Human Trafficking in Southeast Asia: Results from a Household Survey in Vietnam

ABSTRACT

The dimensions, causes and characteristics of the human trafficking tragedy are largely undocumented. Despite over two decades of efforts by national and local governments, the United Nations, academics, NGOs and other interest and advocacy groups, statistically reliable data on the issue are scarce. This lack of reliable data hinders almost every effort to combat human trafficking. Policymakers and researchers cannot answer many basic questions concerning the demographic, occupational, ethnic or other characteristics that might distinguish successful migrants from those more vulnerable to trafficking.

This paper has four principal objectives: First, we review the state of data collection on human trafficking to illustrate the need for better data on human trafficking. Second, we present our household survey methodology, which we seek to validate as a method for collecting statistically reliable data on human trafficking. Third, we present the results of a pilot study using the household survey methodology to gather information on human trafficking in Trà Vinh Province Vietnam. These include estimates on the extent of trafficking in the region, the predictors of trafficking vulnerability and predictors of knowing trafficking victims. Finally, we use the results of our pilot study to draw tentative suggestions for public policy. Our analysis reveals several areas that deserve more attention from governments and other interested organizations as well as areas where resources are not being deployed effectively.

KEYWORDS:

• Human trafficking
• household survey
• Vietnam
• Asia

Acknowledgments

We thank Toman Omar Mahmoud, Matthieu Chemin, Johannes Haushofer, Laura Hackney, Elizabeth Xiao, Mahvish Shaukat, Anna Alekseyeva and students at the Fulbright Economics Teaching Program in Ho Chi Minh City for valuable comments. The following Bates College programs all generously contributed funding to this project: the Phillips Fellowship Program, the Harward Center for Community Engagement and the Faculty Development Fund. Anonymized data and the full questionnaire are available upon request.

Notes

¹ 38 persons were reported as trafficked out of a total of 2,394 individuals. 38/2,394 yields a raw trafficking rate of 15.9 per thousand. When the data are weighted to account for the undersampling of non-migrant households, we find a trafficking rate of 9.6 per thousand households.

² “Rare” here does not imply rare in an absolute sense, but relative to the dataset. For example, a dataset with 100,000 observations, where only 1,000 observations see the event realized, would be considered rare. On the other hand, a dataset with only 50 observations where the event is realized in 25 instances would not be considered rare.

³ In data organized into G groups with N observations per group, the unconditional maximum likelihood estimation of group fixed effects suffers from incidental parameters bias, which can become quite significant for a set N as G increases. Conditional logistic regression estimates eliminate this bias by conditioning each individual contribution to the likelihood function on the sum of measures within each group, which removes individual level parameters from the likelihood function. As we have multiple cases where G (equal to 24 in this case) is greater than the number of households sampled in the village, we use conditional logistic regression when estimating our fixed effects.

⁴ Regression clustering at the household level yield similar results and are not presented here. When clusters are nested, as in the case where households are nested within villages, it is recommended to cluster standard errors at the higher level, in our case the village (see Cameron, Gelbach, & Miller, 2011; Pepper, 2002 for further discussion).

⁵ The inverse hyperbolic sine transform is $\mathrm{I}\mathrm{n}(\mathrm{x}+\sqrt{{\mathrm{x}}^2+1})$, which approximates ln(2) + ln(x), making it interpretable in the same way as a standard log transform, with the important distinction that it is defined at zero. This makes it well suited to dealing with wealth and income data, which often contains a large number of zero values (Burbridge, Magee, & Robb, 1988).

⁶ We follow WHO convention and define children as aged 14 and younger.

⁷ The correlation coefficient between average household education level and the difference between individual and household education level is 0. As a robustness check, we ran specifications including each of the education levels separately. These results are available upon request from the authors.

⁸ As a robustness check, we ran an Extreme Bounds Analysis (EBA) on the educational difference variable. EBA essentially runs a large number of specifications, one for every possible combination of explanatory variables. The largest and smallest estimated coefficient on the variable of interest are the bounds. If they have the same sign, the estimate is generally considered “robust.” The extreme bounds on our educational difference variable are 0.21 and 0.08. As an additional check, we ran a specification using marital status instead of employment as the omitted variable from our selection equation. This yields similar results, which are available upon request from the authors.

⁹ Alpha is calculated as $\alpha =(\frac{\mathrm{k}}{\mathrm{k}-1})(1-\frac{\sum {\sigma }_{\mathrm{Y}\mathrm{i}}^2}{{\sigma }_{\mathrm{X}}^2})$ where K is the number of questions, ${\sigma }_{\mathrm{X}}^2$ is the variance of all responses and ${\sigma }_{\mathrm{Y}\mathrm{i}}^2$ is the variance of each individual response.

Additional information

Funding

This work was supported by the Phillips Fellowship Program, Bates College, Lewiston Maine USA; Bates College Faculty Development Fund, Lewiston Maine USA; Harward Center for Community Engagement, Bates College, Lewiston Maine USA.