Social network- and community-level influences on contraceptive use: evidence from rural Poland

The diffusion of ‘modern’ contraceptives—as a proxy for the spread of low-fertility norms—has long interested researchers wishing to understand global fertility decline. A fundamental question is how local cultural norms and other people's behaviour influence the probability of contraceptive use, independent of women's socioeconomic and life-history characteristics. However, few studies have combined data at individual, social network and community levels to simultaneously capture multiple levels of influence. Fewer still have tested if the same predictors matter for different contraceptive types. Here, we use new data from 22 high-fertility communities in Poland to compare predictors of the use of (i) any contraceptives—a proxy for the decision to control fertility—with those of (ii) ‘artificial’ contraceptives—a subset of more culturally taboo methods. We find that the contraceptive behaviour of friends and family is more influential than are women's own characteristics and that community level characteristics additionally influence contraceptive use. Highly educated neighbours accelerate women's contraceptive use overall, but not their artificial method use. Highly religious neighbours slow women's artificial method use, but not their contraceptive use overall. Our results highlight different dimensions of sociocultural influence on contraceptive diffusion and suggest that these may be more influential than are individual characteristics. A comparative multilevel framework is needed to understand these dynamics.


Data collection
Our sample of communities (21 villages and one town) was randomly drawn from four neighbouring municipalities containing a total of 34 potential study populations. All 22 of our sampled communities agreed to participate. We did not have access to a list of names or a map of households, so within each community every third house was sampled, with every adult woman (≥18 years) present in the house invited to take part in the survey. This strategy was further stratified in the town, by randomly selecting streets from a list obtained from the local government and approaching every third house/apartment on that particular street. We returned to houses that were unoccupied at the time of selection on up to three occasions. All consenting adult women were interviewed. All eligible women who declined to take part were noted as non-responders, as were women ineligible to take part due to age or illness. 52.4% to 89.4% of respondents who were approached in any village agreed to an interview, and the average response rate for the entire sample was 75% (total interviews ÷ [total responders + non-responders]). Our sampling strategy means that important variables such as age are approximately normally distributed in all groups. The sample should not be considered representative of the country as a whole, but of the wider rural population of this particular region.
As part of a wider survey interview, respondents were asked about their current and past use of 15 different contraceptive methods. These were then classified as either natural or artificial methods. Natural methods include calendar and sympto-thermal methods, withdrawal/coitus interruptus and periodic abstinence. Artificial methods include condoms, birth control pills, IUDs, post-coital or emergency contraception ('morning-after' pills), spermicides, sterilisation, hormonal implants and patches, and diaphragms (Table 1).
Respondents were asked to name up to five female social network partners with whom they could talk about "important personal matters". These alters could be kin or non-kin, geographically proximate or living far away, and respondents were encouraged to nominate only those most important to them rather than being forced to name five people. Respondents reported whether each nominated alter in their social network had ever used a natural and/or artificial contraceptive method, and whether a member of their family had ever used any methods of contraception. Where respondents did not know the contraceptive status of their kin or friends, or where they did not give information about them, it was assumed that these individuals had not used any method of contraception. There were almost no cases where a known dyad (i.e. two respondents who nominated each other) gave inconsistent information about each other's contraceptive use. Nonetheless, the data are likely to underestimate the number of kin and alters using contraceptives, and as a result our results should be considered conservative measures of the influence of alters on respondents' contraceptive status.
Table S1 details the observed frequency of kin and social network partners using either any or artificial methods of contraception.

Table S1
Description of the number of respondents who had kin and/or social network partners who used either any method or an artificial method of contraception.
The personal networks reported here are fairly well mixed between kin and non-kin, are characterized by long-term friendships with a high degree of interpersonal contact and are located in a relatively limited geographical area. Table S2 gives an overview of the main characteristics of these personal networks. On average, women have known their network partners for over 23 years (s.d. = 13.57), indicating that the data capture long-term rather than transient personal ties. Approximately 47% of all network partners were kin, with the remaining ~52% being non-kin. 49% of all alters originated in the same village that ego currently lives in, with 52% currently resident in the same village. Most of the remaining alters (~40%) were currently living in the local villages and towns, with fewer than 7% of all alters living in a city or foreign country. Frequency of both face-to-face and telephone/email contact was high among network partners, with over 43% of friends being physically visited and over 66% being contacted by other means (mainly telephone) on a daily basis. Any method Artificial method  mother used contraception  266  23  sister(s) used contraception  382  115  other kin used contraception  287  318   0 alters used contraception  1136  1436  1 alter used contraception  216  204  2 alters used contraception  204  109  3 alters used contraception  128  51  4 Table S4 gives a breakdown of completed fertility differences between users and non-users of the most popular contraceptive methods. Significant differences in completed fertility are mainly driven by use of sympto-thermal methods of contraception. Although the trend is for artificial contraceptive users to have fewer children than non-users of these particular methods, the differences are not statistically significant.

Table S4
Completed fertility of post-reproductive aged women (45+) broken down by method type.

Variables used in the analysis
We used a combination of direct and composite measures in our analyses (see [1] for further details). Education, material wealth, market integration and farming wealth were standardised to have a mean of 0 and s.d. of ~1. Education was measured as the highest level of formal education obtained by the respondent. This ranged from 1 = none/some primary (n = 21), 2 = full primary (n = 317), 3 = vocational (n = 523), 4 = secondary (n = 806) to 5 = tertiary (n = 305).
Household material wealth is a weighted linear combination of six variables, created using Principal Component Analysis (PCA): mean household income (measured across all adults in the household in brackets ranging from <600 Polish Zlotys [PLN] per month, to >2,500 PLN per month), ownership of a computer, an internet connection, a satellite TV, a car, as well as the number of habitable rooms in the house. This measure captures both income and assets, and covers the whole household rather than simply the parents or the husband of the respondent. The variables used to construct this measure show good inter-item scalability (Cronbachs α= 0.71), and the first Principal Component accounted for 42% of the variance in the original items.
Household market integration is a weighted composite measure capturing the extent to which the household is dependent on wage-labour versus farming income. For every householder, occupation (ranging from student to farmer to full-time employed), occupational status (if employed) and an indicator of employment history (ever employed) were weighted to provide a cumulative 'market integration' score. We took the household mean as the measure of overall market integration. The respondent is not included in this measure to avoid confounding co-linearity with the other measures, specifically individual education. We control for farming status (resident in a farming household at the time of the survey or not; n 'farmers' = 1,239, n 'non-farmers' = 733), experience of under-five mortality (n 'yes' = 65, n 'no' = 1,907) and age of the respondent (mean =44, s.d. = 17.83, range = 18-91). We also included age squared and age cubed to account for the non-linear relationship between fertility and age across the lifespan. 23 women who were not currently resident in any of the sampled communities were excluded from the analysis, leaving a total sample of 1,972.
Degree of religiosity was assessed using three variables; frequency of attendance at Mass, a four-point scale ranging from 'never' to 'more than once a week', self-identified religiosity, based on the question "would you say you were a religious person?" and coded as a fourpoint scale ranging from 'not religious at all' to 'very religious', and self-reported importance of religion in the person's life, based on the question "how important would you say religion is in your daily life?", a five-point scale ranging from 'not at all important' to 'very important'. These variables were combined into a single dimension using principal component analysis. The first component explains a large proportion (56%) of the variance in the original data and the variables show good inter-item scalability (Crohnbach's α = 0.61). Our 'religiosity' score has a mean of 0 and a s.d. of 1.
'Social media exposure' is a principal component, made up of three variables indicating personal ownership of a mobile phone and regular use of computers and the internet (i.e. independent of ownership of a computer). Again the variables show good inter-item scalability (Cronbach's α = 0.83) and the dimension explains a large proportion of the variation in the original data (~75%). The score has a mean of 0 and a s.d. of 1. Table S5 gives a correlation matrix of the community level predictors considered in the multilevel models. The table shows that there is low-to-moderate inter-correlation between the measures. There is a moderate correlation between the presence of a church and a health centre (Pearson's R = 0.49, p < 0.001), most likely due to some villages having more amenities overall than others. Note that average religiosity is not highly correlated with the presence of a church in the village.
We control for the presence of a health centre and/or a church in the community in our models because these are proxies both for access to contraceptives and for information on the social, financial and physiological costs and benefits to using artificial methods. This allows us to better estimate whether the effects of average religiosity and education are the result of unobserved aspects of shared environment or cultural norms. Twelve of the 22 study communities contain a church and four communities contain a public health centre, with the main hospital (where most children are born) being in the town. We group-mean centred individual education and religiosity to allow the reintroduction of the uncorrelated group means (average community education and religiosity) later in the analysis.

Statistical analysis
The multilevel logistic regression models estimate the change in the log-odds of having ever used either method based on a unit change in the predictors (Tables S6-S8). All analyses were run in the R environment, version 3.1.1 [2], using the 'arm' [3], 'lme4' [4], 'blme' [5] and 'ggplot2' [6] packages. Model comparison was carried out using the 'MuMIn' package [7].
To avoid boundary problems associated with small variance parameters and a small sample of groups, including incorrect or underestimated uncertainty in the model parameters and covariance matrices, we used maximum penalized likelihood (MPL) [8,9], with the Laplace approximation, to estimate our models. MPL uses 'weakly informative' or 'diffuse' priors to obtain Bayesian modal estimates on the parameters in the model. This ensures that the variance component estimates remain off the boundary of the feasible parameter space (zero), but with weak enough priors so that inferences remain consistent with the data. This method extends standard multilevel modelling techniques without requiring simulation as in fully Bayesian analysis (which obtains posterior mean estimates) and outperforms standard maximum likelihood methods.
We did not include controls for age at first birth or length of first inter-birth interval as this results in the exclusion of women who have not yet had any children. We included unmarried women in the analysis because a sizeable proportion of unmarried women have used artificial methods of contraception (n = 125, 24% of artificial contraceptive users). Our models included only those predictors that have been shown to be salient in previous demographic and evolutionary research on contraceptive uptake. ! !    In the social network and kin models, we controlled for structure and composition of the social network (i.e. size and proportion of kin), because these are assumed to be sources of bias in the estimation of social network effects on contraceptive behaviour [10], but we did not control for network density, as this results in the exclusion of women with only one or no network partners, reducing the sample by over one quarter to 1,516. Since this means comparisons between women with and without network partners cannot be made it biases the sample towards women with at least two network partners. We ran additional analyses to check whether our social network effects would hold among the subset of women for whom we could calculate network density (n = 1,516), but this predictor was neither significant nor affected the model outcomes (Table S7) We also examined whether our network results were driven by homophily along educational lines, which would indicate that clustered contraceptive behaviour is driven by educational similarities among network partners, rather than being the result of social transmission among network partners. However, including average network education (which reduced the sample to 1,773) as an additional control had no effect on the model outcomes (Table S8).

Correlation matrix of community level variables considered in the analysis
We re-ran our model comparisons using a different information criterion, AICc, to see if this affects the model comparison outcomes (Tables S9 and S10). We also examined model comparisons where we first 'dredged' the variables within model classes, i.e. reduced each model to its most parsimonious form before comparison (Tables S11 and S12). All criteria point to the same models being the best approximating models given the data.
As with all cross-sectional analysis, we cannot establish the direction of causality in the models presented here and it is possible that other unobserved factors are important in driving contraceptive uptake. Our models show that contraceptive use is strongly clustered in families and social networks, however we cannot rule out the possibility that individuals simultaneously adopted contraception without making use of any social information from network partners. Longitudinal data are the gold standard for testing hypotheses about the timing of contraceptive uptake and inferring the order in which individuals may influence each other, and since many of the variables used in this analysis could be transformed into a person-period format suitable for such an analysis, a future goal is to replicate this study using a time-to-event approach.  S10 Model comparison and ranking estimates based on all possible combinations of model classes (32 possibilities) using AIC instead of DIC.

Table S11
Model comparison and ranking estimates based on model classes that have undergone a first round of within-class model selection, i.e. when each model has been reduced to the most parsimonious set of variables.

Table S12
Model comparison and ranking estimates based on all possible combinations of model classes (32 possibilities) that have undergone a first round of within-class model selection, i.e. when each model has been reduced to the most parsimonious set of variables.