Mining, deaths and dropouts International evidence on the long-run health and
education effects of mining
4 November 2013 Crawford Ph. D conference
Ryan Edwards Ph. D candidate Arndt-Corden Department of Economics Crawford School of Public Policy
Panel: Paul Burke (Chair), Robert Sparrow and Budy Resosudarmo
Question
What impact does mining have on health and education outcomes?
2
Scope of question
3
What impact does mining have on health and education outcomes?
Effect of X on Y
Holding all else constant
‘Average treatment effect’ (ATE/LATE)
The mining sector and mining growth
Relative to other sectors and non-mining growth
In the long run
Across / between countries
Answer
1. Mining explains substantial long-run differences in health & education outcomes between countries
4
Answer
1. Mining explains substantial long-run differences in health & education outcomes between countries
2. ‘No growth’ is better than ‘mining growth’
5
Answer
1. Mining explains substantial long-run differences in health & education outcomes between countries
2. ‘No growth’ is better than ‘mining growth’
3. In the long run, on average, doubling the mining share of the economy causes the:
- infant death rate to be 11 % higher
- secondary completion rate to be 23 % lower
- % of people with no education to be 75 % higher
6
Answer
1. Mining explains substantial long-run differences in health & education outcomes between countries
2. ‘No growth’ is better than ‘mining growth’
3. In the long run, on average, doubling the mining share of the economy causes the:
- infant death rate to be 11 % higher
- secondary completion rate to be 23 % lower
- % of people with no education to be 75 % higher
(Caveat: ATE and impact heterogeneity)
7
Contributions and motivations 1. Explicitly considers mining
– Mining has not explicitly been considered in any cross-country econometric research on the ‘resource curse’ or human capital
– Only channel for ‘point’ resources to have effects
8
Contributions and motivations 1. Explicitly considers mining
– Mining has not explicitly been considered in any cross-country econometric research on the ‘resource curse’ or human capital
– Only channel for ‘point’ resources to have effects
2. Isolates direct causal impacts of X on Y
– Purges country, time and omitted variable effects with panels
– Introduces new time-variant instruments to ‘resource curse’ lit.
9
Contributions and motivations 1. Explicitly considers mining
– Mining has not explicitly been considered in any cross-country econometric research on the ‘resource curse’ or human capital
– Only channel for ‘point’ resources to have effects
2. Isolates direct causal impacts of X on Y
– Purges country, time and omitted variable effects with panels
– Introduces new time-variant instruments to ‘resource curse’ lit.
3. Disaggregates social development / human capital
– Latest (and only causal) papers use composite indices
10
Contributions and motivations 1. Explicitly considers mining
– Mining has not explicitly been considered in any cross-country econometric research on the ‘resource curse’ or human capital
– Only channel for ‘point’ resources to have effects
2. Isolates direct causal impacts of X on Y
– Purges country, time and omitted variable effects with panels
– Introduces new time-variant instruments to ‘resource curse’ lit.
3. Disaggregates social development / human capital
– Latest (and only causal) papers use composite indices
4. Contributes to the international evidence on the determinants of long-run prosperity and development
– Human development is a long-run phenomena
– More external validity and useful to identify trends
– Useful starting point to develop theory and mechanisms for micro studies
11
Why would mining…
12
Why would mining have anything to do with health and education?
13
?
?
?
Many potential channels: net impact ambiguous
• Positive channels – Income / wealth effects (e.g. van der Ploeg, 2011)
– Endogeneity of human capital (e.g. Easterly, 2001)
– Strengthened fiscal position (e.g. Emerson, 1982; Arezi et al, 2011)
– Spill-overs (Kaplinsky, 2011) and private local investment (MCA, 2012)
14
Many potential channels: net impact ambiguous
• Positive channels – Income / wealth effects (e.g. van der Ploeg, 2011)
– Endogeneity of human capital (e.g. Easterly, 2001)
– Strengthened fiscal position (e.g. Emerson, 1982; Arezi et al, 2011)
– Spill-overs (Kaplinsky, 2011) and private local investment (MCA, 2012)
• Negative channels – Low returns to skills, education and knowledge (e.g. Gylfason, 2001)
– ‘Dutch Disease’ (e.g. Corden & Neary, 1982; Sachs & Warner, 2001)
– Endogenous institutions and conditionality (Mehlum et al, 2006)
– Uncertainty and volatility (van der Ploeg, 2011; Carmiganani, 2013)
15
Many potential channels: net impact ambiguous
• Positive channels – Income / wealth effects (e.g. van der Ploeg, 2011)
– Endogeneity of human capital (e.g. Easterly, 2001)
– Strengthened fiscal position (e.g. Emerson, 1982; Arezi et al, 2011)
– Spill-overs (Kaplinsky, 2011) and private local investment (MCA, 2012)
• Negative channels – Low returns to skills, education and knowledge (e.g. Gylfason, 2001)
– ‘Dutch Disease’ (e.g. Corden & Neary, 1982; Sachs & Warner, 2001)
– Endogenous institutions and conditionality (Mehlum et al, 2006)
– Uncertainty and volatility (van der Ploeg, 2011; Carmiganani, 2013)
• Today, I am only interested in casual effects of X on Y – ‘Black box’ impact evaluation approach (ATE / LATE only)
– Mechanisms, ‘why’ and impact hetero. are 2ndary questions, for future research
16
17
Mining and infant deaths
AFG
AGO
ALB
ARE
ARG
ARM
ATG
AUSAUT
AZE
BDI
BEL
BENBFA
BGD
BGR
BHR
BIH
BLZ
BOL
BRA
BRB
BRN
BTN
BWA
CAF
CANCHE
CHL
CIV CMR
COL
CPV
CRI
CUB
CYPCZE
DEU
DJI
DMA
DNK
DOM
DZA
ECU
ERI
ESPEST
ETH
FIN
FJI
FRA
GAB
GBR
GEO
GHA
GINGNB
GNQ
GRC
GRD
GTMGUY
HND
HRV
HTI
HUN
IDN
IND
IRL
IRQ
ISL
ITA
JAMJOR
JPN
KAZ
KENKHM
KWT
LIE
LKA
LSO
LTU
LUX
LVA
MAR
MDG
MDV
MHL
MLI
MLT
MMR
MNE
MNG
MOZ
MRT
MUS
MWI
MYS
NAM
NERNGA
NIC
NLD
NOR
NPL
NZL
OMN
PAK
PANPER
PHL
PLW
PNG
POL
PRT
PRY
QAT
ROM
RUS
RWA
SAU
SDNSEN
SGP
SLB
SLE
SLV
SMR
SOM
SRB
STP
SUR
SVN
SWE
SWZ
SYR
TCD
TGO
THA
TMP
TON
TTO
TUNTUR
TUV
UGA
UKRURY
USA
VCT
ZAF
ZMB
ZWE
12
34
5
Log in
fant
mort
alit
y ra
te (
per
'000
death
s)
-10 -8 -6 -4 -2 0Log mining share of value added
limr Fitted values
Mining and no schooling
18
AFG
ALB
ARE
ARG
ARMAUS
AUT
BDI
BEL
BENBGD
BGR
BHRBLZ BOLBRA
BRB
BRNBWA
CAF
CAN
CHE
CHL
CMR
COL
CRICUB
CYP
CZE
DEU
DNK
DZAECU
ESP
EST
FIN
FJI
FRA
GAB
GBR
GHA
GRC
GTM
GUY
HND
HRV
HTI
HUN
IDN
IND
IRL
IRQ
ISL
ITA JAM
JOR
JPN
KAZ
KENKHM
KWTLKA
LSO
LTU
LUX
LVA
MAR
MDV
MLI
MLT
MMR
MNG
MOZ
MRT
MUS
MWI
MYS
NAM
NER
NIC
NLD
NOR
NPL
NZL
PAK
PANPER
PHL
PNG
POL
PRT
PRY
QAT
ROM
RUS
RWA
SAU
SDN
SEN
SGP
SLE
SLV
SRB
SVN
SWE
SWZ SYR
TGO
THA
TONTTO
TUN
TUR
UGA
UKRURY
ZAFZMB
ZWE
-50
5
Log p
erc
enta
ge o
f popuatio
n w
ith n
o s
choolin
g
-10 -8 -6 -4 -2 0Log mining share of value added
llu Fitted values
Basic equation
ln Yc = α + β ln mining c+ γ ln GDPc + δ Xc+ ε
19
Infant / U5 mortality
Years education
No education
Primary completion
Secondary completion
Mining value-
added level per capita
Mining share of
GDP
Non-mining value-
added level per capita
GDP per capita
Error
Controls
e.g. latitude
Institutions Fixed effects
Basic equation
ln Yc = α + β ln mining c+ γ ln GDPc + δ Xc+ ε
20
Infant / U5 mortality
Years education
No education
Primary completion
Secondary completion
Mining value-
added level per capita
Mining share of
GDP
Non-mining value-
added level per capita
GDP per capita
Error
Controls
e.g. latitude
Institutions Fixed effects
log β coefficient interpretation: • Long-run health and elasticity of mining in the economy,
holding all else constant; and • Effectively a long-run equilibrium ‘impact estimate’.
Data
• Sample – 99-151 countries, 1970 – 2008, up to 6630 observations
• Mining value-added – United Nations Environmental Accounts, 1992 – 2008
– Plus: mining + utilities: United Nations National Accounts, 1970 - 2010
• Health and education indicators – World Development Indicators (WDI) and Barro and Lee (2010)
• GDP, institutions, latitude and other controls – Penn World Tables (2013)
– Sala-i-Martin et al (2004), Brunschweiller and Bulte (2008)
– World Development Indicators, World Governance Indicators, Resource Governance Index, Corruption Perceptions Index
21
Addressing potential endogeneity
ln Yc,t = α + β ln mining c,t+ … . + εc,t
22
Addressing potential endogeneity
ln Yc,t = α + β ln mining c,t+ … . + εc,t
23
Why might endogeneity be a threat? (Estimates will be biased and inconsistent with
OLS etcetera under endogeneity)
1. Mining value-added is determined by
initial capabilities. E.g. exploration abilities, ability in other sectors
2. Mining value-added is a product of
domestic decisions. E.g. industrial policy, trade policy, firms and individuals
Addressing potential endogeneity
ln Yc,t = α + β ln mining c,t+ … . + εc,t
24
Why might endogeneity be a threat? (Estimates will be biased and inconsistent with
OLS etcetera under endogeneity)
1. Mining value-added is determined by
initial capabilities. E.g. exploration abilities, ability in other sectors
2. Mining value-added is a product of
domestic decisions. E.g. industrial policy, trade policy, firms and individuals
Need to instrument mining (IV estimates will be biased but consistent)
1. Initial country reserves / SS assets - Time invariant - Source: Norman, 2009; World Bank
2. Time variant country reserves - Only for oil and gas - Less country coverage - Source: US Energy Information
Administration, 2013
3. Commodity prices - Time and country variant country
weighted commodity price index - Source: Burke and Leigh, 2010
Empirical approach: Cross section first
Cross-section estimators:
1. Ordinary least squares (OLS) (inconsistent and biased)
2. Instrumental variable estimator (IVE) (consistent; opp. bias) – Instrumented with initial reserves; all use ‘ivreg2’
3. Panel ‘between’ IVE – Instrumented with initial reserves
– Confirms OLS/IVE results over many years by averaging out
• A cross section exploiting between-country variation has a natural long run interpretation – Primary result: long run elasticities and marginal effects
• Omitted, unobserved and other country-specific factors?
25
Estimators:
1. Fixed effects (LSDV/within) (inconsistent and inefficient)
2. Fixed effects (FE) IV (consistent, but both inefficient) – Instrumented by the commodity price index OR reserves
3. FE generalised method of moments (GMM) (efficient) – Instrumented by the commodity price index and both reserves
4. System-GMM (efficient) – Instrumented by system of lags and lagged differences, using ‘xtabond’
• Controls for country/time-specific factors
• Minimises omitted / unobserved variable biases
• Efficient estimation under endogeneity
Empirical approach: extend to panel
26
Rationale behind this dual approach
• Identification rests on holding all everything except for mining constant and finding valid instruments to remove endogeneity
• Many cross-country results do not ‘survive’ fixed effects
• Results consistent in magnitude and significance between cross-section and panel data imply that: – Time-specific effects are not an issue;
– Country specific effects are not an issue;
– Omitted variables are not an issue;
– Unobserved variables are not an issue; and
Primary parsimonious cross-section IV specification is sufficient to obtain consistent and unbiased parameter estimates (simpler, better)
27
Recap: Identification strategy and managing conceivable threats
28
Threats Solutions
Sample is not random. Selection bias is unavoidable
Use IVEs and control for as much as possible
Time-specific effects (‘year selection bias’)
Use between and panel estimators
Small sample bias and IV inconsistency
Use a large sample of countries (excluded none). Panel approach. Rich and non-resource countries important for control group.
Endogeneity of mining IVs and reduced form / conditional equations. Local average
treatment effect
Weak instruments Strong by Stock-Yogo critical values and over-identification tests.
Exclusion restriction Commodity prices exogenous. Commodities pass through mining
to development. A-R and overid. tests.
Omitted and unobserved variables Country and time fixed effects and system-GMM
Mining and infant deaths
29
Dependent variable Log infant mortality (deaths per '000)
Equation (1) (2) (3) (4) (5) (6)
Sample 2005 2005 Panel 2005 2005 Panel
Estimator OLS IV IV Between OLS IV IV Between
Log per capita mining level
value-added
0.07*** 0.08*** 0.11*
(0.02) (0.02) (0.06)
Log mining share of value
added
0.12*** 0.11*** 0.11***
(0.02) (0.03) (0.03)
Log per capita non-mining
level value-added
-0.61*** -0.64*** -0.67***
(0.04) (0.06) (0.10)
Log real GDP per capita -0.56*** -0.56*** -0.55***
(0.05) (0.05) (0.05)
Excluded F statistic
21.99*** 51.82***
Results are similar if I use
• Any of the estimators discussed so far – OLS, IVE, panel BE and IVE BE; panel FE, FE IVE, and SGMM
30
Results are similar if I use
• Any of the estimators discussed so far – OLS, IVE, panel BE and IVE BE; panel FE, FE IVE, and SGMM
• Different variables – Resources: rents share of GDP and point resource exports (WDI)
– Dependent variables: U5MR, life expectancy, WDI education data
– Instruments: disaggregated Norman, WB sub-soil assets and natural capital, interaction instruments (e.g. index * Norman or reserves)
31
Results are similar if I use
• Any of the estimators discussed so far – OLS, IVE, panel BE and IVE BE; panel FE, FE IVE, and SGMM
• Different variables – Resources: rents share of GDP and point resource exports (WDI)
– Dependent variables: U5MR, life expectancy, WDI education data
– Instruments: disaggregated Norman, WB sub-soil assets and natural capital, interaction instruments (e.g. index * Norman or reserves)
• Different functional forms – 1st differences and growth rates over 20 years, cross-section
– 1st differences and growth rates over 12 / 20 years, t = 4; 2 panel FE
– level-level, level-log, log-level
– Different level panels intervals (e.g. 5 and 10 year intervals
32
Does this result fit the ‘real world’? Example: Health in Papua New Guinea
• Compare PNG’s IMR reduction to other countries
• Mining rose from 19% (1994) to 30% (2008) of GDP
33
Does this result fit the ‘real world’? Example: Health in Papua New Guinea
• Compare PNG’s IMR reduction to other countries
• Mining rose from 19% (1994) to 30% (2008) of GDP
• Prediction: This ~50% increase should correspond to around a 5.5 per cent increase in IMR, holding all else constant, or an 0.4% increase p.a. (Hint: all else was not constant!)
34
Does this result fit the ‘real world’? Example: Health in Papua New Guinea
• Compare PNG’s IMR reduction to other countries
• Mining rose from 19% (1994) to 30% (2008) of GDP
• Prediction: This ~50% increase should correspond to around a 5.5 per cent increase in IMR, holding all else constant, or an 0.4% increase p.a. (Hint: all else was not constant!)
• Actual? – East Asia and Pacific average IMR decrease: - 4% p.a.
– World average IMR decrease: - 1.6% p.a.
– PNG IMR decrease: - 1% p.a
– Difference with world: + 0.6
– Estimates seems highly plausible, in this case
35
Issues to be resolved and ongoing work
• Impact heterogeneity – ATE overestimates impacts for resource rich advanced countries
(e.g. Australia and Norway)
– ATE underestimates impacts for developing countries (e.g. Papua New Guinea and Nigeria)
– Global effect (ATE) is robust
• controlling for Africa, regions, country FE and institutions
• using interaction terms
• Possible application to human capital more broadly (e.g. TFP, R&D)
• Further work needed to explain divergent experiences within a cross country framework
36
Australian GDP growth vs. MFP (Source: Karunaratne, 2010)
37
Issues to be resolved and ongoing work
• Mechanisms – First examination used same model as discussed, for simplicity and to
avoid identification issues
– Investment in health (consistently negative)
– Investment in education (insignificant or positive)
– Institutions (consistently negative)
– Gini coefficient (insignificant)
– Growth in other sectors, productivity, others (?)
• Within country spatial and dynamic analysis – Household surveys, administrative data and qualitative mining data
– Papua New Guinea, Indonesia, and Australia
38
Final remarks
• Absence of evidence of non-monetary development impacts to date is absolutely telling
– Income and wealth effects from mining growth on human capital are, on average, non-existent
– Human capital effects undermine long-run growth and development prospects net of this (2/3 of HDI, most of MPI)
39
Final remarks
• Absence of evidence of non-monetary development impacts to date is absolutely telling
– Income and wealth effects from mining growth on human capital are, on average, non-existent
– Human capital effects undermine long-run growth and development prospects net of this (2/3 of HDI, most of MPI)
• May need to rethink any mining-based human development strategy
– Specific policy recommendations require further understanding of SR / LR mechanisms and within-country dynamics - each is likely to be different
40
In the meantime..
• Standard prescriptions (often not followed!) remain a good starting point to deal with these issues.
41
In the meantime..
• Standard prescriptions (often not followed!) remain a good starting point to deal with these issues. E.g.
– Stably invest resource revenues in public HK (WB/IMF)
– Encourage micro-diversity and value-adding (UNIDO, 2011)
– Strengthen institutions / minimise rent-seeking opportunities (Extractive Industries Transparency Initiative, Publish What you Pay, Natural Resource Charter)
– Smooth macroeconomic volatility (van Der Ploeg, 2011)
– Ensure a broad tax and transfer system (IMF)
– Avoid high levels of inequality (Carmignani, 2013)
42
Any questions?
• Thank you for your attention
• Comments are most welcome (it’s a work in progress)
• My contact details are:
43