Sysomos matched these keywords to any Facebook story (a posted item of content by a user) or any type of Facebook interaction that can be a reshare (a reposting of an original story to another user connected to the sharing user), a comment (a text comment made by a connected user on the original story or on a prior comment), or a reaction (one of several emoji representing emotions, such as positive, negative, empathetic, surprise, and love).

All counts for numbers of stories and interactions are unique, by Facebook’s construction of nonoverlapping categories of story, reshare, comment, and reaction. Counts of authors are more complex. Within a category, the number of authors is the unique number of authors. For example, if 45,000 women commented on an article, these are nonduplicated authors. Across categories this may not hold, as the same author may post several stories, comment on other stories, and react to many others.

Accordingly, we cannot add the numbers of authors across the different categories of interactions. For example, 1.1 million unique authors making reactions and the 0.4 million unique authors making comments cannot be added to obtain 1.5 million authors, because this resulting sum double counts women doing both. However, the actual total is no smaller than 1.1 million and no larger than 1.5 million. We conservatively report only the lower number and use phrases such as “…at least…” in reporting these totals.

Sysomos reported to us summary aggregate statistics such as totals, time-based trends such as subtotal by day, content-based subtotals, keyword prevalence, other word prevalence in context of keyword, and most popular website links that were posted or shared. Importantly, Facebook explicitly limits some aggregate data to just the top 10 items within a category and limits all aggregate data to items with at least 100 instances. This is due to confidentiality concerns and the ability otherwise to potentially reidentify individuals. We provide selected excerpts of these data, including tabular and graphical summaries.

In this pilot study, we were most interested in the type of content that was being shared. Links to website content originate in a story. Such stories can be authored by women who embed a link in a posted story, or authored by a marketer or news media organization that uses a shortened (eg, bitly) Web address to allow ease of use and visibility. Sysomos allowed us to identify the actual 10 most popular links and the frequency of each, by interaction type and content of link.

We clicked through all of these links and examined their content in detail. One study team member, a physician scientist (MH), manually categorized their content retroactively. This led to us identifying 5 mutually exclusive and collectively exhaustive categories to which all shared links belonged. These categories were e-commerce related to breast cancer, celebrity breast cancer information, breast cancer advocacy and charity, noncelebrity breast cancer information, and unrelated to breast cancer. This last category arose because, although a user may have been commenting on a breast cancer news story, they may also have been sharing an unrelated news item in the same post, and hence both were captured. These categories have not been externally validated and should be considered hypothesis generating only.

This study was conducted using completely deidentified, aggregated summary data provided by a third party, and accordingly did not involve human participant research and did not require an institutional review board determination or approval in our institution.

Stories and interactions were most often authored by women aged 45-54 years (Figure 2). We observed a substantial spike in volume on November 28, 2016, the Monday on which many news sites shared a picture of a bald Ms Shannen Doherty (an American actress) and her mother, immediately prior to Ms Doherty’s radiotherapy.

A search of mentions of “Doherty” in any link shared, reshared, or otherwise interacted with showed 6700 mentions by 6600 unique authors over the month, respectively 0.4% and 0.6% of the overall totals for the month.

Common terms relating to breast cancer and screening mammography mentioned in any context included “mammogram” (266,000 interactions, or 16% of the month’s total interactions), “lump” (26,600, 1.6%), “abnormal mammogram” (4400, 0.3%), “scars” (4000, 0.2%), “BRCA” (3800, 0.2%), “dense” (3200, 0.2%), “DCIS” (3000, 0.2%), “high risk” (2900, 0.2%), and “compression” (1000, 0.06%).

Across all interactions, the 10 most popular links accounted for a total of 33,600 interactions, or almost 2% of monthly total interactions (Table 1). Multimedia Appendix 1 provides the actual links.

Links to e-commerce related to breast cancer represented the plurality of interacted links at 36% of all top interactions and were the most popular to be reshared with connected users (59% of all top reshares). These tended to represent a for-profit or not-for-profit organization that was selling items in some way connected to breast cancer themes over the Web. The most common of these were thebreastcancersite.com, hopinghand.com, and makaiclothing.com. Other links in this category were also aimed at recruiting email addresses for future direct marketing by offering giveaways (thebreastcancersite.com).

The next largest category was links to celebrity breast cancer information, which represented 26% of all top interactions and originated in stories such as from the television program The Today Show, E! Online, and similar sites. Here the most prominent foci were Ms Shannen Doherty, Ms Danielle Spencer, and Ms Robin Roberts, in that order. Almost all of these interactions were emoji-based reactions.

The third largest category represented links to breast cancer advocacy and charity sites, with 15% of all top interactions concentrated among links to 2 sites. The American Cancer Society donation page (“Making Strides Giving Tuesday”) was the second most popular link interacted with overall, with 4900 interactions or 0.3% of all interactions, was the most reacted to, and was the fifth most often reshared link among all top links. The other site was the advocacy and information site of Susan G. Komen, which did not make the top 10 links overall, and represented only 100 commented-on links (the 10th most commented-on link).

Noncelebrity breast cancer information links constituted only 6% of all top interactions and included mostly personal blog stories, inspirational messages, and some traditional news media. The most prominent themes in this category were positive stories around the use of tattoos to mask the surgical scars associated with breast cancer surgery and a new breast cancer vaccine trial at City of Hope hospitals in Duarte, California.

Importantly, in this category were at least 700 shared links relating to mercola.com, a natural health advocacy site that presented a view against breast cancer screening, including multiple references to scientific studies and a recent article by Welch and colleagues [29]. This single link alone represented nearly 6% of all the top links in the 35- to 54-year age group.

Finally, more than 1 in 6 links were not in relation to breast cancer or screening terms. These presumably were stories, reshares, and comments in which a user conveyed multiple messages, some about breast cancer (hence they were selected by Sysomos) and some not about this.

We repeated our analyses to understand how interest and interactions changed across age groups (Table 2). Examining the totals of the top 10 links by age group, we found a clear increase with age. The most salient findings were the increase by age groups in e-commerce, peaking in the 45- to 54-year age group, and the complete or nearly complete lack of interest in those between 18 and 44 years of age in breast cancer advocacy and charity sites.

Additionally, we noted that noncelebrity-related news and information about breast cancer represented a larger share among the youngest users (50%) than among older users. We also noted the apparent complete lack of interest among the age group 65 years and older in celebrity-related breast cancer information.

While our study had several important strengths, including novelty, exhaustiveness, and national scale in the United States, there are several important limitations. Our data source, Sysomos, is a commercial reseller of data obtained indirectly from Facebook through another intermediary, Datasift. Data provenance, custody, and governance must be assumed but cannot be verified or guaranteed. For example, software errors could occur at each one of these handoffs, as well as within each segment of the data custody chain.

In particular, Facebook is the data owner, whose terms of service do not permit actual visualization of the original post or comment. To protect users’ privacy, all data were aggregated, deidentified, and mapped coarsely into topics. We therefore had no independent ability to confirm whether the reported statistics we obtained were accurate, representative, or exhaustive. Moreover, under our contract, data availability was limited to rolling 1-month lookback periods. Other restrictions motivated by privacy and imposed by the data owner include sampling only high-frequency items, limiting results to the top 10 items in a category, and masking results in which fewer than 100 Facebook users mentioned a term or shared a link. As a result, none of our results were able to provide a full view of the frequency distribution.

Neither we nor other researchers can subsequently return to historical periods beyond examining reports that were downloaded contemporaneously. Similarly, only a 30-day rolling period of aggregated data is made available by Facebook, Datasift, and Sysomos. This clearly further limits replication and error checking. For research purposes, while substantial information abides, much is lost during this process. This weakness does not seem to be one that will be alleviated, given legitimate concerns regarding online privacy [35].

Finally, while our research is internally valid, the extent to which it is externally applicable is not known. The particular month of data we looked at was almost immediately after a polarizing general election in the United States, in which health-related conversations (eg, Affordable Care Act, Planned Parenthood, women’s right to choose) were widely occurring. In other months, there might have been fewer mentions of breast cancer screening terms. Our research also explicitly required women to have access to the Internet, be a member of Facebook, and use English in their interactions. There are clearly large parts of US society in which one or more of these requirements are not met.

Future researchers may exploit other less coarse methods for obtaining online social media and social network data. Companies operating online survey panels such as Knowledge Networks, Inc [36], Qualtrics, and ClearVoice Research [37] can allow more representative surveys and more specific questions as to what women share on Facebook. Free resources can also be accessed through Google’s own Trends data—for example, to analyze searches related to Angelina Jolie’s disclosed prophylactic mastectomies [8], and to understand interest in public hospital quality reports [38]—or by exploiting the freely available Twitter data [39].

Examining novel data from the universe of mentions on the leading online social network regarding breast cancer screening-related terms provided an important but superficial and initial look at topics of great interest among all female Facebook users over 1 month. More work is needed using this novel data source and applying its insights to solving pressing public health problems, including the inadequate screening for breast cancer.