3.1 Networks as maps of baseline attitudes

We first examine differences in network structures across villages and explore whether individual-level network features are related to baseline attitudes toward refugees.

Figure 2 shows the social networks for each of the 16 villages using a Fruchterman–Reingold force-directed layout. Networks vary considerably from one village to another. While some feature most nodes highly interconnected to one another with high density (see villages 7 and 9), others show greater fragmentation into subcommunities (see villages 10 and 14) and some contain isolates (see villages 1 and 16). Isolates here most likely indicate that that node’s primary relationships are not within the village.

With complete views of the networks at the village level, we can explore the relationship between node-level positions and baseline attitudes. Since network structure can be correlated with other traits and behaviors (Atwell and Nathan, Reference Atwell and Nathan2022; Burt et al., Reference Burt, Jannotta and Mahoney1998), it is possible that network positions would also predict attitudes toward an out-group such as refugees.

Table 4.

Node-level network correlates of attitudes at baseline. Only transitivity is statistically significantly related to attitudes, with higher transitivity associated with colder attitudes

Table 4 Long description

A table with seven columns and eight rows, including a header row. The columns are labeled (1), (2), (3), (4), (5), (6), (7), and the rows are labeled Degree, Avg dist, Closeness, Betweenness, Eigenvector, Transitivity, Village FE, Observations, and R squared. Each cell contains numerical values or text. The table presents OLS regressions and village-clustered standard errors in parentheses. Notable values include significant transitivity values of 1.8 and 1.7 with standard errors of 0.66 and 0.63, respectively. The table also includes observations ranging from 1,824 to 1,981 and R squared values around 0.15 to 0.16.

OLS regressions and village-clustered standard errors are in parentheses. Signif. Codes: ***0.01, **0.05, *0.1.

We calculate a node’s degree, average distance to all other nodes in the same component, closeness centrality, betweenness centrality, eigenvector centrality, and transitivity. We regress baseline attitudes (range of 6–30) on these measures one at a time, including village fixed effects (since we observe substantial cross-network variation). As shown in Table 4, the only feature that is statistically significantly related to baseline attitudes is transitivity. On average, the more that a node’s social ties are linked to one another, the lower is that node’s baseline attitude toward refugees. The same relationship holds when we include subsets of the network features or, as shown in column 7, all other network features at once. Transitivity is consistently related to lower baseline attitudes.

This pattern is consistent with foundational ideas about the value of bridging structural holes (Burt, Reference Burt1992) and weak ties (Granovetter, Reference Granovetter1973), which hold that tight-knit pockets within a network can be isolated from new ideas. People in especially tight-knit pockets of their network with few ties outside of it can sustain more prejudicial views about outsiders.

3.2 Attitude homophily increases

Next, we zoom in on the extent to which nodes that are linked to one another hold similar views. Similarity in linked nodes—homophily—is common in social networks with respect to lots of attributes (Jackson et al., Reference Jackson, Nei, Snowberg and Yariv2023; McPherson et al., Reference McPherson, Smith-Lovin and Cook2001), including explicitly political ones (Bond and Sweitzer, Reference Bond and Sweitzer2022; Huber and Malhotra, Reference Huber and Malhotra2017). We might expect homophily on attitudes toward refugees if people have deliberated the issue with their social ties, or if their ties were formed due to their shared views on refugees, or even if their ties were formed because of other attributes in common that lead them to react to common events with similar perspectives.

We measure homophily on attitudes in terms of the absolute differences in attitude index scores between a node and her network neighborhood. Specifically, for node $i$ we calculate

\begin{eqnarray*} NW Dif_{i}^{Bl} = \frac {1}{\#N_{i}}\sum _{j \in N_{i}} |AttitudeBl_{i} - AttitudeBl_{j}| \end{eqnarray*}

where $N_{i}$ is the set of $i$ ’s neighbors in the network, $N_{i} = \{j \, : \, ij = 1\}$ for $j \neq i$ , and $\#N_{i}$ is the number of these neighbors. $NW Dif_{i}^{Bl}$ is then the average absolute difference between $i$ and her network neighbors’ attitudes on the 6–30 scale at baseline. While this value on its own is not very meaningful, we can calculate the same for node $i$ at endline,

\begin{eqnarray*} NW Dif_{i}^{El} = \frac {1}{\#N_{i}}\sum _{j \in N_{i}} |AttitudeEl_{i} - AttitudeEl_{j}|, \end{eqnarray*}

which allows us to compare the extent of network difference in attitudes between the start and the end of the study for a particular village’s network. Figure 3 shows the distribution of $NWDif^{Bl}$ and $NWDif^{El}$ , the network differences throughout a village at baseline and endline for all sixteen study villages.

Figure 3.

Absolute differences in attitudes between network neighbors at baseline and endline in each of the sixteen study villages. These differences shrink considerably by endline.

The plots in Figure 3 show that in all villages, the extent to which nodes hold different attitudes from their network neighbors decreases between baseline and endline. In other words, homophily on attitudes increases. It also shows that the extent to which homophily increases varies across villages, with some ending the study with attitudes that are substantially more similar to one another than at the start.

Of course, this increase in homophily does not tell us that people became more similar to their network neighbors because of the network. An intervention that shifts attitudes in a common direction on a scale that is capped would mechanically induce greater homophiliy. In the extreme case where everyone ended the study with attitudes maxed out at 30, everyone would hold the same attitudes as their neighbors and $NWDif^{El}$ would be 0 for everyone, regardless of whether anyone was compelled by anything their network neighbor said on the matter.

To explore whether these shifts in homophily are consistent with people actively moving their attitudes toward those of their network neighbors, we perform two tests. The first is a placebo test that benchmarks what the average $NWDif^{El}$ would be if everyone in the village simply experienced a level shift in their attitudes the size of the mean observed change in the village, subject to the capped scale. In other words, for all people in a village, their placebo endline attitudes are taken to be their own baseline attitudes plus a level shift the size of the village mean increase. If that would make their placebo endline exceed 30, the score is replaced with 30. This mimics the situation where everyone warms independently of the network and attitudes can be compressed by the cap.

Table 5 shows the results of this exercise. It first reports the average $NWDif^{Bl}$ and $NWDif^{El}$ in each of the 16 villages, quantifying the observed change over time in the bold. It then reports the average network difference that would be observed at endline under the placebo test. In all 16 villages, the amount of incidental change we would see from a level shift in attitudes that occurs absent any network influence is smaller (in most cases much smaller) than the change we do observe.

Table 5.

Change in network differences by study villages. In all villages, individuals hold attitudes that are more similar to their network neighbors at endline than they did at baseline, and this change is much larger than would be observed under a placebo level shift in attitudes

Table 5 Long description

The table presents data on network differences in 16 study villages, comparing baseline (Bl) and endline (El) values. It includes columns for V1 to V16, with rows for NW dif Bl, NW dif El, Change, Placebo dif El, and Placebo change. Each row provides specific values for each village. Notable trends include varying changes in network differences across villages, with some villages showing significant changes while others show minimal changes. The placebo change values are generally smaller than the observed changes.

While this is reassuring that the increase in homophily is not simply mechanical, we perform an additional test that provides a better sense of expected variation in homophily. For each village, we take the network and baseline attitudes as given. We also take the set of observed attitude changes—the extent to which a person’s attitude changed by the end of the study—as given, but not their attribution. Instead, we shuffle these changes on the network, reassigning attitude changes to nodes at random. For example, in a three person village if between baseline and endline someone’s attitude increased by 3 points, someone else’s increased by 5 points, and someone’s decreased by 2 points, we would have a set of attitude changes $\{3,5,-2\}$ . Randomly reassigning these would lead to a new set of simulated endline attitudes for people on the observed network and a new average value of $NWDif^{El}$ for the village.

Iteratively shuffling the attitude changes and recalculating the network similarity that would be observed at endline results in a sampling distribution of a village’s mean $NWDif^{El}$ . Figure 4 shows these for all 16 villages. Vertical lines indicate the actual $NWDif^{El}$ that we observe for that village. In all cases, the observed network difference is substantially smaller than would be expected by chance under this permutation test. In fact, the implied p-values in all 16 villages are less than  $.001$ .

Figure 4.

Sampling distribution of the mean network difference at endline when the observed distribution of long-term changes in the village are shuffled across the network. The observed reduction in network differences (equivalently, the observed increase in homophily) is too large to be expected by chance.

Figure 4 Long description

The image contains sixteen density plots, each representing a different village. Each plot shows the density distribution of mean endline network differences. The x-axis is labeled Value, and the y-axis is labeled Density. Each plot includes a blue density curve and an orange dashed vertical line. The values on the x-axis range from 2.5 to 10.0. The density curves vary in shape and position across the villages, indicating different distributions of network differences. The orange dashed lines represent specific values within each distribution.

Putting these pieces together, it seems that however homophilous a village’s network was in terms of attitudes toward refugees at the start of the study, it became much more so by the end. Attitudes changed to be more similar to network neighbors. The differences are larger than would be expected if all respondents were simply individually reacting to the study with warmer attitudes independent of their network neighbors and are also larger than would be expected if the study generated the same set of reactions within the village but independent of network position.

3.3 Attitudes respond to the treatment of network neighbors

Next, we leverage the study design, which tracks attitudes over time and exogenously induces attitude change by randomly treating approximately 40% of individuals in each village following the baseline survey.

The treatment consisted of a perspective-taking conversation about refugees conducted at the end of the baseline interview. During this conversation, an enumerator presented a narrative about a specific refugee and guided respondents through an exercise designed to encourage them to imagine themselves in a similar situation. Respondents were also given the opportunity to express concerns or reservations, which the enumerator addressed respectfully. Immediately following the intervention, attitudes among treated respondents were measured again.

We focus on the immediate role of treatment and network neighbors’ responses in shaping individuals’ endline attitudes. To the extent that respondents discussed refugees after the intervention, treatment may have generated spillovers that influenced the attitudes of their social contacts. Following Borusyak and Hull (Reference Borusyak and Hull2023), we construct a measure of indirect exposure to treatment through social ties. Specifically, we compare the observed number of treated network neighbors to the number expected under random assignment, given an individual’s number of ties and the probability of treatment assignment (see the Appendix for details).

Extra Trt captures the number of “extra” treated neighbors relative to this expectation. The measure is positive when an individual has more treated neighbors than expected by chance and negative when they have fewer. Extra Trt Strong Tie is defined analogously but is restricted to strong ties, where tie strength is measured using multiplexity (Granovetter, Reference Granovetter1973).

The first two columns of Table 6 suggest that the treatment of network neighbors is indeed related to a respondent’s endline attitudes. Respondents with more (fewer) treated ties than would have been expected by chance end with warmer (colder) endline attitudes, even when accounting for their own treatment status and their number of ties. The same is true when restricting attention to strong ties. Each extra treated tie is associated with endline attitudes 0.15 warmer on the 6–30 attitude scale; the increase is 0.23 for each extra treated strong tie.

Table 6.

Spillover analyses. OLS regressions with village fixed effects and controls that include gender, age, education, religion, economic precarity, experience as and contact with refugees. Full coefficient table and other specifications can be found in the appendix

Table 6 Long description

A table titled ’Attitudes El’ with four columns and twelve rows. The table presents OLS regression results with village fixed effects and demographic controls. The columns are labeled (1), (2), (3), and (4). The rows include Baseline attitude, Treatment, Extra trt, Degree, Extra trt strong tie, Strong degree, Neighbor reaction, Strong tie reaction, Village FE, Demographic controls, Observations, R squared, and Within R squared. Each cell contains regression coefficients and standard errors in parentheses. Notable values include Baseline attitude with coefficients around 0.30 to 0.33, Treatment with coefficients ranging from 0.32 to 0.45, Extra trt with a coefficient of 0.15, and Extra trt strong tie with a coefficient of 0.23. The table also includes significance codes and observations for each column.

Village-clustered standard errors are in parentheses. Signif. Codes: ***0.01, **0.05, *0.1.

These analyses have focused on the treatment status of neighbors. Next we consider not just whether neighbors received treatment, but how neighbors reacted to any treatment they received. Neighbor Reaction is the average attitudes of any treated network neighbors measured right after treatment, in Bl2. Strong Tie Reaction is the average attitudes of any treated strong tie neighbors in Bl2. Columns (3) and (4) of Table 6 show that the reactions are also related to endline attitudes. People whose neighbors reacted positively to treatment right away have warmer attitudes by endline, while people whose neighbors reacted negatively right away have colder endline attitudes. The scale of these variables is also 6–30. A one standard deviation increase in Neighbor Reaction (3.2 points on this scale) is associated with endline attitudes that are 0.26 warmer. Someone whose neighbors’ reaction is at the maximum of the scale has endline attitudes about 2 full points warmer than someone whose neighbors’ reaction is at the minimum. Reactions of strong ties are similarly impactful.

In the appendix we show that these findings are robust to other specifications. Importantly, we also show that they pass a placebo test in which we use as the dependent variable the immediate posttreatment attitudes rather than endline attitudes. Because the respondent would not have had a chance to consult with any of her network neighbors before forming these attitudes, the network neighbors’ treatment status and reactions could not have affected attitudes by this point in time. Any relationships that would appear in these regressions would mean that something spurious is connecting network reactions to attitudes. Reassuringly, we find no such relationships.

In short, long-term attitude change seemed to be actively shaped by the attitudes of network neighbors.

3.4 Attitudes change in response to strong reactors farther away

Finally, we broaden our focus beyond immediate neighbors and ask whether individuals’ attitudes are shaped by others elsewhere in the network.

These analyses exploit substantial heterogeneity in responses to the intervention. Although respondents reacted positively on average, treatment effects varied considerably across individuals. A minority of respondents reported attitudes immediately after the intervention that were less favorable than their baseline attitudes. Among those who responded positively, the magnitude of attitude change also varied widely: some respondents’ attitudes shifted only slightly, while others exhibited much larger changes.

If individuals who reacted most strongly to the intervention were also more likely to participate in subsequent village discussions, their influence may have extended beyond their immediate social contacts. Put differently, respondents who experienced the largest immediate attitude changes may have played a disproportionate role in shaping attitudes throughout the broader village network over time.

To explore evidence for this possibility, we identify the strongest reactors in each village in each direction. Specifically, we identify the three people whose immediate reaction to the intervention was the largest positive attitude change (MostPersuaded) and the three whose immediate reaction to the intervention was the largest negative attitude change (MostBacklashed). Then we calculate the length of the shortest path between each person and their nearest MostPersuaded strong reactor (DistPersuaded) and their nearest MostBacklashed strong reactor (DistBacklashed).

Figure 5 shows the distribution of these distance variables. Someone who is in the set MostBacklashed would have DistBacklashed = 0. Someone who is the immediate neighbor of someone in that set would have DistBacklashed = 1. The largest finite distance away from a member of either set is 5.

Now we regress attitudes at endline on these network distances to strong immediate reactors. The results are shown in column (1) of Table 7. A few results stand out. First, even when controlling for one’s own baseline attitudes and neighborhood size, network neighbors’ baseline attitudes are related to endline attitudes. By the end of the study, a person is drawn closer to her network neighbors’ attitudes.

Table 7.

Above and beyond one’s own and one’s network neighbors’ baseline attitudes, distance to strong reactors is related to endline attitudes. Greater distance from the strong positive reactors predicts more negative baseline attitudes, while greater distance from strong negative reactors (weakly) predicts more positive endline attitudes. Using the immediate posttreatment attitudes measured for the treated as a placebo, none of these factors is related to attitudes at that point in time (before any socializing could take place)

Table 7 Long description

The table has two columns labeled Placebo Attitudes El and Attitudes Bl2, and it includes various rows with corresponding values. Row 1: Baseline attitude, 0.33, 0.50. Row 2: NW Bl attitudes, 0.07, 0.006. Row 3: Degree, 0.02, 0.02. Row 4: DistPersuaded, -0.28, 0.003. Row 5: MostPersuaded, 2.5, 5.2. Row 6: DistBacklashed, 0.07, 0.01. Row 7: MostBacklashed, -1.7, -7.1. Row 8: Village FE, Yes, Yes. Row 9: Observations, 1,819, 880. Row 10: R squared, 0.39, 0.61. Row 11: Within adjusted R squared, 0.18, 0.53.

Village-clustered standard errors are in parentheses. Signif. Codes: ***0.01, **0.05, *0.1.

Figure 5.

Distribution of network distances to those who immediately responded most positively (DistPersuaded, left) and most negatively (DistBacklashed, right). Respondents with distance 0 are members of the reference category (MostPersuaded and MostBacklashed, respectively).

Second, extreme reactions have persistence. Being one of the people with the strongest immediate reactions positively or negatively—MostPersuaded and MostBacklashed—is associated with endline attitudes that are more strongly positive or negative. Strong negative reactions predict negative endline scores and strong positive reactions predict much more positive endline scores.

Finally, distance to strong reactors predicts endline attitudes in the expected direction. For positive reactors, this is estimated precisely. The greater the network distance to a strong positive reactor, the more negative one’s endline scores. That is, being farther from strong positive reactors is associated with more negative attitudes while being closer to strong positive reactors is associated with more positive attitudes.

Strong reactions were not randomly assigned. Those who reacted most strongly to their treatment may be different in other ways, and this difference could induce a spurious correlation between distance to these people and attitude change. We once again turn to a placebo test that takes the attitude measurement that happened right after treatment as the dependent variable to assess this possibility. Importantly for our strategy, this measurement occurred before a respondent would have had a chance to go and talk to their network neighbors to learn how they reacted to our study.

Column (2) of Table 7 reports the results of this placebo test. Reassuringly, none of the measures intended to capture active network influence are associated with the immediate posttreatment attitude outcome. The two variables that do emerge as strongly significant, MostPersuaded and MostBacklashed, are expected to do so because they are defined using extreme values of $Attitudes Bl2$ and are therefore mechanically related to the placebo outcome.

The key results are those highlighted in the table. These coefficients are substantively small—indeed, effectively zero—and statistically indistinguishable from zero. Their lack of association with the placebo outcome strengthens the interpretation that the corresponding relationships with endline attitudes reflect the active use of social networks in shaping attitudes during the intervening period, rather than preexisting correlations present immediately after treatment.

Finally, we include distance from strong reactors as a set of indicator variables, allowing the relationship between distance and attitudes to be estimated flexibly. Specifically, we enter separate indicators for each network distance from an extreme reactor, with distance 0 serving as the omitted reference category.

\begin{align*} Attitudes\_El_{i} &= \beta_{1} \, DistPers=1_{i} + \beta _{2}DistPers=2_{i} + \beta _{3}DistPers \\[4pt] &=3_{i} + \beta _{4} \, DistPers=4_{i} + \beta _{5}DistPers=5_{i} + \ldots + \epsilon _{i}. \end{align*}

We also control for respondents’ average network distance to all other individuals in their village and for baseline attitudes.

Figure 6.

Coefficient plot showing relationship between each distance away from a strong reactor and endline attitudes, controlling for baseline attitudes, average network distance, village fixed effects, and standard errors clustered by village. Being close (low distance) to those who reacted positively and far (high distance) from those who reacted with backlash are associated with warmer endline attitudes.

Figure 6 presents the resulting coefficient estimates. The pattern is broadly consistent with the hypothesized role of network proximity. Attitudes are more favorable among respondents who are closer to strong positive reactors and farther from strong negative reactors. The largest effects appear at greater network distances from the reference category, although these estimates should be interpreted with caution given the smaller number of observations at longer distances (see Figure 5).