I’m a 12th-grade student learning economics and stats. I understand that correlation doesn’t imply causation, but in economics, we can’t do proper experiments. So how is causation actually figured out using data?
I’m a 12th-grade student learning economics and stats. I understand that correlation doesn’t imply causation, but in economics, we can’t do proper experiments. So how is causation actually figured out using data?
Comments
It’s the dismal science. Only a few years ago economists figured out that people behave in ways that don’t always maximize economic value. They won a nobel prize for figuring that out.
In a couple decades when they start catching up to the rest of us maybe economists will start thinking about causation.
Economists, like other researchers that can’t perform a fully randomized experiment, such as doctors, must rely on “natural experiments” where the world presents them with an opportunity to research.
In science, there are 5 elements that prove causality:
Temporal Precedence: The cause must occur before the effect. This can be easily assessed in economics: did the causal variable get applied before the effect was seen in all cases (ex: injecting money into the economic via a tax rebate led to changes in spending behavior)
Covariance: The cause and effect must be related or correlated, and if you change the cause, the effect should change in some predictable way (ex: if you increase the amount of the tax rebate, the magnitude of the spending behavior will also increase).
Nonspuriousness: The relationship between the cause and effect must not be due to a third variable or confounding factor. This can be challenging to reach conclusively, so often relies on designing and running lots of different studies that all point to a consistent cause for the effect while ruling out other variables (ex: we didn’t see this spending behavior when inflation increased/decreased, when the elasticity of goods increased/decreased, when the interest rate changed, etc…the tax rebate was the only factor that explained the effect consistently)
Mechanism: There must be a plausible economic, psychological, biological, chemical, or other mechanism that explains how the cause leads to the effect (ex: a plausible mechanism might be that a tax rebate gives consumers more money, which results in an increase in discretionary spending. An implausible explanation would be that a black hole millions of light years way from us affected the brain chemistry of consumers to shop more).
Plausibility: The proposed causal relationship should be supported by existing scientific knowledge or theory. Importantly, future studies that are run would all continue to support your variable as the causal one (ex: other studies that focus on tax rebates, rebates that are non-tax related, on shopping habits, etc all consistently point to the conclusion that more money in your pocket increases spending habits).
This is a very simplified explanation.
Imagine two towns just on two opposite sides of a state border. One town is subject to state laws about, for example how schools are funded, and the other side isn’t. Now, you want to evaluate how well the school funding law works, you can compare these two towns, which because they’re so close by, are probably similar in most respects EXCEPT the school funding law* and then see how well schools are funded on one side Vs the other. This is called a “natural experiment” where nature itself randomises the treatment between two similar observations.
*The “probably similar” is doing all the heavy lifting, and there are formal ways of testing for this that is beyond the scope of an ELI5 but that I’d be happy to elaborate on if you’d want an ELI15
The field of statistics you’re looking for is Causal Inference, and there’s more than one method under that umbrella, but a popular one is a Natural Experiment/Difference-In-Difference, like another person commented.
A popular example is Card & Krueger’s Paper. I’m just gonna copy/paste part of the abstract:
> On April 1, 1992 New Jersey’s minimum wage increased from $4.25 to $5.05 per hour. To evaluate the impact of the law we surveyed 410 fast food restaurants in New Jersey and Pennsylvania before and after the rise in the minimum. Comparisons of the changes in wages, employment, and prices at stores in New Jersey relative to stores in Pennsylvania (where the minimum wage remained fixed at $4.25 per hour) yield simple estimates of the effect of the higher minimum wage…Relative to stores in Pennsylvania, fast food restaurants in New Jersey increased employment by 13 percent [ed. 2.7 full time employees per store]
The method is, you take two entities that are mostly the same before an observable change to only one of them, in this example a minimum wage increase. Then take the difference of the differences. In the above case, they looked at average number of full-time employees per store:
Before:
NJ: 20.4
PA: 23.3
NJ – PA: -2.9
After:
NJ: 21.0
PA: 21.2
NJ – PA: -0.2
Difference in Difference: -0.2 – (-2.9) = 2.7
This method is imperfect, as the assumption that NJ and PA are comparable can be questioned, but it theoretically implies causation more than a correlation would.
because economics is a more vibes-based area of study than astrology
Experimental economics is a thing and has been at least since the 80s. Vernon Smith was a pioneer. Lab experiments have long been a thing.
There are “natural experiments” that (kind of) allow casual inference.
Most experiments – yes, even in the natural sciences – are far less pure than they seem and rely on various rules and assumptions that are often tiptoed past. Most “empirical knowledge” turns out to be more theory-bound than commonly understood.
This is a persistent challenge with all social sciences, not just economics. You can’t put human behavior in a perfectly controlled testing environment like engineers can. You have to sort of take the data you have, isolate certain variables and account for things that will skew your statistics. This is why in most well-thought out articles about economics, they won’t say “X caused Y” they’ll say “a study conductive by the Z institute suggests that X likely caused Y.” You can’t isolate all variables but you can look at net effect overtime and get a good idea of what major changes happened during that time.