Transcript The Indomitable in Pursuit of the Inexplicable

Economic Development Experts
versus
Economics:
the example of industrial policy
World Bank
Monday September 14, 2009
William Easterly (NYU and NBER)
Outline: 2 Negatives on
Development Experts & 1 Positive
on Economics
• Negative: The failure of the empirical
growth literature
• Negative: How Experts mistake
randomness for evidence (with example of
industrial policy)
• Positive: How Economics is useful after
all.
Reliance on Development Experts
was Legacy of the Great
Depression
• “the Depression led to conclusion: economic
development was not spontaneous, as in the classical
capitalist pattern, but was consciously achieved through
state planning.” (UN History of Development Thinking)
• Gunnar Myrdal (1956): “Super-planning HAS to be
staged by underdeveloped countries with weak
administrative apparatus … the alternative to making the
heroic attempt is acquiescence in economic stagnation
… which is politically impossible …”
• Strong political motive (as opposed to academic
breakthrough) to create a “Development Expert
Economics” to achieve faster growth than capitalist rich
countries had achieved with just “Economics”
Index of per capita income in developing and developed (based on unweighted average per
capita growth)
400
One Big (Only
Half-Serious)
Not getting
development
expert
advice
developing
developed
Stylized Fact
300
250
Getting
development
expert advice
200
150
20
04
20
02
20
00
19
98
19
96
19
94
19
92
19
90
19
88
19
86
19
84
19
82
19
80
19
78
19
76
19
74
19
72
19
70
19
68
19
66
19
64
19
62
100
19
60
Index of per capita income (1960=100)
350
Failure of empirical growth
literature to give expert guidance
• “experience … frustrated the expectations …{that} we had a
good fix on the policies that promote growth.” (Rodrik 2007)
• “It is hard to know how the economy will respond to a policy”
(World Bank Spence Growth Commission 2008)
• predicted in advance when Levine and Renelt (1992) failed to
find any robust determinants of growth
• 145 different variables “significant” in growth regressions
(with approx. 100 observations) --Durlauf, Johnson, and
Temple 2006
• Ciccone and Jarociński 2008: Bayesian model averaging
gave completely different “robust” variables from
Doppelhofer, Miller, and Sala-i-Martin 2004 for different
equally plausible samples
• Therefore, growth regression evidence on trade &
industrial policy is of little value either pro- or con- (e.g.
Rodriguez and Rodrik 2000)
Fail to spread high growth from
East Asia to other countries
• “At any time, some country is doing well, and …
observers … generalize from its choices and
recommend the same to all countries. After a decade or
two, this country ceases to do so well, some other
country using some other policies starts to do well, and
becomes the new star that all countries are supposed to
follow.” Avinash Dixit (2007)
• Development economists have advised “just be like
Korea” for decades, but we do not have any successes
at replicating Korea’s growth rates (not even in Korea,
where growth has now slowed!)
Mistaking randomness for evidence
• Panel regressions of annual growth rates for all
countries show that only 8% of the variance is
permanent cross-country differences, the other
92% is transitory (will disappear next year!)
deviations from world mean of about 1.8% per
capita
• Kahneman and Tversky’s Sarcastically-named
“Law of small numbers”
– Making too much of a small # of episodes with
too few years to give us “secrets to success”,
not sufficiently appreciating that outcomes of a
small sample will have a huge random
component (as we see growth rates do)
.15
Transitory growth: there is very
strong mean reversion in growth
(including East Asia!)
Original
source:
Easterly,
Kremer,
Pritchett,
Summers
1993
LVA
.1
EST
AGO
.05
CMR
ALB
TTO MMR
NIC
SLE RWA TCD
SVK
HUN
ZMBROM
DZA
HTI
MOZ
JOR IRN
COG
SUR
WSM
ARE
GRC
BGR
ETH
MEX
ISL
TUN
SEN
BEN
FIN
MNG DOM
ZAFPER
PAN
NAM
MLI
TGO
MDG
MAR
SDN
GMB
CIV
SAUBFA
VNM
MWI
BGD
NER
SWE
CAN
NZL
COM
CPV
PHL
ECU
CAFGAB
LAO INDIRLBWA
EGY
NGA
GHA
BHR
AUS
LKA
USA
HND
CHE
BOL
UGA
TUR
KEN
GBR
DNK
BDI
MRT
NOR
BRA
FRA
ESP MAC BTN
TON
AUT
KIR
BEL
NLD
ARG
GTM
CRI
VCT LUX
NPL
PAK
VENSYR
DEU
GUY
LSO
ITA
VUT
ISR
JPN PRT MUS
PRY SLV
COL
SYC
URY
BLZ
HKG
CHL
MYS
SGP
PNGSWZ
MLT
ZWE
JAM
GNB
IDN
0
ZAR
-.05
Change
in
growth
from
198695 to
19962005
MDA
MHL
SLB
-.1
-.05
0
g8695
dg9605
CHN
KOR
THA
.05
Fitted values
Previous growth rate (1986-95)
.1
Seeing patterns in randomness:
episode analysis
• Episode analysis: select maximum growth period in each
country, timing flexible, biases episodes to have large
transitory element (analogous to a streak of heads if you flip
a coin long enough, & then study “heads episode”)
• Example: Spence Commission criteria for success stories:
Monte Carlo simulations suggest 37 percent of top
successes using their selection criteria would NOT have the
top long-term growth rates.
• Selecting maximum decade growth on average out of 45
years will discover a growth experience 2.5 pp above the LT
average (Monte Carlo simulations).
• Spence Commission attributes high growth episodes to
whatever “leaders” (advised by experts) happen to be in
power during the episode (sounds plausible but nonfalsifiable)
• Therefore, episode analysis is not reliable guide to whether
industrial policy “works.”
Example of Ha Joon Chang case
for industrial policy
• With a lot of random variation, easy to find
examples to confirm your bias.
• Chang cites high-growth Korea and Taiwan as
evidence for industrial policy, but says free-trade
high-growth Singapore and Hong Kong were
“exceptions,” and he never mentions highgrowth free trade Botswana.
• Small Numbers Problems: Mexico has had only
1.8 percent per capita growth from 1994 to 2002
after NAFTA, so Chang concludes NAFTA isn’t
working.
Another heuristic bias --Attributing
intentional skill to random outcomes
• An experiment in which subjects observed two people
executing a task, rigged so that the two persons’
performance was equal.
• The subjects told that one would receive a large payment
and choice which one would be random.
• The subjects then asked to describe the performance of
the two agents.
• Despite subjects’ knowledge that payment was random,
gave superior marks on performance attributes to the
agent who received the payment.
• A lot of industrial policy “case studies” of success are like
this – we give too much credit to those geniuses in
Korea and Taiwan for industrial policy, or to Spence’s
“leaders” for high growth episodes.
Yet another randomness fallacy:
Confusing conditional probabilities
• Kahneman et al. experiments show we confuse
Prob(X|Y) with Prob(Y|X)
• (A) Probability(If you win big in Vegas|You bet a
large sum at long odds) is high.
• (B) While Probability(If you bet a large sum at
long odds|You win big in Vegas) is low.
• You get in trouble if you decide whether to make
such a bet based on A and not B!
• do we really ever make this obvious mistake?
Example: Dani Rodrik on industrial
policy
• Dani Rodrik:“the countries that have
produced steady, long-term growth during
the last six decades are those that
…promoted… diversification into
manufactured … goods”
• So Dani concludes that developing
countries will have to get busy with “real
industrial policies.”
Yes, we do reverse conditional
probabilities
• Dani is calculating (A) probability(If successful/Then
have industrial policy)
• But this is wrong probability, just like in Vegas!
• We want to know (B) Probability (If have industrial
policy/Then successful)
• There are many examples of failed industrial policy
around Africa ($6 billion Nigerian Ajaokuta Steel Mill that
never produced steel?), Middle East, Former Soviet
Union, and Latin America, so probability (B), the right
probability, seems too low to get enthusiastic about
industrial policy.
• So “success story” evidence for industrial policy is just
reflecting cognitive biases on how we mishandle random
variables and probabilities.
Alternative explanations of data
• (A) Industrial policy worked in Korea and Taiwan,
but not elsewhere, thanks to some unknown
factor.
• (B) Industrial policy didn’t even work in Korea
and Taiwan, which succeeded due to other
reasons.
• (C) Industrial policy worked in other places too,
but other factors made for poor growth
outcomes.
• We don’t have a reliable aggregate empirical
methodology (growth regressions or case
studies) to distinguish (A), (B), or (C)
Why economists can say
something useful after all
Optimal response to “law of small
numbers”
• Get more numbers!
• Empirical literature has shifted towards explaining income
levels rather than growth rates
• Log of per capita income is the sum of all previous percent
growth rates – that takes us away from misleading “law of
small numbers” to the true power of “law of large
numbers.”
• Long-term evidence provides little support for industrial
policy, lots of support for Econ 101 principles:
Entrepreneurship in Markets, Division of Labor, Gains
From Specialization, Comparative Advantage, Gains from
Trade; No such LT evidence for industrial policy.
• Moreover, we are not starting from scratch as economists,
the ideas that made it into Econ 101 are those that have
survived the test of time by many previous generations of
economists testing these ideas.
Long-run evidence on trade and
per capita income
• Levels studies summarized by Harrison and RodriguezClare (2009) suggestive evidence that trade causes
prosperity.
• Stylized Facts: Over the last two centuries, divergence
between (1) Europe and North America (with a lot of
trade) and (2) rest of world (with a lot less trade because
of poor infrastructure and geographic distance).
• Common sense: a small, poor economy can’t make most
goods for itself, it desperately needs trade to get access
to valuable goods (Computers? Cars? Antibiotics?).
Industrial policy is second-order compared to this.
Applying common sense of
comparative advantage to industrial
policy
• Corrupt Indian civil servants give drivers’
licenses to people who can’t drive (Bertrand,
Djankov, Hanna, Mullanaithan 2008) and we
expect them to do industrial policy?
• M.A. Thomas (2009) & Pritchett (2009) on
limited capacity of poor governments.
• Comparative advantage depends on
government too.
• A corrupt, low-skilled, poorly-funded government
does not have a comparative advantage in
finding the country’s comparative advantage.
The Leroy Smith Principle: Success is a
Surprise
• Who predicted cut flowers in Kenya (40% of European
market), or women’s cotton suits in Fiji (42% of the US
market), or bathroom ceramics in Egypt? (30 percent of
manf exports,93% goes to Italy).
• Countries specialize to a remarkable degree by both
product (out of 2985 6-digit manf. possibilities) and
destination (217 possibilities). Out of 647,745 manf.
possibilities, Top 1% of nonzero entries account for 52% of
manufacturing exports (Easterly, Reshef, and
Schwenkenberg 2009)
• Who will do better finding Big Hits: public officials with
limited capacity & information and ambiguous
incentives, or decentralized search by entrepreneurs
with specialized skills, strong incentives and much
more information?
Friedrich Hayek on why the Leroy
Smith principle is itself an argument
for Econ 101
• “It is because every individual knows so
little and… because we rarely know which
of us knows best that we trust the
independent and competitive efforts of
many to induce the emergence of what we
shall want when we see it."
Summary
• Empirical growth literature has failed to produce
useful expert knowledge
• Mistaking randomness for evidence led to the
wrong approach to development – “industrial
policy” which requires “expert knowledge,” which
has little or no evidence base.
• But economists can say something useful about
“big picture” development after all, using
– (1) long run evidence,
– (2) common sense economics that has stood the test
of time,
– (3) following Principles of Economics that produce
prosperity even when the “development experts” can’t
produce exact answers.