When scientists study how we think, they usually design their experiments to control as many variables as possible. Though essential for careful science, these expensive measures restrict research in many ways: most psychology studies are conducted on small groups of Western undergraduates over short periods. Most scientists agree we could learn more from longer studies of larger, more diverse populations, but until recently acquiring such data proved nearly impossible. Now the vast data sets of an online brain-training company, Lumos Labs, are offering insights that have been out of reach for traditional laboratories.

Lumos runs Lumosity.com, a popular brain-training Web site. More than 40 million participants of many ages and nationalities train on the site, some for many years, by taking psychology tests from home. The activities involve, among other things, rapidly solving arithmetic problems or typing words of various lengths when given a word stem. Some participants also volunteer health and lifestyle information.

In a recent paper in Frontiers of Human Neuroscience, scientists at Lumos, the University of Michigan and Duke University analyzed the information gathered from these psychology tests—the largest ever data set on human cognition. In one study, the team looked at data from upward of 100,000 diverse participants and determined that higher levels of daily alcohol consumption are associated with ever steeper declines in performance. Unlike earlier work on the cognitive costs of drinking, the new study revealed that each additional drink beyond two a day causes a greater drop in performance than the one before.

In another analysis, the scientists compared more than 22,000 participants of different ages to see how much they improved on tests of memory and reasoning over long periods. They confirmed previous findings that intuition and reasoning, or fluid intelligence, declines faster with age and improves less with training than memory for facts and skills, known as crystallized intelligence. Yet they also found that the ability to improve on tasks worsens with age regardless of how good you are at the outset—old dogma claims that improvement is easier if you start at a low level.

The scientists are careful to point out that the conclusions reached on this type of data must come with some caveats. Participants train from home instead of in a lab, leaving open to question many variables researchers cannot see or control for, such as dishonesty about age or behavior. In addition, the subjects still may not be representative of the general population—perhaps people who choose to use Lumosity differ in some fundamental way from those who do not. Nevertheless, the Internet offers a unique, cost-effective way of studying human cognition on a huge scale. Big data has already become central to genomics research, explains Lumos data scientist Daniel Sternberg. Cognition “is orders of magnitude more complicated.”