Association Between Social Network Communities and Health Behavior: An Observational Sociocentric Network Study of Latrine Ownership in Rural India

Abstract

Objectives. We identified communities of interconnected people that might serve as normative reference groups for individual-level behavior related to latrine adoption.

Methods. We applied an algorithmic social network method to determine the network community from respondent-reported social ties of 16 403 individuals in 75 villages in rural Karnataka, India; data were collected from 2006 to 2008. We used multilevel modeling to test the association between latrine ownership and community-level and village-level latrine ownership. We also investigated the degree to which network cohesion affected individual latrine ownership.

Results. Three levels of social contacts (direct friends, social network community, and village) significantly predicted individual latrine ownership, but the strongest effect was found at the level of social network communities. In communities with high levels of network cohesion, the likelihood was decreased that any individual would own a latrine; this effect was significant only at lower levels of latrine ownership, suggesting a role for network cohesion in facilitating the nonownership norm.

Conclusions. Although many international health and development interventions target village units, these results raise the possibility that the optimal target for public health interventions may not be determined through geography but through social network interactions.

Interconnected groups of individuals can shape the health-related behavior of the constituent individuals, in part by promulgating and enforcing group-specific social norms. An understanding of the process whereby such groups are assembled and such norms are enforced could be used to improve the health of both individuals and groups. Recently developed social network community detection techniques allow us to discern these groups and suggest ways of exploiting social processes to foster desirable behavior change in diverse settings,^1–3 and other work has begun to use network community detection to understand normative constraints in outcomes as diverse as smoking cessation⁴ and adolescent delinquency.⁵

Descriptive norms are those regular behaviors that can be observed within a community and that can be adopted regardless of the expectations of others.⁶ Injunctive norms, on the contrary, reflect community standards and are enforced through social sanctions; sanctions can be negative for transgressions or positive for conformity and can be as blatant as a public shaming or as subtle as unspoken disapproval.^7–10 Simply observing others’ behaviors can unconsciously motivate action,¹¹ but injunctive norms are more complex, and may even be conflicting, because their primary function is to maintain social relationships.¹²

Medicine and public health have a rich tradition of efforts to change norms and hence move individuals and communities toward better health.^5,13,14 However, a key task of interventions seeking to do this is to identify the valid community or “reference group” for each individual in the population of interest. Normative change hinges on the fact that the individuals involved place a value on the expectations of those with whom they share social connections.^9,15,16 However, from which particular groups of people do individuals take their guidance? How is such a group defined? The theory of planned behavior, for instance, includes self-reported measures of reference groups subjectively defined as “people who are important to you.”^17–19 Most research surveys (e.g., with outcomes as diverse as the use of contraception^20,21 or adolescent smoking²²) have not collected data specific to social norms, requiring researchers instead to create measures of possible norms through the aggregation of existing measures and to infer reference groups from geographic boundaries, such as villages. However, does a valid reference group include (1) just those to whom one is directly connected (friends), (2) those to whom one is more broadly and indirectly connected (e.g., friends of friends), or (3) a still larger aggregation (e.g., the village or town in which one resides)? And are structural features of the network in which individuals are embedded relevant to the ability of reference groups to enforce norms or stimulate the adoption of new behaviors?

Not only may the group one picks as a normative referent be relevant, but also the structure of the ties within that group may be crucial to any normative effect. Network cohesion is an important such structural characteristic. Cohesion can be quantified with transitivity, which is the probability that 2 of someone’s social connections are connected to each other.^5,23–27 Transitivity can either constrain or compel adoption of complex behaviors (see Figure 1 for a theoretical illustration of this concept).^25,28,29 For example, people in more highly interconnected communities may be less likely to embrace behavior change because the strongly clustered nature of their social groups means that they may have little access to new information outside of their networks.^23,30 Furthermore, in highly interconnected communities, the effect of injunctive norms on individuals may be more powerful.^25,31–33 It can be very difficult for an individual to defy a social norm even when the expected outcome may be personally beneficial—because the individual can avoid sanctions only if the entire community is willing to change simultaneously.^34–36 This dynamic also works inversely, such that when a critical mass of a highly interconnected group has adopted a behavior, the probability also increases that any individual in that group will adopt it.^28,29,37,38

FIGURE 1—

Theoretical changes in community latrine proportion as a function of transitivity at 3 different levels of community latrine proportion.

Note. When the overall latrine ownership proportion in a community is low, then the effect of transitivity is to act as a normative constraint on overall latrine ownership within the community. When latrine ownership is medium, then the effect of transitivity is null. At high levels of latrine ownership, transitivity then acts as a catalyst toward adoption, and community-level latrine ownership will increase.

We used algorithmic social network analytic methods to identify community groups and predict a socially influenced health behavior: latrine ownership in rural India. Poor sanitation, including the lack of clean, functioning toilets, is a major contributor to morbidity and mortality from infectious disease, particularly in India.^39–41 Open defecation has been practiced in many communities for centuries.⁴² A large body of qualitative research points to the fact that, above and beyond demographic factors, social dynamics can influence the decision to build a latrine.^40,43 Although the government of India has devoted considerable resources, including subsidies, to latrine-building campaigns over the last decade, results have been mixed, and considerable socioeconomic disparities persist.^44–46 The most successful latrine-building campaigns have been those initiated by Community-Led Total Sanitation; these successfully shift the norms of the community through village-wide agreements to stop open defecation, and these include shaming of those who transgress, and community commitment to invest in building latrines.^47,48

We used so-called community detection methods to mathematically identify relevant social groups of interconnected people. These network communities are subsets of individuals with relatively strong within-group social connections and relatively weak between-group connections.^3,49,50 As such, they include not only direct social contacts (friends) but also friends of friends and friends of friends of friends.⁵¹ Because latrine interventions in the developing world typically are implemented at the village level, we expected that both village-level latrine ownership practices and social network community–level practices would have important implications for individual behaviors.

Our research aims were to (1) quantify groups of interconnected individuals with a social network algorithm; (2) assess whether latrine ownership at the level of direct contacts, community network groups, or villages was associated with individual latrine ownership; and (3) test the degree to which the level of connectivity at the community and village levels was associated with latrine ownership. Research that uses a community-detection approach to test the association between social network features and health behavior is rare; hence, this study offers the opportunity to investigate whether network-derived communities are a good unit of analysis for the measurement of injunctive social norms.

METHODS

We used sociocentric network data from 75 villages in rural Karnataka, India, that were collected as part of an analysis of microfinance diffusion.^52,53 Sociocentric studies focus on a small population and attempt to ascertain all of the social relationships within a set of interconnected individuals.⁵⁴ From 2006 to 2008,⁵² a complete census of each village was taken by interviewing 1 person within each household about household characteristics such as latrine ownership; individual surveys also were used to collect demographic and network data from more than half of the eligible households (those with women between ages 18 and 57 years), which were randomly chosen using stratification by religion and geographic sublocation.^52,53 A total of 16 984 individuals were interviewed. We defined our population as being eligible women and their spouses; links to others outside this population, like any sociocentric network study, were unavoidably excluded. After we removed observations with missing data (and certain other exclusions; see “Network Measures”), our analysis included data on 16 403 individuals in 6786 households. Of these, 6490 were household heads, 5845 were spouses of household heads, and 4068 were other individuals in the household. (Data used for this study are publicly available at: http://thedata.harvard.edu/dvn/dv/jpal/faces/study/StudyPage.xhtml?globalId=hdl:1902.1/16559.)

Individual-Level Measures

We measured age, gender, religion (Hindu or Muslim), native language (Kannada, Tamil, Telugu, or Hindi), and caste (scheduled caste, scheduled tribe, other backward caste, or general). We measured education using 16 levels ranging from none to higher degree. Household quality variables included roof type (6 categories), number of rooms in the home, number of beds in the home, and household electricity (private, government, or none). We used these household quality measures as a proxy for income.⁵⁵ We also included the type of ration card held by each individual: below poverty level, above poverty level, or not holding a card.

Our outcome variable was a binary measure of household latrine ownership.

Network Measures

A name generator is the survey instrument used in social network data collection to elicit individuals’ important connections.⁵⁶ Respondents (egos) were asked to name up to 8 social contacts (alters) for each of 12 name generators. All eligible women and their spouses could be named, thus allowing us to assemble the network. Each network tie indicated that one individual had named the other (or vice versa) in at least 1 name generator. Dyads in which the individual and her social contact were from the same household were excluded from the analysis. For each individual, we then created 1 continuous measure representing the proportion of her alters, including both inbound and outbound ties, that owned a latrine.

The social communities to which people belonged were mathematically generated from sociocentric social network data. To detect communities, we applied the “fast greedy” algorithm, which optimizes a quantity known as modularity.⁵⁷ Modularity is a measured property of a network obtained by calculating the proportion of ties (or “edges”) that fall within groups minus that same proportion calculated for a random network.⁵⁸ The larger the value, the more likely it is that the observed network of ties within a group of individuals is not due to chance.⁵⁰

To ascertain the relation between network connectivity and latrine ownership, we also calculated transitivity at both the community and the village levels. Transitivity measures the number of connected triads divided by the total number of possible triads (in other words, the probability that 2 of a person’s friends are also friends with each other).

Network Community–Level and Village-Level Measures

Our key predictors were the proportion of a respondent’s social network community that owns a latrine and the proportion of a respondent’s village that owns a latrine. All group proportion measures were calculated excluding the participant’s own value.^20,59 Communities that had 10 or fewer individuals were excluded from the analysis to ensure more reliable community-level means.²⁰

Statistical Analyses

We first used logistic regression, including demographic and household covariates, to generate separate models regressing participant latrine ownership on the proportion of direct social contacts that owned a latrine, the proportion of same-community members that owned a latrine, and the proportion of same-village members that owned a latrine.

We specified multilevel models that included covariates at the participant, social contact (friend), community, and village levels. Specifically, we estimated hierarchical logistic regression models to adjust for the clustering of observations at the different levels, and we partitioned the variance of the dependent variables into social contact, community, and village components.⁶⁰

RESULTS

Each of the 16 403 individual participants were in 1 of 396 communities nested within 75 villages. The mean community size was 48 (SD = 27) with a range of 11 to 185. The mean number of respondents per village was 244 (SD = 73) with a range of 90 to 398. Table 1 shows descriptive statistics for the study population. The overall mean rate of latrine ownership within the sample was 30%, although this varied according to community (range = 0%–85%) and village (range = 2%–58%).

TABLE 1—

Demographic and Individual Characteristics of Adult Respondents (n = 16 403) From Rural Karnataka, India: 2006–2008

Characteristic	Mean (SD) or Proportion
Age, y	39.30 (12.50)
Education, y	5.02 (4.65)
No. of rooms in house	2.60 (1.50)
No. of beds in house	1.04 (1.62)
Alter owns a latrine	0.31
Gender (female)	0.551
Caste
Scheduled caste	0.252
Scheduled tribe	0.060
Other backward caste	0.566
General	0.122
Ration card category
Above poverty	0.175
Below poverty	0.669
Does not own card	0.157
Roof type
Thatch	0.020
Tile	0.314
Stone	0.305
Sheet	0.182
Reinforced cement concrete	0.138
Other	0.041
Household electricity
Private	0.657
Government	0.285
No	0.058
Language
Kannada	0.750
Tamil	0.040
Telugu	0.170
Urdu	0.040
Other	0.002
Latrine ownership	0.304

Figure A (available as a supplement to the online version of this article at http://www.ajph.org) shows the distribution of latrine ownership proportions at the village versus community levels. Although these 2 measures are correlated (Pearson r = 0.62), note that, within villages, network communities can vary widely in the proportion of their members who own latrines. Table A (available as a supplement to the online version of this article at http://www.ajph.org) shows the range of values for all measures aggregated at the community and village levels. In addition to the latrine measures, nearly all other measures show greater between-group variation at the community level than at the village level, suggesting the important observation that communities may indeed be better units of analysis for discovering reference groups than are villages. Figure 2 shows the clustering of latrine ownership by community within 1 village in the network.

FIGURE 2—

A network depiction of 1 village in Karnataka, India: 2006–2008.

Note. Objectively identified network communities (using modularity maximization) are differentiated by color. Those who own latrines are depicted with squares, and those who do not own latrines are depicted with circles. Note both the clustering of individuals and the varying distribution of latrine ownership by community. For instance, in the pink community on the right side of the figure, no one has a latrine. The dark blue community toward the top, however, has a latrine ownership proportion of approximately 50%. If latrine ownership within the community was determined solely by geography, then we would expect the distribution of latrines to be uniform throughout the communities.

We found that 11% of the variance of individual latrine ownership can be accounted for at the village level and 29% at the community level. We tested the bivariate relations between individual latrine ownership and the size of the community and the size of the village with a random effects model to account for clustering at both the community and the village levels (not shown); neither association was significant, suggesting that the size of the network community was not driving the results.

Network Predictors of Latrine Ownership

In Table 2, we show models that test the relation between participant latrine ownership and latrine ownership of social contacts, members of the network community, and members of the village; these multilevel models include individual demographic controls and random effects at each level. Model 1 includes separate measures of latrine ownership at the level of direct social contacts, community, and village. All 3 associations were strongly significant, but the association with network community latrine ownership was highest. The odds of individual latrine ownership increased by 1.19 times (95% confidence interval [CI] = 1.05, 1.34) for each 1 SD increase in village latrine ownership and by 1.12 times (95% CI = 1.07, 1.18) for each 1 SD increase in social contact ownership. However, they increased by 1.42 times (95% CI = 1.33, 1.54) for each 1 SD increase in community ownership, suggesting that the best reference group for an individual extends beyond the directly connected social contacts to friends of friends and the broader network community in which he or she is embedded. To frame these results in comparison with demographic effects, an individual in a community in which 70% of people own latrines has a 2.02 times (95% CI = 1.72, 2.39) greater odds of owning a latrine than does a person living in a community in which 30% of people own latrines. This is comparable to the difference in the odds of owning a latrine between someone of general caste and someone of scheduled caste (odds ratio [OR] = 1.90; 95% CI = 1.56, 2.31), between someone with 12 years of education and someone with none (OR = 2.05; 95% CI = 1.62, 2.60), and between someone above the poverty level and someone below (OR = 2.04; 95% CI = 1.81, 2.32).

TABLE 2—

Results of Multilevel Logistic Regression Analysis Showing Association in Social Group Latrine Ownership With Individual Latrine Ownership at the Contact, Community, and Village Levels in Rural Karnataka, India: 2006–2008 (n = 16 403)

Fixed Part	Model 1, OR (95% CI)	Model 2, OR (95% CI)	Model 3 (With Village-Level Fixed Effectsa), OR (95% CI)
Proportion of contacts with latrine	1.48 (1.24, 1.76)	1.48 (1.24, 1.76)	1.45 (1.21, 1.73)
Community latrine proportion	5.87 (3.89, 8.86)	4.53 (2.94, 6.97)	5.00 (3.45, 7.26)
Village latrine proportion	4.18 (1.47, 11.80)	3.53 (1.15, 10.77)
Community transitivity		0.15 (0.05, 0.49)	0.18 (0.07, 0.50)
Village transitivity		0.30 (0.01, 6.60)
Caste
Scheduled caste (Ref)	1.00	1.00	1.00
Scheduled tribe	0.83 (0.63, 1.09)	0.84 (0.64, 1.11)	0.87 (0.67, 1.12)
Other backward caste	1.54 (1.31, 1.80)	1.57 (1.34, 1.83)	1.54 (1.34, 1.76)
General	1.90 (1.56, 2.31)	1.95 (1.61, 2.38)	1.97 (1.62, 2.40)
Household electricity
Private (Ref)	1.00	1.00	1.00
Government	0.37 (0.33, 0.42)	0.37 (0.33, 0.42)	0.38 (0.34, 0.43)
No	0.25 (0.18, 0.34)	0.25 (0.18, 0.34)	0.25 (0.18, 0.34)
No. of rooms in house	1.21 (1.16, 1.26)	1.21 (1.16, 1.26)	1.21 (1.16, 1.26)
No. of beds in house	1.43 (1.38, 1.49)	1.43 (1.38, 1.49)	1.43 (1.38, 1.49)
Roof type
Thatch (Ref)	1.00	1.00	1.00
Tile	0.79 (0.47, 1.31)	0.79 (0.48, 1.32)	0.77 (0.46, 1.28)
Stone	1.77 (1.06, 2.94)	1.75 (1.05, 2.91)	1.80 (1.08, 3.00)
Sheet	1.67 (1.00, 2.77)	1.67 (1.00, 2.77)	1.67 (1.00, 2.77)
Reinforced cement concrete	4.26 (2.51, 7.24)	4.26 (2.51, 7.24)	4.35 (2.56, 7.38)
Other	0.82 (0.46, 1.45)	0.81 (0.46, 1.43)	0.84 (0.47, 1.47)
Years of education	1.06 (1.04, 1.08)	1.06 (1.06, 1.06)	1.06 (1.06, 1.06)
Ration card
Above poverty (Ref)	1.00	1.00	1.00
Below poverty	0.49 (0.43, 0.55)	0.49 (0.43, 0.55)	0.49 (0.43, 0.55)
No card	0.68 (0.59, 0.78)	0.68 (0.59, 0.78)	0.68 (0.59, 0.78)
Language
Kannada (Ref)	1.00	1.00	1.00
Tamil	0.90 (0.68, 1.18)	0.90 (0.68, 1.18)	0.87 (0.66, 1.14)
Telugu	1.17 (1.00, 1.37)	1.16 (0.99, 1.36)	1.17 (1.00, 1.37)
Urdu	1.80 (1.40, 2.33)	1.88 (1.46, 2.42)	1.86 (1.44, 2.40)
Other	1.70 (0.69, 4.19)	1.72 (0.70, 4.23)	1.77 (0.72, 4.36)
Village-level variance	0.19	0.21
Community-level variance	0.12	0.13	0.05
Log likelihood	−6799	−6792	−6682

Note. CI = confidence interval; OR = odds ratio.

^aNot shown.

Social Cohesion

In Table 2, we evaluate 2 measures of network connectedness (transitivity at both village and community levels); both were individually significant in multivariate models including individual attributes and latrine proportion variables (not shown), but when we incorporated both measures, village-level transitivity was not significant at the .05 level. Figure B (available as a supplement to the online version of this article at http://www.ajph.org) shows that individual latrine ownership declined as community transitivity increased. In other words, individuals in more tightly connected communities were less likely to own latrines. The models with controls in Table 2 suggest that the odds of individual latrine ownership decreased by 1.13 (95% CI = 1.03, 1.23) for each 1 SD decrease in community-level transitivity.

Interaction Effect

Finally, we found that the association between community latrine ownership and individual latrine ownership varied according to community level transitivity (Table B and Figure C, available as supplements to the online version of this article at http://www.ajph.org). To explore this association further, we stratified the analyses, running separate models for communities with low and high latrine proportions (less than and greater than the median of 0.30). We found that for communities with low latrine proportions, each 1 SD decrease in community transitivity increased the odds of owning a latrine by 1.58 (95% CI = 1.42, 1.75), whereas for communities with high latrine proportions, the effect of transitivity was not significant.

DISCUSSION

To change the health behavior of individuals, we often must change the social norms within the web of ties in which individuals are embedded, targeting small but well-defined groups of people at the same time. The ability to identify the socially relevant groups within which the norms are held, therefore, is crucial to the science of behavior change and hence to the goal of sustainable health promotion efforts in diverse settings. Identifying reference groups, however, is a complex task. Earlier work involving theory of planned behavior found that group-level norms are predictive of behavior only when the reference group has been correctly identified.^61,62 Asking participants to directly name important others (as done for theory of planned behavior) is a step in the right direction, but it does not capture the multidimensional layers of social connections possible in sociocentric network analysis. Similarly, researchers who use demographic and health survey data have approximated reference groups from sampling clusters, but this method is based solely on geography, with no measures of direct social ties.

These results suggest that organic social network communities are more strongly associated with normatively driven behavior (in this case, latrine ownership) than are direct social contacts (e.g., friends) and contacts bounded geographically within a village. Although current efforts to change norms are typically directed at the village level,⁶³ these results imply that norm-based interventions may be more effective if they target network communities within villages.

The results also showed that network connectivity is inversely associated with individual latrine ownership at the community level. Individuals in less transitive communities are more likely to own latrines than are those within more transitive communities. These results are consistent with previous research suggesting that the degree of connectivity within a network can affect the adoption of novel behaviors.^{31,33,64–67} If a behavior is normatively determined, then those in highly connected reciprocal networks may adopt more slowly, because the change must occur at the community level rather than at the individual level.²³ Centola et al.^28,68 reported that because the transmission of norm-changing behavior often requires multiple reinforced exposures, it can be inhibited initially by highly transitive networks but ultimately requires the reinforcement of such networks to successfully occur. Research on condom use, for instance, has found that individuals in more tightly connected networks are less likely to have sexual partners willing to use condoms.^31,32 Potentially, once a critical mass of partners in tightly knit social groups adopted condom use, it would be reinforced rather than inhibited.

We conjecture that the network-determined communities we identified in our study actually make up the salient reference groups to which individuals turn for information on normative expectations. Descriptive norms can potentially diffuse between people within the same geographic vicinity, such as between strangers at the local market.⁶⁴ Injunctive norms, however, can take place only within communities that are socially relevant to an individual and with whom that individual strives to maintain positive relationships.¹² The relationships we observed in this sample could certainly be the result of both dynamics. However, the fact that socially connected network communities are the most significantly correlated with latrine ownership suggests that an injunctive normative dynamic may be the most significant for latrine adoption. The fact that transitivity at the community level is correlated with a decreased probability of latrine adoption is further evidence of possible normative influence.

Furthermore, the interaction model suggests that the effect of community transitivity is strong in communities with few latrines but is negligible in communities where the proportion of latrines is greater than 30%. If normative influences are working against latrine adoption, we would expect a strong relation between the level of transitivity in the network and the probability of an individual owning a latrine, but we also would expect this effect to lessen as latrine ownership became more accepted within the community. If latrine ownership became the normative standard within communities (e.g., at rates > 70%), we might expect to see a positive correlation between transitivity and community latrine ownership, as the new norm of ownership would be more reinforced. As more and more people within tightly connected communities adopt latrines, individual exposure to latrines would increase, the exposure would occur from multiple close connections, and ultimately the new norm would be established. We cannot evaluate this possible association with these data because we have very few communities at those very high rates of adoption.

Finally, to explain these dynamics, we also must consider the possibility of homophily,^69,70 or the fact that individuals may choose their social relationships based on similarities that would also lead each individual to build a latrine. That our community groups showed a higher degree of within-group demographic homogeneity than villages did is clear evidence of homophily, as well as important evidence that these groups are indeed potential sources of reference for normative beliefs. In our analyses, however, these demographic and other plausible, measured factors were controlled for at the ego level. Thus, although homophily may certainly drive group formation, our evidence suggested that normative forces beyond homophily may be driving latrine-building decisions.

Limitations

This study had several limitations. First, because the data were from a single time point, we were not able to track time-dependent effects that might have shed light on causal explanations for the associations we observed. Although our results were consistent with our causal hypothesis, longitudinal studies with randomly assigned treatment groups would substantially strengthen this analysis. We are undertaking such studies. Second, the data were sociocentric, but only 50% of the population was surveyed. Although exclusions were random, having missing network ties could have biased our results. Third, latrine ownership does not guarantee latrine use. Further analysis could shed light on the degree to which latrine use is socially predicted and whether network-based interventions eventually lead to a decrease in morbidity and mortality from sanitation-related infectious disease.

Conclusions

Understanding the role of network communities in facilitating health behavior change is a new and emerging field of research. This is one of the few studies to show that employing a network community intervention might facilitate adoption, particularly for behaviors that are complex and socially determined. Analysis of large data sets is becoming easier, and health behavior researchers will have increasing opportunities to identify network communities within large populations of people.

Perhaps the optimal target for public health interventions is not determined through geography but through social network communities; similarly, the more optimal target for clinical interventions may not be the individual but, again, the community. Even though many international health and development interventions target village units, if the village unit is not the primary source of normative expectations for individuals, then village-level interventions aimed at changing norms may be inefficient or even ineffective. Future work should explore interventions that focus on identifying and engaging socially meaningful community groups to determine whether this helps maximize the likelihood of widespread and enduring behavior change.