Large datasets are a means to an end; they are not an end in themselves.
The most widely discussed feature of big data sources is that they are BIG. Many papers, for example, start by discussing—and sometimes bragging—about how much data they analyzed. For example, a paper published in Science studying word-use trends in the Google Books corpus included the following (Michel et al. 2011):
“[Our] corpus contains over 500 billion words, in English (361 billion), French (45 billion), Spanish (45 billion), German (37 billion), Chinese (13 billion), Russian (35 billion), and Hebrew (2 billion). The oldest works were published in the 1500s. The early decades are represented by only a few books per year, comprising several hundred thousand words. By 1800, the corpus grows to 98 million words per year; by 1900, 1.8 billion; and by 2000, 11 billion. The corpus cannot be read by a human. If you tried to read only English-language entries from the year 2000 alone, at the reasonable pace of 200 words/min, without interruptions for food or sleep, it would take 80 years. The sequence of letters is 1000 times longer than the human genome: If you wrote it out in a straight line, it would reach to the Moon and back 10 times over.”
The scale of this data is undoubtedly impressive, and we are all fortunate that the Google Books team has released these data to the public (in fact, some of the activities at the end of this chapter make use of this data). But, whenever you see something like this you should ask: is that all that data really doing anything? Could they have done the same research if the data could reach to the Moon and back only once? What if the data could only reach to the top of Mount Everest or the top of the Eiffel Tower?
In this case, their research does, in fact, have some findings that require a huge corpus of words over a long time period. For example, one thing they explore is the evolution of grammar, particularly changes in the rate of irregular verb conjugation. Since some irregular verbs are quite rare, a large amount of data is need to detect changes over time. Too often, however, researchers seem to treat the size of big data source as an end—“look how much data I can crunch”—rather than a means to some more important scientific objective.
In my experience, the study of rare events is one of the three specific scientific ends that large datasets tend to enable. The second is the study of heterogeneity, as can be illustrated by a study by Raj Chetty and colleagues (2014) on social mobility in the United States. In the past, many researchers have studied social mobility by comparing the life outcomes of parents and children. A consistent finding from this literature is that advantaged parents tend to have advantaged children, but the strength of this relationship varies over time and across countries (Hout and DiPrete 2006). More recently, however, Chetty and colleagues were able to use the tax records from 40 million people to estimate the heterogeneity in intergenerational mobility across regions in the United States (figure 2.1). They found, for example, that the probability that a child reaches the top quintile of the national income distribution starting from a family in the bottom quintile is about 13% in San Jose, California, but only about 4% in Charlotte, North Carolina. If you look at figure 2.1 for a moment, you might begin to wonder why intergenerational mobility is higher in some places than others. Chetty and colleagues had exactly the same question, and they found that that high-mobility areas have less residential segregation, less income inequality, better primary schools, greater social capital, and greater family stability. Of course, these correlations alone do not show that these factors cause higher mobility, but they do suggest possible mechanisms that can be explored in further work, which is exactly what Chetty and colleagues have done in subsequent work. Notice how the size of the data was really important in this project. If Chetty and colleagues had used the tax records of 40 thousand people rather than 40 million, they would not have been able to estimate regional heterogeneity and they never would have been able to do subsequent research to try to identify the mechanisms that create this variation.
Finally, in addition to studying rare events and studying heterogeneity, large datasets also enable researchers to detect small differences. In fact, much of the focus on big data in industry is about these small differences: reliably detecting the difference between 1% and 1.1% click-through rates on an ad can translate into millions of dollars in extra revenue. In some scientific settings, however, such small differences might not be particular important, even if they are statistically significant (Prentice and Miller 1992). But, in some policy settings, they can become important when viewed in aggregate. For example, if there are two public health interventions and one is slightly more effective than the other, then picking the more effective intervention could end up saving thousands of additional lives.
Although bigness is generally a good property when used correctly, I’ve noticed that it can sometimes lead to a conceptual error. For some reason, bigness seems to lead researchers to ignore how their data was generated. While bigness does reduce the need to worry about random error, it actually increases the need to worry about systematic errors, the kinds of errors that I’ll describe below that arise from biases in how data are created. For example, in a project I’ll describe later in this chapter, researchers used messages generated on September 11, 2001 to produce a high-resolution emotional timeline of the reaction to the terrorist attack (Back, Küfner, and Egloff 2010). Because the researchers had a large number of messages, they didn’t really need to worry about whether the patterns they observed—increasing anger over the course of the day—could be explained by random variation. There was so much data and the pattern was so clear that all the statistical statistical tests suggested that this was a real pattern. But, these statistical tests were ignorant of how the data was created. In fact, it turned out that many of the patterns were attributable to a single bot that generated more and more meaningless messages throughout the day. Removing this one bot completely destroyed some of the key findings in the paper (Pury 2011; Back, Küfner, and Egloff 2011). Quite simply, researchers who don’t think about systematic error face the risk of using their large datasets to get a precise estimate of an unimportant quantity, such as the emotional content of meaningless messages produced by an automated bot.
In conclusion, big datasets are not an end in themselves, but they can enable certain kinds of research including the study of rare events, the estimation of heterogeneity, and the detection of small differences. Big datasets also seem to lead some researchers to ignore how their data was created, which can lead them to get a precise estimate of an unimportant quantity.