Data source, study design, and study sample
This study is a cross-sectional secondary analysis using data from the STEAM. The study methodology has been published elsewhere . The study sample included 1015 survivors: trafficked males, females, adolescents, and children (aged 10–17 years) who reached the country of exploitation and attended post-trafficking assistance services in Cambodia, Thailand, or Vietnam.
A two-stage sampling strategy was used to identify individuals using post-trafficking services. First, 15 post-trafficking support service organizations were selected across the three countries (6 services in Cambodia, 4 in Thailand, and 5 in Vietnam) based on diversity of clientele (e.g., age, sex, sector of exploitation, and country of origin), service relationship with the International Office of Migration (IOM) country teams, and agreements with government agencies (e.g., support, referral, and service arrangements). The STEAM describes individuals who received post-trafficking services, regardless of differing legal definitions of trafficking and service eligibility criteria between countries .
Second, a consecutive sample of individuals were invited to participate in a structured interview within 2 weeks of admission to the post-trafficking services between October 2011 and May 2013. Participants were recruited only if the locally-trained caseworker or social worker determined that their participation would not cause harm to their well-being. Individuals in the sample were identified as trafficked by the local governmental and non-governmental referral networks and post-trafficking service providers. The response rate for the baseline survey was 98%.
Interviews were conducted by caseworkers or social workers from the agencies providing post-trafficking services. Interviewers received an intense one-week training provided by one of the principal investigators of the STEAM (LK) in collaboration with the IOM partners in each country. Data collection and double data entry were coordinated by IOM country offices, with oversight by the London School of Hygiene and Tropical Medicine (LSHTM).
Development of survey questionnaire and application
The survey questionnaire was based on the instrument used in a previous European study on health and sex trafficking  and adapted by the study team for the different study populations (various labor forms of exploitation) and the regions studied by STEAM. The interviewers also participated in adapting the questionnaire, which was pilot tested in the study settings. The survey included questions about socioeconomic background, pre-trafficking and post-trafficking exposures, living and working conditions during trafficking, violence and coercive factors, mental and physical health outcomes, and future plans and concerns. The instrument was translated into Khmer, Thai, Vietnamese, and Lao in multiple steps: professional translation from English to other languages, group translation-discussion processes with IOM counter-trafficking teams, pilot-testing, and review after back-translation into English.
A strict ethical and safety protocol was implemented based on the World Health Organization (WHO) Ethical and Safety Recommendations for Interviewing Trafficked Women . Ethical approval for the study was granted by the LSHTM and by the National Ethics Committee for Health Research in Cambodia, the Hanoi School of Public Health in Vietnam, and the Ministry of Social Development and Human Security in Thailand. Core ethical guidance included measures to ensure that participation was voluntary and confidential, assurance that declining participation would not affect the provision of support services, avoidance and management of distress, and the offering of options for supported referral for health or other problems. The secondary analysis was approved by the University of Michigan Health Sciences and Behavioral Sciences Institutional Review Board, eResearch ID: HUM00097096.
Anxiety, depression, and post-traumatic stress disorder symptoms measures
Anxiety and depression symptoms in the past week were measured by the Hopkins Symptom Checklist-25, a symptom inventory . It consists of 25 items: 10 for anxiety symptoms and 15 for depression symptoms. The scale for each item includes four categories of response (“Not at all,” “A little,” “Quite a bit,” and “Extremely,” rated 1 to 4, respectively). The anxiety score was calculated as the average of the anxiety items, while the depression score was the average of the depression items. The depression score has been correlated with major depression as defined by the Diagnostic and Statistical Manual of the American Psychiatric Association, 4th edition (DSM-IV) in several populations . A cutoff of 1.625 instead of the established value of 1.75 was used to identify symptoms of depression, as item 12 in the questionnaire (i.e., loss of sexual interest or pleasure) was excluded, given the nature of the study population . For anxiety, a cutoff of 1.75 indicated symptoms of anxiety, based on previous research on individuals using post-trafficking services and on studies of Cambodian, Laotian, and Vietnamese refugees with whom this instrument has been validated [31, 36, 40, 41].
PTSD symptoms in the past week were measured using the Harvard Trauma Questionnaire (HTQ) part IV, which includes 27 trauma symptoms . The first 16 items were derived from the DSM-IV criteria for PTSD and assessed the presence of the main PTSD symptom clusters: intrusive experiencing, avoidance behaviors, hypervigilance, and emotional numbing . The remaining items were developed by the Harvard Program in Refugee Trauma. These PTSD symptom items focus on the impact that the traumatic experiences may have had on the subject’s perception of his or her daily life (e.g., having difficulty dealing with new situations) [42,43,44]. Each question has four response categories: “Not at all,” “A little,” “Quite a bit,” and “Extremely,” rated 1 to 4, respectively. A total score was calculated by averaging the 27 items. A cutoff of 2.0 was used to assess symptoms of PTSD based on previous research on trafficked individuals accessing post-trafficking services [30, 32]. Although the HTQ has not been validated with the study population, it has been used in cross-cultural settings and among Southeast Asian populations (e.g., Cambodians) exposed to trauma [40, 45, 46]. This instrument has shown high sensitivity for identifying persons with PTSD when diagnosed by experienced psychiatrists in a clinical setting and according to DSM criteria . The HTQ has high reliability  and internal consistency [47, 48] and test-retest reliability ranging from 0.89 to 0.92 [48, 49].
Violence and coercion measures
To assess physical and sexual violence, standardized and validated questions from the World Health Organization (WHO) International Study on Women’s Health and Domestic Violence were used . These questions describe acts of physical and sexual violence commonly reported by trafficked individuals in post-trafficking services and shelters [4, 30, 34, 36]. For females, we created a three category indicator of: “no violence”, “physical violence only”, “sexual violence only”, and “physical and sexual violence.” “Physical violence only” indicated the experience of any violent acts such as: being kicked, dragged, or beaten up; being tied or chained, choked, or burned; having a dog released to bite or scratch; being threatened with a weapon, cut with a knife, or being shot at, experiencing punches, slaps, and hits, but no experience of sexual violence. “Sexual violence only” was defined by a positive response to one item asking whether the respondent was forced to have sex, but no experience of physical violence, and lastly, those in the “physical and sexual violence” category were exposed to both types. These categorizations were based on previous research with survivors of trafficking [26, 28, 34]. For males, violence (yes/no) was measured with the variable “physical violence with additional threats made with a gun, knife, or other weapon” as only six males reported sexual violence. This variable included all the acts of physical violence listed above with additional threats made with weapons. We also identified in descriptive analysis that the experience of violence in males was for the most part defined by this type of physical violence.
Two additional questions were used to assess coercion during the trafficked period for females and males: (a) “While you were in this situation, did anyone threaten to hurt you?” (yes or no) and (b) “During this time did anyone threaten to hurt your family or someone you care about?” (yes or no). These questions assess threats commonly made by traffickers that are considered hallmarks of the trafficking experience and are frequently used in studies of interpersonal violence [14, 28, 34, 50].
Covariates in this analysis were theory-driven and based on prior analyses of the STEAM [11, 32, 34, 51], and included age (10–17, 18–25, and 26 or above), country of exploitation and trafficking (Thailand, China, or Other [Cambodia, Malaysia, Vietnam, Indonesia, Mauritius, South Africa, and Russia]), and time in trafficking (1–12 and 13 or more months). Participants were asked which trafficking sector they were exploited in most recently. The grouping of sectors of exploitation was based on similarity of occupational exposures and risks, balanced with the need to group sectors together due to low counts in particular occupations. Sectors for females and males were grouped together as sex work, forced marriage, entertainment, and dancing (sex and entertainment industry); domestic work, cleaning, restaurant work, and begging (hospitality industry and begging); construction and factory work (manufacturing industry); and livestock, meat packing and preparation, agriculture, or fishing (animal and agriculture industry). For males, we further collapsed the sex and entertainment industry with the hospitality industry and begging sectors to be able to make meaningful comparisons, since there were few individuals in those sectors. Groupings of sectors of exploitation were also based on previous research that indicates that some of these labor sectors might share similar levels of violence [32, 34, 52].
Because of important differences in the distribution of violence, coercion and trafficking-related exposures, the analyses were stratified by sex. We calculated frequencies and conducted bivariate analyses with cross tabulations using Rao-Scott chi-square tests to account for the clustered structure of the data (i.e., post-trafficking service organizations) and assess associations between violence, coercion (threats) and covariates with anxiety, depression, and PTSD symptoms . Sex-specific unadjusted and adjusted modified Poisson regression models were conducted to estimate prevalence ratios (PRs) and their 95% confidence intervals (CIs) for the associations between violence and coercion with anxiety, depression, and PTSD .
Generalized estimation equations (GEEs) with an extension of the sandwich variance estimator were used to calculate a robust variance estimation that considers the level of correlation of observations within a cluster and produces standard errors of the estimates accordingly . This statistical approach was chosen because it is considered to be a direct and less-biased approach to estimating the PRs. This method corrects standard errors, considers clustered data [54, 56, 57], and it is robust to the specification of the working correlation structure chosen .
To determine the best fit of the model and the working correlation structure, we used the quasi-likelihood under the independence model criterion (QIC) statistic, which is robust to the selection of correlation structure . We chose an exchangeable correlation structure that assumes that all pairs of observations are correlated within a cluster. We fit separate and sex-specific binary modified Poisson regression models for each of the outcome variables (anxiety, depression, and PTSD). We fit a crude model for females and males with the previously specified demographic covariates only and separate crude and adjusted models for violence (model 1) and coercion (model 2) with each of the mental health outcomes. Given the low number of missing data (e.g., one female and one male were missing for anxiety, depression, and PTSD) we allowed for listwise deletion in all analytical models. All tests were two-tailed and analyses were performed using SAS version 9.4 (SAS Institute, Inc., Cary, NC). PROC SURVEYFREQ with a cluster and chisq statement were used for the descriptive analysis. PROC GENMOD was used with the robust variance estimator provided by the REPEATED statement with a cluster identifier that uses the method of GEE to estimate the model and give a proper estimate of the standard error of the PRs while accounting for clustering in the data.