Population contextual associations with heterosexual partner numbers: a multilevel analysis

Abstract

Objective

The study examines whether an individual's sexual behaviour is associated with the demographic context within which they live.

Methods

Data from a large behavioural survey were matched to the census and the number of opposite sex partners individuals reported having in the year before interview was modelled against a suite of individual characteristics and analogous characteristics for the population in which they lived.

Results

The number of partners reported (none, one, two, three, or more) were variously associated with an individual's gender, age, marital status, sexual identity, and same sex activity in the previous year. Additionally, population age structure, sex ratio, and the proportion of the population reporting specific patterns of sexual activity were associated with the behaviour of individuals.

Conclusions

This study demonstrates that population context is associated with individual behaviour even after individual characteristics have been taken into account. This suggests that multilevel modelling of sexual behaviour data can provide new insights into the pattern of sexual behaviour.

It is well established that individual demographic and social attributes are associated with patterns of sexual activity. In Australia, for example, we know that men are more likely than women to report multiple opposite sex partners in the previous year and that people aged 20–49 were generally more likely to report multiple opposite sex partners than were people aged 16–19 and 50–59.¹,² Moreover, people who identified as bisexual rather than heterosexual or homosexual were more likely to report multiple opposite sex partners in the previous year as were people of higher socioeconomic status and those who were unmarried.¹ These patterns of sexual behaviour appear consistently in Western countries including Britain,³,⁴ Canada,⁵ the United States,⁶ and a range of European countries.⁷

While there has been considerable interest in understanding the patterning of sexual behaviour through social network analysis,⁸,⁹,¹⁰,¹¹,¹²,¹³ and to a less extent geographic analysis of sexually transmitted infections,¹⁴,¹⁵,¹⁶ there has been much less interest to date in understanding the potential contribution of population characteristics to the sexual practices of the individuals who live within those populations. Applications of multilevel analysis are increasingly common in many areas of health research such as smoking initiation,¹⁷ cardiovascular disease,¹⁸ excess mortality,¹⁹ and self rated health²⁰ but have a largely unexplored potential in understanding the patterning of sexual and reproductive health behaviours although the benefits of the approach have recently been stressed²¹ as have the difficulties.²²

Few population based multilevel studies linking context to behaviour have been published. One examined the prevalence of unprotected anal intercourse among 1016 bisexual men in nine areas in Ontario, Canada, and found an association between greater HIV prevention service delivery at the area level and less unprotected sex by individual men.²³ A second, a study of 915 adolescents in 80 Chicago neighbourhoods, found that the timing of the onset of sexual activity was associated with area level poverty even when accounting for a large range of individual characteristics of the adolescents.²⁴ This finding was replicated in a larger multilevel study employing nationally representative data on adolescents.²⁵

The present paper significantly extends this limited, but important, literature by exploring the individual and population contextual associations with sexual activity in a nationally representative sample of people aged 16–59. The aim of the present paper is to determine whether population contextual factors are associated with the number of opposite sex partners people report, an extremely important predictor of the risk of acquiring sexually transmitted infections including HIV.

Methods

Data sources and variables

The primary data for this study come from the Australian Study of Health and Relationships.²⁶ This was a computer assisted telephone interview of a stratified random sample of 19 307 Australians aged 16–59 conducted in 2001–2. The overall response rate was 73.1% and the data were weighted to the 2001 Australian census to account for the differential probability of selection due to the stratification.²⁶ A small subset of variables was selected from that study: the number of opposite sex partners in the year before interview (zero, one, two, three, or more); age; gender; legal marital status (married, separated/divorced/never married/widowed); current sexual identity (heterosexual, homosexual, bisexual); and whether they had reported any same sex partners in the previous year. These variables were chosen as they provide a basic description of the sexual contexts or opportunity structures of areas. Preliminary analyses indicated that they were much stronger predictors of behaviour than socioeconomic status and so individual and area level socioeconomic status was not explored in the models.

Participants were allocated to one of the 206 Australian statistical subdivisions (SSDs) based on their post code using the National Localities Index.²⁷ SSDs cover the entire country without gaps or overlap and are defined by the Australian Bureau of Statistics as “…socially and economically homogeneous regions characterised by identifiable links between the inhabitants. Moreover, in the non‐urban areas, an SSD is characterised by identifiable links between the economic units within the region, under the unifying influence of one or more major towns or cities.”²⁸(p 13) Where possible, variables were extracted from the 2001 Australian census that were exact analogues of those measured at the individual level. They were sex ratio; median age and the ratio of those aged over 65 to those aged under 16; and the percentage of the population aged 15 or more who were legally married. Current sexual identity is not measured in the census and a population level analogue was not derived. Two measures of the sexual environment of the SSDs were derived from the survey data: the percentage of the participants in each SSD who reported any homosexual contact in the year before interview; and the percentage of the participants in each SSD who reported more than five opposite sex partners in the year before interview.

The modelling was performed using MLwiN version 2.02.²⁹

Modelling strategy

The outcome being modelled was the number of opposite sex partners in the year before interview: zero, one, two, three or more. An ordered logistic model presumes that the difference between having no partner and one partner is equivalent to the difference between one partner and two partners to a constraint we thought was inappropriate to the data to we elected to use a multinomial model.

Having selected a multinomial logistic model, we specified the reference category as those people having one partner in the previous year as this was the most common. We then fitted an intercept only random intercept model—that is, we estimated the average likelihood of having zero, two, or three or more partners compared to one partner but allow that average likelihood to vary randomly at the SSD level. In essence, fitting such a model asks the question whether there is any SSD level variability in the outcome.

Having fitted the intercept only random intercept model we then included each of the individual level variables, appropriately parameterised. Finally, each of the SSD level variables were added.

Results

Of the original sample, 18 647 participants provided data for all of the variables in the present study, including a postcode that could be linked to an SSD. Of the 206 Australian SSDs, 200 were represented in the data. The number of participants in an SSD ranged between one and 999 with a median of 50 and mean of 93.23.

The distribution of partner numbers is shown in figure 1. The age groups most likely to report no partners were 16–19 year olds and 50–59 year olds whereas the age groups most likely to report two or more partners were the 20–29 year olds and the 30–39 year olds.

Figure 1 The distribution of opposite sex partner numbers by age group.

The results of the intercept only random intercept model are shown in table 1. Each of the three variances was significantly different from zero as judged by their confidence intervals and the joint test that all three variances were zero was highly significant (χ²₃  =  35.94, 0.001 > p). The variance was higher for no partners and for three partners suggesting that proportion of people reporting those partner numbers exhibited greater variation between SSDs than did the proportion reporting two partners.

Table 1 Results of the intercept only random intercept model.

Parameter	No partners		Two partners		Three or more partners
	Odds ratio	95% CI	Odds ratio	95% CI	Odds ratio	95% CI
Intercept	0.132	0.127 to 0.137	0.055	0.053 to 0.058	0.066	0.063 to 0.069
SSD variance	0.121	0.063 to 0.179	0.076	0.029 to 0.123	0.196	0.113 to 0.284

Having one partner is the reference category and is omitted.

The results of fitting the model including the individual level variables appear in table 2. Each of the individual level variables was highly significant when judged by the joint test that all of the parameters were equal to zero. The least significant was identifying as bisexual (χ²₃  =  15.88, p = 0.0012) and being separated, divorced, widowed, or never married the most significant (χ²₃  =  1677.96, p ≅ 0).

Table 2 Results of the random intercept model including individual level parameters.

Parameter	No partners		Two partners		Three or more partners
	Odds ratio	95% CI	Odds ratio	95% CI	Odds ratio	95% CI
Age	1.033	1.029 to 1.036	0.989	0.983 to 0.995	0.969	0.965 to 0.974
Age2§	1.051	1.051 to 1.052	1.000	0.999 to 1.001	1.001	1.000 to 1.001
Gender
Male	1
Female	1.141	1.069 to 1.218	0.677	0.615 to 0.745	0.358	0.325 to 0.394
Marital status
Married	1
Separated/divorced/widowed/never married	46.960	41.650 to 52.947	12.137	10.696 to 13.773	19.858	16.995 to 23.203
Current sexual identity
Heterosexual	1
Homosexual	85.004	59.151 to 122.156	0.241	0.096 to 0.600	0.175	0.064 to 0.485
Bisexual	1.685	1.290 to 2.201	1.421	0.934 to 2.163	2.027	1.455 to 2.824
Same sex activity in last year
None	1
Some	1.715	1.298 to 2.266	3.499	2.304 to 5.313	4.227	2.858 to 6.251
Intercept	0.001	0.001 to 0.001	0.019	0.014 to 0.025	0.031	0.025 to 0.040
SSD variance	0.097	0.052 to 0.143	0.033	0.000 to 0.066	0.115	0.063 to 0.167

§Calculated as the square of the difference between age and the mean age of the entire sample

Having one partner is the reference category and is omitted.

Generally, the likelihood of having higher numbers of opposite sex partners declined linearly with age. The age² term is significant only for having no partners indicating that the likelihood of having no partners is higher at the end of the age range in the data. Women were more likely than men to report no partners and less likely to report more than one. People who were separated, divorced, widowed, or never married were more likely to report no partners and more than one partner. Those who identified as homosexual rather than heterosexual were much more likely to report no opposite sex partner in the previous year and much less likely to report more than one opposite sex partner. People who identified as bisexual rather than heterosexual were more likely to report no opposite sex partner in the previous year and more likely to report more than one opposite sex partner, although the confidence interval for having two rather than one opposite sex partner overlapped with one. Having accounted for current sexual identity in the model, those participants who reported any same sex contact in the previous year were more likely to report no opposite sex partners in the previous year and more likely to report more than one opposite sex partner. The variances at the SSD level remained significantly different from zero as judged by their confidence intervals and the joint test that all three variances were zero was highly significant (χ²₃  =  37.90, 0.001 > p). They were reduced by between a quarter and a half in magnitude through the inclusion of the individual level variables.

The inclusion of the SSD level variables had little impact on the parameter estimates for the individual level variables (table 3). While at least one of each of the SSD level parameters did not overlap with one, the overall significance of the variables as judged by the joint test had a p value below 0.05 for five: median age (χ²₃  =  9.16, 0.05 > p > 0.01); ratio of those over 65 to those under 16 (χ²₃  = 7.98, 0.05 > p > 0.01); sex ratio (χ²₃  =  12.62, 0.01 > p > 0.001); the percentage of the population reporting any same sex contact in the previous year (χ²₃  =  11.08, 0.01 > p > 0.001); and the percentage of the population reporting more than five opposite sex partners in the previous year (χ²₃  =  35.77, 0.001 > p).

Table 3 Results of the random intercept model including individual level and SSD level parameters.

Parameter	No partners		Two partners		Three or more partners
	Odds ratio	95% CI	Odds ratio	95% CI	Odds ratio	95% CI
Individual level parameters
Age	1.032	1.029 to 1.036	0.989	0.983 to 0.995	0.970	0.965 to 0.974
Age2§	1.005	1.005 to 1.005	1.000	1.000 to 1.001	1.001	1.000 to 1.001
Gender
Male	1
Female	1.141	1.069 to 1.218	0.676	0.614 to 0.745	0.359	0.327 to 0.395
Marital status
Married	1
Separated/divorced/ widowed/never married	47.068	41.754 to 53.059	12.089	10.658 to 13.712	19.402	16.621 to 22.649
Current sexual identity
Heterosexual	1
Homosexual	85.030	59.169 to 122.193	0.234	0.094 to 0.587	0.174	0.064 to 0.473
Bisexual	1.665	1.275 to 2.174	1.426	0.935 to 2.176	2.018	1.473 to 2.764
Same sex activity in last year
None	1
Some	1.703	1.288 to 2.252	3.448	2.254 to 5.275	4.155	2.830 to 6.101
SSD level parameters
Median age	1.003	0.979 to 1.029	0.991	0.968 to 1.016	1.066	1.043 to 1.090
Ratio of over 59 to under 16 (%)	1.001	1.000 to 1.001	1.000	0.999 to 1.001	1.002	1.001 to 1.003
Sex ratio (%male)	0.917	0.894 to 0.941	0.981	0.955 to 1.007	0.954	0.925 to 0.984
Separated/divorced/widowed/ never married (%)	0.987	0.977 to 0.997	0.998	0.987 to 1.009	1.026	1.015 to 1.037
Any same sex activity in last year (%)	1.056	1.038 to 1.075	1.024	1.008 to 1.040	1.002	0.987 to 1.018
More than five opposite sex partners in last year (%)	1.009	0.985 to 1.033	1.005	0.978 to 1.032	1.123	1.101 to 1.146
Intercept	0.088	0.016 to 0.491	0.069	0.011 to 0.436	0.006	0.001 to 0.041
SSD variance	0.096	0.049 to 0.144	0.040	0.003 to 0.0760	0.056	0.014 to 0.097

§Calculated as the square of the difference between age and the mean age of the entire sample

Having one partner is the reference category and is omitted.

In the model for having no sexual partner rather than one sexual partner in the previous year, three population contextual variables were significant (table 3). Those people living in an SSD with a higher proportion of men had a reduced likelihood of having no sexual partners in the previous year. Having a higher percentage of the population separated, divorced, widowed, or never married decreased the likelihood of reporting no opposite sex partners in the previous year, whereas a higher percentage of the population reporting same sex activity in the previous year was associated with a higher likelihood of reporting no opposite sex partners in the previous year.

In the model for having two sexual partners rather than one in the previous year a higher percentage of the population reporting same sex activity in the previous year was associated with a higher likelihood of reporting two opposite sex partners in the previous year. In the model for having three or more sexual partners rather than one in the previous year the likelihood of reporting three or more sexual partners increased with the median age of the SSD and the ratio of those aged over 65 to those aged under 16. It decreased with an increasing sex ratio, increased with the percentage of the SSD that were separated, divorced, widowed or never married, and also with the percentage of the population reporting more than five sexual partners in the previous year.

Each of the three variance remained significantly different from zero as judged by their confidence intervals and the joint test that all three variances were zero was significant (χ²₃  =  16.09, 0.05 > p > 0.001). The variance around having no partners decreased slightly, whereas that for two partners was increased slightly while that for three or more partners was almost halved. This is to be expected given the number and magnitude of SSD level variables influencing those particular outcomes.

Discussion

In this paper we have offered an alternative to more traditional ways of thinking about the distribution of sexual behaviour and sexual risk at the population level. In particular, we have demonstrated the demographic and behavioural context in which people live is associated with the behaviour of individuals even once we have taken into account those same demographic and behavioural variables as they pertain to the individuals themselves.

We have shown that individuals' demographic and behavioural characteristics relate to the number of opposite sex partners they reported in the previous year in a manner consistent with previous studies.¹,²,³,⁴,⁵,⁶,⁷ Having taken those factors into account, we have shown that population level demographic variables are associated with whether people reported none, one, two, or three or more sexual partners in the previous year. Moreover, different population level demographic variables are associated with different reported partner numbers.

Employing a multinomial model allows the emergence of quite complicated patterns of association even in these models, which use only a handful of individual level parameters and their population level analogues. The population level parameters appear to have their strongest and most numerous associations with the difference between having one partner versus three or more opposite sex partners in the previous year. They are weakest in the difference between having one partner versus two opposite sex partners in the previous year. This suggests that uncommon behaviours are particularly dependent on enabling contexts rather than individual desire.

This study is not without its limitations. Firstly, locality was measured by postcode and then mapped to SSD. Postcodes do not map exactly to SSDs and so there is likely to have been some misallocation. However, a recent study has shown that compared to the smallest areal unit at which census data are available (the census collector district), larger areal aggregations such as postcodes reliably capture the relation between variables such as socioeconomic status and health, although the greater the level of aggregation the weaker the strength of those associations.³⁰ In part, we chose to use SSDs because the higher the level of aggregation the lower the likelihood of misallocation. Secondly, we relied on individual level data from the Australian Study of Health and Relationships, which used self report. The validity, reliability, and representativeness of those data have been extensively explored elsewhere and we are confident that they are robust.²⁶ Thirdly, these data are cross sectional and hence the direction of causality cannot be established. Fourthly, in the interests of interpretability we included gender in the model rather than running separate models for women and men. This may have constrained the emergence of gender differences. Finally, we used legal marital status rather than relationship status as legal marital status is available in the census data. This may have led to an overstatement of the importance of legal marital status.

Key messages

• Individual sexual behaviour is associated with the population context in which the individual lives

• Less common behaviours may be more strongly associated with context than are more common behaviours

• Population characteristics may provide guidance for policy makers in identifying sites or areas for intervention

The models presented in this paper are relatively simple. Further modelling needs to be undertaken to explore cross level interactions—for example, whether the SSD sex ratio is more important in the behaviour of men than of women. Also, we need to explore more complex variance structures to examine—for example, whether there is greater variation in women's behaviour among different SSDs than in men's behaviour among different. Regardless, the potential of multilevel modelling for exploring population contextual influences on individual sexual behaviour is clearly established.

Abbreviations

SSDs - statistical subdivisions