This study found small effects for a number of system-level identity variables, as well as post-level variables. Identity attributes relative to group norms, as predicted by SIDE theory, appears to play a role in predicting censorship in this anonymous space. Interestingly, offensive attributes were both static (relating to the person) and dynamic (relating to their behavior). Surveying the results of both the quantitative and qualitative findings reveals the following reoccurring themes:

Gender. Both correlations and regressional statistics indicate a substantial connection between gender, sexual orientation and censorship. Posts targeting (and ostensibly read by) women were much less likely to be censored then those authored by men. This may suggest that women have less rigid views on what qualifies a post’s presence on craigslist, or it may just be that men generally censor more frequently than women.

Sexual Orientation. Statistics showed that posts made to sections intended for sexual minorities (m4m and w4w) were more likely to result in censorship. Findings about the w4w section should be taken with a grain of salt due to the small sample size, but the m4m section proved predictive of censorship. This may relate to the quantity of sexual content (see below), an internalized homophobia, or a more active censoring readership.

Alternately, this may suggest a same-gender bias towards censorship. Both the m4m and w4w section are inhabited by authors and readers of the same gender and sexual orientation. It may be the case that readers who feel some level of familiarity with the author’s identity consider themselves better suited to judge the appropriateness of their posts. Posts that deviate from the reader’s acceptable range of behavior for their shared category may then end up censored.

Explicit Content. Explicit content may support an alternative to the “familiarity” hypothesis of censorship seen above. Sexual and explicit content emerged as a theme during the qualitative analysis, and suggests that reader censor content they do not find appropriate. Mild levels of sexual and explicit content is allowed to persist by the craigslist community, but the censored content seen here was more extreme. Sexual solicitations, often with graphic depictions of sexual relations were common in the censored data set. Readers may enter the craigslist system willing to tolerate some range of inappropriate content. Once that threshold is crossed, the post will be censored.

Anonymity. Two themes related anonymity emerged during this study. Posts that did not contain sufficient identity information as to even identify the author or target as individuals were often censored. In contrast, posts with too much identifying information were censored as well. Anecdotally, Missed Connections with photos of their targets are quickly flagged and removed from the site.

SPAM. One interesting finding was actually discovered while coding the dataset for SPAM. Surprisingly (although not unexpected), much of the SPAM in the Missed Connection section imitates an authentic post, including plasible identities and interactions. One SPAM campaign ended with a request to visit a competitor’s site where, the ad promised, additional information about the interaction was posted. Even though these posts had appropriately been censored, the effectiveness of this “social spam” inside of a social space like craigslist would be interesting to study.

Speed of Censorship. Another finding discovered during the course of this study was the speed with which posts are censored. The majority of posts that were removed from the system were eliminated within the first fifteen minutes of their submission. This makes sense given that the craigslist system shows the posts in reverse chronological order.

Who are these readers? Assessing the intent of the censorer was undoubtedly the most limiting factor of this study. Particularly during the qualitative potion, author behavior, rather than an identity attribute, seemed to more frequently predict censorship. While “actions speak louder than words”, identities such as “sexual” do not account for degree of that attribute, or the context in which it applied.

Conclusion. While I had originally expected post-level identity attributes to play a dominant role in this study, system-level attributes appeared to have a greater influence. This may be the result of the clear identities (particularly gender/sexual orientation) by which the craigslist system is organized. It may also be the result of the complicated array of content that actually exists inside censored posts. However, I suspect that the implicit identities outlined by the system play a substantial role in organizing reader and authors into sub-populations with their own censoring behavior. Future research should compare post-level identity content, comparing across system-level identity groups. This would serve as an excellent follow up study to test the idea of a “familiarity and appropriateness” model of censorship.