Background: Age mixing may explain differences in HIV prevalence across populations in sub-Saharan countries, but the validity of survey data on age mixing is unknown.
Methods: Age differences between partners are frequently estimated indirectly by asking respondents to report their partner's age. Partner's age can also be assessed directly by tracing partners and asking them to report their own age. We use data from 519 relationships, collected in Likoma (Malawi), in which both the partners were interviewed and tested for HIV. In these relationships, age differences were assessed both indirectly and directly, and estimates could thus be compared. We calculate the specificity and sensitivity of the indirect method in identifying age-homogenous/age-disparate relationships in which the male partner is less/more than 5 or 10 years older than the respondent.
Results: Women were accurate in identifying age-homogenous relationships, but not in identifying age-disparate relationships (specificity ≈90%, sensitivity = 24.3%). The sensitivity of the indirect method was even lower in detecting partners older than the respondent by 10+ years (9.6%). Among 43 relationships with an HIV-infected partner included in this study, there were about 3 times more age-disparate relationships according to direct measures of partner's age than according to women's reports of their partner's age (17% vs. 46%).
Conclusions: Women's survey reports of their partner's age significantly underestimate the extent of and the HIV risk associated with age mixing in this population. Future studies of the effect of sexual mixing patterns on HIV risk in sub-Saharan countries should take reporting biases into account.
A study conducted in Malawi in which both sexual partners involved in a relationship were interviewed found that (1) women underestimate the age of their sexual partner(s) during survey interviews and (2) survey data on sexual partnerships may underestimate the risk of HIV infection associated with age-disparate relationships.
From the *Columbia University, Mailman School of Public Health, Heilbrunn Department of Population and Family Health, New York, NY; †University of Pennsylvania, Population Studies Center, Philadelphia, PA; and ‡University of Malawi College of Medicine and Invest in Knowledge Initiative, Zomba, Malawi
Supported by NIH grants RO1 HD044228, RO1 HD/MH41713, and R01HD053781 and Population Aging Center (P30-AG-012836) and the Population Studies Center (R24-HD-044964), University of Pennsylvania through 2 PARC/Boettner/PSC Pilot Grants.
Correspondence: Stéphane Helleringer, PhD, Columbia University, Mailman School of Public Health, 60 Haven Ave, New York, NY 10032. E-mail: email@example.com.
Received for publication March 9, 2011, and accepted June 3, 2011.
Women get infected with HIV at an earlier age than men in sub-Saharan countries.1–4 Although differences in male-to-female versus female-to-male probabilities of HIV transmission may explain some of these patterns,5 age differences between male and female sexual partners can also play a key role in explaining gender differences in age at HIV infection.6–8 There are several reasons as to why having older sexual partners may put young women at a higher risk of HIV infection. First, the prevalence of HIV is likely higher among older partners of young women than among boys of their own age.3,7 Second, older men are generally less likely to use condoms.9 Finally, women engaged in relationships with older partners may have more limited agency to negotiate safe sex practices,10 and may even be particularly vulnerable to domestic violence.11–13 Age differences in sexual partnerships may be accompanied by economic asymmetries,10 which further limit women's agency to adopt safe sex practices. Age-mixing patterns may play a key role in explaining the “uneven spread” of HIV in sub-Saharan populations. Chapman et al.6 recently found that age differences between partners were larger in the most affected southern African populations. Thus, interventions to limit sexual mixing between younger women and older men have been proposed and implemented in a growing number of sub-Saharan countries.14,15
Despite this programmatic focus, mathematical models indicate that the effect of such interventions at the population level could be limited if not accompanied by other behavioral changes.16 However, the conclusions of these models are based on those survey data on partner's age that have not been validated. Survey respondents are usually asked to report the age of their sexual partners,6–8,17 but the accuracy of such reports may be limited in situations where partners are not well aware of each other's age. This may be the case in short-term relationships or in societies where vital registration is limited. Reporting of partner's age may also be affected by large social desirability biases: in contexts where “sugar daddies” are stigmatized as a key epidemic driver, women may be tempted to underestimate the age of their partners during sexual behavior interviews; on the other hand, if women associate older partner's age with increased resources, some men may be tempted to exaggerate their age to their potential partner(s), when they are trying to initiate a new relationship.
In this article, we used unique data, collected on Likoma Island, on 519 sexual relationships (including both marital and nonmarital relations),18 in which both sexual partners were interviewed and reported their own age as well as their partner's age. We used these data to determine the accuracy of women's reports of their partner's age collected during surveys. We then tested whether measurement error may have led to underestimates of the HIV risk associated with age differences between sexual partners.
DATA AND METHODS
The data presented here are derived from the second round of the Likoma Network Study (LNS), a longitudinal study of sexual networks and HIV infection conducted in Malawi. Detailed descriptions of the data collection procedures are available elsewhere.18 Briefly, during the second round of data collection, all adults aged 18 to 49 years living in all but 2 of the island's villages were asked to do the following: (1) complete a socioeconomic survey, (2) complete a sexual network survey during which they identified up to 5 of their most recent sexual partners, and (3) participate in HIV testing and counseling. Information on individual characteristics, sexual partnerships, and HIV status were then linked to produce images of the sexual networks connecting population members.19
Measures of Age Differences Between Partners
In the majority of sexual behavior studies conducted in sub-Saharan countries,6,7,17,20,21 the age difference between a respondent and her partner is measured indirectly by asking the survey respondent to report her partner's age or at least estimate the age difference between her and her partner. Specifically, a respondent may be asked to report the age of her partner in completed years or to first classify her partner(s) as younger, of the same age as, or as older than her. A follow-up question for partners younger or older than the respondent then asks the respondent to specify by how many years the partner is older/younger than the respondent. Sometimes, this follow-up question may provide the respondent with categorical answers to choose from: older/younger by 5 years or less, by 6 to 10 years, or by 11 years and more.
Age differences between sexual partners can also be established directly in situations where both the partners have been interviewed during a survey. Researchers can link respondents' reports, allowing them to compare the report of a respondent's own age to the report of a partner's age made by the partner himself during the survey. The age difference between partners in a sexual relationship is thus equal to the difference between respondent's age and partner's age. This direct approach has occasionally been used in surveys in sub-Saharan countries (e.g., the couple sample of the Demographic and Health Surveys), but unfortunately (1) it has only been used among coresiding, often marital, couples, and (2) it has not been used in conjunction with estimates of age differences based on the indirect method. As a result, the accuracy of the more common indirect method for measuring age differences has not been assessed.
The LNS data permit estimating age differences using both the direct and indirect approaches. Each LNS respondent was asked to classify the age difference between her and her partner using the indirect method. In a subset of the relationships that includes both marital and nonmarital relationships, both the partners were interviewed by the study team thus also enabling direct measurement of age differences. In this article, relationships in which the man (partner) is 6 or more years older than the woman (respondent) are “age-disparate”22; relationships in which the man (partner) is 5 years older or less than the woman (respondent) are “age-homogenous.” The choice of this threshold (5 years) is due to the survey instrument used during the LNS to obtain indirect measures of partner's age: respondents were asked to state whether their older/younger partner was 5 or fewer years, 6 to 10 years, or more than 10 years older/younger than them. In this article, we assess the accuracy of women's reports of their partner's age.
The analyses of data validity are thus based on a subset of relationships in which both the partners were interviewed and agreed that they were in a relationship.23 We compare the characteristics of relationships in this analytical subset to the rest of the relationships using χ2 tests of association. Not all partners included in these analyses were tested for HIV during the LNS. We also compare the characteristics of those with known and unknown HIV status using the same approach.
Accuracy of Self-Reports of Age Differences.
We consider direct estimates of age differences as a “gold standard” against which we evaluate the validity of data obtained indirectly on partner's age. To determine the accuracy of respondents' reports of age differences, we assess whether relations identified as age-disparate (or age-homogenous) in this gold standard are also classified as such in the respondents' indirect reports. The specificity of the indirect method is thus defined as the proportion of all relationships classified as age-homogenous by the direct method that are also reported as such (indirectly) by the respondent. The sensitivity of the indirect method is defined as the proportion of all relationships classified as age-disparate by the direct method that are also reported as such (indirectly) by the respondent. We also assess the accuracy of the indirect method in detecting partners more than 10 years older than the respondent.
Survey respondents may misclassify some of their relations, if the difference between their age and the age of their partner is close to the value of the 5-year threshold (e.g., ±2 years). We thus test whether the sensitivity/specificity of the indirect measures of age differences vary with the “true” age difference between partners (classified as within 0–2 years, 3–5 years, or more than 5 years from the threshold). Survey respondents may also misreport their own age in demographic surveys, often because of age heaping.24–27 This may significantly confound our validation of the indirect method. In order to test the robustness of our findings to misreporting of a respondent's own age, we use 2 strategies. First, we use data from the first round of the LNS18 to identify individuals who reported their age consistently in both survey waves, that is, the age they reported in the second round is 1 or 2 years more than its value in the first round. Second, we consider only respondents and partners whose reported age did not end in 0 or 5. Age heaping, as measured in Figure 1, is indeed significant in the LNS dataset. We then repeat our calculations of the sensitivity and specificity of the indirect method, using the subset of relationships in which (1) both the partners are consistent reporters, and (2) neither the respondent nor her partner have an age that ends in 0 or 5.
Age-Disparate Relationships and HIV Risk.
Because the LNS is based on a partner tracing design,18 we are able to assess the prevalence of HIV among partners of survey respondents. Among all relationships with an HIV-positive partner, we estimate the frequency of age-disparate versus age-homogenous relationships according to both the indirect and direct approaches to measure age differences between partners. We test for systematic differences between the 2 measures using Fisher exact χ2 test of association (2-sided). This comparison provides an indication of the extent of bias in indirect estimates of the relative risk of HIV infection associated with age-disparate relations.
Characteristics of Relationships Included in Analysis.
The analysis focuses on 519 relationships reported by both female and male partners during the LNS. Table 1 describes the characteristics of these relationships. Relationships reported by younger respondents were as likely to be included as those reported by older respondents, but relationships of HIV-infected respondents were slightly less likely to be included (33.5% vs. 27.8%, P = 0.1). Nonmarital and dissolved relationships were less likely to be included than ongoing marital relationships. However, the probability of inclusion did not vary according to indirect measurements of the age difference between a respondent and her partner; 33.5% of the age-homogenous relationships and 34.2% of the age-disparate relationships (P = 0.83) were thus included in the analysis.
Characteristics of Partners With Known HIV Status.
Only 376 of 519 partners (72.5%) participated in HIV testing. However, there were few differences between tested and nontested partners. In particular, partners of HIV-infected respondents were not significantly more likely to be tested for HIV during the study than partners of HIV negatives. Participation in HIV testing appeared slightly higher among partners less than 5 years or older than the respondents, but this difference was not significant (P = 0.11).
Indirect and Direct Measures of Age Differences Between Partners.
Six respondents refused to answer the question about the age of their male partner (6/519, 1.1%). Respondents (indirectly) reported that 15.6% (80/513) of the included relationships involved partners older than them by more than 5 years, and 2.5% (13/513) included partners older than them by more than 10 years.
Direct measures of age differences indicate that the median age difference between a woman and her partner was 4 years (Fig. 2). Figure 2 shows how the distribution of directly measured age differences varied by age of the respondent and marital status of the relationship. The median age difference was significantly smaller in nonmarital relationships than in marriages. Among women more than 25 years, the median age difference with nonmarital partners was 1 year, whereas it was 5 years with spouses (P < 0.01). Among women less than 25 years, nonmarital partners were on average 3 years older than them, whereas spouses were on average 5 years older (P < 0.01). More than 40% of the nonmarital partners of women aged more than 25 were actually younger than them.
Accuracy of Indirect Measures of Age Differences in Sexual Relationships
Women were accurate in identifying age-homogenous relationships, but not in identifying age-disparate relationships. The specificity of the indirect measure of age differences was 89.1% (95% CI: 85.1–92.1; Table 2), but its sensitivity was 24.3% (95% CI: 18.2–31.3). Furthermore, younger respondents were significantly less accurate than older respondents in identifying age-disparate relations. Respondents less than 25 years correctly classified only 8 of 57 (14.0%) of their age-disparate relationships versus 35 of 120 (29.2%) among respondents aged ≥25.
The specificity of indirect reports increased when we sought to identify partners who are more than 10 years older than the respondent. The sensitivity of these reports, on the other hand, deteriorated further. Only 5 of 52 (9.7%) age-disparate relationships identified as such by the direct method were also correctly classified by the indirect method. In robustness tests, the sensitivity and specificity of indirect measures were similar in analyses based on the subset of relationships between consistent respondents or on the subset of relationships without age heaping.
Are Misclassifications Due to the Thresholds Imposed on Respondents' Reports of Partner Age?
We found that respondents' accuracy in classifying age-homogenous relations did not vary with the age difference between partners (Table 3). On the other hand, their ability to correctly identify age-disparate relations increased sharply with the age difference between partners. Respondents were more likely to misclassify relations with partners 6 to 7 years older than they were to misclassify relations with partners more than 7 years older than them (P = 0.05). The sensitivity of the indirect method was highest among relations with partners more than 10 years older than the respondent. However, even in such relationships, only one-third of all age-disparate relationships were correctly classified as such by the respondent.
Misclassifications and HIV Risk in Age-Disparate Relationships
There were 43 relations between respondents and an HIV-infected male partner. Figure 3 shows the proportion of age-disparate relations among these 43 relations. According to indirect measures of age differences between partners, only 7 were age-disparate (Fig. 3; 16.3%). According to direct measures, however, this proportion was much higher because 20 of these relations were age-disparate (46.5%, P < 0.01).
The analyses presented here indicate that the accuracy of indirect estimates of age differences between partners—on which most assessments of the epidemiologic importance of age mixing for HIV spread are based—is limited. We used a study design allowing comparison of indirect measures of age differences between partners with direct measures relying on partners' reports of their own age. We showed that women on Likoma engage in more age-disparate relationships than they reported during the survey. As a result, respondents' reports of their partners' ages led to underestimating the effect of age-disparate relationships on HIV infection risks. Although it initially appeared that only 1 in 6 relationships between a woman and an HIV-infected male partner was age-disparate (according to women's report of age differences between partners), direct measures of partner's age indicated that this proportion was actually closer to 1 in 2. We found that younger women were even less accurate than older respondents in identifying age-disparate relationships. Patterns of age mixing could thus play an even greater role than previously thought7 in explaining the disproportionate HIV burden among female youth in similar populations.3
However, these analyses suffer from several limitations. First, they are only based on a selective subsample of the relationships reported during the study. We are not able to assess the validity of the reports of age differences in relationships where the male partner was not interviewed or tested for HIV during the study. Second, our indirect measure of the age difference between partners was based on broad categories rather than on continuous measures of partner's age. As a result, it prevents a more detailed assessment of the correlation between respondent's reports of their partner's age and the actual age of the partner. It also forced us to adopt an arbitrary definition of age-disparate versus age-homogenous mixing. Rather than focusing on broad classifications of age differences between partners, future studies should seek to measure the accuracy of reports of exact age of partner. Third, our gold standard (partner's reports of their own age) is admittedly imperfect. In the absence of systematic birth registration, individuals may not even accurately report their own age. We showed, however, that the specificity and sensitivity of the indirect method were similar in relationships between consistent age reporters as well as in the subset of relationships without age heaping. But, unfortunately, we have no means of independently confirming a respondent's own age. Future assessments of indirect measures of age differences in sexual partnerships could be conducted in demographic surveillance sites,28 where the quality of data on age is potentially higher.
Despite these caveats, our research on the reporting of partner's age in surveys of sexual behaviors conducted in sub-Saharan populations has important implications. First, it suggests that the preventive benefits associated with effective behavioral change communications aiming to reduce age differences between partners may have been underestimated. If age-disparate relations are more common and account for a larger proportion of all the relationships through which women are possibly exposed to HIV, the epidemiologic benefits of preventing age-disparate sexual mixing could be larger than initially thought. This should be reassessed in mathematical models that account for possible measurement error. Second, it indicates that future research on the effect of age mixing (and other sexual mixing patterns) on HIV transmission should be based on rigorous study designs that allow more objective measurements of mixing patterns than current measurements that are based solely on reports by survey respondents. Taking respondents' reports at face value leads to erroneous inferences about the relative contribution of various risk factors to HIV spread.29 It could thus contribute to misguided policies and interventions, or to lack of support for promising interventions.
1. Pettifor AE, Rees HV, Kleinschmidt I, et al. Young people's sexual health in South Africa: HIV prevalence and sexual behaviors from a nationally representative household survey. AIDS 2005; 19:1525–1534.
2. Heuveline P. HIV and population dynamics: A general model and maximum-likelihood standards for east Africa. Demography 2003; 40:217–245.
3. Glynn JR, Carael M, Auvert B, et al. Why do young women have a much higher prevalence of HIV than young men? A study in Kisumu, Kenya and Ndola, Zambia. AIDS 2001; 15(suppl 4):S51–S60.
4. Taha TE, Dallabetta GA, Hoover DR, et al. Trends of HIV-1 and sexually transmitted diseases among pregnant and postpartum women in urban Malawi. AIDS 1998; 12:197–203.
5. Gray RH, Wawer MJ, Brookmeyer R, et al. Probability of HIV-1 transmission per coital act in monogamous, heterosexual, HIV-1-discordant couples in Rakai, Uganda. Lancet 2001; 357:1149–1153.
6. Chapman R, White RG, Shafer LA, et al. Do behavioural differences help to explain variations in HIV prevalence in adolescents in sub-Saharan Africa? Trop Med Int Health 2010; 15:554–566.
7. Gregson S, Nyamukapa CA, Garnett GP, et al. Sexual mixing patterns and sex-differentials in teenage exposure to HIV infection in rural Zimbabwe. Lancet 2002; 359:1896–1903.
8. Kelly RJ, Gray RH, Sewankambo NK, et al. Age differences in sexual partners and risk of HIV-1 infection in rural Uganda. J Acquir Immun Defic Syndr 2003; 32:446–451.
9. Luke N. Confronting the “sugar daddy” stereotype: Age and economic asymmetries and risky sexual behavior in urban Kenya. Int Fam Plan Perspect 2005; 31:6–14.
10. Luke N. Age and economic asymmetries in the sexual relationships of adolescent girls in sub-Saharan Africa. Stud Fam Plann 2003;34:67–86.
11. Jewkes R, Dunkle K, Nduna M, et al. Factors associated with HIV sero-status in young rural South African women: Connections between intimate partner violence and HIV. Int J Epidemiol 2006; 35:1461–1468.
12. Jewkes R, Levin J, Penn-Kekana L. Risk factors for domestic violence: Findings from a South African cross-sectional study. Soc Sci Med 2002; 55:1603–1617.
13. Jewkes RK, Levin JB, Penn-Kekana LA. Gender inequalities, intimate partner violence and HIV preventive practices: Findings of a South African cross-sectional study. Soc Sci Med 2003; 56:125–134.
15. Clark S, Bruce J, Dude A. Protecting young women from HIV/AIDS: the case against child and adolescent marriage. Int Fam Plan Perspect 2006; 32:79–88.
16. Hallett TB, Gregson S, Lewis JJ, et al. Behaviour change in generalised HIV epidemics: Impact of reducing cross-generational sex and delaying age at sexual debut. Sex Transm Infect
2007; 83(suppl 1):i50–i54.
17. Harrison A, Cleland J, Frohlich J. Young people's sexual partnerships in KwaZulu-Natal, South Africa: Patterns, contextual influences, and HIV risk. Stud Fam Plann 2008; 39:295–308.
18. Helleringer S, Kohler HP, Chimbiri A, et al. The Likoma Network Study: Context, data collection, and initial results. Demogr Res 2009; 21:427–468.
19. Helleringer S, Kohler HP. Sexual network structure and the spread of HIV in Africa: Evidence from Likoma Island, Malawi. AIDS 2007; 21:2323–2332.
20. Ott MQ, Barnighausen T, Tanser F, et al. Age-gaps in sexual partnerships: Seeing beyond “sugar daddies.” AIDS 2011; 25:861–863.
21. Katz I, Low-Beer D. Why has HIV stabilized in South Africa, yet not declined further? Age and sexual behavior patterns among youth. Sex Transm Dis 2008; 35:837–842.
22. Leclerc-Madlala S. Age-disparate and intergenerational sex in southern Africa: The dynamics of hypervulnerability. AIDS 2008; 22(suppl 4):S17–S25.
23. Helleringer S, Kohler HP, Kalilani-Phiri L, et al. The reliability of sexual partnership histories: Implications for the measurement of partnership concurrency during surveys. AIDS 2011; 25:503–511.
24. Zelnik M. Age heaping in the United States census: 1880–1950. Milbank Mem Fund Q 1961; 39:540–573.
25. Stockwell EG, Wicks JW. Age heaping in recent national censuses. Soc Biol 1974; 21:163–167.
26. Feeney G. A technique for correcting age distributions for heaping on multiples of five. Asian Pac Cens Forum 1979; 5:12–14.
27. Heitjan DF, Rubin DB. Inference from coarse data via multiple imputation with application to age heaping. J Am Stat Assoc 1990; 85:304–314.
28. Tanser F, Hosegood V, Barnighausen T, et al. Cohort profile: Africa Centre Demographic Information System (ACDIS) and population-based HIV survey. Int J Epidemiol 2008; 37:956–962.
29. Minnis AM, Steiner MJ, Gallo MF, et al. Biomarker validation of reports of recent sexual activity: Results of a randomized controlled study in Zimbabwe. Am J Epidemiol 2009; 170:918–924.