Skip to content

“Thinking about the possibility of spurious correlation isn’t a matter of liking—it should be pretty much automatic.”

I agree with sociologist David Weakliem when he writes the above sentence.

Here’s the full paragraph:

Krugman says, “you can, if you like, try to argue that this relationship is spurious, maybe not causal.” Actually, I [Weakliem] liked his original figure, since I agree with Krugman on economic policy. But thinking about the possibility of spurious correlation isn’t a matter of liking—it should be pretty much automatic.

And the full story is here.

That last bit (“it isn’t a matter of liking—it should be pretty much automatic”) holds for a lot of statistical issues, not just spurious correlation. I think that Weakliem’s illustrating a general issue, when someone does a statistical analysis and goes into “story time“—they’ll sometimes bring up a potential objection (in this case, spurious correlation) but not take it seriously. Better to bring it up than to not mention it at all. But better still would be to more fully recognize the limitations of one’s analysis. So good catch by Weakliem on this one.


  1. Jeff Walker says:

    Is spurious continuous or discrete (a threshold function of some magnitude of bias)? We tend to think of it as discrete “this is spurious, not causal”, which probably leads to this sloppy talk. But certainly it’s the case that any observational correlation is spurious (in the sense that there must be some bias) we just don’t know how spurious it is.

  2. Jonathan says:

    I’ll mount a mild defense of Krugman here: In the broader context, Krugman was countering a movement that made a direct casual claim that reduction in government spending would cause increase in gdp growth. It’s very hard to explain such a strong negative correlation under a casual model that predicted a strong positive correlation.

    So in that context, the correlation, spurious or not, is a compelling argument against using the competing casual model to inform policy.

    Rhetorically, “if you like” seems perfectly justified in this context.

    Of course, moving from debunking Austerian theories to promoting Keynesian ones requires a more robust approach.

  3. Elin says:

    I’m not at all surprised to see a sociologist make that point. I think most of quantitative sociology, certainly most that is admired in the discipline, is about looking at a correlation and then explaining why it is spurious or at least not as strong as it seems. The post about regression got me thinking about how differently we use regression than other disciplines, for example almost never is it about prediction. It’s usually about making a model more complex by adding variables representing competing hypotheses, thinking about timing, networks, correct specification of the response, and so on based on the idea that observed relationships are possibly completely wrong.

  4. Rahul says:

    Your correlation is spurious. Mine is causal.

    Until we have a good, practical framework to decide causality (apologies to Pearl), isn’t most of this moot?

  5. Mark Palko says:

    A bit more context (thank you, Google)

    “As I said, you can, if you like, try to argue that this relationship is spurious, maybe not causal. But one form of argument that is really illegitimate is to comb through the data, pick out outliers, and claiming that the existence of these outliers — because stuff does, in fact, happen — disproves Keynesian logic. Unfortunately, you see a lot of that, including from economists who really should know better.”

    Also keep in mind, this comes in mid-discussion. We are talking about predictions that were made publicly before the treatments/policies were implemented, not a true natural experiment but closer than we normally get. This would normally be the foundation of a strong causal argument. Krugman is saying that, even after a strong causal argument, claiming relationships are spurious is acceptable, but argument by outlier is not.

Leave a Reply