By: Carl Coates
April 12, 2026—I recently co-presented at a workshop with the top-drawer educators at Data Science for Everyone for some awesome North Carolina econ teachers. I described the presentation as “two great tastes that taste great together”, recalling a candy bar ad from my youth. There were so many positive takeaways from this experience that I wanted to expand on this pairing for our newsletter.
Economics is fundamentally an empirical discipline — its claims are only as strong as the data behind them. Data science is the discipline that teaches us how to curate and frame that data honestly and effectively. This pairing is natural, even necessary. But it comes with a shared vulnerability: both fields (economics and data science) are only as good as the questions they ask and the information they gather. Fundamentally, conclusions, predictions and answers rely on a foundation of data free from fallacies. These two fields, studied together, can build the kind of thoughtful citizen who neither trusts nor dismisses a number but rather interrogates it.
Data That Misleads
As Professor John List discussed just this month in his UChicago E4E micro course, data can be a minefield wherein even smart, credentialed people can get it wrong and will fall victim to fallacies, including:
- Confounding third variable: when a hidden third factor creates a false appearance of a relationship between two variables.
- Selection bias: when your sample doesn’t represent your population of interest, so your conclusions are skewed.
- Goodhart’s Law: when a measure becomes a target, it stops being a useful measure.
Being aware of these data traps is particularly important because data generated in academic research may inform policy development and implementation. Even peer review does not inoculate a claim against misleading data. It’s worth reviewing some examples to remind us of potential negative outcomes of misinterpreted data.
Robert McNamara oversaw the Vietnam War as Secretary of Defense and was, by training, a data systems analyst. His challenge was identifying metrics, which are data selected and framed for a purpose, that could measure whether the war was being won. Geographic dominance, the traditional indicator of success in a war, was useless in a guerrilla-based conflict. McNamara settled on Body Count: deaths of North Vietnamese Army soldiers and Vietcong fighters. Fewer enemy combatants was equated with military progress. As that number climbed, it created the impression of progress. But McNamara had turned a measure into a target, and when a measure becomes a target, it ceases to be a good measure. This was Goodhart’s Law operating at a disastrous scale. The relentless pursuit of Body Count actively undermined a harder-to-measure American goal: winning hearts and minds. Every death, combatant or civilian, generated mistrust and potential new recruits against U.S. forces. Body Count wasn’t tracking success; it was producing its opposite.
A less consequential but instructive misuse of data comes from public schools. In the early 2000s, researchers identified a consistent correlation: students who ate breakfast performed better academically in many measures, including higher grades, better focus, stronger attendance. The finding felt intuitively solid and was replicated across studies. School districts expanded breakfast programs in response, and many schools began providing breakfast on standardized testing days for all students specifically to boost scores.
The correlation was real. The causal inference was not. More rigorous analysis revealed a confounding variable: students who regularly ate breakfast before school were more likely to come from stable, well-resourced homes. Home stability was the stronger driver of positive academic outcomes, not breakfast itself. The breakfast program provided genuine benefits like nutrition and routine, but it addressed a symptom rather than the cause. Given the budget constraints most districts operate under, one might ask whether those resources would have produced stronger outcomes invested in academic intervention programs. Well-intentioned data had led to well-intentioned policy, built on a causal chain that was far murkier than the original study suggested.
Data That Clarifies
Technological progress, and its accompanying disruptions, are a fact of life in market systems. This reality was codified by Josef Schumpeter with his observation of Creative Destruction. The first significant technological change that led to the industrial revolution was the mechanization of the textile industry, from the late 18th century through the mid-19th century. The consequences of this transition from home production to factories as one’s place of work are varied and complicated. The productive capacity of machines such as the spinning mule, spinning jenny and power loom all but ended handweaving and hand spinning as viable employment by the mid-19th century.
History often focuses on the destruction part of Schumpeter’s assertion. The Luddites made their appearance on the stage of history early in the transition to a mechanized textile industry. The Luddites occupied factories and destroyed power looms to preserve their jobs. The pattern of new employment opportunities in yet-to-be-created industries had not been observed yet. The Luddite movement was short-lived (historians suggest 1811-1816 as its heyday) but the violent consequence of the decimation of cottage work that caused the movement only tells part of the story of textile mechanization.
Further research tells a more complete picture. Employment as power loom operators alone increased 100-fold over the same period in the table above. While handweavers’ employment plummeted after the 1835 peak, textile factories continued to grow, particularly in the woolen and worsted textile sector- by 13% and 20% respectively. And while wages collapsed for handweavers, factory workers’ wages increased nominally and in real terms throughout this same period. Prices of raw cloth and finished clothing decreased significantly and prices overall decreased by 37%. (CPI: 1800= 13.5, 1850= 8.4) Further complicating the picture was the growth of urban centers, where lack of facilities and housing would eat into an increase in real wages as city growth outpaced housing capacity.
The data tells something else that I always wanted my students to understand: it turns out, the story is complicated. Asking better questions of data and seeking out more information can produce a more useful picture. Certainly, in this historical case, there’s no question that the outcomes of technological change were varied. They involved winners and losers with individuals and governments taking adaptive actions both in the near term and in the long run.
A more complete yet complicated story powered by data can also help us consider potential predictions based on historical patterns, particularly important as we start to see the impact of increased usage of AI. Data about this newest of technological disruptions is limited, but we can investigate employment sectors that are exposed to AI, just as the handweavers of 19th-century England were exposed to power looms and spinning jennies. We can identify what kinds of employment might be substituted by LLMs (Large Language Models) and what kinds of employment could be complemented or augmented by LLMs. And we can bring some clarity to the anxiety that accompanies such technological disruptions.
Data That Teaches
Effective approaches to using data as a vehicle for learning are inherently constructivist. This is the assertion that people don’t learn passively. They build it actively through experience, inquiry, and challenges to their existing mental models. (See this E4E Lesson on Empiricism about the importance of data and potential data pitfalls.)
You didn’t learn to tie your shoes from an instruction manual; you have to trip over them and tie soggy shoelaces until you learn. You can’t learn to ride a bike by reading about pedaling and balance. You have to fall off a few times. Learning is iterative by nature, and data provides exactly that kind of constructivist opportunity in the classroom. A student who has been told that correlation isn’t causation will forget it. A student who has questioned spurious correlations and hunted for confounding variables develops durable, transferable causal literacy which makes them a better thinker in any context.
When students are given data that isn’t pre-framed for consumption, they can “notice and wonder”, returning to it as their content knowledge deepens to make new observations and revise their initial conclusions. This mirrors how both fields actually operate: economists don’t derive conclusions — they build models, test them against data, and revise. Data scientists don’t find answers — they ask questions, interrogate data, and refine their questions. Both fields model, explicitly, the habit of treating your current understanding as provisional.
The Confounding Variable of Citizenship
McNamara had access to ample data. The Luddites had data from their own observations. The breakfast researchers had data. What was missing, in each case, was the habit of interrogating it. The breakfast study highlighted above is a useful metaphor for much of civic life: we see programs expanded and policies initiated that treat visible symptoms but do not address the underlying cause. Mission creep, misallocation of resources, and reactive and ineffective policies will continue, not out of malice, but out of data illiteracy. A citizenry that is willing to ask, “what’s the confounding variable here?” is more likely to address the harder to measure and more uncomfortable underlying issues. Teaching economics and data science together is, in this sense, not just academic enrichment. It’s building the diagnostic capacity that self-governance actually requires.