congogbv.Rmd

---
title: "Sexual violence, conflict, and female empowerment: Exploratory evidence from a list experiment in Eastern DR Congo"
author: Koen Leuveld
output:
  bookdown::pdf_document2:
    toc: false
    latex_engine: xelatex
bibliography: congogbv.bib
date: "`r Sys.Date()`"
papersize: a4
geometry: "left=2cm,right=2cm,top=2cm,bottom=2cm"
header-includes:
- |
  ```{=latex}
  \usepackage{authblk}
  \author{Koen Leuveld\\ koen.leuveld@wur.nl}
  \affil{Wageningen Economic Research}
  ```

#more info:
#https://www.rstudio.com/wp-content/uploads/2015/03/rmarkdown-reference.pdf
---



```{r setup, eval=TRUE, include = FALSE, echo=FALSE, warning=FALSE, error=FALSE, message=FALSE}

library(ggplot2)
library(dplyr)
library(rlang)
library(readr)
library(stringr)
library(flextable)
library(officer)
library(here)




# get varlables
varlabels <-
  read_csv(here("tables/cleandata_collabels.csv")) %>%
  pull(label) %>%
  set_names(read_csv(here("tables/cleandata_collabels.csv")) %>% pull(Variable))


#standard chunk options
knitr::opts_chunk$set(ft.latex.float = "float",eval=TRUE, include = TRUE, echo=FALSE, warning=TRUE, error=TRUE, message=FALSE)
set_flextable_defaults(font.size = 9, fonts_ignore=TRUE)


#run analysis
#rmarkdown::render(here("congogbv_analysis.Rmd"))

```

<!-- this is the default float positioning in latex -->

```{=latex}

\makeatletter
  \providecommand*\setfloatlocations[2]{\@namedef{fps@#1}{#2}}
\makeatother
\setfloatlocations{figure}{ph}
\setfloatlocations{table}{ph}

```

# Introduction {#introduction .unnumbered}

Over the past decades, tremendous progress has been made worldwide to
improve the lives of the world's poor. The proportion of people living
under the poverty line of \$1.25 per day dropped from over half to just
14%; gender disparity in primary education has been drastically reduced
or even eliminated; under-five mortality rates have been halved
[@UnitedNations2015]. However, such progress has largely bypassed
fragile states, like the Democratic Republic of the Congo (DRC)
[@Asadullah2018; @Samy2011]. Congolese women in particular face economic
hardships and human rights violations, including a high rate of Sexual
and Gender Based Violence (SGBV): estimates of the proportion of women
who have suffered from this range from 15% to 40%
[@Johnson2010; @Peterman2011]. Aid workers have called the country "the
world's worst place to be a woman or a child", and the UN's Special
Representative on Sexual Violence in Conflict, Margot Wallström, even
called the country the "rape capital of the world"
[@HumanRightsWatch2009]. The issue of SGBV is of specific concern, given
its high psychological, social and economic costs
[@Post2002; @Peterson2018]. Consequently, tremendous international
efforts have been made to implement or support projects to assist the
victims of SGBV. The 2018 Nobel Peace prize was awarded to Dr. Denis
Mukwege, for his work on victims of SGBV at Bukavu's Panzi Hospital.

Despite this attention, very little reliable data exists on the topic
[@Palermo2011]. Data collection efforts have been hampered by the
conflicts the country has faced, which have made large-scale data
collection from representative samples difficult. As a result, most data
available on the topic is from surveys within clinics and NGOs aimed at
assisting victims of SGBV, making comparison between victims of SGBV and
non-victims difficult. These samples are obviously skewed, as they only
survey women who have already come forward in search for help. Even when
survey data is available, the sensitive nature of the topic may cause
respondents to withhold information due their unease in discussing
sensitive topics with survey field staff.

In this paper, I explore the characteristics of victims of SGBV as well
as non-victims to study the dynamics and potential drivers of SGBV. I
classify female survey respondents according to potential risk factors
for SGBV, and analyse whether these factors are in fact associated with
the incidence of SGBV. In this way, this paper aims to address the
question what the drivers of SGBV in Eastern Congo are. Specifically, I
consider conflict and the position of women in Congolese society as
potential drivers, and compare their relative contribution.

The conflict that has persisted in the country for the past decades is
the most often-cited driver of the high rate of SGBV. This is
particularly true for policy circles, where the framing of SGBV in Congo
as "weapon of war" is popular [@Baaz2013; @Kirby2015]. There is
empirical evidence to support this notion. [@Johnson2010] carried out a
large-scale survey in Eastern Congo to investigate incidence and
perpetrators of SGBV, and found that the majority of sexual violence
reported by their respondents was conflict-related; of female victims of
sexual violence, 74.3% reported the perpetrators to be conflict-related.
Likewise, @Bartels2013 find that the majority of the victims of SGBV
treated at Panzi Hospital -- in Bukavu, South Kivu -- indicate that the
perpetrators were armed groups. It is therefore not surprising that the
topic of SGBV in Congo has often been analysed within the context of
violent conflict [@Baaz2013]. Conversely, the aspect of the conflict
that has received the most world-wide media attention has been SGBV
[@Autesserre2012a].

However, this view of the central role of conflict in sexual violence in
the DRC has come under increasing scrutiny. It has been argued that this
focus on the relationship between sexual violence and conflict has been
counter-productive, as it has distracted attention from other pressing
problems the DRC faces [@Autesserre2012a; @Hilhorst2018; @Porter2019].
Moreover, it risks missing the civilian perpetrators of SGBV. There is
empirical evidence for this position too. Based on DHS data,
[@Peterman2011] find that rates of Sexual Intimate Partner Violence
(IPV) are higher than rates of other forms of Sexual Violence in Congo.

This increased focus on IPV, rather than conflict, shifts attention from
conflict to the bargaining position of women in Congolese society and
households as a driver of SGBV. The bargaining position of a woman is
the level of autonomy she has, and is determined by things such as her
outside options; its effect on SGBV is ambiguous [@Eswaran2011]. One the
one hand, a woman's welfare may depend on her bargaining position: women
with more income, and better prospects in case of a divorce would thus
face less risk of IPV. On the other hand a woman's partner may use IPV
as an instrument to assert power as a response to her increased
empowerment. The empirical record reflects this ambiguity.
[@Bhattacharya] find that increase in employment, and the increasing of
status of a woman within the household reduces violence. Similarly,
[@Hidrobo2016] find that cash transfers to women, decrease the risk of
violence. However, when specifically looking at sexual IPV in the
Dominican Republic, [@Bueno2017] find that an increase in women's
(economic) empowerment led to an increase in IPV. In Vietnam,
[@Bulte2019] find that a project aiming to increase women's income, may
have led to increased IPV. In Afghanistan, [@Gibbs2020] find no link
(positive or negative) between economic empowerment and IPV. The link
between IPV and women's intra-household bargaining position may be
moderated by local customs, and depend on exactly the type of IPV and
the type of empowerment under consideration.

These two main drivers of SGBV -- conflict and empowerment -- are not
necessarily separate, as conflict may affect empowerment on the long
run. In the short run, conflict may have a direct effect on SGBV through
perpetration by armed groups, potentially in a strategic manner
[@Baaz2013; @Kirby2015]. This caused the topic to be on the
international agenda, as a "weapon of war". However, in the long run
there is a more indirect effect as well: conflict causes the breakdown
of norms, which may have long-lasting effects. For example, [@Kelly2018]
find that that IPV increased in districts that experienced conflict in
Liberia, while [@Muller2019] draw similar conclusions from data from the
Gaza strip. [@Saile2013] investigate the correlates of IPV for a sample
of conflict-exposed women in Northern Uganda. They find that while the
level of conflict exposure predicts physical violence, sexual violence
is more associated with the level of childhood familial violence. This
link between current and past experiences of violence suggests that the
effect of conflict on violence is deeper than just the direct effect.
People traumatized during the conflict (either because they were victims
or perpetrators) are more likely to be victimized later on.

In answering the question what the main drivers of SGBV are, I thus
consider two main drivers: the position of women in Congolese society
and conflict. Within conflict, I distinguish between historic conflict
(here I use pre-2012 data) with long-term, indirect, effects and recent
conflict (up to one year prior to the interview) with short-term, direct
effects. I argue that the indirect long-term effects of conflict are
likely to be related to the position of women, through changing norms,
while more recent conflict events may not have had an additional impact
on norms yet. For empowerment, I use a bargaining game, and survey
questions that determine women's pre-marriage relative status.

This paper contributes to the empirical evidence base on the incidence
of SGBV in Eastern Congo by drawing on a sample of beneficiaries of
development assistance projects in South Kivu province, in the Eastern
DRC. While the selection of respondents was not done to produce a
representative sample for the province, it does not suffer from the same
problems that clinic-based surveys have, allowing me to compare victims
of SGBV with non-victims. Data on SGBV comes from a list experiment, a
technique which has been gaining popularity as a way to obtain
information on sensitive topics [see e.g.
@Sniderman1991; @Holbrook2010; @Bulte2019; @Peterman2018; @LaBrie2000; @Corstange2009].
Put briefly, list experiments allow group-level analysis of SGBV
victimization, without individuals revealing their own victimization
status. This eliminates the need for respondents to withhold information
and thus reduces the social desirability bias that results [@Blair2012].
Such bias may explain the fact that studies on the drivers of SGBV often
contradict each other. While one study finds conflict-related
perpetrators are responsible for the majority of cases of SGBV
[@Johnson2010], another finds intimate partners as the most common
culprits [@Peterman2011]. @Stark2017 provides an example of how
different methodologies can provide different answers: when using Audio
Assisted Self-Interviews (ACASI) they find that intimate partners are
the main perpetrators of SBV. However, in complementary group
discussions, where social desirability bias is likely to be present,
respondents did not bring up intimate partners at all.

I combine list experiments with detailed survey data on the household
and outcomes from behavioural experiments, which allows for a rich
characterization of victims of SGBV. Because such a characterization is
lacking thus far, this data is useful in addressing and preventing SGBV.
Moreover, while the potential drivers of SGBV mentioned above --
conflict and empowerment -- have been studied in isolation, this paper
contributes by analysing these in one framework.

I find high victimization rates in my sample: 30% of the women report
SGBV in the past twelve months. These victims are likely to be married
to higher-status men, have low intra-household bargaining power, and
have been exposed to violent conflict to the extent where they have lost
family or household members before 2012 (two years before the list
experiment). I find no evidence of a link between SGBV and recent
conflict exposure. These findings are consistent with recent findings in
the literature that conflict has long-lasting impact on SGBV through
IPV. This paper is structured as follows: first I describe the research
setting, the sample, and then the various sources of data. The
subsequent section describes my empirical framework, which revolves
around the use of a list experiment. I then present the results of the
analyses. In the concluding section I contextualize the findings and
present policy implications.

# Background {#background .unnumbered}

Congo's 2006 constitution grants equal rights to men and women. In
practice, however, women hold an inferior position in Congolese society.
This is reflected in social and economic outcomes. The literacy rates
among women and girls aged 15-24 is 73.6% (compared to 91.2% among men
and boys of the same age); only 8.5% of women have completed secondary
education (compared to 16.2% of the men); while 67% of women work, only
7.8% work outside of agriculture or trading and services
[@DHSCongoReport]. Within the household, women occupy an inferior
position as well: the husband is the head by law, and marital rape is
not considered a crime [@Kilonzo2009].

In addition to the difficulties inherent to their inferior position,
women have faced widespread human rights abuses during the conflicts
that have swept the country since the mid 1990s. South Kivu (the setting
for the present study) has been greatly affected by these conflicts. The
first Congo war started with an invasion by Rwanda and armed groups
supported by Rwanda to clear perpetrators of the Rwanda genocide from
the refugee camps in the east of the country, putting the province on
the front line. Throughout this First Congo War (1996-1997), the Second
Congo War (1998-2003) and the subsequent fragile peace, ethnic tensions
have remained high throughout the province, resulting in frequent
localized bursts of violence [see e.g. @Verwijen2016]. While some of the
human rights abuses during these phases of the conflict occurred during
large-scale attacks on civilians, often they occurred during ambushes
while women were conducting their day-to-day tasks
[@Freedman2011; @HRW2002]. Women were often assaulted by multiple
perpetrators. These were not only members of rebel groups, but also the
government army [@HumanRightsWatch2009].

The consequences of (conflict-related) SGBV for the victims have been
well-researched. It has severe mental and physical health consequences
[@Johnson2010]. However, due to the remote nature and lack of resources,
victims have difficulty finding professional help, often having to
travel more than a day to clinics
[@HarvardHumanitarianInitiative2009; @Kohli2012]. The negative
consequences persist until long after the event, as victims face
stigmatization within their communities and households
[@Albutt2017; @HarvardHumanitarianInitiative2009].

The adverse consequences of conflict-related SGBV do not remain limited
to the direct victims. The violence against women during the conflict
resulted in a change in norms, where armed groups were no longer the
main perpetrators of SGBV, but civilians (including intimate partners)
[@Freedman2011]. Risk factors for sexual IPV include partner problematic
use of alcohol and partner controlling behaviours [@Babalola2014]. While
the Congolese government has made attempts to address the situation,
such as through the Law on the Suppression of Sexual Violence,
implementation of these measures has been marred by the general lack of
resources state authority in the country [@Steiner2009].

# Sample {#sample .unnumbered}

```{r table1dhs, tab.cap='Comparison of DHS and sample data', tab.id='dhscompare'}

read_csv(here("tables/table1.csv")) %>% 
  flextable() %>%

  #header row
  add_header_row(top = TRUE, values = c("","Age of FR","Tin Roof","Education FR"),colwidths = c(1,1,1,2)) %>%
  mk_par(i = 2, j = 1:3, as_paragraph(""),part="header") %>%
  mk_par(i = 2, j = 4, as_paragraph("Primary"),part="header") %>%
  mk_par(i = 2, j = 5, as_paragraph("Secondary"),part="header") %>%

  #layout
  theme_booktabs() %>%
  align(i = 1,j = 4,align = "center", part = "header") %>%
  autofit() %>%
  colformat_double(digits = 2) %>%

  footnote(i = 1,j = 4, value = as_paragraph("FR = Female respondent. For the DHS data, means are provided for household members satisfying the same criteria as FRs from the sample: female heads of household, or female spouses of household heads"), ref_symbols = "a", part = "header", inline = TRUE, sep = ";")   %>%
  footnote(i = c(1,2),j = 1, value = as_paragraph("DHS = Demographic Health Survey (MPSMRM, 2014)"), ref_symbols = "b", part = "body", inline = TRUE, sep = ".")

```
The main source of data for this study is the gender module from a
household survey that was undertaken in 2014 as the endline survey for
the evaluation of Dutch development aid. This evaluation concerned
projects ran by four NGOs in the territories of Kabare, Fizi and Uvira,
and the commune of Bagira.^[In the remainder of the paper, I will consider Kabare and Bagira
    to be one "territory", since the selected communities in Kabare and
    Bagira are located close together, in the peri-urban zone of Bukavu.] The baseline for this evaluation was done
in 2012. Half of the respondents were selected from communities that
benefited from the projects, the other half were selected from
comparable households in non-intervention communities. These projects
were about agriculture, women's rights and education. Overall, the
beneficiaries of the projects were vulnerable, mostly rural, households.
An indicator for being beneficiary to any of these projects is included
in the full analysis below. In total data was collected in 73
communities. In each community, baseline data was collected on 15
households in 2012; however, due to attrition, 2014 data is available
for an average of 12 households per community, for a total of 889
households.

The sampling procedure outlined here is thus unlikely to have produced a
nationally (or even provincially) representative sample. In Table
\@ref(tab:dhscompare),
I present a comparison across selected demographics between the full
study sample (column 3), and the representative sample from the DHS
Program (columns 1-2). Women in the study sample are older, and less
likely to have finished school, than the provincial average in South
Kivu.

```{r table2sample, tab.cap='Gender module sample make up', tab.id='bargsample'}

 read_csv(here("tables/table2.csv")) %>%
  flextable() %>%

  #header
  add_header_row(top = TRUE, values = c("","Male Repondent"),colwidths = c(1,5)) %>%
  mk_par(i = 2, j = 1, as_paragraph("Female Respondent"),part="header") %>%
  mk_par(i = 2, j = 2, as_paragraph("Consented"),part="header") %>%
  mk_par(i = 2, j = 3, as_paragraph("Refused"),part="header") %>%
  mk_par(i = 2, j = 4, as_paragraph("Absent"),part="header") %>%
  mk_par(i = 2, j = 5, as_paragraph("No husband"),part="header") %>%

  #body
  mk_par(i = 1, j = 1, as_paragraph("Consented"),part="body") %>%
  mk_par(i = 2, j = 1, as_paragraph("Refused"),part="body") %>%
  mk_par(i = 3, j = 1, as_paragraph("Absent"),part="body") %>%
  mk_par(i = 4, j = 1, as_paragraph("No wife"),part="body") %>%

  theme_booktabs() %>%
  align(i = 1,j = 2,align = "center", part = "header") %>%
  autofit()

```

Not all households participated fully in the gender module. Where
possible, it was administered to both the head of the household and
their spouse, so that there were a Female and a Male Respondent to the
interview. In the vast majority of the cases, the husband is
considered the head, but it was left open to the respondents to indicate
the head. Table \@ref(tab:bargsample) displays how the sample is built up. In
total, there were 889 respondents to the survey. In 593 households, the
Female Respondent (the wife of the household head or the female head)
consented to responding to the gender module. In 1 household, the female
respondent refused; in 5, the Female Respondent was absent during the
interview, and in 290 households the head of the household had no wife,
and there was thus no Female Respondent. In 470 households, the Male
Respondent (usually the household head) consented to the module, 6
refused, 255 Male Respondents were absent, and in 158 households the
head of the household was an unmarried woman, meaning that there was no
Male Respondent. In 184 households, both husband and wife responded to
the module. Efforts to increase this number, by tracking down absent
household heads, were constricted by the limited time field teams had in
each community due to the security situation at the time of field work.
As a consequence of this, sample sizes between various analyses are
different: analyses relying on both partners being present - e.g. for
the bargaining game - will have a lower sample size than others.

The selection of respondents to the gender module is unlikely to have
been random. In column 4 of Table \@ref(tab:dhscompare)
selected demographics for the Female Respondents to the gender module
are presented. The respondents are slightly older than the full sample,
and considerably older than the provincial average. They are slightly
more likely to have completed secondary school than the full sample, but
less likely than the provincial average. In Table \@ref(tab:sampleselection) I present results from logit models
to find correlations between household characteristics and participation
in the gender module. The dependent columns of the columns are whether
the wife, the husband and the couple participated in the gender module,
respectively. There are some selection effects. Households that own tin
roofs, are more likely to have a Female Respondent. In households that
own livestock, it was less likely that there was a female respondent to
the gender module, and more likely to have male respondent. The final
analysis below will include these as controls.

# Methods {#methods .unnumbered}

```{r table3barganing, tab.cap='Bargaining game lotteries', tab.id='risklotteries'}
# 
read_csv(here("tables/table3.csv")) %>%
  flextable() %>%
  mk_par(i = 1, j = 1, as_paragraph(""),part="header") %>%
  theme_booktabs() %>%
  autofit()

```

This paper combines data from the 2014 and 2012 rounds of the survey,
with ACLED data. The gender module from the 2014 survey is the main
source of data for this paper. The module was administered separately to
Male and Female Respondent (with a small part being administered
jointly). It contained (i) a list experiment designed to elicit the
incidence of SGBV among Female Respondents; (ii) a risk bargaining game
to elicit the relative intra-household bargaining position of the Male
and Female Respondent; and (iii) a set of propositions to collect
detailed information on gender attitudes. I present the List Experiment,
and the analysis thereof, in more detail in the Empirical Framework
below.

The risk bargaining game in the gender module was modified from
[@Martinsson2009]. In the game, the respondents chose between a set of
six risky lotteries, based on [@Eckel2002]. The lotteries presented
ranged from fairly low-risk ones -- where low and high pay-out were
nearly equal -- to high-risk one -- where there was a large difference 
between high and low pay-outs (see Table \@ref(tab:risklotteries)
for details of the lotteries). The Male and Female Respondents first
chose privately (without knowing their partner's choice), and then
jointly. By comparing the couple decision with the individual decision,
I obtain an indicator for bargaining power: the closer the couple
decision is to the Female Respondent's decision -- relative to the Male
Respondent's decision -- the higher her bargaining power. The difference
between the procedure used by [@Martinsson2009] and the one here, is
that they use a risk experiment based on [@Holt2002]; a more complicated
experiment compared to @Eckel2002. This added complication may cause
some participants to not fully understand the procedure, leading to poor
results [@Dave2010a]. Given the low numeracy of the subjects, I
implemented the simpler of the two experiments.

I draw on two sources for conflict data: data from the 2012 round of the
survey, and ACLED data from 2013-2014. The 2012 data contains detailed
information of the conflict history of the respondents dating back to
the start of the First Congo War in 1996. Among other things,
respondents were asked whether they lost family members, whether they
lost property, and when these events took place. I use this to construct
indicators for historic victimization, which may have indirect effects
on SGBV victimization. Due to time constraints, the 2014 round of the
survey did not contain a detailed conflict exposure module. In order to
get more detailed information on recent victimization, complement the
household-level data with more recent data from the Armed Conflict
Location & Event Data Project [ACLED; @Raleigh2010]. The 2014 data
contains GPS coordinates for all interviewed households. Using these
coordinates, I can link households to nearby conflict events from the
ACLED database that took place within the 12 months preceding the
interview. Because this data then coincides with the window for SGBV
used here, any direct effects from conflict -- like perpetration of SGBV
by armed groups -- will be captured by this indicator. However, while
this data is more recent, it does not capture individual experience;
only exposure based on the distance from the household to conflict
events.

# Empirical Strategy {#empirical-strategy .unnumbered}

A major concern in collecting data on SGBV is reporting bias.
Respondents are unlikely to be comfortable to truthfully answer
questions about SGBV. Respondents may want to hide undesirable answers,
leading to what's called social desirability bias. Not only may this
lead to an underestimate of the incidence of SGBV, the unwillingness to
divulge information may be correlated to the identity of the
perpetrators: people may be more willing to divulge victimization from
armed groups, than from intimate partners [@Stark2017]. This non-random
nature of non-response would thus lead to an underestimate of the
incidence of SGBV, and biased estimates for the correlates of SGBV when
using direct questions. This is why such direct questions are not used
in list experiments. Instead, interviewers present respondents with a
list of issues and ask them to indicate the number of issues from the
list they have faced. By adding the sensitive item to the list of issues
for half of the respondents (randomly selected), estimates for incidence
of the sensitive item can be obtained by comparing the mean number of
issues faced in both groups (hence "item count technique" as an
alternative name for list experiments). The advantage is thus that
answers are guaranteed to be anonymous: the interviewer (or the data
analyst) does not know the number of non-sensitive issues the respondent
faces and so has no way of knowing the answer to sensitive item. This
anonymity removes the need to hide the answer, and thus the social
desirability bias.

Over the past decades, list experiments have grown in popularity as a
way to obtain accurate data on sensitive topics. [@Holbrook2010] review
48 studies using list experiments, and found that they are effective at
decreasing social desirability bias. Comparing studies that use list
experiments with studies that do not, they find that reporting rates of
sensitive items are higher in studies using list experiments. It is
therefore not surprising that this approach has been applied to a wide
range of topics, such as sensitive political opinions
[@Frye2017; @Blair2014; @Meng2017; @Corstange2009], over-reporting of
voting [@Holbrook2010], risky behaviours [@LaBrie2000] and SGBV
[@Bulte2019; @Peterman2018].


For the list experiment in this study, the female respondents were
randomly divided into two groups. This was done by the electronic survey
software (ODK), based on the randomly assigned ID codes. I follow
[@Imai2011] in calling these groups Treatment and Control. An
interviewer told each respondent: "I will read 4 *(or 5)* problems that
women can experience. These can be sensitive problems. When you've
experienced a problem in the last year, please drop one of the balls to
the ground. I will not look at when you drop these balls, and only want
to know the total number of balls at the end. In the past 12 months, did
you experience\...

- Lack of food;
- Lack of money;
- Theft;
- Sterility; and,
- Sexual Violence (Treatment group only)"

The interviewer only presented women randomly selected to be in the
treatment group with the fifth item (Sexual Violence). I selected the
four control items in such a way that it is unlikely women in the sample
face none, or all, of the issues. In such cases the interviewer knows
the respondent's answer to the sensitive issue (\"no\" if the total
number of issues is 0, \"yes\" if the total number is 5). Not all the
control items are non-sensitive, as the item "sterility" is a sensitive
item. This was done to reduce respondent suspicion when one sensitive
items is juxtaposed with a number of completely non-sensitive items (see
@Chuang2019 for a more detailed explanation). After all items were read,
the interviewer asked the respondent to count the number of balls, and
report the number. The questionnaire was field tested prior to field
work to ensure that respondents understood these concepts. All
interviewers were thoroughly trained in the protocols, and the
electronic questionnaire was programmed in such a way to ensure
compliance to the protocol.

A crucial assumption for the list experiment is that the randomization
ensures that Treatment and Control groups are identical. Table
\@ref(tab:summstats) (Column 7) provides a comparison of the two
groups within the sample. The treatment and control group are not
perfectly balanced across some of the variables. However, an F-test on
the differences between treatment and control being jointly equal to
zero fails to reject the null-hypothesis that they are equal (p=0.20).
This suggests that the differences found are due to chance, rather than
any bias in the randomization procedure.

While the indirect nature of list experiments prevents reporting bias,
this does come at a cost of efficiency in statistical analysis. The
incidence is easily computed by subtracting the mean of issues faced in
the control group from the mean number of issues in the treatment group.
This means that sample sizes have to be far larger for list experiments
than for direct questions.

In a regression framework, the incidence would be estimated as follows
[@Holbrook2010]:


\begin{equation} 
NumIssues_i = \beta_0 + \beta_1 Treatment_i + \epsilon_i
(\#eq:basic)
\end{equation}

Where $NumIssues_i$ is the number of issues experienced by respondent
$i$, and $Treatment_i$ is her treatment assignment. Coefficient
$\beta_1$ yields the estimate for the incidence. To find correlates of
SGBV, equation \@ref(eq:basic) can be augmented using interaction terms as
follows:

\begin{equation} 
NumIssues_i = \beta_0 + \beta_1 Treatment_i + \beta_2 X_i + \beta_3 Treatment_i X_i + \epsilon_i
(\#eq:interaction)
\end{equation}

Where $X_i$ is an explanatory variable and coefficient $\beta_3$ gives
the estimate for the additional incidence of SGBV associated with a unit
increase of $X$. This can be easily modified to allow for more
variables. Again, this is much less efficient than when using direct
questioning. By using more sophisticated methods proposed by @Imai2011
(and implemented by [@Tsai2019] in Stata), more efficient estimates can
be obtained.


```{r table4summstats, tab.cap='Descriptive statistics by treatment assignment', tab.id='summstats'}

# flextable orders the variables alphabetically, but I want to preserve the order
# because the variabe order is the same in summstats and difcol, I can just use sequential numbers
# making sure to do the same with the labels 

numvars <- read_csv(here("tables/summstats_diffs.csv")) %>% nrow()
numtreatment <- 3

summstats <- 
  read_csv(here("tables/summstats.csv")) %>%
  left_join(read_csv(here("tables/cleandata_collabels.csv"))) %>%
  mutate(Treatment = ifelse(Treatment != "Overall",paste0(" ",Treatment),Treatment)) %>%
  mutate(var = Variable,
         Variable = str_pad(rep(1:numvars,numtreatment),2,pad="0")) 

difcol <- 
  read_csv(here("tables/summstats_diffs.csv")) %>%
  mutate(Variable = str_pad(1:numvars,2,pad="0"))

labels <- summstats %>% pull(label) %>% unique()
names(labels) <- str_pad(1:numvars,2,pad="0")

summstats %>%
  tabulator(rows = "Variable",
            columns = "Treatment",
            datasup_last = difcol,
            `N` = as_paragraph(as_chunk(n,digits=0)),
            `Mean` = as_paragraph(as_chunk(mean,digits=2)),
            `SD` = as_paragraph(as_chunk(sd,digits=2)))%>%
            #`Mean \n (SD)` =  as_paragraph(as_chunk(mean,digits=2,),as_chunk("\n("),as_chunk(sd,digits=2),")")) %>%
  as_flextable() %>%
  labelizor(j = "Variable", labels = labels, part = "all") %>% 

  #fit to page
  padding(padding = 0.5,part = "all") %>%
  padding(padding.top = 0, padding.bottom = 0) %>%
  fontsize(size = 7, part = "all") %>%
  autofit() %>%

  #add footer
  add_footer_lines(values = "*p < 0.1,**p < 0.05,***p < 0.01", top = FALSE) %>%
  fontsize(size = 7, part = "footer") 


```


# Results {#results .unnumbered}


```{r overallmeans, fig.cap =  "Comparison of means of issues faced: treatment vs. control", out.width = "50%", fig.align = 'center'}

read_csv(here("tables/dataforgraphs1.csv")) %>%
  ggplot(aes(Treatment, numballs, color = Treatment)) +
  geom_pointrange(aes(ymin = ci.lower, ymax = ci.higher)) +
  ylab("Number of reported issues") +
  xlab("")

```

In this section, I will first compare the results of the list experiment
in the whole sample, then in different sub-groups. I then present
results from a full multivariate regression that aims to minimize
potential bias caused by confounding variables.

In the full sample, the difference between the group who were presented
with only four issues (the Control group) and the group who were
presented four issues plus SGBV (the Treatment group) is the estimate of
the incidence of SGBV. The average number of issues reported by the
control group is 2.34, while the number of issues reported by the treatment
group is 2.65 (see Figure \@ref(fig:overallmeans)). The difference of 0.31 issues implies that
the incidence of SGBV is 31% in this sample. This estimate appears substantially higher than
previous estimates. These previous estimates [e.g.
@Peterson2018; @Stark2017; @Johnson2010] arrive at a similar rate of
victimization, but over the life of the respondent, whereas here we only
consider victimization the past twelve months. A higher incidence is
expected, since the sample is non random, drawing mostly from vulnerable
rural households.

## Conflict {#conflict .unnumbered}

```{r confmeans, fig.cap =  "Comparison of means of issues faced across conflict exposure", out.width = "100%", fig.align = 'center'}

read_csv(here("tables/dataforgraphs.csv")) %>%
left_join(read_csv(here("tables/cleandata_collabels.csv")),by = join_by(var == Variable)) %>%
filter(var %in% c("victimproplost", "victimfamlost", "acledviolence10d"))%>%
ggplot(aes(group, numballs, colour = Treatment)) +
  geom_pointrange(aes(ymin = ci.lower, ymax = ci.higher), size = 0.1, position = position_dodge(width=0.5) ) +
  facet_wrap(~label, scales = "free_x", labeller = label_wrap_gen(width=25)) +
  theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1)) +
  xlab("") +
  ylab("Number of reported issues")

```
With respect to conflict, I distinguish between recent conflict (as
indicated by ACLED events that happened within the 12 months before the
list experiment) and historic conflict (1996-2012). Historic conflict
can only have had an indirect effect on SGBV, e.g. through changed
norms, as the list experiment only covers SGBV events within the past 12
months. Recent conflict can have a direct effect through perpetration
during the conflict event.

I first analyse victimization patterns by comparing sub-groups of the
respondents, based on one variable at a time. A full, multivariate
analysis will follow. With respect to conflict, I consider three ways of
splitting the sample in sub-groups: (i) respondents who live in
households that indicated (or not) in 2012 to have suffered loss of (or
damage to) property, including agricultural fields, due to conflict;
(ii) whether the respondent's household indicated in 2012 to have lost
any household members or family as a consequence of the conflict (or
not); and (iii) whether number of instances of violence against
civilians in ACLED data within a 10km radius during the past twelve
months was higher than the number of instances for the median household
(nor not)^[The results presented here are robust to using number of battles
    or number of fatalities rather than the instances of violence
    against civilians; using 5,15,20, 25 or 30km as a radius; and using
    a continuous variable, rather than a binary variable.].

Conflict exposure was high in the sample (see Table \@ref(tab:summstats)):
77% of the respondents reported having lost
property due to conflict between 1996 and 2012. 49% of the respondents
reported the loss of a family or household member. Again, conflict
exposure was high, even when limiting the time-span to one year prior to
the data collection. The mean number of violent conflicts within a 10km
radius was 6.73. This exposure differs across the territories (Table
\@ref(tab:confxterr)). While respondents in all territories
were greatly affected by conflict prior to 2012, those in Fizi were hit
harder. In the 12 months before the survey however, Uvira was in the
midst in an outbreak of violence, related to conflicts surrounding the
succession of traditional rulers in the chefferies of Bafuliiro and
Plaine de la Ruzizi. In fact, weeks before data collection in 2014 took
place, 30 civilians were killed in Mutarule, a village in the Plaine,
but not in my sample. This difference in recent and historic conflict
patterns means that households with conflict exposure pre-2012 are not
more likely to be victimized in 2013-2014 (see also Table \@ref(tab:determinants). Associations between pre-2012
violence and SGBV will thus not be the result of re-targeting of the
same households.


```{r table5confxterr, tab.cap='Conflict exposure by territory', tab.id='confxterr'}

read_csv(here("tables/violence_by_territory.csv")) %>%
  mutate(Territory = ifelse(Territory == "Overall", Territory,paste0(" ",Territory))) %>%
  tabulator(rows = "Variable",
            columns = "Territory",
            `Mean \n (SD)` =  as_paragraph(as_chunk(mean,digits=2,),as_chunk("\n("),as_chunk(sd,digits=2),")")) %>%
  as_flextable() %>%
  labelizor(j = "Variable", labels = varlabels, part = "all") %>% 

  #fit to page
  padding(padding = 1,part = "all") %>%
  padding(padding.top = 0, padding.bottom = 0) %>%
  autofit() 

```

The results of the sub-group analysis is displayed graphically in Figure 
\@ref(fig:confmeans). From the top two panels, it can be
seen that the difference between treatment and control is greater among
conflict-victimized respondents than among non-conflict victimized
respondents. The size of these differences is listed in Table
\@ref(tab:meandiffconf). Among those that indicated not having
lost property, the difference in number of issues faced between
treatment and control is 0.19, implying a SGBV victimization rate of
19%. The difference between Treatment and Control among respondents who
did lose property was 0.38. The difference in the differences between
these groups of 0.19 issues (this corresponds to coefficient $\beta_3$
in equation \@ref(eq:interaction) above) is not statistically significant.
When splitting the sample by households indicating having lost a family
or household member to conflict before 2012, the
difference-in-difference estimate is 0.37, indicating that incidence of
SGBV among respondents who lost family due to conflict is 37 percentage
points higher than among those who have not. This effect is significant
at the 5% level. Note that the SGBV could not have happened during the
same time as the conflict event(s): the SGBV happened twelve months
before the interview in 2014, while the conflict events happened before
2012.

When looking at more recent exposure to conflict, no clear patterns
emerge (see bottom panel of Figure
\@ref(fig:confmeans). SGBV incidence among women who have
more instances of violence near them than the median is 7 percentage
points lower than women who do not (top row of Table
\@ref(tab:meandiffconf)). However, this is not statistically
significant. I thus find no evidence of large-scale direct perpetration
of SGBV by armed groups in the one year before data collection, but also
no evidence of indirect effects of recent conflict.

The fact that conflict before 2012 correlates with SGBV, but recent
conflict does not, points at a more complex relationship between
conflict and SGBV than a simple direct effect due to perpetration by
armed groups. It is more likely that violence has an indirect effect
through changed norms. The fact that recent conflict seems not to have
an indirect effect either, may mean that this change of norms takes
time, or that the nature of recent conflict is different from historic
conflict.

```{r table6meandiffconf, tab.cap='Differences in numbers of issued faced in the list experiment, across conflict indicators', tab.id='meandiffconf'}

p2stars <- function(p){
  case_when(p < 0.01 ~ "***",
                    p < 0.05 ~ "**",
                    p < 0.1 ~ "*",
                    .default = "" )
}

format_stderror <- function(s){
  ifelse(is.na(s),"",sprintf("\n(%.2f)",s))
}


meandifftab <-
  read_csv(here("tables/meandifftab.csv")) %>%
  mutate(stat_subgroup = ifelse(stat_subgroup == "Control", " Control",stat_subgroup ),
        stat = ifelse(stat == "Mean", " Mean",stat),
        stat_subgroup = ifelse(is.na(stat_subgroup)," ",stat_subgroup))



meandifftab %>%

  #content
  filter(var %in% c("victimproplost", "victimfamlost", "acledviolence10d")) %>%
  tabulator(rows = c("var","group"),
            columns = c("stat","stat_subgroup"),
            `Mean` =  as_paragraph(as_chunk(estimate,digits=2),
                                   as_chunk(p2stars(p.value)),
                                   as_chunk(format_stderror(std.error)))) %>%
  as_flextable(spread_first_col = TRUE) %>%

  #labelling, alginment and formatting
  labelizor(labels = varlabels, part = "all") %>%
  mk_par(i = 1, j = c(1,2), as_paragraph(''),part = "header") %>%
  bold(i = ~ !is.na(var), bold = TRUE) %>%
  align(i = ~ !is.na(var), align = "left") %>%  

  #fit to page
  padding(i = c(2:4,6:8,10:12), j=1, padding.left=20) %>%
  padding(padding.top = 0, padding.bottom = 0) %>%

  autofit()


```

## Intra-household bargaining position {#intra-household-bargaining-position .unnumbered}

```{r marmeans, fig.cap =  "Comparison of means of issues faced across intra-household status", out.width = "100%", fig.align = 'center'}

read_csv(here("tables/dataforgraphs.csv")) %>%
left_join(read_csv(here("tables/cleandata_collabels.csv")),by = join_by(var == Variable)) %>%
filter(var %in% c("statpar", "bargresult"))%>%
ggplot(aes(group, numballs, colour = Treatment)) +
  geom_pointrange(aes(ymin = ci.lower, ymax = ci.higher), size = 0.1, position = position_dodge(width=0.5) ) +
  facet_wrap(~label, scales = "free_x", labeller = label_wrap_gen(width=25)) +
  theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1)) +
  xlab("") +
  ylab("Number of reported issues")


```

I then create sub-groups based on the intra-household bargaining
position of the respondents. I compare women across two variables.
First, I compare women across the relative status of the partners at the
time of marriage, by using family land-holdings as a proxy for status.
The 2014 survey contained a section on the marriage of the (spouse of
the) household head. In this section, respondents were asked whose
family owned more land, prior to the marriage: the wife's, the
husband's, or whether they had equal land. This choice of proxy was made
in consultation with local partners (including NGOs and universities),
and based on the importance of agriculture in the area. In 33% of the
cases, the husband's family had more land, in 21% of the cases the wife's
family did. Note that only 226 households responded to this question, as
some refused to give a definite answer (Table \@ref(tab:summstats)).

The second intra-household aspect I explore is derived from the results
of the bargaining game played with couples during the 2014 survey. I
create three groups, based on whether the joint decision is closer to
the husband's decision, to the wife's, or if the distance is equal. The
mean choice of the Female Respondents in the sample was 3.58; the Male
Respondents were slightly more risk-averse: their mean choice was 3.45 In 40%
of the cases, the couple decision was closest to the Male Respondent's
choice. In 27% it was closer to the Female Respondent's. Note that the
size of the sample here is smaller (only 87) than for the other variables
presented, as it was not always possible to have both the Male and
Female Respondent present at the same time for the interview.

Figure \@ref(fig:marmeans) displays the results from the sub-group
analysis. Overall, the difference between treatment and control is
larger for the sub-groups of respondents with a worse intra-household
bargaining position, indicating that the incidence of SGBV is higher
among these respondents. As suggested by the large size of the 95%
confidence intervals, some of these sub-groups are small. In Table
\@ref(tab:meandiffmar), these differences are tabulated,
including the sizes of the sub-groups. However, the variable definitions
are slightly different, due to the difficulties in interpreting
difference-in-differences between three sub-groups. For each variable,
two comparisons are tabulated: one, comparing households where the
female respondents had the better bargaining position with the two other
sub-groups, and one comparing households where the male respondent had
the better bargaining position with the two other sub-groups. Female
respondent in households where the family of the husband had the most
land prior to marriage were victims of SGBV in 50% of the cases, while
16% of the other respondents were. The difference of 33 percentage
points is statistically significant at the 10% level. In the other
comparison for the same variable, the difference is even larger, but not
statistically significant; perhaps due to the low number of women with
more pre-marital status than their husbands. The differences when split
by results from the bargaining game are larger still: 57 or 61
percentage points, depending on the groups used.

While these results may suggest that IPV is an important driver of SGBV,
the fact that I have no information on perpetrators means that this is
not certain.

```{r table7meandiffmar, tab.cap='Differences in numbers of issued faced in the list experiment, across intra-household status', tab.id='meandiffmar'}


meandifftab %>%

  #content
filter(var %in% c("husbmoreland", "wifemoreland", "barghusbandcloser", "bargwifecloser")) %>%
  tabulator(rows = c("var","group"),
            columns = c("stat","stat_subgroup"),
            `Mean` =  as_paragraph(as_chunk(estimate,digits=2),
                                   as_chunk(p2stars(p.value)),
                                   as_chunk(format_stderror(std.error)))) %>%
  as_flextable(spread_first_col = TRUE) %>%

  #labelling, alginment and formatting
  labelizor(labels = varlabels, part = "all") %>%
  mk_par(i = 1, j = c(1,2), as_paragraph(''),part = "header") %>%
  bold(i = ~ !is.na(var), bold = TRUE) %>%
  align(i = ~ !is.na(var), align = "left") %>%  

  #fit to page
  padding(i = c(2:4,6:8,10:12), j=1, padding.left=20) %>%
  padding(padding.top = 0, padding.bottom = 0) %>%

  autofit()


```

## Multivariate Regression analysis {#multivariate-regression-analysis .unnumbered}

In the preceding sections, I examined univariate relations between
variables of interest and the incidence of SGBV. However, such analysis
may suffer from omitted variables and spurious correlations. Here I move
to a richer specification, in order to prevent such biases, and assess
the relative importance of each driver. I expand equation \@ref(eq:interaction)
to simultaneously include indicators for
conflict and intra-household bargaining position. To reduce the risk of
multi-collinearity, I do not include the full set of variables discussed
above, but select one indicator for each, guided by the results obtained
above. A key criterion for selection is the number of respondents for
each indicator. The analysis of list experiments suffers from rapid loss
of power due to the indirect nature of the analysis. To mitigate this,
indicators that are available for large groups of respondents were
selected. For conflict, I include both the indicator for household
member killed before 2012 (as an indicator for historic conflict) and
violence against civilians from the ACLED data (as an indicator for
recent conflict); and for intra-household bargaining position a dummy
for the husband's family having the most land. I use the KICT Stata
package developed by @Tsai2019 to estimate these models. Interpretation
of the coefficients is the same as equation
\@ref(eq:interaction), but estimation is more efficient.

In order to reduce missing variable bias, I include a set of controls
that likely (co-) determine SGBV and the right-hand side variables
listed above. A full analysis of these determinants is provided in the
Appendix, Table
\@ref(tab:determinants). In addition, I include
variables that determine sample selection, as displayed in Table
\@ref(tab:sampleselection). In particular, I include the age of
the Female Respondent; indicators for the education of the Male and
Female respondents; asset holdings of the household, including livestock
and a tin roof; territory dummies; and an indicator for being in the
treatment group of any of the projects under evaluation for the survey.

In Table \@ref(tab:regress), I display the results of these
regressions. In columns 1-3 I rerun the univariate models from above.
Results are the same as before: both conflict history and
intra-household bargaining are associated with increased incidence of
SGBV. I In column 4 I present the full model. I find that women in a
marriage where their husband's family had more land before the marriage,
are 45 percentage points more likely than other women to be victim of SGBV.
Note that the pre-marriage status of women within the household is
uncorrelated to conflict (see Table \@ref(tab:determinants). Women who live in households
that lost a family or household member due to conflict prior to 2012 are
37 percentage points more likely to be victimized by SGBV than other women.
Of note is also the negative associated of the Female Respondent having
a secondary education: in this linear model, women with secondary
education are percentage points 125 less likely to be victimized. The fact
that the absolute value of this coefficient is higher than 1 is due to
the fact that linear models do not constrain predictions of
probabilities between 0 and 1.

The finding that conflict history is associated with an increase in
SGBV, while recent conflict is not, points to the indirect relationship
between conflict and SGBV, where conflict may affect SGBV rates not
through perpetration by armed groups, but by an increase in IPV. The
notion that IPV is a major driver of SGBV is reinforced by the fact that
both intra-household bargaining position and secondary education are
negatively associated with SGBV. This suggests that the position of
women is important in protecting them from human rights violations.

Caution should be taken with this interpretation, as results presented
here are not necessarily causal: women with higher education may differ
from other women in non-observable ways, and face lower victimization
because of that, rather than education. Furthermore, no data exists on
the perpetrators of the violence. The method of a list experiment does
not allow for follow-up questions to victimized women, as the
interviewer cannot know who to ask these follow up questions to.

```{r table8regress, tab.cap='Multivariate regression Results', tab.id='regress'}


add_scalars <- function(x,label,values){

  row_vector <- vector(mode="character", length = 1 + length(values) * 2)

  row_vector[1] <- label
  for (i in 1:length(values)) {
      footercolumn <- 1 + i * 2
      row_vector[footercolumn] <- values[i]
  }

  x %>%
    add_footer_row(top = FALSE, values = row_vector, colwidths = rep(1,ncol_keys(.)))

}


regs <-
  read_csv(here("tables/regression_results.csv")) %>%

  #add in varlables, and then change terms to manipulate ordering by tabulator
  left_join(read_csv(here("tables/cleandata_collabels.csv")),by = join_by(term == Variable)) %>%
  mutate(label = ifelse(term == "(Intercept)","Constant",label)) %>%
  mutate(order =case_when(term == "husbmoreland" ~ "01",
                          term == "victimfamlost" ~ "02",
                          term == "acledviolence10" ~ "03",
                          term == "attwifetotal" ~ "04",
                          term == "(Intercept)" ~ "99",
                          .default = "90"),
        term = paste0(order,term))


n <- c(449,402,379,350)

labels <-
  regs %>%
  distinct(term,label) %>%
  pull(label, name = term) 

model_labels <- paste0("(",1:4,")")
names(model_labels) <- paste0("reg",1:4)


regs %>%
  tabulator(rows = c("term"),
          columns = c("model"),
          `Beta` =  as_paragraph(as_chunk(estimate,digits=2),
                                   as_chunk(p2stars(p.value)),
                                   as_chunk(format_stderror(std.error)))) %>%
  as_flextable() %>%
  add_scalars("N",n) %>%

  labelizor(labels = labels, part = "all") %>%
  labelizor(labels = model_labels, part = "header") %>%
  mk_par(i = 1, j = 1, as_paragraph(''),part = "header") %>%
  
  theme_booktabs() %>%
  autofit()  


```

# Conclusion {#conclusion .unnumbered}

In this paper, I analysed the results from a list experiment, in order
to identify potential drivers of SGBV in Eastern Congo. Prevalence of
SGBV is high in Congo, however little is known about the victims, and
the drivers of victimization. In order to address this, I combined the
results from the list experiment with rich data, including a household
survey, a bargaining game, and conflict data.

The incidence rates I find are very high: 30% of the women in the sample
report having been the victim of SGBV in the past twelve months. Most
data collected on lifetime victimization arrives at similar rates,
suggesting that this estimate for a one-year window is high. The rate
found here may thus not be nationally, or regionally, representative.
This is likely due to the fact that women in the sample were recruited
among beneficiaries and potential beneficiaries of programs aimed at
assisting the most vulnerable women and households. It is to be expected
that incidence rates in this group are higher than for other groups. In
fact, I find that secondary schooling rates among women in my sample is
lower than the national or provincial average, and that incidence of
SGBV among women who have attended secondary school are significantly
lower than among other women.

When examining the backgrounds of the victims, I find that they are
likely to be married to higher-status men, have low intra-household
bargaining power, and have been exposed to violent conflict to the
extent where they have lost family or household members before 2012 (two
years before the list experiment). When comparing these effects in one
analysis, I find that the effect of intra-household dynamics is larger
than the effect of conflict. This contrasts with popular frames where
the conflict is seen as the primary driver of SGBV, but is in line with
previous literature suggesting that intimate partners are more likely
perpetrators of SGBV than members of armed groups [see e.g.
@Peterman2011].

Taken together, these finding imply that human rights violations do not
end when the conflict ends. The disruption of social norms may cause
women (and perhaps men, but the present data set does not cover them) to
suffer from violence long after the last shot has been fired. A focus of
rape as a "weapon of war" may thus be too narrow to address these
violations. This is not to say there direct perpetration of SGBV by
armed forces is not a problem in Congo. There is ample proof that
large-scale violations have been committed by armed forces, especially
historically. The conflict has undergone changes throughout the years,
and with it the kinds of human rights violations perpetrated. The
massacre in Mutarule in the weeks before data collections did see 30
innocent civilians murdered, but there are no reports of rape.
Furthermore, focusing efforts to assist women on the victims from such
attacks risks missing women victimized in their homes, far away from any
fighting. Structural changes encouraging women's education and tangibly
raising their status are needed to protect these women as well.

There are three large caveats with these findings: (i) causal
interpretation is difficult due to the cross-sectional nature of the
data; (ii) little analysis could be done on the perpetrators of the
violence, as indirect questioning precludes probing into this. More
research is needed to to address these important issues; and (iii) I did
not collect data on the victimization of men. More research is needed to
address these.

\clearpage

# Appendix {#appendix .unnumbered}

```{r tablea1sampleselection, tab.cap='Sample Selection for the gender module', tab.id='sampleselection'}



coeffs <-
  read_csv(here("tables/sampleselection_coeffs.csv")) %>%
  left_join(read_csv(here("tables/cleandata_collabels.csv")),by = join_by(term == Variable)) %>%
  mutate(label = ifelse(term == "(Intercept)","Constant",label)) %>%
  mutate(order =case_when(term == "husbmoreland" ~ "01",
                          term == "victimfamlost" ~ "02",
                          term == "acledviolence10" ~ "03",
                          term == "attwifetotal" ~ "04",
                          term == "(Intercept)" ~ "99",
                          .default = "90"),
        term = paste0(order,term)) 

labels <-
  coeffs %>%
  distinct(term,label) %>%
  pull(label, name = term) 


scalars <-
read_csv(here("tables/sampleselection_scalars.csv"))


n <- scalars[["nobs"]]



model_labels <- c("(1)\nWife","(2)\nHusband","(3)\nCouple")
names(model_labels) <- paste0("reg",1:3)

coeffs %>%
  tabulator(rows = c("term"),
          columns = c("model"),
          `Beta` =  as_paragraph(as_chunk(estimate,digits=2),
                                   as_chunk(p2stars(p.value)),
                                   as_chunk(format_stderror(std.error)))) %>%
  as_flextable() %>%
  add_scalars("N",n) %>%
  labelizor(j = 1, labels = labels, part = "all") %>%
  labelizor(labels = model_labels, part = "header") %>%  
  mk_par(i = 1, j = 1, as_paragraph(''),part = "header") %>%

  theme_booktabs() %>%
  autofit()


```



```{r tablea2determinants, tab.cap='Determinants of Violence and Bargaining Power', tab.id='determinants'}




coeffs <-
  read_csv(here("tables/determinants_coeffs.csv")) %>%
  left_join(read_csv(here("tables/cleandata_collabels.csv")),by = join_by(term == Variable)) %>%
  mutate(label = ifelse(term == "(Intercept)","Constant",label)) %>%
  mutate(order =case_when(term == "husbmoreland" ~ "01",
                          term == "victimfamlost" ~ "02",
                          term == "acledviolence10" ~ "03",
                          term == "(Intercept)" ~ "99",
                          .default = "90"),
        term = paste0(order,term))


labels <-
  coeffs %>%
  distinct(term,label) %>%
  pull(label, name = term) 


scalars <-
read_csv(here("tables/determinants_scalars.csv"))


n <- scalars[["nobs"]]


model_labels <- c("(1)\nFamily MR \nhad more land","(2)\nBargaining:\ncloser to FR","(3)\nConflict\npre-2012:HH\nmember killed
")
names(model_labels) <- paste0("reg",1:3)

coeffs %>%
  tabulator(rows = c("term"),
          columns = c("model"),
          `Beta` =  as_paragraph(as_chunk(estimate,digits=2),
                                   as_chunk(p2stars(p.value)),
                                   as_chunk(format_stderror(std.error)))) %>%
  as_flextable() %>%
  add_scalars("N",scalars[["nobs"]]) %>%
  labelizor(j = 1, labels = labels, part = "all") %>%
  labelizor(labels = model_labels, part = "header") %>%
  mk_par(i = 1, j = 1, as_paragraph(''),part = "header") %>%
  theme_booktabs() %>%

  padding(padding.left = 0.1, padding.right = 0.1, part = "header") %>%
  fontsize(size = 7, part = "all") %>%

  autofit() 

```

\clearpage

# References