Study setting

Our data arise from the Manicaland HIV Prevention Study, a general population cohort study carried out between 1998 and 2011, with participants recruited in five rounds: 1998 to 2000 (Round 1), 2001 to 2003 (Round 2), 2003 to 2005 (Round 3), 2006 to 2008 (Round 4), and 2000 to 2011 (Round 5) [36]. The study was conducted in 12 sites, differing in terms of socioeconomic level: two small townships, four forestry/tea/coffee estates, two roadside trading settlements, and four subsistence farming areas. In the area, only starting with the years 2000s, services aimed at HIV prevention and treatment were implemented. In particular, prevention mother-to-child transmission (PMTCT) services started around 2003. Antiretroviral therapy (ART) services started around 2005, but they were really scaled-up from 2009 due to the economic crisis that hit the country between 2006 and the early 2009.

Study design

At each study round and in each selected study site, men and women aged 15–54 years were recruited based on a household census, and were asked to answer a structured interview, aimed at retrospectively collecting socio-demographic and health-related information. The interview included a number of questions related to sexual behaviour that were asked using a blend of face-to-face interview (FTFI) and self-completion methods, namely, the Informal Confidential Voting Interview (ICVI), so as to reduce social desirability bias [39]. Even though ICVI’s efficacy seems to be highly dependent on temporal [40] and local [41] circumstances, it was proved to be more effective than FTFI for the reporting of the number of sexual partners [39].

At the end of the interview, respondents were given the opportunity to ask questions. Quite often, they would ask for the correct answers to survey questions on HIV awareness. Moreover, they were asked to take an anonymous HIV serological test. Written informed consent was sought, blood spots were collected on filter paper from a finger prick, and dried blood spots were analysed using a highly sensitive and specific antibody dipstick assay [42]. Being the test anonymous, survey respondents were not provided with their results. However, independently from the study, they were offered free HIV voluntary counselling and testing (VCT). The service was made available at a mobile clinic that was present within the study site at the time the survey was being conducted [42]. Participants who opted for VCT could do so after completing the survey questionnaire. Prior ethical approval was granted by the Medical Research Council of Zimbabwe, Harare-based Biomedical Research and Training Institute’s Institutional Review Board, and the Imperial College London Research Ethics Committee. The collected data were anonymised before performing any analysis.

The overall response rates in the study varied between rounds from 73% to 87%, with response rates being similar in peri-urban and rural areas. Overall follow-up rates between rounds varied from 48% (over 3 years) to 61% (over 2 years). Excluding deaths and out-migrants (who were not eligible for follow-up), these rates increased to 77% (3 years) and to 97% (2 years). Of course, quite a few new people migrated into study areas, which, to some extent, might have compensated for those leaving.

All the individuals who ever participated in the study were initially considered as eligible for inclusion in our sample. Additional criteria for inclusion in our analysis were (i) being sexually experienced at the time of the last interview; (ii) having provided information on HIV testing history (irrespective of whether or not the respondent had been previously tested or not); (iii) having provided dried blood spots for HIV testing at least once during the study, and having a clear test result. The total number of eligible respondents fulfilling our inclusion criteria was 27,073, of which 15,949 women and 11,124 men (S1 Table).

Retrospective self-reported data provided by respondents on their age (in years) at the events of interest were used in the current analysis. As already noted for the Manicaland Study and for other similar studies on SSA countries, a high proportion of unreliable reports of both age at first sex and age at first union was found among individuals attending multiple surveys [43,44]. The same problem was found for the age at first child. Inaccurate reporting of ages at events might depend on several factors. Questions on sexual debut might be at risk of social desirability bias, with women tending to under-report sexual activity, and men over-reporting it [45–48]. Among older people, recall bias resulting in inaccurate reporting across rounds may be most common due to the longer gap between the time when the events actually occurred and the date of the interview [44]. In those cases when it was possible to suitably address inconsistencies (for example, when the inconsistency was found only in one wave), individuals were kept in the sample. Otherwise, they were excluded. Nevertheless, this action should have not resulted in artificial generation or suppression of trends [43,44].

After the removal of inconsistent data, our sample was made up of 11,647, of which 6,829 women and 4,818 men, with their information updated at the last available interview (S1 Table).

HIV testing

The HIV test taken as part of the study, based on the dipstick assay applied to the respondents’ dried blood spots, was used for research purposes. Being the testing procedure anonymous, and not deemed accurate enough for diagnostic purposes, survey respondents were not provided with their results. Therefore, a respondent’s consent to the test does not imply the knowledge of her HIV status.

However, respondents might have already taken an HIV test elsewhere through the available VCT services. Therefore, we had to consider the possible issue of reverse causality, with people making decisions on their partnership or parenthood based on the knowledge of their own HIV status. However, considering that at each study round, starting with the third one, respondents were asked whether they had ever thought about taking an HIV test, we managed to distinguish between those respondents who reported having already taken an HIV test (people who, most likely, knew their current HIV status), and those who declared that they never thought about taking the test or that they thought about taking it, but had not done it yet (S1 Text). We decided to exclude from the sample those individuals who were tested for HIV infection, reducing our sample to 7,905 respondents, of which 3,927 women and 3,978 men (S1 Table).

Construction of event sequences

In order to construct the life course trajectories leading to parenthood, we considered the reported age (in years) when sexual debut (S), union–marriage, cohabitation or long-term relationship (more than twelve months) (U)–and parenthood (C) were firstly experienced. Information is available only about the year of age at the events. Therefore, when two or more events were experienced at the same year of age, we defined a combined event. For example, (SU) indicates that sexual debut and union occurred in the same year of age.

Each life course trajectory was described by the considered events in the order they were experienced for the first time. For example, the life course of a woman who had sexual debut at 17 years of age, married at 18 years, and had her first child at 20 years is represented by the ordered combination of the experienced events, the so-called event sequence (S)(U)(C). This sequence can be detailed by registering also the age at the events, obtaining the time-and-event sequence (S)(U)(C)(17,18,20). Whilst the number of event sequences is clearly limited, the number of time-and-event sequences is generally high, since each individual experiences the events at different ages. To simplify the latter sequences, one could refer to event sequence analysis [49,50], which defines suitable criteria to measure the dissimilarity between two time-and-event sequences. Cluster analysis can be applied to such dissimilarities, to obtain groups of individuals who are similar with respect both to the ordering and to the timing of the events. Unfortunately, in our case such approach led to unsatisfactory results (reported for completeness in S2 Text). First, the obtained clusters generally included individuals who experienced the same event sequence, without properly distinguishing between different time-and-event sequences. Second, the less frequent event sequences were not identified (even imposing a higher number of clusters), and they were rather assigned to clusters dominated by the most frequent sequences.

We therefore decided to focus only on the individuals’ event sequences (from now on, simply referred to as sequences), taking explicitly into account all the observed ordered combinations of events. The timing of events, which is clearly a relevant aspect of the life courses, was included in our analysis using the ages at the three events as independent variables in the statistical models for the association between HIV infection and sequences (illustrated in the next section).

Even when focusing only on the observed combinations, there is a relevant aspect to consider when dealing with time at event data. Indeed, some individuals presented complete sequences (i.e., they experienced all the considered events by the last interview). Instead, other individuals experienced only some events (at least by the last interview), and their life courses are consequently censored, since the inexperienced events might still occur later. Even so, the probability of experiencing such events clearly depends on the individual’s age at the time of the last interview. For example, consider two women, aged respectively 19 and 30 years at the moment of the last interview, who are characterised by the same sequence, (S)(U), and who experienced the two events at the same ages. For the first woman a problem of censoring arises, because she might become a mother after the last interview. Instead, the second woman had no children in the relatively large interval between her first union and the last observation period. Hence, her sequence can be reasonably considered as complete: the probability that she will give birth after the interview can be expected to be low.

To accommodate the incomplete sequences (thus limiting the problem of censoring) accounting for the individuals’ ages, we defined a new set of sequences built as follows. Consider the incomplete sequence (S)(U), which can clearly only evolve into (S)(U)(C). For each individual presenting the complete sequence (S)(U)(C), we derive δ_UC, the observed waiting time from (U) to (C), and fit a generalised linear model (with gamma distribution and an inverse gamma link [51]) relating δ_UC to the age at (U). We then consider the residuals, , and define the threshold as the 90^th percentile of their distribution. Clearly, cases with a residual are characterised by a time to the last event, here (C), much higher than the predicted one, , and thus experienced a delayed last event. Based on this reasoning, we further considered the time between the age at (U) and the age at the last interview (T_F), , and derived the predicted waiting time, , and the residual . Cases characterised by , were assigned the delayed sequence (S)(U) → (C). Such sequence indicates that, irrespective of the possible experience of event (C) after the last year of observation, these individuals delay childbearing compared to the other individuals in the sample with the same initial track. Conversely, we had to remove from the analysis those cases observed for a period shorter than , since the incompleteness of their sequence could not be accommodated.

The procedure just described was applied to all the incomplete sequences, using thresholds which, albeit different, all vary between 2 and 3 years. Note that the incomplete sequence (S) can evolve into different complete sequences, characterised by different waiting times between (S) and the subsequent event. Thus, to avoid an excess of intervention on data, we had to exclude cases presenting this specific sequence from the sample (S2 Table).

The final sample, selected and filtered as described in the previous sections, consisted of 5,307 individuals, of which 3,269 women and 2,011 men (S1 Table).

Statistical analysis

We modelled the probability of being HIV-positive as a function of the obtained event sequences using logistic regression and adjusting for background socio-demographic effects. The individual HIV status was derived based on the application of the dipstick assay to the dried blood spots. The socio-demographic effects included in the model were birth cohort (1961–1970, 1971–1980, and 1981–1990), urban (small township) or rural site (estate, roadside settlement or subsistence farming area), current age of respondent at the last available round (included up to a quadratic term), and age at school leaving (a proxy for the socioeconomic information on the level of education). To account for the age at the focal events, the ages at first union and at the first child (set equal to zero for those individuals who did not experience such events) were also inserted into the model. We ran separate analyses for women and men, as we expected the sequences to have structural gender-based differences in terms of ordering and timing [9].

To explore the interaction between sequences and known transmission mechanisms for HIV, we also fitted two additional models including two intermediate HIV risk factors. Specifically, we considered the years since sexual debut (calculated as the difference between the current age of respondent and the age at first sexual intercourse) and the type of experienced non-regular sexual partnerships (none, only premarital–before the first union, only extramarital–after entering the first union, and both premarital and extramarital).

We also performed a robustness check to investigate the effect of the incomplete sequences on HIV infection, focusing only on individuals with complete trajectories (details and results are reported in S3 Text). This approach tends to exclude younger people, who are more likely to experience incomplete and therefore censored sequences. Even so, it provides useful additional insights: in the case when a delayed sequence is associated with HIV infection, it is possible to assess whether the association is driven by individuals who actually experienced the event with delay, or by those who never experienced it.

All data cleaning and analyses were performed using the software R. The algorithm to build the sequences was written using SAS.

Sample description

As described before, our final sample consists of 3,269 women and 2,011 men, all sexually experienced, whose characteristics are reported in Tables 1 and 2, respectively. Men and women in the sample showed different socio-demographic patterns, mainly because of the older age of men (S3 Table). Included participants belonged mostly to the birth cohort 1971–1980 (43.1% of women vs. 52.8% of men, P<0.001), as reflected by the median age at last interview (28 years for women vs. 32 years for men, P<0.001). Moreover, women differed from men by living setting (19.1% of women living in an urban area vs. 23% of men, P<0.001), by age at school leaving (16 years old for women vs. 18 years old for men, P<0.001), and by HIV prevalence (28% of women vs. 24.1% of men, P<0.001). Prior research relates the latter difference to higher AIDS-related mortality among men [52,53], to higher women’s susceptibility to HIV infection [54], and to the possibility that women acquire HIV infection from committed partners who often also engage in extramarital sexual relationships [55]. Gender-based differences were also observed in the ages at events, with women showing a higher probability of experiencing events earlier than men, and in the exposure to non-regular sexual partnerships. In particular, the majority of women (63.8%) did not report any non-regular sexual relationship, whereas the majority of men (55.5%) reported at least one premarital sexual partner (S3 Table). Moreover, the proportion of women reporting both premarital and extramarital sexual partners was much lower than for men (2.5% for women vs. 22.4% for men, P<0.001), which might also be a consequence of social desirability bias [45–48].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Characteristics of women included in the sample.

Descriptive statistics on socio-demographic and sexual behaviour characteristics for women, stratified by sequence, and overall, Manicaland (Zimbabwe), 2000–2011. Sequences are sorted in descending order by HIV prevalence.

https://doi.org/10.1371/journal.pone.0163730.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Characteristics of men included in the sample.

Descriptive statistics on socio-demographic and sexual behaviour characteristics for men, stratified by sequence, and overall, Manicaland (Zimbabwe), 2000–2011. Sequences are sorted in descending order by HIV prevalence.

https://doi.org/10.1371/journal.pone.0163730.t002

Description of event sequences

Our results show that women and men experience similar sequences, characterised nonetheless by different frequencies and by different levels of HIV prevalence, which range from 20% to 54% for women (Table 1), and from 0 to 50% for men (Table 2). We decided to avoid the aggregation of low-frequency sequences with those with higher frequency, since this would lead to extremely heterogeneous groups of individuals in terms of both HIV prevalence and other important risk factors.

The predominant sequence for women is (SU)(C) (Table 1), characterised by a very short interval between the events, the lowest HIV prevalence (24%; average prevalence 28%), low levels of early sexual debut, moderate levels of early union and childbearing, and a low percentage of non-regular sexual partnerships (29%). This sequence can be considered as the normative transition to parenthood for women in Zimbabwe, as the average ages at events are consistent with those reported for the entire country [56]. On the other hand, we notice that this sequence, as well as the other sequences presenting a short span between sexual debut and union and very low proportions of non-regular sexual partnerships, report a relatively low prevalence, although actually not small in magnitude. This fact could be related to the sexual behaviour of their male partners [55], who, particularly when older than their spouses, might have been exposed for a longer time to HIV risk and have accumulated large numbers of non-regular sexual partners [12,57].

Conversely, sequences with higher HIV prevalence (above 35%) are generally characterised by a gap between sexual debut and first union, delayed (or inexperienced) union or motherhood, but also premarital pregnancy (Fig 1). More than one half of the women presenting these sequences show patterns of non-regular sexual intercourse (Table 1). In particular, women with both a gap between sexual debut and first union and premarital motherhood report a high average number of premarital sexual partners, around 2.4 (S4 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Age at events by sequence and HIV prevalence for women.

Boxplot of the age at sexual debut, age at first union, and age at first child, by sequence, sorted by median age at sexual debut (in ascending order) and coloured according to the HIV prevalence, Manicaland (Zimbabwe), 2000–2011.

https://doi.org/10.1371/journal.pone.0163730.g001

For men, the two most frequent sequences are (S)(U)(C) and (SU)(C), with HIV prevalence levels close to the average or lower (Table 2). Both the sequences are experienced by men younger than the average, but who differ in terms non-regular sexual partnerships. Indeed, a high percentage of men with the first sequence had premarital sexual partners, while almost half of the men with the second sequence never engaged in non-regular sex. On the other hand, in a similar way to what was observed for women, the sequences presenting the highest HIV prevalence (above 30%) are those characterised by delayed (or inexperienced) union or fatherhood, but also premarital parenthood (Fig 2). Moreover, most of men characterised by these high HIV-prevalent sequences experienced non-regular sex, with a high average number of premarital sexual partners (S4 Table), high rates of both premarital and extramarital sex [33], and long intervals between sexual debut and union (Fig 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Age at events by sequence and HIV prevalence for men.

Boxplot of the age at sexual debut, age at first union, and age at first child, by sequence, sorted by median age at sexual debut (in ascending order) and coloured according to the HIV prevalence, Manicaland (Zimbabwe), 2000–2011.

https://doi.org/10.1371/journal.pone.0163730.g002

Relationship between event sequences and HIV infection

Table 3 reports the adjusted odds ratios from the logistic regression relating HIV infection to the explanatory variables for women. The most frequent sequence, (SU)(C), was used as a reference category. According to Model 1, four sequences show a significantly positive odds ratio, namely, (S)(U)(C), (S)(UC), (S)(C)(U), and (SC)(U), with the last two sequences presenting premarital childbearing. These sequences are characterised by an early sexual debut (except for (SC)(U), where sexual debut occurs later), and by a median gap between sexual debut and first union of one year, which increases to four years for (S)(C)(U) (Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Adjusted prevalence odds ratios for women.

Adjusted odds ratios and 95% confidence intervals for the probability of being HIV-infected, based on logistic regression models, among women, Manicaland (Zimbabwe), 2000–2011. Sample size is varying because of missing values.

https://doi.org/10.1371/journal.pone.0163730.t003

As for the socio-demographic factors, living in an urban area is associated with greater HIV risk, while birth cohort, age at school leaving, and age at first union or first child turned out to be not significant. A significant negative curvilinear effect of the age at interview shows that HIV risk peaks at around 35 years of age and drops afterwards. Considering that the predicted HIV-adjusted life expectancy in Zimbabwe was between 36 and 38 years for both women and men [52], the observed result can be regarded as the consequence of HIV/AIDS-related mortality.

The results obtained for men are displayed in Table 4. The sequence (SU)(C) was used as reference also in this case, because of its low HIV prevalence and high frequency. The sequences positively associated with HIV infection are (S)(U) → (C) and (S)(C)(U), characterised respectively by delayed fatherhood (as confirmed by the robustness check performed on the complete sequences, reported in S3 Text) and premarital fatherhood. Moreover, both of them present an early sexual debut (Fig 2), and a gap between sexual debut and first union. Protective factors from HIV infection are being of a younger birth cohort, living in a rural area, and having a longer education. In a similar way to what was observed for women, we did not find any significant evidence of association between HIV infection and the age at first union or at the first child. As for the age at interview, its curvilinear effect indicates that HIV risk increases up to 32 years of age and then drops. Again, such decrease can be a consequence of HIV/AIDS-related mortality [52].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Adjusted prevalence odds ratios for men.

Adjusted odds ratios and 95% confidence intervals for the probability of being HIV-infected, based on logistic regression models, among men, Manicaland (Zimbabwe), 2000–2011. Sample size is varying because of missing values.

https://doi.org/10.1371/journal.pone.0163730.t004

Interaction between event sequences and HIV proximate determinants

We found that the length of sexual exposure and the type of non-regular sexual partnerships allow to understand the association between HIV risk and the sequences. Indeed, when these two additional variables are included in the model, most of the significant associations between sequences and HIV risk observed in Model 1 are no longer significant.

For women, there is no significant association between the length of the sexual exposure and HIV infection (Model 2, Table 3). Nonetheless, its inclusion in the model leads to a loss of the statistical significance for the sequences (S)(U)(C) and (SC)(U).

Conversely, a longer sexual exposure for men is associated with a higher HIV risk (Model 2, Table 4). When the variable is included, the association between HIV infection and both sequences (S)(C)(U) and (S)(U) → (C) is no longer significant. This may indicate that a long sexual exposure, likely initiated at an early age (Fig 2), moderates the association between the sequences and the HIV infection.

Nonetheless, the most relevant proximate determinant turns out to be the type of non-regular sexual relationships: its inclusion in Model 3 causes the loss of statistical significance for all sequences. For women, the experience of non-regular sex is an important transmission mechanism, with its odds ratios increasing in magnitude from premarital to extramarital sex, up to the combination of both types (Table 3). Along with the loss of significance, we observe a shrinking in the values of the odds ratios. This suggests that the experience of non-regular sexual relationships may explain part of the association between the sequences and HIV infection.

For men, we find a different pattern of association between non-regular sex and HIV infection. Indeed, the large majority of men experienced premarital sexual relationships, either alone or followed by extramarital relationships (Table 2). The high rates of non-regular sex, in combination with a longer exposure to sexual activity, produce the shrinking in the magnitude of the odds ratios, and this might explain the significance of the sequences in Model 1 (Table 4).

Sequences and HIV infection across birth cohorts: a sign of behaviour change?

In our models, birth cohort was inserted as a control variable. Nonetheless, the relevance of the different sequences might vary for different birth cohorts, and this could explain the secular decline in HIV in Zimbabwe from 24% to 14% over the study period [37]. Therefore, we investigated the possible change in the sequences’ distribution across the different birth cohorts, and studied the relationship between such differences and HIV infection (Fig 3). We noticed that, for both women and men, the frequency of normative and “safer” sequences, e.g., (SU)(C) and (SUC), increased over birth cohorts, while the reverse holds for less normative sequences, such as (S)(U) → (C), (S)(C) → (U), or (SU) → (C) (Fig 3, top row). Interestingly, we also found that the HIV prevalence characterising normative and safer sequences showed a decreasing trend over cohorts (Fig 3, bottom row). This might be due to a change in the sexual behaviour across birth cohorts, which favoured more normative and less risky transitions to parenthood (e.g., (SU)(C) and (SUC)), with a consequent decrease in the HIV prevalence. On the other hand, we observed that for some less normative and riskier sequences, e.g., (S)(U) → (C) and (SC) → (U) for women, or (S)(C) → (U) for men, the associated HIV prevalence rose, despite its decrease in frequency across cohorts. Even though this change was not statistically significant (as indicated by the confidence intervals in the bar plots reported in Fig 3, bottom row), still it might provide some evidence of the risk associated in more recent times with non-normative transitions to parenthood and to adulthood, in particular those characterised by delayed or inexperienced union or parenthood, or premarital parenthood.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Relationship between sequences and HIV across birth cohorts.

Change in the distribution across birth cohorts of the sequences and of their HIV prevalence, Manicaland (Zimbabwe), 2000–2011. Top row: bar plot of the distribution of sequences across birth cohorts, for women (left panel) and for men (right panel). Bottom row: bar plot (with 95% confidence interval) of the HIV prevalence associated with each sequence across birth cohorts, for women (left panel) and for men (right panel).

https://doi.org/10.1371/journal.pone.0163730.g003