Does causation imply correlation?

Correlation does not imply causation, as there could be many explanations for the correlation. But does causation imply correlation? Intuitively, I would think that the presence of causation means there is necessarily some correlation. But my intuition has not always served me well in statistics. Does causation imply correlation?

• Problem is, if you look up "imply" in a dictionary you'll see both "suggest" and "necessitate." – rolando2 Feb 1 '13 at 23:54
• Correlation doesn't imply causation, but it does waggle its eyebrows suggestively and gesture furtively while mouthing 'look over there'. xkcd.com/552 – jchristie Feb 28 '14 at 6:52
• The question itself doesn't appear to be looking for a specific, factual answer, as indicated by the use of the word imply. The reference above is like an ultimate maybe. Or more like a probably but I can't prove it. – jchristie Feb 28 '14 at 7:21

As many of the answers above have stated, causation does not imply linear correlation. Since a lot of the correlation concepts come from fields that rely heavily on linear statistics, usually correlation is seen as equal to linear correlation. The wikipedia article is an alright source for this, I really like this image: Look at some of the figures in the bottom row, for instance the parabola-ish shape in the 4th example. This is kind of what happens in @StasK answer (with a little bit of noise added). Y can be fully caused by X but if the numeric relationship is not linear and symmetric, you will still have a correlation of 0.

The word you are looking for is mutual information: this is sort of the general non-linear version of correlation. In that case, your statement would be true: causation implies high mutual information.

• It's usually but not always true that high mutual information accompanies causation. See @gung's answer where "if the cause is perfectly correlated with another causal variable with exactly the opposite effect." – Neil G Apr 12 '12 at 21:02
• The argument of two causes with opposite effects that always cancels each other doesn't make much sense to me as a cause. I can always assume there are unicorns causing something, and gremlins cancelling their efforts perfectly; I avoid this since it's silly. But maybe I am misunderstanding your point. – Artem Kaznatcheev Apr 12 '12 at 22:41
• His example is more extreme than it needs to be. It's possible for you to have Boolean variables $A, B$, and $C$ such that $A$ and $B$ are causes of $C$, and $C = A + B$ (mod 2). Then, absent of knowledge of $B$, $A$ and $C$ have no mutual information. $B$ is an undiscovered confounder — what you are calling "gremlins" even though it is something very common. – Neil G Apr 12 '12 at 23:15
• @NielG I agree with your first sentence, but not the second. Just because A & B causes C, doesn't mean that A causes C and B causes C. I don't see why cause has to be distributive over &. – Artem Kaznatcheev Apr 13 '12 at 21:45
• The reason that A is nevertheless a cause of C is because changing A will still change C. So, C is dependent on A even when we don't observe B. – Neil G Apr 13 '12 at 21:51

The strict answer is "no, causation does not necessarily imply correlation".

Consider $X\sim N(0,1)$ and $Y=X^2\sim\chi^2_1$. Causation does not get any stronger: $X$ determines $Y$. Yet, correlation between $X$ and $Y$ is 0. Proof: The (joint) moments of these variables are: $E[X]=0$; $E[Y]=E[X^2]=1$; $${\rm Cov}[X,Y]=E[ (X-0)(Y-1) ] = E[XY]-E[X]1 = E[X^3]-E[X]=0$$ using the property of the standard normal distribution that its odd moments are all equal to zero (can be easily derived from its moment-generating-function, say). Hence, the correlation is equal to zero.

To address some of the comments: the only reason this argument works is because the distribution of $X$ is centered at zero, and is symmetric around 0. In fact, any other distribution with these properties that would have sufficient number of moments would have worked in place of $N(0,1)$, e.g., uniform on $(-10,10)$ or Laplace $\sim \exp(-|x|)$. An oversimplified argument is that for every positive value of $X$, there is an equally likely negative value of $X$ of the same magnitude, so when you square the $X$, you can't say that greater values of $X$ are associated with greater or smaller values of $Y$. However, if you take say $X\sim N(3,1)$, then $E[X]=3$, $E[Y]=E[X^2]=10$, $E[X^3]=36$, and ${\rm Cov}[X,Y]=E[XY]-E[X]E[Y]=36-30=6\neq0$. This makes perfect sense: for each value of $X$ below zero, there is a far more likely value of $-X$ which is above zero, so larger values of $X$ are associated with larger values of $Y$. (The latter has a non-central $\chi^2$ distribution; you can pull the variance from the Wikipedia page and compute the correlation if you are interested.)

• @DQdlM: The standard random variable has vanishing odd central moments, due to the evenness of the density. Matthew: The answer is no, as StasK has demonstrated, because correlation is not the only type of dependence. – Emre Apr 12 '12 at 1:15
• @DQdlM: see the bottom middle graph in the first image on the Wikipedia correlation page. That's StasK's case. It only works when x is equally distributed about the origin (ie, if $X\sim N(3,1)$, correlation will be fairly high) – naught101 Apr 12 '12 at 1:17
• P.S. I'm so glad you posted this answer. It was hard to believe the question went so long without this answer. This was the exact example that came to my mind when I saw this question, but didn't get the time to write it up. I'm glad you did take the time. Cheers. – cardinal Apr 12 '12 at 20:40
• @cardinal: yeah, I guess we all learned these kinds of simple counterexamples in grad school... and yes, from the derivation of the covariance, you only need the first and the third moments to be zero. If you have a non-trivial example of an asymmetric distribution that has a zero third moment (finely tuned probability masses over five or six points do not count), I'd be very curious to see it though. – StasK Apr 12 '12 at 20:42
• Here 'causality' is being assumed to be expressable as a function. This is, $X$ causes $Y$ if and only if there exists a measurable function, $f$, such that $Y=f(X)$. I guess we could spend the rest of our lives discussing about the validity of this argument. – user10525 Apr 13 '12 at 10:51

Essentially, yes.

Correlation does not imply causation because there could be other explanations for a correlation beyond cause. But in order for A to be a cause of B they must be associated in some way. Meaning there is a correlation between them - though that correlation does not necessarily need to be linear.

As some of the commenters have suggested, it's likely more appropriate to use a term like 'dependence' or 'association' rather than correlation. Though as I've mentioned in the comments, I've seen "correlation does not mean causation" in response to analysis far beyond simple linear correlation, and so for the purposes of the saying, I've essentially extended "correlation" to any association between A and B.

• I tend to reserve the word correlation for linear correlation, and use dependence for nonlinear relations that may or may not have linear correlation. – Memming Apr 11 '12 at 20:34
• @Memming I would too, save for the fact that people trot out "Correlation does not imply causation" re: fairly complex non-linear association. – Fomite Apr 11 '12 at 20:36
• Memming is right. You need to define correlation if you don't mean Pearson correlation. – Neil G Apr 11 '12 at 20:52
• @NeilG Or for that matter, one may be able to get a linear Pearson correlation by transforming one variable or the other. The problem is the adage itself is over-simplified. – Fomite Apr 11 '12 at 20:59
• @EpiGrad: Both good points. In common parlance, correlation is just more of A coincides with more B. I think your answer would benefit from making your use of a broad definition of correlation clear. – Neil G Apr 11 '12 at 21:01

Adding to @EpiGrad 's answer. I think, for a lot of people, "correlation" will imply "linear correlation". And the concept of nonlinear correlation might not be intuitive.

So, I would say "no they don't have to be correlated but they do have to be related". We are agreeing on the substance, but disagreeing on the best way to get the substance across.

One example of such a causation (at least people think it's causal) is that between the likelihood of answering your phone and income. It is known that people at both ends of the income spectrum are less likely to answer their phones than people in the middle. It is thought that the causal pattern is different for the poor (e.g. avoid bill collectors) and rich (e.g. avoid people asking for donations).

Things are definitely nuanced here. Causation does not imply correlation nor even statistical dependence, at least not in the simple way we usually think about them, or in the way some answers are suggesting (just transforming $X$ or $Y$ etc).

Consider the following causal model:

$$X \rightarrow Y \leftarrow U$$

That is, both $X$ and $U$ cause $Y$.

Now let:

$$X \sim bernoulli(0.5)\\ U \sim bernoulli(0.5) \\ Y = 1- X - U + 2XU$$

Suppose you don't observe $U$. Notice that $P(Y|X) = P(Y)$. That is, even though $X$ causes $Y$ (in the non-parametric structural equation sense) you don't see any dependence! You can do any non-linear transformation you want and that won't reveal any dependence because there isn't any marginal dependency of $Y$ and $X$ here.

The trick is that even though $X$ and $U$ cause $Y$, marginally their average causal effect is zero. You only see the (exact) dependence when conditioning both on $X$ and $U$ together (that also shows that $X\perp Y$ and $U\perp Y$ does not imply $\{X, U\} \perp Y$). So, yes, one could argue that, even though $X$ causes $Y$, the marginal causal effect of $X$ on $Y$ is zero, so that's why we don't see dependence of $X$ and $Y$. But this just illustrates how nuanced the problem is, because $X$ does cause $Y$, not just in the way you naively would think (it interacts with $U$).

So in short I would say that: (i) causality suggests dependence; but, (ii) the dependence is functional/structural dependence and it may or may not translate in the specific statistical dependence you are thinking of.

• Carlos, it's correct to says that if we know the full set of variables involved in the causal model this problem (statistical invisibility) disappear? – markowitz Dec 10 '18 at 0:07
• @markowitz you would need to observe everything to the deterministic level, thus not a very realistic scenario. – Carlos Cinelli Dec 11 '18 at 1:19
• I interpret your answer as “yes”. You are right, the situation that I supposed is unrealistic; I’m aware about it. However the question was related only about the logic that you described and the finality was to grasp it. My conviction was something like “causation imply statistical association” and others answers in this page sound like this. After all also your example is slight unrealistic but not for this reason in uninteresting. It seems me that, also in general, causation without statistical association is slight unrealistic but theoretically interesting. – markowitz Dec 14 '18 at 13:38
• @markowitz the "statistical invisibility" happens when the model is not faithful to the graph. For exact cancelation, this depends on a specific choice of parameterization, so some people argue it is indeed unlikely. However, near cancelation might be plausible since it depends on a neighborhood of parameters, so it all depends on context. The point here is just that you need to make your causal assumptions explicit because, logically, causation does not imply association by itself -- you need extra assumptions. – Carlos Cinelli Dec 21 '18 at 18:20

The cause and the effect will be correlated unless there is no variation at all in the incidence and magnitude of the cause and no variation at all in its causal force. The only other possibility would be if the cause is perfectly correlated with another causal variable with exactly the opposite effect. Basically, these are thought-experiment conditions. In the real world, causation will imply dependence in some form (although it might not be linear correlation).

• @NeilG, I indulged my addiction to italics. – gung Apr 11 '12 at 21:35
• Some theories actually imply this, e.g. many game theory models. Some empirical situations where you can't discern a difference (although there would actually be one 'in gung-italics' as it were :-) include 'neutral' no gene change scenarios when evolutionary selection pressure at two levels point in different directions. – conjugateprior Apr 11 '12 at 23:38
• I like the first exception, but not the second exception. I like to think that flipping the switch causes the light to go on, but if I happen to only flip the switch during a blackout nothing happens. Perhaps there was not really a causal relation. – emory Apr 12 '12 at 1:17
• @naught101, you raise a good point, which has been discussed elsewhere on this page. I have edited my answer. However, when I've worked with people, I don't think they have a strong conception of correlation as necessarily linear, even though I tell them that. Although they wouldn't put it in these terms, I think most people understand 'correlation' as closer to 'function of'. Nonetheless, I should be clearer in my use of terms, and should've been from the start. – gung Apr 12 '12 at 3:40
• @emory: the cause of the light coming on is actually the closure of the electrical circuit (which is caused by the flicking of the switch, with the environmental conditions including a functioning grid). During a blackout, flicking the switch doesn't close the circuit, because it's broken elsewhere. So in a sense, the blackout is the "opposite" effect that gung was talking about (ie. light is on, blackout turns it off). It could also be thought of as a nullifying effect. – naught101 Apr 12 '12 at 3:50

There are great answers here. Artem Kaznatcheev, Fomite and Peter Flom point out that causation would usually imply dependence rather than linear correlation. Carlos Cinelli gives an example where there's no dependence, because of how the generating function is set up.

I want to add a point about how this dependence can disappear in practice, in the kinds of datasets that you might well work with. Situations like Carlos's example are not limited to mere "thought-experiment conditions".

Dependences vanish in self-regulating processes. Homeostasis, for example, ensures that your internal body temperature remains independent of the room temperature. External heat influences your body temperature directly, but it also influences the body's cooling systems (e.g. sweating) which keep the body temperature stable. If we sample temperature in extremely fast intervals and using extremely precise measurements, we have a chance of observing the causal dependences, but at normal sampling rates, body temperature and external temperature appear independent.

Self-regulating processes are common in biological systems; they are produced by evolution. Mammals that fail to regulate their body temperature are removed by natural selection. Researchers who work with biological data should be aware that causal dependences may vanish in their datasets.

Wouldn't a cause without any correlation be an rng?

Unless, like the accepted answer implies, you're using an incredibly limited interpretation of the word 'correlation', it's a silly question- if one thing 'causes' another, it is by definition affected by it in some way, whether it's an increase in population, or just intensity.

right?

Then again, you could be discussing something more like, the visibility of something being affected by something else, which I guess would look like causation, but really you aren't measuring what you think you're measuring...

So yeah, I guess the short answer would be, "Yes, so long as you can't create entropy."

protected by Community♦Feb 28 '14 at 15:58

Thank you for your interest in this question. Because it has attracted low-quality or spam answers that had to be removed, posting an answer now requires 10 reputation on this site (the association bonus does not count).

Would you like to answer one of these unanswered questions instead?