
 the
 95%
 highest
 posterior
density
 interval
 (HPDI).
 The
 p
 >
 0
 statistic
 provides
 an
 at-a-glance
assessment
 if
 the
 ATT
 is
 unambiguously
 negative
 (p
 >
 0
 is
 close
 to
zero),
 positive
 (p
 >
 0
 is
 close
 to
 one),
 or
 indeterminate
 in
 sign
 (p
 >
 0
is
 close
 to
 0.5).
 Thus,
 it
 can
 be
 used
 as
 an
 informal
 indication
 of
significance,
 in
 slight
 abuse
 of
 the
 classical
 definition.
 The
 HPDI,
 in
turn,
 gives
 the
 narrowest
 possible
 set
 of
 bounds
 that
 contain
 95%
 of
the
 posterior
 distribution.
 While
 conceptually
 and
 computationally
different,
 it
 can
 be
 interpreted
 akin
 to
 a
 classical
 confidence
 interval
in
 the
 sense
 that
 there
 is
 a
 95%
 chance
 that
 the
 (presumable
 fixed
but
 unknown)
 ATT
 lies
 between
 these
 bounds.
 The
 last
 column
 in
the
 table
 gives
 percentage
 deviation
 of
 estimates
 flowing
 from
 the
flawed
 remote
 sensing
 approaches
 relative
 to
 those
 generated
 by
the
 benchmark.
As
 can
 be
 seen
 from
 the
 table,
 all
 three
 benchmark
 ATTs
 are
clearly
 negative,
 though
 the
 HPDI
 does
 include
 zero
 in
 the
 case
of
 Boulder,
 which
 has
 the
 smallest
 sample
 size.
 Generally,
 though,
this
 supports
 existing
 research
 that
 finds
 that
 MPB-induced
 tree
damage
 has
 a
 detrimental
 effect
 on
 home
 values
 (Price
 et
 al.,
2010;
 Cohen
 et
 al.,
 2016).
 For
 the
 sample
 at
 large,
 the
 average
loss
 is
 $10,471
 per
 home,
 which
 amounts
 to
 approximately
 2.5%
of
 the
 sample
 mean,
 counting
 all
 treated
 and
 all
 eligible
 controls.
In
 contrast,
 average
 losses
 in
 Golden
 amount
 to
 over
 $23,000,
 or
5.1%
 of
 the
 sample
 mean.
 Boulder
 posts
 the
 highest
 losses
 with
a
 posterior
 mean
 of
 over
 $124,000,
 or
 close
 to
 19%
 of
 the
 sample
mean.
Turning
 to
 the
 “PI,
 ABG”
 remote
 sensing
 version,
 the
 swapping
effect
 dominates
 any
 potential
 selection
 effects
 through
 omission
for
 Golden
 and
 Boulder,
 and
 thus
 introduces
 a
 pronounced
 positive
deviation
 from
 the
 benchmark
 for
 the
 posterior
 mean
 for
 both
 areas
(108%
 and
 35%,
 respectively).
 The
 opposite
 holds
 for
 the
 full
 sample,
where
 the
 “PI,
 ABG”
 approach
 produces
 a
 posterior
 mean
 that
 shows
a
 negative
 deviation
 of
 −79%
 from
 the
 benchmark.
 Thus,
 in
 this
 case


===== PAGE 8 (score 30, numeric-paren 12, table-words 1, var-words 10) =====

K.
 Moeltner
 et
 al.
 / Journal
 of
 Forest
 Economics
 26
 (2017)
 46–55
 
53
Table  4
Estimated
 losses
 in
 home
 values
 ($000’s).
Method
 
Mean
 
std
 
p
 > 0
 
Low
 
Up
 
Diff.
 bench.
 (%)
Full
 sample
PI,
 visual
−10.471
(2.926)
 
0.000
 
−16.517
 
−5.110
 
–
PI,
 ABG
 
−18.746
 
(3.192)
 
0.000
 
−24.828
 
−12.387
 
−79.03%
ADS,
 visual
 
−23.835
 
(2.299)
 
0.000
 
−28.312
 
−19.376
 
−127.63%
ADS,
 ABG
 
15.506
 
(55.595)
 
0.607
 
−91.265
 
123.971
 
248.09%
Golden
PI,
 visual
 
−23.097
 
(5.871)
 
0.000
 
−34.791
 
−11.562
 
–
PI,
 ABG
1.912
 
(6.252)
 
0.621
 
−10.563
13.898
 
108.28%
ADS,
 visual
−14.430
 
(5.130)
 
0.004
 
−24.869
 
−4.573
 
37.52%
ADS,
 ABG
 
−91.314
 
(112.915)
 
0.204
 
−311.240
 
136.497
 
−295.35%
Boulder
PI,
 visual
 
−124.338
 
(81.445)
 
0.058
 
−285.075
 
40.801
 
–
PI,
 ABG
 
−81.503
 
(40.797)
 
0.024
 
−164.931
 
−0.792
 
34.45%
ADS,
 visual
 
−173.582
 
(26.180)
 
0.000
 
−227.700
 
−124.026
 
−39.60%
ADS,
 ABG
 
−159.567
 
(22.131)
 
0.000
 
−201.598
 
−114.280
 
−28.33%
PI:
 damage
 assessment
 via
 expert
 photo-interpretation;
 ADS:
 damage
 assessment
 via
 Aerial
 Detection
 Survey;
 ABG:
 address-based
 geocoding
 using
 generic
 GIS
 software;
vis.:
 expert
 geocoding
 photo-imagery;
 p
 >
 0:
 proportion
 of
 posterior
 distribution
 greater
 than
 zero;
 low
 (up):
 lower
 (upper)
 bound
 of
 95%
 highest
 posterior
 density
 interval;
diff.
 bench:
 %
 difference
 relative
 to
 benchmark
 (PI,
 visual).
intensity
 effects
 via
 involuntary
 selection
 of
 retained
 treated
 appear
to
 outweigh
 swapping
 effects.
The
 “ADS,visual” approach
 leads
 to
 negative
 deviations
 from
 the
benchmark
 for
 the
 full
 sample
 (−128%)
 and
 Boulder
 (−40%),
 and
 a
positive
 deviation
 for
 Golden
 (38%).
 The
 “ADS,ABG”
 version,
 in
 turn,
yields
 a
 negative
 deviation
 for
 Boulder
 (−28%)
 and
 Golden
 (−295%,
although
 with
 20%
 of
 posterior
 draws
 in
 the
 positive
 domain),
 and
a
 positive
 deviation
 for
 the
 full
 sample
 (248%,
 though
 40%
 of
 the
posterior
 is
 located
 in
 the
 negative
 domain).
 Thus,
 while
 some
 of
these
 estimates
 exhibit
 less-than-desirable
 posterior
 precision,
 we
obtain
 a
 mixed
 picture
 for
 these
 ADS-based
 effects,
 with
 selection
effects
 potentially
 dominating
 over
 swapping
 effects.
Our
 benchmark
 results
 for
 MPB-induced
 losses
 in
 home
 values
are
 within
 the
 ballpark
 range
 of
 those
 found
 by
 Cohen
 et
 al.
 (2016)
for
 Boulder
 county
 and
 its
 northern
 neighbor,
 Larimer
 county.
Specifically,
 Cohen
 et
 al.
 (2016)
 estimate
 MPB-induced
 losses
 of
12.5%
 for
 the
 typical
 Boulder
 county
 home
 located
 near
 host
 trees
(damaged
 or
 not)
 compared
 to
 a
 home
 outside
 of
 the
 host
 tree
 zone
for
 the
 2011
 calendar
 year.
 In
 contrast,
 our
 analysis
 de
 facto
 com-
pares
 homes
 with
 damaged
 host
 trees
 to
 homes
 with
 healthy
 host
trees.
 To
 the
 extent
 that
 homes
 in
 a
 healthy
 forest
 surrounding
 are
typically
 more
 highly
 priced
 than
 homes
 outside
 the
 host
 tree
 zone
in
 that
 area,
 our
 larger
 loss
 estimate
 of
 19%
 for
 Boulder
 appears
 rea-
sonable.
 It
 also
 comes
 close
 to
 Cohen
 et
 al.
 (2016)’s
 estimate
 of
 a
22%
 erosion
 of
 home
 values
 in
 2011
 for
 the
 typical
 host-tree-zone
property
 in
 adjacent
 Larimer
 county.
Conclusion
We
 take
 a
 closer
 look
 at
 the
 possible
 estimation
 biases
 than
 can
occur
 in
 a
 property
 valuation
 study
 with
 focus
 on
 localized
 for-
est
 health
 due
 to
 imprecise,
 address-based
 geocoding
 and
 overly
coarse
 tree
 damage
 assessment
 via
 the
 Forest
 Service’s
 Aerial
Detection
 Survey.
For
 our
 sample
 of
 over
 900
 single
 family
 residences
 in
 the
Colorado
 front
 range,
 we
 find
 that
 both
 geocoding
 and
 damage
assessment
 errors
 can
 introduce
 substantial
 biases
 for
 residential
capitalization
 rates.
 In
 theory,
 either
 type
 and
 combination
 of
 spa-
tial
 mis-information
 can
 introduce
 selection
 bias
 by
 retaining
 only
a
 subset
 of
 treated
 that
 is
 no
 longer
 representative
 of
 the
 broader
population,
 as
 well
 as
 swapping
 bias
 by
 sorting
 homes
 into
 the
wrong
 bin
 in
 an
 econometric
 matching
 framework.
 In
 our
 appli-
cation,
 we
 find
 that
 ADS-based
 approaches
 to
 damage
 assessment
are
 susceptible
 to
 both
 types
 of
 errors,
 leading
 to
 substantial
 over-
or
 under-estimation
 of
 lost
 values.
While
 our
 empirical
 application
 is
 germane
 to
 forest
 health,
the
 importance
 of
 accurate
 geocoding
 likely
 carries
 over
 to
 other
property
 valuation
 contexts
 that
 rely
 on
 localized
 distances
 to
environmental
 (dis-)amenities
 or
 hot
 spots,
 such
 as
 polluted
 river
segments
 (Cho
 et
 al.,
 2011),
 hazardous
 waste
 sites
 (Guignet,
 2013;
Mastromonaco,
 2014),
 noise
 sources
 (Pope,
 2008),
 urban
 recreation
sites
 (Phaneuf
 et
 al.,
 2008),
 or
 air
 and
 water
 quality
 monitor-
ing
 stations
 (Leggett
 and
 Bockstael,
 2000;
 Poor
 et
 al.,
 2007;
 Neill
et
 al.,
 2007;
 Anselin
 and
 Lozano-Gracia,
 2008).
 It
 may
 be
 pru-
dent
 in
 such
 cases
 to
 obtain
 a
 small
 sample
 of
 expertly
 geocoded
properties
 and
 decide
 if
 ABG-related
 position
 errors
 are
 small
and/or
 random
 enough
 to
 be
 ignorable
 for
 a
 primary
 valuation
analysis.
A
 natural
 extension
 of
 our
 work
 would
 be
 a
 more
 careful,
imagery-based
 measurement
 of
 the
 intensity
 of
 infestation
 sur-
rounding
 a
 given
 property,
 in
 terms
 of
 the
 number
 or
 proportion
of
 diseased
 trees
 within
 a
 close
 buffer.
 This
 would
 allow
 for
 a
more
 rigorous
 examination
 and
 quantification
 of
 our
 presumed
selection
 effect
 towards
 more
 heavily
 impacted
 areas
 associated
with
 the
 ADS
 efforts.
 It
 would
 also
 pave
 ways
 towards
 investi-
gating
 if
 the
 relationship
 between
 infested
 trees
 and
 home
 values
may
 be
 nonlinear
 or
 exhibit
 threshold
 effects.
 Another
 interest-
ing
 extension
 might
 be
 the
 outsourcing
 of
 image
 interpretation
for
 dead
 trees
 to
 online
 labor
 markets,
 as
 recently
 implemented
by
 Yu
 et
 al.
 (2015)
 in
 the
 context
 of
 cloud
 detection
 in
 Landsat
scenes.
Acknowledgements
This
 work
 was
 supported
 by
 the
 United
 States
 Forest
 Service,
Southern
 Research
 Station
 [cooperative
 agreement
 number
 12-JV-
11330143-106].
Appendix
 A.
 Econometric
 details
The
 GenMatch
 balancing
 algorithm
 returns
 the
 original
 sample
of
 treated,
 i =
 1
 . .
 . nT, and
 a
 matched
 sample
 of
 unique
 controls,
j
 =
 1
 .
 . . nM, with
 possible
 replications,
 such
 that
 nM ≤
 nT. Each
matched
 control
 receives
 a
 weight
 Kj that
 indicates
 for
 how
 many
treated
 it
 serves
 as
 a
 match.
 For
 most
 matched
 observation,
 Kj will
be
 unity.


===== PAGE 9 (score 17, numeric-paren 0, table-words 3, var-words 8) =====

54
 
K.
 Moeltner
 et
 al.
 / Journal
 of
 Forest
 Economics
 26
 (2017)
 46–55
Next,
 we
 build
 the
 auxiliary
 regression
 model,
 using
 only
 the
matched
 controls.
 The
 models
 follows
 a
 basic
 weighted
 OLS
 speci-
fication,
 that
 is:
˜yj = ˜x′
jˇ
 +
 j, 
with
˜yj = 
Kjyj,
˜xj = 
Kjxj
and
j∼n
0,
 2
(A.1)
where
 yj is
 the
 observed
 sales
 price
 of
 home
 j,
 xj is
 a
 vector
 of
 home
characteristics,
 wildfire
 indicator,
 and
 quarter-of-sales
 indicators
as
 described
 in
 the
 main
 text.
 The
 error
 term
 j follows
 the
 usual
normal
 distribution
 with
 zero
 mean
 and
 common
 variance
 2.
For
 the
 entire
 sample
 of
 matched
 controls,
 the
 likelihood
 func-
tion
 is
 thus
 given
 as:
p
˜y|ˇ,
 2, ˜X
= (2)−nM /2
2−nM /2
exp

−1
22

˜y −˜Xˇ′ 
˜y −˜Xˇ
,
 
(A.2)
where ˜y
 denotes
 the
 entire
 vector
 of
 weighted
 home
 prices
 for
 the
matched
 controls,
 and ˜X is
 the
 weighted
 sample
 matrix
 of
 explana-
tory
 variables.
Next,
 we
 employ
 the
 standard
 normal
 prior
 for
 ˇ
 and
 inverse-
gamma
 prior
 for
 2, i.e.:
ˇ∼n (0,
 V0) ,
 
or
p
ˇ
= (2)−k/2|V0|−1/2 exp

−1
2

ˇ
 −
 0
′(V0)−1 
ˇ
 −
 0

2∼ig (0, 0) ,
 
or
p
2
=
0
0
 (0)

2−(0+1) exp

−0
2

, 
with
E 
2
=
0
0 −
 1 ,
 
V 
2
=
2
0
(0 −
 1)2 (0 −
 2)
,
(A.3)
where
 k
 is
 the
 dimension
 of
 ˇ. As
 is
 evident
 from
 (A.3)
 we  para-
meterize
 the
 inverse-gamma
 with
 shape
 0 and
 scale
 0, following
Gelman
 et
 al.
 (2004).
Combining
 the
 likelihood
 with
 these
 priors
 and
 dropping
 all
nuisance
 terms
 yields
 the
 posterior
 kernel
p
ˇ,
 2|˜y, ˜X
∝
2 −nM−20−2
2
exp

−1
2
 1
2

˜y
 −˜Xˇ′ 
˜y
 −˜Xˇ
+
ˇ
 −
 0
′(V0)−1 
ˇ
 −
 0

(A.4)
Following
 standard
 procedures,
 we
 break
 this
 posterior
 kernel
 into
conditional
 densities
 for
 ˇ
 and
 2,
 given
 as:
p
ˇ, |2,
 y, ˜X
∼n (1,
 V1) , 
with
V1 =

V−1
0
+ 1
2 ˜X
′ ˜X
−1
,
 
and
1 =
 V1

V−1
0 0 + 1
2 ˜X
′˜y

(A.5)
and
2|ˇ, ˜y, ˜X∼ig (1,
 1) ,
 
with
1 = 20 +
 n
2
,
 
and
1 =
 0 + 1
2

˜y
 −˜Xˇ′ 
˜y
 −˜Xˇ
(A.6)
Our
 Gibbs
 Sampler
 then
 draws
 sequentially
 and
 repeatedly
 from
these
 two
 conditionals.
 as
 discussed
 inter
 alia
 in
 Koop
 (2003)
 and
Table
 B.1
Balancing
 results
 for
 benchmark
 case.
Mean
 
max
Variable
 
Before
 
After
 
Before
 
After
Full
 sample
Age
 
6.07
 
0.94
 
59.00
 
5.00
sqft00
 
2.56
 
0.91
 
6.15
 
8.96
lot000
 
58.86
 
20.63
 
542.40
 
152.10
Golden
Age
 
5.79
 
1.68
 
47.00
 
5.00
sqft00
 
2.68
 
1.01
 
6.15
 
5.23
lot000
 
67.40
 
19.10
 
542.40
 
114.75
Boulder
Age
 
11.43
 
6.15
 
61.00
 
17.00
sqft00
 
6.60
 
0.90
 
36.70
 
2.38
lot000
 
120.32
 
54.12
 
825.90
 
387.68
mean
 (max)
 = mean
 (maximum)
 quantile
 deviation
 (5%
 steps);
 sqft00
 =
 square
footage
 in
 units
 of
 100;
 lot000
 = lot
 size
 in
 units
 of
 1000
 square
 feet.
Koop
 et
 al.
 (2007), after
 a
 sufficient
 number
 of
 iterations
 the
 sam-
pler
 will
 converge
 to
 the
 desired
 joint
 posterior
 given
 in
 (A.4).
As
 described
 in
 the
 main
 text,
 for
 each
 retained
 draw
 of
 ˇr from
the
 rth
 round
 of
 the
 sampler,
 we
 compute
 the
 predictive
 constructs
ˆ
r(˜xi)
 = ˜x′
iˇr and
 ˆ
r(˜xj)
 = ˜x′
jˇr. These
 then
 feed
 into
 the
 rth draw
of
 the
 ATT
 via
 Eqs.
 (2)
 and
 (3)
 of
 the
 main
 text.
Appendix
 B.
 Balancing
 results
To
 check
 on
 improvements
 in
 distributional
 similarity
 between
treated
 and
 controls
 before
 and
 after
 balancing
 the
 sample,
 we
compare
 the
 difference
 in
 5%-incremented
 quantiles
 between
treated
 and
 original
 and
 matched
 controls,
 respectively.
 The
 table
below
 shows
 these
 diagnostics
 for
 the
 benchmark
 remote
 sensing
approaches
 for
 each
 sample
 used
 in
 our
 main
 analysis,
 and
 for
 the
three
 continuous
 variables
 used
 in
 matching.
 Clearly,
 both
 mean
and
 maximum
 quantile
 deviations
 decrease
 markedly
 after
 balanc-
ing
 for
 all
 variables
 and
 applications.
The
 only
 exception
 is
 a
 slight
 increase
 in
 the
 maximum
 deviation
of
 square
 footage
 (from
 6.15
 to
 8.25)
 for
 the
 full
 sample.
 How-
ever,
 this
 comes
 at
 the
 gain
 of
 dramatic
 reductions
 in
 maximum
deviations
 for
 age
 and
 lot
 size.
Table
 B.1
References
Abadie,
 A.,
 Imbens,
 G.,
 2006.
 Large-sample
 properties
 of
 matching
 estimators
 for
average
 treatment
 effects.
 Econometrica
 74,
 235–267.
Abadie,
 A.,
 Imbens,
 G.,
 2011.
 Bias-corrected
 matching
 estimators
 for
 average
treatment
 effects.
 J.
 Bus.
 Econ.
 Stat.
 29,
 1–11.
Abadie,
 A.,
 Drucker,
 D.,
 Herr,
 J.,
 Imbens,
 G.,
 2002
 October.
 Simple
 and
bias-corrected
 matching
 estimators
 for
 average
 treatment
 effects,
 nBER
Technical
 Working
 paper
 283.
Abbott,
 J.,
 Klaiber,
 H.,
 2013.
 The
 value
 of
 water
 as
 an
 urban
 club
 good:
 a
 matching
approach
 to
 community-provided
 lakes.
 J.
 Environ.
 Econ.
 Manage.
 65,
 208–224.
Anselin,
 L.,
 Lozano-Gracia,
 N.,
 2008.
 Errors
 in
 variables
 and
 spatial
 effects
 in
hedonic
 house
 price
 models
 of
 ambient
 air
 quality.
 Empir.
 Econ.
 34,
 5–34.
Backsen,
 J.,
 Howell,
 B.,
 2013.
 Comparing
 aerial
 detection
 and
 photo
 interpretation
for  conducting
 forest
 health
 surveys.
 West.
 J.
 Appl.
 For.
 28,
 3–8.
Bentz,
 B.,
 Jacques,
 R.,
 Fettig,
 C.,
 Hansen,
 E.,
 Hayes,
 J.,
 Hicke,
 J.,
 Kelsey,
 R.,
 Negrn,
 J.,
Seybold,
 S.,
 2010.
 Climate
 change
 and
 bark
 beetles
 of
 the
 Western
 United
States
 and
 Canada:
 direct
 and
 indirect
 effects.
 BioScience
 60,
 602–613.
Burra,
 T.,
 Jerrett,
 M.,
 B.R.T.,
 Anderson,
 M.,
 2002.
 Conceptual
 and
 practical
 issues
 in
the
 detection
 of
 local
 disease
 clusters:
 a
 study
 of
 mortality
 in
 Hamilton,
Ontario.
 Can.
 Geogr.
 46,
 160–171.
Carroll,
 A.,
 Taylor,
 S.,
 Régnière,
 J.,
 Safranyik,
 L.,
 2004.
 Effects
 of
 climate
 change
 on
range
 expansion
 by
 the
 mountain
 pine
 beetle
 in
 British
 Columbia.
 In:
 Shore,
 T.,
Brooks,
 J.,
 Stone,
 J.
 (Eds.),
 Mountain
 Pine
 Beetle
 Symposium:
 Challenges
 and
Solutions.
 October
 30–31,
 2003,
 Kelowna,
 British
 Columbia,
 Natural
 Resources
Canada,
 Canadian
 Forest
 Service,
 Pacific
 Forestry
 Center,
 Information
 Report
BC-X-399,
 pp.
 223–232.
Cayo,
 M.,
 Talbot,
 T.,
 2003.
 Positional
 error
 in
 automated
 geocoding
 of
 residential
addresses.
 Int.
 J. Health
 Geogr.
 2,
 1–12.