10
$\begingroup$

Lets say I have a table with columns "A", "B"

Is there a statistical method to determine if "A" causes "B" to happen? One can't really use Pearson's r, because:

  • it only tests the correlation between values
  • correlation is not causation
  • Pearson's r can only correlate linear relationships

So what other options do I have here?

$\endgroup$
4
  • 1
    $\begingroup$ There is not. From such data you can demonstrate a high degree of correlation; you cannot demonstrate causation. $\endgroup$
    – Brian M. Scott
    Jun 26 '12 at 9:40
  • $\begingroup$ check it out: cms.ieis.tue.nl/Beta/Files/WorkingPapers/Beta_wp166.pdf $\endgroup$
    – Avatar
    Jun 26 '12 at 10:39
  • 1
    $\begingroup$ Causation is just not something you can squeeze from the numbers... so, repeat after me: causation is not correlation, causation is not correlation... $\endgroup$ Aug 5 '12 at 12:12
  • 1
    $\begingroup$ See "Causality" by Judea Pearl (2011 Turing Award winner). $\endgroup$
    – user52423
    Jul 20 '14 at 19:34
4
$\begingroup$

The answers and comments so far are basically correct at the practical level, but for completeness, there is research into so-called causality models that are based upon Bayesian statistics and graph theory. So although in general correlation indeed does not imply causation, there are more complex models that do attempt to tease out causation. See the book Causality by Judea Pearl for more details, but this is very heavy-duty mathematics and is probably not what you want.

$\endgroup$
2
$\begingroup$

There are many so-called quasi-experimental methods with which you can credibly argue about causality, even though your data are observational. These methods typically rely on finding a source of exogenous variation in your variable of interest.

I think a good and accessable overview is given in the book "Mostly Harmless Econometrics". They cover basically all quasi-experimental methods that people (meaning: economists) believe in (at least sometimes). They do not cover the methods mentioned by for instance trb456 (for the same reason: not many believe in them).

$\endgroup$
1
$\begingroup$

To determine causation you need to perform a randomization test. You take your test subjects, and randomly choose half of them to have quality A and half to not have it. You then see if there is a statistically significant difference in quality B between the two groups.

It is important that you do the randomization before you do any measurement. In particular, if you are given a data set with $A$ and $B$ already measured, then it is impossible to determine causation.

Note that it may be impossible to do the randomization test that you want to do. For example, how could you test if being tall causes you to weigh more? Certainly there is a correlation between height and weight, but you can't randomly assign one group of people to a 'tall' group and one to a 'short' group. In this case, the randomization test can't be done.

$\endgroup$
0
0
$\begingroup$

Somers' d works for explaining the relationship between ordinal variables in a way that pearson's correlation coefficient does for data sets.

$\endgroup$
2
  • 1
    $\begingroup$ I agree that it takes more than numbers to establish causation. How does the use of ordinal variables enter into the question? $\endgroup$ Dec 17 '17 at 22:57
  • 1
    $\begingroup$ @MichaelChernick Somers' D is an asymmetric measure of association. It can distinguish between "if it is raining, then it is cloudy,' from "if it is cloudy, then it is raining." It works for ordinal or higher data. It does not establish causation, but it does establish directionality. $\endgroup$ Dec 18 '17 at 0:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.