Data

The Haiti DHS is a nationally representative household survey. Target households were identified using a multi-stage stratified random design, and the survey included detailed interviews with men and women of reproductive age. Data from 2005–06 were used for female respondents and those from 2000 for male respondents. The 2005–06 survey did not collect male migration data, and both are the most recent surveys to include migration questions. Because of the focus on youth and the limited range of migration history, the analysis only included respondents aged 15–20 at the time of the survey who were regular residents of the selected household.

Migration history was ascertained by responses to questions on the length of time individuals had lived in their current location, their previous type of place of residence (i.e. rural, town or city) and their current type of place of residence. From this information, age at migration and migration pattern (e.g. rural-to-city, town-to-city, rural-to-town) were determined. Two sets of migration-related time-varying covariates were coded. The first classified each person-year into one of five responses: city non-migrant, town non-migrant, rural non-migrant, city post-migration or town post-migration. The second included binary variables to indicate the year the individual migrated, one year post-migration, and two years post-migration.

Migration events, which re-occur and may be circular, are difficult to account for in cross-sectional surveys. The DHS only asked about the most recent migration episode, which means two individuals of the same age who both experienced a rural-to-urban migration event at the same time would appear identical in the data, even if one had experienced multiple migration events, and the other only one. However, evidence from the nationally representative 2009 Haiti Youth Survey, which collected complete migration histories (but not age at first sex), suggests that return migration and multiple migration episodes among adolescents are not common; among 15- to 19-year-olds, only 6% returned home after migrating, 14% migrated twice and 7% experienced three or more migration episodes (Lunde, Reference Lunde2009). Moreover, given migration motives and opportunity structures in Haiti, it is likely that most subsequent migration events after migrating to an urban area are lateral (e.g. one urban location to another) (Lunde, Reference Lunde2010). To create an analytic sample that corresponded to the study’s focus on migration to more-urban areas and to partially attenuate the consequences of incomplete migration histories that do not fully detail smaller residential changes and circular migration, inclusion was limited using several criteria. Lateral migrants were excluded. This eliminated 13.4% of the potential female sample and 6.7% of the male sample. Also excluded were those who migrated to less-urban areas (4.8% of the potential female sample and 2.3% of the male sample). The final sample included 1215 female youth (6185 person-years) and 829 male youth (3586 person-years). Excluded respondents were few at less than 20% and 10% of female and male respondents respectively, but if sexual initiation was earlier among those who were excluded it would bias results against finding significant associations between migration and the timing of first sex.

Respondents reported their age at first sex. Accurate reporting of sensitive information is challenging, and, most notably, female adolescents more often under-report their sexual experiences, whereas male adolescents exaggerate them (Zaba et al., Reference Zaba, Pisani, Slaymaker and Boerma2004). However, the same study found that the largest reporting discrepancies were in the number of sexual partners, whereas age at first sex was less subject to reporting bias. Moreover, particular survey formats may facilitate more accurate reporting. Survey formats such as that used in the DHS, in which age at first sex was asked without first asking if the respondent has ever had sex, are thought to provide more plausible responses (Curtis & Sutherland, Reference Curtis and Sutherland2004). Furthermore, it is unlikely that response bias differs by migration status. Therefore, the data’s usefulness in addressing this study’s primary research questions is not severely limited.

A number of characteristics could confound the association between the timing of first sex and migration. Affiliation with Protestant or Evangelical religious groups may be associated with later sexual initiation, and analyses controlled for Protestant/Evangelical versus other religion (primarily Catholic). Educational attainment is associated with the liberalization of sexual values, but it may also provide an incentive to postpone early sexual activity. Likewise, migration may facilitate educational attainment or be undertaken due to the lack of educational opportunities. Many respondents were still enrolled in school, and thus educational attainment was limited to whether they completed primary school. The economic and educational status of the family may also be associated with the timing of first sex and whether youth migrate, and controls for household wealth and parental education were included. The household wealth index was calculated using information about asset ownership reported by the household head (Rutstein & Johnson, Reference Rutstein and Johnson2004). Female respondents reported their parents’ educational attainment, which was classified by whether they completed primary school (not available for male respondents).

Analysis

Discrete-time event history analysis was used to analyse the relationship between age at first sex and migration. This approach models the probability that an event (first sex) will occur at any given age (Singer & Willett, Reference Singer and Willett2003). Respondents retrospectively reported age at first sex, which eliminates the problem of left-censoring. Within the sample, approximately half of respondents had experienced first sex at the time of the interview, and information about age at first sex for remaining respondents was unknown (i.e. right-censored). Event-history analysis offers advantages in such situations, because it incorporates all available information from right-censored observations and does not erroneously assume either that the event does not occur or that it occurs at the point of censorship (Singer & Willett, Reference Singer and Willett2003). First sex before age 12 was extremely rare (<1% of the sample); therefore the exposure period in all models began at age 12.

The basic form of the model predicts logit h(t _ij ), which is an approximation of the hazard of experiencing first sex for individual i during period j. To test whether the relationship between first migration and timing of first sex can best be described by the adaptation, disruption or selection hypotheses, three separate models with differing specifications of migration were estimated for male and female adolescents. Model 1:

$$\eqalignno{ {\rm logit} \ h(t_{{ij}} )=\ &#x0026; \beta _{0} {\plus}\beta _{1} {\rm Town}_{{ij}} {\plus}\beta _{2} {\rm Cit}y_{{ij}} {\plus}\beta _{3} {\rm TownMig}_{{ij}} \cr &#x0026; {\plus}\beta _{4} {\rm CityMig}_{{ij}} {\plus} \beta _{5} {\rm AgeGrps}_{{ij}}^{{}} {\plus}\beta _{6} {\rm Char}_{i}^{{}} $$

includes a series of time-varying dummy variables for which the coefficients indicate whether the individual is currently living as a non-migrant in a rural area (reference group), town (β ₁) or city (β ₂) as a non-migrant or is a migrant living in a town (β ₃) or city (β ₄). The coefficient β ₅ corresponds to a vector of time-varying dummy variables for age groups (ages 12–14, 15–17, 18–20); β ₆ corresponds to a vector of time-invariant characteristics that may potentially confound the association between migration and timing of first sex, i.e. religion, educational attainment, household wealth and parental education (female models only). As a robustness check to directly compare models for male and female adolescents, the female models were estimated excluding the variables for parental education. The effect sizes in the models without parental education did not differ significantly from the models that included parental education (results are not shown.)

In Model 1 the coefficients β ₁–β ₄ describe the likelihood of experiencing first sex in any particular year compared with rural non-migrants. The coefficient for non-migrant town residence (β ₁) is then compared with the one for town residence post-migration (β ₃), and the coefficient for non-migrant city residence (β ₂) is compared with city residence post-migration (β ₄). The difference in these values compares sexual initiation for life-time town residents and migrants to towns and also compares life-time city residents with migrants to cities. Significant differences between these coefficients would suggest that migrants behave differently from similar youth who already reside in these destinations, thus providing evidence for either the disruption or selection hypotheses. A lack of significant differences between life-time residents and migrants suggests that the two groups behave similarly, which provides evidence for the adaptation hypothesis or suggests that migration and the timing of first sex are unrelated.

In Model 2:

$$\eqalignno{ {\rm logit} \ h(t_{{ij}} )= \ &#x0026; \beta _{0} {\plus}\beta _{1} {\rm MigYr}_{{ij}} {\plus}\beta _{2} {\rm Post}1{\rm Yr}_{{ij}} {\plus}\beta _{3} {\rm Post}2{\rm Yr}_{{ij}} \cr &#x0026; {\plus}\beta _{4} {\rm AgeGrps}_{{ij}}^{{}} {\plus}\beta _{5} {\rm Char}_{i}^{{}} $$

coefficients are estimated for three time-varying binary variables, including the year of migration (β ₁), one year post-migration (β ₂) and two years post-migration (β ₃). As with Model 1, coefficients are also estimated for time-varying indicators for age groups (β ₄) and time-invariant characteristics (β ₅). If the effects of the migration timing variables (β ₁–β ₃) are significant, it suggests that young people are modifying their sexual behaviour according to the timing of migration, which provides support for the disruption hypothesis. More specifically, if the coefficients are significant and positive, indicating that they are more likely to begin sexual activity during this time, it suggests that migration may foster a risky environment for adolescents’ sexual well-being. If the coefficients are significant and negative, consistent with delayed sexual activity near the time of migration, it suggests that migration may guard against earlier sexual debut.

In Model 3:

$${\rm logit} \ h(t_{{ij}} )= \beta _{0} {\plus}\beta _{1} {\rm RtoT}_{i} {\plus}\beta _{2} {\rm RtoC}_{i} {\plus}\beta _{3} {\rm TtoC}_{i} {\plus}\beta _{4} {\rm AgeGrps}_{{ij}}^{{}} {\plus}\beta _{5} {\rm Char}_{i}^{{}} $$

coefficients are estimated for binary indicators representing the three migration patterns, including rural-to-town (β ₁), rural-to-city (β ₂) and town-to-city migrants (β ₃). The variables are time-invariant and represent whether the individual ever migrated; non-migrants are the reference group. Again, coefficients are estimated for time-varying age groups (β ₄) and time-invariant characteristics (β ₅). Significant effects for the migration coefficients (β ₁–β ₃) suggest that individuals who engage in the specified migration patterns experience unique patterns of sexual debut, which supports the selection hypothesis.

Data

After examining how migration is associated with the timing of sexual initiation, longitudinal data from the HYTS, a panel study that recruited a cohort of rural youth as they completed primary school (6^th grade), was used to examine if sexual and reproductive health knowledge and endorsement of premarital sex change as young migrants are adapting to their new environments. The end of primary school is a point at which limited opportunities lead many youth to migrate, and at the second wave of the survey, migrant youth (21.5% of the sample) were re-interviewed at the destination. This design affords comparisons between migrants and non-migrants from the same point of origin and offers advantages over designs that simply compare rural-to-urban migrants with rural non-migrants.

Participants were recruited from school records of all students who completed their sixth grade exams in the summer of 2011, within a defined geographical area in rural south-east Haiti. Initial interviews occurred between August and September 2011, when a parent or guardian was also interviewed and background information was collected. The second interview occurred between February and March 2012. Youth who were eligible to be included during the first wave and were not located were invited to participate in the second wave (eleven respondents). A total of 223 youth participated; 213 were interviewed in the first wave and 215 in the second. In the case of eight respondents who could not be interviewed face-to-face during the second wave, interviews were conducted over the telephone. Overall, 91% of respondents were interviewed in both waves, and multiple imputations were used to account for the remaining respondents with incomplete data on the key indicators.

Each interview lasted approximately one hour, and participants were interviewed in Haitian Creole, their native language. To balance accurate data collection on sensitive topics with the protection of confidentiality and reassuring parents of their child’s welfare, interviews were conducted in the view of family members at a distance where responses would not be audible (e.g. far end of the yard).

Comparing the HYTS and the 2005–06 DHS, HYTS respondents were on average 15.9 years old, and in the DHS rural sixth graders were on average 16.5 years old. Both were well over their age-for-grade, but HYTS respondents were somewhat younger, possibly a consequence of educational expansion during the 5-year interim. The HYTS households were larger than households in the DHS (6.4 vs 4.7 members), probably due to difference in the sampling frames (i.e. all household versus those with a sixth grader present). To compare household wealth between the two surveys, HYTS households were matched to DHS households with identical asset profiles based on asset ownership, housing quality and sanitation quality. Rural DHS households were distributed among the five quintiles from wealthiest to poorest in the following proportions: 5%, 10%, 34%, 30%, 31%. The HYTS households were distributed as 6%, 10%, 6% 71%, 8%, respectively. The HYTS households were notably poorer than the national population and were heavily concentrated in the fourth poorest quintile.

Variables

To measure sexual and reproductive health knowledge and endorsement of premarital sex during the early migration period, two indices were constructed. Sexual and reproductive health knowledge – which focused on how to prevent HIV and pregnancy – was assessed by coding responses to open-ended questions. Youth were asked, (i) ‘Please tell me what you know about HIV and AIDS’; (ii) ‘What can a person do to prevent HIV and AIDS?’; (iii) ‘Sometimes a person may want to have sex without becoming pregnant. Can you tell me what someone can do to avoid becoming pregnant?’ (iv) ‘What do you think you might do in order to have the number of children you want?’ and (v) ‘When you hear “family planning” (planin in Haitian Creole), what is it?’

For each open-ended question, responses were recorded by the interviewer, and after each response, the interviewer probed the respondent for additional responses to the same question. Responses to the questions were coded by the author to reflect the knowledge domains that are typically taught in national sexual health curricula for youth.

Sexual and reproductive health knowledge scores could range from 0 to 12. A point was awarded for each of the following components of HIV knowledge: (i) identifying that HIV is an incurable illness, (ii) knowing that HIV could be transmitted by sex, (iii) knowing that HIV could be transmitted by blood/needles, (iv) knowing that HIV could be prevented by using condoms, (v) knowing that HIV could be prevented by maintaining a mutually monogamous sexual relationship, (vi) knowing that HIV could be prevented by avoiding sexual activity, and (vii) not providing any incorrect information (e.g. HIV is transmitted by mosquito bites). A point was also awarded for each of the following components of pregnancy prevention knowledge: (i) identifying condoms as a means of pregnancy prevention, (ii) recognizing that family planning or contraceptives are a means of preventing pregnancy, (iii) knowing that a doctor could provide you with means of preventing pregnancy, (iv) identifying any specific family planning method, and (v) not providing any incorrect information (e.g. that pregnancy could be prevented by not having sex during menstruation). Initially, HIV and pregnancy prevention knowledge were two different indices, but they behaved similarly in the analysis and were therefore combined.

To assess respondents’ endorsement of premarital sex, respondents were asked the extent to which they agreed with a series of seven statements on a four-point Likert scale: if it was acceptable for a female youth to have sex with a friend, someone she loved and with a boyfriend, the same questions from the perspective of a male youth, and if was acceptable for female youth to buy condoms. The items were generated based on beliefs articulated during preliminary fieldwork. Respondents could indicate their responses using a touch card that showed the responses in written format along with corresponding images of different quantities of beans and allowed respondents to indicate their response without stating it aloud. Each response was coded from 0 to 3 with higher numbers indicating more agreement, and the seven responses were summed for a total score ranging from 0 to 21. Internal reliability was calculated using Cronbach’s alpha and found to be good for both time points (T ₀ α=0.72; T ₁ α=0.78).

Interviewers used a life-history calendar to solicit information about the timing of respondents’ important life events, including all migration episodes, romantic partnerships, school enrolment and progression, work experiences and significant family events. At the first wave, the life-history calendar included information about all events during the respondent’s lifetime, and at the second wave it was updated to include events between the two waves. Migration between the two waves was limited to rural-to-urban migration by study design (recruitment only in rural areas), and did not include lateral migration episodes.

Having a boyfriend or girlfriend and exposure to media sources were additional factors that could influence the key dependent variables and vary over the 5-month interval. Whether he/she had a significant-other at each time point was coded from the life-history calendar. Whether the individual had watched television during the previous week had a ‘yes’ or ‘no’ response. In both rural and urban areas, television is often watched in public locations where a television owner uses a generator and charges admission to see a movie or show. In many urban areas, television service is more widely available, and adolescents may watch television at home or at a friend’s home.

Analysis

To examine whether there were differences in the overall level of knowledge and attitudes between migrants and non-migrants, whether knowledge and attitudes changed between the two time points and whether change in these domains occurred differently for male and female migrants and non-migrants, multilevel models for examining change in individuals over time were estimated (Raudenbush & Bryk, Reference Raudenbush and Bryk2002). First, a model that included main effects for time, migration and gender, along with time-varying controls for having a significant-other and television viewing, was estimated. Then interactions, including a three-way interaction among time, migration and gender and all lower-order interactions, were added to the model. The full model, including potential interactions, is written as:

$$\eqalignno{ Y_{{ij}} =\ &#x0026; \beta _{0} {\plus} \beta _{1} T_{{ij}} {\plus}\beta _{2} {\rm MIG}_{j} {\plus}\beta _{3} {\rm MALE}_{j} {\plus}\beta _{4} T_{{ij}} {\rm MIG}_{j} \cr &#x0026; {\plus}\beta _{5} {\rm MALE}_{j} {\rm MIG}_{j} {\plus}\beta _{6} {\rm MALE}_{j} T_{{ij}} {\plus}\beta _{7} T_{{ij}} {\rm MALE}_{j} {\rm MIG}_{j} \cr &#x0026; {\plus}\beta _{8} {\rm OTH}_{{ij}} {\plus}\beta _{9} {\rm TV}_{{ij}} {\plus} \rmu _{j} {\plus} {\rvarepsilon}_{{ij}} $$

where $$Y_{{ij}} $$ represents the dependent variable (either sexual and reproductive health knowledge or endorsement of premarital sex) at time i for person j. β ₁ is the coefficient for a time-varying covariate for survey wave (T=0 or 1) and in the absence of interactions indicates the degree to which the dependent variable changed across the two waves. When all interactions are included in the model, β ₁ indicates change over time for female non-migrants. The variable MIG indicates whether the individual ever migrated and is time-invariant. In the absence of interactions β ₂ indicates whether migrants differ from non-migrants at the first wave; when all hypothesized interactions are included, it indicates whether there are differences between female migrants and non-migrants at the first wave. The value of β ₃ indicates whether there are gender differences, and when all interactions are included in the model, it indicates gender differences among non-migrants at the first wave. β ₄ is the coefficient for a migration by time interaction, and in the context of all other interactions it indicates whether the change between the two survey waves differs between female migrants and non-migrants. β ₅ describes the combined effect of being both male and a migrant at the first wave, β ₆ describes the change over time among male non-migrants and β _7, the coefficient for the three-way interaction among time, gender and migration status, describes change over time among male migrants. The coefficients β ₈ and β ₉ represent the effects of two time-varying control variables, i.e. whether the individual has ever had a boyfriend or girlfriend and whether the individual watched television during the previous week. The individual specific error term $${\rm \rmu }_{j} $$ accounts for unobserved characteristics of individuals that do not vary across the two time points (e.g. education, parental characteristics). Because of the short interval between surveys waves, there are a limited number of potential time-varying characteristics that can be included as control variables.