Population Studies, Vol. 59, No. 3, 2005, pp.

265 /282

Social networks, perceptions of risk, and changing

attitudes towards HIV/AIDS: New evidence from a
longitudinal study using fixed-effects analysis

Stéphane Helleringer and Hans-Peter Kohler

University of Pennsylvania

The study presented here is an investigation of the importance of social interactions to perceptions of the
risk of AIDS, and explores spousal communication about the AIDS epidemic in rural Malawi. A fixed-
effects analysis based on longitudinal data collected in 1998 and 2001 shows that social interactions on the
subject of HIV/AIDS have significant and substantial effects on respondents’ perceptions of the risk of
HIV/AIDS, even after controlling for unobserved factors that affect the selection of social networks. These
effects are more complex than previously thought. The dominant mechanisms */social learning and social
influence */are found to vary by sex and by region, because of regional variations in the marriage pattern
and the resulting implications for the formation of local social networks. The conclusion of the study is that
rather than fostering denial and inaction, social interactions are an important vector of change in the face of
the HIV/AIDS epidemic.

Keywords: HIV/AIDS; social interactions; risk perceptions; Malawi; unobserved heterogeneity;

longitudinal studies

[Submitted June 2004; Final version accepted April 2005]

It is well known that AIDS is a major cause of adult Blaikie 1992; Orubuloye et al. 1997). Moreover,
mortality in sub-Saharan Africa. Initially, research- fatalism and a lack of perceived self-efficacy also
ers attributed the spread of HIV primarily to a lack appear as important elements in the failure to
of correct information, or to myths and rumours control the AIDS epidemic (Campbell 2003). The
disseminated informally (Bond and Dover 1997). belief that ‘death comes when it is due’ is said to be
Now, however, the conventional view is that the widespread. Caldwell (2000, p. 124) argues that men
severity of the AIDS epidemic in Africa is due to involved in high-risk practices tend to justify their
social and cultural ‘barriers’ that inhibit behavioural behaviour in terms of competing risks: ‘if the latency
change in spite of widespread knowledge about the period for AIDS is a decade, then they are not
fatal consequences of the disease and about recom- worried because they are likely to die of something
mended methods of prevention. Such barriers are else in such a long time’.
thought to include the following: a political economy According to Caldwell (2000), such beliefs and
that stimulates labour migration, which in turn judgements are actively maintained by varied forms
separates spouses (Setel 1999); patterns of breast- of social interaction, and individuals may experience
feeding and postnatal abstinence that make women pressure from peers to engage in risky behaviour.
unavailable for sexual relations with their husbands Much has also been said about mechanisms of taboo,
for extended periods of time (Caldwell et al. 1989; stigma, and denial that generate an ‘extraordinary
Caldwell et al. 1992); and other beliefs that are said silence around the causes of the disease’ and prevent
to inhibit change */for example, that ‘males are open discussion of AIDS (Nzioka 1996; Caldwell
biologically programmed to require sex with more 2000). However, the view that social interactions
than one woman’, that sexual intercourse with a constrain the adoption of new and safer behaviour in
condom is not ‘sweet’, or that it is necessary to response to the epidemic in Africa has recently been
preserve risky circumcision rituals (Barnett and questioned. Low-Beer and Stoneburner (2004) argue

ISSN 0032-4728 print/ISSN 1477-4747 online/05/030265-18 # 2005 Population Investigation Committee

DOI: 10.1080/00324720500212230
266 Stéphane Helleringer and Hans-Peter Kohler

that in Uganda, the only country in sub-Saharan response to this uncertainty, individuals have turned
Africa that has experienced a dramatic decline in to others for help in evaluating the risks and benefits
HIV prevalence, social interactions have facilitated associated with the ‘new’ practices. For example,
the adoption of new behaviour. In particular, more analyses of both qualitative and aggregate data from
survey respondents in Uganda than in neighbouring Thailand and Kenya show that women have informal
countries reported personal channels of communica- discussions among themselves about family planning
tion about AIDS and more reported knowing some- and family size (Montgomery and Casterline 1993;
one with AIDS or who had died of AIDS. The Entwisle et al. 1996; Rutenberg and Watkins 1997;
authors suggest that communication may have pro- Watkins 2000), and that these informal conversa-
vided greater personal exposure to the fear-evoking tions have influenced attitudes and behaviour relat-
consequences of the epidemic, and thus catalysed the ing to fertility (e.g., Entwisle et al. 1996;
process of behaviour change. Others have reached Montgomery and Chung 1998; Kohler et al. 2001;
similar conclusions using qualitative data. Setel Munshi and Myaux 2002). Even in countries like
(1999, p. 185), drawing on participant observation Italy where fertility is very low, social interactions
in Tanzania to trace the construction of local knowl- appear to be influential (Bernardi 2003).
edge about HIV, finds that the process ‘required While there is less evidence for the role of social
nothing other than an ability to engage in collective interactions in shaping responses to heightened
creative thought about a disease that remained a mortality from AIDS than there is for the new
chimerical presence for most, long after scientific methods of fertility control, it seems reasonable to
authorities had announced its epidemiologic arrival’. expect social interactions to be at least as important
Watkins (2004) has used data from observational in shaping responses in this arena, and perhaps more
field journals to predict that the epidemic in Ma- so. Again, there is great uncertainty surrounding
lawi */and possibly also other sub-Saharan African HIV/AIDS, since the ability of individuals to learn
countries */is at a turning point. The journals in- directly from the deaths of friends, relatives, and
dicate that social interactions have played an essen- neighbours is constrained by several factors. First,
tial role in forming individuals’ perceptions of their the long latency period of the disease makes it
risk of HIV/AIDS and their evaluation of the almost impossible to link symptoms and eventual
benefits (and costs) of behavioural changes such as death to particular sexual encounters or to risky
the adoption of condom use, increased marital behaviour that took place as long as a decade, or
fidelity, or more careful selection of sexual partners. even more, before death. Second, it is likely that
The present study makes use of a unique long- many infections occur during the migration of men
itudinal set of data collected in rural Malawi since beyond their villages, and during premarital or
1998 to investigate the effects of social interactions extramarital relationships that the participants at-
on perceptions of risk and on the attitudes of tempt to keep secret. Nonetheless, men brag about
individuals towards AIDS. In particular we ask their affairs, women confide in friends (Tawfik
whether social interactions play an active role in 2003), and those infected with a sexually transmitted
maintaining beliefs that sustain high-risk practices, disease seek treatment advice from others. Thus, as
or whether they foster and encourage attitudes Watkins (2004, p. 673) has shown, when someone in
conducive to behavioural change. a village dies, members of social networks contribute
to the diagnosis of the cause of death by contributing
fragments of local knowledge. In the same networks,
Background men and women are ‘collectively navigating AIDS’:
they ‘evaluate the sources of risk and varieties of
A growing literature on the diffusion of family sexual pleasure, debate global and national prescrip-
planning in less developed countries provides evi- tions and formulate local and sometimes innovative
dence that social interactions are an important strategies of prevention’ following conversations
determinant of the adoption of new demographic with relatives, friends, neighbours, acquaintances,
behaviour (Bongaarts and Watkins 1996; Casterline and even strangers.
2001; Kohler 2000, 2001; Watkins et al. 2003). The Despite the presumption that social interactions
key theoretical idea driving these empirical investi- influence perceptions of the risk of AIDS and the
gations is that the global promotion of fertility adoption of prevention strategies, it is difficult to
control has provoked considerable uncertainty about make convincing causal inferences from either
the risks of adopting the small-family model and qualitative or survey data (e.g., Manski 1995; Behr-
new methods of family planning, and that, in man et al. 2002). First, individuals’ engagement in
 Social interactions and AIDS in rural Malawi 267

social interactions is often determined by their own influence allows us to test Caldwell’s hypothesis
attitudes and preferences. Social-network partners about the role of social interactions in the African
are therefore not chosen randomly from the popula- AIDS epidemic and their importance in maintaining
tion, but are selected systematically according cultural ‘barriers’ that prevent behavioural change.
to observed and unobserved characteristics, a pro- For example, whereas sexual intercourse outside
cess often resulting in homophily (social interactions marriage is widely practised in sub-Saharan Africa
among persons sharing common characteristics). because of long durations of postnatal abstinence
Research designs that do not address this selection and high levels of polygyny, Caldwell (2000, p. 127)
issue are likely to misrepresent the role and argues that ‘wives are not supposed to recognize
relevance of social interactions. Second, hardly their husbands’ extramarital activities and the pre-
any studies have tried to identify the underlying tence is well maintained’. According to this point of
mechanisms that link social interactions and view, the risk perceptions of women, or the possibi-
individual behaviour in the context of HIV/AIDS lity of spousal communication about AIDS-related
(the few that have done so include Behrman et al. matters, are constrained by the actions and judge-
2004). ments of relatives and other relevant reference
Theories of social interaction and demographic groups. Evidence for social learning about risk
change usually distinguish between two processes perceptions or attitudes towards prevention strate-
through which the social environment exerts an gies would thus conflict with this characterization of
influence on individuals’ behaviour or attitudes: the social environment and its consequences for
social learning and social influence (Montgomery individual attitudes; indeed, when social interactions
and Casterline 1996; see also Axinn and Yabiku operate through learning, social networks do not
2001; Kohler et al. 2001; Behrman et al. 2002; constrain but rather represent a valuable tool that
Kravdal 2002; Moursund and Kravdal 2003). We individuals use to learn about and cope with the
use the term ‘social influence’ to refer to the uncertainty of their environments. Evidence for
processes by which preferences for sexual behaviour, social influence, on the other hand, would provide
gender relations, or other behaviour relevant to some support for Caldwell’s view.
AIDS are potentially affected by the opinions and
attitudes that prevail in an individual’s social envir-
onment. These processes may lead individuals to Data and context
change their preferences in response to normative
pressures after interactions with others about the The analysis presented here is based on data from
threat of AIDS */for example, if network partners the Malawi Diffusion and Ideational Change Project
express their disapproval of a man having sexual (MDICP). The project’s goal is to evaluate and
intercourse with sex workers on the ground that it analyse the role of informal conversations in chan-
exposes himself and his spouse to the risk of HIV ging behaviour and attitudes towards family plan-
infection. On the other hand, ‘social learning’ is seen ning and HIV/AIDS in rural Malawi, and the data
in individuals’ use of social interactions to gather collected as part of the project consist of a long-
information that helps them cope with the uncer- itudinal household survey and a set of semi-struc-
tainties brought about by changing environments, tured interviews, focus groups, and observational
such as those arising from a new epidemic like HIV/ field journals. The panel survey was initiated in 1998
AIDS. Thus network partners may be sources of with 1,541 women of childbearing age who had ever
information about, for example, modes of transmis- married and 1,065 male spouses, and was conducted
sion of the disease, adequate prevention strategies, in three districts of Malawi: Rumphi (in the north),
or the relevant characteristics and partnership his- Mchinji (centre), and Balaka (south). A follow-up
tories of potential sexual partners. survey that re-interviewed 1,200 women and 841
The empirical distinction between the two me- men was conducted in 2001. A detailed description
chanisms */social learning and social influence */is of the survey, including aspects of data quality, is
relevant to the present study for several reasons. available in a special collection of papers in Demo-
First, because the existence of social interaction may graphic Research (Watkins et al. 2003).
modify the evaluation of the impact of AIDS-related Subsistence agriculture and small-scale retail out-
policy interventions, understanding the mechanisms lets are common in rural Malawi, but there is also
of social interaction could facilitate the optimal extensive wage labour, some of which requires
design and use of such policy interventions. Second, migration to urban areas. Although each of the
the distinction between social learning and social three sites of the survey has the foregoing features in
268 Stéphane Helleringer and Hans-Peter Kohler

Table 1 Socio-economic characteristics of respondents to a survey of the effects of social interactions on attitudes to
AIDS. Malawi 1998

Women Men

South/centre North South/centre North

Age 31.2 30.9 37.2 37.1

(9.5) (8.6) (10.1) (11.1)
Children ever born 4.4 4.3 5.4 5.1
(3.2) (2.8) (4.2) (4.2)
Proportion of respondents
Who ever used family planning 45.9 61.6* 58.4 62.5
With at least some primary education 47.2 96.8* 71.6 97.5*
With at least secondary education 1.7 10.6* 7.3 30.1*
Who speak English 0.6 10.6* 8.3 35.4*
With metal roof on house 5.2 10.8* 5.9 12.1*
Who own a radio 50.5 70.4* 63.5 76.2*
Who own a bicycle 55.7 43.4* 64.7 48.1*
Whose household has a pit latrine 66.1 88.8* 79.8 93.7*
Engaged in wage work 72.1 80.5* 98 95.8
In polygamous union 8.5 25.9* n/a n/a
N 814 386 575 240

*The difference between respondents in the south/centre and respondents in the north is significant at p B/0.05.
Note : Standard deviations are in parentheses.
Source : Malawi Diffusion and Ideational Change Project (MDICP).

common, there are also some differences (Table 1). respondent’s perception of the risk of AIDS. Inter-
For example, the north is wealthier, less densely estingly, as shown in Table 2, this level of worry
populated, and has a higher level of education than declined by 2001, especially in the south, perhaps
the centre or the south (probably as a result of the being associated with the adoption of prevention
early establishment of Christian missions in this strategies that the respondents perceived to be
area). Marriage patterns also differ across the effective (Smith and Watkins 2005). In addition to
regions. The Tumbuka ethnic group living in the recognizing the risk of infection, respondents were
north is typical of the ‘father-right’ societies de- also aware of different ways by which HIV/AIDS is
scribed by Gluckman (1950): inheritance is primarily transmitted, and most knew several methods of
patrilineal, and residence of spouses after marriage protection. Asked about the best way to protect
is patrilocal. Polygyny is widespread and the age themselves against AIDS, both men and women
difference between spouses is large (Reniers 2003). frequently cited marital fidelity as a strategy to
In contrast, the main ethnic group (Yao) in the south prevent HIV infection. When husbands and wives
follows a matrilineal system of inheritance, and a talked with each other about AIDS, they often
man usually moves into his spouse’s household after encouraged each other to ‘take care’, a euphemism
marriage. The population in the central district is for avoiding extramarital relations and other risky
mostly Chewa and also follows a matrilineal system behaviour (Zulu and Chepngeno 2003). In addition,
of descent, although less strictly than in the south. although attitudes towards condom use were gen-
Indeed, the Chewas have been affected by an influx erally negative, some men advised each other to use
of patrilineal groups from southern Africa (Ngonis) condoms with casual partners and criticized those
and from the coast (Swahilis), and residence after who did not.
marriage may thus be either matrilocal or patrilocal The epidemic in Malawi has also turned AIDS
(Phiri 1983; Reniers 2003). into an urgent topic of conversation, and discussions
Estimates of HIV prevalence often indicate that about HIV occur frequently and in a broad range of
close to 15 per cent of the Malawi population is contexts. In order to measure the extent of these
currently infected with HIV (UNAIDS 2004), and social interactions, the MDICP survey asked for
concerns about HIV/AIDS are widespread. For information on up to four persons with whom the
instance, in the first wave of the MDICP survey 61 respondent had ‘chatted’ about AIDS. These per-
per cent of the women reported being ‘very wor- sons are subsequently referred to as ‘network
ried’; this expression of worry is taken as the partners’, and the group that includes the respon-
 Social interactions and AIDS in rural Malawi 269

Table 2 Attitudes towards HIV/AIDS among respondents to a survey of the effects of social interactions. Malawi 1998 /
2001. All men and women interviewed in both waves

Women Men

Wave 1 Wave 2 Wave 1 Wave 2

Proportion of respondents worried about getting AIDS

Highly worried 61.5 47.0* 53.3 36.9*
A little worried 21.0 23.4 19.2 20.6
Not worried at all 17.5 28.6* 27.3 42.4*
Assessments of various HIV/AIDS risk-reduction strategies
Proportion of respondents who consider efficient:1
Spousal communication 46.6 37.8* 39.8 33.3*
Condom use with all partners except spouse 11.9 26.6* 17.4 40.8*
Marital fidelity 65.1 70.4 74.2 72.0
Proportion of respondents
Most worried about getting HIV from spouse 30.9 50.5* 13.5 23.2*
Consider condom use with spouse acceptable 14.2 28.5* 11.1 24.4*
AIDS prevention effort
Proportion of respondents who had heard of AIDS
At the clinic 85.7 95.5* 83.3 95.7*
On the radio 89.7 96 97.8 99.0
From CBD2 agent 23.6 35.9* 39.6 44.5
Proportion of respondents without AIDS network 16.1 5.2* 7.3 3.2*
N 1,200 1,200 815 815
1
The survey question asked: ‘What do you think are the best ways to protect yourself from AIDS?’ without prompting.
Multiple responses were allowed.
2
Community-based distribution.
*The difference between the two waves is significant at p B/0.05.
Source : As for Table 1.

dent and his or her four network partners is referred asked whether each of her/his network partners
to as his or her ‘social network’. The term ‘chat’ was knew each of the others, and how well. The MDICP
used in the interview to indicate to respondents that data thus include information about the ties among
the research team was primarily interested in all members of the respondent’s AIDS discussion
informal conversations rather than the lectures on network. Descriptive statistics for these networks
AIDS given at government clinics and elsewhere. are reported in Table 3.
The network data were collected by first asking the
respondents for the number of network partners
with whom they had chatted about AIDS (i.e., the Methods
uncensored size of the respondent’s AIDS network).
Between 1998 and 2001, the percentage of female While there are almost no quantitative studies of the
respondents who reported having had such discus- impact of social networks on the diffusion of AIDS-
sions with at least one network partner increased related behaviour or attitudes (one exception is
from 85 to 95 per cent. After the respondents had Behrman et al. 2003), there are a number of studies
listed all network partners with whom they had of fertility-related behaviour that are based on
chatted about AIDS, they were asked a series of empirical models of the form
questions about up to four of these partners (the Y  aX bdY o (1)
censored size of the network). These questions
included the relationship to the respondent (co- where X is a vector of observed covariates, Y
wife, sibling, etc.), residence, and degree of closeness represents the average outcome in a reference group
(confidant, friend, acquaintance), as well as educa- (e.g., social network, village, sampling cluster), and d
tion, condom usage, and worry about HIV/AIDS is interpreted as the effect of social interactions on
and current likelihood of HIV infection (using the the outcome variable Y.
wording ‘How worried is [name of network partner] An important limitation of relation (1) is that it
about AIDS?’). In addition, the respondent was does not allow for the possibility that the effects of
 270 Stéphane Helleringer and Hans-Peter Kohler
Table 3 Characteristics of social networks, by region and sex, of respondents to a survey of the effects of social interactions on attitudes to HIV/AIDS. Malawi 1998 /2001.
Respondents with 2/ network partners in both waves

Women Men

South/centre North South/centre North

Wave 1 Wave 2 Wave 1 Wave 2 Wave 1 Wave 2 Wave 1 Wave 2

(Uncensored) size of network 5.77 6.08$ 5.32 7.58* 7.44$ 7.34$ 6.02 7.95*
(5.31) (5.04) (5.30) (7.06) (6.65) (6.65) (6.25) (8.09)
Density
Average density of links among network partners 0.78$ 0.87* 0.83 0.88* 0.78$ 0.86* 0.85* 0.9
(0.33) (0.27) (0.29) (0.24) (0.33) (0.24) (0.27) (0.22)
Average proportions of network partners who are
Males 0.06 0.06 0.04 0.05 0.95 0.94 0.93 0.94
Females 0.94 0.94 0.96 0.95 0.05 0.06 0.07 0.06
Male relatives of respondent 0.02 0.02 0.01 0.04* 0.17$ 0.16$ 0.27 0.28
Female relatives of respondent 0.29$ 0.25$ 0.42 0.39 0.01 0.01 0.02 0.01
Members of the spouse’s family (in-laws) 0.09$ 0.10$ 0.21 0.21 0.05 0.05 0.06 0.04*
Relationship of the network partners to the respondent
Per cent confidants 0.21 0.20$ 0.23 0.25 0.24 0.23 0.25 0.23
Per cent friends 0.53 0.57 0.55 0.59 0.53 0.57$ 0.51 0.63*
Per cent acquaintances 0.17 0.15$ 0.18 0.10* 0.14 0.13$ 0.17 0.09*
Per cent met just once 0.09$ 0.08 0.04 0.06 0.09 0.07 0.07 0.05
Average proportion of network partners who live
On the same compound 0.07$ 0.13* 0.17 0.15 0.03$ 0.06$ 0.10 0.10
In the same village (includes network partners living on same compound) 0.52 0.58* 0.56 0.60 0.40 0.47$ 0.47 0.53*
In the same district (includes previous categories) 0.91$ 0.94 0.95 0.93 0.84$ 0.91* 0.92 0.92
Elsewhere (e.g., in the city, abroad, etc.) 0.09$ 0.06* 0.05 0.07 0.16$ 0.09* 0.08 0.08
N 548 256 548 256 462 462 186 186

*The difference between waves is significant at p B/0.05.

$The difference between respondents in the south/centre and respondents in the north is significant at p B/0.05.
Note : Standard deviations are in parentheses.
Source : As for Table 1.
 Social interactions and AIDS in rural Malawi 271

social interactions may be modulated by the struc- The information in the MDICP data about the
ture of social networks. For instance, the effects may relationships of the respondent’s network partners
differ depending on whether network partners among themselves allows us to infer whether, and
represent a densely knit group of friends or family through which processes, networks affect respon-
members, or a relatively loose group of individuals dents’ risk perceptions and attitudes towards AIDS.
who may not even know each other. Moreover, the Consider for instance the social network in Figure 1,
effect of the network structure is likely to vary depicting a respondent who has ties to four network
fundamentally with the kind of mechanism through partners. The MDICP data contain information on
which social interactions affect individual behaviour whether A, B, C, and D know each other directly, or
and attitudes. This observation has previously been only indirectly through their relation to the respon-
used to identify patterns of social learning and social dent. Density is the formal concept used to represent
influence in the diffusion of family planning meth- the degree of social connectedness among the net-
ods in Kenya (Kohler et al. 2001), and the present work partners. It is defined as the proportion of all
study extends this approach to the diffusion of possible social links that actually exist in the
AIDS-related behaviour. observed network. In a network of four, for exam-
In situations where it is crucial to learn about ple, a total of six possible links exists; the density is
opportunities, new practices, or potential health the number of existing links divided by six. We will
threats, individuals can ‘search’ their networks for call a network very dense if the density equals 1. We
the relevant information. For instance, a woman who will designate it as very sparse if the density is 0, that
suspects her husband of infidelity will ask her net- is, if A, B, C, and D do not share any direct links.
work partners if they know or have heard something Social learning and social influence can be dis-
(rumours, gossip, etc.) about her spouse’s where- tinguished from each other by using the distinct
abouts. The implications of network structure for this implications of social networks with different den-
learning process are complex: Granovetter (1973), sities (Kohler et al. 2001). If social learning dom-
and also Burt (1992), have argued that in situations inates, the effect of density should be small; this is
where learning is the main focus, tightly knit network because both sparse and dense networks can be
structures can significantly reduce information flows. ‘searched’ for information (Watts et al. 2002). It
Indeed, if the network partners all know each other, should thus be expected that the proportion of
they may provide redundant information. network partners very worried about AIDS should
On the other hand, networks may also impose be the main determinant of the probability of the
constraints on individual attitudes and behaviour, respondent reporting being very worried. If social
and the implications of network structure for pro- influence dominates, on the other hand, a different
cesses of social influence have been well identified. pattern should be expected: when a respondent’s
Indeed, tightly knit structures such as ‘cliques’ exert network is sparse, variations in the proportion of
a stronger normative influence on an individual’s network partners who are very worried should exert
behaviour and attitudes (Moscovici 1985; Burt 1992), only small effects on the respondent’s probability of
and are thus more efficient at defining what con- being very worried, whereas these variations should
stitutes a socially acceptable attitude or behaviour. be associated quite strongly with a woman’s risk
perception in dense networks.
An important concern in identifying these net-
B C
work influences is that estimates of the effects of
social interactions are potentially subject to omitted-
variable bias. In particular, both the attitudes of the
A D members of a respondent’s social network and the
structure within which these interactions occur, are
likely to be correlated with unobserved factors that
R
also determine the respondent’s risk perception.
R Respondent
Behrman et al. (2002) point to two reasons why
this might occur. The first is homophily: respondents
Network partner with high
AIDS risk perception interact primarily with others much like
themselves*/people who share some of their char-
Network partner with moderate
AIDS risk perception acteristics or attitudes. For example, in rural Malawi,
because of transportation and communication con-
Figure 1 Social network with four network partners straints, interaction occurs mostly with those who
272 Stéphane Helleringer and Hans-Peter Kohler

live in the same village, or in the villages nearby, thus of primary importance. Coefficient d1 would there-
constraining the networks towards higher homoge- fore be expected to be quite large, positive, and
neity and density. The second type of selectivity is significant, with d2 and d3 of little importance in
strategic: respondents may deliberately choose net- explaining variations in risk perceptions. On the
work partners who are potentially valuable sources other hand, when social influence is the most
of information about the risk of infection, preven- relevant mechanism in social interactions, d2 and
tion strategies, or similar AIDS-related topics. d3 would be expected to be the coefficients of
The longitudinal nature of the MDICP data, in primary importance, since sparse networks should
combination with fixed-effects estimation, allows the have little effect on the respondents’ attitudes while
analyst to control for observed and unobserved variation in the network partners’ risk assessments
heterogeneity among MDICP respondents that af- in dense networks would be likely to exert strong
fects the selection of social networks. In particular, pressure towards conformity of attitudes.
by using fixed-effects estimation it is effectively This analysis of social-interaction effects using the
possible to control for all time-invariant character- fixed-effects logit model above focuses on two
istics of respondents and their communities. To independent variables: the respondent’s perception
illustrate this approach to the analysis of AIDS- of the risk of AIDS, and the respondent’s assessment
risk perceptions, we posit a logit model for the of spousal communication about AIDS, in particular
respondent’s subjective HIV-risk perception that whether the respondent thinks an effective strategy
depends on social networks, observed individual for preventing AIDS is to advise the spouse to ‘take
characteristics, and unobserved fixed effects as care’ and avoid extramarital relations and other
Yit  Xit bd1 ×Nit d2 ×densityit d3 ×Nit risky behaviour. This variable is given as 1 if the
respondent gave appropriate answers to the open-
×densityit fi o it ended question ‘What do you think are the best ways
to protect yourself from getting AIDS?’ (multiple
Yit  1 if Yit  0 and zero otherwise (2)
responses were allowed), and 0 otherwise. Summary
where Yit is subjective AIDS-risk perception by statistics for these two dependent variables are
individual i at time t based on the question ‘How reported in Table 2.
worried are you that you might catch AIDS?’, with The central explanatory variables for the analysis
responses coded as 1/worried a lot and 0 /either presented in this paper are measures of the respon-
not worried at all or worried a little. Yit is the dent’s social interactions regarding AIDS. Based on
corresponding latent variable, Nit is the average the theoretical discussion above, social interactions
level of perceived risk of AIDS among the network are represented through three variables: the propor-
partners of individual i before time t data (the tion of a respondent’s network partners who are
subscript t / for predetermined variables empha- very worried about HIV; the density of a respon-
sizes that these variables refer to the time before t ), dent’s social network; and an interaction between
and the measure densityit describes the structure of both. In order to compute these different network
a social network. The interaction term between Nit measures, especially density, the analysis is restricted
and densityit captures the fact that the effects of to the respondents recorded as having chatted with
social interactions may be modulated by the struc- at least two network partners in both waves of the
ture of a network. In addition, Xit is a vector of data (see also related discussions above and below).
observed variables for individual i (e.g., age, marital Summary statistics for these variables are reported
status, parity, wealth); fi is a vector of unobserved in Table 3.
fixed factors that determine an individual’s AIDS The fixed-effects approach presented here controls
worries (e.g., unobserved community characteris- for all time-invariant characteristics of the respon-
tics); and oit is a random, logistically distributed dent and his or her household and community. We
disturbance term that affects the risk perception by therefore include only a small number of additional
individual i at time t . time-varying characteristics of the respondent: parity
Social learning and social influence can be dis- (number of children ever born); marital status (a
tinguished in this specification because of a different dummy variable for being currently married); socio-
pattern of coefficients implied by the two mechan- economic status (based on proxy variables*/whether
isms. When social learning is the most relevant the household has a metal roof or a radio or both */
mechanism through which networks affect attitudes that indicate a household’s wealth); and economic
and risk perceptions, the percentage of network activity (a dummy variable entered as 1 if the
partners who worry a lot about AIDS is the variable respondent worked for income). Exposure to
 Social interactions and AIDS in rural Malawi 273

AIDS-prevention programmes is also controlled for. An important corollary of this approach is that all
The measure of programme activity shown here is individuals who show no change in the value of the
based on the sum of three dummy variables, indicat- dependent variable between waves make no con-
ing whether or not the respondent has heard of tribution to the likelihood function of the fixed-
AIDS-prevention strategies during a talk at a clinic effects model and the identification of the coeffi-
or hospital, on the radio or TV, or from a community- cients of interest. This is because, for the respon-
based distribution worker*/that is, by the main dents whose risk perceptions remain constant
channels through which prevention programmes between the two waves, it is impossible to disen-
are conducted in rural Malawi. However, this vari- tangle the effects on respondents’ risk perceptions
able does not capture occasional prevention efforts that are attributable to social interactions (or other
that occur outside these channels, such as dramas time-varying variables) from those that are attribu-
about AIDS that may sometimes be performed in table to unobserved fixed characteristics (see Hsiao
villages, or in the nearby trading centres. ‘Chishango’ 2003). Fixed-effects models can thus be used to
(‘shield’ in Chichewa) condoms may also be distrib- estimate the probability of an increase in worry in
uted in bars, local beer stores, or other outlets. instances where there has been either an increase or
The random-effects models reported here for a decrease in risk perceptions. This probability is
comparison with the fixed-effects models addition- conditional on the values of the fixed effects (fi ),
ally include age and education in 1998 (the latter and is solely a function of the changes in the values
represented by dummy variables for completion of of the variables used to describe a respondent’s risk
primary or secondary schooling or both). Because perception.
there are also secular trends in the dependent In this model, therefore, the parameters of inter-
variables over time, our fixed-effects and random- est (d1, d2, and d3) are consistently estimated on the
effects analyses include a time trend that is given as assumption that stochastic shocks that affect the
0 for the 1998 wave and 1 for the 2001 wave of data representation of network partners before time t do
collection. not affect the current perception of the AIDS risk at
The major advantage of our preferred fixed- time t . Qualitative evidence gathered as part of the
effects models is the ability to control for unob- MDICP suggests that the assumption of a simple
served heterogeneity among respondents that affects fixed effect representing the unobserved heteroge-
the selection of social networks. In particular, fixed- neity in the selection of network partners is reason-
effects models ‘condition out’ the fixed component fi able. When asked during in-depth interviews why
from the residual and from all explanatory variables they talked to a particular person about AIDS,
included on the right-hand side of model (2) (Hsiao women seldom described strategic search behaviour,
2003). This approach can eliminate a potential and instead typically provided answers of the form
correlation between the indicators of social interac- ‘she was at the borehole today’ (Watkins 2004). The
tions (N and density) and the residual that is due to opportunity structure for social interactions, which is
unobserved fixed individual characteristics. In con- related to fixed characteristics such as personality,
trast, the residual in standard logit estimates of household location, and so on, thus seems to be the
relation (2) includes the fixed effects (fi ), as well as principal source of heterogeneity in the process of
oit . Standard logit analyses may thus misrepresent network selection and is adequately accounted for
the role of social interactions if there is a correlation by the fixed-effects specification of relation (2).
between the residual (fi ,/oit ) and the right-hand- The advantage of longitudinal data in controlling
side variables of interest. This correlation occurs, for for unobserved heterogeneity and the endogenous
instance, if respondents select network partners who selection of network partners is often thought to be
are similar to themselves with respect to character- limited by attrition, especially in less developed
istics that are unobserved in the data (homophily), countries where attrition rates are likely to be high
or if the social context of individuals differs in because of mortality and extended periods of
(unobserved) aspects that provide differential op- absence from the household. However, recent stu-
portunities for social interactions about HIV/AIDS dies have concluded that attrition in panel data, even
(e.g., the presence or absence of other women, when substantial, does not cause serious bias (Fitz-
particularly female relatives, in the compound). In gerald et al. 1998; Alderman et al. 2001; Bignami et
both of these cases, the composition and the density al. 2003; Falaris 2003). Moreover, attrition is likely to
of the networks will depend on the fixed effects (fi ), be substantially determined by characteristics of the
resulting in biases of standard logit estimates of the individual that do not vary across time (Ziliak and
parameters. Klisner 1998), and fixed-effects models that control
274 Stéphane Helleringer and Hans-Peter Kohler

for these time-invariant observed and unobserved Results of basic analyses

characteristics are likely to be considerably less
affected by bias than standard analyses. This discussion begins with some basic estimates of
A further potential problem for the present study relation (1). For comparability with subsequent
follows from the fact that the analysis is based on analyses, these initial estimates are based on the
respondents who report having discussed AIDS with subset of respondents who reported talking about
at least two network partners. This could induce a AIDS with at least two network partners in both
sample selection bias because participation in social survey waves. The results in Table 4 show that the
interactions determines inclusion in the study. How- two dependent variables */the respondent’s subjec-
ever, to the extent that a social network with fewer tive worries about getting AIDS (‘risk perception’)
than two members is determined by observed and the respondent’s assessment of communication
characteristics or unobserved fixed characteristics, about AIDS with his or her spouse as an effective
such as age or constraints on social interactions AIDS prevention strategy (‘spousal commu-
caused by the location or demographic composition nication’)*/are, with few exceptions, not strongly
of the respondent’s compound or village, the fixed- predicted by socio-economic characteristics such as
effects analysis controls for this selection and the age, parity, schooling, marital status, or the indica-
estimated coefficients will remain unbiased, despite tors for wealth and earning activity (see Montgom-
the focus on respondents with at least two network ery 2000 for similar remarks). This basic analysis,
partners. however, reveals two potentially important conse-

Table 4 Estimates of the associations between attitudes to HIV/AIDS and individual and network characteristics, using
random-effects logit models. Malawi 1998 /2001. Respondents with 2/network partners in both waves1

Risk perception2 Spousal communication3

Women Men Women Men

Intercept /3.08*** /3.22*** 0.15 0.70

(0.39) (0.71) (0.32) (0.71)
Age 0.01 /0.01 0.01 /0.01
(0.01) (0.01) (0.02) (0.01)
Parity /0.03 /0.01 /0.01 0.02
(0.03) (0.03) (0.02) (0.03)
Primary schooling 0.29* 0.25 /0.07 /0.35**
(0.15) (0.23) (0.11) (0.17)
Currently married /0.21 0.44 /0.77*** 0.15
(0.21) (0.45) (0.19) (0.46)
House has a metal roof 0.31 0.38 /0.04 0.03
(0.22) (0.28) (0.18) (0.21)
Doing wage work 0.29* 0.40 /0.15 /1.18**
(0.15) (0.33) (0.11) (0.48)
AIDS prevention effort 0.17* /0.01 0.18** /0.01
(0.09) (0.01) (0.08) (0.02)
Per cent network partners highly 2.97*** 3.61*** 0.11 0.04
worried about AIDS (0.18) (0.22) (0.14) (0.16)
Time trend 0.44*** 0.56*** /0.58*** 0.06
(0.15) (0.16) (0.11) (0.15)
Dummy for south 4.28*** 3.71*** /4 1.51***
(0.64) (0.72) (0.41)
Time * south /2.42*** /1.83*** /4 /0.95***
(0.35) (0.41) (0.26)
N 814 648 814 648
1
Standard errors are adjusted for the correlation of residuals across waves.
2
The survey asked ‘How worried are you that you might catch AIDS?’
3
The survey asked ‘What do you think are the best ways to protect yourself from AIDS?’ Answers were not prompted.
4
The dummy variable and interaction effect were not included because it was not significant for women in the south.
***p B/0.01; **p B/0.05; *p B/0.1.
Source : As for Table 1.
 Social interactions and AIDS in rural Malawi 275

quences for risk perceptions of the respondent’s model above persist or change once we control for
exposure to discussions about AIDS. First, more unobserved characteristics and potentially endogen-
exposure to the efforts of organized programmes is ous network selection? Second, can our analysis
associated both with more pessimistic perceptions of distinguish between the information a social net-
the risk of HIV and a higher probability of spousal work provides and the normative influence it exerts?
communication about AIDS */but only for women.
Second, social interactions with others seem to affect Self-perceived risk of HIV infection
risk perceptions for both men and women. Table 4
shows that the proportion of social-network partners Estimates for women . Table 5 presents estimates
who are very worried about AIDS is significantly using model (2) for individual risk perceptions. For
related to the respondent’s own risk perception, but simplicity, we focus on the network variables of
not to the probability of spousal communication primary relevance for the study presented here.
about HIV/AIDS. Initially it was found that the pattern of coefficients
differed across regions, suggesting important local
variation in the mechanisms underlying the effects
Fixed-effects analysis of social interactions. Specifically, the coefficients of
the network variables in relation (2) were allowed to
In the following section, we turn to our preferred differ between the pooled districts of the south and
fixed-effects analysis of relation (2), which avoids the centre and the north.
important limitations of the basic analysis of relation The coefficients reported in Table 5 for the south/
(1). In this analysis we focus on two issues. First, do centre are typical of a situation in which social
the relations that can be estimated using the simple learning is relatively more important. The coefficient

Table 5 Estimates of the effects of social interactions on the probability of worrying about AIDS, using random-effects
and fixed-effects logit models. Malawi 1998 and 2001. Respondents with 2/ partners in both waves

Random effects1 Fixed effects1,2,3,4

Women5 South/centre North South/centre North

Per cent network partners worried a lot about AIDS (d1) 2.51*** /0.60 2.56*** /0.45
(0.57) (0.80) (1.02) (1.20)
Density (d2) 0.39 /1.24** 0.91 /2.45**
(0.38) (0.5) (0.71) (0.87)
Per cent network partners worried a lot about AIDS * density (d3) 0.59 2.48*** /0.29 3.86**
(0.68) (0.89) (1.11) (1.52)
N 548 256 236 111

Men All districts All districts

Per cent network partners worried a lot about AIDS (d1) 3.06*** 3.88***
(0.45) (1.01)
Density (d2) /0.26 0.60
(0.32) (0.85)
Per cent network partners worried a lot about AIDS * density (d3) 0.86 /0.21
(0.75) (1.10)
N 648 283
1
All models include controls for age, education, marital status, AIDS prevention effort, economic activity, a dummy for
metal roof on the house, and parity.
2
Education and age are time-invariant and are subsumed under the fixed effect.
3
The fixed-effects logit model is based only on respondents who changed their risk perceptions between the two waves.
4
The main difference between fixed-effects and random-effects estimates is that fixed-effects models allow for the
possibility that the fixed effects (fi ) are correlated with the independent variables of the logit model. Such a correlation is
expected whenever network partners are not randomly assigned to the respondent. Random-effects models, just like
ordinary models, assume that the fixed effects (fi ) are not correlated with the variables included in the analysis.
5
A Wald test indicated that results for women could not be pooled across districts.
***p B/0.01; **p B/0.05; *p B/0.1.
Source: As for Table 1.
276 Stéphane Helleringer and Hans-Peter Kohler

d1 is large, positive, and highly significant, and the dent’s risk perception of additional network partners
density of the network structure has no significant with AIDS worries diminishes.
effect on the likelihood of worrying about AIDS. The fixed-effects estimates for the north (Figure
Figure 2 (for south/centre) shows the correspond- 3) show a strikingly different pattern. In particular,
ing predicted effect that changes in that proportion the interaction between the network partners’ risk
of social-network partners who are highly worried assessments and the density of the social network is
about AIDS have on the respondent’s own risk positive and significant. On the one hand, a dense
perception. The graph displays the probability that network with no network partners very worried
a woman worries a lot (vs. worrying a little or not at about AIDS exerts a strong negative effect on the
all) as a function of, first, the proportion of network probability that the respondent is very worried;
partners who worry about AIDS (measured on the x- correspondingly, an increase in the level of concern
axis), and second, the density of the social network, in such a dense network has a strong positive effect
ranging from a relatively sparse network with a on the respondent’s risk assessment. On the other
density of 0.5 (dashed line) to a closely knit network hand, in a sparse network (density below 0.5),
with a density of 1 in which all network partners are variation in the level of concern across networks
tied to each other (solid line). has only a weak effect on the respondent’s risk
Several important conclusions follow from this perception, or none at all. Thus, it is primarily when
analysis. First, the probability of a respondent being the network is relatively dense that the level of
very worried about AIDS increases with the pre- concern in the network matters. In addition, in-
valence of AIDS concern in her network, and this creases in density have a positive effect on the
effect does not vary significantly across network respondent’s AIDS worries only if the social net-
structures, contrary to what would be expected if work consists primarily of network partners who are
social influence was the relevant mechanism. Sec- very worried about AIDS. This pattern of social-
ond, as the level of concern in a network increases, a interaction effects is typical of social influence:
further worried network partner has a diminishing social interactions seem to exert pressure towards
effect on the respondent’s risk perceptions. These conformity in risk perceptions in a social network.
findings are consistent with a dominance of social Since these results are derived from our fixed-
learning in the social interactions about HIV/AIDS: effects estimates, it is important to assess whether
in situations where it is primarily social networks controls for all time-invariant characteristics of
that provide information about a new health threat, individuals and communities make a difference. In
having one very worried network partner is likely to the south/centre, the difference between model
be enough to provide the basic information that is specifications is very small. The point estimate for d1
known throughout the community (e.g., that HIV is is quite similar in both specifications, and while the
sexually transmitted and that AIDS is fatal). Addi- estimate for d3 differs, it is not statistically significant
tional partners are unlikely to provide much new in either specification. This suggests that network
information, so the marginal effect on the respon- variables in the south/centre are not strongly

1.0 0.8
Probability of worrying about AIDS

Probability of worrying about AIDS

0.7 Density = 0.50

0.9
Density = 0.75
0.6 Density = 1.00
0.8
0.5
0.7 0.4
0.3
0.6 Density = 0.50
Density = 0.75 0.2
0.5 Density = 1.00
0.1
0.4 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Proportion of network partners highly Proportion of network partners highly
worried about AIDS worried about AIDS

Figure 2 The effect of network partners’ risk perceptions Figure 3 The effect of network partners’ risk perceptions
on the probability of worrying about AIDS, based on on the probability of worrying about AIDS, based on
fixed-effects estimates for women reported in Table 5, fixed-effects estimates for women reported in Table 5,
south/centre regions northern region
Source : As for Table 1 Source : As for Table 1
 Social interactions and AIDS in rural Malawi 277

correlated with unobserved individual and commu- south/centre and the north for females, and the
nity characteristics. In the north, however, the uniform pattern of social learning across all three
differences between the random-effects and fixed- regions for males, it is important to consider how the
effects models are substantial, suggesting that ran- north differs from the other regions. Two possible
dom-effects models greatly underestimate the effect explanations seem most pertinent. First, social-
of density and its interaction with the proportion of interaction effects may differ because the networks
network partners who are very worried. This may be in the north have a different composition and
due to a stronger selection of network partners by structure from those in the south/centre. Second,
women that results in underestimating the extent of the differences in social-interaction effects may
social influence in this region. result from contextual differences in the nature of
the interactions between the respondent and his or
her network partners.
Estimates for men. The qualitative data collected Socio-economic contexts and the size and struc-
as part of the MDICP show that not only women but ture of social networks do indeed differ between the
also men spend considerable time engaging in north and south/centre. Table 1 shows that there
informal conversations about HIV/AIDS, the threats are several systematic socio-economic differences
posed to their health and their communities, and the between the regions, some of which may have
possibilities of reducing infection risks by beha- implications for network dynamics. For instance,
vioural changes and partner selection (Watkins the average education level is much higher in the
2004). In the lower panel of Table 5 we therefore north, as is the proportion of respondents speaking
report the fixed-effects estimates of relation (2) for English. A higher proportion of households there
males. Because Wald tests for the differences in the also have a radio or pit latrine, or a house roofed
coefficients across regions show that these estimates with metal sheeting. Our qualitative data, as well as
do not exhibit the same regional variation shown for our participation in the collection of the survey data,
females, in the present study the three regions are led us to expect that different settlement patterns
pooled to give a single set of coefficients. The results and different ways of supplementing subsistence
in Table 5 show that the coefficient d1 is sizable and agriculture may be particularly important. Because
highly significant, and that the effect of worries the north is less densely settled than the south/
among social-network partners on the respondent’s centre, people are more likely to go to a trading
risk perception does not vary much across different centre for a casual conversation with a friend in the
network structures. This pattern suggests that social latter two regions. In addition, households in the
learning is the main mechanism behind social south use small businesses as a way of supplement-
interactions for men in all three regions. This ing subsistence agriculture. This leads to greater
interpretation is corroborated by numerous descrip- diversity in daily activities, and thus to more inter-
tions of conversations in the observers’ field reports: actions with acquaintances than with confidants. In
there are frequent conversations among males about the north, on the other hand, most men find
whether a potential, current, or past sexual partner is employment on tobacco plantations (the main
likely to be infected. Topics include whether or not a export crop in Malawi) and thus have fewer incen-
woman has other partners, and whether one of her tives to look for work or income elsewhere. Table 3
previous partners died of AIDS. (See Kaler 2004 and shows that networks in the south/centre tend to
Watkins 2004 for detailed descriptions of these include more network partners who are acquain-
conversations.) It is also worth noting that the point tances, or have just met the respondent once,
estimates for d1 vary greatly between model speci- whereas networks in the north include a higher
fications, and that random-effects models tend to proportion of friends and confidants of the respon-
underestimate the effects of social interactions on dents. Responses to a question about the residence
men’s perceptions of risk. of network partners show that more respondents in
the south/centre than in the north engage in
conversations about AIDS with network partners
Local variations in the effects of social from the cities or other regions of Malawi. As a
interaction: the role of marriage patterns corollary to these observations, the density of social
networks tends to be lower in the southern and
To understand the distinct differences in the patterns central regions, even though this difference tends to
of social learning and social influence between the attenuate over time. This may indicate a higher
278 Stéphane Helleringer and Hans-Peter Kohler

prevalence of ‘weak ties’ (Granovetter 1973) in the ability to talk with unrelated women also impede the
two districts of the south and centre. flow of information that would be a cause for
However, these regional differences in network concern. In contrast, semi-structured interviews
composition are similar for men and women, and with women in the south have shown that they
thus do not provide an explanation for the differ- know a great deal about the sexual partners of
ences by sex in the patterns of social interactions in others in the community (Tawfik and Watkins 2003),
Table 5. It is likely that this pattern of the effects of and our observers’ field reports show that informa-
social interactions on respondents’ risk perceptions tion about a husband’s infidelity reaches his wife
is primarily related to the distinctive marriage quickly. Formal tests of these hypotheses are not
patterns found in the north and in the south/centre, possible with the present data, for such tests would
and the manner in which these different marriage require complete (‘sociocentric’) network data, and
patterns alter the composition of social networks. especially information about the centrality of female
Marriage in the north is primarily patrilocal: women relatives and how this determines their ability to
leave their families once marital unions are formed control information flows and act as cut points
and start living in the husband’s family, which means between various social groups (Granovetter 1973;
that women have to form new local networks after Burt 1992).
marriage. Marriage in the south/centre, on the In contrast to the pattern for women, there are no
other hand, is matrilocal, and it is thus men’s detectable differences in network composition for
networks that are altered at marriage. Table 3 men between the three districts. Of particular note is
supports the expectation that differences in marriage the finding that men in the matrilineal regions of
patterns between the north and the south/centre south/centre do not talk about AIDS with mem-
will be associated with different patterns of social
bers of their spouse’s family more often than do men
interactions. Women’s social networks in the north
in the north (Table 3). Phiri (1983) shows that men
contain a large proportion of female relatives of the
in the matrilineal societies of Malawi have ‘invented’
respondent (39 per cent in 2001), and a large fraction
several institutions that protect them from the
of female in-laws (21 per cent), whereas these
influence of their in-laws. Because men’s AIDS
proportions are much lower in the other two
conversation networks are less embedded in the
districts.
wife’s family, and because behavioural expectations
Moreover, the dispersed settlement pattern in the
are different for men, social-interaction effects on
north makes it more difficult for women to talk to
men’s risk perceptions are not strongly affected by
others outside the compound, and because the
the different marriage patterns.
distances to a trading centre or clinic are greater,
women from the same compound are likely to walk
together, thus enabling a wife’s interactions to be
more readily influenced by her husband’s female kin Social interactions and spousal communication
(Schatz 2002). In contrast, the denser settlement
patterns in the south/centre make possible a wider Spousal communication about HIV/AIDS is often
range of casual network partners, with little or no seen as an important determinant of behavioural
supervision by a husband’s female relatives. responses to the epidemic. In particular, the pre-
While direct evidence is not available in the ventive strategies promoted by the international
quantitative data used for the present study, we health community, which encourage the use of
believe that, in the north, the over-representation of condoms or abstinence from sexual relationships
members of the husband’s family among women’s altogether, do not appear adequate for married
network partners for conversations about AIDS couples. Spouses in rural Malawi have also come to
constrains their opportunities for learning and understand that their fates are joined (Watkins
increases the extent to which networks exert social 2004), and that the management of the risk of
influence. In particular, the husband’s female rela- HIV/AIDS requires the cooperation of husbands
tives (his mother and his unmarried sisters, as well as and wives. Initiating a discussion about the threat of
his sisters-in-law, who are all competing for com- AIDS within one’s partnership is therefore seen as
pound resources) may effectively hinder a wife from the first step in trying to persuade one’s spouse to be
learning about behaviour of her husband that would faithful or to adopt risk-reduction strategies in
increase her worry about becoming infected, such as extramarital relations (Zulu and Chepngeno 2003).
gossip about his extramarital relationships. In addi- However, open spousal communication about sexual
tion, the constraints that distance places on the behaviour and potential HIV infection is often
 Social interactions and AIDS in rural Malawi 279

Table 6 Estimates of the effects of social interactions on the probability of spousal communication about AIDS, using
random-effects (RE) and fixed-effects (FE) logit models. Malawi 1998 and 2001. Respondents with 2/network partners in
both waves

Women1,2,3 Men1,2,3

RE FE RE FE

Per cent network partners worried a lot about AIDS (d1) /0.65* /0.72* /0.14 /0.77
(0.36) (0.43) (0.40) (0.60)
Density (d2) /0.26 /1.01* /0.47 /0.84*
(0.25) (0.61) (0.31) (0.51)
Per cent network partners worried a lot about AIDS * density (d3) 0.76* 1.55** 0.36 1.07
(0.39) (0.66) (0.37) (0.70)
N 814 387 648 301
1
All models include controls for age, education, marital status, AIDS prevention effort, economic activity, a dummy for
metal roof on the house, and parity.
2
Education and age are time-invariant and are subsumed under the fixed effect.
3
The fixed-effects logit model is based only on respondents who changed their attitudes between the two waves.
***p B/0.01; **p B/0.05; *p B/0.1.
Source : As for Table 1.

contrary to long-standing practices in sub-Saharan different between patrilineal and matrilineal socie-
Africa, and discussing the likelihood of HIV infec- ties (Caldwell et al. 1992).
tion with one’s spouse is likely to provoke issues of The fixed-effects estimates for men show a similar
trust or imply accusations of infidelity that can pattern for the probability that a husband considers
threaten the viability and stability of unions. Cald- spousal communication an effective prevention
well et al. (1992, p. 385) thus describe the ‘near- strategy. In particular, social-interaction effects are
prohibition on the discussion of sexual activity dominated by social influence, and interactions with
between . . . spouses’. densely knit networks can reduce the probability
Table 6 shows the results of the fixed-effects that men consider spousal communication important
model (2) for the probability that a respondent for AIDS prevention if the network is characterized
considers spousal communication to be among the by a generally low level of AIDS concern. The effect
best ways to protect oneself from getting AIDS. The operates in the opposite direction if a high level of
results for women, displayed in Figure 4, indicate AIDS concern is found in a dense network. How-
ever, these results are only marginally significant.
that attitudes towards spousal communication vary
The relative weakness of social-interaction effects
strongly with network density. For instance, the
on men’s attitudes towards spousal communication
proportion of network partners who are very wor-
may be due to the fact that they do not perceive
ried about AIDS has almost no effect on the
respondent’s attitude toward spousal communica-
Probability of spousal communication

0.50
tion if the network is relatively sparse (density 5/ Density = 0.50
0.5), while it has very strong effects if the network is 0.45 Density = 0.75
Density = 1.00
dense (density /1). This pattern is typical of social
about AIDS

0.40
influence. As a consequence, women indeed appear
to be socially constrained towards not discussing 0.35
HIV/AIDS with their husbands when the prevalence
of AIDS concerns in their network is relatively low, 0.30
and when their network partners form a closely
0.25
linked group. On the other hand, when the pre- 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
valence of concern is high among densely knit Proportion of network partners highly
worried about AIDS
network partners, women may be encouraged to
discuss the risks of HIV infection with their hus- Figure 4 The effect of network partners’ risk perceptions
on the probability of spousal communication about AIDS,
bands. These results are also uniform across districts,
all regions, based on fixed-effects estimates for women
even though underlying sexual behaviour and pat- reported in Table 6
terns of extramarital relationships may be very Source : As for Table 1
280 Stéphane Helleringer and Hans-Peter Kohler

themselves as being put at risk of infection by their spouse’s family and therefore impose tighter forms
spouse: in 1998 only 13 per cent of male respondents of social control and social influence.
(25 per cent in 2001) cited their spouse as the
potential source of infection that most worried
them (Table 2). For comparison, in 2001 half the Notes
women in our sample reported being worried about
getting infected by their spouse. In casual conversa- 1 Stéphane Helleringer is a PhD candidate in the Popula-
tions about AIDS, men appear to discuss primarily tion Studies Center, University of Pennsylvania, 3718
the possibility of remaining faithful, or better ways Locust Walk, Philadelphia, PA 19104-6299, USA. E-
of assessing the risk associated with certain types of mail: hellerin@[Link]
partnerships. 2 Hans-Peter Kohler is Associate Professor of Sociology
and Research Associate in the Population Studies
Center, University of Pennsylvania, 3718 Locust Walk,
Conclusions Philadelphia, PA 19104-6299, USA. E-mail: hpkohler@-
[Link]
The aim of the study presented in this paper was to 3 We are grateful for the many helpful comments that we
investigate whether social interactions among have received from Susan Watkins, Jere Behrman, Irma
friends, peers, relatives, and community members Elo, Georges Reniers, Philip Anglewicz, and two
play an important role in developing strategies to anonymous referees. We also gratefully acknowledge
reduce the risk of HIV/AIDS, and in evaluating the the support for this research through NIH grants RO1
HD044228 and RO1 HD/MH41713.
benefits of ‘new’ behaviour such as the adoption of
condom use, increased marital fidelity, or more
selective partner choices. The study provides clear
evidence that social interactions do indeed have
References
these functions. First, it shows that social interac-
Alderman, Harold, Jere R. Behrman, Hans-Peter Kohler,
tions on the subject of HIV/AIDS have significant
John A. Maluccio, and Susan C. Watkins. 2001. Attri-
and substantial effects on respondents’ perceptions tion in longitudinal household survey data: some tests
of the risk of HIV/AIDS, even after controlling for for three developing-country samples, Demographic
unobserved characteristics affecting both respon- Research 5: 79 /123.
dents’ risk perceptions and their social networks, Axinn, W. G. and Scott T. Yabiku. 2001. Social change, the
and that a larger proportion of network partners social organization of families, and fertility limitation,
who are worried about HIV/AIDS tends to increase American Journal of Sociology 106: 1219 /1261.
the respondent’s own perception of risk. Second, it Barnett, Tony and Piers Blaikie. 1992. AIDS in Africa: Its
suggests that social networks do not necessarily Present and Future Impact . New York: Guilford Press.
constrain individuals towards the formation of fatal- Behrman, Jere R., Hans-Peter Kohler, and Susan C.
istic risk perceptions, or the adoption of risky Watkins. 2002. Social networks and changes in contra-
behaviour. On the contrary, respondents in the ceptive use over time: evidence from a longitudinal
sample tended to use their social networks to gather study in rural Kenya, Demography 39: 713 /738.
information about the disease and to assess their Behrman, Jere R., Hans-Peter Kohler, and Susan C.
own risk of infection. Rather than a constraint or a Watkins. 2004. Social networks and HIV/AIDS risk
‘barrier’, social networks appear as a resource for perceptions. Mimeo, University of Pennsylvania, Phila-
individuals to learn about and evaluate new beha- delphia.
vioural strategies in the face of the epidemic. Third, Bernardi, Laura. 2003. Channels of social influence on
we find that the dominant mechanisms */social reproduction, Population Research and Policy Review
learning and social influence */through which social 22(5/6): 527 /555.
interactions shape attitudes towards AIDS vary by Bignami-Van Assche, Simona, Georges Reniers, and Alex
sex and by region, and that marriage patterns play a Weinreb. 2003. An assessment of the quality of KDICP
major role in defining the mechanisms of social and MDICP data, Demographic Research , Special
interactions. In particular, our analysis suggests that Collection 1: 32 /69.
women in the patrilocal region of the study have Bond, Virginia and Paul Dover. 1997. Men, women and
fewer opportunities to make use of social interac- the trouble with condoms: problems associated with
tions to learn about threats to their sexual health condom use by migrant workers in rural Zambia,
because networks are deeply embedded in the Health Transition Review 7(Suppl.): 377 /391.
 Social interactions and AIDS in rural Malawi 281

Bongaarts, John and Susan C. Watkins. 1996. Social Kohler, Hans-Peter. 2001. Fertility and Social Interaction:
interactions and contemporary fertility transitions, An Economic Perspective. Oxford: Oxford University
Population and Development Review 22(4): 639 /682. Press.
Burt, Ronald S. 1992. Structural Holes: The Social Struc- Kohler, Hans-Peter, Jere R. Behrman, and Susan C.
ture of Competition . Cambridge, MA: Harvard Uni- Watkins. 2001. The density of social networks and
versity Press. fertility decisions: evidence from South Nyanza Dis-
Caldwell, John C., Pat Caldwell, and Pat Quiggin. 1989. trict, Kenya, Demography 38(1): 43 /58.
The social context of AIDS in sub-Saharan Africa, Kravdal, Øystein. 2002. Education and fertility in sub-
Population and Development Review 15(2): 185 /234. Saharan Africa: individual and community effects,
Caldwell, John C., Pat Caldwell, and I. O. Orubuloye. Demography 39(2): 233 /250.
1992. The family and sexual networking in sub-Saharan Low-Beer, Daniel and Rand L. Stoneburner. 2004.
Africa: historical regional differences and present-day Uganda and the challenge of AIDS, in K. Poku Nana
implications, Population Studies 46(3): 385 /410. and Alan Whiteside (eds.), The Political Economy of
Caldwell, John C. 2000. Rethinking the African AIDS AIDS in Africa . Aldershot: Ashgate, pp. 165 /190.
epidemic, Population and Development Review 26(1): Manski, Charles. 1995. Identification Problems in the
117 /135. Social Sciences. Cambridge, MA: Harvard University
Campbell, Catherine. 2003. ‘Letting Them Die’: Why HIV/ Press.
AIDS Prevention Programmes Fail . Bloomington and Montgomery, Mark R. and John B. Casterline. 1993. The
Indianapolis: Indiana University Press. diffusion of fertility control in Taiwan: evidence from
Casterline, John B. 2001. The pace of fertility transition: pooled cross-section time-series models, Population
Studies 47(3): 457 /479.
national patterns in the second half of the twentieth
Montgomery, Mark R. and John B. Casterline. 1996. Social
century, Population and Development Review 27(S2:
learning, social influence, and new models of fertility,
Suppl.: Global Fertility Transition): 17 /52.
Population and Development Review 22(Suppl.: Ferti-
Entwisle, Barbara, Albert I. Hermalin, Peerasit Kamnuan-
lity in the United States: New Patterns, New Theories):
silpa, and Apichat Chamratrithirong. 1996. A multilevel
151 /175.
model of family planning availability and contraceptive
Montgomery, Mark R. and W. S Chung. 1998. Social
use in rural Thailand, Demography 21(4): 559 /574.
networks and the diffusion of fertility control in the
Falaris, Evangelos M. 2003. The effect of survey attrition
Republic of Korea, in R. Leete (ed.), The Dynamics of
in longitudinal surveys: evidence from Peru, Côte
Values in Fertility Change. Oxford: Oxford University
d’Ivoire and Vietnam, Journal of Development Eco-
Press, pp. 179 /209.
nomics 70: 133 /157.
Montgomery, Mark R. 2000. Perceiving mortality decline,
Fitzgerald, John, Peter Gottschalk, and Robert Moffitt.
Population and Development Review 26(4): 795 /819.
1998. An analysis of sample attrition in panel data: the
Moscovici, Serge. 1985. Social influences and conformity,
Michigan Panel Study of Income Dynamics, Journal of
in G. Lindzey and E. Aronson (eds.), Handbook of
Human Resources (Special Issue: Attrition in Long- Social Psychology, Volume 2. New York: Random
itudinal Surveys) 33(2): 251 /299. House, pp. 347 /412.
Gluckman, Max. 1950. Kinship and marriage among the Moursund, Anne and Øystein Kravdal. 2003. Individual
Lozi of Northern Rhodesia and the Zulu of Natal, in A. and community effects of women’s education and
R. Radcliffe-Brown and Daryll Forde (eds.), African autonomy on contraceptive use in India, Population
Systems of Kinship and Marriage. London: Oxford Studies 57(3): 285 /301.
University Press (published for the International Afri- Munshi, Kaivan and Myaux Jacques. 2002. Development
can Institute). as a process of social change: an application to the
Granovetter, Mark S. 1973. The strength of weak ties, fertility transition. Mimeo, University of Pennsylvania,
American Journal of Sociology 78(6): 1360 /1380. Philadelphia.
Hsiao, Cheng. 2003 (1st edn 1986). Analysis of Panel Data . Nzioka, Charles. 1996. Lay perceptions of risk of HIV
Cambridge: Cambridge University Press. infection and the social construction of safer sex: some
Kaler, Amy. 2004. AIDS-talk in everyday life: the pre- experiences from Kenya, AIDS Care 8(5): 565 /579.
sence of HIV/AIDS in men’s informal conversation in Orubuloye, I. O., John C. Caldwell, and Pat Caldwell.
Southern Malawi, Social Science & Medicine 59: 285 / 1997. Perceived male sexual needs and male sexual
297. behaviour in southwest Nigeria, Social Science &
Kohler, Hans-Peter. 2000. Fertility decline as a coordina- Medicine 44(8): 1195 /1207.
tion problem, Journal of Development Economics 63: Phiri, Kings M. 1983. Some changes in the matrilineal
231 /263. family system among the Chewa of Malawi since the
282 Stéphane Helleringer and Hans-Peter Kohler

nineteenth century, Journal of African History 24(2): the Population Association of America Annual Meet-
257 /274. ing, Minneapolis, MN.
Reniers, Georges. 2003. Divorce and remarriage in rural UNAIDS. 2004. Epidemiological Fact Sheet, Malawi, 2004
Malawi, Demographic Research , Special Collection 1: update.
174 /202. Watkins, Susan C. 2000. Local and foreign models of
Rutenberg, Naomi and Susan C. Watkins. 1997. The buzz reproduction in Nyanza province, Kenya, 1930 /1998,
outside the clinics: conversations and contraception in Population and Development Review 26(4): 725 /759.
Nyanza Province, Kenya, Studies in Family Planning Watkins, Susan C., Jere R. Behrman, Hans-Peter Kohler,
28(4): 290 /307. and Eliya M. Zulu. 2003. Introduction to ‘Research
Schatz, Enid. 2002. Numbers and narratives: making sense on Demographic Aspects of HIV/AIDS in Rural
of gender and context in rural Malawi. Unpublished Africa’, Demographic Research , Special Collection 1:
PhD dissertation, University of Pennsylvania, Philadel- 1 /30.
phia. Watkins, Susan C. 2004. Navigating the AIDS epidemic in
Setel, Philip W. 1999. A Plague of Paradoxes. AIDS, rural Malawi, Population and Development Review
Culture and Demography in Northern Tanzania . Chi- 30(4): 673 /706.
cago: University of Chicago Press. Watts, Duncan J., Peter S. Dodds, and Mark E. J. Newman.
Smith, Kirsten P. and Susan C. Watkins . 2005. Perceptions 2002. Identity and search in social networks, Science
of risk and strategies for prevention: responses to HIV/ 296: 1302 /1305.
AIDS in rural Malawi, Social Science & Medicine Ziliak, James P. and Thomas J. Kniesner. 1998. The
60(3): 649 /660. importance of sample attrition in life cycle labor
Tawfik, Linda. 2003. Soap, sweetness and revenge: patterns supply estimation, Journal of Human Resources (Spe-
of sexual onset and partnerships amidst AIDS in rural cial Issue: Attrition in Longitudinal Surveys) 33(2):
southern Malawi. Unpublished PhD dissertation, 507 /530.
Bloomberg School of Public Health, Johns Hopkins Zulu, Eliya M. and Gloria Chepngeno. 2003. Spousal
University, Baltimore, MD. communication about the risk of contracting HIV/
Tawfik, Linda and Susan C. Watkins. 2003. Sex in Geneva, AIDS in rural Malawi, Demographic Research , Special
sex in Lilongwe and sex in Balaka. Paper presented at Collection 8: 247 /278.