Blog

Meal-Sharing and Social Network Formation

Thanks to the generous support of the Adrianne Hall Fund at the Women and Public Policy Program at Harvard Kennedy School

16 May, 2024 Blog Post 1: Study Design

In seventy-one percent of societies worldwide, women move to the area where her husband and his kin live at the time of marriage (Murdock 1967). I seek to evaluate how women’s ability to enter social networks after moving to a new village for marriage affects their mental health and integration into social and financial networks in rural Malawi villages. While a growing literature shows mental health disorders and loneliness are highly prevalent among low-income people, including within developing countries (Banerjee et al. 2023), there is limited evidence documenting causal pathways and mitigating strategies in this population. Furthermore, although evidence documents that social networks facilitate financial risk-sharing and information transmission (Breza et al. 2019), little is known about how social networks originate or evolve. My intervention will help young, low-income female marriage migrants to initiate meal-sharing with other women, a common way to engage socially. My study design will allow me to understand the causes and consequences of sparse and homophilic networks within a socially isolated and vulnerable population. I will explore the roles of social frictions and budget constraints as potential inhibitors of homophilic and non-homophilic network formation among young female marriage migrants. Finally, I will investigate how encouraging meal-sharing relationship affects social network formation and characteristics, psychological well-being, and socio-economic inclusion.

I am conducting a three-month randomized controlled trial in Malawi, which includes two intervention arms that each attempt to solve barriers to meal-sharing: a food price subsidy to solve budgetary concerns, and information about who else in the village would like to share meals to solve social concerns, such as fear that an invitation might be rejected. I will first recruit a sample of 2,600 women who moved to the village after the age of 14 but no more than 20 years ago (Baseline Sample). I will select a 1,600-person subset of the Baseline Sample for the intervention (Intervention Sample), selected if they have a straw roof (proxy for low-income). I will offer Baseline Sample participants who are not selected for the Intervention Sample the opportunity to add their name to a list of women who are interested in sharing meals with other women in the village more often (Sharing List). The rest of the study will be conducted with the Intervention Sample, who will be randomized into three groups:

1. Sharing Information (n=600) – This group will view ten names from the Sharing List and can invite up to five women on the list to share a meal (invitation delivered by enumerators).

2. Vouchers + Sharing Information (n=600) – Recipients will receive 7 USD/month (8% median monthly income) vouchers, redeemable at several local vendors selling perishable produce, and the Sharing Information intervention.

3. Control (n=400) – This group will not receive any intervention.

Within treatments 1 and 2, participants view one of three types of Sharing Lists: all low-SES (homophilic), all high-SES (non-homophilic), or a random selection of women. While cross-social-class friendships are highly correlated with economic mobility in the United States (Chetty et al. 2022), economists know little about the causal impacts of these relationships, catalysts to forming these relationships, and the nature of these relationships in developing countries. This second-stage randomization allows me to compare the consequences of cross-SES interactions relative to within-SES, and the ability higher-quality foods (Vouchers) to facilitate these relationships.

Furthermore, this design allows me to test if homophily is caused by preferences or constraints. Homophily – or like-seeking-like – is, in my setting, equivalent to within-SES relationship. If individuals have a strict preference for spending time with people within their own socio-economic status, then the opportunity to invite high-SES individuals in addition to the opportunity to invite low-SES individuals should not change people’s preferences relative to the condition when women only have the opportunity to invite low-SES women. In other words, invitation choices in the random-selection-list condition should look no different from choices in the low-SES-list condition. Conversely, if individuals have a strict preference for cross-SES relationships, then they find themselves in homophilic relationships because they are constrained from entering cross-SES relationships. If the invitation lits and the voucher are sufficient to overcome these constraints, then invitation choices in the random-selection-list condition should look no different from choices in the high-SES-list condition (in the Voucher group).

1 July, 2024 Blog Post 2: Sample of female marriage migrants

Sample

My full sample consists of 3600 women from Mchinji district in Malawi. Women were selected if they were at least 18 years old, had moved to the village at no younger than 14 years old, and had lived in the village for no more than 20 years. Eighty-five percent of women moved to the village for the purpose of getting married, and the remainder mostly moved to live with extended family. Ninety-nine percent of women who moved to the village for marriage were still married, whereas only half of the women who moved to be with family were married (the remainder were divorced, separated, or widowed – only seven people in the entire sample were never-married). While this indicates that some women move away from their husband’s village after divorce or separation, the vast majority of female rural-to-rural migrants are married women who came to the village for marriage.

The median age was 27 years old (mean: 29), the median age of women when they moved to the village was 20 (mean: 22), and the median number of years in the village was 5 (mean: 7). The majority of these women are in their child-bearing years. Thirty-eight percent of the sample was pregnant or breast-feeding at the time of being surveyed, 13% of the sample had a child in the past 2 years, 24% had a child in the past 3-5 years, and the remaining 21% had gone more than five years without having a child.

5 July, 2024 Blog Post 3: Loneliness among female marriage migrants

Measuring loneliness

I used the UCLA-3 scale to assess loneliness. This scale consists of three questions:

1. How often do you feel that you lack companionship?
2. How often do you feel left out?
3. How often do you feel isolated from others?

Respondents select “Hardly ever” (1 point), “Some of the time” (2 points), or “Often” (3 points) for each question. Someone is considered “lonely” if their total score is 6 or more points. Eighteen percent of the sample was lonely.

Predictors of loneliness

I assess three predictors of loneliness: women’s relationships with their husbands, women’s relationships with other people, and the difficulty women face in returning to their home villages.

1. Women’s relationships with their husbands: Women report having very strong relationships with their husbands in terms of their ability to trust them with their thoughts and feelings, and their ability to laugh with them. However, only 16% of women unprompted mention their husbands as somebody who they engage with in any activities, and only 5% of women unprompted offer their husband as someone who they trust to tell as a secret. In an index of a variable on the relationship with their husbands, a one-standard deviation increase in the strength of the relationship (i.e. more trusting) was associated with a 1.6 percentage point increase in loneliness (12% increase over the mean; p-value: 0.054). It is unclear why stronger marital relationships are associated with more loneliness.

2. Women’s relationships with other people: Women list on average 6 people in their social network. Each additional person in the social network is associated with a 2 percentage point decrease in loneliness (15% decrease over the mean; p-value: <0.001). The average strength of the network is not associated with loneliness. The strength of each network connection was assessed in the same why that the relationship with the husband was assessed, and then the average was drawn from all the network connections for each woman without weighting by the number of network connections. However, the interaction between the size of the network and the average strength of the network was highly significant, statisically and in magnitude, for decreasing loneliness. This indicates that people with small, strong networks, are still quite lonely. People with larger, weaker networks are on average somewhat less lonely, but people with larger, stronger networks are on average much less lonely.

3. Difficulty of returning home: I asked women three questions to determine how difficult it is for them to return to their home village: how long does the trip take them, how much does the trip cost them, and how many times do they return per year. On average, it takes women 2.3 hours to return home (the most common mode of transport, used by 50% of the sample, is a bicycle taxi), and the trip costs women on average 3,800 MWK (2.38 USD; the median cost is 2,000 MWK, approximately 1.25 USD). Given these constraints, women visit home on average 8 times per year, though the median woman visits far fewer: 4 times per year. I use these three variables to construct an index of difficulty returning home. A one standard deviation increase in the difficulty of returning home is associated with a 3.5 percentage point (20%) increase in loneliness. Only 8% of women in the lowest decile of difficulty returning home lonely, whereas 24% of women in the highest decile of returning home are lonely. This relationship is blunted somewhat by larger networks. While the positive correlation still exists among people with larger networks, there is a levels shift downward in loneliness among people with larger networks. While it is unfortunate that there is no discernible antidote for loneliness associated with homesickness, it is encouraging the larger networks are associated with a reduction in loneliness even among the people who are very far from home.

15 July, 2024 Blog Post 4: Strength of network connections

In this post, I discuss (1) how I measure the strength of ties, (2) which relationship characteristics and individual characteristics are associated with stronger ties.

Measuring Strong Ties

I measure the strength of ties using five questions that women answer about their relationship with a specific individual: the Inclusion of Self in Others score, ability to trust that person with thoughts and feelings, how easy it is to laugh with that person, whether or not you consider that person to be someone you tell everything to, whether or not you believe that person tells everything to you, and whether or not you can trust that person with a secret. I conduct principal components analysis (PCA) on these variables and use the first component as my measure of strong ties.

The Inclusion of Self in Others (ISO) score is a psychology instrument to determine closeness of relationships. Respondents see two circles, where one circle represents oneself, and the other circle represents the other individual. There are seven images, where in each image the circles between increasingly more overlapping. When the circles are completely non-overlapping, this represents a relationship where the two individuals feels like completely separate entities; and when the circles are almost entirely overlapping, this represents a relatinoship where the other person is an important part of one’s identity, or highly “included in the self”. Interestingly, women responded very differently to this question depending on whether the other person they were answer the question about was their husband or someone else. The level of trust and ability to laugh with their husbands was more or less uncorrelated with the ISO score they gave for their husband (if anything, it was negatively correlated). Conversely, levels of trust, ability to laugh, and the ISO score were all positively correlated with people who were not the husband. Because of these differences in how women responded I calculated the index of strong ties with friends separately from the index of strong ties with husbands.

What Makes a Strong Tie

I use lasso to determine what makes a strong tie. Lasso is an algorithm that will select variables that are most important for explaining variation in a variable of interest.

Husbands: Lasso only selects two variables to explain the strength of ties with husbands. The first is if the respondent gave her husband’s name when describing the people with whom she does a variety of activities (positively correlated with strength of the relationship). I asked women to report the names of people with whom she shares meals, lends and borrows with, asks for advice, trusts to tell a secret, tells everything to, conducts business with together, or engages in other activities. Only 19% of women gave their husband’s name at some point during this exercise. The other variable that lasso selects is a binary variable indicating that the woman listed her husband’s name as someone who she visits as a guest to eat with (negatively correlated with strength of the relationship). I interpret this to mean that these are cases where the woman does not live together with her husband, which may be true in cases of polygamy or separation due to work migration. However, while this variable is highly negatively correlated with the strength of the relationship, it is exceedingly rare and affects less than 1% of the sample.

Friends: Lasso selects 66 variables to explain the strength of ties with friends. I will focus on the five most important in explaining variation in the strength of ties with friends: usually seeing that person alone (+), being blood relatives (as opposed to relatives through her husband’s side) (+), being a non-relative friend who the woman met herself rather than through her husband (+), having a larger share of the network be comprised by friends with whom the woman engages in social activities rather than business activities (+), and being someone who employs the woman in piecework activities (-). Three main patterns emerge from these results: (1) more social activities and fewer business activities are important for the strength of networks, (2) seeing people alone, rather than in groups, is important for the strength of networks, and (3) knowing someone oneself, rather than through the husband, is important for the strenght of networks.

25 July, 2024 Blog Post 5: Depression and occupations among female rural-to-rural migrants

Measuring Depression

To measure depression, I use the CESD-R 10, a 10-item version of the revised Center for Epidemiologic Studies Depression Scale. This scale consists of ten statements, for which respondents indicate if, in the past week, they have experienced these feelings “Rarely or Never”, “Sometimes”, “Often”, or “Always”.

The ten statements are:

1. I was bothered by things that usually don’t bother me.
2. I had trouble keeping my mind on what I was doing.
3. I felt depressed.
4. I felt that everything I did was an effort.
5. I felt hopeful about the future.
6. I felt fearful.
7. My sleep was restless.
8. I was happy.
9. I felt lonely.
10. I could not “get going.”

Predictors of depression

I assessed depression in my sample of 1,600 low-income women. Twenty-seven percent of respondents in my sample had a CESD-R 10 score predictive of clinical depression (score >=10). I assess the relationship between depression and loneliness, social networks, and consumption. I find that consumption is the strongest predictor of depression.

I regress depression scores on the following variables: number of animal protein meals consumed in the last month (fish, meat, and eggs); number of meals consumed yesterday; number of meals consumed the day before yesterday; UCLA-3 loneliness score; number of friends in the social network; average strength of network connections. A one-standard deviation increase in animal protein meals consumed last month is associated with a 0.54 standard-deviation decrease in depression scores (p-value < 0.001). While loneliness and depression are related, the relationship is weak – a one-standard deviation increase in loneliness scores is only associated with a 0.01 standard deviation increase in depression scores (p-value < 0.001). Sixty-four percent of the sample is neither at risk of loneliness or depression; 5.6% of people are at risk of both; 21.5% are at risk of depression, but not loneliness; and 8.6% are at risk of loneliness, but not depression. Social networks are not meaningfully correlated with depression scores.

Occupations

This is an agricultural setting. Eighty-three percent of respondents report that they cultivate crops, and 26.7% of these women cultivate for subsistence only (the rest cultivate for subsistence and sale). Twenty-seven percent of participants report working in piecework employment (casual day labor), with the vast majority of them working in agricultural piecework. Twenty-nine percent of respondents earn income from business. For most respondents, these businesses are informal, and are not brick-and-mortar establishments. The most common businesses are selling prepared foods (46% of businesses), selling vegetables (17% of businesses) and selling kachasu (12% of businesses), a traditional distilled liquor. Only sixteen respondents (less than 1%) have a contract job, and six participants earn income by renting land that they own. Thirty respondents do not do any income-earning activities (almost all of whom are pregnant or have a young child).

Fifty-nine percent of respondents earn income from one occupation, 37% of respondents earn income from two sources, and 4% of respondents earn income from 3 or 4 jobs. By endline, people were working fewer jobs – 75% of respondents earned income from one source, and 15% from 2 sources, and 1% from 3 sources. The baseline survey was at the end of the harvest, meaning there was recently plenty of agricultural work available, people were cultivating their own crops, and people were selling their crops. The endline survey was conducted right before the land preparation for the rainy season begins, meaning that job opportunities are less available.

7 August, 2024 Blog Post 6: Migration and kinship structure

Migration

In one month, between the baseline and endline surveys, several households had already migrated. A major reason was because they were searching for part-time seasonal employment. These visits were during the dry season, during which agricultural jobs are scarce. In the endline survey, 1.5% of surveys were conducted on the phone because the respondents had moved sufficiently far away that it would not be feasible to conduct the survey in person. A small sample of households (less than 1%) had moved abroad (neighboring countries of Zambia and Mozambique), seeking seasonal employment. Another 2.5% of respondents had moved somewhere else within our study district, and we were about to interview them in person.

While 5% of the sample moving is not a concerning rate of attrition, the fact that 5% of people move within just one month does speak to the high level of mobility in this sample. If people so easily move between villages, it may reduce the incentives to invest in relationships with neighbors.

Kinship

Ninety-two percent of respondents belong to the Chewa tribe, a traditionally matrilocal tribe. However, as my findings indicate, these traditional kinship structures are no longer universally followed. Indeed, in the listing survey, 45% of households that respondents visited were eligible for the study, meaning that they did not grow up in the village that they lived. Some of these women (13%) belong to the village – meaning that the village takes ownership of them as members through their lineage, but they did not grow up in that village. Most women in this sample moved to their husband’s home village (77%).

Another 11% live in a village to which neither the respondent or her husband traditionally belong. Anecdotally, land scarcity is driving people to seek out areas with plentiful or cheap land. Indeed, low income is the strongest predictor of living in a village to which neither the respondent nor her husband belongs. Interestingly, religion is also a strong predictor – women who belong to the largest churches in this area (Catholicism and the Church of Central African Presbytarian) are less likely to live in a village that is not a home village to either the husband or wife. This may be because churches are a common place to engage socially and find community in this setting.

15 August, 2024 Blog Post 7: Intervention

Ninety-nine percent of women signed up for the Sharing List. Each Inviter sent an invitation to an average of 1.4 Invitees. The experiment created 1730 invitations across the whole sample, only 24 which were sent to someone from an Inviter’s baseline network. Aggregated, the study created 0.47 new invitations per person in the sample.

I randomly varied the saturation of the treatment (proportion of Control participants relative to Inviters and Invitees) across geographic clusters. I defined geographic clusters using k-means clustering, an unsupervised learning technique that groups units according to similarities in latitude and longitude, because villages are very large in this setting (Poll, 2024). I specified two parameters in my algorithm: each cluster should have at least 20 people in it, and clusters should have 36 people each on average. This resulted in 97 clusters, where 44 clusters (719 Control participants) were randomized to low-saturation, and 53 clusters (688 Control participants) were randomized to high-saturation. To ensure that these clusters are a relevant unit for social networks, I evaluated how often network links were within the same cluster, using network links where both individuals in the link were in my sample. Both sides of a network link were within my sample for 9% of network links. Of these relationships, 76% were within-cluster. Although these clusters are not completely socially disconnected from one another, randomizing treatment saturation at the cluster level led some Control participants to randomly have more treated neighbors than others – and therefore, more treated women from their existing or potential social networks – to be a part of the intervention.

I randomized half of the clusters to high-saturation, and half to low-saturation. In high treatment saturation villages, 13% of women were randomized to Control, 45% of women were randomized to be Invitees, and 55% were randomized to be Inviters. In low treatment saturation villages, 65% of women were randomized to Control, 20% to be Invitees, and 25% to be Inviters. I will estimate the treatment effect of the number of treatment participants within a certain radius (radius chosen to minimize the Schwarz BIC following Egger et al. (2022)), recentering the measure following Borusyak and Hull (2023) to purge it of bias from endogenous features.

There was a 48% increase in the number of invitations sent and received through the treatment within 100 meters of Control participants in high-saturation geographic clusters relative to Control participants in low-saturation geographic clusters (1.4 invitations sent and 1.3 invitations received, versus 0.89 invitations sent and 0.93 invitations received).

29 August, 2024 Blog Post 8: Reflections

Reflections from running this experiment:

1. Deep-diving into the anthropological and historical literature about Malawi has proven invaluable in conducting this research. I would encourage anyone conducting development research to engage deeply with the work produced by scholars from the country in which they conduct their research, particularly with work outside of their discipline, and meet with these scholars if possible.

2. One of the most important findings for me in this study has been documenting the way in which, even in societies that are almost ethnically and religiously homogenous, lower-income women are left out of financial and social groups. In my sample, a one-standard deviation increase in a wealth index is associated with a 10 percentage point increase in the probability of being a member of a village savings and loan association (women’s self-help savings groups), and an 8 percentage-point increase in the probability of belonging to a church-based women’s group. This relationship is true even among women who are all below the global poverty line.

3. Thank you to everyone who assisted me in this project. I would especially like to thank Invest in Knowledge Initiative (IKI), the survey firm with whom I partnered to conduct this project. The IKI leadership provided invaluable assistance throughout the project. The project field staff conducted excellent fieldwork, including field project management, enrolling vendors, and survey enumeration.

References

Banerjee, A., et al. 2023. Depression and Loneliness among the Elderly in Low- and Middle-Income Countries. Journal of Economic Perspectives, 37(2): 179-202.

Borusyak, Kirill, and Peter Hull. 2023. “Non-Random Exposure to Exogenous Shocks.” Econometrica, 91(6): 2155-2185.

Breza, E., A. Chandrasekhar, B. Golub, and A. Parvathaneni. 2019. “Networks in economic development.” Oxford Review of Economics Policy, 34(4): 678-721.

Chetty, Raj, et al. 2022. “Social capital I: measurement and associations with economic mobility.” Nature 608: 108-121.

Egger, Dennis, Johannes Haushofer, Edward Miguel, Paul Niehuas, and Michael Walker. 2022. “General Equilibrium Effects of Cash Transfers: Experimental Evidence from Kenya.” Econometrica, 90(6): 2603-2643.

Murdock, George Peter. 1967. Ethnographic Atlas. Pittsburgh: University of Pittsburgh Press.

Poll, Moritz. 2024. “Micro-enterprise saturation and poverty graduation at high frequency.” AEA RCT Registry. June 21. https://doi.org/10.1257/rct.11789-1.1