Indicators of Sex Trafficking in Online Escort Ads

Case file sampling

Case data from closed cases were eligible for inclusion based on several criteria: (1) case year (year the investigation began) was 2013 or later; (2) case falls into the categories of sex trafficking or trafficking-adjacent activities (e.g., prostitution, promoting prostitution); (3) case must involve online escort ads; (4) ads associated with the case were not ads for recruitment of victims; and (5) case detail must be available such as police report(s), prosecutorial files, indictments, and/or interview transcripts upon which to base a determination of trafficking. The research team used these case details to identify elements of trafficking that comport with the federal definition of human trafficking (see below), so that determination of positive and negative case status did not rely solely on charging decisions that can be impacted by outside factors such as plea agreements. For this study, a single case may contain multiple related perpetrators and/or victims even if the defendants were prosecuted separately. If selecting all available cases from fieldwork sites or all ads associated with each case was a possibility, this option was chosen. In instances where this was not possible, a process as close to random sampling as possible was employed with the ads, with some purposive oversampling of cases to ensure representation of important subcategories of trafficking victims and of cases negative for trafficking.

Ad sampling within cases

While most cases were associated with 10 or fewer ads, a few cases returned hundreds of related ads. Therefore, sampling of ads within the outlier cases was undertaken to reduce potential skew. First, we conducted a manual deduplication process in which identical ads that had been copied and reposted were removed. Next, a power analysis was conducted to determine the appropriate number of ads per case to achieve statistical power, while reducing potential skew. Based on a general linear model, which includes logistic regressions, a random sample of 30 ads for each outlier case was selected to reach a power of 0.8. Since the independent variables are binary, normality measures of skewness and kurtosis are not required (Seltman, 2018).             The final case file sample included 114 cases investigated in four states (California, Georgia, Texas, and Nebraska), involving 1,263 unduplicated ads. The ads within these cases were posted in 35 states and Ontario, Canada, indicating extensive travel involved in most cases. Seventy-six (76) percent of ads were classified as trafficking cases, 9% were negative for trafficking, and 15% of ads could not be classified due to insufficient case file information.

Data Source:  Case file data from five jurisdictions in four states. Online escort ads collected from three web scraper archives.

Collection Mode(s):   coded on-site observation; face-to-face interview; mixed mode; telephone interview; web scraping

Weights:   State weights were calculated to mitigate selection bias at the criminal case level and must be used in the analysis.

Unit(s) of Observation:   cases, ads within cases

Geographic Unit:   State