« It's time to start worrying about productivity again | Main | "Canada's managers are under-achievers" »


Feed You can follow this conversation by subscribing to the comment feed for this post.

Looks like you had an example floating round in your mind when you wrote this. Penny for that thought?

Heh. Yes, yes I did. Here and here.

You are not a very careful reader. I was using the "correlation is not causation" argument against a view of the cause of health care inflation to which I am sympathetic, but which I do not think is proven by the correlation that my reader found.

Seems to me Mr Kling was making more than just a simple statement about his sympathies with respect to health care. Quote:

"That is why most macro-econometrics is junk science. That is one reason I would tend to suspect that Larry Bartels' work on Presidential party and income inequality is junk science."

He's used the correlation-causation distinction to malign a field or research and a particular research topic he apparently does not like or that argue for opinions he finds at odds with his own ideological leanings.


The fact that Kling is right doesn't seem to faze you. Nor does the fact the opinions and ideological leanings that he would tend to favor, such as the health care inflation argument, he also "maligns" because the "evidence" is based on spurious correlation.

Actually, the findings of Mr Kling's reader isn't suprising at all. It is a result found in many health economics paper that look into determinants of health care spending. Some due in fact use macro time series, but it isn't like the problem of spurious regressions is new or unknown to researchers. The junk science part of M. Kling's comment is pretty pompous if you ask me.


If you came to me and said, "I have a hypothesis that A causes B. I have plotted the two series on a graph and have discovered that the correlation coefficient between the two series is very high. Therefore, because the correlation coefficient is so high, this must mean that A causes B."

Is the proper response to say, "Well, there you have it. Based on the very high correlation coefficient between the two variables, you have proved A causes B." Or should one say, "Correlation does not mean causation. You have not proved your hypothesis BASED ON THE EVIDENCE PROVIDED."

The burden of proof is on the person running the hypothesis. No legitimate study would claim causation based on a high correlation coefficient. Yet, Kling is suffering the slings and arrows regarding the health care example for just that.

If Kling was presented the results of a multiple regression model that included other variables and one that didn't have autocorrelation and the researcher had first differenced the economic data, then Kling might change his mind if the evidence pointed that way. But to condemn Kling for point out the obvious fact that correlation does not mean causation is outrageous.

True: Correlation does not mean causation. But the only form of causation we know, as Hume rightly points out, is correlation, one thing following another. All causation--from an epistemological point of view--is nothing more than observed correlation. Then it's just a matter of analayzing the matrix of correlations in order to determine which of them we would like to describe as hypothetically causal at the ontological level.

In some cases we can perform laboratory experiments. In other cases, we have to rely on historical data.

My argument is with his junk science comment. It seemed like a cheap way to take his ideological axe against (a) macro-econometrics and (b) presidential party and income inequality.

Yet Kling makes statements such as the following: "But the reality is that the intelligences that feed into IQ are what drive economic success."

The only way he can make such a statement is through correlations. Yet he says they "drive" economic success. I.e. "cause."

All so-called fundamental physical laws are actually (more or less dense) scatter plots. Statistical estimates, including correlation, provide some values for subjective judgement only. So, "complete" causality in scientific sense is an illusion in real world, except banal cases.

Formally, macroeconomic variables (being measured values) have the same right as physical ones to be treated seriously and modeled in the same way. Statistics for economics should not be much different from that for super-collider experiments.


I wasn't entirely clear. I am not debating the point that his reader's simple regression was actually correct or demonstrate causality. My "issue" is more with the "most macro-econometrics is junk science" quote. I do not know your background, but pretty much all macro papers that I have read that are dated after 1990-1992 that use macroeconometrics take into account the paper of Phillips in 1986. My interests aren't even macroeconomics, I just think that flatly saying that most of macro-econometrics is junk science is really pompous.
Although, I have no presumption of expertise and compared to M. Kling, my knowledge of econometrics is most likely inadequate to even debate him on that subject.

Okay, I'm no economist, but I'll soon be an MMath in Biostatistics. First, the statement that correlation does not imply causation is not true in general - correlation may indeed imply causation, for the simple reason that if A causes B (and so "not A" implies "not B"), then the presence of A should certainly be correlated with the presence of B (setting aside whether they occur simultaneously or not - if they do, causation would be rather difficult to establish!). On the other hand, if A is found to be significantly correlated with B, the observed association may reflect an underlying causal relationship. Correlation does not imply that B follows A, but merely that there is an association between the two variables - this association can't be arbitrarily dismissed as "spurious", though, particularly given an observed value of 0.92. It is, of course, just a descriptive statistic, but I'd say it merits further study.

Concerning this:
If Kling was presented the results of a multiple regression model that included other variables and one that didn't have autocorrelation and the researcher had first differenced the economic data, then Kling might change his mind if the evidence pointed that way. But to condemn Kling for point out the obvious fact that correlation does not mean causation is outrageous.

Why difference it immediately? Kling's comments are simply irrelevant to the issue at hand - if you observe a high correlation between two series, you have just observated a clear (and probably significant) association between said series. Causation might be inferred if there is a clear lag between the "cause" in one series and the "effect" in another, but the association stands all the same. The appropriate response to such correspondence is to investigate further - pithy pseudo-methadological comments aren't altogether helpful.

Can I try this one out? Just stumbled upon this (interesting) bit from the New York Times link to Canadian topics.

As I recall, before one can test relationships between various variables, FIRST one has to identify the dependent and independent variables. Correlations can be VERY misleading. Now, I am jumping into this thread from out of nowhere, so please forgive me if anyone think me some sort of troll (I'll keep this web page open and refresh it time to time over the next few days).

Here is one almost cliche example that demonstrates why the thinking is faulty: It is historically evident that as the number of churches in a metropolitan area grows, so does crime. In fact, the correlation could be convincingly high in some cases. The trouble with this, of course, is that both of these issues are probably more dependent on population growth, not each other (for the most part).

Hypotheses have to be educated from the start; applying pure mathematics and statistical analysis BLINDLY to any study can easily mislead us. Determining which variables are dependent or independent is, naturally, a skill honed from experience and wisdom. To have one hard and fast rule that a seeming "correlation" is automatically evidence of a trend or pattern or causation, etc., could lead to many disasters.

I think I see examples of this stateside. Some believe that US-led bombing of Iraq has prevented another 9/11 scale attack on the US. And if we plotted appropriate points on a graph, it probably does look so. But is bombing Iraq really the independent factor that has prevented another disaster? Maybe our fighting in Afghanistan could be credited for that, especially since that is where Al Quaida likes to hang out (along with Pakistan, which we are now invading with that government's blessings). Also consider the fact that along with Iraqi men who may be fighting us, we are indiscriminately bombing women and children in Iraq as well. Shall I believe that the correlation proffered by some means that the US has been kept safe from terror because we have slaughtered Iraqi women and children? Again, we failed to identify the independent variable. And, again, that is not something that can be derived purely mathematically. I think tax cuts for the wealthy fall into this category somewhere, too.

I wasn't around here for the discussion on ... "presidential party" and income inequality? I am not sure if I am familiar with the former, and I don't see a link back to the original issue. But sounds interesting, and I may follow up on this.

Awww ... go ahead. Beat me up. I'm just a yankee anyway.

You can certainly test for associations among variables without labelling some as "independent" and others as "dependent" - but such correlational studies would focus purely on associations. In general, though, something like linear regression is misused when it's assumed that the response is "independent" and the covariates are "dependent" - all regression does is use a correlation structure to study the relation between a particular response and a set of explanatory variables. Any causal relationships inferred require either an experimental study or else a very careful and robust examination of possible "causes" and "effects". Some sort of mechanism wouldn't hurt either; in fact, one can examine potential causal relationships by constructing and testing different mechanisms.

The comments to this entry are closed.

Search this site

  • Google

Blog powered by Typepad