Correlation Does Not Imply Causation (When It Comes to Childhood Obesity)

By Reid Ewing

The June 2014 issue of the Journal of Planning Education and Research contains four articles on "healthy schools," meaning, in part, schools whose location and design promote physical activity. The cover of the issue, accordingly, has a graphic that compares childhood obesity and walking-to-school rates at two points in time, 1969 and 2012.

In 1969, 87 percent of U.S. students living close to school walked to school, and childhood obesity stood at seven percent. By 2012, the percentage of students walking to school had dropped to 31 percent, while the childhood obesity rate had risen to 17 percent. One might conclude (perhaps incorrectly) that one variable, a decline in walking to school, has caused another, a rise in childhood obesity.

For those not steeped in methodology, this is a teachable moment. The cover leads us to ponder the perils of causal inference. Correlation may not imply (or equal) causation. I suspect that practitioners are more likely to fall into this trap than academics are. The source of the figure, Nisha Botchwey, an associate professor of city and regional planning at Georgia Tech, writes in that issue: "The built environment CAN affect children's behavior, such as walking to school, which has dramatically declined in recent decades, and [CAN] contribute to childhood obesity." Botchwey's JPER cover image is not intended to equate causation with proximate events; however, its combination of two findings reminds us of a general human tendency. The wonderfully insightful book Thinking, Fast and Slow, by the winner of a Nobel Prize in Economics, Daniel Kahneman, puts it this way: "Our mind is strongly biased toward causal explanations and does not deal well with 'mere statistics.'"

To assert causation, three conditions must be met. First, variable A and variable B must be related to one another. This condition is met in the JPER example. From time series data, walking to school and childhood obesity have been inversely related to one another. Second, a proper time order must be established, with the causal variable A preceding the effect variable B. We cannot tell if this condition is met in the above example, as the two trends have occurred simultaneously.

Third, and most importantly, the relationship between variable A and variable B must not be due to some confounding, extraneous, or "third" variable. Again, we cannot tell if this condition is met as we have no information about other variables that may confound the relationship between school-related travel and childhood obesity. A confounding variable is one that is correlated with variable A and causally related to variable B. Potential confounders abound, as trends in mode of travel to school have been accompanied by many other lifestyle changes.

Students' environment has changed dramatically in many ways over 43 years, and we cannot be sure which contribute more to childhood obesity. Students take less physical education at school than they used to. They spend more time watching TV and playing video games after school. They eat more fast food and less home cooking than before. I could go on, but you get the idea.

So what would it take to draw causal inferences about the built environment, travel to school, and childhood obesity? First and foremost, it would require disaggregate data for a decent-sized sample of individual children in order to avoid a common statistical problem called aggregation bias. Then it would require socio-demographically similar students living in very different built environments vis-a-vis school access. Then it would require objectively measured physical activity levels for different students at different times of day, including just before and after school, so the researcher could distinguish between students walking to school and those being driven. Finally, it would require objectively measured weight and height for students in order to link sociodemographics, the built environment, and physical activity to body mass index.

In fact, a quasi-experimental study that meets most of these criteria has already been done by a colleague at the University of Utah, Barbara Brown, in the Department of Family and Consumer Studies. Brown's sample consisted of 187 fifth graders from two schools representing three communities: (1) a walkable new urbanist community, Daybreak; (2) a mixed community (where students live in a less walkable community but attend the walkable school so that part of the route to school is walkable), and (3) a less walkable community.

She and her graduate student coauthor, Robert Stevens, tested whether students living in the new urbanist community had greater accelerometer-measured moderate-to-vigorous physical activity (known as MVPA) compared to students from other communities. Community walkability was, in fact, related to more MVPA during the half-hour before and after school and, among boys only, more MVPA after school generally. In a related study, physical activity proved to be inversely correlated with student body mass index.

So the simple cover graphic of the most recent JPER issue, with its evidence of correlation, may eventually find support in carefully controlled studies like Brown's. Let's conduct more of those studies to see if there is causation.

Image: Walking to school and obesity rates of children ages 2-19 living within one mile of school. Source: Nisha Botchwey, Georgia Tech; courtesy JPER.

Reid Ewing is a professor of city and metropolitan planning at the University of Utah and an associate editor of the Journal of the American Planning Association. Past columns are available at mrc.cap.utah.edu/publications/research-you-can-use, including "The Perils of Causal Inference" from May 2007. The Stevens and Brown article, "Walkable New Urban LEED Neighborhood-Development (LEED-ND) Community Design and Children's Physical Activity: Selection, Environmental, or Catalyst Effects?" appears in International Journal of Behavioral Nutrition and Physical Activity, 2011.