The Nature of Causal Problems in Social Inquiry

Every year, doctors, teachers, and government officials make choices that affect millions of people. To make good choices, they need to know what works. But just watching what happens in the world rarely gives the full story.

The Big Picture#

We want to make the world better. To do that, we need to know which actions actually lead to the results we care about. Does a new reading program raise test scores? Does raising the minimum wage reduce poverty? At first, these questions seem simple: just compare places that did the thing with places that didn’t. But those simple comparisons are full of traps. This chapter explains the most basic trap: the difference between seeing two things move together and knowing that one caused the other. We’ll see why our everyday observations often fool us, and why policymakers can’t afford to be fooled.

Why Causal Answers Matter for Policy#

Imagine you are a public health official in a small city. Doctors report that an unusually high number of babies are being born with serious neurological problems. Parents are frightened. You need to act quickly. Your first instinct is to look for something those families have in common — perhaps they all drank water from the same municipal well, or lived near a certain factory, or bought the same brand of baby food. But here is the problem: even if you find that every affected family bought the same brand of baby food, does that mean the food caused the harm? Or did those families all shop at the same store because it was the cheapest in their neighborhood, and something else about the neighborhood — old lead pipes, perhaps — is the real culprit?

This is not a made-up disaster. Real crises have happened because people confused correlation with causation. For decades, doctors told parents to put babies to sleep on their stomachs. They thought this reduced choking, based on early observations. Later, large careful studies showed the opposite: putting babies to sleep on their backs greatly reduces the risk of sudden infant death syndrome. A mistaken causal belief, built into official advice, likely cost thousands of lives.

So causal questions are not just puzzles for academics. They sit at the heart of every decision that aims to protect, heal, teach, or employ. When a government picks a school reform, a tax policy, or a vaccination campaign, it is betting that the action will cause a better result. Getting that bet wrong because of bad causal thinking can waste huge resources — and sometimes cause real harm. That’s why we need a clear understanding of what causation means and how we can (and cannot) spot it.

📝 Section Recap: Policy lives or dies by causal claims; confusing mere association with genuine cause-and-effect can lead to costly, even deadly, mistakes.

Correlation is Not Causation: The Common-Cause Fallacy#

Almost everyone has heard the phrase “correlation does not imply causation.” But what exactly is a correlation? In plain language, two things are correlated when they tend to happen together. If we notice that on days when ice cream sales go up, so do rates of drowning, we have a correlation. Does eating ice cream make people drown? Surely not. But then why do the two move together?

The answer is that a hidden third factor — hot weather — causes both. People eat more ice cream when it is hot, and they also go swimming more, which unfortunately leads to more drownings. Ice cream and drowning are both “downstream” effects of a common cause. In causal reasoning, this is the common-cause fallacy, and it is one of the most frequent ways we get fooled.

Common-cause fallacy: The error of believing that because two things occur together, one must cause the other, when in fact a hidden third factor is driving both.

A classic analogy is the relationship between a barometer and a storm. Right before a storm hits, the barometer’s needle drops sharply. No one would try to stop a hurricane by pushing the barometer needle back up. We immediately understand that the falling needle does not cause the storm; the drop in atmospheric pressure causes both the storm and the movement of the barometer. The barometer is merely a signal — a useful forecaster, but not a lever. In the social world, many of the patterns we see are like barometer readings. For example, we might observe that countries with more television sets per person tend to have longer life expectancies. Does owning a TV set add years to your life? Unlikely. Instead, both television ownership and health are products of a country’s overall wealth and infrastructure. Wealth is the common cause.

Sometimes the hidden third factor is not a single thing but a whole set of conditions. People who go to university tend to earn more over their lives than those who don’t. But university-goers differ from non-goers in many ways before they ever set foot on campus: they often come from wealthier families, had better early schooling, and have skills and motivations that push them toward higher education. If we simply compare earnings, we give the degree too much credit because we are mixing up the effect of education with the effect of all those pre-existing advantages. The barometer-storm logic applies: the degree might add value, but the earnings gap we see is not a clean measure of that value.

The fallacy becomes more dangerous when the common cause is not obvious. Suppose a charity runs a job-training program in one neighborhood and then measures employment rates. If they pick the neighborhood because it was especially poor (which is likely, since they want to help), the initial high poverty level — the very reason the program existed — will make it look like the program did little good, even if workers actually benefited enormously. The selection process itself is a hidden common cause of both the treatment (getting the program) and the starting condition (being worse off). Untangling that requires more than just after-the-fact comparisons.

📝 Section Recap: When two things vary together, a third factor often drives both; the barometer-storm analogy reminds us that signals and causes are not the same.

Two Ways to Think About Causality: Ontic and Epistemic#

When we say “X causes Y,” what do we actually mean? Philosophers have wrestled with this for centuries, but for our purposes it is useful to split the idea of causation into two complementary views: one about what causation is out in the world, and one about how we discover it.

Ontic causality: Causation as it exists in reality — the actual physical, biological, or social processes through which one event produces another.

Think of ontic causality as the gears and levers of the universe. If you drink a cup of coffee, the caffeine molecules bind to receptors in your brain, blocking a neurotransmitter called adenosine, which makes you feel more alert. That chain of events is the ontic causal story. It does not depend on anyone measuring it, believing it, or understanding it. It just is.

Epistemic causality: Causation from the point of view of our knowledge — the evidence and reasoning we use to infer that one thing causes another.

Epistemic causality is about how we, as limited observers, come to learn what causes what. We rarely have a perfect x-ray view of all the gears. Instead, we gather data, run tests, listen to experts, and build arguments. When a public health agency says “smoking causes lung cancer,” that statement reflects decades of accumulated evidence — laboratory experiments on animals, population studies, and finally a strong mechanistic understanding of how tobacco chemicals damage DNA. The agency did not need to watch every single cancer cell form; it pieced together an epistemic case strong enough to guide action.

Why does this distinction matter for policy? Because the two can come apart. An education minister might believe, based on sound ontic theory, that smaller class sizes help children learn, because teachers can give more individual attention. But if every study that tried to estimate the effect of class size on test scores is flawed (maybe because budget decisions created spurious correlations), then the minister lacks strong epistemic backing. She may believe in a true cause, but she cannot prove it. Conversely, a policy might have strong epistemic evidence — several careful studies showing that a certain after-school program boosts graduation rates — even though the exact ontic mechanism (is it the mentoring, the extra study time, or the snack?) is unclear. For a policymaker, epistemic causality is what justifies action. But a deep ontic understanding lets us predict whether the policy will work in a new setting.

The ideal is when ontic and epistemic lines of evidence converge. When we not only have a rigorous experiment showing that a malaria-prevention net saves lives (epistemic), but we also understand the biological mechanism — the net physically blocks the mosquitos that carry the parasite (ontic) — our confidence in the causal claim soars, and we are better equipped to adapt the intervention to different environments.

📝 Section Recap: Ontic causality describes the real-world cause-effect machinery; epistemic causality describes our evidence for it; good policy requires both, but epistemic strength is what gives us permission to act.

Why Observations Alone Often Mislead#

If the common-cause fallacy were the only obstacle, finding causes would still be hard, but doable. Unfortunately, data from just watching what happens — without stepping in — has several other traps.

First, there is confounding. Confounding is a broader version of the common-cause problem. A confounder is any variable that influences both the cause we are studying and the outcome we care about. In the earlier university-earnings example, family wealth is a confounder: it affects who goes to university and also affects later earnings (through networks, inheritance, and other channels). When confounders are present, a simple side-by-side comparison of outcomes gives a biased estimate of the true effect. The bias can make a harmless intervention look beneficial, or a beneficial one look worthless.

Confounding: A distortion of the apparent relationship between a cause and an effect, created by a third factor that influences both.

Second, we face reverse causality. Sometimes it is not obvious which way the arrow points. Consider the relationship between mental health and unemployment. Being unemployed certainly wears down mental health. But having poor mental health also makes it harder to find and keep a job. If we look at a snapshot of survey data and see that unemployed people report higher depression, we cannot tell whether unemployment caused depression, depression caused unemployment, or (likely) both. Observational data alone rarely gives us the direction of the causal arrow with certainty.

Third, selection bias distorts the picture. This happens when the individuals we observe are not a random slice of the population, but have sorted themselves (or been sorted) into groups in ways that are linked to the outcome. A classic example: studying the effect of a gym membership by comparing the health of gym members to non-members. Gym members are healthier to begin with, on average, because they are the sort of people who choose to exercise. The observed difference between the two groups overstates the benefit of the gym because it mixes the effect of the membership with the effect of being the kind of person who joins a gym.

Even the simple act of measurement can mislead. Suppose a city installs more police officers in neighborhoods with rising crime. After a year, crime in those neighborhoods is still high compared to the rest of the city. Does that mean the extra policing failed? Not necessarily. The high-crime areas were chosen precisely because they were already on an upward crime trajectory. If we fail to account for that pre-existing trend, we may conclude that policing does not work, when in fact it might have slowed the rise substantially. Purely observational data rarely hands us a clean comparison group that is identical to the treated group in all ways except the treatment itself.

These limitations are not just academic footnotes. They have repeatedly led well-meaning analysts to claim that a new policy “failed” when it actually succeeded, or to promote a “successful” program that would have happened anyway even without the intervention. The history of medicine is littered with treatments that were hailed on the basis of observational evidence — bloodletting, hormone replacement therapy for menopause — only to be overturned when better causal methods were applied.

📝 Section Recap: Observational data are riddled with confounders, reverse causality, and selection biases that make it impossible, by themselves, to separate true cause-and-effect from accidental patterns.

Pathways to Reliable Causal Knowledge#

If simple observation cannot get us to the truth, what can? There are two broad strategies, and they work best in tandem.

The first is building a strong mechanistic understanding of the causal process. This means opening up the “black box” between cause and effect and tracing each link in the chain. In the case of the stomach-sleeping advice for infants, a purely statistical look at population data had suggested an apparent benefit. But once researchers began to study the physiology of sleep, they discovered that when babies sleep on their stomachs, they are more likely to rebreathe their own exhaled air, leading to a dangerous buildup of carbon dioxide. The ontic mechanism pointed squarely toward an increased risk of death. That mechanistic insight overturned a decades-old policy. When we know exactly how a cause produces its effect, we are far less likely to be tricked by spurious correlations.

In the social world, mechanistic knowledge often comes from qualitative research: talking to people, understanding their decision-making, watching a program operate on the ground. If a new math curriculum raises test scores, is it because of the new textbooks, the training the teachers received, or the fact that the program brought more parents into the classroom? Knowing the mechanism helps us replicate the success in other districts that may lack one of those ingredients.

The second strategy, and in many ways the gold standard for epistemic causality, is the randomized experiment. In a randomized experiment, we assign people (or schools, or villages) to either receive a new intervention or be in a comparison group purely by the flip of a coin. Because the assignment is random, the two groups will, on average, be identical in every way — both measured and unmeasured — before the intervention begins. Any differences we see afterwards can therefore be attributed to the intervention itself, not to pre-existing differences. The coin flip breaks the back of confounding and selection bias, because it ensures that the only systematic difference between the groups is the treatment.

Of course, randomized experiments are not always possible. We cannot randomly assign presidents to countries, randomly cause financial crises, or randomly expose some children to toxic pollution just to measure the effects. When experiments are impossible or unethical, we must turn to careful observational designs that mimic the logic of an experiment as closely as possible — topics we will explore later. But even then, the gold-standard ideal gives us a benchmark against which to judge weaker methods.

For the policymaker, the practical lesson is this: the strongest case for a causal claim rests on both pillars. A randomized experiment showing that a job-training program raises wages becomes far more convincing when we also understand the mechanism — for example, that the program taught a specific software skill that local employers desperately need. The mechanistic knowledge makes us confident that the result is not a fluke and that it will travel to other contexts. Conversely, a rich mechanistic theory without any rigorous test of its real-world impact remains a hypothesis, not a basis for spending public money.

Think back to the public health official at the start of the chapter. She cannot simply look at overlapping maps and declare a cause. She needs to follow the chain of evidence: rule out common causes through careful comparison groups, test the biological mechanism in a lab, and — if ethically possible — run a trial. Only then can she act with the confidence that lives are being saved rather than time and money wasted.

📝 Section Recap: Reliable causal knowledge combines a clear ontic story of how a cause works with an epistemic design — ideally randomized — that unambiguously isolates the effect, allowing policymakers to act on evidence rather than hunch.

Summary#

We began with a simple truth: lots of things move together, but very few of those movements are genuine cause-and-effect links. Spurious correlations are everywhere, and relying on them can turn well-intentioned policies into failures, or worse, into tragedies. We saw how the common-cause fallacy hides a hidden third variable, much like a barometer prophesies a storm without causing it. We then distinguished between the real, physical nature of causation (ontic) and the evidence we build to establish it (epistemic), and we learned that policy decisions ultimately ride on the strength of our epistemic case. Observational data, for all their abundance, are haunted by confounders, reverse causality, and selection — forces that render them unreliable guides when used alone. Finally, we glimpsed the way out: combining a deep understanding of the mechanistic chain with the unconfounded clarity that a well-designed experiment provides. With these tools, we start to take causal claims seriously — and we begin to protect the people who depend on those claims.