IFPRI Discussion Paper 02232 

December 2023 

Gender, Deliberation, and Natural Resource Governance 

Experimental Evidence from Malawi  

Amanda Clayton 

Boniface Dulani 

Katrina Kosec 

Amanda Lea Robinson 

 

Poverty, Gender, and Inclusion Unit 



 

 

INTERNATIONAL FOOD POLICY RESEARCH INSTITUTE 

The International Food Policy Research Institute (IFPRI), a CGIAR Research Center established in 1975, 

provides research-based policy solutions to sustainably reduce poverty and end hunger and malnutrition. 

IFPRI’s strategic research aims to foster a climate-resilient and sustainable food supply; promote healthy 

diets and nutrition for all; build inclusive and efficient markets, trade systems, and food industries; 

transform agricultural and rural economies; and strengthen institutions and governance. Gender is 

integrated in all the Institute’s work. Partnerships, communications, capacity strengthening, and data and 

knowledge management are essential components to translate IFPRI’s research from action to impact. 

The Institute’s regional and country programs play a critical role in responding to demand for food policy 

research and in delivering holistic support for country-led development. IFPRI collaborates with partners 

around the world.  

 

AUTHORS 

Amanda Clayton (a.bclayton@berkeley.edu ) is an assistant professor in the Department of Political 

Science at the University of California, Berkeley. 

 

Boniface Dulani (bdulani@ipormw.org) is an associate professor in the Department of Politics and 

Government at the University of Malawi and Director of Research at the Institute of Public Opinion and 

Research. 

 

Katrina Kosec (k.kosec@cgiar.org)  is a senior research fellow in the Poverty, Gender, and Inclusion Unit 

at the International Food Policy Research Institute. 

 

Amanda Robinson (robinson.1012@osu.edu) is an associate professor in the Department of Political 

Science at the Ohio State University.  

 

 

 

 

 

 

 

 

 

 
Notices  

 
1 IFPRI Discussion Papers contain preliminary material and research results and are circulated in order to stimulate discussion and 
critical comment. They have not been subject to a formal external review via IFPRI’s Publications Review Committee. Any opinions 
stated herein are those of the author(s) and are not necessarily representative of or endorsed by IFPRI.  
 
2 The boundaries and names shown and the designations used on the map(s) herein do not imply official endorsement or 
acceptance by the International Food Policy Research Institute (IFPRI) or its partners and contributors. 
 
3 Copyright remains with the authors. The authors are free to proceed, without further IFPRI permission, to publish this paper, or any 
revised version of it, in outlets such as journals, books, and other publications. 

mailto:a.bclayton@berkeley.edu
mailto:bdulani@ipormw.org
mailto:k.kosec@cgiar.org
mailto:robinson.1012@osu.edu


iii 

 

ABSTRACT 
 

Initiatives to combat climate change often strive to include women’s voices, but there is limited evidence 

on how this feature influences program design or its benefits for women. We examine the causal effect of 

women’s representation in climate-related deliberations using the case of community-managed forests in 

rural Malawi. We run a lab-in-the-field experiment that randomly varies the gender composition of six-

member groups asked to privately vote, deliberate, then privately vote again on their preferred policy to 

combat local over-harvesting. We find that any given woman has relatively more influence in group 

deliberations when women make up a larger share of the group. This result cannot be explained by 

changes in participants’ talk time. Rather, women’s presence changes the content of deliberations towards 

topics on which women tend to have greater expertise. Our work suggests that including women in 

decision-making can shift deliberative processes in ways that amplify women’s voices. 

 

Keywords:  Gender, women’s empowerment, communal forest management, decision-making, poverty 

 

  



iv 

 

ACKNOWLEDGMENTS 
 

Research funding was provided by the Balzan Foundation (under the terms of a prize awarded in 2017 to 

Professor Robert O. Keohane and administered by Princeton University and the Center for Advanced 

Study in the Behavioral Sciences at Stanford University under his supervision) and Vanderbilt University. 

It was undertaken as part of the CGIAR Research Initiative on Gender Equality and the CGIAR Research 

Initiative on Fragility, Conflict, and Migration. The CGIAR Initiative on Gender Equality aims to use 

impactful gender research to address the four dimensions of gender inequality by applying gender-

transformative approaches to harmful norms, bundling socio-technical innovations for women’s 

empowerment, leveraging social protection to increase women’s access to and control over resources, and 

promoting inclusive governance and policies for increased resilience. The CGIAR Initiative on Fragility, 

Conflict, and Migration addresses challenges to livelihood, food, and climate security faced by some of 

the most vulnerable populations worldwide. The Initiative focuses on building climate resilience, 

promoting gender equity, and fostering social inclusion. 

We would like to thank all funders who supported this research through their contributions to the CGIAR 

Trust Fund: https://www.cgiar.org/funders/. Data were collected in collaboration with the Institute for 

Public Opinion and Research (IPOR) in Zomba, Malawi. We thank Jan Duchoslav, Jordan Kyle, Graeme 

Blair, Bonnie Meguid, Ruth Meinzen-Dick, and Claudia Ringler for helpful comments; Hannah Swila and 

Funny Muthema for their project management; Juliet Magombo and Sanudi Maoni for their field 

supervision; and Lucia Carrillo and Juan Manuel Peréz for their research assistance. We are also grateful 

for feedback that we received at Princeton University; Toronto Metropolitan University; the UCLA 

Conference on the Politics of Climate Change and the Environment; the 2023 APSA Annual Meeting; the 

USAID DRG Evidence and Learning Talk Series; the EGAP Meeting on Climate Governance; an EGAP 

feedback session, and the 2023 Empirical Study of Gender (EGEN) Meeting. 

 

 

https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.cgiar.org%2Finitiative%2Fgender-equality%2F&data=05%7C02%7CP.Fowlkes%40cgiar.org%7C854e6b7d109744e047e208dc0e0f2528%7C6afa0e00fa1440b78a2e22a7f8c357d5%7C0%7C0%7C638400705837920579%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=HVfsKaQvLMJH9%2B%2BlgrfXN7fD80J%2BMfMfuBKS8j%2FqbEI%3D&reserved=0
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.cgiar.org%2Finitiative%2Ffragility-conflict-and-migration%2F&data=05%7C02%7CP.Fowlkes%40cgiar.org%7C854e6b7d109744e047e208dc0e0f2528%7C6afa0e00fa1440b78a2e22a7f8c357d5%7C0%7C0%7C638400705837920579%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=xXjNci4G3fdcGAaH2gOjrfcxOmdZK%2BykHB2%2Br04VqPQ%3D&reserved=0
https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.cgiar.org%2Finitiative%2Ffragility-conflict-and-migration%2F&data=05%7C02%7CP.Fowlkes%40cgiar.org%7C854e6b7d109744e047e208dc0e0f2528%7C6afa0e00fa1440b78a2e22a7f8c357d5%7C0%7C0%7C638400705837920579%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=xXjNci4G3fdcGAaH2gOjrfcxOmdZK%2BykHB2%2Br04VqPQ%3D&reserved=0
https://www.cgiar.org/funders/


Introduction

Women’s inclusion is now the norm in global and local initiatives to combat climate change.1

Interventions that target women’s participation often reference the disproportionate burden women

face from a warming planet, particularly in rural agricultural settings (Brulé 2022; Deininger

et al. 2023). Yet how women’s representation affects climate-related governance is still poorly

understood. We examine the causal effect of women’s representation in deliberations to address

the over-harvesting of community-managed forests in rural Malawi. The rapid decline of Malawi’s

forest cover makes it a typical case of tropical deforestation—the second largest source of greenhouse

gas emissions after fossil fuel combustion, and estimated to cause up to a quarter of anthropogenic

carbon emissions worldwide (Kindermann et al. 2008).

We run a lab-in-the-field experiment that randomly varies the gender composition of six-member

groups asked to privately vote, deliberate, and then privately vote again on their favored solution

from a set of policies aimed to combat deforestation in nearby community-managed forests. We

worked with community leaders to assemble groups ranging from all women to all men with each

of the seven possible permutations randomly assigned. We measure influence in several ways,

including through participants’ assessments of their own influence, through the assessment of a

hypothesis-blind enumerator observing group discussions, and through a secret vote by all group

members to select the most influential participant. Across measures, we find that women’s relative

influence increases when there are more women in the group. Put another way, women’s inclusion

does more than increase women’s aggregate influence, it also increases the likelihood that any given

woman will influence group deliberations. Our data also reveal that peer assessments of women’s

influence increase particularly starkly among men participants, who become much more likely to

recognize women’s influence in settings with more women. Finally, we find that when women are in

the majority, group decisions are more likely to match women’s pre-treatment policy preferences.

1For example, within the United Nations Framework Convention on Climate Change (UNFCCC) there
is a formal “Women and Gender Constituency” and a “Gender Action Plan.” Likewise, the Green
Climate Fund has a “Gender Policy and Action Plan.” See http://womengenderclimate.org and
http://www.unwomen.org/en/news/stories/2017/11/announcement-first-ever-gender-action-plan-on-

climate-action-adopted

1



We next test observable implications of potential mechanisms that may explain our findings.

Contrary to our pre-specified expectations, we do not find that women speak significantly more

in the company of more women. We also do not find evidence that group dynamics become

more collaborative as women’s representation increases. Rather, we find that the substance of

group discussions changes as women’s presence grows. Using structural topic models based on

the discussion transcripts, we find that group discussions tend to include more time on topics on

which women have greater socially recognized expertise (cooking methods and replanting incentives)

and less time on topics for which men have greater socially recognized expertise (community and

government enforcement) in groups with more women, likely granting women more authority in these

settings. Extensive qualitative evidence also supports this interpretation. From interviews with key

stakeholders, separate focus groups with villagers, and observations of actual village meetings on

natural resource management, we find evidence that policies to combat deforestation tend to affect

men and women in different ways because of longstanding gender roles in forest management.

Moreover, the gendered division of forest-related labor maps onto the gender differences that we

observe in our discussion transcripts.

Our findings speak to a growing body of work that investigates how women’s representation

shapes deliberative processes and outcomes. This research collectively demonstrates that women’s

presence matters; it shapes policy decisions in ways that tend to reflect women’s preferences

and priorities. For instance, in legislative settings, women’s representation is associated with

the increased prioritization of goods women tend to prefer, including public health (Clayton and

Zetterberg 2018), drinking water in rural areas of the Global South (Bratton and Ray 2002;

Chattopadhyay and Duflo 2004), women’s rights (Franceschet and Piscopo 2008), and policies that

broadly support working mothers (e.g., Kittilson 2006; Weeks 2022). Moreover, observational work

suggests that women change their behavior in the presence of more women. For instance, in New

Zealand, Grey (2002) finds that women parliamentarians verbally represent themselves as women

more often after surpassing fifteen percent of the legislature, and in Argentina, Barnes (2016) finds

that women’s relative propensity to collaborate with other women increases as they comprise a larger

2



share of the legislative chamber. Our results lend new insights into the causal dynamics behind these

findings. We find that women’s presence in community-managed environmental bodies shapes the

content of discussions in ways that likely grant women more authority, particularly in the eyes of

men.

Our work also builds on an influential body of scholarship that specifically examines how a

group’s gender composition causally affects women’s willingness to participate in group discussions

(Karpowitz and Mendelberg 2014; Karpowitz et al. 2023; Born, Ranehill and Sandberg 2018).

Collectively, this scholarship suggests that women are constrained by social expectations about

who should participate in group decisions, and these constraints seem to be ameliorated as the

number of women in the group increases. These insights inform our theorizing as we apply them

for the first time outside of the Global North university setting.

Finally, our work speaks to a growing research agenda on how the climate crisis is gendered (see,

e.g., Brulé 2022; Bush and Clayton 2022). In the Global South, women face the precarious position

of both being more affected by extreme weather events caused by climate change and having little

say in local decision-making focused on mitigation and adaptation (Kumar and Quisumbing 2014;

Agarwal 1992). We draw on a rich literature in political science and economics on interventions

to improve the management of common pool resources (Ostrom 1990; Slough et al. 2021; Samii

et al. 2014; Meinzen-Dick et al. 2022). We also build on observational work that has examined how

the inclusion of marginalized resource users affects natural resource management. For instance,

Agarwal (2009, 2010) studies forest user groups in India and Nepal and finds a positive correlation

between the proportion of women on governing bodies and improved forest governance and resource

sustainability. Recent work from India further shows that the inclusion of marginalized ethnic

groups (scheduled tribes) improves forest conservation (Gulzar, Lal and Pasquale 2023). We move

this literature forward by causally examining how women’s representation in decision-making around

common pool resources causally affects deliberation itself. Our findings provide new evidence that

including women in local initiatives to combat climate change can shift group deliberations in ways

that amplify women’s voices.

3



Gender and Deforestation Preferences

We are interested in whether and how women are able to influence deliberations over the management

of community forests. An initial question is whether men and women have different preferences

on community forest management. Whereas a robust literature across lower-income countries

suggests that men and women tend to hold different policy priorities (Chattopadhyay and Duflo

2004; Gottlieb, Grossman and Robinson 2018), this work rarely examines issues related to the

environment. Moreover, related research presents mixed expectations. On the one hand, some

recent experimental work suggests that women might be more conservation-minded than men. For

example, through behavioral game experiments conducted in Peru, Tanzania, and Indonesia, Cook,

Grillos and Andersson (2019) find that gender-balanced groups indicate a greater willingness to

reduce tree harvesting compared to men-majority groups. Moreover, women’s traditional roles in

very low-income countries like Malawi—including gathering firewood, tending crops, and collecting

drinking water—may mean that they are especially affected by the depletion of natural resources,

which could motivate resource conservation (Deininger et al. 2023; Mawaya and Kalindekafe 2010).

However, other studies find that men are better stewards of the environment in low-income

agricultural societies, either because they are more likely to adopt new technologies and resource

monitoring practices that are associated with improved sustainability (Mai, Mwangi and Wan 2011;

Mwangi, Meinzen-Dick and Sun 2011), because they have greater interactions with conservation

agencies (Villamor et al. 2014), or because men are more likely to exploit forest resources for

commercial purposes (Mawaya and Kalindekafe 2010). Women in these societies also face many

informal barriers to accessing information and fully participating in decision-making due to lower

education levels when compared to men, gender norms, and gendered differences in access to

resources (Mawaya and Kalindekafe 2010; Alkire et al. 2013; Mudege et al. 2017). Additionally,

in very low-income countries like Malawi, men tend to know more about climate change and feel

more strongly that actions should be taken to stop climate change than do women. In the 2022

Afro-Barometer survey, 79 percent of Malawian men reported having heard about climate change

versus 69 percent of women. Moreover, among those aware of climate change, 50 percent of men

4



strongly agreed that “ordinary Malawians can do a lot to stop climate change,” whereas only 39

percent of women chose this response option.

These mixed findings leave an open question about the size and scope of gender gaps on this

issue, and we have no a priori expectations here. As a first step below, we inductively explore

whether and how men’s and women’s preferences diverge in this policy area before investigating

how women’s representation influences group deliberations.

How Gender Composition Shapes Group Deliberations

We are interested in the content of men’s and women’s preferences on deforestation policy and the

extent to which women’s representation affects their relative influence in group deliberations on this

topic. Our focus on women’s influence relative to their share of the group is a hard test. Most work

on how women’s representation affects policy outcomes examines whether women’s presence matters

in the aggregate sense, sifting policy outcomes to be more aligned with women’s preferences. Here,

however, we are interested in the influence of the average woman, and how this changes in settings

with different gender compositions. Our primary pre-registered expectation is that women will have

more relative (not just absolute) influence in group decisions as their representation increases. As a

result, we anticipate that group deliberations and group decisions will be more reflective of women’s

preferences in increasingly women-majority groups.2

We expect this process to occur through three potential channels. First, women may participate

more actively in group deliberations as their share of the group increases. This may be because

women feel more confident in voicing their opinions in settings with more women. When women

are in the minority, they may feel more reticent to share their views in the presence of men

because social norms tend to ascribe men with more authority, particularly in the realm of political

decision-making. There may also be strong social stigmas against being overly outspoken or

disagreeing openly with men. If participation is positively correlated with influence (Karpowitz,

2Our pre-analysis plan (PAP) is included in SI K. The PAP also registers additional expectations (e.g.,
expectations related to women’s perceptions of self-efficacy) that we hope to test in future work.

5



Mendelberg and Shaker 2012)—that is, if talking more makes one more influential—then we expect

that any given woman will speak more and thus have more influence over group deliberations as

the number of women grows.

The second mechanism through which women may gain more influence in settings with more

women is if they are increasingly recognized by others for the contributions that they do make.

There are many subtle ways in which speech can be either recognized or ignored, and women are

often not acknowledged for their contributions to group deliberations to the same extent as men

(Parthasarathy, Rao and Palaniswamy 2019; Clayton, Josefsson and Wang 2014). One way that

any given woman might gain influence in settings with more women is if other women are more

likely to acknowledge her contributions. Men may be more likely than women to “talk past” a

woman speaker, pivoting the subject or interrupting her (Holmes 2013). In this case, we would

expect women’s relative influence to increase in groups with more women simply because there are

more group members predisposed to recognize a fellow woman’s contributions (i.e., more women).

Another possibility is that other group members (men or women) change their behavior in the

presence of more women. This could occur if norms of speech change in groups with more women

causing decision-making bodies to generally become more collaborative (Barnes 2016; Holman and

Mahoney 2018). In this instance, the average tone of discussions might change towards greater

recognition of everyone’s contributions, including those made by women.

The final channel through which women might gain more authority in group settings is by

altering the substantive content of group discussions. Above, we described how preferences towards

deforestation policy might be gendered. In such instances, conversations with more women may

focus on different aspects of the problem of overharvesting of forest products and its solutions. As

an example, women are responsible for cooking in daily life in Malawian villages and one of the main

uses of forest resources is firewood harvesting and charcoal production, both of which women use in

cooking (Mawaya and Kalindekafe 2010). If a group with more women spent more time discussing

cooking methods as a dimension of forest management, women’s substantive expertise could lend

them more authority than they might have in a group discussing other policy dimensions.

6



In sum, we theorize that increases in women’s representation may change group dynamics in

three distinct ways that could result in women having more relative influence in group deliberations:

(1) by compelling women to talk more, (2) by making groups more collaborative and thus receptive

to women’s views, and/or (3) by changing the content of group discussions. While these mechanisms

are likely interdependent—for instance, women may talk more on subjects on which they feel

confident—for the purposes of testing distinct implications associated with each mechanism, we

consider each separately below.

Our final pre-registered expectation pertains to the durability of gender gaps in participation

and influence more broadly. Although we expect that women’s relative influence will increase in

groups composed of greater shares of women, we also expect that across all mixed-gender groups,

men will participate more actively and have relatively more influence than women, on average. In

Malawi, as in most other settings around the world, patriarchy is an organizing feature of daily

life. In our pre-treatment survey, we find that men have more interest in politics than women, have

more confidence in their own political abilities, and are more likely to have recently contacted a local

or national leader (see Supplementary Information [SI] Table SI.1). The persistence of entrenched

gender roles and pre-existing gender gaps in political participation means that women are unlikely

to be as active as men in group discussions. Thus, we expect that gender gaps will shrink as women’s

presence grows, but they are unlikely to fully close in any mixed-gender setting.

Malawi’s Communally Managed Forests

Malawi is in the midst of a deforestation crisis. Sixty-five percent of Malawi’s forests are located

on customary land, and communities overharvest these forests for timber, charcoal production,

firewood, and livestock grazing (Ngwira and Watanabe 2019). The results of deforestation and

forest degradation have been devastating. Between 1972 and 1992, Malawi’s total forest cover fell

from 47 percent of total land cover to 20 percent. Estimates of the current rate of deforestation

are between an annual average loss of 164,000-460,600 hectares of forest cover, the highest rate of

7



deforestation in the Southern African Development Community (UN-REDD Programme 2017).

The over-exploitation of forest reserves threatens the livelihoods of communities that depend

on them (Mawaya and Kalindekafe 2010). Yet avoiding the over-harvesting of community-managed

forests in Malawi and other developing countries is extremely difficult as community members

face enormous incentives to over-harvest. Households often rely on selling forest products like

wood and charcoal for subsistence and cannot afford to forgo the income generated by current

harvesting behavior despite its high long-term costs (Ngwira and Watanabe 2019). The relatively

slow regeneration rates of forests also mean that future income streams can take years to materialize

(Eisenbarth, Graham and Rigterink 2021).

In Malawi, forests are governed by multiple actors. The main bodies responsible for day-to-day

forest management are called Village Natural Resource Management Committees (VNRMCs).

These groups manage protected forest areas within or adjacent to their communities. Their main

mandate is to set penalties for those who harvest beyond their family allotment, and these penalties

are typically enforced in collaboration with the local chief. For some forests, there are also elected

groups called Block Management Committees (BMCs), whose function is to oversee the management

of government-owned forests. These forests tend to be bigger and are the site of most large-scale

illegal harvesting used for timber and charcoal production. The BMCs tend to have more authority

than the VNRMCs, including working with forest guards who are employed by the Ministry of

Natural Resources and Climate Change and monitoring and enforcing penalties for illegal harvesting.

Our research question on the role of gender in forest management is of particular relevance in

Malawi, where women are formally required to occupy up to fifty percent of the positions on these

governance bodies; however, these requirements are often loosely enforced, if at all (UN-REDD

Programme 2017).

8



Study Design and Estimation Strategy

Our study sites comprise 90 communities adjacent to Zomba-Malosa Forest Reserve in southern

Malawi. We randomly selected these communities from among the 216 communities within three

kilometers of Zomba-Malosa Forest Reserve boundary, a distance that implies that activities within

the reserve are accessible and thus relevant to them. Figure 1 shows a map of the study area.

Between 2000 and 2021, Zomba-Malosa Forest Reserve lost 25 percent of its tree cover, both due

to commercial timber loggers and local harvesting for cooking fuel (Global Forest Watch 2021).

The surrounding communities rely upon wood harvested from the reserve for their cooking fuel and

heating. For instance, 94 percent of the respondents in our survey indicated relying on local forest

stock for their cooking needs. Moreover, many residents’ livelihoods depend on harvesting and

reselling wood and charcoal in local markets (Moyo, Chikuni and Chiotha 2016). At the same time,

communities near the reserve also pay the short- and long-term costs of deforestation, including

devastating mudslides and flash floods, reduced water supply, worse air quality, soil degradation,

increased disease burdens, and less overall forest land for future use. Past research demonstrates that

most residents understand the negative implications of deforestation but face significant incentives

to exploit the resource beyond sustainability (Moyo, Chikuni and Chiotha 2016). Virtually all

participants in our pre-treatment survey (97 percent) indicated that the over-harvesting of nearby

forests was a “big problem” in their community.3

In the communities adjacent to the forest reserve, we worked with a Malawian research firm to

lead group discussions with villagers on the issue of deforestation. These citizen groups serve two

purposes. First, this exercise allows us to establish whether and how deforestation preferences and

deliberations are shaped by participant gender, laying the groundwork for future work to interrogate

this question within actual governance bodies. Second, Zomba-Malosa Forest Reserve is co-managed

by multiple actors (as described above), including citizen governance bodies. We designed our study

to mimic the actual forest governance structures in place such that our results might be applicable

to these bodies, as well as to the many other areas around the world where conservation efforts

3Response options were: not a problem at all, somewhat of a problem, or a big problem.

9



District	Boundary
Towns
Forest	Reserve
Study	Area

Zomba-Malosa
Forest
Reserve

Figure 1 Map of the Study Area within 3km of the Zomba-Malosa Forest Reserve in Zomba and
Machinga Districts, Malawi. Map created by authors using GIS files.

are managed in part or in full by local communities (see, e.g., Slough et al. 2021; Gulzar, Lal and

Pasquale 2023; Ostrom 1990).

In each community, we worked with community leaders to assemble seven groups of six members

each, with every possible combination of women and men represented (i.e., a group of zero women,

six men; a group of one woman, five men; . . . , and a group of six women, zero men).4 The study

proceeded in the following steps:

1. Introduction of the study to the whole group (42 participants across seven groups
of six members each)

2. Consent process with each participant, one by one

3. Randomization into groups by drawing cards

4Our study took place in August, 2022. We visited each of our study sites beforehand and asked the local
chief to assemble ≈ 50 villagers, equally men and women, on the study day. From this group, men and women
picked randomization numbers out of separate buckets to ensure the desired group compositions until there were
no numbers left. Refreshments were provided (both to study participants and to the “extras”) but, in compliance
with local research regulations, no compensation was given. More details are provided in SI J. Men and women
participants across groups had similar socio-economic characteristics, suggesting that randomization worked.

10



4. Pre-treatment survey, including a secret vote on preferred policy from among a set
of seven options

5. Group discussion (i.e., deliberation over policies)

6. Secret vote over the same set of policies

7. Post-treatment survey

Our strategy has several advantages. Step (4) allows us to collect a pre-treatment measure of

men’s and women’s preferences before any deliberation takes place. Through the secret vote in step

(6), we can gauge whether and how women and men were differently persuaded by the preceding

discussion. And step (7) allows us to gauge how influential men versus women were in shifting the

group’s decisions to reflect their own preferences, and allows us to collect responses on whom each

participant viewed as the most influential member of their group. By recording and transcribing

the discussions, the design also allows us to measure how various group compositions affect women’s

participation as well as the content of group discussions across treatment groups.

The key feature of our intervention is that the gender composition of deliberating groups is

randomly assigned. The resulting sample includes 630 groups of 6 individuals each (n=3,780; 1890

men and 1890 women), with 90 groups (540 participants) in each of the seven treatment conditions.

By blocking at the community (i.e., village) level, we have one group in each of the seven treatment

conditions within each of the 90 communities. This allows us to use community fixed effects to

account for any village-specific characteristics, and thus lends greater precision to our estimates.

After administering the pre-treatment survey to each respondent, the facilitator led the group in a

discussion about deforestation policy. The discussions were all done in Chichewa, Malawi’s national

language, which is widely spoken in the area. All groups followed the same format, responding to

and deliberating over two prompts. The first prompt asked the group to reflect on the issue of

climate change generally. The facilitator asked the group:

First let’s start with discussing the issue of climate change. Do you think climate change
will affect this community? If so, how?

This first prompt allows us to better understand how respondents view the issue of climate

change in general, and whether and how they connect it to the very local problem of deforestation.

11



Overwhelmingly, respondents saw the degradation of communal forests as connected to both local

climate issues, such as soil erosion and mudslides, and to more generalized climate change issues,

such as unstable weather patterns. The response below is representative:

Climate change will affect us so much with things like disasters, like floods, scarcity of
rainfall, and soil degradation. This means we will not have food. Hunger will hit us. If
rain is scarce then we will not have water. Drought will be everywhere and it will be
hard on us. Hunger will be everywhere. (Man, Matapwe Village, Zomba District)5

The second prompt introduced several policies to combat deforestation. The policy options were

drawn from a review of the literature on deforestation in Malawi and tropical deforestation in other

developing countries and from eight initial scoping focus group discussions in communities adjacent

to the study sites. They would thus likely be options with which participants would be familiar.

The prompt read as follows:

Now we will shift to discussing the problem of deforestation. As we explained earlier,
we want to understand how Malawians think about potential solutions to the problem of
over-harvesting of forest products. Before this discussion, we asked each of you about
your personal opinion on some solutions that others in the country have suggested. Now,
we’d like you to come together as a group to discuss which solution you think will be most
effective to stop the problem. After this discussion, each of you will vote on your preferred
solution. We will collect each group’s vote and share this information anonymously with
officials in the local forestry offices. The proposed solutions are: [Moderator shows cards
with pictures depicting each solution while describing each, shuffling cards so that the
order of introducing each solution is random.]

• Community Enforcement: Set rules/by-laws against over-harvesting and charcoal
production which are monitored and enforced by a community committee or the
chief.

• Government Enforcement: Set rules/by-laws against over-harvesting and charcoal
production which are monitored and enforced by government-employed forest guards.

• Replanting Incentives: Create an incentive program that pays community conservation
groups for each seedling that is planted in communal forests and survives the first
year.

• Civic Education: Offer training to make members of the community aware of the
consequences of over-harvesting.

5All village names attached to direct quotations are pseudonyms.

12



• Alternative Cooking Methods: Provide materials and training to use alternative
cooking techniques (e.g., chitetezo mbaula stove) or alternative fuel (e.g., briquettes)
to reduce demand for wood.

• Jobs Training: Provide small business training for those individuals who currently
engage in over-harvesting, so that they can provide for their families without harvesting
trees.

First, please go through and discuss each proposed solution as a group, touching on the
pros and cons of each proposed solution. I will give you time to discuss among yourselves
without weighing in. When you are done discussing, we will ask each of you one-by-one
in private to tell us your vote for the solution you think is most likely to be effective, and
then I will tally the votes and report the solution(s) with the most votes.

We recorded, transcribed, and translated all group deliberations and merged each respondent’s

contributions with his/her pre- and post-treatment survey responses. In total, our data comprise

approximately 20,000 unique statements (utterances) across the 3,780 participants.6 Each group

had a facilitator who led the group discussion and a separate research assistant (note taker) who,

in addition to administering the pre- and post-treatment surveys, also observed and recorded

the group dynamics through an enumerator survey. Our analyses include both individual (i.e.,

respondent) and group-level outcomes.7 For individual-level outcomes, we use OLS models with

survey enumerator fixed effects, village (community) fixed effects, and standard errors clustered by

group-village (i.e., discussion groups nested within villages). For group-level outcomes, we use OLS

models with discussion facilitator fixed effects, village (community) fixed effects, and standard errors

clustered by village.8 On average, men and women participated quite actively, with 90 percent of

men participants and 89 percent of women participants speaking at least once during the group

deliberations. On average, group discussions of the two prompts lasted about 35 minutes.

6A statement is something that is preceded by and then followed by another speaker. It can thus be as short as
one word or as long as several paragraphs.

7Enumerators administered the surveys and observed focus group dynamics, while facilitators ran the discussion
groups. Enumerators were trained to understand that the study was about deforestation, not about gender, thus
reducing incentives to report what they think the researchers hope to find (i.e., demand effects).

8Note that for the individual-level outcomes, we use enumerator fixed effects and for the group-level outcomes, we
use facilitator fixed effects. The enumerator delivered the surveys, so fixed effects here capture any enumerator-specific
variation in survey administration or response biases. Facilitators moderated the focus groups, so fixed effects here for
group-level outcomes capture any facilitator-specific variation in moderating the focus group discussions or ensuing
differences in group dynamics.

13



For completeness, for each measure of influence that we describe below, we run models with

three different specifications of the independent variable (the group’s gender composition): (1) as a

factor variable with each of the group compositions coded as dummy variables, (2) as a continuous

measure based on the count of women (i.e., from zero to seven), and (3) as a dummy variable to

indicate whether the group is majority-women or not. All specifications are included in SI E.9

Unless otherwise noted, in the main text, we present graphical depictions of our results based on

models that treat our independent variable as a factor variable with a dummy variable for each

treatment. This operationalization is the most informative graphically as it does not force linearity.

However, using specifications that we pre-registered, and based on the theorizing that we have used

throughout, we generally gauge statistical significance based on the continuous measure of gender

compositions; that is, whether an increase in the number of women is associated with more relative

influence.

Results

Gender differences in pre-treatment preferences

We first test whether and how men and women differ in their opinions about the most effective ways

to combat deforestation. Before group deliberations, we presented respondents with a list of policies

to curb over-harvesting in the nearby Zomba-Malosa Forest Reserve. Enumerators first showed each

respondent cards with an image depicting each policy (see Figure SI.1) as they explained the details

of each. Enumerators then asked each respondent to privately select the policy that they thought

would work best in their community to prevent over-harvesting by pointing to the associated card.

Figure 2 shows that men and women have the same ranking of policy responses. Both groups

tend to prefer policies offering remuneration or services (replanting incentives and job trainings)

over those aimed at altering behavior or stepping up enforcement actions. Still, we identify some

moderate gender differences (a chi-squared test is significant at p ≤ 0.10). Men are significantly

9SI E also includes a more in-depth write up of our results for each model specification.

14



0 10 20 30 40
%

Jobs Trainings

Alternative Cooking Methods

Civic Education

Replanting Incentives

Government Enforcement

Community Enforcement

Pre-Discussion Policy Preferences

Man
Woman

Figure 2 Gender differences in pre-treatment preferences on deforestation policy

more likely than women to prefer community enforcement policies and civic education programs

about the consequences of over-harvesting. Women, in contrast, are slightly more likely than men

to prefer government enforcement and replanting incentives, although these differences do not reach

traditional significance levels (see SI C). This suggests that some preferences around combating

deforestation are gendered, and that women’s inclusion might influence group decisions. However,

these preferences diverge less than we anticipated and we discuss what this might mean for the

scope conditions of our findings below.

We also underscore that even if men’s and women’s preferences differ only moderately, the fact

that we are considering an issue with six possible options—none of which received more than a

third of participants’ pre-treatment votes—makes the likelihood that any given woman and any

given man disagree is quite high. Based on Figure 2 (and also summarized in Table SI.2), we

calculate that the probability that any randomly-selected man’s preferred policy matches that of

any randomly-selected woman is 0.20; put another way, it is four times more likely that a given

man and a given woman disagree than it is that they agree.

15



0

.1

.2

.3

.4

Pr
ed

ic
te

d 
Va

lu
es

0W-6M 1W-5M 2W-4M 3W-3M 4W-2M 5W-1M 6W-0M
Gender Composition

Man
Woman

Self Most Influential (R)

Figure 3 Average likelihood that men and women participants rate themselves as the most influential
group member by treatment condition.

Measuring women’s influence

We measure women’s relative influence in three ways that we pre-specified. The first is based on

responses to a survey question asked after the group deliberations and secret vote. The question

read: “Which one person was the most influential in the group’s discussions and decisions?” We

analyze how well group gender composition predicts the likelihood that women and men respondents

rated themselves as the most influential person in their group. Self perceptions of influence may

accurately reflect group dynamics and/ or they may capture one’s self-confidence in his/ her ability

to sway decisions. The objective odds that the respondent was actually the most influential person

in their six-member group are 1 in 6, or 0.167, and this serves as a helpful benchmark for us to

assess whether men or women are more or less likely to vote for themselves than they would by

picking the “most influential” person at random. We plot these results in Figure 3. Black lines and

circles correspond to men participants, gray lines and squares indicate women participants, and we

visualize 95 percent confidence intervals around the estimated value of each treatment condition.

16



Consistent with our expectations, Figure 3 shows that women’s likelihood of rating themselves

as the most influential group member increases as the number of women in the group grows. This

increase is substantively meaningful. Women move from an average likelihood of just over 10 percent

in the one woman condition to a 20 percent likelihood of rating themselves as the most influential in

the all-women condition. A continuous measure of the number of women in the group is statistically

significant at the p ≤ 0.05 level for women group members (β = 0.015, p=0.035; estimates based

from Model 2 of Table SI.3).10 Men’s likelihood of rating themselves as most influential varies less

clearly with group gender composition, always exceeds that of women, and is always above the 16.7

percent benchmark of random chance.

We next turn to a measure of influence from our enumerator survey. Here, we use whether

the enumerator observing group dynamics selected a woman as the most influential member of the

group.11 To capture women’s relative influence within each treatment condition, we divide the

total number of votes women received by the total number of women in that study arm (across

groups). If there were no gender differences in influence as perceived by the enumerator, this

number would, as above, equal 0.167 (i.e., 1/6) for all treatment conditions. In contrast, as

Figure 4 shows, across all mixed-gender groups, any given woman’s likelihood of being rated by

the enumerator as the most influential in the group is always below 0.167. That is, men always

have relatively more influence than women in mixed-gender groups. Yet, as women’s representation

grows, enumerators are increasingly likely to rate any given woman member as influential. Again,

note that here we are normalizing by the number of women in each treatment condition, so this

is not a mechanical relationship: any given woman becomes relatively more influential as she has

more women around her. Table SI.4 (Model 2) shows the causal effect of a continuous measure of

the number of women in the group. Here we see a positive coefficient—having one more woman

increases any given woman’s likelihood of being rated by the enumerator as the most influential

by 1.2 percentage points—which is just outside the 10 percent significance threshold. However,

our alternative specification, whether the group is majority women (Model 3 in Table SI.4), does

10We calculate this estimate based on the margins command in Stata.
11The question on the enumerator survey reads: “Of the six participants, which one was the most influential?”

17



.05

.1

.15

.2

Pr
ed

ic
te

d 
Va

lu
es

1W-5M 2W-4M 3W-3M 4W-2M 5W-1M
Gender Composition

Women's Relative Influence in Discussion (E)

Figure 4 Women’s relative influence in group discussions from enumerator ranking. Calculated by
dividing the total number of votes that women received from the enumerator by the total number of
women in that study arm (across groups). If there were no gender bias, the ratio would be 0.167

show a statistically significant increase by 3.3 percentage points. Overall, we consider this evidence

weak but suggestive; relying on enumerators’ impressions of discussions, women’s relative influence

is either not worse or significantly better when there are more women in the group.

As a final measure of women’s influence in groups, we gauge how respondents rated their peers.

We consider this the most straightforward and objective measure of influence. For this, we again

use the question that asked participants to select in private the most influential group member

after the deliberations concluded. We measure relative influence at the group level by counting the

total number of votes that women members received and dividing that count by the total number

of women in the group. When this measure equals one, women have influence in the discussion in

proportion to their share of the group. Values less than one indicate that men are more influential

than women, and values greater than one indicate that women are more influential than men. Figure

5 reveals that women’s relative influence increases monotonically as the number of women in the

group increases, though men are always more influential than women. A linear measure of women’s

18



.2

.4

.6

.8

1

Pr
ed

ic
te

d 
Va

lu
es

1W-5M 2W-4M 3W-3M 4W-2M 5W-1M
Gender Composition

Women's Relative Influence in Discussion (Votes)

Figure 5 Women’s relative influence in group discussions from peer assessments. Calculated as the
sum of votes that women members received divided by the total number of women in the group. Values
less than one indicate that men are more influential than women.

representation is significant at the p ≤ 0.01 level (see Model 2 in Table SI.5). Again, this increase is

substantively meaningful. When a woman is alone in a group of men, her relative influence is about

half of what it should be if each group member picked the most influential member at random.

When women make up 5 of the 6 group members, they become closer to reaching parity, closing 90

percent of the gender gap in influence.

All three of our pre-specified influence measures—group members’ assessment of their own

influence, enumerators’ assessment of participants’ influence, and group members’ assessments of

their peers’ influence—point in the same direction and in combination offer strong support for our

core hypothesis: when women are surrounded by more women, the relative influence of any given

woman in group deliberations increases.

Yet perceptions of influence may not necessarily equate with actually changing the group’s

vote. Recall that after group deliberations, we asked each group member to vote in private on

their preferred policy solution to combat deforestation (repeating their pre-deliberation, private

19



.2

.3

.4

.5

Pr
ed

ic
te

d 
Va

lu
es

0W-6M 1W-5M 2W-4M 3W-3M 4W-2M 5W-1M 6W-0M
Gender Composition

Man
Woman

Prefered Policy (Pre-Treatment) Won in Group Vote

Figure 6 Gender differences in the likelihood of one’s pre-discussion policy preference winning the group
vote by gender and gender composition of group

vote). To analyze the extent to which women and men influence group votes across study arms, we

turn to each respondent’s pre-deliberation policy preference, which we compare with the group’s

post-deliberation vote. Figure 6 plots the predicted likelihood that a respondent’s pre-discussion

policy preference won in the post-discussion group vote by treatment condition for both men and

women participants. In men-majority groups, men and women are similarly likely to have their

(pre-treatment) preferred policy win in the group vote. In evenly split groups, men have a higher

likelihood of having their preferred policy selected. When women become majority group members,

they become more likely than men to have their preferred policy win the group vote. Regression

results presented in Table SI.6 (Model 3) reveal that the key interaction term of interest here, woman

member × majority-women group, is significant at the p ≤ 0.001 level.12

12A similar pattern emerges when we consider not whether a respondent’s pre-treatment policy preference won
the vote, but how many votes their pre-treatment policy preference received (i.e., the intensive margin of influence
as opposed to the extensive margin; see Figure SI.3 in SI G.)

20



Why are women more influential in the presence of other women?

Above, we theorized that there are three ways whereby women’s relative influence in group decisions

might increase with their share of the group: (1) women may speak more, (2) women may be

acknowledged for their contributions more often, and/ or (3) the substantive content of discussions

may move toward topics on which women have greater perceived authority. We now analyze

implications associated with each mechanism in turn.

Talk time

The first mechanism rests on the assumption that influence is positively correlated with talk time;

one cannot influence a group deliberation by remaining silent. Our data support this intuition.

From our transcript data, we count the number of words each respondent contributes to his/

her group’s deliberation and merge these data with our influence measures above. The number

of words spoken is indeed correlated with our influence measures; for instance, the number of

words spoken by a respondent is positively correlated with the enumerator picking the respondent

as the most influential person in the group (ρ = 0.42, p ≤ 0.001). Yet, while this correlation

is significant, it is substantively moderate, and appears to be smaller than that identified by

Karpowitz, Mendelberg and Shaker (2012) in their foundational study on gender and speech patterns

(Karpowitz, Mendelberg and Shaker 2012, 542). This suggests that while talk time does translate

to speaker influence, other factors might also affect who has influence in group deliberations in our

context.

Figure 7 shows the average number of words spoken by men and by women during group

discussions in each study arm.13 As above, we depict women’s average levels in gray, and men’s

average levels in black. We find little evidence that women participate more actively when they are

surrounded by more women. Women’s participation increases steadily from the one woman condition

to the three women condition, but then dips back down and remains low in the four, five, and six

13We are counting total contributions, including to both prompts that made up our focus group protocol (i.e., also
including responses to the prompt about climate change). We get similar outcomes when we just analyze responses
to the deforestation prompt.

21



100

200

300

400

500

Pr
ed

ic
te

d 
Va

lu
es

0W-6M 1W-5M 2W-4M 3W-3M 4W-2M 5W-1M 6W-0M
Gender Composition

Man
Woman

Number of Words Spoken in Discussion (T)

Figure 7 Gender differences in number of words spoken in discussion by gender composition. Women’s
average levels in orange, and men’s average levels in blue. Circle size indicate number of respondents per
treatment condition.

women conditions. Our regressions confirm a null finding for both genders; no measure of group

gender composition or its interactions with gender is a statistically significant predictor of respondent

talk time (see Table SI.7). Across study arms, men always speak more than women. These speech

patterns suggest that the core result that we find above—that women’s relative influence grows with

their presence—cannot be explained by women speaking more in settings with more women.

Recognition

Our second theorized mechanism is that women are increasingly acknowledged for their contributions

in groups with more women. This may occur because women more than men tend to recognize

women’s contributions across settings, or because men and/or women change their behavior in

groups with more women. This is particularly pertinent in rural Malawi, where it is customary

for women to defer to men in mixed group settings. To test for gender differences in assessments

of influence, we replicate our figure above that shows how group composition affects the likelihood

22



.2

.4

.6

.8

1

Pr
ed

ic
te

d 
Va

lu
es

1W-5M 2W-4M 3W-3M 4W-2M 5W-1M

Men
Women

Gender Composition

Women's Relative Influence in Discussion (Votes)

Figure 8 Women’s relative influence in group discussions by gender and group gender composition.
Influence is measured by the number of votes that men versus women received as the most influential
group member. Values greater than one indicate women have more influence, values less than one indicate
that men have more influence.

that a woman is voted by her peers as the most influential group member (Figure 5), but now we

assess this likelihood among men and women participants separately. Figure 8 shows these results.

As above, values can be interpreted as women’s influence relative to their representation: values

equal to one mean that women have influence in the discussion in proportion to their share of the

group, values less than one indicate that men are more influential than women, and values greater

than one indicate that women are more influential than men.

Figure 8 shows that women generally are more likely than men to see women as influential across

treatment conditions, especially when there are few women in the group. But, more strikingly,

Figure 8 reveals that group gender composition has a dramatic effect on men’s perceptions of

women’s influence: men steadily increase their perceptions of women’s influence in the presence of

more women. Thus, the patterns we observed in Figure 5 appear to be primarily driven by changes

in men’s behavior: men seem to take women’s contributions more seriously in groups with more

women. These findings are confirmed in our regressions. For men participants, Table SI.8 (Model 2)

23



reveals a positive and highly significant effect of being in a group with more women (p ≤ 0.01). For

women participants, the coefficient is positive, but not statistically significant (Table SI.9, Model

2).

Above, we theorized that one way in which group behavior might change in groups with more

women is through a general change in discussion tone in which group members are more likely

to acknowledge and engage with each other’s contributions. To test whether the tone of group

discussions changes across treatments, we code whether or not each of the approximately 20,000

statements that participants said across groups contains an expression of agreement.14 Statements of

agreement are fairly common in group discussions, representing about 10 percent of all participants’

statements. The following examples, taken verbatim from our transcripts, are representative.

I agree with what the brother and sister have said. . . People cut down trees, make
charcoal and start selling while police officers are just watching. (Woman, Chimbende
village, Zomba District)

To also add on what she has said, we should go for other alternative cooking methods.
Using chitetezo mbaula is the best option since it reduces demand for wood. (Woman,
Chagwira village, Zomba District)

We examine whether the proportion of statements conveying agreement increases in groups with

more women. Yet, as Figure 9 shows, we find no evidence that conversations generally become more

agreeable as women’s representation increases. In fact, regression results in Table SI.10 (Model 2)

show a small and significant negative effect.15 We also do not find evidence that relative references

specifically to women speakers (e.g., “I am in agreement with what auntie has just said.”) increase

in group settings with more woman (see SI H, Figure SI.4).

14We experimented with automating whether we could code expressions of agreement by training a model on
certain agreement phrases. However, this method produced very inaccurate results. So, instead, we hand-coded each
statement as to whether it contained an agreement expression or not. At the same time, we also coded instances of
interruptions and statements of disagreement. We did not find one instance of the former (interruptions) and very
few instances of the latter (explicit disagreement).

15We hope to explore this result in future work. Our initial speculation is that it is due to women’s lower levels
of political efficacy going into the group discussions. As we report above, women talk less than men across group
compositions, and examining the transcripts, it seems that women in general are less likely to engage in back-and-forth
dialogue than are men, likely because women are less accustomed to engaging in political debates.

24



.08

.1

.12

.14

.16

Pr
ed

ic
te

d 
Va

lu
es

0W-6M 1W-5M 2W-4M 3W-3M 4W-2M 5W-1M 6W-0M
Gender Composition

Share of Statements Expressing Agreement (T)

Figure 9 Percentage of agreement statements across treatment conditions. The denominator is the
total number of statements made across groups in each treatment condition.

In sum, although we find that women’s relative influence grows in group deliberations with

more women, and particularly in the eyes of men, we do not find any evidence that this can be

explained by increases in women’s talk time or by the increased recognition of women’s contributions

in settings with more women.

Discussion content

A final dynamic that we theorized might change women’s relative influence in group discussions

relates to possible variation in the substance of discussion topics across group compositions. It is

possible that groups with more women tend to focus on different dimensions of deforestation policy

than groups with more men, and moreover that the gendered nature of the ensuing discussions

lends greater authority to one gender or the other. To test these expectations, we first run a series

of structural topic models (STMs) on the total corpus of statements made by participants across

groups when discussing the prompt of deforestation policy.16 STMs involve a semi-automated form

16Here we examine only the discussions specifically following the prompt asking respondents to debate deforestation
policy, not on the first prompt about climate change. This make the total corpus of text ≈16,500 rather than ≈20,000.

25



of text analysis that enables researchers to inductively discover key topics within open-ended texts

(Roberts et al. 2014, 1066).

As is standard, we leverage the same information that we use for our model specifications

to structure the number and content of topics, namely respondent gender, treatment group, and

facilitator and village fixed effects. In our data, STM diagnostics suggest that responses maximize

semantic coherence when they are grouped into seven topics. We label each of the seven topics based

on the model-generated key words and the representative responses associated with each topic. The

seven topics that emerge are: replanting incentives, community enforcement, income generation,

forest guards, cooking methods, charcoal markets, and job training. Several of the topics seem to

align with the policy options that we presented to groups, which gives us confidence that the STMs

are picking up plausible groupings of topics relevant to the deliberations.

Figure 10 (left panel) shows the words or stems most associated with each topic (e.g., “busi”

is the stem of business and businessman and “enforc” is the stem of enforcing and enforcement)

and the prevalence of each topic in our dataset (as indicated by expected topic proportions on the

x-axis). This graphic reveals that the two most frequently mentioned topics during discussions were

replanting incentives and community enforcement. Figure 10 (right panel) also shows the marginal

effect of respondent gender on the frequency with which each topic is mentioned from models that

also include village and facilitator fixed effects.

Figure 10 (right panel) reveals that all seven topics have significant gender differences; that

is, that one gender is more likely than the other to contribute to the groups’ discussions on that

particular theme. Two topics are more frequently mentioned by women than by men. The first

is replanting incentives, which includes the words or stems: “tree,” “plant,” “care,” “take,” and

“pay.” The sentiments associated with this topic tend to express support for replanting initiatives

(i.e., paying villagers to plant new trees) and/ or are about the general need to take care of seedlings

and new growth. For example, model diagnostics suggest that the most representative statement

on this topic is the following:

The advantage of planting trees and caring for trees [is that] trees help to regain soil

26



0.0 0.2 0.4 0.6 0.8 1.0

Top Topics on Deforestation Policy

Expected Topic Proportions

Jobs: solut, find, train, teach, may

Char. mkts: govern, charcoal, enforc, think, stop

Cooking: use, cook, busi, firewood, method

Guards: good, protect, know, destroy, guard

Income: forest, cut, help, thing, way

Com. enforce: law, communiti, say, work, mountain

Replanting: tree, plant, care, take, pay

-0.04 -0.02 0.00 0.02 0.04

Marginal Effect of Gender on Topic Prevelance

Men ... Women

Jobs:

Char. mkts:

Income:

Cooking:

Replanting.:

Guards:

Com. enforce:

Figure 10 Left panel: Words and stems associated with the seven “topics” in participant contributions
to group discussions. Right panel: Marginal effect of gender on topic prevalence. Data are from the STM
analysis of 16,577 statements made discussing the prompt of policies to curb deforestation made by 3,749
study participants.

fertility and protect soil from erosion. . . People can be encouraged to plant and take care
of the trees but the the disadvantage comes when they are not paid and cannot take
good care of the trees. As a result, trees can wilt. (Woman, Kalanje village, Zomba
District)

The second topic more frequently brought up by women than by men is cooking, which includes

the words or stems: “use,” “cook,” “busi,” “firewood,” and “method.” Intuitively, the sentiments

associated with this topic all involve cooking methods. The following statement is representative:

When we use methods for cooking like chitetezo mbaula stove, this stove does not use
a lot of charcoal. There are other mbaula stoves that are molded, they do not use a lot
of firewood. (Woman, Singala village, Machinga District)

The remaining five topics are more frequently mentioned by men participants than by women.

Here we focus on the two with the largest gender differences favoring men, and contain a further

description of all topics in SI I. The topic with the largest gender gap in favor of men is community

enforcement, which contains the words or stems: “law,” “communiti,” “say,” “work,” and “mountain.”

Model diagnostics indicate that the following response is the most representative of this topic:

27



Community enforcement [means] creating by-laws that must be reinforced and followed.
We are the ones who are staying in the community and we are supposed to come up with
the rules because no one can come from outside our community and create the rules for
us. . . (Man, Kampaka village, Machinga District)

The second topic on which men tend to speak about more than women relates to forest guards

that are employed by the Malawian government to patrol Zomba-Malosa Forest Reserve. This topic

contains the words or stems: “good,” “protect,” “know,” “destroy,” and “guard.” These statements

tend to contain descriptions of the responsibilities of forest guards, and generally pertain to the

policy choice of government enforcement. The following statement is representative:

The government indeed needs to employ guards in this country so that our forest should
not be destroyed. It is good [and] a very wise decision to do that so our forests are
protected. . . (Man, Musowa village, Zomba District)

While we did not pre-specify any expectations about how men and women might differ in the

ways in which they tend to think about the problem of deforestation and its solutions, over the

course of this study, we accumulated a large amount of qualitative evidence that the differences we

find in Figure 10 map in predictable ways onto the gendered nature of harvesting and deforestation

policy in our study sites. Our qualitative data take the following forms. First, we return to the initial

eight focus groups that we did with men and women villagers in communities adjacent to our study

sites. We ran these focus groups to better inform the policy choices that we ultimately presented to

our study participants, but here we turn to a question that we asked specifically about the gendered

nature of forest management. Second, we conducted eighteen interviews with key stakeholders,

including Government Forestry Officers, local traditional authorities (chiefs), VNRMC and BMC

chairs, and local staff of NGOs working to curb over-harvesting. Local members of our research

team also attended, recorded, and transcribed the proceedings of VNRMC and BMC meetings in

ten villages near our study sites.

Our qualitative work supports our finding that replanting and cooking tend to be gendered

female. To begin, we found that existing replanting initiatives in our study areas primarily rely

on the work of women volunteers. The tree seedlings for these initiatives are typically funded by

28



government initiatives, donor groups (e.g., the EU), or non-profits working in this space (e.g., the

One Acre Fund). The labor of planting and tending to the new seedlings is typically done by

women. As one informant put it, “Women tend to spearhead voluntary efforts like tree planting.”17

Another informant said plainly: “Planting is seen as women’s work.”18 Cooking, the second topic

more frequently brought up by women, is also, as one would expect, a highly gendered activity

in Malawian village life (Mawaya and Kalindekafe 2010). Throughout our interviews, informants

mentioned that the primary use of forest resources by women is firewood for cooking. In our initial

scoping focus groups, cooking was a frequent topic discussed by women participants. For instance,

as one woman told us: “Women are the most affected [by over-harvesting] because for a house to

be called a home it needs a woman who cooks, and she requires firewood.”19

The topics that are more often discussed by men are also consistent with the gendered nature of

forest management. The two areas where we see the biggest gender gaps in favor of men are around

enforcement, both community enforcement and forest guards sent by the government of Malawi to

patrol the forest reserve. Again, our qualitative work is informative. From our initial scoping focus

groups, we find that men have more experience with community monitoring and enforcement, which

is typically done in conjunction with the local chief, who is also usually a man. As one man told us:

I am among the people that look after the forest and my colleague [another man] here too.
We help each other. . .We always respond if some people come and start cutting down
trees in the forests. . .We do it voluntarily, and we get nothing out of the initiative.20

Government-employed forest guards are also typically men, and this is perceived as potentially

dangerous work. As one woman in an initial focus group told us, “I think if we can have brave men

guarding the forest, [enforcement] can work.”21

Our exercise above was inductive; we sought to identify whether men and women make different

substantive contributions to group discussions and why this might be. Having done this, we now

17Interview with Jephthah Maliro, Agribusiness Officer & Project Manager of GIZ, interview by IPOR staff, July
29, 2023.

18Interview with Mr. Henry Utila; Forestry Research Institute of Malawi; July 10, 2022.
19Scoping focus group 2, T/A Malemia, Zomba District, Malawi.
20Scoping focus group 1, T/A Malemia, Zomba District, Malawi.
21Scoping focus group 2, T/A Malemia, Zomba District, Malawi.

29



0 1 2 3 4 5 6

0.
14

0.
17

0.
20

Replanting

Treatment (Number of Women)

E
xp

ec
te

d 
To

pi
c 

P
ro

po
rti

on
s

0 1 2 3 4 5 6

0.
10

0.
12

0.
14

0.
16

Cooking

Treatment (Number of Women)

E
xp

ec
te

d 
To

pi
c 

P
ro

po
rti

on
s

0 1 2 3 4 5 6

0.
13
0

0.
14
5

Guards

Treatment (Number of Women)

E
xp

ec
te

d 
To

pi
c 

P
ro

po
rti

on
s

0 1 2 3 4 5 6
0.
18

0.
20

0.
22

Community Enforcement

Treatment (Number of Women)

E
xp

ec
te

d 
To

pi
c 

P
ro

po
rti

on
s

Figure 11 Frequency of gendered topics across treatment conditions. X-axis indicates the number of
women in the group.

seek to identify how the prevalence of each of the four topics that we identified as most significantly

gendered varies across the treatment conditions. Figure 11 shows these results in models that, as

above, also control for village and facilitator fixed effects.

Consistent with our expectations, the two topics that are gendered female, cooking and replanting,

increase in frequency with more women in the group, while the two topics that are most gendered

male, community enforcement and forest guards, decrease in frequency. The treatment variable

is a significant predictor in all four models at the p ≤ 0.01 level (see Table SI.17).22 In sum, the

substance of group conversations are changing in groups with more women; they tend towards topics

that are gendered male in groups with more men and towards topics that are gendered female in

groups with more women. This evidence is thus most consistent in explaining our core finding

above: women likely have more authority in groups with more women (and particularly in the eyes

22Groups with more women are also more likely to select replanting incentives and alternative cooking methods in
their post-deliberation vote and less likely to select community enforcement and government enforcement. However,
only one of these (community enforcement) is significant at traditional significance levels (p ≤ 0.10), see SI F.

30



of men) because these groups tend to discuss issues on which women have more socially-recognized

expertise.

Conclusion

We find that women have more influence in deliberations on combating deforestation as their

representation grows. Across influence measures, women’s relative influence tends to increase with

their share of the group, and this pattern is strongest when we measure influence by peer assessments.

Moreover, the change in women’s perceived authority seems particularly driven by changes in men’s

assessments. Contrary to our expectations and to previous work from the United States, we do

not find that women participate more actively in discussions with more women, nor do we find

evidence that the tone of group discussions becomes more collaborative. Rather, our results seem

best explained by the changing substance of the deliberations themselves. Groups with more women

tend to talk about different dimensions of deforestation policy compared to groups with more men.

Moreover, our qualitative research suggests that this change in discussion content may affect how

other group members perceive women’s authority. Women’s relative influence grows in discussions

that spend more substantive time on dimensions of deforestation policy that are considered women’s

work, and decreases in discussions that focus on men’s work.

Our findings speak to a growing body of experimental research that is interested in how a

group’s gender composition causally affects participants’ behavior (Karpowitz and Mendelberg 2014;

Karpowitz et al. 2023). Like this literature, we find that gender composition powerfully predicts

who has influence in group deliberations. Yet, we also find that even if women are not speaking

at greater length, the substantive content of deliberations can change who is perceived to have

influence. Our results are from small group discussions, but it seems plausible that our findings

might also hold in larger political arenas. For instance, previous work has found that men and

women members of parliament bring up different issues in legislative debates (Bäck, Debus and

Müller 2014; Clayton, Josefsson and Wang 2016). It is impossible to know the counterfactual in

31



these settings, but it seems plausible that legislative bodies with more women both discuss different

types of topics and that women’s perceived authority grows as a result.

Our work also suggests some important scope conditions. While we find that men and women

tend to discuss deforestation differently, we also find only moderate differences in men’s and women’s

actual pre-treatment preferences about how to best address the problem. Our case seems to be one

in which there is nearly universal agreement in the community about the extent of the problem,

and similar ideas among men and women about best practices to address it. Our results might be

different in cases where preferences between men and women are more divergent—for instance on

issues such as which public goods to prioritize (Gottlieb, Grossman and Robinson 2018) or about

challenging patriarchal practices, such as child marriage or land rights (see, e.g., Benstead, Muriaas

and Wang 2022; Muriaas et al. 2019). This represents an important extension of our work, and one

for which we can envision competing expectations. On the other hand, women may find that there

is more at stake on issues for which there are significant gender gaps in preferences, and thus make

an even greater effort to influence group deliberations, particularly when settings are in their favor

(i.e., in groups with more women). On the other hand, issue areas perceived as zero sum or those

that threaten gender hierarchies may be settings in which men feel more emboldened to preserve

their power and authority. Studies that seek to vary either the issue area or the stakes of the groups’

decisions are promising areas for future research.

Our results further speak to efforts by international donors to include gender mainstreaming in

climate interventions (and other issue areas) and suggest that such efforts can elevate women’s voices

in community-led development. However, at the same time, we find that women’s pre-treatment

preferences are only more likely to win in the group vote when women are in the majority. This

presents a potential challenge to policy interventions that often strive for women to comprise a

“critical mass” (often conceptualized as 30 percent) or at most gender parity, but seldom promote

women to majority status. Future work might seek to test interventions that grant women greater

decision-making authority even when they are in the minority, such as training programs for women

citizens, candidates, or politicians (see, e.g., Hyde et al. 2022).

32



Finally, across all study arms, we continue to find that men participate more actively and have

relatively more influence than women in group deliberations. For deliberative democracy to be fully

realized, these gaps must close. Put another way, as it stands, our results show that men have an

oversized influence on how to address an acute community problem with global significance even

when they are in the absolute minority. At the same time, in the short run, our results suggest that

institutional features that promote women’s inclusion, such as gender quotas, may give women more

say in deliberative processes than they otherwise would. Including women in climate interventions,

thus, promotes fairer deliberative processes.

References

Agarwal, Bina. 1992. “The gender and environment debate: Lessons from India.” Feminist studies
18(1):119–158.

Agarwal, Bina. 2009. “Gender and forest conservation: The impact of women’s participation in
community forest governance.” Ecological economics 68(11):2785–2799.

Agarwal, Bina. 2010. “Does women’s proportional strength affect their participation? Governing
local forests in South Asia.” World development 38(1):98–112.

Alkire, Sabina, Ruth Meinzen-Dick, Amber Peterman, Agnes Quisumbing, Greg Seymour and Ana
Vaz. 2013. “The women’s empowerment in agriculture index.” World development 52:71–91.

Bäck, Hanna, Marc Debus and Jochen Müller. 2014. “Who takes the parliamentary floor? The role
of gender in speech-making in the Swedish Riksdag.” Political Research Quarterly 67(3):504–518.

Barnes, Tiffany D. 2016. Gendering Legislative Behavior. Cambridge University Press.

Benstead, Lindsay J, Ragnhild L Muriaas and Vibeke Wang. 2022. “Explaining Backlash: Social
Hierarchy and Men’s Rejection of Women’s Rights Reforms.” Social Politics: International Studies
in Gender, State & Society .

Born, Andreas, Eva Ranehill and Anna Sandberg. 2018. “A man’s world: The impact of a male
dominated environment on female leadership.” Working Paper .

Bratton, Kathleen A and Leonard P Ray. 2002. “Descriptive Representation, Policy Outcomes, and
Municipal Day-Care Coverage in Norway.” American Journal of Political Science pp. 428–437.

Brulé, Rachel. 2022. “Climate Shocks and Gendered Political Transformation: How Crises Alter
Women’s Political Representation.” Politics & Gender pp. 1–7.

Bush, Sarah Sunn and Amanda Clayton. 2022. “Facing change: Gender and climate change attitudes
worldwide.” American Political Science Review pp. 1–18.

33



Chattopadhyay, Raghabendra and Esther Duflo. 2004. “Women as Policy Makers: Evidence from
a Randomized Policy Experiment in India.” Econometrica 72(5):1409–1443.

Clayton, Amanda, Cecilia Josefsson and Vibeke Wang. 2014. “Present without presence? Gender,
quotas and debate recognition in the Ugandan parliament.” Representation 50(3):379–392.

Clayton, Amanda, Cecilia Josefsson and Vibeke Wang. 2016. “Quotas and Women’s Substantive
Representation: Evidence from a Content Analysis of Ugandan Plenary Debates.” Politics &
Gender pp. 1–29.

Clayton, Amanda and Pär Zetterberg. 2018. “Quota Shocks: Electoral Gender Quotas and
Government Spending Priorities Worldwide.” The Journal of Politics 80(3):916–932.

Cook, Nathan J, Tara Grillos and Krister P Andersson. 2019. “Gender quotas increase the equality
and effectiveness of climate policy interventions.” Nature Climate Change 9(4):330–334.

Deininger, Franziska, Andrea Woodhouse, Anne T. Kuriakose, Ana Gren and Sundas Liaqat.
2023. “Placing Gender Equality at the Center of Climate Action: A Review of Programmatic
Experience.” World Bank Group Gender Thematic Policy Notes Series .

Eisenbarth, Sabrina, Louis Graham and Anouk S Rigterink. 2021. “Can community monitoring save
the commons? Evidence on forest use and displacement.” Proceedings of the National Academy
of Sciences 118(29):e2015172118.

Franceschet, Susan and Jennifer M Piscopo. 2008. “Gender Quotas and Women’s Substantive
Representation: Lessons from Argentina.” Politics & Gender 4(03):393–425.

Global Forest Watch. 2021. https://www.globalforestwatch.org/dashboards/country/MWI/.
Accessed: 2023-01-18.

Gottlieb, Jessica, Guy Grossman and Amanda Lea Robinson. 2018. “Do men and women have
different policy preferences in Africa? Determinants and implications of gender gaps in policy
prioritization.” British Journal of Political Science 48(3):611–636.

Grey, Sandra. 2002. “Does Size Matter? Critical Mass and New Zealands Women MPs.”
Parliamentary Affairs 55(1):19–19.

Gulzar, Saad, Apoorva Lal and Benjamin Pasquale. 2023. “Representation and Forest Conservation:
Evidence from Indiaâs Scheduled Areas.” American Political Science Review p. Online First.

Holman, Mirya R and Anna Mahoney. 2018. “Stop, collaborate, and listen: Women’s collaboration
in US state legislatures.” Legislative Studies Quarterly 43(2):179–206.

Holmes, Janet. 2013. Women, men and politeness. Routledge.

Hyde, S, E Malesky, A Coppock, M Poertner, L Young, C Adida, L Arriola, A Cheema, A Chiweza,
A Clayton et al. 2022. “Meta-Preanalysis Plan for Metaketa V: Women’s Action Committees and
Local Services.”.

Karpowitz, Christopher F and Tali Mendelberg. 2014. The silent sex. Princeton University Press.

34

https://www.globalforestwatch.org/dashboards/country/MWI/


Karpowitz, Christopher F, Tali Mendelberg and Lee Shaker. 2012. “Gender inequality in deliberative
participation.” American Political Science Review 106(3):533–547.

Karpowitz, Christopher, Stephen D O’Connell, Jessica Preece and Olga Stoddard. 2023. “Strength
in Numbers? Gender Composition, Leadership, and Women’s Influence in Teams.”.

Kindermann, Georg, Michael Obersteiner, Brent Sohngen, Jayant Sathaye, Kenneth Andrasko,
Ewald Rametsteiner, Bernhard Schlamadinger, Sven Wunder and Robert Beach. 2008. “Global
cost estimates of reducing carbon emissions through avoided deforestation.” Proceedings of the
national Academy of Sciences 105(30):10302–10307.

Kittilson, Miki Caul. 2006. Challenging Parties, Changing Parliaments: Women and Elected Office
in Contemporary Western Europe. Ohio State University Press.

Kumar, Neha and Agnes Quisumbing. 2014. Gender, shocks, and resilience. Vol. 11 International
Food Policy Research Institute.

Mai, Yen Hoang, Esther Mwangi and Melinda Wan. 2011. “Gender analysis in forestry research:
looking back and thinking ahead.” International Forestry Review 13(2):245–258.

Mawaya, Chimwemwe and Meya P Kalindekafe. 2010. “Access, control and use of natural resources,
in southern Malawi: A gender analysis.” Natural Resource Management: The Impact of Gender
and Social Issues pp. 88–125.

Meinzen-Dick, Ruth, Thomas Falk, Katrina Kosec, Pratiti Priyadarshini, Richu Sanil, Mequanint
Melesse, Wei Zhang, Lucia Carrillo and Hagar ElDidi. 2022. “The Impact of Experiential Learning
Interventions on Sustainable Water Management: Evidence from a Randomized Control Trial in
India.” EGAP Registry .

Moyo, Gift Gladson, Marlene Chikuni and Sosten Chiotha. 2016. “Community Knowledge in
Restoring Trees in Controversial Forest Hot Spots: Case of Nkanya-Lusewa rivers catchment
area in Zomba-Malosa Forest Reserve, Malawi.” Open Science Journal 1(3).

Mudege, Netsayi N, Norita Mdege, Putri E Abidin and Sandra Bhatasara. 2017. “The role of gender
norms in access to agricultural training in Chikwawa and Phalombe, Malawi.” Gender, Place &
Culture 24(12):1689–1710.

Muriaas, Ragnhild L, Vibeke Wang, Lindsay Benstead, Boniface Dulani and Lise Rakner. 2019.
“Why the gender of traditional authorities matters: Intersectionality and womenâs rights
advocacy in Malawi.” Comparative Political Studies 52(12):1881–1924.

Mwangi, Esther, Ruth Meinzen-Dick and Yan Sun. 2011. “Gender and sustainable forest
management in East Africa and Latin America.” Ecology and society 16(1).

Ngwira, Susan and Teiji Watanabe. 2019. “An analysis of the causes of deforestation in Malawi: a
case of Mwazisi.” Land 8(3):48.

Ostrom, Elinor. 1990. Governing the commons: The evolution of institutions for collective action.
Cambridge university press.

35



Parthasarathy, Ramya, Vijayendra Rao and Nethra Palaniswamy. 2019. “Deliberative democracy
in an unequal world: A text-as-data study of south India’s village assemblies.” American Political
Science Review 113(3):623–640.

Roberts, Margaret E, Brandon M Stewart, Dustin Tingley, Christopher Lucas, Jetson Leder-Luis,
Shana Kushner Gadarian, Bethany Albertson and David G Rand. 2014. “Structural Topic Models
for Open-Ended Survey Responses.” American Journal of Political Science 58(4):1064–1082.

Samii, Cyrus, Matthew Lisiecki, Parashar Kulkarni, Laura Paler, Larry Chavis, Birte Snilstveit,
Martina Vojtkova and Emma Gallagher. 2014. “Effects of payment for environmental services
(PES) on deforestation and poverty in low and middle income countries: a systematic review.”
Campbell Systematic Reviews 10(1):1–95.

Slough, Tara, Daniel Rubenson, Roâee Levy, Francisco Alpizar Rodriguez, Maŕıa Bernedo del
Carpio, Mark T Buntaine, Darin Christensen, Alicia Cooperman, Sabrina Eisenbarth,
Paul J Ferraro et al. 2021. “Adoption of community monitoring improves common pool
resource management across contexts.” Proceedings of the National Academy of Sciences
118(29):e2015367118.

UN-REDD Programme. 2017. “Tenure and REDD+ in Malawi.”.

Villamor, Grace B, Meine van Noordwijk, Utkur Djanibekov, Ma Elena Chiong-Javier and Delia
Catacutan. 2014. “Gender differences in land-use decisions: shaping multifunctional landscapes?”
Current Opinion in Environmental Sustainability 6:128–133.

Weeks, Ana Catalano. 2022. Making Gender Salient: From Gender Quota Laws to Policy.
Cambridge University Press.

36



Supplementary Information for:

“Gender, Deliberation, and Natural Resource Governance:

Experimental Evidence from Malawi”

A Policy cards 2

B Pre-treatment political attitudes 3

C Pre-treatment policy preferences 4

D Post-treatment Preferences 5

E Main results: regression tables 6

F Policy choices by treatment condition 15

G Group decision-making: alternative specification 21

H Deliberation tone: alternative specification. 22

I STM Tables and Plots Across Topics 23

J Human Subjects Research 26

K Pre-analysis plan 28

1



A Policy cards

Figure SI.1 Cards used in the survey and discussion to represent the six policies

 

 
Boma kukhazikisa 

Malamulo oti Azisatidwa 
 

 
 
 

 
Dela Kukhazikitsa 

Malamulo oti 
Azisatidwa 

 
 

 

 

 
Kulipira Anthu 
Ozala/Kusamala Mitengo 

 

 
 
 

 
Kuphunzitsa Njira zina 

Zopezela Ndalama 
 

 
 
 

 
Kuwaphunzitsa Anthu 

Zokhuzana ndi Mitengo 
 

 
 

 
Kugwiritsa Ntchito Njira 

zina zophikira 
 

 

2



B Pre-treatment political attitudes

Table SI.1 Pre-treatment summary statistics by gender

Variable Men Women Difference
Age 34.833 36.794 1.961***

(15.204) (13.608) (0.472)
Discusses Politics 0.796 0.711 -0.084***

(0.690) (0.677) (0.022)
Interested in Politics 2.334 2.226 -0.108***

(1.109) (1.071) (0.036)
Attended Community Meeting 0.796 0.846 0.051***

(0.403) (0.361) (0.012)
Raised an Issue with Others 0.753 0.745 -0.008

(0.431) (0.436) (0.014)
Voted in Last Election 0.643 0.706 0.064***

(0.479) (0.456) (0.015)
Contacted Local Councilor 0.158 0.101 -0.057***

(0.365) (0.301) (0.011)
Contacted Member of Parliament 0.165 0.112 -0.053***

(0.371) (0.315) (0.011)
Contacted Traditional Authority 0.573 0.527 -0.046***

(0.495) (0.499) (0.016)
Politics Too Complicated 3.389 3.339 -0.050

(0.930) (0.929) (0.030)
People Like Me Can Participate 1.864 1.764 -0.100***

(1.059) (1.036) (0.034)
Political Ability 2.390 2.101 -0.289***

(1.232) (1.183) (0.039)
Political Confidence 2.619 2.218 -0.401***

(1.218) (1.167) (0.039)
Observations 1,886 1,882

Notes: Age is a continuous varaible, while five variables range from 0 to 5: Interested in Politics,
Politics Too Complicated, People Like Me Can Participate, Political Ability, and Political
Confidence. Remaining variables are binary.

3



C Pre-treatment policy preferences

Table SI.2 Pre-treatment policy preference by gender

Variable Men Women Difference (p-value)
Community Enforcement 0.083 0.066 -0.017 (0.047)
Government Enforcement 0.191 0.206 0.014 (0.264)
Replanting Incentives 0.302 0.316 0.014 (0.353)
Civic Education 0.082 0.066 -0.016 (0.062)
Alternative Cooking Methods 0.114 0.118 0.003 (0.755)
Job Training 0.227 0.226 0.001 (0.934)
Observations 1,886 1,882

Notes: Individuals had to choose exactly one preferred option prior to deliberation; here we
display means by gender.

4



D Post-treatment Preferences

Figure SI.2 Gender differences in post-treatment preferences on deforestation policy

0 10 20 30 40
%

Jobs Trainings

Alternative Cooking Methods

Civic Education

Replanting Incentives

Government Enforcement

Community Enforcement

Post-Discussion Policy Preferences

Man
Woman

Notes: Individuals had to choose exactly one preferred option following deliberation; here we
display means by gender.

5



E Main results: regression tables

Table SI.3 Likelihood of rating oneself the most influential group member, by group
composition and gender

Self Most Influential (R)

(1) (2) (3)

Woman −0.142∗∗∗ −0.096∗∗∗ −0.081∗∗∗

(0.039) (0.029) (0.021)
1W-5M 0.029

(0.022)
2W-4M 0.047∗∗

(0.024)
3W-3M 0.047∗

(0.028)
4W-2M 0.108∗∗∗

(0.033)
5W-1M 0.002

(0.044)
6W-0M 0.124∗∗∗

(0.043)
Woman × 2W-4M 0.073

(0.059)
Woman × 3W-3M 0.036

(0.054)
Woman × 4W-2M −0.024

(0.056)
Woman × 5W-1M 0.120∗∗

(0.061)
Number of Women 0.014∗∗

(0.006)
Woman × Number of Women −0.003

(0.008)
Majority Women 0.046∗

(0.026)
Woman × Majority Women −0.021

(0.035)
Constant 0.177∗∗ 0.185∗∗∗ 0.204∗∗∗

(0.069) (0.068) (0.071)

Enumerator FEs Yes Yes Yes
Village FEs Yes Yes Yes

Observations 3768 3768 3768
Adjusted R2 0.030 0.030 0.029

Notes: The dependent variable is an indicator for the respondent (R) rating themself as the most influential
individual in their group following deliberation. For estimates of the interactions between gender and group
composition, the 1W-5M condition serves as the omitted category. Standard errors clustered at the village level are
reported in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

6



Table SI.4 Relative likelihood of enumerator selecting a woman as the most influential
group member, by group composition

Women’s Relative Influence in Discussion (E)

(1) (2) (3)

2W-4M 0.019
(0.051)

3W-3M 0.019
(0.043)

4W-2M 0.038
(0.039)

5W-1M 0.052
(0.036)

Number of Women 0.012
(0.007)

Majority Women 0.033∗

(0.017)
Constant 0.206∗∗∗ 0.195∗∗∗ 0.219∗∗∗

(0.075) (0.066) (0.059)

Facilitator FEs Yes Yes Yes
Village FEs Yes Yes Yes

Observations 433 433 433
Adjusted R2 0.015 0.023 0.022

Notes: The dependent variable is an indicator for the enumerator selecting a woman as the most
influential group member divided by the number of women in the group; this captures a measure of
women’s perceived (by the enumerator) influence that removes the mechanical increase in women’s
likelihood of being the most influential member that comes with having more women in the group. This
analysis drops single-gender groups, as we cannot divide by zero to construct this variable (for groups
with 0 women), and cannot talk about women’s relative influence in groups with 0 men. Standard errors
clustered at the village level are reported in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

7



Table SI.5 Number of votes women received divided by number of women in the group, by
group composition.

Women’s Relative Influence in Discussion (Votes)

(1) (2) (3)

2W-4M 0.128
(0.172)

3W-3M 0.230
(0.162)

4W-2M 0.330∗∗

(0.150)
5W-1M 0.481∗∗∗

(0.127)
Number of Women 0.116∗∗∗

(0.029)
Majority Women 0.286∗∗∗

(0.076)
Constant 1.244∗∗∗ 1.132∗∗∗ 1.365∗∗∗

(0.207) (0.202) (0.203)

Facilitator FEs Yes Yes Yes
Village FEs Yes Yes Yes

Observations 433 433 433
Adjusted R2 0.050 0.059 0.043

Notes: The dependent variable is the number of votes for a woman as the most influential member in the
group divided by the number of women in the group; this captures a measure of women’s perceived (by
peers) influence that removes the mechanical increase in women’s likelihood of being the most influential
member that comes with having more women in the group. This analysis drops single-gender groups, as
we cannot divide by zero to construct this variable (for groups with 0 women), and cannot talk about
women’s relative influence in groups with 0 men. Standard errors clustered at the village level are
reported in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

8



Table SI.6 Likelihood that respondent’s pre-discussion policy preference won in the
post-discussion group vote, by group composition and gender

Prefered Policy (Pre-Treatment) Won in Group Vote

(1) (2) (3)

Woman 0.004 −0.037 −0.037
(0.055) (0.040) (0.024)

1W-5M −0.014
(0.035)

2W-4M −0.005
(0.036)

3W-3M 0.025
(0.039)

4W-2M −0.108∗∗

(0.042)
5W-1M −0.092∗

(0.053)
6W-0M −0.035

(0.064)
Woman × 2W-4M −0.016

(0.071)
Woman × 3W-3M −0.091

(0.068)
Woman × 4W-2M 0.089

(0.071)
Woman × 5W-1M 0.080

(0.077)
Number of Women −0.014∗

(0.008)
Woman × Number of Women 0.015

(0.011)
Majority Women −0.101∗∗∗

(0.031)
Woman × Majority Women 0.121∗∗∗

(0.039)
Constant 0.470∗∗∗ 0.483∗∗∗ 0.470∗∗∗

(0.145) (0.144) (0.142)

Enumerator FEs Yes Yes Yes
Village FEs Yes Yes Yes

Observations 3768 3768 3768
Adjusted R2 0.020 0.019 0.021

Notes: The dependent variable is an indicator for the respondent’s preferred policy pre-treatment being
that selected by the secret ballot vote held in the group following deliberation. For estimates of the
interactions between gender and group composition, the 1W-5M condition serves as the omitted category.
Standard errors clustered at the village level are reported in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗

p < 0.01.

9



Table SI.7 Number of words respondent contributes to their group’s deliberation, by group
composition and gender

Number of Words Spoken in Discussion (T)

(1) (2) (3)

Woman −125.603∗∗∗ −96.513∗∗∗ −97.227∗∗∗

(28.518) (30.994) (19.602)
1W-5M 4.329

(33.619)
2W-4M 30.789

(34.963)
3W-3M 45.595

(43.988)
4W-2M −25.985

(42.699)
5W-1M 26.903

(56.388)
6W-0M 25.007

(41.926)
Woman × 2W-4M 0.941

(42.347)
Woman × 3W-3M 19.308

(47.947)
Woman × 4W-2M 26.501

(38.311)
Woman × 5W-1M −19.968

(57.368)
Number of Women 3.804

(8.215)
Woman × Number of Women −6.961

(10.126)
Majority Women −24.331

(31.587)
Woman × Majority Women −7.041

(33.189)
Constant 45.094 78.118 79.817

(165.035) (169.904) (160.135)

Enumerator FEs Yes Yes Yes
Village FEs Yes Yes Yes

Observations 3515 3515 3515
Adjusted R2 0.118 0.117 0.118

Notes: The dependent variable is a continuous measure of the number of words spoken, coded from
analysis of deliberation transcripts. For estimates of the interactions between gender and group
composition, the 1W-5M condition serves as the omitted category. Standard errors clustered at the village
level are reported in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

10



Table SI.8 Number of votes women received from male participants divided by number of
women in the group, by group composition

Women’s Relative Influence in Discussion (Votes by Men)

(1) (2) (3)

2W-4M 0.029
(0.172)

3W-3M 0.228
(0.165)

4W-2M 0.260∗

(0.154)
5W-1M 0.541∗∗∗

(0.132)
Number of Women 0.168∗∗∗

(0.016)
Majority Women 0.437∗∗∗

(0.063)
Constant 1.062∗∗∗ 0.641∗∗∗ 0.955∗∗∗

(0.199) (0.179) (0.181)

Facilitator FEs Yes Yes Yes
Village FEs Yes Yes Yes

Observations 433 520 520
Adjusted R2 0.099 0.151 0.075

Notes: The dependent variable is the number of votes for a woman as the most influential member in the
group cast by male participants divided the number of women in the group; this captures a measure of
women’s perceived (by male peers) influence that removes the mechanical increase in women’s likelihood
of being the most influential member that comes with having more women in the group. This analysis
drops single-gender groups, as we cannot divide by zero to construct this variable (for groups with 0
women), and cannot talk about women’s relative influence in groups with 0 men. Standard errors
clustered at the village level are reported in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

11



Table SI.9 Number of votes women received from female participants divided by number
of women in the group, by group composition.

Women’s Relative Influence in Discussion (Votes by Women)

(1) (2) (3)

2W-4M 0.198
(0.282)

3W-3M 0.089
(0.247)

4W-2M 0.210
(0.231)

5W-1M 0.313
(0.217)

Number of Women 0.064
(0.045)

Majority Women 0.169
(0.103)

Constant 1.811∗∗∗ 1.788∗∗∗ 1.914∗∗∗

(0.446) (0.404) (0.365)

Facilitator FEs Yes Yes Yes
Village FEs Yes Yes Yes

Observations 433 433 433
Adjusted R2 0.002 0.008 0.007

Notes: The dependent variable is the number of votes for a woman as the most influential member in the
group cast by female participants divided the number of women in the group; this captures a measure of
women’s perceived (by female peers) influence that removes the mechanical increase in women’s
likelihood of being the most influential member that comes with having more women in the group. This
analysis drops single-gender groups, as we cannot divide by zero to construct this variable (for groups
with 0 women), and cannot talk about women’s relative influence in groups with 0 men. Standard errors
clustered at the v