"Nightlights and Economic Activity in Military Combat Zones of Former USSR Countries".
Bachelor's Thesis Presentation 2019.
Grade: 10/10.
Presented by Rufat Mustafaev (https://www.linkedin.com/in/rufat-mustafaev/).
Supervised by Konstantin Kholodilin (https://www.hse.ru/en/org/persons/26723814).
Higher School of Economics (HSE), St Petersburg, Russian Federation, Economics Faculty.
Estimating Economic Loss from War Using Satellite Images
1. NIGHTLIGHTS
AND ECONOMIC ACTIVITY
IN MILITARY COMBAT ZONES
OF FORMER USSR COUNTRIES
Presented by:
Supervisor:
Rufat Mustafaev
Konstantin Kholodilin, PhD
Department of Economics
3. 2
- 10 local conflicts since 1991
- No less than 200,000 casualties
The paper attempts to estimate economic costs of conflicts in ex-USSR countries
using nightlights as a proxy for economic activity to challenge the lack of official data
- Costs of conflict analysis is in
demand due to militaristic agenda
- Statistical services are close to
non-existent in war conditions
- No reliable data for estimating
economic costs of conflict
Background Research
Estimate economic activity and
costs of conflict in ex-USSR
Use a novel method in Russian
literature (nightlights data)
How did conflicts affect former
USSR countries?
Are night lights data a viable
proxy for economic activity?
5. 4
There has been a number of papers concerning costs of conflict estimation and most
of them use GDP per capita as a proxy of economic activity
1 Costs of conflict
Paper
Treated
area
Time
frame
Dependent
variable
Method Result
Collier (1999) Uganda 1987-1999
Decade average
ΔGDP per capita
Pooled OLS
panel regression
Long: -2.2%/year
Short: -2.1%/year
Imai & Weinstein (2000) Global 1960-1996
Decade average
ΔGDP per capita
Random effects
panel regression
-1.25%/year
Bluszcz & Valente (2019)
Eastern
Ukraine
1995-2017 GD(R)P per capita
SCM,
Placebo tests
DID
Ukraine: -12.7%
Donetsk: -43%
Luhansk:- 52%
6. 5
Nightlights data present a valuable tool which available for all locations at any time
and is not vulnerable to imperfections of governmental statistical services
2 Nightlights and economic activity
Paper
Treated
area
Time
frame
Dependent
variable
Method Result
Henderson et al. (2012)
Sub-Sahara
Africa
1992-2008 ΔGDP Fixed effects
Up to 3.2%
difference from
official GDP
Bertinelli & Strobl (2013) Carribean 1993-2009 ΔGDP Fixed effects -0.7%/year
Kochnev (2019)
Eastern
Ukraine
2013-2016 Log Average light Fixed effects
-38% and -51% in
separatist-
controlled parts
of DNR and LNR
10. 8
1 Data description and sources
Nightlights data as well as economic and demographic data are obtained from 3
major sources
NASA satellite images
Nightlights data
National statistical bureaus
Economic data (regional level)
World Bank Data
Economic data (national level)
South Ossetia Chechnya Donetsk & Luhansk
2007
2008
1992
1998
Feb 2014
Nov 2014
Luminosity for conflict regions and their neighbors before and during conflicts
Data
sources
Values range
0-63
Regional average
yearly luminosity
725.8 mln
pixels
350 Gb
of raster images
11. Conflict
Neighbors
Control
7
1 Data description and sources
Country Regions Avg light 1992 Avg light 2018
Kazakhstan 14 0.43 2.44
Uzbekistan 13 9.38 3.14
Lithuania 10 1.44 6.15
Azerbaijan 9 1.96 3.89
Kyrgyzstan 9 9.46 2.29
Belarus 6 2.14 4.74
Latvia 5 1.26 5.95
Armenia 1 0.72 3.08
Estonia 1 1.49 6.45
Moldova 1 6.23 2.93
Tajikistan 1 0.86 2.95
Turkmenistan 1 0.39 3.12
Russia 80 3.50 4.96
Ukraine 27 5.89 4.34
Georgia 11 4.69 7.07
Time frame
1992-2018
15 countries
189 regions
The paper uses panel data with 27 time periods and 189 region units 4 of which are
conflict affected and 11 are neighboring regions potentially prone to spillovers
12. 11
𝑙𝑜𝑔 𝑜𝑙𝑑_𝑓𝑜𝑟𝑚𝑎𝑡 = 1.5612 + 0.7347 ∗ 𝑙𝑜𝑔 𝑛𝑒𝑤_𝑓𝑜𝑟𝑚𝑎𝑡
Observed
values
Fitted values
Lin-Lin Lin-Log
Log-Log
Log-Lin
Lin-Lin Lin-Log Log-Lin Log-Log
AIC 2017.9 2106.1 876.8 783.2
BIC 2029.3 2117.4 888.1 794.6
Norm RMSE 8.46% 9.69% 12.50% 7.28%
Adjusted R^2 0.7518 0.6748 0.2815 0.4608
Ramsey's test
p-value
0.000 *** 0.000 *** 0.000 *** 0.0603 .
Simple linear regression model helps establish relationship between the values from
the old satellites and from the new ones and convert the data to a common format
2 Converting data to a common format
13. 12
Based on formal tests, prediction power and intuition we choose to proceed with a
panel data OLS fixed effects regression model with log-log specification
3 Estimating GRP using nightlights values
Time Region Country Time + Region Time + Country
Norm. RMSE 26% 7% 16% 9% 12%
Adjusted R2
0.210 0.737 0.428 0.964 0.666
β coefficient -0.050 0.141 0.121 0.035 0.039
p-value 0*** 0 *** 0 *** 0 *** 0 ***
Test p-value
BPLM 0.000 ***
Var(ui) ≠ 0,
RE is preferred to pooling
Hausman 0.000 ***
RE estimator is inconsistent,
FE is preferred to RE
𝒍𝒐𝒈 𝑮𝑹𝑷𝒊𝒕 = 𝜷 C 𝒍𝒐𝒈 𝑵𝑳𝒊𝒕 + 𝜸𝒊 + 𝒖𝒊𝒕,
where 𝐺𝑅𝑃KL – GRP PPP per capita in
constant 2011 USD, 𝑁𝐿KL - region average
yearly nightlights intensity, 𝛾K - region fixed
effect, 𝑢KL - error term
6 12
Fixed effects
for every region
OLS Region Fixed Effects
14. 13
Costs of conflict estimation is conducted by constructing a synthetic control for every
analyzed region and calculating the difference between potential and actual values
4 Costs of conflict estimation (SCM and DID)
Source: Abadie & Gardeazabal (2003)
Synthetic Control Method Difference-in-Differences
SCM estimator: Q
𝑎RL = 𝑌RL − ∑V 𝑤V
∗
𝑌
VL,
where 𝑌RL – treated unit outcome, 𝑌
VL – control unit
outcome, 𝑤V
∗
– weights that are included in the model:
W
V
𝑤V𝑌KL
X
= 𝛿L + 𝜃L W
V
𝑤V𝑍K + 𝜆L W
V
𝑤V𝜇K + W
V
𝑤V𝜖KL
The goal is to set such 𝑤V
∗
that ∑V _
𝑤V𝑌
VL = 𝑌
VL as well as
∑V _
𝑤V𝑍V = 𝑍R
Variables used: nightlights, Δnightlights, pop.density,
unemployment, distance to treated region, ethnic
fractionalization
𝑙𝑜𝑔 𝑒𝑠𝑡. 𝐺𝑅𝑃 = 𝛼 + 𝛽R𝑡𝑟𝑒𝑎𝑡𝑒𝑑 + 𝛽y𝑐𝑜𝑛𝑓𝑙𝑖𝑐𝑡
+𝛽|𝑡𝑟𝑒𝑎𝑡𝑒𝑑 ∗ 𝑐𝑜𝑛𝑓𝑙𝑖𝑐𝑡,
where 𝐷𝑖𝐷 = 𝛽|
𝐷𝑖𝐷
= 𝑌~•€L••‚
ƒ•„…~•€†‚K~L
− 𝑌L•„‡L„ˆ
ƒ•„…~•€†‚K~L
− 𝑌~•€L••‚
ƒ•‰L…~•€†‚K~L
− 𝑌L•„‡L„ˆ
ƒ•‰L…~•€†‚K~L
15. 14
In case of South Ossetia conflict, there was no significant effect for South Ossetia
and its Georgian neighbors while North Ossetia suffered a 13.5% decrease in GRP
South Ossetia
4.1
Path plot Gaps plot
Estimated
GRP
Treated
-
Synthetic
DID est p-value
South
Ossetia
+13.42% 0.4151
Georgian
neighbors
-3.57% 0.7506
Russian
neighbors
-13.48% 0.0203 *
Treated region
Synthetic region
Conflict in South Ossetia (2008)
16. 15
In case of Chechen wars, there was a statistically significant 21.1% drop in GRP for
Chechnya while the region’s neighbors were not affected in both Georgia and Russia
Chechnya
4.2
Path plot
Estimated
GRP
DID est p-value
Chechnya -21.10% 0.0091 **
Georgian
neighbors
-5.49% 0.8312
Russian
neighbors
-0.02% 0.9988
First Chechen war (1992-1994)
Implicit conflict (1995-1999)
Treated region
Synthetic region
Gaps plot
Treated
-
Synthetic
17. 16
Due to the war in Eastern Ukraine Donetsk lost more than a half and Luhansk – a
quarter of their GRPs. The conflict also significantly affects their Ukrainian neighbors
Eastern Ukraine
4.3
Path plot
Estimated
GRP
DID est p-value
Donetsk -58.56% 0.0000 ***
Luhansk -26.51% 0.0000 ***
Ukrainian
neighbors
-26.39% 0.0106 *
Russian
neighbors
-2.71% 0.7264
Donbass conflict (since 2014)
Euromaidan (2013)
Treated region
Synthetic region
Gaps plot
Treated
-
Synthetic
18. 17
Our results for Eastern Ukraine are comparable with the ones presented by Kochnev
and Bluszcz & Valente. All conflict regions suffer economic losses in terms of USD
Comparison
4.4
Region Loss, % p-value Loss, GRP PC p-value
South Ossetia + 13.42% 0.4151 - $ 2988.3 0.000 ***
GE neighbors - 3.57% 0.7506 - $ 305.4 0.472
RU neighbors - 13.48% 0.0203 * - $ 617.3 0.008 **
Chechnya - 21.10% 0.0091 ** - $ 518.2 0.041 *
GE neighbors - 5.49% 0.8312 + $ 28.2 0.976
RU neighbors - 0.02% 0.9988 - $ 43.6 0.931
Donetsk - 58.56% 0.0000 *** - $ 3252.4 0.000 ***
Luhansk - 26.51% 0.0000 *** - $ 1357.6 0.002 **
UA neighbors - 26.39% 0.0106 * - $ 1721.5 0.008 **
RU neighbors - 2.71% 0.7264 - $ 262.0 0.627
Paper Loss
Collier
- 2.1% Short wars
- 2.2% Long wars
Imai & Weinstein - 1.25%
Kochnev
- 38% Donetsk
- 51% Luhansk
Bluszcz & Valente
- 12.7% Ukraine
- 43% Donetsk
- 52% Luhansk
20. 19
The paper presents worthy empirical results despite the limitations associated with
lack of data and provides a wide area for improving and enhancing the analysis
Conclusions Limitations Prospects
Long-term conflicts caus-
ed significant economic
losses, a short one did not
Only one robust evidence
of spillovers effect on
neighboring regions
Estimation with nightlights
data shows comparable
results to a similar paper
Lack of pre-treatment
economic data for South
Ossetia and Chechnya
Short pre-treatment period
with available nightlights
data in case of Chechnya
Lack of resources and
computational power
Estimate costs of conflict
for different periods after
a conflict
Use spatial regression
models
Include city-level analysis
and/or monthly data for
2012-2018
25. Per capita GRP trends
in Donetsk vs. its synthetic counterpart
Per capita GRP trends
in Luhansk vs. its synthetic counterpart
Per capita GRP trends
in the Ukraine vs. its synthetic counterpart
23
Source: Bluszcz & Valent (2019)
Bluszcz &
Valent
Our
results