Evaluating DeforestationCarbon Impacts of Protected Areas: challenge, approach, Costa Rican case

1
Evaluating Deforestation/CarbonImpacts of
Protected Areaschallenge, approach, Costa
Rican case two applications to the Brazilian
Amazon
Alexander Pfaff Duke University prepared
for WWF / Moore / Linden workshop Palo Alto,
California February 11, 2009
2
Acknowledgements ( policy/land-use
research)Costa Rica Team (NSF-MMIA, NCEAS,
Tinker, SSHRC, IAI)Suzi Kerr, Arturo Sanchez,
David Schimel, Shuguang Liu, Boone Kauffman,
Flint Hughes, Vicente Watson, Joseph Tosi, Juan
Robalino, F. Alpizar, C. Leon, C.M.
RodriguezInterOceanic Team (funding Duke
University Nicholas Institute)Cesar Delgado
(lead), Dalia Amor, Joseph Sexton, Fernando
Colchero, with assistance from Juan Robalino,
Diego HerreraMexico Team (funding The Tinker
Foundation, IAI, RFF)Allen Blackman, Yatziri
Zepeda, Juan Robalino, Laura Villalobos
3
Acknowledgements ( roads/land-use research)
Brazil Team (funding NASA LBA (II/III), Tinker
Foundation, IAI) Eustaquio Reis, Claudio Bohrer,
Robert Walker, Steve Perz, Juan Robalino, James
Gibbs, Robert Ewers, Bill Laurance, Steven
Aldrich, Eugenio Arima, Marcellus Caldas,
othersMayan Team (funding Mesoamerican
Biological Corridor, Mexico Unidos para la
Conservacion, CONABIO, Conservation
International, Conservation Strategy Found
(CSF))Dalia Amor (lead), Fernando Colchero,
Norman Christensen, with data help from the
Mexican Ministry of Transportation, Victor Hugo
Ramos (WCS), UNAM, Jaguar Conservancy
4
POINT forest/carbon policy design matters !!

• Costa Rica deserves gratitude as very public
  pioneer
• the lessons learned from Costa Rica are
  misleading
• protection ecopayments CAN have large impacts
  that does NOT mean ANY policy has a big impact
• policies details really can matter they can
  change dismal evaluations of the past -gt
  positive advice !!

5
IMPACT non-random location confounds

• much evaluation of policy ex-post, little about
  where
• where is a choice that responds to factors we
  observe those factors in turn affect the
  deforestation outcome
• examine protection empirically ( payments
  roads)
• 1) LOCATIONS NON-RANDOM WORSE LAND
• 2) CORRECTED IMPACT IS CLEARLY LOWER
• 3) BUTTARGETING CAN INCREASE IMPACT

6
Reserves optimal location(s) literature(s)
e.g. spatial/ecological GRUAS rationale
7
GRUAS plus Deforestation Pressure Index
(from Pfaff Sanchez 2004)
8

• EXPECT Non-Random Reserve Location
• an agency might target the threatened forest if
  the goal is the largest change relative to the
  baseline
• an agency might target the nonthreatened forest
  if goal is a long-lasting or uncontentious/cheap
  park
• an agency might target forest with high benefits,
  hoping for but ignoring threat (correlation /-
  ?)
• we would not expect a random park distribution
  can significantly bias impact estimates (flip
  sign)

9
Matching to Address Non-randomness

• compare treated to similar subset of the
  untreated
• define similar by plot characteristics PSM, CM
  (Rosenbaum Rubin 1983 and Abadie Imbens 2006)
• choose a rule for how to select matched untreated
• choose how to compare the treated with matched
• standard errors from bias-correction regressions
  require correction, e.g. for re-use of any
  control

10
Land-Use Analysis In Background

• examine deforestation over time (e.g. for C
  baseline) 1963 aerial photos 1979, 1986, 1997,
  2000, 2005 satellites
• initially used district and sub-district
  observations still some, but focusing on more
  recent pixel data
• biophysical proxies yield expected results as do
  the socioeconomic covariates rain, temperature,
  slope, soil, distances to markets and national /
  local roads all of these observable factors
  useful for matching

11
Matching (PSM n4) protection 86-97 defor.

• Dist. San Jose Treated 101 All
  Untreated 90 Matched Untreated 98
• Dist. Natl Road Treated 15 All
  Untreated 6 Matched Untreated 13
• Dist. Local Road Treated 11 All
  Untreated 5 Matched Untreated 10
• Dist. Wide River Treated 3.9 All
  Untreated 3.2 Matched Untreated 4.0
• Dist. Cleared Area Treated 3.2 All
  Untreated 0.7 Matched Untreated 2.8
• Slope Treated 14
  All Untreated 9 Matched Untreated 15
• Altitude Treated 1.3
  All Untreated 0.6 Matched Untreated 1.4
• Rain Treated 4.0
  All Untreated 3.8 Matched Untreated 4.1

12
Treated and Untreated In All Propensity Bins
Number of Observations
Propensity Score Likelihood of treatment
13
Though Poorer Matches for Higher Scores
14
Park Impacts within boundaries PSM CM
Park Effects on 86-97 Deforestation, n 4 in
each method (Andam et al. (PNAS) examines 1963
forward as per trends, estimated impact is
higher but matching reduces similarly)
Adj. Diff. in Means
Difference in Means
Strategy
-1.99
-9.38
Using All of the Untreated (Naive)
-1.37
-0.05
Propensity Score Matching (PSM)
-0.85
-2.19
Covariate Matching (CM)
PSM vs. CM
3.5
Treated Observations with the same match
33.4
Similarity between Control Groups
15
Matching applied to Brazilian Amazon Protected
Areas (Federal, State, Indigenous 2000-04
deforestation)
16
The Chico Mendes Extractive Reserve
17
2007 Evidence of External Pressures On Chico
Mendes Extractive Reserve
Chico Mendes Extractive Reserve
18
Matching Applied To InterOceanic Highway Region
Protected Areas (Chico Mendes Extractive Reserve
Acre / InterO Protected Areas) (impacts on 1989
2000 deforestation 2000 2007 deforestation)

19
Park Impacts within boundaries -- comparisons

• Targeting Variable distance to San Jose
• over 85km, essentially none
• under 85km, greater than average
• Targeting Variable distance to national roads
• over 7.53km, insignificant
• under 7.53km, greater than twice the average
• Targeting Variable slope
• over 7.12 degrees, insignificant
• under 7.12 degrees, greater than five times the
  average

20
blocking?
blocking?
blocking
nothing
blocking
leaking?
21
Matching BY ranges of start-of-period road
distances deforestation rates for treated (road
investments)and untreated/control (no new
roads) pixels
DEFORESTATION 76-87
Note , , and represent 10, 5 and 1
respectively.
22
Matching BY ranges of start-of-period road
distances deforestation rates for treated (road
investments)and untreated/control (no new
roads) pixels
DEFORESTATION 00-04
Note , , and represent 10, 5 and 1
respectively.
23
Brazilian Amazon -- Temporally Rich Data (below
for Amazonia, census tract or pixel data)

• Road Changes 1968 1975, 1975 1987, 1985 -
  1993
• from maps, so roads can be mapped to census
  tracts
• amazingly, separating Fed/State Paved/Unpaved
  (important to consider which types follow
  others..)
• Forest Changes 1976 - 1987, 1986 - 1992, 1992
  - 2000
• remotely sensed unlike census, can map to tracts
• blending pairs of Diagnostico and TRFIC/Prodes

24
Brazilian Amazon -- Spatially Rich Data
25
Non-Road Factors (Amazon Mayan) (units vary)

• Distances to large and medium and small cities
• evolution determined by Perz demographic
  projections (though small city group only
  shrinks, not adding new)
• clearing frontiers move away from big cities over
  time?
• Biophysical constraints on production (for us,
  fixed)
• - amount of rain, several categories of slope,
  soil fertility
• Prior Clearing represents all sorts of possible
  changes
• Census data (counties) changes in population
  output