Local Industrial Shocks and Infant Mortality

Anja Tolonen*

August 4, 2016

Abstract

Local industrial development has the potential to improve well-being, increasingpopulation health, while also damaging health through harmful pollution. It is an em-pirical question which of these effects will dominate. Exploiting the quasi-experimentalexpansion of African large-scale gold mining, I find that local infant mortality ratesdecrease by more than 50% alongside rapid economic growth. The almost instan-taneous reduction is comparable to overall gains in infant survival rates in the studycountries from 1970 to today. The results are robust to migration. Spurring industrialdevelopment—despite risk of pollution—is of first-order importance for reducing infantmortality in developing countries.

Keywords: Industrial Development, Gold Mining, Infant Mortality, Sustainable Devel-opment GoalsJEL classification: O12, O13, I15, J11

*Corresponding author. atolonen@barnard.edu, Department of Economics, Barnard College, ColumbiaUniversity; 3009 Broadway, NY 10027, New York; (212) 854-3454; Fax: (212) 854-8947. I would like tothank Klaus Deininger, James Fenske, Michael Greenstone, Johannes Haushofer, Randi Hjalmarsson, SolomonHsiang, Amir Jina, Andreas Kotsadam, Kyle Meng, Edward Miguel, Andreea Mitrut, Hanna Muhlrad, RandallReback, Rajiv Sethi, Carolin Sjoholm, Mans Soderbom, Kristina Svensson, and Arthi Vellore for commentsand support, as well as seminar participants at Barnard College, Columbia University, University of Californiaat Berkeley, University of Chicago, University of Gothenburg and University of Oxford. I am grateful toOxCarre, University of Oxford for access to the mineral data and sponsoring of field work in 2013, and to TheLars Hierta Memorial Foundation for a research scholarship. The results in this paper were previously includedin ”Local Industrial Shocks, Female Empowerment and Infant Health: Evidence from Africa’s Gold MiningIndustry” (2014).

1

1 Introduction

One in nine children born in developing countries dies before its fifth birthday1 and Sub-Saharan Africa has among the highest rates of child mortality in the world. Achieving thesustainable development goals for infant mortality—12 neonatal deaths and 25 deaths underthe age of 5 per 1000 live births—by 2030 poses a significant challenge in the region asseveral countries in West and Central Africa are off-track to meet the target (Wang et al.,2014).

Lack of economic development is one reason for the high child mortality rates: curableand preventable conditions such as lower respiratory infections, diarrheal diseases, malaria(Black et al., 2003; Dupas, 2011), and undernutrition (Black et al., 2013a) contribute to thehigh health burden in developing countries. In fact, maternal and early-life undernutrition isa leading cause of global child mortality, responsible for up to 45% of child deaths (Blacket al., 2013a). Analysis of global data shows that country-wide infant mortality rates increasewith negative aggregate income shocks (Baird et al., 2011).

On the other hand, poor environmental conditions, such as airborne and waterborne pol-lution are threats to infant health in developing countries (Greenstone and Hanna, 2014;Jayachandran, 20092). The low environmental quality and high disease burden in devel-oping countries come from high marginal utility of income: when facing a health-wealthtradeoff, poor households prefer consumption today over investing in environmental quality(Greenstone and Jack, 2015).

I explore if investment in extractive industries, in particular the large-scale gold miningsector, changes local infant mortality rates in Sub-Saharan Africa. The effect is a priori am-biguous: on the one hand it can increase infant mortality rates by polluting the environment,affecting agricultural production (Aragon and Rud, 2015). On the other hand it can reduce itby bringing economic development (Aragon and Rud, 2013), and jobs in manual labor andservices (Kotsadam and Tolonen, 2016). The effect of the health-wealth trade-off on infantmortality in the region is an empirical question. It will depend not only the magnitudes ofthe environmental change and wealth gains but the margins for health improvements. Thehealth-wealth tradeoff generated by a polluting industry is bound to be weaker in contextswith high child mortality due to poverty, and in particular, malnutrition.

1Millennium Development Goals, Child Mortality, The World Bank, 2014. Seehttp://www.worldbank.org/mdgs/child mortality.html)

2There is a larger literature on the health costs of exposure to pollution in developed countries, e.g. Almondet al., 2009; Black et al., 2013b; Chay and Greenstone, 2003; Currie and Schmieder, 2009; Currie et al., 2011and Moretti and Neidell, 2011.

1

01

23

4G

old

prod

uctio

n (m

ton)

050

010

0015

00

1020

3040

50Ac

tive

gold

min

es (c

ount

)

1970 1980 1990 2000 2010

Years

Gol

d pr

ice

(USD

/oz)

Active gold mines (count)

Gold production (mton)

Gold price (USD/oz)

Data from IntierraRMG. Calculations author’s own.

Figure 1: Illustration of the natural experimentNotes: The data in Figure 1A comes from IntierraRMG and the calculations are the author’s own.

The recent gold mining boom in nine African countries—Burkina Faso, Cote d’Ivoire,Democratic Republic of Congo, Ethiopia, Ghana, Guinea, Mali, Senegal, Tanzania, illus-trated in Figure 1—serves as a quasi-experiment to understand how the risk of infant mor-tality changes with local industrial development. Open pit gold mining—the most com-mon form of large-scale gold mining in the region—is capital intensive and dominated bylarge multinational firms (see Table A4) previously not integrated into the local economy(Gajigo et al., 2012). I focus on large-scale gold mining because this part of the miningindustry expanded rapidly during the sample period (Figure 1), has a dominating productiontechnology—open pit mining—in the sample region making its effects plausibly compara-ble across localities, and is less reliant on infrastructure connectivity than other more bulkynatural resources (Weng et al., 2013). For these reasons, I argue that the mine openings areplausibly exogenous to the local economies. Large gold mines often open up in relativelypoor, rural areas: in our sample, the infant mortality rate is higher (with an average of 151 per1000 live births) and the urbanization rate lower in the communities where the gold mineswill eventually open.

I construct a dataset of 37,365 children born within 100 km of a mine by combiningdata on women’s fertility records from Demographic and Health Survey (DHS) and large-scale gold mining data. Combining the two data sources using geographic information atthe village/neighborhood (henceforth called DHS cluster) and mine level, I construct sev-eral measures of proximity to mines. I utilize the exogenous increase in large-scale gold

2

mining to estimate the causal effect of local industrialization by defining treatment and con-trol groups based on proximity measures. Outcomes in the treatment group—children bornclose to mines—are contrasted with children born further away in a before-after analysis. Ishow that pre-mining trends are similar across the treatment and control groups. Importantly,the method flexibly controls for unobservable differences between countries and districts—such as culture, religion and ethnicity—countrywide shocks—e.g., policy or governmentchanges—and temporal trends within subnational districts.

Infant mortality rates decrease with more than 50% of baseline within a few years fromthe start of the industrial gold mining. The effects are concentrated within 10 km from themine center point, corresponding to the area with the starkest changes in local economicgrowth and job creation. Moreover, the results are robust to different assumptions abouttrends, fixed effects and clustering. I find no significant changes in child health care access,but mother’s have better access to fertility information on the radio. Women are also morelikely to work in the service sector. These findings are supported by evidence showing thatlarge-scale gold mining increases women’s empowerment in Sub-Saharan Africa (Tolonen,2016).

Importantly, the results are robust to the exclusion of children born to recent migrants.Excluding all mothers who migrated after the mine opening year, or in the four years priorto the mine opening year, reduces the treatment effect from 7.9 percentage points to 6.8percentage points. This indicates that children born to mothers who have lived in the com-munities for a long time also benefit from the mine opening. This is in line with previousevidence from large-scale mining in Africa, confirming that both women born in the min-ing communities and migrant women benefit from the expansion in mining (Kotsadam andTolonen, 2016; Tolonen, 2016).

The treatment effect of a new large-scale gold mine on local infant mortality is largecompared with countrywide trends in infant mortality rates. The drop in the infant mortalityrate occurring within one to two years from the first year of production is twice as largeas the reduction in the infant mortality rate experienced in Singapore during two decadesof high economic growth, and equivalent to total gains in infant survival rates in the Africansurvey countries since 1970 to today. The results illustrate that industrial development bringssignificant and rapid gains in infant survival rates in high mortality areas.

The paper discusses remainder of the paper is organized as follows. Section 2 presentsthe data and context of gold mining in Sub-Saharan Africa. Section 3 describes the empir-ical strategy. Section 4 presents the results and discusses potential mechanisms. Section 5

3

100km Buffer

DHS Cluster

Country

10km Buffer

Gold mineStudy Country

A

B

Lake Victoria

A

B

Figure 2: Map of Gold Mines and DHS Clusters in Northwestern Tanzania

describes the robustness analysis. Section 6 provides a brief discussion of the magnitude ofthe results in relation to global trends in infant mortality and concludes.

2 Data and Context

The paper creates a unique record of births across mining areas in Africa and extractiveindustry data. Specifically, I combine demographic data from 30 individual nationally repre-sentative surveys conducted in 9 countries. The dataset contains 37,365 births from 1987 to2012 recorded to mothers living within 100 km of an industrial gold mining site. Figure 2Ashows the geographic location of the gold mines used in the analysis, and Figure 2B zoomsin on Tanzania, highlighting mine locations (yellow dots), the survey area (the green circles)and DHS clusters (blue dots).

The Demographic and Health Survey (DHS) collects data on health and fertility in de-veloping countries (see Appendix for more detail). The final data set consists of 30 cross-sectional DHS datasets: four survey rounds each for Burkina Faso, Ghana, Guinea, Mali,and Tanzania, and three survey rounds for Cote d’Ivoire, Ethiopia, and Senegal, and onesurvey round for Democratic Republic of Congo. This corresponds to all the DHS survey

4

rounds that have household location data for countries in which there is at least one large-scale gold mine that was ever active between 1987 and 2012. Appendix Figure A1 shows thesample years for each country (dashed line), and the years for which we have sampled births(shaded grey). All women aged 15-49 in the randomly selected households are selected forthe sample, and the outcome of all pregnancies in the last five years are recorded. Descrip-tive statistics for children aged 0 to 5 as well as for their mothers are presented in Table1. Variable descriptions can be found in Table A2. For certain outcomes such as women’semployment and fertility, I use the women’s recode instead of the child recode. Women whohave given birth in the last five years prior to the survey year are included in the child recode,however, to understand for example the effect of industrial mining on fertility we need toanalyze the full sample of women – also those who had no children.

The large-scale gold mining data comes from IntierraRMG (see Appendix Text) and con-tains all African large-scale gold mines with geographic coordinates and historic productionvolumes. The data has previously been used in research on the effects of large-scale mining(Aragon and Rud, 2015; Kotsadam and Tolonen, 2016; von der Goltz and Barnwal, 2014).The geo-coordinates provided by the data have been updated to correspond to the mine centerpoint (see Appendix for further details).

The data set contains only large-scale mining operations, most of them owned by foreignowned companies (see Table A4). The context is highly relevant: extractive industries re-ceive a large share of total foreign direct investment. In particular, large-scale gold mininghas rapidly expanded across African countries: Africa currently produces 20% of the worldproduction of gold and 34 countries in Africa have significant gold deposits that could beextracted in large-scale operations in the future (Gajigo et al., 2012).

5

Table 1: Extensive Summary Statistics

(1) (2) (3) (4) (5) (6) (7)

Mean value Min Max

Sample whole control treatment control treatment

sample group group group group

Time period pre pre post post

Characteristics

mother age 29.17 29.54 28.86* 28.73 28.27 15 49

mother education 1.88 1.333 1.876* 2.483 4.880 0 21

household is urban 0.187 0.194 0.044* 0.181 0.236 0 1

child is male 0.506 0.506 0.512 0.506 0.550 0 1

birth number 3.976 3.996 4.290* 3.96 3.676 1 17

birth year 2000 1999 1997* 2003 2003 1987 2012

Infant mortality

1 month 0.038 0.042 0.066* 0.035 0.024 0 1

6 months 0.057 0.070 0.108* 0.057 0.030 0 1

12 months 0.097 0.106 0.151* 0.091 0.048 0 1

12 months boys 0.103 0.110 0.150* 0.097 0.063 0 1

12 months girls 0.093 0.102 0.154* 0.085 0.030 0 1

Treatment variables

industrial*mine dep. 0.010 0 1

mine dep. (10km) 0.023 0 1

Observations 48,151 25,583 639 21,438 491

Notes: Control group is within 10-100 km from a mine

Treatment group is 0-10 km from a mine

pre-treatment, control group has mine = 0, and no active mine within 100 km

pre-treatment, treatment group has mine = 1, but no active mine within 10 km

post-treatment, control group has mine = 0, and at least 1 active mine within 100 km

post-treatment group has mine = 1, and at least 1 active mine within 10 km

* p < 0.05 for t-test between control group (2) and treatment group (3), pre-treatment

6

The industrial mining database only contains information on large-scale gold operations.Artisanal and small-scale mining (ASM) may be a confounding factor but the lack of de-tailed, time-varying records of legal and illegal ASM activities makes it impossible to dis-entangle the two sectors. In some instances, artisanal and small-scale mining is part of theland use prior to the establishment of a large-scale mine. The establishment of a large minemay (1) crowd out small-scale activities through enforcing property rights, (2) not affect theASM sector, especially if the small-scale mining is illegal and property rights are not en-forced, (3) increase ASM activities if the latter uses the scrap material from the large-scalemine. It is, to my knowledge, unknown which one of these three scenarios is the most com-mon. Small-scale gold mining is associated with mercury pollution—the mercury is usedin the traditional amalgamation process to separate the gold from the ore—which can be athreat to fetal development and child health. It is possible that large-scale gold mining couldimprove local child health by crowding out the small-scale sector, if the latter is associatedwith more adverse health effects.

The expansion of the large-scale mining industry, however, also causes concern aboutenvironmental safety and sustainability (Economic Commission for Africa, 2011). The pol-lution burden from gold mining is mainly cyanide used in the amalgamation process to sepa-rate the gold from the ore, as well as arsenic found naturally in the gold ore, and other heavymetals such as lead, cadmium, chromium and nickel. Cyanide and arsenic exposure in utero,both lethal at high doses, are associated with adverse birth outcomes (Milton et al., 2005),and lead exposure with increased risk of premature birth, low birth weight, and retardedgrowth (Iyengar and Nair, 2000).

Despite these concerns, estimated pollution levels in mining areas are generally withinWHO thresholds, and the literature is inconclusive regarding the health effects below theWHO thresholds (ATSDR, 2007b; ATSDR, 2007a). However, arsenic levels above the WHOthresholds have been confirmed in mining areas in Tanzania and Ghana (Almas and Manoko,2012; Serfor-Armah et al., 2006), for example after mining accidents. There is suggestiveevidence that pollution from mining activities reduces health: lead pollution from miningactivities has been linked to stunting in children (von der Goltz and Barnwal, 2014), and across-sectional study in Spain found higher incidences of cancer in the population within 5km from mining sites (Fernandez-Navarro et al., 2012).

This study does not expand much beyond exploring infant mortality rates. There aretwo reasons for this. First, the DHS data contains a limited number of reliable child healthindicators, such as cough, diarrhea and fever, and they are only collected in the sample year.

7

Second, because the analysis is conducted on countries with high infant mortality rates, Iargue that exploring effects of industrial development on infant mortality is justifiable andof first-order importance. It remains plausible that more children survive their first birthday,but face poorer long run health.

3 Empirical Strategy

The strategy of the paper follows Kotsadam and Tolonen (2016)3 who estimate local em-ployment effects from industrial mining in Africa, and links to the field of economic geogra-phy measuring agglomeration economies, e.g., local multipliers (Moretti, 2010), and healthnear industrial sites (Currie et al., 2011; Currie et al., 2015). The empirical strategy is adifference-in-difference analysis, comparing outcomes before and after in a treatment and acontrol group.

Treatment is based on proximity to mining site interacted with activity measures of themine. Previous literature on mining (Aragon and Rud, 2013; Aragon and Rud, 2015; Kot-sadam and Tolonen, 2016; von der Goltz and Barnwal, 2014) and commuting behavior inTanzania, Ghana and Cote d’Ivoire (Shafer, 2000; Amoh-Gyimah and Aidoo, 2013; Kunget al., 2014), indicate that mine treatment effects should be concentrated to communitieswithin 5-20 km from the mine. Kotsadam and Tolonen (2016) find the strongest treatmenteffects within 10 km from the mine center point, but use a 20 km distance in their baselinespecification. DHS household coordinates are displaced by 1 to 5 km, and up to 10 km in1% of the cases to ensure that individuals cannot be identified, and DHS recommends us-ing binned thresholds larger than 5 kilometers. The baseline minimum distance used in thispaper is 10 km. A contribution of the present paper is its empirical approach to estimat-ing distance effects in spatial analyses where a radius of influence is not known a priori bycarefully mapping the spatial decay function with binned distance thresholds (equation 2) todetermine the treatment distance.

The main specification is the following:

In f antmortalityicdt = b0 +b1industrialct ·minedepositc +b2minedepositc

+ad +sdtrend + dkt +Xi + eicdt (1)

where the outcome, infant mortality in the first 12 months, is regressed on an indicator for3With a trivial modification to the variable specification to facilitate the interpretation of the coefficients.

8

if there is a known gold deposit (known by 2012) within 10 km (called mine deposit), andan interaction effect capturing if this mine deposit was in production in the year of the birth(industrial*mine deposit). The regression controls for sub-national district fixed effects ad ,district linear time trends sdtrend , and country-year fixed effects (birth year or survey year),dkt , and a vector of mother and birth controls, Xi. Subscript i indicates an individual birth,c DHS cluster, k country, and t year. Individual level controls include mother’s age, agesquared, education, if the household lives in an urban area, and child’s birth order. In allregressions, I have limited the sample to within 100 km from a deposit, and I cluster thestandard errors at the DHS cluster level (unless otherwise stated).

To allow for non-linear effects with distance and better understand the geographic distri-bution of effects, I implement a spatial lag model. This specification will map the geograph-ical distribution of effects across a 100 km plane from the mine center point. This strategywill inform about the usefulness of the chosen threshold distance in the baseline strategy (10km). I allow for two sets of non-linear spatial structures:

Yicdt = b0 +Âs

bd industrialct ·minedepositc +Âs

bd minedepositc (2)

+ad +sdtrend + dkt +Xi + eicdt

for s 2 {0�10, 10�20, . . . ,60�70,70�100}

This spatial lag model allows for non-linear effects with distance from the mine. Eachbirth is recorded to a distance bin—0-10 km, 10-20 km, etc.—and compared with the refer-ence category 70-100 km away. The specification controls for the same fixed effects, trendsand individual level controls as the baseline specification.

To estimate the causal effect of an industrial-scale gold mine opening on infant mortal-ity, the timing and placement of the opening must be exogenous to local changes in infantmortality. The necessary condition for a gold mine is a gold deposit, which is a geologi-cal anomaly and random (Eggert, 2002). Moreover, multinational gold mining firms preferto invest in regions with low corruption, and stable and transparent business environments(Tole and Koop, 2011). Such factors could vary at the national, province, or at most, districtlevel and should pose no problem for the estimation strategy as it controls for unobserveddifferences between districts (by using district fixed effects), national-level policy changesor other unobserved factors changing annually (by using country-year fixed effects), anddistrict-specific linear time trends.

Access to infrastructure is one factor that varies within administrative districts. High-

9

volume, bulky resources such as coal and iron are heavily reliant on good infrastructure—railways, road networks and ports—but the high-value commodity gold is less dependenton regional infrastructure as it can be transported using air traffic (Weng et al., 2013).

Parallel trends

Both empirical strategies rely on the assumption that in absence of the mine opening, thetreatment and the control communities would have been on similar trajectories. The validityof this assumption is examined in Figure 3, and appendix Figures A2, A3, and A4.

The mining communities and non-mining communities were on similar trends in infantmortality in the years before the mine openings, which is confirmed using linear trend pre-diction in Figure 3A4 and using local polynomial smooth in Figure 3B. Both specificationsallow for a trend break at year -2—the start of the investment phase—of the mine lifetime,in contrast to Figure 3C which shows one smoothed trend. In the Appendix I explore therobustness of these results. I redefine the investment period to one year, allowing for a breakin year -1 (Figure A2B), or at year 0 (Figure A2C) which changes the slopes of the pre-trendsso they are not parallel. I conclude that the mine starts affecting local infant mortality ratesat year -2 of the mine lifetime.

The incidence of infant mortality is higher in communities with a future mine. Thesecommunities—located within 10 km of a mine—start seeing reductions in mortality ratesduring the investment period which is the years immediately preceding the mine openingyear, whereas communities 10-100 km away continue along the same trend. I will provideregression results with and without the investment period: both methods yield similar results.In general, the effects are weaker if the investment period is included in the control groupsample, as expected. An alternative specification would be to include the investment periodin the treatment period. However, the investment period is not associated with the samepollution risk as the production phase but may have significant economic effects. If so,including it in the treatment period could lead to an underestimation of the pollution channel.

Figure A3 explores the same pattern for a sample which excludes migrants. I excludeall migrants that moved to the community after the mine opening year, or in the four yearsprior to the mine opening year. I do this to exclude anticipatory migration that may lead toselection bias. The patterns observed when excluding migrants are similar to those in thefull sample. Row A—which excludes the two year investment phase—provides the most

4A t-test performed confirmed that the pre-mine opening slopes are not significantly different from eachother.

10

suitable quasi-experimental setting. The linear trends are not perfectly parallel, but the localpolynomial smoothing allows us to see that the two groups are tracking each other over time,especially in the last few years before mine opening.

The same figures are produced for boys (A4A), although the mortality rate seems to beincreasing in the treatment group pre-treatment, and girls (A4B) where it was slightly de-creasing in the treatment group, pre-treatment in the linear specification. This may be drivenby the unbalanced sample, and that there are no controls for country, district or calendar year,which is supported by the fact that the downward sloping trend for girls is hard to observewhen using local polynomial smooth.

The trends for infant mortality in the first 6 months of life (A4C) are parallel for thetreatment and control group, with a clear jump in the treatment group post-treatment. Theevidence is less convincing for neonatal mortality (first month of life) using linear trends(A4D), possibly because of something happening in the treatment group in the years -10 to-8. The polynomial trends look, however, parallel in the five years before the mine opening.

An aternative investigation of the parallel trends assumption is using night lights (Figure4). The graphs show a stark deviation from the trend in year -2 of the mine lifetime, both inthe linear specification (A) and using local polynomial smooth (B).

4 Results and Mechanisms

Table 2 shows the regression results for infant mortality using a 10 km distance cutoff: mineopening is associated with a 5.5 percentage point decrease in mortality rates (column 1).Taking potential spillovers into account by excluding the 10-30 km area, and the investmentphase (two years prior to mine opening), the coefficient is estimated at 7.9 percentage points(column 2). Infant mortality decreases for boys (6.3 percentage points), and girls (9.5 per-centage points). The effects are economically significant: the 5.5 percentage point decreasein mortality is equivalent to 50% decrease in mortality rates, compared with the samplemean.

Selective fertility— There is a noted association between fertility and infant mortality. Onthe one hand, a reduction in fertility can lead to higher investment per child, and improvedchild survival rates. On the other hand, high infant mortality rates can motivate high fertility,ensuring a minimum number of surviving children. The drop in fertility rate (-0.217 childrenper woman, see Table 2) is, however, insignificant and too small to explain the whole drop inthe infant mortality rate: an accounting exercise shows that the counterfactual mortality rate

11

0.05.1.15.2.25

-10 -5 0 5 10

0.05.1.15.2.25

-10 -5 0 5 10

0.05.1.15.2.25

-10 -5 0 5 10Year since mine opening

Infa

nt m

orta

lity

rate

Investment periodMine opening

Control

Treatment

Linear trend prediction: drop year -2 to 0

Control

Treatment

Investment periodMine opening

Control

Treatment

Control

Treatment

Control

Treatment

Control

Treatment

Kernel weighted local polynomial smooth: drop year -2 to 0

Kernel weighted local polynomial smooth

A

B

C

Infa

nt m

orta

lity

rate

Infa

nt m

orta

lity

rate

Figure 3: Linear trend (A), local polynomial smooth (B), and local polynomialsmooth (C) for infant mortality in first 12 months

Notes: Years since mine opening is on the horizontal axis, ranging from ten years before mineopening to ten years after mine opening. The treatment group is drawn within 10 km from theclosest mine, and the control group10-100 km from the closest mine, excluding births with a secondmine within 20 km. Figure A and B allow for a break at year -2. Figure C provides 95% confidenceintervals. In contrast to the main specification there are no control variables or fixed effects, and onlythe opening year of the closest mine is considered. Mine closing year is not considered.

12

Tabl

e2:

Mai

nR

esul

ts

Dep

ende

ntva

riab

le:

infa

ntm

orta

lity

first

12m

onth

sto

tal

liste

nto

radi

ose

rvic

efe

rtilit

yfa

mily

plan

ning

sect

orjo

bSa

mpl

e:ch

ildre

nch

ildre

ngi

rlsbo

ysw

oman

wom

anw

oman

drop

spill

over

(1)

(2)

(3)

(4)

(5)

(6)

(7)

indu

stria

l*m

ine

depo

sit(

atbi

rth)

-0.0

55**

*-0

.079

***

-0.0

95**

*-0

.063

**(0

.019

)(0

.026

)(0

.036

)(0

.032

)in

dust

rial*

min

ede

posi

t(at

surv

ey)

-0.2

170.

147*

**0.

064*

*(0

.141

)(0

.046

)(0

.024

)m

ine

depo

sit(

with

in10

km)

0.03

4**

0.04

8**

0.06

4**

0.03

20.

107

-0.0

70-0

.036

(0.0

17)

(0.0

24)

(0.0

29)

(0.0

28)

(0.1

05)

(0.0

46)

(0.0

24)

Cou

ntry

-birt

hye

arFE

Yes

Yes

Yes

Yes

No

No

No

Cou

ntry

-sur

vey

year

FEN

oN

oN

oN

oYe

sYe

sYe

sD

istri

ctlin

eart

rend

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Dro

p10

-30

kmaw

ayN

oYe

sYe

sYe

sN

oN

oN

oD

rop

inve

stm

entp

hase

No

Yes

Yes

Yes

No

No

No

Mea

nof

outc

ome

0.09

70.

102

0.09

80.

105

3.26

00.

470

0.23

3M

ean

(trea

tmen

t,pr

e-tre

atm

ent)

0.15

10.

151

0.15

00.

154

3.90

50.

354

0.12

0O

bser

vatio

ns37

,365

29,2

2114

,863

14,3

5857

,581

46,0

2855

,944

R-s

quar

ed0.

101

0.10

40.

119

0.12

30.

673

0.20

40.

183

Not

es:*

**p<

0.01

,**

p<

0.05

,*p<

0.1

Clu

ster

edst

anda

rder

rors

clus

tere

dat

DH

Scl

uste

rlev

el.L

inea

rpro

babi

lity

mod

els.

All

regr

essi

ons

inco

lum

n1

to5

cont

rolf

orm

othe

r’sag

e,ag

esq

uare

,mot

her’s

educ

atio

n,ur

ban,

child

’sbi

rthnu

mbe

r,an

dfix

edef

fect

sfo

rdi

stric

t,bi

rthm

onth

,and

coun

try-b

irth

year

.Out

com

eis

infa

ntm

orta

lity

inth

efir

st12

mon

thss

ince

birth

.All

regr

essi

onsc

olum

ns5-

7co

ntro

lfor

wom

an’s

age,

educ

atio

n,ur

ban,

and

fixed

effe

cts

ford

istri

ctan

dco

untry

-sur

vey

year

,as

wel

las

dist

rictl

inea

rtim

etre

nd.M

ine

depo

sitc

aptu

res

ifth

ere

isa

gold

depo

sitw

ithin

10km

from

the

hous

ehol

d.In

dust

rialc

aptu

res

ifth

ego

ldde

posi

twas

activ

ely

extra

cted

inth

ech

ild’s

birth

year

,ori

nth

esu

rvey

year

.The

inve

stm

entp

hase

isde

fined

asth

etw

oye

ars

prec

edin

gth

eop

enin

gye

aran

dis

drop

ped

inco

lum

ns2-

4.M

ean

(trea

tmen

t,pr

e-tre

atm

ent)

isth

esa

mpl

em

ean

fort

hetre

atm

entg

roup

befo

reth

em

ines

wer

eac

tive.

13

would need to exceed 80% among the children that were not born post treatment to explainthe whole reduction in mortality5.

Women’s employment— Women’s employment and access to information about familyplanning may be important determinants of infant health. Results in Table 2 confirm thatwomen in mining communities are 14.7 percentage points more likely to listen to radioshows discussing family planning than their peers, and 6.4 percentage points more likely towork in the service sector. This result is supported by previous research on African min-ing (Kotsadam and Tolonen, 2016; Tolonen, 2016) finding that women gain jobs in servicesectors in mining areas and are more empowered.

Access to health care— Corporate social responsibility programs from the mining com-panies and the change in local economic growth could change access to health care and earlyhealth-behavior. I consider this in Table 3, which shows that there are no significant changesin health care behavior, such as received any antenatal care (column 1), has health card (col-umn 4), was ever vaccinated (column 5), nor were the children less likely to be very smallat birth—a proxy measure for inadequate in utero growth or preterm birth—or have cough,diarrhea or fever in the last 2 weeks before surveying.

Local economic development— The results on infant mortality are supported by similarnon-parametric analysis of night light density (Figure 4), a proxy for economic development.Night lights have been used as a proxy for urbanization and economic performance in Africa—cf. Michalopoulos and Papaioannou (2013)— although caution is advisable as the proxymay be less reliable in areas with low level of economic development. Additionally, themines may omit lights at night due to around the clock operations. The extent to which mea-sures of economic growth from mining would suffer from this bias is not, to my knowledge,known. The figure illustrates that mining areas (within 10 km) are on a similar night lighttrend as areas further away (30-100 km from the mine location) in the pre-treatment period,although at a higher level. However, starting in the investment phase (mine year -2) there isa clear divergence with average light sharply increasing in mining areas and it stabilizes at ahigher level. There is a drop in night lights at year 7, which could be caused by some minesclosing down. Such mine closures are controlled for in the regression analysis.

5There are 3.260 births per woman, so 30.7 women give birth to 100 children. The regression results showthat 0.217*30.675 (6.656) fewer children will be born. The counterfactual mortality would need be 5.5 childrenper 6.656 children born, above 80%, for the drop in fertility to account for the whole drop in mortality rate.

14

Tabl

e3:

Alte

rnat

ive

mec

hani

sms

Dep

ende

ntva

riab

le:

rece

ived

very

was

has

ever

coug

hdi

arrh

eafe

ver

birth

ante

nata

lsm

all

brea

stfe

dhe

alth

vacc

inat

edla

stla

stla

stor

der

care

atbi

rthm

onth

sca

rd2

wee

ks2

wee

ks2

wee

ks

Sam

ple:

all

all

all

all

all

surv

ivin

gsu

rviv

ing

surv

ivin

gal

lbi

rths

birth

sbi

rths

birth

sbi

rths

child

ren

child

ren

child

ren

birth

s(1

)(2

)(3

)(4

)(5

)(6

)(7

)(8

)(9

)

indu

stria

l*m

ine

depo

sit

0.02

2-0

.032

0.29

0-0

.004

0.00

4-0

.050

-0.0

19-0

.024

-0.2

06(0

.075

)(0

.022

)(0

.601

)(0

.047

)(0

.105

)(0

.034

)(0

.027

)(0

.042

)(0

.165

)m

ine

depo

sit

-0.0

280.

012

-0.0

31-0

.059

-0.1

20-0

.032

-0.0

15-0

.027

0.16

7(0

.062

)(0

.021

)(0

.504

)(0

.051

)(0

.074

)(0

.039

)(0

.026

)(0

.046

)(0

.117

)

Birt

hm

onth

FEYe

sYe

sYe

sYe

sYe

sN

oN

oN

oYe

sD

istri

ctFE

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Cou

ntry

-birt

hye

arFE

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Cou

ntry

-sur

vey

year

FEN

oN

oN

oN

oN

oYe

sYe

sYe

sN

oD

istri

cttim

etre

ndYe

sYe

sYe

sYe

sYe

sN

oN

oN

oYe

sD

istri

ctfix

edef

fect

No

No

No

No

No

Yes

Yes

Yes

No

Dro

p10

-30

kmaw

ayN

oN

oN

oYe

sYe

sYe

sYe

sYe

sYe

sD

rop

inve

stm

entp

hase

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Mea

nof

outc

ome

0.78

0.05

16.3

0.81

0.69

0.19

0.17

0.29

3.97

Obs

erva

tions

28,5

8835

,510

30,6

3132

,613

12,6

9530

,475

31,3

3133

,595

38,4

14R

-squ

ared

0.32

20.

065

0.63

50.

214

0.31

70.

073

0.07

20.

095

0.72

9

Not

es:*

**p<

0.01

,**

p<

0.05

,*p<

0.1

Clu

ster

edst

anda

rder

rors

clus

tere

dat

DH

Scl

uste

rlev

el.A

llre

gres

sion

sco

ntro

lfor

mot

her’s

age,

age

squa

re,

mot

her’s

educ

atio

n,ur

ban,

child

’sbi

rthor

der(

notc

olum

n9,

whe

reit

isth

eou

tcom

e),a

ndfix

edef

fect

sfo

rbirt

hm

onth

,dis

trict

,and

coun

try-b

irth

year

,as

wel

las

linea

rdis

trict

time

trend

s.C

olum

ns1-

5,an

dco

lum

n9

uses

the

sam

ple

ofal

lbirt

hsin

the

last

five

year

s,an

dco

lum

ns6-

8is

run

onth

esa

mpl

eof

child

ren

aliv

eat

the

time

ofth

esu

rvey

.Out

com

eva

riabl

esar

e(1

)ant

enat

alca

rere

ceiv

ed,(

2)ch

ildw

asve

rysm

alla

tbirt

h,(3

)the

child

was

brea

stfe

din

mon

ths,

capp

edat

12,a

ndsa

mpl

elim

ited

toch

ildre

nw

hosu

rviv

edto

thei

rfirs

tbirt

hday

,(4)

the

mot

herh

asa

heal

thca

reca

rdfo

rthe

child

,(5)

child

was

ever

vacc

inat

ed,t

hech

ildha

dco

ugh

(6),

diar

rhea

(7),

feve

r(8)

inth

ela

sttw

ow

eeks

,and

the

child

’sbi

rthor

der(

9).

15

Tabl

e4:

Sens

itivi

tyA

naly

sis

infa

ntin

fant

infa

ntm

orta

lity

mor

talit

ym

orta

lity

Dep

ende

ntva

riab

le:

first

12m

onth

s1

mon

th6

mon

ths

Spec

ifica

tion:

noco

untry

-di

stric

tdr

opdi

str.

min

em

ine

base

line

base

line

FEye

arFE

trend

spill

over

clus

ter.

clus

ter.

FE(1

)(2

)(3

)(4

)(5

)(6

)(7

)(8

)(9

)

indu

stria

l*m

ine

depo

sit

-0.0

60**

*-0

.060

***

-0.0

55**

*-0

.079

***

-0.0

79**

*-0

.079

***

-0.0

84**

*-0

.028

**-0

.062

***

(0.0

20)

(0.0

19)

(0.0

19)

(0.0

26)

(0.0

26)

(0.0

27)

(0.0

26)

(0.0

14)

(0.0

19)

min

ede

posi

t0.

039*

*0.

036*

*0.

034*

*0.

048*

*0.

048*

0.04

80.

051*

*0.

015

0.03

4(0

.016

)(0

.016

)(0

.017

)(0

.024

)(0

.029

)(0

.029

)(0

.024

)(0

.014

)(0

.018

)

Birt

hm

onth

,yea

rFE

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Dis

trict

FEYe

sYe

sYe

sYe

sYe

sYe

sYe

sYe

sYe

sC

ount

ry-b

irth

year

FEN

oYe

sYe

sYe

sYe

sYe

sYe

sYe

sYe

sD

istri

cttim

etre

ndN

oN

oYe

sYe

sYe

sYe

sYe

sYe

sYe

s

Dro

p10

-30

kmaw

ayN

oN

oN

oYe

sYe

sYe

sYe

sYe

sYe

sD

rop

inve

stm

entp

hase

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Dis

trict

clus

terin

gN

oN

oN

oN

oYe

sN

oN

oN

oN

oM

ine

clus

terin

gN

oN

oN

oN

oN

oYe

sN

oN

oN

oM

ine

FEN

oN

oN

oN

oN

oN

oYe

sN

oN

o

Mea

nof

outc

ome

0.09

70.

097

0.09

70.

102

0.10

20.

102

0.10

20.

038

0.05

7M

ean

(trea

tmen

t,pr

e-)

0.15

10.

151

0.15

10.

151

0.15

10.

151

0.15

10.

066

0.10

8O

bser

vatio

ns37

,365

37,3

6537

,365

29,2

2129

,221

29,2

2129

,221

38,4

1434

,016

R-s

quar

ed0.

038

0.09

90.

101

0.10

40.

104

0.10

40.

105

0.02

50.

046

Not

es:*

**p<

0.01

,**

p<

0.05

,*p<

0.1

Clu

ster

edst

anda

rder

rors

clus

tere

dat

DH

Scl

uste

rlev

el,e

xcep

tcol

umn

5an

d6

whi

chcl

uste

rsat

the

dist

rict

leve

land

min

ele

velr

espe

ctiv

ely.

All

regr

essi

ons

cont

rolf

orm

othe

r’sag

e,ag

esq

uare

,mot

her’s

educ

atio

n,ur

ban,

child

’sbi

rthnu

mbe

rand

birth

mon

th.

Col

umn

2co

ntro

lsfo

rcou

ntry

-birt

hye

arfix

edef

fect

s,an

dco

lum

n3

adds

dist

rictl

inea

rtim

etre

nd,c

olum

n4

drop

sth

ein

vest

men

tpha

sean

din

divi

dual

sliv

ing

10-3

0km

away

.Col

umn

7ad

dsm

ine

fixed

effe

ct.O

utco

mes

are

infa

ntm

orta

lity

inth

efir

st12

mon

ths

(col

umns

1-7

,one

mon

th(c

olum

n8)

orsi

xm

onth

sof

life

(col

umn

9),c

ondi

tiona

lon

live

birth

.Min

ede

posi

tcap

ture

sif

ther

eis

ago

ldde

posi

twith

in10

kmfr

omth

eho

useh

old.

Indu

stria

lcap

ture

sif

the

gold

depo

sitw

asac

tivel

yex

tract

edin

the

child

’sbi

rthye

ar.T

hein

vest

men

tpha

seis

defin

edas

the

two

year

spr

eced

ing

open

ing

year

.Mea

n(tr

eatm

ent,

pre-

treat

)sho

ws

the

mea

nva

lue

inth

etre

atm

entg

roup

,bef

ore

treat

men

t.

16

01

23

4

-10 -5 0 5 10

01

23

4

-10 -5 0 5 10

Nig

ht li

ghts

(mea

n pe

r are

a)

year since mine opening

Investment periodMine opening

Treatment

Control

year since mine opening

Investment periodMine opening

Treatment

Control

A B

Figure 4: Night lightsNotes: The figure shows a twoway line plot (A) and kernel-weighted local polynomial smoothestimation (B) of night lights in mining areas (within 10 km from a mine) compared with the controlarea 30-100 km away from a large-scale gold mine.

5 Sensitivity analysis

Model sensitivity tests are found in Table 4. From a parsimonious specification, the modelis modified sequentially to allow for country-birth year fixed effects (column 2), district-linear time trends (column 3), spillover effects (column 4), as well as clustering of standarderrors at the district level (column 5), and mine level (column 6), and closest-mine fixedeffects (column 7). A limitation of the specifications using mine fixed effects is that if abirth happened near several gold mines, the fixed effect will only control for the nearest goldmine. The regression results are stable both in magnitudes and significance levels across thespecifications.

Neonatal mortality— Table 4 columns 8 and 9 provide estimates using the baseline spec-ification for infant mortality measures in the first month, and first 6 months as outcomevariables. The sample sizes differ across the regressions for 12, 6, and 1 month. This isbecause a birth is only kept in the sample if the child was at least 12, 6, or 1 month at thetime of the interview. A birth that happened 8 months before the interview will be includedin the sample for infant mortality 1 month and infant mortality 6 months, but not for infantmortality 12 months. The sample sizes are 38,414 for neonatal mortality (1 month), 34,016(for 6 months), and 29,221 (12 months). Note that this sample size is smaller than the samplesizes columns 1 to 3 that does not use the spillover specification introduced in column 4.

The mean values are also different: average infant mortality rate in the first month is5.7%, in the first 6 months it has increased to 10.8% of all births, and in the first year it

17

increases to 15% of births. The mean values indicate that the neonatal period — just afterbirth up to the first month — is a crucial time for a child. The result (column 8) indicates thatmine opening reduces neonatal mortality by 2.8 percentage points, or 42%. The mortalityrisk decreases also at later stages, indicated by the 6.2 percentage points (or 57%) reductionin mortality in the first 6 months (column 9).

Selective migration— Selective inward migration of women with better child survival is apotential concern. Removing women who migrated during the investment phase (defined asthe two years preceding the mine opening year) or later from the sample marginally reducesthe estimated effect to 7.3 percentage points (Table 5 column 3, compared with the mainresult in Table 2 column 2 of 7.9 percentage points). The effect is still economically andstatistically significant. However, it is unclear from what year women may start moving to amining area in anticipation of it opening, so I also drop earlier migrants in Table 5. Excludingall women who ever migrated is overly conservative: the majority of women migrants indeveloping countries migrate for marriage (Rosenzweig and Stark, 1989). Excluding latemigrants, however, reduces the concern that selective inward migration of women drivesthe result. Excluding women who migrated one year before the first year of productionof the mine, or later, reduces the coefficient size marginally (7.7 percentage points, Table 5column 2). Reducing the sample size by dropping earlier migrants (columns 3-6) reduces thecoefficient size further, although the magnitudes are still large (5.3 to 7.3 percentage points)and significant (except for column 6). Women who are born in the mining communities,or who moved to the communities several years before the mine opening year thus alsoexperience a drop in infant mortality risk.

The analysis, however, does not solve a potential issue of selective outward migration.Women who have children with worse survival chances could move out from mining areas asa response to the start of gold mining, which would reduce the incidence of infant mortality.DHS does not collect data on place of origin, so it is not possible to test this hypothesis.However, no negative change in infant health is observed in the control communities, whichcontinue along the same trend (Figure 3A-C). Such an hypothesis is also contradicted byregression results in Figure 6A, which detects no deterioration in the distance bins furtheraway from the mines.

Mother fixed effects— Mother fixed effects—previously used by Kudamatsu (2012)—can reduce the concern that the effect is driven by selective inward migration. I use motherfixed effects in Table A1. The mother fixed effect sharply reduces the sample size by im-

18

Table 5: Anticipatory migration

Dependent variable: Infant mortality (12 months)(1) (2) (3) (4) (5) (6)

industrial*mine deposit (at birth) -0.079*** -0.077** -0.073** -0.070** -0.068* -0.053(0.026) (0.033) (0.034) (0.035) (0.037) (0.035)

mine deposit (within 10 km) 0.048** 0.043 0.043 0.043 0.045* 0.039(0.024) (0.027) (0.027) (0.027) (0.027) (0.025)

Sample limit:migrated [...] yearsbefore mine opening - 1 2 3 4 5

Birth month FE Yes Yes Yes Yes Yes YesDistrict FE Yes Yes Yes Yes Yes YesCountry-birth year FE Yes Yes Yes Yes Yes YesDistrict time trend Yes Yes Yes Yes Yes YesDrop 10-30 km away Yes Yes Yes Yes Yes YesDrop investment phase Yes Yes Yes Yes Yes Yes

Observations 29,221 20,630 20,623 20,613 20,605 20,590R-squared 0.104 0.104 0.104 0.104 0.104 0.103

Notes: *** p < 0.01, ** p < 0.05, * p < 0.1. Clustered standard errors clustered at DHS cluster level. Allregressions control for mother’s age, age square, mother’s education, urban, child’s birth number and birthmonth and fixed effects for country-birth year and district linear time trend. The regressions exclude migrants.Column 2 exclude migrants who migrated one year before the mine’s first active year or later. Column 6exclude migrants who migrated 5 years before the first mine opening year or later.

19

posing that a woman has more than two births in the last five years before surveying, andthat at least one birth happens on either side of the mine opening year. In total, there are 94women in the sample who fulfill these criteria. 368 women living within 10 km from a minedeposit have more than one birth, but do not necessarily experience the mine opening duringthe time period. In total, there are fifteen thousand women who have more than one birth.The magnitude of the results using mother fixed effects are comparable to the main results,although statistically insignificant. However, a power calculation indicated that this samplesize is insufficient to detect changes in an event as rare as infant mortality.

Timing of effects— The economic effects and pollution effects need not coincide: thepollution burden might build up over time mitigating the positive effects on infant mortalitystemming from economic growth. We find no evidence for this hypothesis by splitting thedata into time bins. The coefficients for infant mortality are negative for all years subsequentto mine opening (years 1-2, 3-4, 5-6, and so on up to year 15-16) (see Figure 5A), althoughsome coefficients are not significant. Figure 5B allows for the effect on lifetime fertility tochange over time. If a new mine opens, it might take a few years to adjust fertility preferencesand outcomes. However, there is little evidence of the effect of the mine on fertility varyingover time. The coefficients remain stable, negative and for the most part insignificant, withthe exception of years 5-6 and 13-14.

As an alternative specification to explore timing of the reduction in infant mortality, thesample is limited by the number of years active (Table 6). In this strategy, I compare thecontrol group to children born within 2 years of mine opening (column 1) up to within thefirst 6 years since mine opening (column 6). There are only slight changes in coefficient sizeand it peaks with specification 4, which allows for the effect to happen during the five yearsafter mine opening. The exercise illustrates that the drop in infant mortality occurs alongsideeconomic growth from the mine opening, and remains lower during the mine’s productivephase. We can exclude the possible hypothesis that the reduction in infant mortality is offsetduring the mine lifetime as the pollution levels build up.

20

Table 6: Timing of Effects

Dependent variable: infant mortality first 12 months(1) (2) (3) (4) (5) (6)

industrial*mine -0.061** -0.067*** -0.070*** -0.093*** -0.061* -0.068***(0.030) (0.024) (0.024) (0.026) (0.035) (0.024)

mine deposit 0.036 0.044** 0.046** 0.068*** 0.038 0.047**(0.027) (0.021) (0.021) (0.024) (0.035) (0.022)

Sample limit:years after mine opening 2 3 4 5 6 10

Birth month FE Yes Yes Yes Yes Yes YesDistrict FE Yes Yes Yes Yes Yes YesCountry-birth year FE Yes Yes Yes Yes Yes YesDistrict time trend Yes Yes Yes Yes Yes YesDrop 10-30 km away Yes Yes Yes Yes Yes YesDrop investment phase Yes Yes Yes Yes Yes Yes

Observations 22,197 23,722 24,702 25,420 26,517 31,601R-squared 0.112 0.110 0.109 0.118 0.120 0.108

Notes: *** p < 0.01, ** p < 0.05, * p < 0.1. Clustered standard errors clustered at DHS cluster level. Allregressions control for mother’s age, age square, mother’s education, urban, child’s birth number and birthmonth and fixed effects for country-birth year and district linear time trend. The different regressions limit thetreatment years to first 2 years, 3, 4, 5, 6, or 10 years after the first mine opening year. It imposes no restrictionon the pre-opening time period.

21

-1 -.5 0 .5-.3 -.2 -.1 0 .1

Infant mortality FertilityMine lifetime

years 1-2

years 3-4

years 5-6

years 7-8

years 9-10

years 11-12

years 13-14

years 15-16

Mine lifetime

years 1-2

years 3-4

years 5-6

years 7-8

years 9-10

years 11-12

years 13-14

years 15-16

A B

Coefficient size Coefficient size

Figure 5: Infant Mortality and Fertility by Mine Lifetime

Notes: Each graph shows results from 8 different regressions, separate for infant mortality in the first12 months (A) and women’s lifetime fertility (B). The specification in A is the baseline specificationwhich does not exclude spillovers. The vertical axis shows the sample used in the treatment group:births happening within the first two years after mine opening (1-2), in year 3 and 4 (3-4), in year 5and 6 and so on up to mine life year 15-16. The horizontal axis shows the coefficient estimate and95% confidence intervals. The control group remains the same in all regressions. Figure B shows theregression results using the same specification but for the women’s sample and with lifetime fertilityas the outcome variable.

Geographic distribution of effects— An alternative hypothesis is that the estimated treat-ment effect comes from an increase in infant mortality in the control group. The reasonscould be because of displacement of artisanal mining from the large-scale mining site tonearby communities, or relocation of households with worse health outcomes. However,Figure 3A shows that while the treatment group sees a large reduction in infant mortality, thecontrol group remains on a weakly negatively sloping trend. To allow for non-linear effectswith distance and better understand the geographic distribution of effects, I implement a spa-tial lag model using 10 km distance bins (Figure 6A). Mortality rates decrease sharply in theclosest bin (0-10 km), but there is no significant effect beyond 10 km. We can thus excludethat the estimated treatment effect stems from a deterioration of health in the control group.

Spatial randomization test— A spatial randomization inference test is used to show thatthe main results are not spurious because of a mis-specified model. If the mine locationis offset between 0 and 50 km in any direction while the mine keeps its de facto openingyear, the results attenuate toward zero. Figure 6B shows the distribution of treatment effects(industrial*minedeposit) when the mine location was randomized 1,500 times. The dashedline shows the initial treatment effect using the baseline model. A distribution centered atzero is not expected as the random mine placement (offset up to 50 km) will partially overlap

22

-0.1

-0.08

-0.06

-0.04

-0.02

0

0.02

0.04

0.06

0-10 10-20 20-30 30-40 40-50 50-60 60-70

Distance from mine (kilometers)

Infa

nt m

orta

lity

A

05

1015

-.1 -.05 0 .05 .1

p=0.0327

Distribution of coefficients from 1500 iterations

Dist

ribut

ion

of c

oeffi

cien

ts

B

Figure 6: Spatial lag model (A) and Spatial randomization test (B)Notes: Figure A shows the results from a spatial lag model with 10 km distance bins using thebaseline set of control variables and 95% confidence intervals. Figure B shows the distribution of1,500 coefficients from a spatial randomization test. The true mine location was offset by 0-50 km,and the main estimation model was re-estimated. The red line is the baseline estimate and the exactp-value shows the likelihood of the original estimate being drawn from this distribution.

with the 10 km treatment area. The exact p-value is presented in the figure and shows that itis unlikely that a misspecification of the model (Equation 1) is driving the results for infantmortality.

6 Discussion

The large-scale gold mining industry is rapidly expanding in Africa, yet we have limited un-derstanding of how large-scale mining operations affect local communities. In this analysis,I have attempted to fill the knowledge gap regarding infant mortality in the wake of the min-ing boom. The analysis contributes to the understanding of a more general question: whetherlocal industrial shocks can reduce infant mortality despite pollution risks. Infant mortalityis an important parameter in the study context: the sample used in this study has an averagerate of infant mortality of 9.7%, with most districts recording average rates of 80-90 deathsper 1000 births (Figure 7).

When a gold mine opens, the infant mortality rate decreases by more than 50% amongchildren who live within 10 kilometers6. The effects are geographically concentrated to the

650% is the less conservative measure using the main treatment effect (-0.055) and the mean in the overallsample (0.097). The most conservative measure using the treatment group, pre-treatment mean (0.151) resultsin a 36.4% change. The least conservative estimate is using the preferred specification (Table 2, column 2) of

23

0

20

40

60

80

100

120

140

160

180

China

United States

India

1965-1970

1975-1980

1985-1990

1995-2000

2005-2010

1955-1960

Infa

nt m

orta

lity

rate

(dea

ths /

100

0 liv

e bi

rths

)

(conservative) estimateindustrial x mine deposit

020

4060

80N

umbe

r of d

istric

ts

0 100 200 300

B

Years

(upper bound) estimateindustrial x mine deposit

A

Data from UN Population Division. Calculations author’s own. Data from DHS. Calculations author’s own.

Infant mortality rate (deaths / 1000 live births)

200

in sample African country average

Singapore

Figure 7: Global trends in infant mortality compared with treatment effect (A),District distribution of infant mortality (B)

Notes: The data in Figure A comes from UN Population Division and the calculations are theauthor’s own. The data are provided with 5 years averages. The data in Figure B comes from DHSand shows the distribution of infant mortality by district within the sample.

vicinity of the mine: at 20 km distance, there is no longer a detectable change in survivalrates (Figure 6A). Importantly, selective migration of mothers with better child survival skillsdoes not explain the whole effect: children born to never or early migrants, who migrated thelatest 4 years before the mine opening year experience a drop in mortality risk comparativeto that of the full sample.

The large, significant and almost instant drops in infant mortality experienced by miningcommunities around the time of mine opening may come from increases in local welfare.The mining spurs local economic growth—proxied by night-lights—and women living closeto mines are more 27% more likely to work in the service sector. Local women also havebetter access to media discussing matters of reproductive health. The effects of large-scalegold mining on women’s empowerment is further discussed in Tolonen (2016). I excludethe possibility that the effects are driven by a decrease in fertility among women in the localcommunities or an increase in health seeking behavior.

The trend break in mortality happens in the years immediately preceding the first pro-duction year, which corresponds to the capital investment period. The investment period isknown to generate local employment. Sensitivity analysis shows that infant mortality dropsaround the time of mine opening and remains persistently lower during the mine’s productivephase. The results are robust to the inclusion of mine fixed effects, the exclusion of district

7.9 percentage points and the mean of the treatment group, resulting in a 77.4% decrease in mortality.

24

linear time trends and birth-year fixed effects, and to different levels for clustering of thestandard errors. Furthermore, a spatial randomization test shows that artificially offsettingtrue mine location by 0 to 50 km creates null-results.

The estimated effects are economically important. The average morality rate in the min-ing communities before the mines open is 151 deaths per 1000 births. The infant mortal-ity rates drop by around 55 to 79 deaths per 1000 live births with the onset of large-scalegold mining. The drop in mortality is comparable to historic reductions in infant mortality:China’s infant mortality rate dropped by 58 deaths per 1000 births between 1960 and 1970,or by 79 deaths between 1960 and 1980, from an average of 121 deaths per 1000 births (seeFigure 7B). During the peak of its development phase, the decades of 1950 and 1960, Sin-gapore decreased its mortality rate by 32 deaths per 1000 births. In contrast, in the UnitedStates during the same period, mortality rates dropped from 25 to 14 deaths per 1000 births,but from a lower level. The magnitude of the effect estimated in this paper corresponds to thetotal gains in infant survivals achieved in Sub-Saharan Africa since the 1970s to today. Thisillustrates that the reductions in infant mortality rates spurred by large-scale gold mining inAfrica are comparatively large and important.

The data used in this project comes from 1987-2012. A glance at the UN statistics forinfant mortality shows that infant mortality was a relevant factor also in the last recordedtime period (2005-2010). All study countries report high levels of mortality: 79 deaths (per1000 live births) in Burkina Faso, 77 in Cote d’Ivoire, 116 in Democratic Republic of Congo,72 in Ethiopia, 50 in Ghana, 93 in Guinea, 101 in Mali, 55 in Senegal and 64 in Tanzania(Table A3). If a treatment effect such as the one estimated in this paper occurred today,countrywide, in Ghana, Senegal or Tanzania, the target rate of infant mortality specified bythe sustainable development goals would be reached.

Industrially advanced countries have lower infant mortality rates than less developedcountries. However, there are few micro-economical studies that causally estimate the linkbetween industrial development and infant health. Using the boom in large-scale gold min-ing in Africa, I find that local industrial development reduces infant mortality rates by 50%within two to four years. The research highlights that industrial development have an impor-tant role to play in the achievement of the sustainable development goals for infant mortalityin Sub-Saharan Africa. Importantly, the research illustrates that the return to industrial de-velopment with respect to infant mortality is high in high-mortality areas despite fears ofenvironmental degradation.

25

References

Almas, A. R. and Manoko, M. L. (2012). Trace element concentrations in soil, sediments,and waters in the vicinity of Geita Gold Mines and North Mara Gold Mines in NorthwestTanzania. Soil and Sediment Contamination: An International Journal, 21(2):135–159.

Almond, D., Edlund, L., and Palme, M. (2009). Chernobyl’s subclinical legacy: prenatalexposure to radioactive fallout and school outcomes in Sweden. The Quarterly Journal ofEconomics.

Amoh-Gyimah, R. and Aidoo, E. N. (2013). Mode of transport to work by governmentemployees in the Kumasi metropolis, Ghana. Journal of Transport Geography.

Aragon, F. M. and Rud, J. (2013). Natural Resources and Local Communities: Evidencefrom a Peruvian Gold Mine. American Economic Journal: Economic Policy, 5(2):1–25.

Aragon, F. M. and Rud, J. P. (2015). Polluting Industries and Agricultural Productivity:Evidence from Mining in Ghana. The Economic Journal.

ATSDR, U. (2007a). Toxicological profile for arsenic. US Department of Health and HumanServices, 1.

ATSDR, U. (2007b). Toxicological profile for cyanide. US Department of Health and HumanServices, 1.

Baird, S., Friedman, J., and Schady, N. (2011). Aggregate income shocks and infant mortal-ity in the developing world. Review of Economics and Statistics, 93(3):847–856.

Black, R. E., Morris, S. S., and Bryce, J. (2003). Where and why are 10 million childrendying every year? The Lancet, 361(9376):2226–2234.

Black, R. E., Victora, C. G., Walker, S. P., Bhutta, Z. A., Christian, P., De Onis, M., Ezzati,M., Grantham-McGregor, S., Katz, J., Martorell, R., et al. (2013a). Maternal and childundernutrition and overweight in low-income and middle-income countries. The Lancet,382(9890):427–451.

Black, S. E., Butikofer, A., Devereux, P. J., and Salvanes, K. G. (2013b). This is only atest? Long run impacts of prenatal exposure to radioactive fallout. NBER Working Paper18987.

26

Chay, K. Y. and Greenstone, M. (2003). The impact of air pollution on infant mortality: Evi-dence from geographic variation in pollution shocks induced by a recession. The QuarterlyJournal of Economics, pages 1121–1167.

Currie, J., Davis, L., Greenstone, M., and Walker, R. (2015). Environmental health risks andhousing values: evidence from 1,600 toxic plant openings and closings. The AmericanEconomic Review, 105(2):678–709.

Currie, J., Greenstone, M., and Moretti, E. (2011). Superfund Cleanups and Infant Health.American Economic Review, 101(3):435–41.

Currie, J. and Schmieder, J. F. (2009). Fetal exposures to toxic releases and infant health.American Economic Review, 99(2):177–83.

Dupas, P. (2011). Health behavior in developing countries. Annu. Rev. Econ., 3(1):425–449.

Economic Commission for Africa (2011). Minerals and Africa’s development: The interna-tional study group report on Africa’s mineral regimes. UNECA, Addis Ababa.

Eggert, R. G. (2002). Mining and Economic Sustainability: National Economies and LocalCommunities. MMSD Paper, 19.

Fernandez-Navarro, P., Garcıa-Perez, J., Ramis, R., Boldo, E., and Lopez-Abente, G. (2012).Proximity to mining industry and cancer mortality. Science of the total environment,435:66–73.

Gajigo, O., Mutambatsere, E., and Mdiaye, G. (2012). Mining in Africa: Maximizing eco-nomic returns for countries. African Development Bank Group Working Paper No 147.

Greenstone, M. and Hanna, R. (2014). Environmental regulations, air and water pollution,and infant mortality in India. American Economic Review, 104(10):3038–72.

Greenstone, M. and Jack, B. K. (2015). Envirodevonomics: A research agenda for an emerg-ing field. Journal of Economic Literature, 53(1):5–42.

Iyengar, G. V. and Nair, P. P. (2000). Global outlook on nutrition and the environment:Meeting the challenges of the next millenium. The Science of the Total Environment, 249.

Jayachandran, S. (2009). Air quality and early-life mortality evidence from Indonesia’swildfires. Journal of Human Resources, 44(4):916–954.

27

Kotsadam, A. and Tolonen, A. (2016). African mining, gender, and local employment. WorldDevelopment, 83:325 – 339.

Kudamatsu, M. (2012). Has democratization reduced infant mortality in sub-saharan africa?evidence from micro data. Journal of the European Economic Association, 10(6):1294–1317.

Kung, K. S., Greco, K., Sobolevsky, S., and Ratti, C. (2014). Exploring universal patterns inhuman home-work commuting from mobile phone data. PLoS ONE, 9(6).

Michalopoulos, S. and Papaioannou, E. (2013). Pre-colonial ethnic institutions and contem-porary African development. Econometrica, 81(1):113–152.

Milton, A. H., Smith, W., Rahman, B., Hasan, Z., Kulsum, U., Dear, K., Rakibuddin, M., andAli, A. (2005). Chronic arsenic exposure and adverse pregnancy outcomes in Bangladesh.Epidemiology, 16(1):82–86.

Moretti, E. (2010). Local Multipliers. American Economic Review. Papers and Proceedings,100:1–7.

Moretti, E. and Neidell, M. (2011). Pollution, health and avoidance behavior. Evidence fromthe ports of Los Angeles. The Journal of Human Resources, 46.

Rosenzweig, M. R. and Stark, O. (1989). Consumption smoothing, migration, and marriage:Evidence from rural india. The Journal of Political Economy, pages 905–926.

Serfor-Armah, Y., Nyarko, B., Dampare, S., and Adomako, D. (2006). Levels of arsenicand antimony in water and sediment from Prestea, a gold mining town in Ghana and itsenvirons. Water, Air, and Soil Pollution, 175(1-4):181–192.

Shafer, A. (2000). Regularities in travel demand: An international perspective. Journal ofTransportation and Statistics.

Tole, L. and Koop, G. (2011). Do environmental regulation affect the location decisions ofmultinational gold mining firms? Journal of Economic Geography, 11:151–177.

Tolonen, A. (2016). Local industrial shocks and endogenous gender norms. Mimeo.

von der Goltz, J. and Barnwal, P. (2014). Mines, the local welfare effects of mines in devel-oping countries. Columbia Department of Economics Discussion Papers, 1314-19.

28

Wang, H., Liddell, C. A., Coates, M. M., Mooney, M. D., Levitz, C. E., Schumacher, A. E.,Apfel, H., Iannarone, M., Phillips, B., Lofgren, K. T., et al. (2014). Global, regional, andnational levels of neonatal, infant, and under-5 mortality during 1990–2013: a systematicanalysis for the global burden of disease study 2013. The Lancet, 384(9947):957–979.

Weng, L., Boedhihartono, A. K., Dirks, P. H., Dixon, J., Lubis, M. I., and Sayer, J. A. (2013).Mineral industries, growth and agricultural development in Africa. Global Food Security,2:195–202.

29

Appendix

Demographic and Health Survey

The Demographic and Health Survey collects data on health and fertility in developingcountries and is funded by USAID (see www.dhsprogram.com for more information). Thedataset used in this paper consists of 30 cross-sectional datasets joined as a repeated cross-section. The paper uses two sets of data: child recode and women’s recode. The women’srecode samples women aged 15-49 in eligible households. The women’s recode providesinformation of her marital status, occupation and fertility. For example, the women’s recodeis used to estimate changes in local fertility levels. The main analysis on infant mortalityis done on the child recode. The child recode provides an entry for each birth to a womansurveyed in the women’s recode, limited to births in the last five years prior to the interview.The subset of women are therefore different in the two recodes: only women who have givenbirth in the last five years will figure in the child recode. Extensive summary statistics areavailable in Table 1, and variable definitions in Table A2 in the Appendix.

Gold Mining Data

The gold mine data comes from IntierraRMG. The data is licensed and obtained by sub-scription. The subscription was obtained in 2013 from Oxford Center for Research RichEconomies (OxCarre) where the author is an affiliated researcher. More information aboutIntierraRMG can be found at http://www.intierrarmg.com. The dataset contains all large-scale gold mines globally, with geographic coordinates and historic production volumes.Due to missing information and potential misreporting of the data provided by IntierraRMG,all gold mining entries in the study countries were double checked against Mining Atlas - Ex-plore the World of Mining (www.mining-atlas.com). Using Mining Atlas and Google Earth,the true geographic location was extracted and updated to reduce measurement bias frommisreporting of the geographic coordinates. The coordinates were chosen to represent thecenter point of the mining locations on Google Earth. These coordinates were then matchedwith the coordinates at the village or neighborhood level provided by DHS.

Night lights

The raw night lights data come from NOAA (see http://ngdc.noaa.gov/eog) and uses Version4 DSMP-OLS Nighttime Lights. The night light measures are stable lights by calendar year.

30

The data come at 20 arc second grids and were used to create area weighted average annulusby distance from a gold mine.

UN Population Division

The country level statistics on infant mortality used to compare the estimates from theanalysis with the global trends in infant mortality, comes from UN Population Division,Department of Economic and Social Affairs. The data are taken from World PopulationProspects: The 2015 Revision, File MORT/1-1: Infant mortality rate (both sexes combined)by major area, region and country, 1950-2100 (infant deaths per 1,000 live births). The datawere used to calculate the average infant mortality rate among the countries analyzed in thepaper. The raw data for selected countries are presented in Table A3.

GADM database of Global Administrative Areas

The DHS data provides region (subnational level 1) information. However, in this projectwe have used district level information. The district data comes from GADM database ofGlobal Administrative Areas which provides the polygons for administrative areas globally.The name and size of subdivisions vary across countries. Subnational level 2 was chosenfor this project, regardless of the nationally specific name for this level and the averagegeographic size of such a subdivision. The data were accessed from DivaGIS in August2013 (www.diva-gis.org/datadown). The subdivision level 2 data were then extracted foreach DHS cluster, which is a village or a neighborhood in a city, by overlaying the DHScluster data and the GADM data using ArcGIS. The DHS clusters are offset with 1 to 10 kmto ensure privacy of respondents. This will introduce some random error in the assignmentof districts to individual observations as individuals living close to a district border may havebeen assigned the neighboring district instead.

31

Appendix: Figures and Tables

02

46

80

24

68

02

46

8

1970 1980 1990 2000 2010 1970 1980 1990 2000 2010 1970 1980 1990 2000 2010

Burkina Faso Cote d'Ivoire Dem. Republic of Congo

Ethiopia Ghana Guinea

Mali Senegal Tanzania

Gol

d pr

oduc

tion

(mto

n)

Year

DHS surveyGold production (mton)

05

100

510

150

510

15

Activ

e G

old

Min

es

Active gold mines (count)

15

Figure A1: The Evolution of the Production of Gold and the Demographic and HealthSurveys and Sample Years by Country

32

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

Two way linear prediction

Kernel-weighted local polynomial smoothing

Break: year -2Exclude: -2 to 0

Break: year -1Exclude: -1 to 0

Break: year 0Exclude: -

No time breakExclude: -

Year since mine opening (horizontal axis)Treatment groupControl group

A

B

C

D

Figure A2: Parametric and non-parametric estimation of trends for infant mortality

Notes: The graphs show results for two-way linear prediction (column 1) and kernel-weighted

polynomial smoothing (column 2) across four specifications (rows 1-4). The mine opening year is

the first year of production for the mine that is closest to the individual. The treatment group (red

line) is drawn from within 10 km for the closest mine. The control group (blue line) is drawn 10-100

km away, but excludes births with the second closest mine within 20 km. Row 1 uses the preferred

specification and allows for a break at year -2, and excludes children born year -2 to year 0. Row 2

allows for a break at year -1, row 3 at year 0, and row 4 does not allow for a break.

33

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

0.1

.2.3

-10 -5 0 5 10

Two way linear prediction

Kernel-weighted local polynomial smoothing

Break: year -2Exclude: -2 to 0

Break: year -1Exclude: -1 to 0

Break: year 0Exclude: -

No time breakExclude: -

Year since mine opening (horizontal axis)Treatment groupControl group

A

B

C

D

Figure A3: Parametric and non-parametric estimation of trends for infant mortality.Excluding migrants who moved 4 years prior to mine opening or later

Notes: The graphs show results for two-way linear prediction (column 1) and kernel-weighted

polynomial smoothing (column 2) across four specifications (rows 1-4). The mine opening year is

the first year of production for the mine that is closest to the individual. The treatment group (red

line) is drawn from within 10 km for the closest mine. The control group (blue line) is drawn 10-100

km away, but excludes births with the second closest mine within 20 km. Row 1 uses the preferred

specification and allows for a break at year -2, and excludes children born year -2 to year 0. Row 2

allows for a break at year -1, row 3 at year 0, and row 4 does not allow for a break. All births from

mothers who migrated 4 years before the mine opening year, or later, are excluded.

34

0.1.2.3.4

-10 -5 0 5 10

0.1.2.3.4

-10 -5 0 5 10

0.1.2.3.4

-10 -5 0 5 10

0.1.2.3.4

-10 -5 0 5 10

0.1.2.3.4

-10 -5 0 5 10

0.1.2.3.4

-10 -5 0 5 10

0.1.2.3.4

-10 -5 0 5 10

0.1.2.3.4

-10 -5 0 5 10

Two way linear prediction

Kernel-weighted local polynomial smoothing

Break: year -2Exclude: -2 to 0Outcome: Infant mortalityfirst 12 monthsSample: Boys

Year since mine opening (horizontal axis)Treatment groupControl group

Break: year -2Exclude: -2 to 0Outcome: Infant mortalityfirst 12 monthsSample: Girls

Break: year -2Exclude: -2 to 0Outcome: Infant mortality first 6 months

Break: year -2Exclude: -2 to 0Outcome:Infant mortalityfirst 1 month

A

B

C

D

Figure A4: Parametric and non-parametric estimation of trends for infant mortalityfor boys (A), girls (B), mortality first 6 months (C), and neonatal mortality (D)

Notes: The graphs show results for two-way linear prediction (column 1) and kernel-weighted

polynomial smoothing (column 2) across four samples (rows 1-4). The mine opening year is the first

year of production for the mine that is closest to the individual. The treatment group (red line) is

drawn from within 10 km for the closest mine. The control group (blue line) is drawn 10-100 km

away, but excludes births with the second closest mine within 20 km. All rows (1-4) use the

preferred specification and allows for a break at year -2. Row 1: infant mortality first 12 months,

boys only. Row 2: infant mortality first 12 months, girls only. Row 3: infant mortality first 6 months.

Row 4: infant mortality first 1 month.

35

Table A1: Mother fixed effects

Dependent variable: infant mortality infant mortality neonatal mortality12 months 6 months 1 month

(1) (2) (3)

industrial*mine deposit -0.121 -0.093 -0.089(0.181) (0.125) (0.101)

birth order -0.020* -0.014 -0.011(0.012) (0.009) (0.007)

Mother FE Yes Yes YesCountry-birth year FE Yes Yes YesDistrict time trend Yes Yes YesBirth month FE Yes Yes YesDrop 10-30km away Yes Yes YesDrop investment phase Yes Yes Yes

Sample sizeTreated mothers before-after birth 94 94 94Treated mothers with >1 birth 368 368 368Treated children with >1 sibling 730 730 730Mothers with >1 birth 15,093 15,093 15,093

Observations 29,234 34,030 38,428R-squared 0.755 0.703 0.646

Notes: *** p < 0.01, ** p < 0.05, * p < 0.1. Clustered standard errors clustered at DHS cluster level. Allregressions control for mother fixed effect, child’s birth number and birth month, and fixed effects for

country-birth year and district linear time trend. Mine deposit captures if there is a gold deposit within 10 kmfrom the household. Industrial captures if the gold deposit was actively extracted in the child’s birth year.

36

Tabl

eA

2:Va

riabl

eD

efini

tion

(1)

(2)

Varia

ble

Defi

nitio

nD

ata

sour

ce

infa

ntm

orta

lity

Chi

lddi

edw

ithin

first

12m

onth

sof

life

DH

Sch

ildre

code

infa

ntm

orta

lity

(1m

onth

)C

hild

died

with

infir

stm

onth

sof

life

DH

Sch

ildre

code

infa

ntm

orta

lity

(6m

onth

s)C

hild

died

with

infir

st6

mon

ths

oflif

eD

HS

child

reco

dem

othe

r’sag

eA

geof

mot

heri

nth

ech

ildre

code

DH

Sch

ildre

code

mot

her’s

educ

atio

nN

umbe

rofy

ears

ofsc

hool

ing

DH

Sch

ildre

code

urba

nho

useh

old

Hou

seho

ldliv

esin

anur

ban

dwel

ling

DH

Sch

ildre

code

birth

num

ber

Chi

ldnu

mbe

rin

the

birth

orde

rofm

othe

rD

HS

child

reco

debi

rthye

arof

child

Cal

enda

rbirt

hye

arof

child

DH

Sch

ildre

code

ante

nata

lcar

eA

tlea

ston

ean

tena

talc

are

visi

tdur

ing

preg

nanc

yD

HS

child

reco

deha

she

alth

card

Mot

herc

ansh

owch

ild’s

heal

thca

rdto

inte

rvie

wer

DH

Sch

ildre

code

very

smal

latb

irth

Chi

ldve

rysm

alls

ize

atbi

rth,m

othe

rest

imat

ion

DH

Sch

ildre

code

ever

vacc

inat

edC

hild

ever

rece

ived

vacc

inat

ion

DH

Sch

ildre

code

coug

hC

hild

had

coug

hin

the

2w

eeks

prio

rto

surv

eyin

gD

HS

child

reco

defe

ver

Chi

ldha

dfe

veri

nth

e2

wee

kspr

iort

osu

rvey

ing

DH

Sch

ildre

code

diar

rhea

Chi

ldha

ddi

arrh

eain

the

2w

eeks

prio

rto

surv

eyin

gD

HS

child

reco

dem

igra

nt(2

year

s)If

mot

herm

oved

loca

tion

2ye

ars

prio

rto,

oraf

terc

lose

stm

ine

open

ing

DH

Sch

ildre

code

mig

rant

(5ye

ars)

Ifm

othe

rmov

edlo

catio

n5

year

spr

iort

o,or

afte

rclo

sest

min

eop

enin

gD

HS

child

reco

deto

talf

ertil

ityLi

fetim

ebi

rths,

wom

enD

HS

wom

anre

code

tota

lfer

tility

unde

r22

Life

time

birth

s,w

omen

aged

less

than

22at

time

ofin

terv

iew

DH

Sw

oman

reco

delis

ten

tora

dio

FPH

eard

radi

ofa

mily

plan

ning

mes

sage

inla

stfe

wm

onth

spr

iort

oin

terv

iew

DH

Sw

oman

reco

dese

rvic

ejo

bM

ain

occu

patio

nif

wor

ked

inla

st12

mon

ths,

serv

ice

sect

orD

HS

wom

anre

code

earn

sca

shW

oman

earn

sca

shor

cash

and

inki

ndfo

rwor

kD

HS

wom

anre

code

poly

gam

ous

mar

riage

Isin

apo

lyga

mou

sm

arria

gecu

rren

tlyD

HS

wom

anre

code

barr

ierh

ealth

acce

ssIn

dex

rang

ing

0-1,

whe

re1

isis

leas

tacc

ess

tohe

alth

care

fors

elf

DH

Sw

oman

reco

dedu

eto

mon

ey,p

erm

issi

onor

dist

ance

indu

stria

l*m

ine

depo

sit(

atbi

rth)

1if

anac

tivel

ypr

oduc

ing

min

ew

ithin

10km

from

dwel

ling

inth

eye

arof

the

child

’sbi

rth,

Ow

nco

nstru

ct0

othe

rwis

ein

dust

rial*

min

ede

posi

t(at

surv

ey)

1if

anac

tivel

ypr

oduc

ing

min

ew

ithin

10km

from

dwel

ling

inth

eye

arof

the

inte

rvie

w,

Ow

nco

nstru

ct0

othe

rwis

em

ine

depo

sit

1if

ago

ldm

ine

with

in10

km,r

egar

dles

sof

prod

uctio

nst

atus

,O

wn

cons

truct

0ot

herw

ise

37

Tabl

eA

3:G

loba

lTre

nds

inIn

fant

Mor

talit

y

1950

-195

519

55-1

960

1960

-196

519

65-1

970

1970

-197

519

75-1

980

1980

-198

519

85-1

990

1990

-199

519

95-2

000

2000

-200

520

05-2

010

Ang

ola

230.

4921

4.76

199.

9118

5.99

172.

8516

0.8

156.

9715

2.99

150.

4713

7.9

115.

9810

4.3

Ben

in20

9.89

194.

9318

3.78

172.

5415

1.97

136.

4112

6.42

119.

610

8.19

97.8

90.6

285

.11

Bot

swan

a13

5.09

123.

6711

3.29

103.

9990

.27

74.6

663

.49

54.0

251

.45

61.2

158

.77

40.6

9B

urki

naFa

so30

7.92

258.

421

7.13

183.

515

6.73

135.

8312

0.75

110.

5510

1.42

93.2

585

.86

78.9

Bur

undi

166.

515

7.3

148.

514

013

712

711

8.4

116.

312

6.13

116.

9310

7.09

101.

14C

amer

oon

169.

4315

7.72

144.

8613

3.01

119.

9310

7.51

97.9

493

.05

93.5

496

.196

.67

94.0

5C

ape

Verd

e13

8.79

131.

9912

4.75

116.

8995

.61

78.4

464

.66

53.5

644

.42

36.5

728

.03

20.6

3C

AR

203.

8818

9.37

175.

6615

9.6

138.

3111

9.09

109.

7110

9.24

112.

911

5.91

113.

7110

5.38

Cha

d19

1.37

181.

4917

1.81

162.

6615

3.63

145.

5413

5.61

129.

0712

8.63

131.

2813

3.35

131.

17C

ongo

142.

0612

2.72

106.

9496

.688

.81

83.0

278

.25

73.4

373

.58

75.0

575

.372

.43

DR

C16

6.6

157.

5415

1.19

143.

4613

3.7

129.

1312

4.86

121.

1311

8.58

128.

4711

9.9

115.

81C

ote

d’Iv

oire

167.

4816

0.39

154.

714

7.37

126.

8510

8.54

96.1

693

.37

94.6

791

.93

84.7

77.1

7D

jibou

ti22

2.47

202.

6518

5.25

168.

8615

3.88

140.

5912

5.1

116.

9510

9.17

99.9

891

.08

82.0

8Eq

uato

rialG

uine

a19

6.38

186.

0217

6.11

166.

6215

7.36

148.

613

8.12

127.

6111

8.14

113.

6511

0.66

102.

45Er

itrea

176

162.

815

0.5

139.

113

312

711

5.8

104.

4889

.79

73.2

161

.75

53.8

8Et

hiop

ia19

9.27

181.

4315

9.98

147.

9114

0.03

134.

6214

0.15

126.

9511

4.54

101.

0487

.12

72.4

6G

abon

179.

3416

6.6

157.

5413

4.28

113.

5494

.99

77.6

663

.15

58.8

858

.28

57.7

251

.15

Gam

bia

220.

8621

0.08

199.

1518

4.71

164.

313

9.59

117.

3110

1.48

92.8

786

.41

79.9

873

.84

Gha

na15

0.09

132.

6312

2.87

117.

2110

7.5

97.5

90.4

883

.93

72.8

870

.63

61.8

549

.61

Gui

nea

217.

7721

3.63

209.

6220

5.61

190.

6917

415

8.77

144.

6713

0.9

119.

1510

4.17

93.1

7G

uine

a-B

issa

u21

0.77

200.

9719

1.37

182.

2817

3.55

165.

3715

2.96

145.

1714

0.51

134.

112

5.64

118.

7K

enya

147.

0113

3.67

116.

9610

4.39

91.5

80.4

470

.11

66.7

366

.48

69.1

469

.54

64.7

2Le

soth

o16

9.27

150.

3513

4.28

130.

2612

3.09

109.

7293

.95

83.8

369

.65

81.2

986

.15

76.8

6Li

beria

223.

7220

8.36

194.

2718

1.17

168.

6715

8.48

159.

3416

3.67

167.

7215

4.93

115.

7588

.62

Mad

agas

car

180.

8516

6.88

154.

8814

3.2

132.

4912

2.01

110.

9810

9.85

95.6

376

.29

58.3

44.7

8M

alaw

i19

7.89

192.

4918

5.86

180.

0416

8.43

159.

2715

0.95

143.

4213

3.1

120.

7810

7.41

95.2

3M

ali

175.

417

2.61

168.

1516

2.36

155.

7514

8.7

140.

7813

4.62

127.

1411

9.08

110.

5110

1.35

Mau

ritan

ia15

1.09

149.

3314

7.02

142.

7113

3.32

113.

2693

.07

80.9

476

.71

75.7

76.5

177

.33

Mau

ritiu

s10

3.09

73.8

460

.63

67.2

460

.76

37.6

726

23.1

118

.47

19.6

913

.35

12.7

6M

ozam

biqu

e22

0.12

201.

3918

4.79

172.

0615

8.4

145.

7914

3.06

142.

8413

3.95

114.

699

.09

88.0

3N

amib

ia17

1.61

148.

9513

0.4

114.

9910

0.67

87.6

475

.49

68.5

461

.754

.55

48.2

137

.76

Nig

er17

4.04

170.

7316

7.3

164.

2816

2.15

160.

7915

9.18

155.

114

5.73

130.

6711

2.72

95.9

2N

iger

ia18

8.93

175.

9916

415

2.87

140.

9513

1.73

126.

9912

6.58

126.

4812

2.19

107.

2396

.14

Rw

anda

160.

2815

1.56

143.

0413

713

4.29

132.

312

4.14

119.

9712

7.92

117.

6810

7.9

100.

15Sa

oTo

me

&Pr

inci

pe12

4.13

112.

1398

.83

88.0

179

.67

69.5

966

.41

63.4

760

.62

57.9

255

.35

51.7

1Se

nega

l13

1.4

128.

4412

3.4

116.

5610

8.34

98.7

288

.54

79.9

572

.39

65.8

260

.16

55.2

4Si

erra

Leon

e24

2.37

234.

7622

6.71

214.

1419

4.7

164.

1813

5.32

153.

5316

5.6

156.

7613

3.27

113.

68So

mal

ia20

7.23

192.

8817

9.34

166.

5915

4.55

148.

7613

7.77

127.

1614

1.3

122.

5811

0.58

106.

67So

uth

Afr

ica

96.1

91.3

86.5

83.5

977

.13

70.6

60.7

352

.57

50.6

55.8

559

.21

54.8

1Sw

azila

nd17

3.74

160.

3115

0.39

140.

5312

4.04

107.

7790

.44

77.3

469

.29

80.1

886

.64

75.8

7Ta

nzan

ia15

3.22

143.

1213

6.12

128.

0611

8.62

108.

0710

3.79

100.

6599

.790

.38

77.3

564

.45

Togo

191.

3817

1.94

154.

3113

8.36

123.

5811

3.23

104.

7997

.17

90.5

283

.76

76.6

973

.96

Uga

nda

160.

4314

5.29

130.

1711

6.46

102.

7610

5.98

107.

6410

8.45

110.

4310

4.63

91.0

379

.16

Zam

bia

148.

2913

6.93

126.

5911

7.88

106.

9410

0.19

98.5

210

2.98

107.

1810

5.46

102.

794

.91

Zim

babw

e11

4.51

105.

1396

.67

89.5

682

.63

74.2

663

.77

56.3

355

.564

.58

68.4

559

.28

SSA

aver

age

177.

9716

5.01

153.

4614

3.21

131.

1811

9.50

109.

8110

3.99

100.

1295

.19

87.2

078

.52

Stud

yco

untry

aver

age

185.

4617

2.02

160.

3515

0.23

137.

5812

6.12

118.

2511

0.65

103.

5897

.75

87.9

678

.68

Indi

a16

5.00

153.

0814

0.12

128.

4611

7.99

106.

4595

.02

85.0

776

.45

68.8

660

.752

.91

Chi

na12

1.64

117.

9212

1.17

63.1

546

.93

41.9

637

.59

33.7

530

.34

27.3

24.6

21.9

9Si

ngap

ore

60.6

943

.18

28.6

23.7

819

.34

12.8

58.

77.

794.

493.

332.

551.

92U

nite

dSt

ates

30.4

627

.33

25.3

822

.67

18.3

914

.34

11.6

10.3

78.

817.

496.

925.

40N

otes

:Dat

aco

mes

from

UN

Popu

latio

nD

ivis

ion

and

pres

ents

trend

sin

infa

ntm

orta

lity

fors

elec

ted

coun

tries

from

1950

to20

10.A

vera

ges

pres

ente

dar

eba

sed

onau

thor

’sow

nca

lcul

atio

ns.

38

Table A4: Mining companies

Company name Country origin Mines (in dataset) Countries activeAkrokeri-Ashanti Gold Mines, Inc. Canada 2 GhanaAl-Amoudi Family Saudi Arabia 1 EthiopiaAmara Mining PLC UK 2 Burkina Faso, Cote d’IvoireAnglogold Ashanti Ltd. South Africa 12 Ghana, Guinea, Mali, TanzaniaAvnel Gold Mining Ltd. UK 1 MaliAvocet UK 1 Burkina FasoBanro Corp. Canada 1 Congo (Dem Rep)Barrick Gold Corp. Canada 4 TanzaniaBassari Australia 1 SenegalEden Roc Mineral Corp. Canada 1 Cote d’IvoireEndeavour Mining Corp. Canada 3 Burkina Faso, Ghana, MaliGhana National Petroleum Corp. Ghana 1 GhanaGold Fields Ltd. South Africa 2 GhanaGolden Star Resources Ltd. USA 2 GhanaIamgold Corp. Canada 3 Burkina Faso, MaliKinross Gold Corp. Canada 1 GhanaLionGold Corp. Singapore 1 GhanaMDN Inc. Canada 1 TanzaniaNewcrest Australia 1 Cote d’IvoireNewmont Mining Corp. USA 1 GhanaNoble Mineral Resources Ltd. Australia 1 GhanaPerseus Mining Ltd. Australia 1 GhanaPrestea Resource Ghana 1 GhanaPMI Gold Corp. Canada 1 GhanaRandgold Resources Limited South Africa 4 Cote d’Ivoire, MaliResolute Mining Ltd. Australia 2 Mali, TanzaniaSemafo Inc. Canada 2 Burkina Faso, GuineaSeverstal Russia 2 Burkina Faso, GuineaShanta Gold Ltd. UK 1 TanzaniaState of Cote d’Ivoire Cote d’Ivoire 2 Cote d’ IvoireState of Mali Mali 6 MaliTeranga Gold Corp. Canada 1 SenegalWeather II Egypt 1 Cote d’Ivoire

Notes: Some mines are double counted if the ownership is shared. This is true for operations by State ofCote d’Ivoire, State of Mali, Ghana Petroleum, Iamgold, Anglogold, Barrick, MDN, Eden Roc, Randgold Res,Resolute and Weather II. Missing company information for Esasse and Dunkwa mines in Ghana and Poura inBurkina Faso.

39