The Internet, it seems, can reflect the health of the body politic, though we may not like what we see. Consider, for instance, an analysis done by economist Seth Stephens-Davidowitz, as described by him in The New York Times earlier this month. More than 700,000 Google searches, he found, were performed last year by individuals seeking information on how to perform their own abortion. Some, he notes, used the phrase “how to have a miscarriage,” others inquired into the use of Vitamin C or bleach or a blow to the abdomen as an abortifacient, and a small minority queried the very words “how to do a coat hanger abortion.”
More disconcerting, however, are the trends he traces: those states with the highest frequency of these sorts of searches were (for the most part) the very same states that heavily restrict access to legal abortion. At the same time, he describes a sharp increase in such searches following 2011, a year when legal abortions began to be increasingly restricted in many places. Admittedly, he acknowledges, this is a preliminary analysis, deserving more academic study. Still, it suggests how this method of research can be used as a type of population-level microscope—in this case, so as to observe the alarming state of reproductive health care access in the United States.
This—the use of “big data” culled from the Internet for medical research—is the precise topic of a stimulating new book, Crowdsourced Health: How What You Do On the Internet Will Improve Medicine. It is written by Elad Yom-Tov, an academic who works from within the world of tech. Yom-Tov previously held positions in the research outfits of IBM and Yahoo, and is now a researcher at Microsoft. His book makes the case that what we divulge in cyberspace—whether through searches or posts on social media—constitutes an invaluable trove of data that can shed light on a wide range of medical issues. It’s a slim but effective volume, albeit more a description of his own research than a synthesis of the field. Nonetheless, it reveals both the promise and the pitfalls of such methods, and—just as notably—their potential perils.
At first blush, the overall premise seems sound. When medical illness—or, more commonly, apprehension thereof—strikes, the first place many (most?) of us turn is the Internet. Yom-Tov explores some of the reasons for this. Compared with a health care professional, for instance, the Internet is always available, has no other patients waiting, and—critically—permits a degree of anonymity when we are at our most vulnerable. (I can personally recall my first experience with this as a young teenager: a game of “spin-the-bottle” had me later neurotically combing the Internet for days on end in pursuit of definitive facts on the possibility of contracting HIV through kissing. For better or worse, it was many years before any similar concerns would be again raised).
Perhaps just as important, the Internet doesn’t charge copayments. Early in his book, Yom-Tov observes the role this factor plays when he looked at what people search for after they (or a relative) are given a diagnosis of cancer. “One of the terms that appear most often in queries in conjunction with different types of cancers,” he notes, “is ‘free.’” A cancer specialist tells him that this could be the result of people searching for “disease free survival,” a common metric of efficacy for cancer treatments. Alas, that was not the case: “[W]hen we looked at the queries closely,” he notes, “it became apparent that the context was ‘free diagnosis’ and ‘free treatment,’ probably because the people who submitted the queries had no other option for obtaining treatment.” As with Stephens-Davidowitz’s analysis, such a finding helps reveal the gaps within our health care system. Yom-Tov rightly points to uninsurance and underinsurance—serious and persistent deficiencies of the American health care system—as potential factors contributing to this phenomenon.
His larger point, however, is that we can draw on this deluge of Internet data to understand—and indeed improve—our health. Take, for instance, his research on anorexia nervosa, which occupies one chapter. He argues that the Internet is key to understanding anorexia, not merely because we can learn about anorexia-sufferers from what they share on it, but because the Internet may itself play a role in the disease’s propagation. He cites, for instance, one estimate putting the number of “pro-anorexia” websites at over 600. He also notes that other general use websites (like Flickr) contain what he calls “pro-anorexia content.” Such websites and postings are not simply supportive of anorexia sufferers, he describes, but instead explicitly promote the condition as a “lifestyle” choice instead of a disease: For instance, “thinspiration” sites provide pictures of anorexic celebrities that positively accentuate their gaunt bodies, while other sites provide tips on how to contend with meddling family members. Given the role that such content could potentially play in the development of anorexia, Yom-Tov fairly argues that the Internet is an important arena in which to study it.
He and his team ask some interesting questions. One example: Does Internet exposure to anorexic celebrities lead, in some sense, to anorexia? To answer this, they looked at the Internet searches of more than 9 million people and identified certain search terms that plausibly suggested a diagnosis of anorexia (for instance, such phrases as “tips for anorexia” or “how to become anorexic?”). They also assigned “anorexia” scores to various celebrities. They found that searching for more than one celebrity with a high anorexia score increased the hazard of later developing possible anorexia (again, based on suggestive search terms) by about nine-fold. “Apparently,” Yom-Tov concludes, “viewing celebrities who are perceived as anorexic is correlated with later making queries that are typical of an anorexic.” Whether this is actually a cause-and-effect relationship seems questionable, though Yom-Tov draws on data from Twitter—in which media reports on anorexic celebrities instigate “waves” of tweets that in turn inspire searches suggesting the development of anorexia—to argue that it is. He contends that the way the media describes these individuals—whether they are portrayed as ill or not—is a critical factor in this process. As he concludes,
information on the Internet aids people in becoming anorexic. A major part of this information, and especially being aware to it, is driven by media reports about celebrities who are perceived as suffering from an eating disorder. Positive or indifferent reporting on these celebrities seems to cause people to develop anorexia. However, when these reports indicate that this celebrity is sick with anorexia, such a path does not appear, perhaps because people do not aspire to be sick—they want to be thin.
He likens this to the long-described phenomenon in which news reports about suicide provoke more suicides by similar methods (the so-called “Werther Effect”), and concludes from this that the media should approach anorexia similarly to how it approaches suicide: with an awareness of its potentially infectious nature. Specifically, he argues, reporting about celebrities with anorexia should emphasize that they are ill.
These are interesting findings, though I’d venture that the provocative suggestion that the Internet plays a truly causal role in the development of anorexia requires more examination. If the Internet were to disappear tomorrow—if the “end of Twitter” were finally realized—would the incidence of the disease actually fall? This isn’t Yom-Tov’s point, but it’s a hypothetical worth considering. Anorexia, after all, predated the Internet. And overall incidence rates of the disease, according to one 2012 review, remained relatively unchanged for the last few decades of the twentieth century. It’s possible, but not entirely clear, that things could now be changing: Another study, for instance, found a significant rise in overall eating disorders, but not anorexia itself, in the first decade of this century in the United Kingdom. Is technological change contributing to such trends (assuming they are borne out elsewhere), or are other social or cultural dynamics the primary drivers? More work is needed.
In other chapters, Yom-Tov describes findings that range from the provocative to the pedestrian. One example of the former, for instance, is his study into whether Elisabeth Kübler-Ross’s famous “five stages of grief” actually can actually be observed in some quantitatively-provable realm. To do this, his team looked for people who “suddenly developed an intense interest in a specific type of cancer,” as evidenced by their Yahoo searches, and reasonably concluded that those individuals (or their friends or family) had cancer. They then looked at the types of webpages these 20,808 people subsequently visited. They divided these pages into eleven discrete categories (using an odd form of piecework human labor available through the Internet, Amazon’s so-called “Mechanical Turk”), and then divided the categories into “hidden stages” (using mathematical modeling). Coincidentally or not, their findings conform, in some very rough quantitative sense, to Kübler-Ross’s schema: “The best model turns out to be the one that has exactly five hidden stages. This, though indirect, is one of the first quantitative pieces of evidence for a 40-year-old model.” They make some other interesting observations using this approach, for instance that people move through these “stages” differently depending on whether they are looking for information on acute or chronic forms of cancer.
There are, however, some important caveats here. First, of course, these were individuals with (presumably) new diagnoses of cancer (or their family and friends), not individuals who learned they were dying or even necessarily in grief (there is, of course, sadly some overlap). Nor are his stages the same, in a qualitative sense, as those of Kübler-Ross: They were determined by mathematical modeling, and not associated with particular human emotions. Such work therefore reveals both the power of this mode of research, and it’s limits. Who really are these individuals, and what are they actually experiencing? Web searches can only say so much.
Finally, there is also the important issue of the potential dangers of this sort of research, an issue Yom-Tov is appropriately attentive to. “The data we leave on the Web when we ‘surf,’” he admits, “are highly revealing of who we are and what we are. Anonymity of data is no guarantee of complete anonymity.” He describes, for instance, the story of how the search information of some 650,000 AOL users became publicly available for research uses. Though the data was theoretically made anonymous, by adding up the many highly personal searches of users, he notes, it became clear that it was possible to identify individuals. Though quickly taken off the web, the data was immediately duplicated, and according to Yom-Tov, it is still available online today (where it will presumably persist in perpetuity). The very permanence of such data breaches is what makes them so frightening: Once our privacy is violated, it can never truly be restored.
On a similar note, he describes the controversial “Facebook Experiment,” in which the company changed the order of users’ feeds (some saw happier posts on top, others saw sadder ones) in order to investigate how it impacted their moods. This, he notes, provoked a great deal of discussion as to the ethics of the experiment. (It wasn’t done with the typically-required institutional review board approval needed for interventional studies, for instance). What are we to make of this? My own immediate take is strong skepticism towards research that basically amounts to corporate medical experimentation. Yom-Tov, however, makes a somewhat compelling counterpoint. As he describes, all of these companies are essentially already conducting these sorts of experiments all the time, adjusting various features of their sites and examining how it affects customer satisfaction. There is little controversy in this as long as it is done strictly for the purpose of profit production, and not for research. Isn’t this a bit of an odd dual standard? “If we allow such experimentation,” he asks, “shouldn’t we allow the use of data collected from such experiments to be used for medical research?” Perhaps, but this then raises the question of whether Facebook should have been allowed to perform the experiment to begin with. After all, this isn’t simply an instance of observing people on the Internet: It’s an intervention, however minor, in their actual lives.
This matters, because it is not entirely inconceivable that such online mood interventions could have potential real world consequences. For instance, the Guardian quoted tech thinker and entrepreneur Clay Johnson, who wondered on Twitter whether a social media company could aim to alter people’s moods online so as to sway political events: “Could the CIA incite revolution in Sudan by pressuring Facebook to promote discontent?” he tweeted. “Should that be legal?” There are also health consequences to consider. Imagine, for instance, a hypothetical and more drastic version of this experiment in which hundreds of millions of users are randomized into one of two groups: In one, breaking news about violent events went straight to the top of the feed, whereas in the other, posts that had pictures of baby animals had priority. Over time, might one of these groups experience greater anxiety or depression? Could there even be a statistically demonstrable difference in suicides between the two groups? It seems unlikely, but it’s worth pondering—particularly if it were conducted on a population-level scale.
Still, I’d contend that we should be even more concerned about the potential use of these types of methods for various non-scientific purposes. For instance, this type of internet-data research is increasingly being examined as a potential tool for political surveillance. An article headlined “Military-Funded Study Predicts When You’ll Protest on Twitter,” in Defense One, a defense-focused publication, describes how one Office of Naval Research-funded study sought to figure out ways to predict whether one’s “next tweet will be part of a protest.” “The real value, according to the researchers,” the article’s author Patrick Tucker writes, “lies in predicting how big a political storm could be before it hits.” As he notes, some have seen a potential for serious abuse in such research: Such data could theoretically be used to squelch political dissent, or for other unsavory purposes.
This is, of course, the fine line we always walk with science. The Internet no doubt constitutes a massive, if messy, repository of our thoughts, neuroses, desires, and unspoken fears. It would be folly not to study it. At the same time, we must be aware of its limits—intrinsic to its aggregated, impersonal, disembodied form—while at the same time keeping a watchful eye on those who would misuse it.