AIDDATA
A Research Lab at William & Mary
FINAL REPORT: 
Evaluation of the Infrastructure 
Needs Program II
A geospatial impact evaluation 
of the effect of INP II road 
improvements on economic 
development
Ariel BenYishay, 
Rachel Trichler, 
Dan Runfola, 
Seth Goodman
Author Information 
Ariel BenYishay 
AidData, William & Mary 
abenyishay@aiddata.org 
Rachel Trichler 
AidData, William & Mary 
rtrichler@aiddata.org 
Dan Runfola 
AidData, William & Mary 
drunfola@aiddata.org 
Seth Goodman 
AidData, William & Mary 
sgoodman@aiddata.org 
The views expressed in this report are those of the authors and should not be 
attributed to AidData or funders of AidData’s work, nor do they necessarily reflect the 
views of any of the many institutions or individuals acknowledged here. 
Citation 
BenYishay, A., Trichler, R., Runfola, D., & Goodman, S. (2018). Final Report: Evaluation 
of the Infrastructure Needs Program II. Williamsburg, VA: AidData at William & Mary. 
Acknowledgments 
This evaluation was funded by USAID/West Bank and Gaza through a buy-in to a 
cooperative agreement (AID-OAA-A-12-00096) between USAID's Global 
Development Lab and AidData at the College of William and Mary under the Higher 
Education Solutions Network (HESN) Program. The authors would like to 
acknowledge the research assistance of John Lakwijk and Carey Glenn at AidData 
and the contributions of Tayseer Edeas, Reem Jafari, and their colleagues at USAID/
West Bank and Gaza.
Executive	Summary
AidData	and	USAID/West	Bank	and	Gaza	have	collaborated	to	conduct	an	evaluation	of	the	
Infrastructure Needs	Program	(INP)	II,	which	funded	the	construction	of	water	
infrastructure,	road	networks,	and	schools	with	the	ultimate	goal	of	improving	the	
Palestinian	economy.	We	conduct	a	geospatial	impact	evaluation,	which	utilizes	program	
data	provided	by	the	Mission	on	the	timing	and	location	of	road	improvements	and	
remotely-sensed	nighttime	lights	outcome	data	to	determine	the	program’s	impact	on	
economic	development.
We	employ	a	quasi-experimental	panel	framework	with	monthly	and	cell-level	fixed	effects	
to	rigorously	estimate	program	impacts due	to	completion	of	road	improvements	
(“treatment”).	To	do	so,	we	construct	a	dataset	with	monthly	outcome	measures	between	
April	2012	(11	months	prior	to	the	first	road	improvements)	and	December	2016	(five	
months	after	the	last	road	improvements)	for	750m	square	grid	cells	within	5km	of	an	
improved	road	segment.	We	compare	post-treatment nighttime	lights	in	each	cell	to	
counterfactual	outcomes	obtained	from	that	cell’s	own	preceding	nighttime	light	levels	and	
trends,	as	well	as	the	outcomes	of	cells	near	not-yet-improved	road	segments.	The	
variation	in	the	timing	of	road	improvements	across	cells	and the	inclusion	of	time	and	cell￾level	fixed	effects	at	fine	geographic	levels	address	concerns	about	confounding	and	
omitted	variables.
We	find	a	substantive	and	statistically	significant	increase	in	nighttime	lights	due	to	INP	II.	
We	also	find	that	proximity	to	multiple	improved	road	segments	further	increases	
nighttime	lights	values and	that	cells	with	higher	baseline	luminosity	experience	relatively	
higher	gains.	These	findings	seem	to	point	to	the	importance	of	improving	roads	in	areas	
with	multiple	potential	access	points	to	the	larger	road	network	and	in	more	urban	or	
densely	populated	areas.	It	is	worthwhile	to	both	further	examine	the	positive	treatment	
effects	and	qualitatively	understand	the	larger	context	of	the	INP	II	program	in	which	these	
road improvements	occurred.
1
1.	Introduction and	Background
Program	Overview
The	Infrastructure	Needs	Program	(INP)	II	funded	by	USAID	West	Bank	and Gaza was	part	
of	a	larger	effort	to	support	the	Palestinian	Authority’s	High-Impact	Micro-Infrastructure	
Initiative	(HIMII). One	of	several	infrastructure	projects	funded	by	USAID,	it	began	in	2010	
with	the	goal	of	developing	the	water	and	wastewater	infrastructure,	road	networks,	and	
schools	of	the	region.	The	project	was	created	with	the	purpose	of	improving the	access	of	
Palestinians	to	basic	infrastructure	needs,	in	hopes	of	ultimately	improving	the	Palestinian	
economy.
While	INP	II	included	a	number	of	interventions,	this	evaluation	focuses	only	on	the	road	
improvements	funded	by	the	project.	The	road	network	component	targeted	small	sections	
of	road	(ranging	from	1-15	km)	scattered	throughout	the	West	Bank	with	the	objective	of	
improving	connectivity	to	main	roads	and	the	larger	road	network.	The	intervention	itself	
varied	by	the	needs	of	each	road	segment,	and	included	both	new	construction	and	
rehabilitation	of	existing	road,	such	as	milling	and	overlay	and	adding	or	improving	
retaining	structures,	water	drainage,	shoulders,	intersections,	and	road	signs.	This	
evaluation	does	not	distinguish	among	various	types	of	improvement,	but	rather	considers	
any	type	of	improvement	as one	broad	intervention.
Geospatial	Impact	Evaluation
The	collaboration	between	AidData	and	USAID/West	Bank	and	Gaza began	as	a	result	of	an	
expression	of	interest	in	geospatial	impact	evaluation	methods. Geospatial	Impact	
Evaluations	(GIE)	utilize	spatial	program	and	outcome	data	to	examine	the	change	in	an	
outcome	over	time	that	is	attributable	to	a	specific	project.	It	requires	data	on	the	timing	of	
project	implementation	across	locations	and	outcome	measures	before,	during,	and	after	
the	project,	and	uses	statistical techniques	to	create	a	counterfactual,	or	what	would	have	
happened	in	the	absence	of	the	project.	
It	became	clear	during	our	discussions	with	the	Mission	that	the	road	improvement	
component	of	INP	II	would	lend	itself	well	to	a	GIE.	The	Mission	had	collected	spatial	data	
on	the	locations	of	road	segments	and	the	timing	of	improvement,	which	would	serve	as	
the	“treatment”	in	the	impact	evaluation.	The	road	segments	were spread	throughout	the	
West	Bank	(rather	than	clustered	in	a	small	area)	and	the	timing	of	improvements	varied	
within	the	4	years	of	the	project	(i.e.	improvements	were	made	throughout	the	project	
period,	rather	than	simultaneously	for	all	roads)	– two	features	that	are	useful	for	this	type	
of	evaluation.
The	Mission	was	interested	in	determining	the	impact	of	road	improvements	on	economic	
development.	We	examined	several	potential	sources	of	outcome	data,	looking	for	data	at	
high	spatial	resolution	(municipality	or	smaller)	and	that	would	include	pre-project	
measures	as	well	as	frequent	measures	throughout	the	life	of	the	project. Satellite	data	on	
2
nighttime	lights,	used	as	a	proxy	for	economic	development	(see	further	discussion	below),	
provided	a	relevant	outcome	measure	at	high	resolution	with	monthly	measurements	
before,	during,	and	after	the	project	period	for	all	areas	in	the	West	Bank.	
The	Evaluation	Design	section	details	the	process	by	which	we	merged	these	sources	of	
existing	data	to	conduct	an	impact	evaluation	of	INP	II.
Nighttime	Lights	and	Economic	Development
We	use	remotely-sensed	nighttime	lights	as	a	proxy	outcome for	economic	development.	
Many	researchers	have	demonstrated	that	night	lights	indicate	economic	activity	(Sutton	
and	Costanza	2002,	Ebener	et	al.	2005,	Doll,	Muller,	and	Morley	2006,	Sutton,	Elvidge,	and	
Ghosh	2007,	and	Ghosh	et	al.	2010),	and	more	recently	it	has	been	shown	to	be	a	strong	
proxy	for	GDP	and	an	accurate	measure	of	short-term	and	long-term	economic	growth	at	
the country	level	(Henderson,	Storeygard,	and	Weil	2012).	Important	to	this	work,	higher	
levels	of	luminosity	correlate	strongly	with	greater	subnational	economic	activity	as	well	
(Michalopoulos	and	Papaioannou	2014). The	availability	of	nighttime	lights	data at	higher	
resolution	and	with	greater	frequency	than	traditional	forms	of	household	economic	data	
make	it	possible	to	conduct	this	impact	evaluation	of	INP	II	as	the	program	nears	
completion.
2.	Evaluation	Design
A	Geospatial	Impact	Evaluation	aims	to	determine	a	program’s	impact	on	an	outcome	over	
time,	merging	multiple	sources	of	spatial	data	to	do	so.	We	employ	a	quasi-experimental	
panel	framework	to	assess	economic	productivity	before	and	after	INP	II	road	
improvements	(or	“treatment”)	for	areas	within	5km	of	the	improved	road	segments.	The	
unit	of	analysis	for	this	evaluation	is	a	750m	x	750m	grid	cell,	which	is	the	unit	of	
measurement	for	the	nighttime	lights	satellite	data.	The	evaluation	period	begins	in	April	
2012	and	ends	in	December	2016.
In	this	section,	we	first	describe	the	identification	of	the	evaluation	sample	and	the	sources	
of	geospatial	data	used	for	the	analysis.	We	then	explain	the	quasi-experimental	panel	
methodology	in	greater	detail.
Data
Sample
USAID/West	Bank	and	Gaza	provided	a	shapefile	of	59	road	segments to	be	improved	
through	INP	II.	Of	these,	improvements	to	30	segments	were	completed	by	December	2016	
and	improvements	to	the	remaining	29	segments	were	completed	after	2016.	We	draw	a	
5km	buffer	around	each	road	segment	to identify	a	“catchment	area”	with	the	potential	to	
receive	economic	benefit	from	the	road	improvements,	and	include	the	750m square grid	
cells (from	the nighttime	lights	satellite	data) that	fall	within	this	buffer (See	Figure	1). In	
some	cases,	portions	of	the	5km	buffer	and	its	enclosed	cells	would	fall	beyond	the	
3
administrative	boundaries	of	the	West	Bank.	We	do	not	include	any	grid	cells	that	fall	
outside	of	the	West	Bank’s	administrative	boundaries.
Figure	1:	Creating	the	Evaluation	Sample
Our	sample	includes	6,729	unique	grid	cells. Using	this	sample	of	grid	cells,	we	construct	a	
dataset	that	includes	monthly	treatment	and	outcome	data	for	each	cell	from	April	2012	
through	December	2016,	the	evaluation	period. Figure	1 demonstrates	that many	of	the	
5km	buffers overlap.	As	such,	about half	(53%)	of the	grid	cells	fall	within	the	buffer	of two	
or	more	road	segments.	We	address	the	implications	of	this	on	our	treatment	measure in	
the	next	section.
Treatment	Data
In	this	impact	evaluation,	“treatment”	occurs	when	the	planned	improvements	to	a	road	
segment	through	INP	II	are	completed.	USAID/West	Bank	and	Gaza	provided	information	
on	the	month	and	year	of	completion	for	improvements	to	the	59 road	segments included	
in	the	project.	Only	29	of	these	segments	actually	had	completion	dates,	as	the	other	30	
segments	were	completed	after	December	2016	and	outside	of	the	evaluation	period.	The	
earliest	treatment	date	is	March	2013 and	the	latest	is	July	2016.
We consider	a	grid	cell	to	be	treated	in	the	month	during	which	improvements	were	
completed	for	its	associated	road	segment. Prior	to	treatment,	a	cell	is	part	of	the	control	or	
untreated	group.	Our	sample	does	include	grid	cells	that	only	fall	within	the	5km	buffer	for	
one	of	the	30	road	segments	scheduled	for	improvements	after	December	2016.	These	cells	
comprise	about	30%	of	our	sample	and,	as	they	are	never	treated,	always	remain	in	the	
4
control	group.	The	remaining	70%	of	the	cells	in	our	sample	move	from	the	control	to	the	
treatment	group	at	the	time	of	treatment.	This	differs	from	other	impact	evaluation	
methods,	in	which	an	observation	is	always	part	of	the	control	group	or	always	part	of	the	
treatment	group.	We	discuss	the	panel	methodology	that	allows	for	dynamic	treatment	and	
control	groups	in	greater	detail	below.
It	is	also	worth	noting	that	about	half	of	the	grid	cells	in	our	sample	fall	within	the	5km	
buffer	for	multiple road	segments.	It	is	easy	to	imagine	that	proximity	to	two	or	three	
improved	roads,	instead	of	just	one,	might	have	an	additional	impact	on	economic	
development	in	area.	To	explore	the	effects	of	exposure	to	multiple	treatments,	we
construct	multiple	treatment	measures in	our	regression	models	and	we	do so	
chronologically	– i.e.	Treatment	1 marks	the	earliest	date	of	improvements,	Treatment	2	
marks	the	second	earliest (and	that	a	cell	is	doubly	treated),	and	Treatment	3	marks	the	
third	earliest (and	that	a	cell	is	triply	treated).	The number	of	cells	that	fall	into	4	or	more	
buffers	is	sufficiently	small	that	we	do	not	include	any	treatment	effects	beyond	Treatment	
3.	The	maximum	number	of	buffers	for	one individual	cell	is	nine.
Outcome	Data
We	use	remotely-sensed	satellite	imagery	of	nighttime	lights	as	a	proxy	outcome	for	
economic	development.	Nightly	satellite	images	collected	by	the	National	Oceanic	and	
Atmospheric	Administration	(NOAA)	Visible	Infrared	Imaging	Radiometer	Suite	(VIIRS)	
measure	nighttime	light	activity	starting	in	April 2012	for	750m square	pixels.	The	raw	
nightly	data	measure	radiance	or	brightness	in	a	given	pixel	and	are	aggregated	into	
monthly	composites that	are	also processed	to	address	background	noise	and	exclude	
exceptional	instances	of	brightness	(e.g.	fires). Table	1 provides	summary	statistics	for	the	
nighttime	lights	values	at	baseline	(April	2012)	and	endline	(December	2016).	The	unit	of	
measure	is	not	particularly	intuitive	and	the	values	are	best	understood	in	relation	to	each	
other	(e.g.	the	median	value	in	relation	to	the	minimum	and	maximum	values,	the	amount	
of	change	experienced	by	the	average	grid	cell	during	the	evaluation	period,	etc.).
While	earlier	nighttime	lights	products	provide	annual	measures,	the	monthly	intervals	of	
this	outcome	data	lend	greater	precision	as	we	begin	to	measure	the	impact	of	INP	II	in	the	
exact	month	of	treatment. The	frequency	of	collection	also	provides	multiple	pre-project	
measures	to	establish	baseline	trends	for	each	individual	grid	cell	and	multiple	post-project	
measures	to	detect	delayed	treatment	effects.	In	our	sample,	there	exists	at	least	11	months	
of	pre-treatment	outcome	measures	and	5	months	of	post-treatment	outcome	measures.
We	also	utilize	another	remotely-sensed	nighttime	lights measure	as	part	of	a	robustness	
check.	Nightly	satellite	images	collected	through	the	Defense	Meteorological	Satellite	
Program	(DMSP)	provide	yearly	composite	measures	from	1km	square	grid	cells	between	
1992	and	2013.	
5
Methodology
Our	primary estimation	of	treatment	impact employs	a	panel	structure	with	monthly	
treatment	and	outcome data	for	each	of	the	6,729 grid	cells	in	our	sample.	The	panel	
dataset	includes	monthly	data	beginning	in	April	2012	and	ending	in	December	2016. We	
opt	for	grid	cells	as	the	unit	of	analysis	to	improve	the precision	of	our	outcome	measures	
and	validity	of	our	estimates.	
Our	panel dataset	includes	57	monthly	measures	(April	2012	– December	2016)	of	
nighttime	lights	for	each	of	6,729	cells,	or	383,553	observations	total.		The	dataset	also	
identifies the	month	in	which	a	cell	is	treated	(or	when	improvements	are	completed	for	its	
associated	road	segment) and	considers	a	cell	treated	for	every	month	thereafter.	The	
timing	of	treatment	differs	for	each	of	the	29	road	segments and thus	varies	across	grid	
cells.	Panel	methods	essentially	align	the	timing	of	treatment	and	aim	to	identify	a	pattern	
of	change	in	nighttime	lights	for	all	cells	relative	to	each	cell’s treatment,	e.g.	one	month	
after	treatment,	five	months	after	treatment,	etc. Figure	3 provides	an	example	of	this	
process	using	illustrative	data.
The	variation	in	the	actual	calendar	date of	treatment across	cells helps	to	address	
concerns	about	confounding	variables,	or	other	fixed	factors	specific	to	each grid	cell	that	
may	also	correlate	with	changes in	nighttime	lights.	In	a	panel	model	at	fine	geographic
scale	and	with	varied	timing	of	treatment,	these	alternative	explanations are	very	unlikely	
to	begin	affecting each	cell	at	the	exact	same	time	as	INP	II-funded improvements	are	
completed.	For	example,	if	an	employment	program	is	rolled	out	throughout	the	West	Bank	
in	May	2014,	it	would	likely	benefit	communities	near	roads	treated	late	in	INP	II	as	well	as	
those	treated	earlier.		The	threat	to	our	causal	attribution	is	thus	only	from	other	programs	
or	factors	that	differentially	affected	newly	improved	road	corridors	with	the	very	same	
timing	as	INP	II.		This	threat	appears	quite	small	given	the	dispersion	of	INP	II-improved	
road	segments	over	both	time	and	space.		We	are	thus	confident	that our	empirical	
approach	produces	casual	estimates	that	can	be	attributed	to	INP	II.
Our	models	also	include	month	and	cell-level	fixed	effects.	Month	fixed	effects	control	for	
general	temporal	trends	throughout	the	evaluation	period	(e.g.	the	average	nighttime	lights	
values	increase	between	2012	and	2016,	lights	become	cheaper,	or	minimum	wage	
increases),	as	well	as	changes	in	nighttime	lights	experienced	by	all	cells	in	a	specific	month	
(e.g.	nighttime	light	usage	is	more	common	in	certain	months	of	the	year,	a	massive	power	
outage	occurs	in	a	specific	month,	or	a	month	includes	more	unpaid	holidays which	lowers	
average	wages).	Cell-level	fixed	effects	control	for	the	time-invariant	characteristics	of	each	
cell,	many	of	which	might	also	impact	nighttime	lights	values	(e.g.	proximity	to	a	town	or	
city,	population	density,	proximity	to	water,	slope,	or	elevation).	Controlling	for	the	time￾invariant	features	of	each	cell	is	a	key	advantage	of	a	panel	model	and	allows	each	cell	to	
serve	as	its	own	counterfactual	(rather	than	trying	to	match	similar	control	and	treatment	
cells	with	limited	data).	Applying	month	and	cell-level	fixed	effects	at	high	levels	of	spatial	
resolution	helps	to	control	for	potential	confounds	and	omitted	variables	that	would	
otherwise	bias	our	results.
6
Using	the	panel	framework	with	fixed	effects,	we	estimate	the	following	equation:
𝑁𝑖𝑔ℎ𝑡𝑡𝑖𝑚𝑒 𝐿𝑖𝑔ℎ𝑡𝑠!"# = ∝ + 𝛽 𝑅𝑜𝑎𝑑 𝐼𝑚𝑝𝑟𝑜𝑣𝑒𝑚𝑒𝑛𝑡𝑠!"# + 𝐷! + 𝐷! + 𝜖!"#
Where	𝑅𝑜𝑎𝑑 𝐼𝑚𝑝𝑟𝑜𝑣𝑒𝑚𝑒𝑛𝑡𝑠!" indicates	whether	cell	i	in	municipality	m has	been	treated	
by	month	t,	𝐷! is	a	vector	of	grid-cell	fixed	effects,	and	𝐷! is	month	fixed	effects.	We	
estimate	treatment	effects	via	ordinary	least	squares.		We	also	use	two-way	clustering	of	
standard errors	by	municipality	and	month.
3.	Findings
Main	Treatment	Effect
Table	2 presents	the main	model	of treatment	effects	of	road	improvements	on	nighttime	
lights.	All	models	include	cell-level	fixed	effects,	which	control	for	all	time-invariant	
features	of	each	individual	cell.	The	pre-treatment	measures	of	nighttime	lights	included	in	
the	panel	dataset	establish	baseline	values	and	trends.	We	do	not	include	any	additional	
covariates,	as	time-varying	data	is	limited	at	such	high	resolution,	and	the	panel	methods	
with	time	and	cell-level	fixed	effects	help	to	address	concerns	about	confounding	factors	
(or	other	variables	that	might	be	causing	the	observed	changes	in	nighttime	lights
attributed	to	this	program).
Columns	1	and	2	both	estimate	the	effect	of	a	cell’s	earliest	treatment	(or	the	earliest	
improvements	made	to	a	road	segment	within	5	km),	with	the	only	difference	being	the	
measurement	of	time	trends.	Column	1	includes	a	linear	time	effect, which	measures	the	
average	effect	of	the	passage	of	one	month	on	nighttime	lights, while	Column	2	instead	
includes	a	more	nuanced	monthly	fixed	effect	that	accounts	for the	effects	of each	month
individually.	Columns	3-5	also	include	month fixed	effects.	Column	2	shows	a	significant	
and	substantive	treatment	effect.	The	coefficient	indicates	that	treated	cells	experienced	an	
increase	of	0.341	for	nighttime	lights	measures	during	the	evaluation	period.	
The	meaning	of	a	numeric	change	in	nighttime	lights	is	not	particularly	intuitive,	but	the	
median	lights	values	and	the	average	amount	of	change	experienced	by	all	cells	during	the	
program	shed	some	light.	The	median	lights	value	at	endline	is	3.3 (see	Table	1),	so	a 0.341	
increase	is	roughly	10%	of	the	median	value.	From	baseline	to	endline,	the	median	value	
increased	from	2.6	to	3.4 (see	Table	1).	Given	this	change	of	0.8	experienced	by	the	median	
cell	during	the	evaluation	period,	a	0.341	increase	over	that	same	period	is	substantive
(about	40%	of	the	change	for	the	median	cell).	To	be	clear,	we	cannot	attribute	40%	of	the	
change	over	the	time	period	to	the	road	improvements	made	through	INP	II.	However,	a	
comparison	of	the	treatment	effect and	the	total	increase	for	the	median	cell	during	the	
same	time indicates that	our	estimate	of	treatment	impact	is	not	only	statistically	
significant,	but	substantively	significant	as	well.
Secondary	Treatment	Effects
7
We	next	examine	the	impact	of	exposure	to	multiple	treatments.	About	half	of	the	cells	in	
our	sample	fall	within	the	5km	buffer	of	multiple	road	segments	and	are	treated	at	least	
twice.	In	Column 3	of	Table	2,	we	add	in	Treatment	2,	which	turns	on	for	the	second	
earliest	set	of	road	improvements	for	a	cell.	Interestingly,	when	we	add a	second	treatment,
Treatment	1	maintains	a	statistically	significant	effect,	though	the	magnitude	of	that	effect	
lessens	(from	0.341	to	0.2)	compared	to	our	earlier	model. Treatment	2 demonstrates	a	
statistically	significant	effect	that	is	even	higher	than	the	main	treatment	effect	in	our	initial	
model,	with	a	coefficient	of	0.571.	Returning	to	the	context	provided	by	the	median	cell,	for	
the	median	nighttime	lights	value	of	3.3	at	endline (see	Table	1),	a	0.571	increase	is	equal	
to	17%	of	the	total lights value (compared	to	10%	of	the	total	value	when	only	Treatment	1	
is	included).	The	Treatment	1	and	Treatment	2	estimates	hold	when	a	third	treatment	is	
added	in	Column	4,	though	Treatment	3	does	not	cause any	statistically	significant change	
in	nighttime	lights	values	for	treated	cells.
Figure	2 visually	displays	the	treatment	coefficients	from	our	main	models	in	Columns	2-4	
of	Table	2.	The	effect	of	Treatment	1	decreases	when	we	add	a	second	treatment	(Models	3	
and	4),	Treatment	2	demonstrates	the	largest	effect	on	nighttime	lights,	and	Treatment	3	
does	not	cause	any	statistically	significant	change	in	the	outcome	for	treated	cells	(shown	
by	the	line	representing	the	95%	confidence	interval	crossing	0).
Figure	2:	Graph of	Treatment	Coefficients	for	Main	Models
8
The	identification	of	what	might	be	driving	the larger	increase	in	nighttime	lights	that	
results	from	a	second	treatment	(as	compared	to	and	in	addition	to	the	first	treatment) is	
beyond	the	scope	of	this	analysis,	but	certainly	worthy	of	pursuit. It	is	possible	that	the	
later	improvements	made	to	roads	were	better	in	some	way.	A	more	likely	explanation	is
that	a	second	treatment indicates	greater	connectivity	to	multiple	roads.	Due	to	the	
networked	nature	of	roads,	it	seems	possible that	the	marginal	effect	of	a	second	
connection	point	could	be greater	than	the	effect	of	the initial	connection	and	would	lead	to	
greater	economic	opportunity.	Not	only	are	two	connections	to	a	road	network	better	than	
one,	later	improvements	would	benefit	from	a	better	and	more	established	road	network	to	
connect	to	(as	a	result	of	earlier	improvements).
Using	the	road	network	explanation	of	the	importance	of	the	second	treatment	above,	we	
would	expect	to	see	an	additional	effect	from	Treatment	3	in	Column	4	of	Table	2.	However,	
we	do	not.	About	30%	of	our	sample	of	cells	fall	within	the	buffers	of	3	or	more	road	
segments	and	the	standard	error	on	the	coefficient	is	still	quite	small,	indicating	that	we	do	
have	enough	statistical	power	to	detect	any	treatment	effect	that	exists	for	our	sample.	It	is	
possible	that	these	cells	have	reached	a	saturation	point,	in	which	any	impact	on	economic	
activity occurred	from	the	first	or	second	treatment	and	the	third	treatment	does	not	
provide	any	added	benefit (at	least	as	measured	by	nighttime	lights) for	a	cell	of	that	size.
Further	examination	of	the	effects	of	exposure	to	multiple	improved	road	segments	would	
provide	useful	insight	for	this	evaluation	and	the	siting	and	implementation	decisions	of	
future	road	projects.
Heterogeneous	Effects:	Nighttime	Lights	at	Baseline
We	consider	whether	results	vary	by	the	baseline	levels	of nighttime	lights	in	each	cell	– in	
other	words,	do	we	see	a	differential	effect	for	cells	that	were	darker	at	baseline	(in	less	
populated areas)	versus	those	that	were	brighter	at	baseline	(in	more	populated	areas)?
Nighttime	lights	at	baseline	serve	as	the	best	available	measure	of	population	density.	In	
most	settings,	there	is	a	high	correlation	between	nighttime	lights	and	population	density.	
While	other	sources	of	population	data	do	exist	for	our	study	area,	they	are	available	much	
less	frequently	(yearly,	at	best),	and	at	much	coarser	resolution	(often	only	available	by	
governorate).	This	exploration	of	heterogeneous	effects also	addresses	whether	the	INP	II	
improvements	shifted	economic	activity	from initially	less	active	to	more	active	areas.
We	first	categorize	nighttime	lights values	at	baseline	by quartiles.	The	darkest	25%	have	
nighttime lights	values	ranging	from	0.42	to	1.27,	while	the	brightest	25%	have	nighttime	
lights	values	ranging	from	6.56	to	148.099 (see	Table	1).	These	numeric	values	are	a	
measure	of	brightness that	is	not	particularly	intuitive	and are	best	understood	in	relation	
to	each	other.	While	there	is	a	large	gap	between	the	smallest	value,	0.42,	and	the	largest,	
148.099,	the	middle	half	of	the	data	(from	25%	to	75%)	only	spreads	across	5 points,	
indicating	that	small	changes	in	brightness are	meaningful. The	range	in	the	top	quartile
(6.56	– 148.099) reveals	the	existence	of	a	small	number	of	very	bright	outlier	cells	in	our	
sample.	
9
Column	5	of	Table	2 presents	the	impact	of	the	earliest	treatment (Treatment	1) on	cells	
with	differing	levels	of	baseline	luminosity.	The	impacts	are	statistically	significant	for	all	
quartiles. In	cells	that	were	darkest	at	baseline,	we observe	a	coefficient	of	-0.31,	or	an	
actual reduction	in	nighttime	lights	values	for	treated	cells.	For	cells	that	were	semi-dark,	
the	impact	on	nighttime	lights	is	still	negative	but	smaller	than	the	darkest	cells,	and	for	
cells	that	were	semi-bright,	we	see	a	positive	effect of	0.38.	For	those	cells	that	were	
brightest	at baseline,	we	observe a	large	increase	in	nighttime	lights (1.11) during	the	
evaluation	period.	It	should	be	noted	that	while	the	baseline	values	in	this	brightest	
quartile	are of	course	grater than	the	median	value,	the	magnitude	of	the	change (1.11) is	
similar percentage-wise to that	of Treatment	2 (about	17%	for	the cells	with	baseline	
luminosity	of	6.56).	We	consider	a	17%	increase	in	brightness	to	be	a	substantive	change.
The	differential	impacts	of	INP	II	are	interesting	– the	cells	that	are	brightest	at	baseline	
drive	the	main	effect	we	see	in	Column	2	of	Table	2.	The	cells	that	are	darkest	at	baseline	
actually	grow	darker.	Importantly,	we	do	not	see	evidence	that	the	main	treatment	effect	
merely	reflects	a	shift	in	nighttime	lights	values	from	darker	to	brighter	areas	(suggesting	a	
shift	of	the	same	level	of	economic	activity	from	one	area	to	another)	but	in	fact indicates	
an	overall	increase	in	nighttime	lights	and	economic	activity within	the	study	area	as	a	
whole.
Heterogeneous	Effects:	2km	Buffers
Table	3 presents	the	treatment	effects	for	a	sample	of	cells	that	limited	to	2km	buffers	
around	the	improved	road	segments (rather	than	5km).	The	smaller	buffers	help	to	
determine	if	the	estimates	of	treatment	impacts	hold	for	the	areas	closest	to	the	roads.	This	
is	of	particular	interest	in	this	setting	due	to	concerns	that	increases	in	settlement	
populations	(located	farther	from	the	road	segments)	might drive the	observed	increase	in	
nighttime	lights, rather	than	the road	improvements.	
Our	main results	generally	hold	when	we	use this	smaller	sample	of	82,814	cells that	fall	
within	2km	of	improved	road	segments.	While	the	statistical	significance	of	the main	
treatment	effect disappears	(Columns	1	and	2	of	Table	3),	the	inclusion of	the	second	
treatment	in	Column	3	identifies	an	increase	of	0.9	in	nighttime	lights values for	treated	
cells	during	the	study	period.	This	is	a	substantial	effect.	The	effect	of	a	third	treatment	in	
Column	4	is	actually	a	decrease	in	nighttime	lights	values.	This	result	is	a	bit	puzzling,	
although	it	is worth	noting	that	the	magnitude	of the	decrease	is	smaller	than	the	increase	
from	the	second	treatment,	which	still	indicates	an	overall	gain	in	nighttime	lights	values	
for	cells	with	three	treatments.
Heterogeneous	effects	due	to baseline values	of	nighttime	lights	also	demonstrate	a	similar	
pattern	to	the	differential	effects	found	in	the main	model.	In	Column	5	of	Table	2,	we	
observe a	decrease	in	luminosity for	the	darkest	cells	at	baseline.	Semi-bright	cells
experience	virtually no	change,	while	lights	values in	the	brightest	cells	increase	by	about	
1.	As	in	our	main	model,	we	conclude	that	while	the	darkest	cells	get	darker,	indicating	
some	shift	of	economic	activity	within	the	study	area,	we	also	find	evidence	of	an	overall	
increase	in	economic	activity	in	areas	close	to	the	improved	roads.
10
In	both	this	model	and	the	main	model,	nighttime	lights	at	baseline	serve	as	the best
available measure	of	population	density.	The	results	of	the	interaction	between	treatment	
and	nighttime	lights	indicate	that	treatment effects	are	positive	and	strong	in	the	most	
populated	areas – a	pattern demonstrated	when	limiting	the	study to	a	narrow	2km	buffer	
around	roads	and	at	a	larger 5km	distance	as	well.
Robustness	Checks
We	conduct	several	robustness	checks	to	help	validate	the	results	of	our	main	models. We	
first	consider	whether	our	results	vary	based	on	the	size and	placement of	our	unit	of	
analysis.	We	draw	the	unit	of	analysis	in	our	main	models	from	the	grid	cells	corresponding	
to	the	VIIRS	data,	the	size	and	placement	of	which	are determined	by	the	way	in	which	the	
satellite	data	is	collected.	We	would	not	expect	the	grid	cells	to	align	with roads,	political	
boundaries,	or	any	other	divisions	in	a	way	that	would	bias	our	results.	As	a	check,	we	
aggregate	outcome	and	treatment	values to	a	1km	square	grid	cell	and	re-run	our	models	
using	this	sample	of	cells.	As	we	would	expect,	Table	4 demonstrates	results	at	the	1km	
grid	cell	that	are	consistent	with	those	of	the	smaller	grid	cells	in	our	main model.
We	also	assess	whether	our	results	are	driven	by	a	trend that	takes	place over	time
regardless	of	our	treatment.	We	conduct	a	placebo	test	in	which	we arbitrarily	shift	the	
timing	of	treatment	earlier for	all	cells	– e.g.	if	a	cell	was	treated	in	September	2013,	we	
construct	a	dataset	that	instead	indicates	treatment	in	March	2013.	We	would not	expect	to	
see	a	treatment	effect	in	these	models	as	no	treatment	actually	takes	place	at	that	time.
We	construct	two	different	placebo	tests.	In	the	first,	we	construct	a	treatment	date	that	is	
six	months	earlier	than	the	actual	treatment.	We	use	the	same	VIIRS	monthly	nighttime	
lights	outcome	data.	Table	5 displays	the	results.	We	find	no	statistically	significant	effect	in	
Column 1,	which	includes	a	linear	time	trend.	When	we	replace	that	with	month	fixed	
effects	we	actually	do	observe	an effect,	though	one	that	is negative	and barely	statistically	
significant.	This	indicates	that	in	the	six	months	prior	to	the	actual	date	of	treatment,	
treated	cells	may	actually	experience	a	slight	decrease	in	nighttime	lights	values.	If	
anything,	these	findings	strengthen	the	results	demonstrated	in	our	main	model,	as	the	
observed	increase	in	nighttime lights	reversed	the	trend	in	those	cells	in	the	six	months	
prior	to	treatment.
For	the	second	placebo	test,	we	construct	a	treatment	date	that	is	5	years	earlier	than	the	
actual	treatment	date.	As	the	VIIRS	nighttime	lights	satellite	data	only	begins	in	April	2012,	
we utilize	an	alternate	satellite outcome	measure	in	these	models.	The	Defense	
Meteorological	Satellite	Program	(DMSP)	provides	yearly	aggregates	of	nighttime	lights	
between	1992	and	2013	at	a	1km	square	grid	cell	unit.	As	the	DMSP	data	is	only	available	
yearly,	the	timing	of	treatment	is also	assigned	yearly	– e.g.	if	a	cell	was	actually	treated	in	
September	2013,	the	placebo	treatment	is	2008,	and	this	would	be	true	for	any	cell	treated	
in	any	month	of	2013.	As	we	would	expect,	we	do	not	find	evidence	of	a	treatment	effect	for	
this	second	placebo	test.	
11
4.	Conclusions	and	Next	Steps
In	this	evaluation, we	do	find evidence	that	the	road	improvements	made	through	INP	II	
increase	nighttime	lights,	and	thus	economic	development,	for	areas	within	5km	of	the	
improved	road	segments.	We	find	that	the	level	of	brightness	at	baseline	differentially	
affects	the	change	in	nighttime	lights,	and	these	results	hold	when	we	vary	the	size	of	the	
buffer	and	the	size	of	the	grid	cell. Darker	areas at	baseline get darker	as	a	result	of	
program	treatment,	while	areas	that	are	brighter	at	baseline	get	brighter.	The increases	to	
nighttime	lights	in	these	brighter	cells	more	than	offset	the	decreases	in	the	darker	cells,	
suggesting	an	overall	increase	in economic	activity	across	the	study	area	as	a	whole,	rather
than	a	mere shift of	activity from	one	area	to	another.	
Despite	the	overall	gains,	the	differential	effects	by	level	of	brightness	at	baseline	are	of	
interest	and	deserve	additional exploration.	It	is	important	to note	that	nighttime	lights	
measures the production	of	economic	activity	and	these	findings	suggest	that	production	
has	increased	overall,	but	decreased	in	some	areas	and	increased	in	others.	However,	we	
cannot	conclude	that	welfare for	those	living	in	the	darker	areas	at	baseline has	necessarily	
decreased	in	the	darker	areas,	as	our	models	do	not	account	for	the	movement	of	people.	It	
is	certainly	possible	that	the	road	improvements	increased mobility	and	allowed	those	
living	in	darker	areas	to	obtain	economic	opportunities elsewhere that	maintained	or	
increased	their	welfare	levels.	The	opposite	is	also	possible,	in	which	both	economic	
production	and	welfare	levels	decreased	in	the	areas that	were	darkest	at	baseline.	The	
distinction is	important,	though	the	data	and	design	of	this	evaluation	are	limited	in	
shedding	greater	light	on	these	heterogeneous	impacts.	Additional	exploration	through	
household	surveys	or	other	on-the-ground	estimates	of	economic	wellbeing can	provide	
further	understanding	of	the	effects	on	beneficiaries to	contribute	to	future	planning	and	
program	decisions.
Overall,	the	positive	impacts	we	measure on	economic	activity	due	to	the road	
improvements	made	through	INP	II are	encouraging.	INP	II	is a	component	of	a	broader	
initiative	improving	access	to	infrastructure,	and	this	evaluation	indicates	that	gains	from	
this	investment	are	likely	to	be	substantive and	important.	In	an	effort	to	better	learn	from	
this	analysis,	it	is	worthwhile	to	both	further	examine	the treatment	effects	and	
qualitatively	understand	the	larger	context and	implementation of	the	INP	II	program.	
While	the	latter	is	probably	best	done	by	Mission	staff,	there	exist a	number	of	ways	to	
build	on	this	analysis	to	accomplish	the	former.	One	clear next	step	is	to	re-run	the	analysis	
with	the	information	on	road	improvements	that	were	completed	after	December	2016.	
These	improvements	were	not	included	in	our	analysis,	but	given	the	availability	of	
nighttime	lights	satellite	data	beyond	2016	and	the	ability	to	reapply the	same	design	and	
even	code	for	the	analysis,	this	is	relatively	low-hanging	fruit	that	would	serve	as	another	
robustness	check	to validate	the	results	or	identify	greater	nuance	within	the	program	
impacts.	
Building	on	the	foundation	of	this	evaluation,	new	analysis	could	include	incorporating	
data	on	the	larger	road	network	to	better	understand	how	the	INP	II	improvements	
connected	our	sample	of cells	with	roads	that	connect	to	other	areas	of	the	West	Bank	
12
(including	considerations	of	number	of	connection	points,	travel	time	to	major	roads,	etc).
It	would	also	be	instructive to	connect	the	satellite	measures	of	economic	activity	with	on￾the-ground	estimates	of	wellbeing	through	census	or	household	survey	data,	in	part	to	
better	understand	the	heterogeneous	effects	we	observe	due	to	baseline	luminosity.	Both	
of	these	options	require	additional	data	that	may	not	be	available,	at	all	or	for	the	needed	
time	periods,	or	new	data	collection.	Rigorous	evaluations	of	ongoing	or	future	
infrastructure	projects	(beyond	roads)	also	offer	opportunities	to	contribute	to	learning	
and	targeting	of	resources	to	improve	the	economic	development	in	the	region.
13
References
Doll, C. N., Muller, J., & Morley, J. G. (2006). "Mapping regional economic activity from 
night-time light satellite imagery." Ecological Economics 57, (1), 75-92. 
doi:10.1016/j.ecolecon.2005.03.007.
Ebener, S., Murray, C., Tandon, A., & Elvidge, C. C. (2005). “From wealth to health: modelling 
the distribution of income per capita at the sub-national level using night-time light imagery.” 
International Journal of Health Geographics, 4(5). doi:10.1186/1476-072X-4-5
Ghosh, T., Anderson, S., Elvidge, C., & Sutton, P. (2013). “Using Nighttime Satellite Imagery as 
a Proxy Measure of Human Well-Being.” Sustainability, 5(12), 4988-5019. 
doi:10.3390/su5124988
Henderson, J. V., Storeygard, A., & Weil, D. (2012). "Measuring Economic Growth from Outer 
Space," American Economic Review, American Economic Association, 102(2), 994-1028. 
doi:10.1257/aer.102.2.994.
Michalopoulos, S., & Papaioannou, E. (2014). "National Institutions and Subnational 
Development in Africa," The Quarterly Journal of Economics, Oxford University Press, vol. 
129(1), 151-213.
Sutton, P. C., & Costanza, R. (2002). “Global estimates of market and non-market values derived 
from nighttime satellite imagery, land cover, and ecosystem service valuation.” Ecological 
Economics, 41(3), 509-527. doi:10.1016/s0921-8009(02)00097-6
Sutton, P. C., Elvidge, C. V., & Ghosh, T. (2007). “Estimation of Gross Domestic Product at 
Sub-National Scales using Nighttime Satellite Imagery.” International Journal of Ecological 
Economics & Statistics, 8(S07), 5-21.
14
Figure	3:	Construction	of	Panel	Dataset (Using	Illustrative	Data)
1.	Collect	monthly	measure	of	nighttime	lights	data	for	each	750m	square	grid	cell	in	our	sample,	
beginning	in	April	2012	and	ending	in	December	2016	(57	months).	
2.	Compile	monthly	nighttime	light	measurements	for	all	cells and	identify	month	and	year	of	
treatment	(when	improvements	were	made	to	associated	road	segments	through	INP	II),	marked	
by	red	line.
April	
‘12
May Jun Jul Aug Sept Oct Nov Dec Jan	
’13
Feb Mar Apr May Jun
Cell	1
Cell	2
Cell	3
Cell	4
3.	Align	relative	timing	of	treatment	to	examine	patterns	of	change	in	nighttime	lights	across	all	
cells	and	relative	to	each	cell’s	timing	of	treatment	(i.e.	one	month	after	treatment,	five	months	after	
treatment,	etc.).
15
Table	1:	Summary	Statistics
Notes:	Table	1 displays	descriptive	statistics	for	the	VIIRS	cell-month	panel	dataset	used	to	estimate	the	
main	model	for	750	m	square	grid	cells	(see	Table	2 for	main	model	regression	results).	The	first	two	
rows	summarize	nighttime	lights	values	at	baseline	(April	2012)	and	endline	(December	2016).	The	
nighttime	lights	values	are	a	measure	of	radiance	or	brightness,	which	are	not	intuitive	and	best	
understood	in	relation	to	each	other.	First	Treatment	Length	refers	to	the	duration	of	a	cell’s	earliest	
treatment	(or	the	earliest	road	within	5	km	to	receive	improvements).	The	distance	measures	reflect	the	
average	distance	of	a	grid	cell to	the	improved	road	segments	(first	being	earliest).
16
Table	2:	VIIRS	Cell-Month	Panel	Model	Results	(Main	Model)
(1) (2) (3) (4) (5)
VARIABLES Model	1 Model	2 Model	3 Model	4 Model	5
Treatment	1 0.321** 0.341*** 0.200** 0.199**
(0.129) (0.116) (0.0926) (0.0926)
Treatment	2 0.571*** 0.533***
(0.172) (0.167)
Treatment	3 0.152
(0.237)
Treatment	1	x	NTL	Baseline 		-0.326***
(1st quartile;	Darkest) (0.0983)
Treatment	1	x	NTL	Baseline 0.146***
(2nd quartile) (0.0397)
Treatment	1	x	NTL	Baseline 0.414***
(3rd quartile) (0.0832)
Treatment	1	x	NTL	Baseline 1.584***
(4th quartile;	Brightest) (0.404)
Month 0.00259***
(0.000411)
Observations 383,525 383,525 383,525 383,525 383,525
R-squared 0.973 0.974 0.974 0.974 0.976
Grid	cell	FEs Y Y Y Y Y
Month	FEs N Y Y Y Y
Robust	standard	errors	in	parentheses
***	p<0.01,	**	p<0.05,	*	p<0.1
17
Table	3:	VIIRS	Cell-Month	Panel	Model	Results,	2km	Buffer
(1) (2) (3) (4) (5)
VARIABLES Model	1 Model	2 Model	3 Model	4 Model	5
Treatment	1 0.276 0.298 0.243 0.255
(0.213) (0.204) (0.193) (0.195)
Treatment	2 0.900** 0.931**
(0.445) (0.432)
Treatment	3 -0.763**
(0.356)
Treatment	1	x	NTL	Baseline
(1st quartile;	Darkest)
Treatment	1	x	NTL	Baseline
			-0.481***
					(0.143)
			0.195***
(2nd quartile) (0.0572)
Treatment	1	x	NTL	Baseline 0.468***
(3rd quartile) (0.108)
Treatment	1	x	NTL	Baseline 1.471***
(4th quartile;	Brightest) (0.548)
Month 0.00344***
(0.000534)
Observations 82,814 82,814 82,814 82,814 82,814
R-squared 0.969 0.972 0.972 0.972 0.974
Grid	cell	FEs Y Y Y Y Y
Month	FEs N Y Y Y Y
Robust	standard	errors	in	parentheses
***	p<0.01,	**	p<0.05,	*	p<0.1
18
Table	4:	VIIRS	Cell-Month	Panel	Model	Results,	using	1km	Grid	Cells
(1) (2) (3) (4) (5)
VARIABLES Model	1 Model	2 Model	3 Model	4 Model	5
Treatment	1 0.317** 0.336*** 0.196** 0.195**
(0.125) (0.112) (0.0882) (0.0882)
Treatment	2 0.547*** 0.498***
(0.168) (0.161)
Treatment	3 0.194
(0.238)
Treatment	1	x	NTL	
Baseline
-0.286***
(1st quartile;	Darkest) (0.0977)
Treatment	1	x	NTL	
Baseline
0.256***
(2nd quartile) (0.0563)
Treatment	1	x	NTL	
Baseline
0.691***
(3rd quartile) (0.128)
Treatment	1	x	NTL	
Baseline
1.326***
(4th quartile;	Brightest) (0.401)
Month 0.00250***
(0.000409)
Observations 218,196 218,196 218,196 218,196 218,196
R-squared 0.975 0.976 0.976 0.976 0.977
Grid	cell	FEs Y Y Y Y Y
Month	FEs N Y Y Y Y
Robust	standard	errors	in	parentheses
***	p<0.01,	**	p<0.05,	*	p<0.1
19
Table	5:	VIIRS	Placebo	Treatment	Panel	Model	Results	(6	months)
(1) (2)
VARIABLES Model	1 Model	2
Treatment	1	Placebo -0.00892 -0.180*
(0.102) (0.0967)
Month 0.00319***
(0.000457)
Observations 383,525 383,525
R-squared 0.972 0.974
Grid	cell	FEs Y Y
Month	FEs N Y
Robust	standard	errors	in	parentheses
***	p<0.01,	**	p<0.05,	*	p<0.1
Table	6:	DMSP	Placebo	Treatment	Panel	Model	Results	(5	years)
(1) (2)
VARIABLES Model	1 Model	2
Treatment	1 Placebo 0.250 -0.737
(2.204) (0.834)
Year -0.506
(0.696)
Observations 33,645 33,645
R-squared 0.893 0.923
Grid	cell	FEs Y Y
Month	FEs N Y
Robust	standard	errors	in	parentheses
***	p<0.01,	**	p<0.05,	*	p<0.1