(This is a sample cover image for this issue. The actual cover is not yet available at this time.

This article appeared in a journal published by Elsevier. The attached

copy is furnished to the author for internal non-commercial research
and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elsevier’s archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/copyright
 Author's personal copy

Journal of Health Economics 32 (2013) 304–319

Contents lists available at SciVerse ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Cash transfers and domestic violence夽

Melissa Hidrobo a,∗ , Lia Fernald b
a
International Food Policy Research Institute, 2033 K St, NW, Washington, DC 20006, USA
b
School of Public Health, University of California, Berkeley, CA 94720, USA

a r t i c l e i n f o a b s t r a c t

Article history: Violence against women is a major health and human rights problem yet there is little rigorous evidence
Received 8 November 2011 as to how to reduce it. We take advantage of the randomized roll-out of Ecuador’s cash transfer program
Received in revised form 20 October 2012 to mothers to investigate how an exogenous increase in a woman’s income affects domestic violence.
Accepted 2 November 2012
We find that the effect of a cash transfer depends on a woman’s education and on her education relative
Available online xxx
to her partner’s. Our results show that for women with greater than primary school education a cash
transfer significantly decreases psychological violence from her partner. For women with primary school
JEL classification:
education or less, however, the effect of a cash transfer depends on her education relative to her partner’s.
I10
I38
Specifically, the cash transfer significantly increases emotional violence in households where the woman’s
D13 education is equal to or more than her partner’s.
© 2012 Elsevier B.V. All rights reserved.
Keywords:
Cash transfers
Domestic violence
Household bargaining models
Education
Women’s income

1. Introduction Violence toward women is both a cause and a consequence of

gender inequality, and thus creates a significant obstacle for achiev-
Violence against women is a major health and human rights ing the Millennium Development Goals (García-Moreno et al.,
problem that is wide-spread and common (García-Moreno et al., 2005). The consequences of domestic violence are extensive and
2005, 2006). Recent multi-country studies on domestic violence range from the direct physical and mental harm for women and
conducted by the World Health Organization (WHO) and the their children to economic losses at the community and national
Demographic Health Survey (DHS) show that lifetime prevalence level. Physical violence is associated with victims in both devel-
of physical and sexual violence toward women by a spouse or part- oped and developing countries having poor health, depressive
ner is anywhere between 15% and 75%, with the highest prevalence symptoms, substance abuse problems, and developing chronic
found in Peru (69%), Ethiopia (71%), and Bangladesh (75%) (García- diseases and mental illnesses (Coker et al., 2002; Ackerson and
Moreno et al., 2006; Hindin et al., 2008). In Ecuador the lifetime Subramanian, 2008; Ellsberg et al., 2008). For children, expo-
prevalence of physical violence (31%), sexual violence (12%), and sure to domestic violence is associated with lower birth weight
psychological violence (41%) is also high (CEPAR, 2005). and IQ scores, and increased emotional and behavioral problems
(Sternberg et al., 1993; Koenen et al., 2003; Wolfe et al., 2003; Aizer,
2011). Furthermore, children who suffer from social and emo-
夽 We would like to thank John Hoddinott, Agnes Quisumbing, Amber Peterman, tional problems caused by domestic violence often create spillover
Elizaveta Perova, Jeremy Magruder, Alain de Janvry, Elisabeth Sadoulet, Emily Flynn, effects such as decreasing the academic achievement of classroom
Ann Weber, Jessica Jones-Smith, Rachel Gardner, Rachel Heath, Patricia Kariger, and peers (Carrell and Hoekstra, 2010). The economic costs of domestic
three anonymous referees for all their valuable comments. We gratefully acknowl- violence are also high. In the United States, the cost of domestic vio-
edge the collaboration and contributions of Norbert Schady, Caridad Araujo, and
Christina Paxson. We are grateful to our Ecuadorian collaborators in Ecuador, and
lence is estimated at $5.8 billion per year, which includes the cost
thank Santiago Izquierdo, Mauricio León, Ruth Lucio, Juan Ponce, José Rosero and of medical services and the cost of lost productivity (CDC, 2003).
Yajaira Vázquez. All errors remain our own. In Colombia the estimated cost of domestic violence due to lost
∗ Corresponding author. Tel.: +1 202 862 4612.
labor earning of women is 7.1 billion pesos ($2.5 million USD) or
E-mail address: m.hidrobo@cgiar.org (M. Hidrobo).

0167-6296/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.jhealeco.2012.11.002
 Author's personal copy

M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319 305

approximately 3.2% of GDP in 2003 (Ribero and Sánchez, 2005).1 a woman’s income. We focus on a woman’s absolute and relative
These estimates most likely represent a lower bound given that education and find that for women with more than primary school
they do not calculate indirect costs such as those due to behavioral education (more than 6 years of schooling), the BDH significantly
problems of children who witness domestic violence. decreases emotional violence by 8 percentage points and control-
Although the consequences of domestic violence are well doc- ling behaviors by 14 percentage points. For women with primary
umented, there is less evidence about public policies or programs education or less, we find that the effect of the BDH depends on
that could help reduce domestic violence. In the United States, stud- whether or not the husband or partner has more years of schooling
ies have found that unilateral divorce laws (Stevenson and Wolfers, than his wife or partner. For households where the husband does
2006), reductions in the male–female wage gap (Aizer, 2011), and not have more schooling than his wife, we find that the BDH signif-
community advocacy services for women (Sullivan and Bybee, icantly increases emotional violence by 9 percentage points. While
1999) reduce domestic violence. In South Africa, a micro-finance the direction of impact on physical violence is similar to those on
program for women that was combined with health education the psychological indicators, the impacts are never significant.
led to reductions in domestic violence (Pronyk et al., 2006), and
a community-based educational program (Stepping Stone) led to
reductions in men reporting they had committed domestic violence 2. Theories and evidence on income and domestic violence
but no corresponding reduction in woman reporting experiencing
domestic violence (Jewkes et al., 2008). Theories on domestic violence span many disciplines and con-
In developing countries, a handful of studies on policies that sequently there is no unified understanding across and within
reduce domestic violence focus on the impacts of conditional cash disciplines as to the association between income and domestic
transfer (CCTs) programs, which have become a very popular tool violence (for overview on theories see Heise, 2012). For example,
for reducing poverty and increasing human capital (Fiszbein et al., marital dependency theory in sociology asserts that women who
2009). CCTs operate by giving cash payments to mothers only if they are economically dependent on their partner are at greater risk of
comply with a certain set of requirements relating to the health domestic violence (Vyas and Watts, 2009). Similarly, feminist the-
and education of family members. Since CCTs give money to the ories assert that institutions that promote masculinity and gender
mothers and not the fathers, the programs provide an opportu- inequality trigger domestic violence. Consequently, a cash transfer
nity to study how increases in women’s incomes affect domestic that empowers women and leads to a greater degree of financial
violence. Bobonis et al. study the effects of Mexico’s CCT program, equality could lower the prevalence of domestic violence. On the
Oportunidades, and find that in the short run (2–6 years) physi- other hand some sociologists argue that domestic violence is a dis-
cal violence decreased by 3.6 percentage points (2009), but that order of power and control where a man may resort to violence
these effects disappeared 5–9 years after the program was imple- as a means for reinstating his authority over his wife (Pence and
mented (Bobonis and Castro, 2010).2 In Peru, Perova (2010) finds Paymar, 1993; Macmillan and Gartner, 1999). Along similar lines,
that in the short run its CCT program, Juntas, significantly decreased status inconsistency theories argue that stress and tensions within
physical violence by 9 percentage points and emotional violence by relationships are a consequence of husband and wife occupying
11 percentage points. Programs in both Peru and Mexico are atypical combinations of status characteristics (Hornung et al.,
conditional cash transfer programs and thus it is impossible to com- 1981). Consequently, an increase in her income could increase
pletely isolate the income effect on domestic violence from the domestic violence if it leads to men feeling powerless or to atypi-
conditions. In particular, the health requirements of CCT programs cal roles within the household such as the woman being the main
could lead to decreases in domestic violence because increased income earner.
exposure to health facilities makes it harder to hide physical Economic models of domestic violence expand on household
and sexual abuse, and health care providers are often trained to bargaining models where individual control of resources is impor-
deal with domestic violence. Furthermore, the required health tant because bargaining outcomes depend on threat points such
and nutrition monthly information sessions empowered women as divorce (Manser and Brown, 1980; McElroy and Horney, 1981)
(Adato et al., 2000a,b) and thus could also lead to reductions in or non-cooperative equilibriums (Lundberg and Pollak, 1993). The
domestic violence. more attractive an individual’s opportunities outside the family,
By focusing on an unconditional cash transfer to mothers in the more credible the threat point, and therefore, the more likely
Ecuador we are able to isolate the income effect and thus contribute that the intra-family distribution of resources will align more
to the existing literature. We take advantage of the randomized closely with that individual’s preferences. In these models vio-
roll-out of Ecuador’s unconditional cash transfer program, the Bono lence is either (i) expressive – improves aggressor’s self-esteem
de Desarrollo Humano (BDH), to study the impact of an exogenous and releases frustration (Farmer and Tiefenthaler, 1997); (ii) instru-
increase in a woman’s income on domestic violence. In our analy- mental – used as an instrument to align resources more closely
ses, we concentrate on physical violence and psychological violence with the aggressor’s preferences or to control victim’s behavior
(categorized into emotional violence and controlling behaviors by a (Eswaran and Malhotra, 2011); (iii) extractive – used as a tool
husband or partner). Given the lack of consensus on the direction of to extract resources (Bloch and Rao, 2002); or (iv) a combination
the associations between a woman’s income and domestic violence of expressive and instrumental (Tauchen et al., 1991). Predictions
(explained in more detail below), we investigate not only the aver- on how an increase in a woman’s income affects domestic vio-
age impact of the BDH but also factors that are more likely to lead to lence depend not only on how violence is modeled but also on the
an increase or decrease in domestic violence as a result of increasing assumptions made within the models.
Economic models that predict a decrease in violence: In Farmer and
Tiefenthaler (1997) violence is expressive and a man chooses the
level of violence and the level of income to transfer to his wife that
1
Estimated cost of intra-family violence is 9.4 billion pesos or approximately 4.2% maximizes his utility. A woman’s outside of marriage utility (also
of GDP, and this includes lost labor earning of women due to violence toward their known as reservation utility) determines the amount of violence
children and themselves.
2
The main reason given for the discrepancy between shorter and longer term
she will tolerate for a given transfer. As a wife’s income rises, the
impacts was marital selection, and in particular, the types of couples choosing to utility she receives from his transfer declines, so his ability to “buy”
remain in a relationship longer term. violence from her decreases, and thus violence falls.
 Author's personal copy

306 M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319

Economic models that predict an ambiguous relationship: (BDH), to study the impact of an exogenous increase in a woman’s
Tauchen’s model is similar to that of Farmer and Tiefenthaler, income on domestic violence. We contribute to the existing lit-
except that he allows for violence to also be an instrument to erature by exploring factors that contribute to the heterogeneous
control behavior. In his model the impact of an increase in a relationship between income and domestic violence. We focus on
woman’s income on domestic violence depends on whether or not a woman’s education because it is a factor that is likely to affect a
the woman’s utility outside of marriage or partnership equals her woman’s reservation utility. Women with more education are more
utility from marriage or partnership. If her utility from marriage likely have better out-of-marriage options and thus higher reser-
equals her reservation utility (the binding case), then violence is vation utilities because they are more likely to find work, receive a
only expressive and the predictions from the model are similar to higher wage rate, be financially independent, and face a better mar-
those of Farmer and Tiefenthaler. However, if a woman’s utility riage market (Lundberg and Pollak, 1996; Farmer and Tiefenthaler,
from marriage is greater than her reservation utility (non-binding 1997; Pollak, 2005; Eswaran and Malhotra, 2011). However, we
case), then an increase in a woman’s income will generally decrease focus not just on women’s absolute education, but also on their rel-
domestic violence except in cases when a husband’s or partner’s ative education with respect to their partner because many studies
marginal utility for violence increases with her consumption (i.e., find that women with more education than their partners experi-
a preference for violence against a woman with more purchasing ence more domestic violence (Hornung et al., 1981; Flake, 2005;
power). In their analysis they find that a husband’s marginal util- Ackerson et al., 2008).
ity for violence increases with her consumption when the woman
is the main income earner; and consequently, in these cases an 3. Study design and data
increase in her income increases domestic violence. Similarly, in
Eswaran and Malhotra’s model violence is instrumental and used 3.1. The Bono de Desarrollo Humano
to bring the allocation of goods more in line with a husband or
partner’s preferences. Violence is non-monotonic in a wife’s reser- In 1998 Ecuador created its largest social assistance program
vation utility because a woman might find it worthwhile to allocate called Bono Solidario. The Bono Solidario consisted of an uncon-
resources more in line with her preferences, and so much so, that ditional cash transfer of 100,000 sucres (approximately $15 USD)
it might increase violence. per month to poor mothers and 50,000 sucres (approximately $7.5
Economic models that predict an increase in domestic violence: USD) per month to elderly and handicapped. Due to substantial
Increases in domestic violence as a result of increases in women’s leakage of the program to non-poor families, the Bono Solidario
income are explained in extraction models. For example, Bloch and was phased out and replaced by the Bono de Desarrollo Humano
Rao (2002) explain dowry violence in India using a model where (BDH) in 2003. The major improvement of the BDH over the Bono
violence is used as a bargaining instrument by a husband to extract Solidario was that it was means-tested. Thus, households who were
rents from the wife’s family. Their model predicts that women from in the bottom two poverty quintiles according to the Sistema de
richer families are at increased risk of violence because there are Selección de Beneficiarios (SELBEN) index qualified for the transfer.
more resources to extract. Using this logic, Bobonis et al. (2009) In practice the change from Bono Solidario to the BDH meant that
develop a model for Mexico’s CCT program that predicts that 60% of families who were receiving the transfer under the Bono Sol-
cash transfers paid to women increase the threat of domestic vio- idario continued to receive it under the BDH. The other 40% stopped
lence with no associated physical violence because there are more receiving the transfer and a new set of families who were eligible
resources to extract. A crucial assumption of these extraction mod- were brought into the program. Our study focuses on this new set
els is that wives prefer an intact marriage over anything else. Thus, of families who were gradually brought into the program.
in contexts where divorce and separation are feasible and accept- The BDH was initially designed as a CCT program that consisted
able options, the conclusions of their models might not hold. of giving mothers a monthly stipend if their children met certain
Given that the predictions of the theoretical models are mixed, schooling and health care requirements. However, due to adminis-
it is not surprising that the empirical evidence on the associ- trative constraints the conditions were never implemented, thus
ation between wealth or socio-economic status and domestic making the BDH equivalent to an unconditional cash transfer.
violence is also mixed. In two recent studies on low and middle Another feature of the BDH is that the transfer size does not depend
income countries that investigated the association between socio- on the number of children in the household. The size of the transfer
economic status and having ever experienced domestic violence, in 2005 was $15/month for all households, which was approxi-
some countries reported a significant protective association, others mately 6–10% of an average household’s pre-transfer expenditure
revealed no significant trend or association, and others showed that for households in the bottom two poverty quintiles (Fiszbein et al.,
higher income was associated with more physical violence (Hindin 2009; Paxson and Schady, 2010).
et al., 2008; Vyas and Watts, 2009). The association between female
income and domestic violence is also ambiguous, with it showing 3.2. Study design
a positive association in some countries, no association in others
and a negative association in others (Vyas and Watts, 2009). The Beginning in 2003, the BDH was rolled out slowly across the
main reasons given for the heterogeneous findings have to do with country to new families that qualified for the transfer. This gradual
temporal and contextual factors, such as whether or not it is com- roll-out allowed two separate randomized evaluations to be con-
mon for women to work outside the home. Female education and ducted: a health evaluation (Paxson and Schady, 2010; Fernald and
age at marriage are also factors that have been shown to affect the Hidrobo, 2011) and an education evaluation (Schady and Araujo,
relationship between work and domestic violence (Heath, 2012). 2008). The study reported here uses the data collected for the
The previous studies that have looked at the association health evaluation where parishes (the smallest administrative unit
between income and domestic violence often fail to establish a in Ecuador) from six provinces (three coastal and three highland
causal relationship due to issues of omitted variable bias or reverse provinces) were randomized into treatment and control groups.
causality inherent in cross-sectional studies (Vyas and Watts, 2009; Parishes were stratified into rural and urban groups and from each
Aizer, 2011). We are able to overcome these issues by taking advan- group, treatment and control parishes were randomly selected. In
tage of the randomized roll-out of Ecuador’s unconditional cash total, 118 parishes were selected for the experiment: 79 in the
transfer program to mothers, the Bono de Desarrollo Humano treatment group and 39 in the control group.
 Author's personal copy

M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319 307

The BDH became available to rural treatment parishes in June reports yes to having been pushed, hit, or physically harmed by
2004 and to urban treatment parishes in November 2004. The her partner.6
transfers were not available to control parishes until after 2006. Psychological violence is defined by the WHO as “Any act or
Transfers were administered by the Government of Ecuador and omission that damages the self-esteem, identity, or development
distributed through a large network of private banks (Banred) of the individual. It includes, but is not limited to humiliation,
and through the National Agricultural Bank (Banco Nacional de threatening loss of custody of children, forced isolation from fam-
Fomento). According to survey response data from 2005, there was ily or friends, threatening to harm the individual or someone they
very little contamination of the experiment: take-up of the BDH care about, repeated yelling or degradation, inducing fear through
program was 73% for the treatment group and only 3% for the intimidating words or gestures, controlling behavior, . . .” (Ellsberg
control group. and Heise, 2005, p. 93). Operationally, WHO divides this definition
Baseline (October 2003–March 2004) and follow-up (September of psychological violence into “emotional violence” and “control-
2005–January 2006) surveys were conducted by the World Bank ling behavior by a partner” (García-Moreno et al., 2005). In the
and the Government of Ecuador. In order to be included in the base- follow-up survey there are four questions that can be categorized as
line survey sample, households had to meet the following criteria: “emotional violence”, three that can be categorized as “controlling
have at least one preschool age child, have no children older than behavior” and two that can be categorized as “emotional support”
6 years old, be eligible for the cash transfer program, and not have (see Table A1 in Appendix A for questions and categories). Given
been recipients of the previous welfare program, Bono Solidario.3 that the WHO only uses emotional violence and controlling behav-
Households in treatment and control parishes that met the criteria iors in their definitions, we concentrate on these two categories
above were then randomly sampled for the evaluation. The total for the construction of our psychological violence variables. For
sample at baseline consisted of 3426 households. Given that our each of the emotional violence or controlling behavior questions,
study focuses on spousal domestic violence, our analysis is con- the survey asks if a husband or partner “frequently”, “sometimes”,
ducted only on mothers who were living with their husbands or or “never” exhibits a specific behavior. We create an indicator for
partners at the baseline survey. Thus, of the 3426 mothers, we use emotional violence that equals one if the respondent answered
2354 for our analysis (Fig. 1).4 frequently or sometimes to any of the four emotional violence
questions. For controlling behavior, we create an indicator that
equals one if the respondent answered “frequently” or “sometimes”
3.3. Data to any of the three controlling behavior questions.
Although we follow WHO and DHS protocol and construct
The baseline and follow-up surveys collected information on binary indicators for each type of violence (controlling behaviors,
children and mothers’ health status and health care usage. For this emotional violence, and physical violence) (García-Moreno et al.,
study we concentrate on the spousal domestic violence data that 2005; Hindin et al., 2008), we also analyze each of the seven con-
was collected in the baseline and follow-up surveys. A list of the trolling behaviors or emotional violence indicators separately. This
domestic violence questions that were administered can be found gives us a concrete sense of individual indicators and their mag-
in Appendix A (Table A1). Unfortunately, the questions on domestic nitude. Each indicator takes the value of one if the respondent
violence were not restricted to a specific time period of reference. replied “frequently” or “sometimes”. As a check for robustness,
However, the questions on psychological violence were phrased in we also conduct our analysis using ordered probits to capture
the present instead of past tense thus eliciting current reports of the intensity of domestic violence (i.e., 0 = never, 1 = sometimes,
abuse instead of past reports.5 2 = frequently) and our results are consistent (results presented in
Domestic violence is a multidimensional and complex issue Appendix A).
that is usually categorized into physical violence, psychological One problem that we encounter when performing separate
violence, and sexual violence. We concentrate on physical and hypothesis testing for each of the seven psychological indicators,
psychological violence because data on sexual violence was not however, is that the probability that we wrongly reject a true null
collected. Physical violence is defined by the World Health Orga- hypothesis for at least one outcome is greater than the significance
nization (WHO) as: “The intentional use of physical force with level used for each test (Duflo et al., 2007; Kling et al., 2007). Conse-
the potential for causing death, injury, or harm. Physical violence quently, we also aggregate the information from all 7 psychological
includes, but is not limited to scratching, pushing, shoving, throw- violence indicators to create a summary index. Similar to Kling et al.
ing, grabbing, biting, choking, shaking, poking, slapping, punching, (2007), the summary index is constructed by averaging the z-scores
hitting. . .”(Ellsberg and Heise, 2005, p. 93). In the follow-up survey of its components. The z-scores for each component are calculated
there are two questions on physical violence. The first question using the control group’s mean and standard deviation. Higher val-
asks whether the mother has ever been pushed, hit or physically ues of the summary index indicate more violence. We construct the
harmed by her husband or partner and the second asks whether psychological summary index twice, once using the 7 dichotomous
it occurred frequently or sometimes. We use the first question to indicators and again using the 7 ordinal variables that are meant
create a physical violence indicator that equals one if a mother to capture the frequency of violence. We present the results for the
latter in Appendix A.

3.4. Attrition
3
Paxson and Schady (2010) investigated whether there were systematic differ-
ences across those who were “newly eligible” and those who were already receiving
the BDH. They find no differences in terms of poverty but newly eligible households Of the 2354 mothers living with their partners at baseline, 2190
are younger. While our results cannot be generalized to older households, there (93.0%) were re-interviewed at follow-up. However, the domestic
is no concern with the internal validity of our estimates because households were
randomly drawn from the same sample frame.
4
Results are similar if we conduct the analysis on all mothers in relationships at
6
follow-up and do not limit it to mothers that were in relationships at baseline. We also construct a variable that equals 0 for mothers never physically abused,
5
More importantly, as Section 3.5 will reveal, at baseline the domestic violence 1 for mothers sometimes physically abused and 2 for mothers that are frequently
indicators were balanced across treatment and control arms, and therefore, we can abused. Results however are similar to those for the dichotomous indicator and thus
conclude that any changes we might observe are due to the introduction of the BDH. we only report the dichotomous indicator which is more commonly used.
 Author's personal copy

308 M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319

Fig. 1. Flow of participants.

violence questions could only be administered to mothers if they analysis. Of this “study sample”, 1236 have complete data for the
were still living with their partners (n = 2028). Given the highly emotional violence questions, 1218 for the controlling behaviors
sensitive nature of the domestic violence questions, only 1254 questions, and 1246 for the physical violence questions.
(or 61.8%) of the 2028 mothers that were in relationships were As a consequence of the sensitivity and requirements for being
administered the questions (Fig. 1). The reasons provided for not administered the domestic violence questions described above,
answering the domestic violence questions were: husband was attrition in our sample is high. If attrition is correlated with treat-
present (80.1%), husband’s relative was present (11.5%), another ment assignment, then this could potentially bias the estimates of
person was present (4.7%), did not want to answer or unsure about the impact of a cash transfer on domestic violence. Table 1 how-
answering (3.1%), and other (0.6%). Although survey enumerators ever reveals that attrition is not correlated with treatment at any
were trained to not administer the domestic violence questions if of the three time points when attrition occurred: from baseline
a partner was present, four mothers were administered the ques- to follow-up; at follow-up for mothers who were still living with
tions with their partners present. Given that the presence of their their partners; and at follow-up for mothers who were adminis-
partners could influence their response, we drop these 4 mothers tered the domestic violence questions without their husband or
from our sample and thus we are left with 1250 mothers for our others present.
 Author's personal copy

M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319 309

Table 1
Attrition analysis.

Dependent variable = 1 if mother is. . .

In follow-up In follow-up + lives with partner In follow-up + lives with partner + administered

Treatment 0.01 0.01 0.01

(0.01) (0.02) (0.03)

Observations 2354 2354 2354

Marginal effects from probit regressions reported. Standard errors clustered at the parish level in parentheses. Column 1: dependent variable equals one if mother is in the
follow-up. Column 2: dependent variable equals one if mother is in the follow-up and is living with husband or partner. Column 3: dependent variable equals one if mother
is in the follow-up, is living with husband or partner, and is administered the domestic violence questions without her husband present.
*
p < 0.10.
**
p < 0.05.
***
p < .01.

Table 2
Attrition analysis by baseline characteristics (means).

Control group Treatment group Difference

(1) (2) (3) (4) (5) (6) (7) (8)

Attrited In study p-Value Attrited In study p-Value Col(1)-Col(4) p-Value

Mother’s age (years) 23.22 23.67 0.22 23.34 23.57 0.32 −0.11 0.72
Mother is married 0.45 0.47 0.64 0.40 0.45 0.04 0.05 0.09
Mother is indigenous 0.05 0.04 0.65 0.06 0.06 0.53 −0.01 0.45
Mother is Afro-Ecuadorian 0.11 0.10 0.61 0.10 0.10 0.84 0.01 0.56
Mother has no health insurance 0.93 0.95 0.16 0.94 0.94 0.87 −0.01 0.41
Years of schooling (mother) 6.86 7.72 0.00 7.20 7.45 0.11 −0.34 0.08
Years of schooling (partner) 6.86 7.39 0.02 6.97 7.28 0.05 −0.11 0.56
Schooling gap (partner’s schooling − mother’s schooling) −0.01 −0.33 0.16 −0.24 −0.18 0.68 0.24 0.22
Mother has had a child die 0.08 0.07 0.80 0.06 0.07 0.45 0.01 0.38
Mother uses birth control methods 0.50 0.55 0.13 0.53 0.57 0.19 −0.04 0.26
Mother has poor health 0.10 0.09 0.66 0.09 0.08 0.99 0.01 0.47
Mother is currently pregnant 0.15 0.13 0.46 0.12 0.12 0.88 0.03 0.23
Number of children 0–5 years old 1.77 1.80 0.49 1.76 1.77 0.71 0.01 0.89
Asset index −0.32 0.10 0.01 −0.16 0.06 0.04 −0.16 0.26
Kitchen is used for sleeping 0.27 0.25 0.37 0.29 0.26 0.20 −0.01 0.66
Urban 0.46 0.54 0.03 0.51 0.49 0.44 −0.05 0.10
Husband or partner. . ..
Is not affectionate with you 0.32 0.32 0.90 0.37 0.33 0.21 −0.05 0.28
Doesn’t spend free time with you 0.45 0.42 0.54 0.35 0.37 0.58 0.10 0.03
Ignores you 0.30 0.31 0.79 0.32 0.34 0.59 −0.02 0.58
Doesn’t allow you to study or work 0.31 0.41 0.05 0.34 0.31 0.35 −0.03 0.50
Doesn’t allow you to see friends or family 0.21 0.21 1.00 0.24 0.24 0.81 −0.03 0.44
Yells at you 0.35 0.40 0.26 0.41 0.39 0.62 −0.06 0.17
Tells you your worthless 0.27 0.27 0.94 0.27 0.29 0.49 −0.00 0.97
Threatens to leave you 0.19 0.25 0.15 0.23 0.21 0.43 −0.05 0.22
Threatens to take your children away 0.30 0.29 0.80 0.29 0.27 0.50 0.01 0.73
Physical violence 0.29 0.30 0.83 0.27 0.27 0.95 0.02 0.70
Emotional violence 0.54 0.56 0.74 0.51 0.52 0.71 0.03 0.52
Controlling behaviors 0.51 0.57 0.22 0.53 0.55 0.44 −0.02 0.66

In columns 3 and 6 p-values are reported from t-tests on the equality of means for each variable between the “In Study” and “Attrited” groups. Column 7 reports the difference
in means between the “Attrited” group in the control arm and the “Attrited” group in the treatment arm. Column 8 reports the p-values for the difference in means between
the two “Attrited” groups. “In study” sample consists of mothers that are in the follow-up study and are eligible and willing to be administered the domestic violence
questions. “Attrited” refers to mothers that were living with their partners in the baseline survey but were not in the follow-up or were not administered the domestic
violence questions at follow-up. Emotional violence indicator equals one if mother answered frequently or sometimes to any of the four emotional violence questions (see
Appendix A for classification). Controlling behavior equals one if the respondent answered frequently or sometimes to any of the three controlling behavior questions.

Table 2 examines baseline characteristics of those who were we focus on columns 7 and 8 of Table 2. With the exception of an
administered the domestic violence questions at follow-up (“In indicator for husband or partner not spending free time with his
study”) and those who were not (“Attrited”). Given the require- wife or partner,7 there are no statistically significant differences
ments for being administered the domestic violence questions, we at the 5% level in baseline characteristics for those not in the study
expect there to be differences between those in the study sample
and those not in the study sample; however, differential attrition
across treatment and control arms would threaten the internal
7
The differential attrition with respect to spending free time with wife or partner
validity of the study. In particular, if mothers who experience more
could potentially bias the treatment effect toward zero because not spending free
domestic violence leave the treatment arm in greater proportions time with wife is positively correlated with domestic violence. In other words, the
than the control arm, then our treatment estimates would be higher rate of partners not spending time with their wives that leave the control
biased because any decrease in domestic violence would be due could bias our estimates because those that remain in the control group would
to both treatment and differential attrition. In order to examine display lower rates of domestic violence. However, for our study sample domestic
violence indicators are balanced at baseline, which reduces our concern of bias due
if differential attrition threatens the internal validity of the study,
to differential attrition.
 Author's personal copy

310 M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319

Table 3
Descriptive statistics (means) of study sample from baseline.

N Study sample Control Treatment p-Value

Mother and household characteristics

Mother’s age (years) 1250 23.60 23.67 23.57 0.72
Mother is married 1250 0.45 0.47 0.45 0.53
Mother is indigenous 1250 0.05 0.04 0.06 0.38
Mother is Afro-Ecuadorian 1250 0.10 0.10 0.10 0.88
Mother has no health insurance 1250 0.94 0.95 0.94 0.36
Years of schooling (mother) 1250 7.54 7.72 7.45 0.15
Years of schooling (partner) 1250 7.31 7.39 7.28 0.53
Schooling gap (partner’s schooling − mother’s schooling) 1250 −0.23 −0.33 −0.18 0.46
Mother has had a child die 1250 0.07 0.07 0.07 0.97
Mother uses birth control methods 1249 0.56 0.55 0.57 0.60
Mother has poor health 1250 0.09 0.09 0.08 0.79
Mother is currently pregnant 1250 0.12 0.13 0.12 0.63
Number of children 0–5 years old 1250 1.78 1.80 1.77 0.51
Asset index 1250 0.08 0.10 0.06 0.77
Kitchen is used for sleeping 1250 0.25 0.25 0.26 0.65
Urban 1250 0.51 0.54 0.49 0.15
Domestic violence indicators
Husband or partner. . .
Is not affectionate with you 653 0.32 0.32 0.33 0.78
Doesn’t spend free time with you 653 0.39 0.42 0.37 0.27
Ignores you 653 0.33 0.31 0.34 0.46
Doesn’t allow you to study or work 653 0.34 0.41 0.31 0.01
Doesn’t allow you to see friends or family 653 0.23 0.21 0.24 0.29
Yells at you 652 0.39 0.40 0.39 0.78
Tells you your worthless 653 0.29 0.27 0.29 0.60
Threatens to leave you 653 0.22 0.25 0.21 0.29
Threatens to take your children away 653 0.28 0.29 0.27 0.53
Physical violence 653 0.28 0.30 0.27 0.52
Emotional violence 652 0.53 0.56 0.52 0.42
Controlling behaviors 653 0.56 0.57 0.55 0.71

Study sample refers to mothers who lived with their partners at baseline, are in the follow-up, and are eligible and willing to answer the domestic violence questions. Only
a subset of mothers in the study sample answered the domestic violence questions at baseline. Emotional violence indicator equals one if mother answered frequently or
sometimes to any of the four emotional violence questions (see Appendix A for classification). Controlling behavior equals one if the respondent answered frequently or
sometimes to any of the three controlling behavior questions.

across treatment and control arms (column 8). The differences treatment and control arms, and this is the variable we use in our
in spending free time with wife or partner across treatment and analysis.9
control arms for our study sample however are not significant The prevalence of domestic violence is high, with 28% of mothers
(Table 3), suggesting that the bias due to the differential attrition in the study sample experiencing physical violence; 53% experienc-
of this variable is likely to be very small. ing emotional violence and 56% experiencing controlling behaviors
by husband or partner (Table 3). Of the mothers that report expe-
3.5. Baseline analysis riencing emotional violence, 47% also report experiencing physical
violence and 71% also report experiencing controlling behavior by
Given the restrictions for being administered the domestic vio- husband or partner.10 Mothers in the study sample are young with
lence questions only a fraction of mothers that were administered a mean age of 23.6 years; they have more years of schooling than
the questions at follow-up were also administered them at base- their partners (7.5 years versus 7.3 years), and only 45% of mothers
line. Although we do not have complete baseline information on living with their partners are married.
domestic violence for our study sample, we do have complete
information on all other household and mother characteristics. 4. Empirical analysis and results
Table 3 shows that random assignment was effective at balanc-
ing baseline characteristics of treatment and control groups for 4.1. Empirical strategy
our study sample of 1250 mothers.8 Of the 28 variables exam-
ined, 27 are not different between groups; the only variable that To estimate the treatment effect of the BDH on domestic vio-
is significantly different between treatment and control groups is lence we take advantage of the randomized roll-out of the program
the indicator for whether or not the husband or partner allows and conduct an intent-to-treat analysis. This approach avoids any
his wife or partner to study or work. In the control group, 41% of
husbands or partners do not allow their wives to study or work
compared to 31% in the treatment group. Although this control- 9
While we use all three controlling behavior questions to create the aggregate
ling behavior indicator shows a significant difference between
“controlling behavior” variable, our results are robust to creating the variable using
groups, the aggregate “controlling behavior” variable (which is only the two controlling behavior questions (husband or partner ignores you; hus-
made up of the 3 controlling behavior questions) is balanced across band or partner does not allow you to see friends or family) that are balanced across
treatment and control groups.
10
Similarly, of the mothers that report experiencing controlling behaviors, 43%
also report experiencing physical violence and 73% also report experiencing emo-
8
Paxson and Schady (2010) and Fernald and Hidrobo (2011) also show that ran- tional violence. Of the mothers that report experiencing physical violence, 77% also
dom assignment was effective at balancing baseline characteristics of treatment and report experiencing emotional violence and 68% also report experiencing control-
control groups for the full sample of households. ling behaviors.
 Author's personal copy

M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319 311

bias that might occur due to selection into and out of the program. adding control variables leads to minimal changes in the size of
We conduct all estimations using linear probability models; how- the coefficients on treatment. Being in the treatment group has
ever, results are robust to using probit models. The specific equation no effect on emotional and physical violence, and a negative and
that we estimate is the following: significant impact on controlling behaviors. Table 5 examines the
impact of the BDH on each of the 7 psychological violence indi-
DVip = ı + ˇTp + ˛Xip + εip (1)
cators used to construct the emotional violence and controlling
where DVip is the domestic violence indicator (physical violence, behaviors indicators. While the treatment effect is negative for all
emotional violence, or controlling behavior) for mother i from indicators, the BDH only leads to a significant decrease in the prob-
parish p. Tp is an indicator for whether parish p is in the treatment ability that a partner does not allow his wife or partner to see her
group and Xip is a vector of control variables. ˇ measures the aver- friends or family; and a marginally significant decrease in the prob-
age effect of the BDH on domestic violence. Given the success of the ability that a partner does not allow his wife or partner to study or
random assignment, the inclusion of baseline controls is not neces- work. The results from the ordered probit (which captures the fre-
sary to obtain unbiased estimates of ˇ. In most estimates, however, quency of abuse) are slightly stronger and demonstrate a significant
we include baseline mother (age, years of schooling, race, mari- decrease in all three controlling behaviors, and consequently, there
tal status, partner’s years of schooling, whether mother has had a is also a significant decrease in the aggregate psychological index
child die, whether a mother is currently pregnant) and household (Appendix A, Table A2).
(number of children 0–5 years old, urban indicator, asset index,11 Table 4 also reveals that domestic violence is more likely to
province indicators, and an indicator for whether or not kitchen is occur if a woman is indigenous, not married, and younger; if her
used for sleeping) control variables in order to increase the pre- husband or partner has less schooling; and if she lives in rural areas,
cision of the estimates. In all regressions we cluster the standard has many children 0–5 years old, and her kitchen is used for sleep-
errors at the parish level. ing. The positive correlation between domestic violence and having
In addition to estimating the average effect of the BDH on the kitchen used for sleeping suggests that household crowding and
domestic violence, we estimate the differential effect with respect poverty are stressors on relationships. Even though the provinces
to a mother’s education because women with more education are in the study are not composed of large indigenous populations (the
likely to have higher reservation utilities. In Ecuador, secondary percent of indigenous mothers in our sample is 5%), we still observe
education begins after 6 years of formal schooling, and in our a large and positive correlation of indigenous status on domestic
study sample, 46% of mothers have more than primary educa- violence. Specifically, indigenous women are 10 percentage points
tion. Moreover, the returns to an additional year of schooling in more likely to experience controlling behaviors from their partners
secondary school are much higher than the returns to an addi- and 16 percentage points more likely to report physical abuse than
tional year of schooling in primary school (9% versus 4.2% in 2005) non-indigenous women. There are many socio-economic factors
(Figueroa and Miranda, 2009). Consequently, we use this cut-off to that could explain this strong positive relationship such as the fact
separate mothers into “more educated” and “less educated” groups. that indigenous women are poorer, have less education, and have
However, our results are robust to using a cutoff of 8 years of school- less access to health, legal services, and employment opportunities
ing. We also estimate the differential effect of the BDH with respect (Hughes, 2004; Hall and Patrinos, 2006).
to a mother’s relative education to her partner’s because studies
have found that women with more education or income than their 4.3. Heterogeneous effects
partners experience more domestic violence. The equations that
we estimate are the following: Table 6 presents the heterogeneous impact of the BDH with
respect to a mother’s education. Given that a mother’s education
DVip = ı + ˇ1 Tp + ˇ2 Tp × Sip + ˇ3 Sip + ˛Xip + εip (2) is highly correlated with wealth and race, in columns 4–6 we con-
where Sip is either an indicator that equals one if a mother has trol for the interaction of these variables with treatment, in order
more than primary education (more than 6 years of schooling) or to isolate the heterogeneous impact with respect to a mother’s
an indicator that equals one if a mother has less schooling than her education. The BDH has no effect on any of the domestic violence
partner. ˇ1 measures the effect of the BDH for the omitted category indicators for mothers with 6 years or less of schooling, but for
(mothers with 6 years or less of schooling or mothers with equal to mothers with more than 6 years of schooling the effect of the BDH
or more schooling than their partners), while ˇ1 + ˇ2 measures the is negative and significant for emotional violence and controlling
effect of the BDH for mothers with more than 6 years of schooling behaviors, even after controlling for other possible interactions.
or mothers with less schooling than their partners. Thus, ˇ2 is the Specifically, for mothers with more education the BDH decreases
differential impact of the BDH with respect to a mother’s absolute the probability that a husband or partner engages in emotional
or relative education. violence by 8 percentage points and controlling behaviors by 14
percentage points. The treatment effect for mothers with more
4.2. Average effects education is significantly different than that for mothers with less
education for both emotional violence and controlling behaviors.
Table 4 presents the results of estimating the effect of a cash The significant decrease in emotional violence for mothers with
transfer on emotional violence (columns 1 and 4), controlling more than 6 years of schooling is due to significant decreases in 2
behavior (columns 2 and 5), and physical violence (columns 3 and out of the 4 emotional violence indicators (Table 7). Similarly, the
6). In the first three columns we estimate the treatment effect with- significant decrease in the probability that a partner exhibits con-
out any controls, and in the last three columns we add control trolling behaviors for mothers with more than 6 years of schooling
variables. As expected given the successful random assignment, is due to significant decreases in all 3 controlling behavior indica-
tors. The results from the specifications that capture the frequency
of violence are similar to those in Tables 6 and 7 and reveal no
11
impact of the BDH for mothers with 6 years or less of schooling,
Asset index is constructed using the first principal from a Principal Component
Analysis. Variables used to construct the index are housing infrastructure indicators
and a negative and significant impact for mothers with more than
(e.g. type of floor, roof, toilet, light, fuel, and water source) and asset indicators (e.g. 6 years of schooling (Table A3). There is no significant effect of the
refrigerator, small stove, TV, car, motorcycle, and computer). BDH on physical violence for any group of mothers.
 Author's personal copy

312 M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319

Table 4
Average effect of the BDH on domestic violence.

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

Emotional Controlling Physical Emotional Controlling Physical

Treatment −0.01 −0.05* −0.01 −0.02 −0.06** −0.02

(0.03) (0.03) (0.03) (0.03) (0.03) (0.03)
Years of schooling (mother) −0.01 −0.00 0.01
(0.01) (0.01) (0.00)
Years of schooling (partner) −0.01 −0.01** −0.01**
(0.01) (0.00) (0.00)
Mother’s age (years) 0.01 −0.00 −0.01**
(0.00) (0.00) (0.00)
Mother is married −0.08** −0.02 −0.02
(0.03) (0.04) (0.03)
Mother is Afro-Ecuadorian −0.10* −0.06 −0.03
(0.06) (0.05) (0.05)
Mother is indigenous 0.09 0.10* 0.16***
(0.07) (0.06) (0.05)
Mother is currently pregnant 0.02 −0.07 −0.05
(0.05) (0.04) (0.04)
Mother has had a child die −0.07 0.02 0.01
(0.05) (0.05) (0.05)
Number of children 0–5 years old 0.03 0.00 0.05***
(0.02) (0.02) (0.02)
Kitchen is used for sleeping 0.09*** 0.01 0.06*
(0.03) (0.04) (0.03)
Asset index 0.01 0.01 0.01
(0.01) (0.01) (0.01)
Asset index squared −0.01*** −0.01** −0.00
(0.00) (0.00) (0.00)
Urban −0.05* −0.04 −0.05*
(0.03) (0.03) (0.03)
Constant 0.53*** 0.55*** 0.29*** 0.54*** 0.71*** 0.39***
(0.04) (0.04) (0.03) (0.11) (0.11) (0.10)

Observations 1236 1218 1246 1236 1218 1246

Standard errors clustered at the parish level. All estimations contain province indicators.
*
p < 0.10.
**
p < 0.05.
***
p < .01.

Table 5
Average effect of the BDH on psychological violence.

Does not allow you to Does not allow you to Ignores Yells at Tells you your Threatens to Threatens to Psych.
see friends or family study or work you you worthless leave take children index

Treatment −0.06*** −0.05* −0.04 −0.03 −0.03 −0.03 −0.03 −0.09**

(0.02) (0.03) (0.03) (0.03) (0.03) (0.03) (0.02) (0.04)

Observations 1245 1229 1238 1246 1247 1245 1240 1209

In first 7 columns, outcome variables equal 0 if mother replies never and 1 if mother replies sometimes or frequently. Aggregate psychological index is a z-score constructed
by averaging the z-scores from each of the 7 psychological outcome indicators. Standard errors are clustered at the parish level. All estimations control for baseline mother
characteristics (age, indicator for whether mother is married, indicators for whether mother is indigenous or Afro-Ecuadorian, mother’s years of schooling, partner’s years of
schooling, indicator for whether mother has had a child die, indicator for whether mother is currently pregnant), and baseline household characteristics (asset index, asset
index squared, urban, number of children 0–5 years old, whether kitchen is used for sleeping, province indicators).
*
p < 0.10.
**
p < 0.05.
***
p < .01.

Given that the impact of the BDH on domestic violence depends the increase in emotional violence for this group of mothers is due
on a mother’s education, we estimate the heterogeneous effect of to a significant increase in the probability that a husband or partner
the BDH with respect to a mother’s relative education separately for yells at her (results not shown). The BDH has the largest negative
mothers with more than primary education and those with primary impact on domestic violence for mothers with more than 6 years
education or less. The differential impact of the BDH with respect to of schooling who still have less schooling than their partners. For
mother’s having less schooling than their partners is negative and this group of mothers, the BDH leads to a significant decrease in
significant for the emotional violence indicator for both mothers emotional violence by 27 percentage points, controlling behavior
with more than six years of schooling and less than or equal to six by 17 percentage points, and to a large but non-significant decrease
years of schooling (Table 8). However, the differential impact is not in physical violence by 10 percentage points.
significant for the controlling behavior indicator or for physical vio-
lence. More importantly, the BDH significantly increases emotional 4.4. Robustness
violence by 9 percentage points for mothers with 6 years or less of
schooling in households where they have at least as much schooling One concern with the estimates reported above is that they
as their partners (column 1). Looking at each indicator individually, may be confounded by self-reporting bias. Domestic violence is a
 Author's personal copy

M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319 313

Table 6
Heterogeneous effect of the BDH with respect to mother’s education.

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

Emotional Controlling Physical Emotional Controlling Physical

Treatment 0.02 −0.01 0.00 0.03 0.02 0.01

(0.04) (0.04) (0.04) (0.04) (0.04) (0.04)
More than 6 years of schooling (mother) × treatment −0.10* −0.11 −0.04 −0.11* −0.16** −0.04
(0.06) (0.07) (0.06) (0.06) (0.07) (0.06)
More than 6 years of schooling (mother) 0.08 0.07 0.08 0.09* 0.09 0.08
(0.05) (0.06) (0.05) (0.05) (0.07) (0.05)

Controls for interaction of treatment with indigenous No No No Yes Yes Yes

indicator and asset index
Observations 1236 1218 1246 1236 1218 1246
Treatment effect for mothers with more than 6 years of −0.07* (0.04) −0.12** (0.05) −0.04(0.04) −0.08* (0.04) −0.14*** (0.05) −0.03(0.04)
schooling

Standard errors clustered at the parish level. All estimations control for baseline mother characteristics (age, indicator for whether mother is married, indicators for whether
mother is indigenous or Afro-Ecuadorian, partner’s years of schooling, indicator for whether mother has had a child die, indicator for whether mother is currently pregnant),
and baseline household characteristics (asset index, asset index squared, urban, number of children 0–5 years old, whether kitchen is used for sleeping, province indicators).
*
p < 0.10.
**
p < 0.05.
***
p < .01.

Table 7
Heterogeneous effect of the BDH with respect to mother’s education on psychological violence.

Does not allow you to Does not allow you Ignores Yells at Tells you your Threatens to Threatens to Psych.
see friends or family to study or work you you worthless leave take children index

Treatment −0.04 0.04 0.01 0.04 −0.00 0.03 −0.02 0.02

(0.03) (0.04) (0.04) (0.05) (0.04) (0.04) (0.04) (0.06)
More than 6 years of schooling −0.05 −0.16*** −0.13** −0.16*** −0.06 −0.11** −0.04 −0.23***
(mother) × treatment (0.06) (0.06) (0.06) (0.06) (0.06) (0.05) (0.06) (0.09)
More than 6 years of schooling 0.02 0.08 0.08 0.09* 0.03 0.10** 0.02 0.14*
(mother) (0.05) (0.05) (0.05) (0.05) (0.04) (0.04) (0.04) (0.07)

Observations 1245 1229 1238 1246 1247 1245 1240 1209

Treatment effect for mothers with −0.09** −0.12*** −0.12*** −0.12*** −0.06 −0.09** −0.05 −0.21***
more than 6 years of schooling (0.04) (0.04) (0.04) (0.04) (0.04) (0.04) (0.04) (0.06)

In first 7 columns, outcome variables equal 0 if mother replies never and 1 if mother replies sometimes or frequently. Aggregate psychological index is a z-score constructed
by averaging the z-scores from each of the 7 psychological outcome indicators. Standard errors clustered at the parish level. All estimations control for baseline mother
characteristics (age, indicator for whether mother is married, indicators for whether mother is indigenous or Afro-Ecuadorian, partner’s years of schooling, indicator for
whether mother has had a child die, indicator for whether mother is currently pregnant), and baseline household characteristics (asset index, asset index squared, urban,
number of children 0–5 years old, whether kitchen is used for sleeping, province indicators).
*
p < 0.10.
**
p < 0.05.
***
p < .01.

sensitive issue and thus subject to a high degree of under-reporting measures were taken to reduce under-reporting (such as ensuring
(Ellsberg et al., 2001). The most important factors influencing dis- that the partner/husband was not in the household at the time
closure are the adequate training of interviewers, safety measures of the interview), we cannot dismiss this issue completely. If
that ensure privacy during interviews, and administering multiple under-reporting is associated with treatment, and specifically,
behaviorally specific questions (Ellsberg et al., 2001). Although if those in the treatment group are more likely to under-report,

Table 8
Heterogeneous effect with respect to a mother’s relative education by a mother’s absolute education level.

Mothers with 6 years or less of school Mothers with more than 6 years of school

Emotional Controlling Physical Emotional Controlling Physical

Treatment 0.09* 0.02 0.01 −0.01 −0.11** −0.02

(0.05) (0.05) (0.05) (0.05) (0.05) (0.05)
Mother has less schooling than −0.17** −0.05 −0.03 −0.26** −0.06 −0.08
partner × treatment (0.07) (0.10) (0.08) (0.11) (0.10) (0.10)
Mother has less schooling than partner 0.13** 0.05 0.04 0.24** −0.04 −0.01
(0.06) (0.09) (0.06) (0.10) (0.09) (0.09)

Observations 668 653 672 568 565 574

Treatment effect for mothers with less −0.08 −0.03 −0.01 −0.27*** −0.17* −0.10
schooling than partner (0.06) (0.08) (0.06) (0.09) (0.10) (0.08)

Standard errors clustered at the parish level. All estimations control for baseline mother characteristics (age, indicator for whether mother is married, indicators for whether
mother is indigenous or Afro-Ecuadorian, partner’s years of schooling, indicator for whether mother has had a child die, indicator for whether mother is currently pregnant),
and baseline household characteristics (asset index, asset index squared, urban, number of children 0–5 years old, whether kitchen is used for sleeping, province indicators).
*
p < 0.10.
**
p < 0.05.
***
p < .01.
 Author's personal copy

314 M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319

which might occur if individuals believe their response to the his education relative to his wife or partner’s that is driving the
survey affects whether or not they received a transfer, then our positive relationship between an increase in a woman’s income
estimates would be biased away from zero for the following three and emotional violence. In order to rule out this interpretation, we
groups: (1) mothers with more than 6 years of schooling who estimate the effect of the BDH for partners with 6 years or less of
have less years of schooling than their partners, (2) mothers with schooling and we find no significant effect on emotional violence
more than 6 years of schooling whose years of schooling are equal for both mothers with more or less schooling (results not shown).
to or more than their partners, and (3) mothers with 6 years or Although treatment was not significantly associated with attri-
less of schooling who have fewer years of schooling than their tion (Table 1) and those who attrited are similar across treatment
partners. For the fourth group – mothers with 6 years or less of and control arms (Table 2), attrition in our sample is high and may
schooling whose years of schooling are equal to or more than their be correlated with unobserved heterogeneity in outcomes. Thus
partners – the estimates would be lower bound estimates. If those we conduct the following checks to address the issue nonrandom
that are in the control group are more likely to under-report, then sample selection: (1) we first explore the potential pathways
the opposite pattern in bias would occur. Given that there are no through which treatment could affect non-response; (2) we then
strong arguments for why we should expect those in the treatment re-run all our specifications using inverse probability weights
or control group to be more likely to under-report, any bias from (IPW); and (3) lastly we bound our estimates using Lee’s trimming
differential under-reporting is likely to be small. procedure (Lee, 2009).
Another concern with our inference is that while the BDH was The two main pathways through which treatment could affect
randomized across treatment and control parishes, the research non-response is by affecting the probability that a woman is not in a
design was not stratified across education groups. Consequently, relationship at follow-up or by affecting the probability that a part-
there could be pre-existing differences between treatment and ner is present at the time of the interview. While attrition due to the
control arms across the four groups analyzed. For example, for latter is much larger in our sample, we explore both pathways. If
those in the fourth group (mothers with 6 years or less of schooling, the BDH leads to women leaving their relationship and women who
mother’s education ≥ partner’s education), if the treatment arm have better outside options are the ones who most likely leave their
had a higher proportion of indigenous mothers than the control relationships as a result of the BDH, then our estimates would be
arm, then the higher probability of emotional violence in the lower bound estimates. In order to assess the magnitude of this bias
treatment arm could be due to this higher proportion of indige- we estimate the impact of treatment on the probability of women
nous mothers. Similarly, our previous attrition analysis was only leaving their partners and find no significant effect for any group
conducted on the overall sample and not across subsamples. To of mothers, and thus, bias due to women leaving their relationship
show that attrition and preexisting differences are not a concern, is likely to be small.12 Similarly, if the BDH leads to higher rates of
we compare attrition rates and baseline characteristics across unemployment by men and thus the increased likelihood of men
treatment and control arms for all four education groups that we being present at the time of the interview, then our estimates will
analyze. be biased. Although we do not have information on whether or not
Table 9 reveals that treatment is correlated with attrition for a partner is employed, we can estimate whether or not the BDH
mothers in the third group (mothers with 6 years or less of school- is associated with a partner’s presence at the time of the inter-
ing, mother’s education < partner’s education) but not for the other view. We find no significant associations of the BDH with a partner’s
three groups. If those that attrite are different from those that stay, presence for any group of mothers. Given that the presence of part-
then our estimates for this group of mothers may be biased. How- ners is the main cause of attrition in our sample, we also explore
ever, for this group of mothers, there are no observable differences other factors that are significantly correlated with a partner being
in baseline characteristics at the 5% level. Although attrition does present. We find that only interview month and province are sig-
not lead to observable differences, it may lead to unobservable dif- nificant predictors which suggest that interview timing more than
ferences, and thus we also bound our estimates (discussed in more systematic differences across households is driving attrition. In
detail below). particular, being interviewed in January or in Loja province signifi-
Across the 76 difference-of-means tests (4 groups × 19 vari- cantly increases the probability of a partner being present by 16 and
ables) there are only 3 significant differences in means between 9 percentage points respectively (results not shown).13 Neither of
treatment and control arms at the 5% level. For the group of these factors is correlated with any measure of domestic violence.
mothers with “more than 6 years of schooling, mother’s school- Although the findings above provide evidence for the exogene-
ing < partner’s schooling”, the control arm has a higher proportion ity of selection, we can further examine the robustness of our
(22% versus 6%) of Afro-Ecuadorians. This larger proportion of Afro- results using inverse probability weights. Under an ignorability
Ecuadorians in the control arm is most likely the reason that for this assumption, Wooldridge (2002) shows that inverse probability
group of mothers, there is also a significant difference in means weighting produces a consistent normal estimator which is iden-
for the asset index and having had a child die. Given that being tical to an unweighted estimate if there is no attrition problem.
Afro-Ecuadorian is negatively correlated with domestic violence, In the first stage of IPW, we predict the probability of being
having more Afro-Ecuadorians in the control arm will bias our esti-
mates toward zero for this group of mothers. Consequently, in all
our regressions we control for being Afro-Ecuadorian, the asset
index, and having a child die, and thus, the bias is likely to be small. 12
For a subsample of mothers who have answered the domestic violence questions
More importantly, Table 9 also shows that there are no significant at baseline we investigate whether the BDH leads to mothers leaving a relationship
for mothers in violent relationships at baseline and find no significant effect across
differences between treatment and control arms on any of the base-
any group of mothers.
line domestic violence indicators (emotional violence, controlling 13
Probit regressions are conducted of the probability that a partner is present
behaviors, and physical violence) for any of the four groups studied. on month of interview indicators, baseline mother characteristics (age, indicator
A close examination of Table 9 reveals that the average years of for whether mother is married, indicators for whether mother is indigenous or
schooling for partners is very low (5 years) for households where Afro-Ecuadorian, mother’s years of schooling, partner’s years of schooling, indicator
for whether mother has had a child die, indicator for whether mother is currently
mothers have 6 years or less of schooling and whose years of pregnant), and baseline household characteristics (asset index, asset index squared,
schooling are equal to or more than their partners’. Consequently, urban, number of children 0-5 years old, whether kitchen is used for sleeping,
it could be a partner’s low education levels and not necessarily province indicators).
 Author's personal copy

M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319 315

Table 9
Baseline comparison of treatment and control stratified by education groups.

Mothers with more than 6 years of school

Mother’s edu < partner’s edu Mother’s edu ≥ partner’s edu

Control Treat p-Value Control Treat p-Value

Attrition rates 0.40 0.47 0.28 0.41 0.44 0.46

Baseline characteristics
Mother’s age (years) 22.88 23.33 0.56 23.61 23.66 0.90
Mother is married 0.44 0.43 0.91 0.43 0.43 0.89
Mother is indigenous 0.02 0.06 0.39 0.02 0.02 0.92
Mother is Afro-Ecuadorian 0.22 0.06 0.01 0.10 0.17 0.06
Mother has no health insurance 0.95 0.99 0.21 0.96 0.94 0.36
Years of schooling (mother) 9.00 9.25 0.39 10.71 10.46 0.23
Years of schooling (partner) 11.80 11.95 0.72 7.78 7.49 0.31
Schooling gap (partner’s schooling − mother’s schooling) 2.80 2.70 0.76 −2.93 −2.97 0.87
Mother has had a child die 0.07 0.00 0.01 0.07 0.04 0.25
Mother uses birth control methods 0.63 0.70 0.45 0.66 0.64 0.71
Mother has poor health 0.05 0.07 0.63 0.07 0.06 0.72
Mother is currently pregnant 0.17 0.07 0.09 0.07 0.11 0.15
Number of children 0–5 years old 1.68 1.80 0.41 1.75 1.61 0.06
Asset index 0.98 1.75 0.04 0.79 0.68 0.57
Kitchen is used for sleeping 0.15 0.17 0.77 0.27 0.29 0.66
Urban 0.56 0.60 0.72 0.65 0.57 0.10
Domestic violence indicators
Physical 0.18 0.19 0.93 0.35 0.27 0.24
Emotional 0.50 0.43 0.59 0.55 0.49 0.40
Controlling 0.55 0.45 0.49 0.56 0.55 0.90

Mothers with 6 years or less of school

Mother’s edu < partner’s edu Mother’s edu ≥ partner’s edu

Control Treat p-Value Control Treat p-Value

Attrition rates 0.56 0.45 0.02 0.50 0.49 0.84

Baseline characteristics
Mother’s age (years) 23.04 23.57 0.42 24.30 23.55 0.17
Mother is married 0.45 0.41 0.52 0.52 0.50 0.71
Mother is indigenous 0.09 0.07 0.60 0.05 0.08 0.28
Mother is Afro-Ecuadorian 0.09 0.08 0.72 0.06 0.06 0.88
Mother has no health insurance 0.96 0.96 0.99 0.94 0.91 0.32
Years of schooling (mother) 4.48 4.35 0.61 5.74 5.73 0.89
Years of schooling (partner) 8.36 8.20 0.68 5.17 5.09 0.66
Schooling gap (partner’s schooling − mother’s schooling) 3.88 3.85 0.94 −0.57 −0.64 0.65
Mother has had a child die 0.07 0.12 0.21 0.08 0.10 0.56
Mother uses birth control methods 0.48 0.50 0.74 0.45 0.49 0.39
Mother has poor health 0.11 0.11 0.87 0.11 0.10 0.56
Mother is currently pregnant 0.16 0.11 0.25 0.16 0.14 0.62
Number of children 0–5 years old 2.00 1.85 0.17 1.80 1.89 0.30
Asset index −0.57 −0.50 0.79 −0.57 −0.73 0.45
Kitchen is used for sleeping 0.23 0.24 0.84 0.26 0.26 0.89
Urban 0.59 0.46 0.07 0.38 0.40 0.59
Domestic violence indicators
Physical violence 0.26 0.33 0.49 0.29 0.27 0.69
Emotional violence 0.58 0.60 0.82 0.57 0.53 0.57
Controlling behaviors 0.65 0.60 0.69 0.56 0.56 0.99

p-Values are reported from t-tests of the equality of means for each variable between treatment and control arms for each of the four groups in the study sample. Only a
fraction of mothers in the study at follow-up answered the domestic violence questions at baseline.

administered the domestic violence questions using all baseline which rely on more plausible assumptions. The identifying assump-
control variables from the previous models in addition to the fol- tion for calculating Lee bounds is monotonicity, which implies that
lowing baseline variables: indicators for month of survey, indicator treatment assignment affects sample selection only in one direc-
for mothers who are not household heads or spouses of household tion (Lee, 2009). In our case, this means that treatment cannot
heads, household size, indicator for farming households, indica- simultaneously induce some mothers to stay in the sample and
tor for whether mother answered domestic violence questions at others to leave the sample. Focusing on the main cause of attri-
baseline, indicator for whether mother was in the hospital, and a tion – a partner’s presence – it is unlikely that treatment increases
mother’s CES-D depression score. We then use the inverse of the the probability of a partner’s presence for some and decreases
predicted probabilities as weights in all our specifications and we it for others, especially within the different education groups of
find that estimates from IPW (results not shown) are very similar mothers.
to the unweighted estimates in Tables 4–8, and thus, bias due to The idea behind Lee bounds is to trim the upper and lower tails
selective attrition within our population of interest is minimal. of the distribution of the dependent variable in order to construct
One criticism of IPW is the strong ignorability assumption, and an upper and lower bound of the treatment effect (Lee, 2009).
thus we also calculate non-parametric bounds for our estimates The trimming proportion is the difference in the proportion of
 Author's personal copy

316 M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319

Table 10
Treatment effect estimates and bounds.

Emotional Controlling Physical

Treatment −0.023 −0.064** −0.016

(0.028) (0.030) (0.027)
Lower bound −0.037 −0.071** −0.025
(0.028) (0.030) (0.027)
Upper bound −0.008 −0.057* −0.012
(0.028) (0.030) (0.028)

Standard errors are clustered at the parish level. Treatment effect estimations control for baseline mother characteristics (age, indicator for whether mother is married,
indicators for whether mother is indigenous or Afro-Ecuadorian, years of schooling, partner’s years of schooling, indicator for whether mother has had a child die, indicator
for whether mother is currently pregnant), and baseline household characteristics (asset index, asset index squared, urban, number of children 0–5 years old, whether
kitchen is used for sleeping, province indicators). Bounds are calculated by trimming the top and bottom tails of the distribution of the dependent variable. The trimming
proportion is the difference in the proportion of non-missing outcomes between the treated and control group over the proportion non-missing in the treatment group.
*
p < 0.10.
**
p < 0.05.
***
p < .01.

Table 11
Treatment effect and bounds by mother’s education level.

Mothers with 6 years or less of school Mothers with more than 6 years of school

Emotional Controlling Physical Emotional Controlling Physical

Treatment 0.029 −0.001 0.005 −0.069 *

−0.122 **
−0.039
(0.039) (0.042) (0.039) (0.040) (0.047) (0.037)
Lower bound −0.015 −0.038 −0.050 −0.101** −0.161*** −0.093***
(0.040) (0.042) (0.038) (0.041) (0.047) (0.035)
Upper bound 0.084** 0.045 0.035 −0.041 −0.093** −0.011
(0.039) (0.043) (0.043) (0.041) (0.046) (0.037)

Standard errors are clustered at the parish level. Treatment effect estimations control for baseline mother characteristics (age, indicator for whether mother is married,
indicators for whether mother is indigenous or Afro-Ecuadorian, years of schooling, partner’s years of schooling, indicator for whether mother has had a child die, indicator
for whether mother is currently pregnant), and baseline household characteristics (asset index, asset index squared, urban, number of children 0–5 years old, whether
kitchen is used for sleeping, province indicators). Bounds are calculated by trimming the top and bottom tails of the distribution of the dependent variable. The trimming
proportion is the difference in the proportion of non-missing outcomes between the treated and control group over the proportion non-missing in the treatment group.
*
p < 0.10.
**
p < 0.05.
***
p < .01.

non-missing outcomes between the treated and control group very small. When we split the sample into more and less educated
over the proportion non-missing in the treatment group. Table 10 mothers, the difference in the proportion non-missing is larger and
shows tight negative bounds for the average treatment effect. The thus the bounds widen a bit, but still remain negative for mothers
reason for such tight bounds is that the difference in the proportion with more than 6 years of schooling (Table 11). Treatment effects
of non-missing outcomes between the treated and control group is and bounds are also calculated separately for each of the four

Table 12
Treatment effect and bounds by mother’s absolute and relative education.

Mothers with 6 years or less of school Mothers with more than 6 years of school

Emotional Controlling Physical Emotional Controlling Physical

Mothers with equal to or more schooling than partner

Treatment 0.094* 0.027 0.008 −0.009 −0.107** −0.018
(0.049) (0.048) (0.053) (0.044) (0.050) (0.045)
Lower bound 0.090* 0.021 0.007 −0.028 −0.132** −0.053
(0.050) (0.048) (0.053) (0.045) (0.050) (0.046)
Upper bound 0.099**
0.028 0.009 0.006 −0.089* −0.002
(0.049) (0.048) (0.053) (0.044) (0.050) (0.045)
Mothers with less schooling than partner
Treatment −0.106 −0.049 0.006 −0.390*** −0.189 −0.055
(0.065) (0.091) (0.066) (0.094) (0.114) (0.077)
Lower bound −0.270*** −0.174* −0.189*** −0.508*** −0.295** −0.208***
(0.069) (0.092) (0.068) (0.094) (0.114) (0.073)
Upper bound 0.063 0.081 0.084 −0.350*** −0.127 −0.021
(0.080) (0.089) (0.065) (0.096) (0.115) (0.083)

Standard errors are clustered at the parish level. Treatment effect estimations control for baseline mother characteristics (age, indicator for whether mother is married,
indicators for whether mother is indigenous or Afro-Ecuadorian, years of schooling, partner’s years of schooling, indicator for whether mother has had a child die, indicator
for whether mother is currently pregnant), and baseline household characteristics (asset index, asset index squared, urban, number of children 0–5 years old, whether
kitchen is used for sleeping, province indicators). Bounds are calculated for each of the four group of mothers by trimming the top and bottom tails of the distribution of
the dependent variable. The trimming proportion is the difference in the proportion of non-missing outcomes between the treated and control group over the proportion
non-missing in the treatment group.
*
p < 0.10.
**
p < 0.05.
***
p < .01.
 Author's personal copy

M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319 317

groups of mothers (Table 12). For mothers with more than 6 years consistent with our findings that the BDH led to the largest
of schooling the upper and lower bounds are all negative with the decrease in physical violence for this group of mothers. Given that
exception of the emotional violence indicator for mothers with (a) physical violence is more likely to be under-reported, (b) there
equal to or more schooling than their partners. For mothers with is a high correlation between physical violence and psychological
6 years or less of schooling who have equal to or more schooling violence, and (c) we find decreases in the probability of having
than their partners the upper and lower bounds are positive. For pains and going to the hospital for mothers with more schooling
mothers with 6 years or less of schooling who have less schooling but whose schooling is still less than their partners, our results
than their partners the bounds are very wide and cross over zero, provide suggestive evidence for the direction in which an increase
and thus the direction of impact of the BDH for this group of in a woman’s income would affect physical violence for specific
mothers is ambiguous. The results from bounding our estimates groups of mothers.
reinforce our conclusions that the impact of the BDH is negative Our findings that the BDH leads to reductions in domestic vio-
for mothers with more than 6 years of schooling. For mothers with lence for women with more education but not for women with less
6 years or less of schooling the impact of the BDH depends on a education go against theories that predict that violence will always
mother’s relative education, and in particular, the BDH increases either increase or decrease as a result of increasing a woman’s
domestic violence for mothers in this group who have equal to or income. Instead, our results are consistent with economic house-
more schooling than their partners. hold bargaining models that predict an ambiguous relationship
between a woman’s income and domestic violence. In these mod-
5. Conclusion els violence is an instrument used to align resources more closely
with the aggressor’s preferences or to control a woman’s behavior.
In the existing literature, there is no consensus on the direc- Consequently, increasing a woman’s income leads to the opposing
tion of the associations between a woman’s income and spousal forces of improving her threat point while at the same time creating
domestic violence, and the contradictory evidence is due to contex- situations where partners may exert more violence in order to try
tual factors related to culture, community, timing, and household to control the situation. Our finding suggest that situations where
dynamics. In this paper we explore how household dynamics influ- violence is likely to increase are those where outside-of-marriage
ence these associations, and specifically, we investigate whether options are not a credible threat point and a power imbalance exists
the effects of an increase in a woman’s income on domestic violence among the couple. The latter part is consistent with psychological
depend not only on her education level, but also on her education theories of power, control, and status inconsistency.
relative to her partner’s education. We take advantage of the ran- In order to better understand our findings we explore whether
domized roll-out of a cash transfer program to mothers and find the BDH also affected other dimensions of relationships or labor. In
that for mothers with more than 6 years of schooling, the BDH particular we explore whether the BDH increased the probability
significantly decreases the probability that a husband or partner that mothers left their partners and we find no significant impact
engages in emotional violence by 8 percentage points and control- for any group of mothers. This is not surprising since the BDH led to
ling behaviors by 14 percentage points. For mothers with 6 years improvements in the relationship for most mothers, and in cases
or less of schooling, the effect of the BDH on domestic violence where the relationship deteriorated mothers had little schooling
is ambiguous and depends on the difference in levels of formal and thus low outside of marriage options. We also explore whether
schooling between the mother and her partner. In particular, for the BDH increased the probability of marriage for mothers who
households where the mother’s years of schooling is equal to or were living with their partners at baseline but were unmarried and
more than her partner’s, the BDH significantly increases emotional find no significant impact for any group of mothers. Interestingly,
violence by 9 percentage points. we find that for our study sample the BDH also increased the prob-
Even though we find that the BDH decreases psychological vio- ability that a mother worked for pay. Thus the impact on domestic
lence for certain households and increases it for others, we never violence could be due to both the direct increase in income from
observe a significant corresponding decrease or increase in phys- the BDH and the indirect increase in income from working for pay.
ical violence. One likely reason for this lack of significant impact is It is important to emphasize that the results of the impact of a
that physical violence is more likely to be under-reported. Not only cash transfer on domestic violence are specific to the population
are women more likely to be afraid for their safety if they have been studied: poor young mothers who have young children and who
physically abused, but the physical violence indicator came from live with their husbands or partners. While these results cannot be
one aggregate question as opposed to multiple behaviorally specific generalized to Ecuador’s population as a whole, they do provide
questions which are known to be more effective in eliciting disclo- insight for a large, vulnerable segment of the population. It is also
sure of violence (Ellsberg et al., 2001). Although physical violence is important to stress that the results of an increase in a woman’s
likely to be under-reported, other health indicators that are corre- income on domestic violence are specific to an increase in income
lated with physical violence are not. Thus we investigate whether that is due to the creation of a cash transfer program, and thus, the
the BDH impacts the probability of mothers going to the hospital results may be different if the increase in income is work related.
in the last year, having severe physical pain in the last month, and To our knowledge, this is the first paper in a low- or middle-
having joint pains in the last month. We find that for mothers with income country to investigate the effects of an unconditional cash
more than 6 years of schooling who have less schooling than their transfer to mothers on domestic violence, and adds to the existing
partners, the BDH significantly decreases the probability of all but small literature on the effects of conditional cash transfers on
three health indicators (results available upon request).14 This is domestic violence. Given that the cash transfer is not tied to health
and education requirements, we are able to isolate the income
effect of the cash transfer program. In our analysis we show that
although an increase in a woman’s income leads to a decrease
14
The survey also asks whether a mother went to the health center in the last 30 in domestic violence for many households, there are vulnerable
days and the main reason for going to the health center (with accident and injury households where domestic violence actually increases as a con-
being an option). However, only 3 mothers responded that accident or injury was
the main reason they went to the health center. Thus we do not include this variable
sequence of the income. In particular, these vulnerable households
in our additional analysis. Unfortunately, for the question on going to the hospital in are those where a woman has little schooling, but her schooling is
the last year, the survey does not ask about the main reason for going to the hospital. still greater than or equal to her partner’s. In light of these results,
 Author's personal copy

318 M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319

policy makers are challenged to identify these vulnerable women Appendix A.

and take measures to reduce their risk. Although the primary goal
of the Bono de Desarrollo Humano was not to decrease domestic See Tables A1–A3.
violence, we show yet another dimension in which a cash transfer
can greatly affect the everyday lives of poor women.

Table A1
Psychological and physical violence questions (English translation).

Question Answer Category

Q64. When two people get married or live together they share Many times Sometimes Never NA
good and bad moments. I am going to read to you some
actions or attitudes of husbands or partners. Tell me if the
situation occurs in your relationship many times,
sometimes, or never.
Your husband spends his free time with you Emotional support
He is affectionate with you Emotional support
He ignores you or he is indifferent with you Controlling behavior
He does not allow you to study or work Controlling behavior
He does not allow you to see your friends or family Controlling behavior
He yells at you Emotional violence
He says things like “You are worthless” or similar phrases Emotional violence
He tells you: “I am tired of you, I am thinking of leaving you” Emotional violence
He tells you: “If you leave me I will take our children” Emotional violence
Q66. Has your husband or partner ever pushed Yes No Not sure NA Physical violence
you, hit you or attacked you physically?
Q67. Does your husband or partner physically Frequently Sometimes Not sure NA Physical violence
attack you frequently or only sometimes?
NA indicates no answer.

Table A2
Ordered probit and OLS estimates of BDH on psychological violence variables.

Ordered probit OLS

Does not allow you to Does not allow you Ignores you Yells at you Tells you your Threatens Threatens to Psych.
see friends or family to study or work worthless to leave take children index

Treatment −0.18*** −0.15** −0.14* −0.14* −0.11 −0.07 −0.09 −0.09**

(0.07) (0.08) (0.08) (0.08) (0.08) (0.08) (0.08) (0.04)

Observations 1245 1229 1238 1246 1247 1245 1240 1209

In first 7 columns, outcome variables equal 0 if mother replies never, 1 if mother replies sometimes, and 2 if mother replies frequently. Psychological index is a z-score
constructed by averaging the z-scores from each of the 7 psychological outcome variables. Standard errors clustered at the parish level. All estimations control for baseline
mother characteristics (age, indicator for whether mother is married, indicators for whether mother is indigenous or Afro-Ecuadorian, mother’s years of schooling, partner’s
years of schooling, indicator for whether mother has had a child die, indicator for whether mother is currently pregnant), and baseline household characteristics (asset index,
asset index squared, urban, number of children 0–5 years old, whether kitchen is used for sleeping, province indicators).
*
p < 0.10.
**
p < 0.05.
***
p < .01.

Table A3
Ordered probit and OLS estimates of heterogeneous effects of the BDH with respect to mothers education on psychological violence variables.

Ordered probit OLS

Does not allow you to Does not allow you Ignores Yells at Tells you your Threatens Threatens to Psych.
see friends or family to study or work you you worthless to leave take children index

Treatment −0.13 0.12 −0.02 0.01 −0.02 0.10 −0.03 0.01

(0.10) (0.12) (0.11) (0.12) (0.11) (0.11) (0.13) (0.06)
More than 6 years of schooling −0.16 −0.53*** −0.31* −0.33** −0.19 −0.35** −0.16 −0.22**
(mother) × treatment (0.16) (0.17) (0.17) (0.15) (0.17) (0.16) (0.17) (0.09)
More than 6 years of schooling 0.06 0.24 0.14 0.14 0.12 0.33*** 0.09 0.11
(mother) (0.13) (0.15) (0.13) (0.12) (0.14) (0.13) (0.13) (0.07)

Observations 1245 1229 1238 1246 1247 1245 1240 1209

Treatment effect for mothers with −0.29*** −0.41*** −0.33*** −0.32*** −0.20* −0.25** −0.18* −0.21***
more than 6 years of schooling (0.11) (0.11) (0.12) (0.09) (0.12) (0.11) (0.11) (0.06)

In first 7 columns, outcome variables equal 0 if mother replies never, 1 if mother replies sometimes, and 2 if mother replies frequently. Psychological index is a z-score
constructed by averaging the z-scores from each of the 7 psychological outcome variables. Standard errors clustered at the parish level. All estimations control for baseline
mother characteristics (age, indicator for whether mother is married, indicators for whether mother is indigenous or Afro-Ecuadorian, partner’s years of schooling, indicator
for whether mother has had a child die, indicator for whether mother is currently pregnant), and baseline household characteristics (asset index, asset index squared, urban,
number of children 0–5 years old, whether kitchen is used for sleeping, province indicators).
*
p < 0.10.
**
p < 0.05.
***
p < .01.
 Author's personal copy

M. Hidrobo, L. Fernald / Journal of Health Economics 32 (2013) 304–319 319

References Heise, L., 2012. Determinants of Partner Violence in Low and Middle-Income
Countries: Exploring Variation in Individual and Population-Level Risk. Depart-
Ackerson, L.K., Kawachi, I., et al., 2008. Effects of individual and proximate educa- ment of Global Health and Development/London School of Hygiene and Tropical
tional context on intimate partner violence: a population-based study of women Medicine, London.
in India. American Journal of Public Health 98 (3), 507. Hindin, M.J., Kishor, S., Ansara, D.L., 2008. Intimate Partner Violence among Couples
Ackerson, L.K., Subramanian, S., 2008. Domestic violence and chronic malnutrition in 10 DHS Countries: Predictors and Health Outcomes. DHS Analytical Studies No.
among women and children in India. American Journal of Epidemiology 167 18. Calverton, Maryland. Macro International Inc., USA, pp. 1-97.
(10), 1188. Hornung, C.A., McCullough, B.C., et al., 1981. Status relationships in marriage: risk
Adato, M., Coady, D., et al., 2000a. An Operations Evaluation of PROGRESA from factors in spouse abuse. Journal of Marriage and the Family 43 (3), 675–692.
the Perspective of Beneficiaries, Promotoras, School Directors, and Health Staff. Hughes, J., 2004. Gender, Equity, and Indigenous Women’s Health in the Americas.
International Food Policy Research Institute, Washington, DC. PAHO, Washington, DC.
Adato, M., de la Brière, B., et al., 2000. The impact of PROGRESA on women’s status Jewkes, R., Nduna, M., et al., 2008. Impact of stepping stones on incidence of HIV and
and intrahousehold relations. Final Report. Food Consumption and Nutrition HSV-2 and sexual behaviour in rural South Africa: cluster randomised controlled
Division, International Food Policy Research Institute, Washington, DC. trial. BMJ: British Medical Journal 337, a506.
Aizer, A., 2011. Poverty, violence, and health: the impact of domestic violence during Kling, J.R., Liebman, J.B., et al., 2007. Experimental analysis of neighborhood effects.
pregnancy on newborn health. Journal of Human Resources 46 (3), 518–538. Econometrica 75 (1), 83–119.
Bloch, F., Rao, V., 2002. Terror as a bargaining instrument: a case study of dowry Koenen, K.C., Moffitt, T.E., et al., 2003. Domestic violence is associated with environ-
violence in rural India. The American Economic Review 92 (4), 1029–1043. mental suppression of IQ in young children. Development and Psychopathology
Bobonis, G.J., Castro, R., 2010. The role of conditional cash transfers in reducing 15 (02), 297–311.
spousal abuse in Mexico: short term vs long term effects, http://homes.chass. Lee, D.S., 2009. Training, wages, and sample selection: estimating sharp bounds on
utoronto.ca/∼bobonis/BC dviolence2 mar10.pdf treatment effects. Review of Economic Studies 76 (3), 1071–1102.
Bobonis, G.J., Castro, R., et al., 2009. Public Transfers and Domestic Violence: The Lundberg, S., Pollak, R.A., 1993. Separate spheres bargaining and the marriage mar-
Roles of Private Information and Spousal Control. Working paper 362. Depart- ket. Journal of Political Economy 101 (6), 988–1010.
ment of Economics, University of Toronto. Lundberg, S., Pollak, R.A., 1996. Bargaining and distribution in marriage. The Journal
Carrell, S.E., Hoekstra, M.L., 2010. Externalities in the classroom: how children of Economic Perspectives 10 (4), 139–158.
exposed to domestic violence affect everyone’s kids. American Economic Jour- Macmillan, R., Gartner, R., 1999. When she brings home the bacon: labor-force par-
nal: Applied Economics 2 (1), 211–228. ticipation and the risk of spousal violence against women. Journal of Marriage
CDC, 2003. Costs of Intimate Partner Violence Against Women in the United States. and the Family 61 (4), 947–958.
Centers for Disease Control and Prevention: National Center for Injury Preven- Manser, M., Brown, M., 1980. Marriage and household decision-making: a bargain-
tion and Control, Atlanta, GA. ing analysis. International Economic Review 21 (1), 31–44.
CEPAR, 2005. Encusta demografica y de salud materna e infantil: ENDEMAIN 2004, McElroy, M., Horney, M., 1981. Nash-bargained household decisions: toward a gen-
Informe final. Centro de estudios de poblacion y desarrollo social, Quito, Ecuador. eralization of the theory of demand. International Economic Review 22 (2),
Coker, A.L., Davis, K.E., et al., 2002. Physical and mental health effects of intimate 333–349.
partner violence for men and women. American Journal of Preventive Medicine Paxson, C., Schady, N., 2010. Does money matter? The effects of cash transfers
23 (4), 260–268. on child development in rural Ecuador. Economic Development and Cultural
Duflo, E., Glennerster, R., et al., 2007. Using randomization in development eco- Change 59 (1), 187–229.
nomics research: a toolkit. Handbook of Development Economics 4, 3895–3962. Pence, E., Paymar, M., 1993. Education Groups for Men Who Batter: The Duluth
Ellsberg, M., Heise, L., 2005. Researching Violence against Women: A Practical Guide Model. Springer Publishing Company, New York, NY.
for Researchers and Activists. World Health Organization, PATH, Washington, Perova, E., 2010. Buying out of abuse: how changes in women’s income affect domes-
DC. tic violence.
Ellsberg, M., Heise, L., et al., 2001. Researching domestic violence against women: Pollak, R., 2005. Bargaining Power in Marriage: Earnings, Wage Rates and Household
methodological and ethical considerations. Studies in Family Planning 32 (1), Production. National Bureau of Economic Research, Cambridge, MA, USA.
1–16. Pronyk, P.M., Hargreaves, J.R., et al., 2006. Effect of a structural intervention for the
Ellsberg, M., Jansen, H.A.F.M., et al., 2008. Intimate partner violence and women’s prevention of intimate-partner violence and HIV in rural South Africa: a cluster
physical and mental health in the WHO multi-country study on women’s randomised trial. The Lancet 368 (9551), 1973–1983.
health and domestic violence: an observational study. The Lancet 371 (9619), Ribero, R., Sánchez, F., 2005. Determinants, effects and costs of domestic violence.
1165–1172. Documento, CEDE 38 http://www.google.com/url?sa=t&rct=j&q=&esrc=s&
Eswaran, M., Malhotra, N., 2011. Domestic violence and women’s autonomy source=web&cd=1&ved=0CDUQFjAA&url=http%3A%2F%2Feconomia.uniandes.
in developing countries: theory and evidence. Canadian Journal of Eco- edu.co%2Fcontent%2Fdownload%2F2087%2F12358%2Ffile%2Fd2005-38.pdf&ei
nomics/Revue canadienne d’économique 44 (4), 1222–1263. = yq0UOm4COjV0QGCw4DIDw&usg=AFQjCNFZNUZ ROAxCsKjz5Rn-
Farmer, A., Tiefenthaler, J., 1997. An economic analysis of domestic violence. Review ckBhsVmSw
of Social Economy 55, 337–358. Schady, N., Araujo, M.C., 2008. Cash transfers, conditions, and school enrollment in
Fernald, L.C.H., Hidrobo, M., 2011. Effect of Ecuador’s cash transfer program (Bono Ecuador. Economía 8 (2), 43–70.
de Desarrollo Humano) on child development in infants and toddlers: a ran- Sternberg, K.J., Lamb, M.E., et al., 1993. Effects of domestic violence on children’s
domized effectiveness trial. Social Science & Medicine 72 (9), 1437–1446. behavior problems and depression. Developmental Psychology 29 (1), 44.
Figueroa, M., Miranda, G., 2009. Las rentas laborales a partir de la educación y la Stevenson, B., Wolfers, J., 2006. Bargaining in the shadow of the law: divorce laws
experiencia laboral en el Ecuador en el periodo 2003–2008, utilizando el modelo and family distress. The Quarterly Journal of Economics MIT Press 121 (1),
de Mincer y las encuestas de empleo y desempleo urbano del INEC. Facultad de 267–288.
Ciencias/Escuela Politecnica Nacional, Quito. Sullivan, C.M., Bybee, D.I., 1999. Reducing violence using community-based advo-
Fiszbein, A., Schady, N., et al., 2009. Conditional Cash Transfers: Reducing Present cacy for women with abusive partners. Journal of Consulting and Clinical
and Future Poverty. The World Bank, Washington, DC. Psychology 67 (1), 43.
Flake, D.F., 2005. Individual, family, and community risk markers for domestic vio- Tauchen, H.V., Witte, A.D., et al., 1991. Domestic violence: a nonrandom affair. Inter-
lence in Peru. Violence against Women 11 (3), 353. national Economic Review 32 (2), 491–511.
García-Moreno, C., Jansen, H.A.F.M., et al., 2005. WHO Multi-Country Study on Vyas, S., Watts, C., 2009. How does economic empowerment affect women’s risk
Women’s Health and Domestic Violence against Women: Initial Results on of intimate partner violence in low and middle income countries? A system-
Prevalence, Health Outcomes and Women’s Responses. World Health Organi- atic review of published evidence. Journal of International Development 21 (5),
zation, Geneva, Switzerland. 577–602.
García-Moreno, C., Jansen, H.A.F.M., et al., 2006. Prevalence of intimate partner vio- Wolfe, D.A., Crooks, C.V., et al., 2003. The effects of children’s exposure to domestic
lence: findings from the WHO multi-country study on women’s health and violence: a meta-analysis and critique. Clinical Child and Family Psychology
domestic violence. The Lancet 368 (9543), 1260–1269. Review 6 (3), 171–187.
Hall, G., Patrinos, H.A., 2006. Indigenous Peoples, Poverty and Human Development Wooldridge, J.M., 2002. Econometric Analysis of Cross Section and Panel Data. The
in Latin America: 1994–2004. Palgrave Macmillan, New York. MIT Press, Cambridge, MA.
Heath, R., 2012. Women’s access to labor market opportunities, control of household
resources, and domestic violence. World Bank Policy Research Working Paper
(6149).