5
Collecting Human Trafficking Prevalence Data
The difficulty in collecting human trafficking prevalence data extends beyond the problem of varied definitions. As noted in Chapter 2, different types of trafficking are prominent in different sectors, and populations respond differently to certain data collection methods. Recognizing the problems in definitions and contextual differences, this chapter considers the two general approaches to collecting prevalence data: point-of-crisis contact and surveys and sampling.
DATA FROM POINT-OF-CRISIS CONTACT
The U.S. National Human Trafficking Hotline
Sara Crowe (Polaris) told participants that the U.S. National Human Trafficking Hotline, a program run by Polaris and funded by the U.S. Department of Health and Human Services (HHS), generates one of the largest lists of human trafficking victims in the country. The hotline operates 24 hours a day, and since 2007, has received reports on more than 50,000 trafficking victims, mostly in the United States. In 2018 alone, the hotline received information on more than 11,000 victims. Crowe considers data that come into the hotline as operational (or administrative) data: that is, even though the data are self-reported and useful in understanding the nature of human trafficking, the primary function of the hotline is to link victims with support services, not to collect data. However, the benefit of the hotline is that, unlike discrete or point-in-time surveys, the operational data received are continuously updated and can provide researchers with
unique insight, including the effectiveness of counter-trafficking communications, based on which populations are using the hotline. The significant drawback is that the data only reflect people who are aware of and have access to the support hotline; thus, the data are not representative of prevalence.
Crowe reflected back to the earlier question about the most common forms of labor trafficking (see Chapter 2) and how she listed agriculture as a primary industry when Sheldon Zhang had not. She noted that the U.S. Department of State gives the hotline number and employment rights information to everyone entering the United States on a temporary work visa, which may account for the relatively high number of calls from agricultural workers. But the hotline seldom receives calls from undocumented workers, who do not have access to the same information, which creates a marked gap in the data. Crowe also mentioned that there is a difference between the data needed to carefully define an issue and the data needed to respond to it. In addition, data collected during a point of crisis will often be different in nature from information obtained when a survivor is surveyed in a more stable situation. The hotline also fields calls from health care providers, who can use the hotline to report data on individuals or to obtain useful information to share with their patients.
Health Care Situations
Hanni Stoklosa (HEAL Trafficking) said that she has called the U.S. National Human Trafficking Hotline while acting as a service provider at HEAL. Health care providers are able to report data to the hotline that are compliant with the standards of the Health Insurance Portability and Accountability Act. Sometimes providers call the hotline so they can obtain useful information to share with their patients.
HEAL is a network of more than 2,500 professionals and survivors around the globe who are combating trafficking from a public health approach. Stoklosa noted how the intersection of human trafficking and health presents unique opportunities for prevalence estimation, but that it also presents several challenges. As with other vulnerable populations, trafficking victims often come to the attention of health care providers through emergency departments like the one in which Stoklosa works as a physician. She gave a scenario in which a forced laborer comes to the emergency department, is treated for a work-related injury, identified as a trafficking victim, and provided support. Stoklosa noted, however, that there are several assumptions relevant to prevalence estimation within this scenario: the worker having the opportunity to access health care; the health care professional being aware of and knowing how to assess trafficking victimization; the worker disclosing the trafficking to the health
care professional; and the instance of trafficking being captured in the patient’s record. She said that even when disclosure is made, adding the information to a patient’s record can prove challenging. HEAL has partnered with HHS on several initiatives that increase the capacity of health care providers to identify trafficking victims. These initiatives supplement some states’ mandated trafficking training.
Stoklosa said that despite receiving education about trafficking, doctors are human, and their unconscious biases have the potential to interfere with their ability to identify a victim in a clinical setting. She noted how trauma presents itself differently in different individuals and said that the media can also shape people’s perceptions of what a victim looks like. If the provider is expecting a victim to look a certain way—for example, thinking that sex trafficking victims are only female—or if a person is agitated or aggressive rather than sad and helpless, the doctor or other health care provider may not perceive the patient as a victim during a medical assessment.
Stoklosa told the group that her use of the word “assessment” rather than alternatives like “screening” is intentionally done, to move away from the mentality that a person must be diagnosed and rescued and toward a more survivor-centered approach that creates an atmosphere in which a victim would feel more comfortable disclosing sensitive information. She said it is important to note that many individuals who disclose being victims of trafficking are not in a place where they can or want to immediately leave their situation, and they may not be able to share all of the details of their condition. She added that the key is to respect the survivor’s innate wisdom, understand where they are coming from, and trust their personal safety assessment of their situation, because there may be things they are not able to share in that moment.
Stoklosa said that the language providers use during victim assessments is important. For example, instead of asking about trafficking outright, she encourages the use of open-ended questions, such as, “We’re seeing a lot of folks that are paying for their drugs by selling their bodies; has that ever happened to you or anybody you know?” Compassionate statements can also help build trust: “I care about your health, and I see your health is related to your relationships and the work that you do, which is why I am asking these questions.” However, she noted that health care providers still need to be cognizant of other factors that may inhibit disclosure, such as if a person’s basic living needs are not being met, if the person is withdrawing from drug use, or if there have been threats to that person or the person’s family. Survivors may also simply feel ashamed. The survivor-centered approach requires patience, and Stoklosa said that the demands on clinicians’ time do not always lend itself to that, but there should be a balance between the dimensions of care, with the ultimate goal being protecting a patient’s safety. Stoklosa uses the PEARR guidelines—provide
privacy, educate, ask, respect, and respond—when speaking to patients in the emergency department.
A Method for Hidden Populations: Key Findings
Michael Shively and Ryan Kling (Abt Associates) reported on a study funded by the National Institute of Justice that looked at the feasibility of collecting human trafficking prevalence data from vulnerable populations in such settings as homeless shelters, emergency departments, and correctional facilities. They conducted interviews in Hennepin County, Minnesota, and another city in the southwest border region of the United States.1 For this study, they did not preselect their sample: they talked to everyone 18 years and older in the selected shelters and jails and interviewed a randomized sample of emergency department patients (who were not in need of critical care).
Using the definitions of trafficking in the Trafficking Victims Protection Act (see Chapter 1), the study’s screener questions asked whether the respondent had been subjected to force, but it did not ask about fraud or coercion because of the myriad ways those two conditions can be interpreted. The survey asked about contact with various social systems (i.e., shelters, drug and alcohol treatment, child and family services, faith-based organizations, and immigration) over the individuals’ lifetime, with specific focus on the past year. In the analysis, the researchers accounted for the fact that some of the individuals interviewed may have had contact with more than one system over the past year.
The questionnaire also asked about trafficking victimization, initially modeled after the Vera Institute’s Trafficking Victim Identification Tool.2 Over time, however, Shively and Kling came to realize that the Vera tool was designed for clinical diagnostic use and that it worked best after the interviewer had established a relationship with the potential victim. Shively and Kling’s study involved screening individuals at their first meeting, so the tool did not serve them as well as they had hoped. They ultimately revised and condensed the interview questions and made them closed-ended. Shively said their priority was to do no harm to participants and to not cause distress or violate people’s privacy.
Of the 42 people they interviewed in Hennepin County, all of them had made contact with at least one system in the past 12 months. When the researchers broke the prevalence estimate down across the three venues,
___________________
1 Shively and Kling chose not to name the second city because data accessibility issues prevented them from completing the requisite number of cases there.
2 See https://www.vera.org/publications/out-of-the-shadows-identification-of-victims-ofhuman-trafficking.
they found that an emergency department was the most prominent contact point. The study found that 7 percent of those surveyed identified as trafficking victims, and they had all made contact in the past 12 months with one or more of the following three settings: the health care system, social services (such as the Temporary Assistance for Needy Families Program), and the justice system. Shively called these prior points of contact missed opportunities—earlier instances when individuals could have been identified as trafficking victims if routine screening had been conducted and when they could have received support services. He acknowledged that 7 percent is a very conservative figure and noted that he and Kling are aware that the method they chose can potentially work against finding people who are currently trafficked. Current victims may have less access to care, less mobility, and still actively under the control of a trafficker. They extrapolated an estimate from the data from these interviews and estimated that approximately 7,300 people in Hennepin County (population 1.25 million) have been trafficking victims in their lifetime.3
Kling said the data from this study are still fresh, and he noted some of the challenges to this approach, such as having to use different weights in the estimate for different settings. He and Shively are experimenting with different distributions, including negative binomial regression, and learning more about the venues they chose for personal interviews. They found that shelters do not see a significant rate of turnover—people are there for approximately 180 days—so the sampling frame remains relatively static over the course of the year. Jails, on the other hand, experience high turnover rates, but people captured there have a high likelihood of being captured in another service frame.
DATA FROM SURVEYS AND SAMPLING
Multiple Systems Estimation
Bernard Silverman (member, Planning Committee) posed four questions to the workshop participants: How does one analyze data when many characteristics about them are not available? How might results be presented to policy makers and the wider public? Is there any advantage to making data available in an easily accessible public repository so that there can be further research? What are the features of the data in this area that demand specific methodology?
In his role as the chief scientific adviser to the U.K. Home Office from 2010 to 2017, Silverman helped develop central components of the British
___________________
3 Shively and Kling are still working on the confidence interval for these data.
government’s 2014 modern slavery4 strategy, which led to the creation of the U.K.’s Modern Slavery Act. Silverman’s team applied multiple systems estimation to data from the National Referral Mechanism—a program run by the U.K.’s National Crime Agency that collects data on potential victims from many centralized sources. In the National Referral Mechanism’s original design, duplicate cases from multiple data sources were removed; Silverman and his team requested that, for the purposes of conducting multiple systems estimation, duplicate cases be added back to the database so they could study the sources and the groupings. They arranged the points of human trafficking contact into five primary source groupings: local authorities, nongovernmental organizations, governmental organizations, law enforcement agencies, and the general public. Silverman said confidentiality restrictions prevented them from obtaining additional (covariate) information that may have proven useful in understanding the cases that did not map to any of the five groupings. They used a mathematical model to estimate the number of cases not captured on any of the lists—the “dark figure.”5
Silverman said that using multiple systems estimation essentially quadrupled the initial estimate of human trafficking victims because it included as-yet-unidentified victims. The model he used estimated 10,000 to 13,000 potential victims. When a newspaper did a story on the findings, the title of the article noted 13,000 and did not mention the confidence interval, which he says is a common occurrence. The article itself did mention the lower-end estimate of 10,000, but it also noted that the National Referral Mechanism had initially estimated the prevalence to be less than 3,000. When the relevant government minister was asked in an interview why they thought there were 2,744 but suddenly recalculated to 13,000, she said, “Those were the victims we knew about, and now these are the ones we don’t know about.” He said that he believes that researchers can effectively communicate this kind of uncertainty to policy makers if they work closely together and build trust.
Silverman discussed his attempt to conduct similar research in a city in the United States and how trepidation from agencies over releasing their data prevented the researchers from obtaining a robust dataset. Because multiple systems estimation requires the examination of multiple data sources, he noted that getting buy-in from various organizations, though tough, is essential to the process. Silverman talked about specific
___________________
4 Silverman noted that although the term “modern slavery” is used more frequently in the United Kingdom than “human trafficking,” he would use the terms interchangeably throughout his presentation because the definitions are so similar.
5 See Biderman, A., and Reiss, A. (1967). On exploring the “dark figure” of crime. The ANNALS of the American Academy of Political and Social Science, 374(1), 1–15.
methodologies he used for sparse data tables and emphasized the importance of making data accessible in the public domain. He concluded by saying that researchers “mustn’t let the perfect be the enemy of the good or even the fairly good,” and he emphasized the importance of making data accessible in the public domain for transparency and future research. He feels it is important for researchers to socialize providers to the notion of sharing data; be clear about the uncertainties where they exist; and maintain high standards of replicability, transparency, and confidentiality.
Capture-Recapture Heterogeneity
James Johndrow (Stanford University) discussed his work on casualty estimation, conflict mortality, and methods to measure hidden populations. He used the term “capture-recapture heterogeneity” to describe how, in multiple systems estimation, not all of the individuals have the same probability of being observed by the sampling mechanism. Certain individuals are more visible than others, perhaps because some people are more socially visible and therefore likely to be known about by peers, or perhaps because some administrative lists collect data from certain subgroups more than others. Thus, if someone with a low likelihood of being observed does appear in the data, it could be an indication that several more similar cases exist.
Johndrow said it is not always easy to apply multiple systems estimation to human trafficking or to other phenomena affecting socially vulnerable populations, such as conflict mortality. Recording differences and errors, confidentiality restrictions, and incomplete information can make it difficult to identify the same record across multiple administrative datasets. He explained how his method of capture-recapture assumes no interaction between each list or dataset and that it provides two probabilities: the likelihood the individual is captured on any list at all and the likelihood that an individual is captured on each particular list. He referred to the unknown number of people who are not captured in the data at all as “Big K” (the equivalent of the “dark figure,” discussed above). Once the probability of appearing on any list is calculated, Johndrow applies the Horvitz-Thompson estimator6 to determine the appropriate multiplier to calculate the total size of that particular population. Horvitz-Thompson asserts that the overall population size is inversely proportionate to an individual’s or group’s probability of being observed. If individuals appear on the list, even though their probability of being observed is low, the multiplier is high; if the probability is high, meaning individuals with a given characteristic are likely to show up in the data, the multiplier decreases proportionately.
___________________
Johndrow said that the reality of measuring hidden populations is that if people are close to invisible, they cannot be estimated—but their possible existence needs to be explicitly incorporated into research models in order to achieve reasonable confidence intervals.
When asked whether this type of analysis can still be conducted if the probability of being on different lists varies greatly for each individual, Johndrow said yes, but he doubted there would be a significant difference between the lists; if a person has a low probability of appearing on one list in the set, it is likely that the probability will be low for other lists. Another participant asked how Johndrow accounts for the fact that the population is not static; he responded that he regards the lists as a point-in-time snapshot, and he factors migration into the data model.
Network Sampling for Estimating Hard-to-Reach Populations
Kyle Vincent (independent researcher and consultant) talked about his work estimating hard-to-reach populations using network/referral-based sampling, also known as link-tracing sampling. Vincent explained that link tracing seeks to illuminate and exploit potentially untapped social networks by selecting a subset of the population and asking them to nominate additional individuals for the study with whom they are directly linked. He said network sampling has the capacity to capture individuals who may be missed through traditional sampling designs (e.g., simple random or stratified sampling, systematic or cluster sampling). Network sampling designs occur in waves; for example, 10 initial participants will each be asked to provide a set of their contacts who meet the study criteria, then those contacts will be asked, and so forth. Because of the nature of these networks, Vincent said that the process often results in a heavily connected final sample because of overlaps across referral networks.
Vincent listed several contemporary network-based strategies, including snowball sampling; respondent-driven sampling (which is the most widely used technique); and the network scale-up method, which involves casting a wider net in the initial wave, asking respondents how they interact with the subpopulation of interest, and using administrative data to scale up and get an estimate for the target population size and characteristics. He then described the new method he developed to address the need for an efficient estimate for the population size—something he thinks has been lacking in the research literature but is of utmost importance to studying hidden populations. His goal for the new method was to generate a dataset robust enough for more sophisticated network analyses, which could give insight into population structure, shape, and topology. Vincent said his sampling strategy is resource intensive at the onset, but he emphasized the importance of starting link tracing with a high-quality initial sample.
Vincent’s strategy uses administrative data from partnering agencies to facilitate selecting the initial sample, looks at links both within and projecting from the initial sample, and extrapolates those characteristics to the larger population. He then looks at the characteristics using different network start points, such as treating the individuals identified in the first wave as if they made up the original sample. The link-tracing process can be adapted with each wave to capture individuals of higher interest. He said this approach works well for transient populations, and it can be applied in a spatial setting, similar to adaptive cluster sampling. Vincent and his colleagues tested his method on existing datasets in both the United States and abroad, such as empirical datasets based on injection drug users and commercial sex workers, and found it to provide efficient estimates of population size.
This page intentionally left blank.
