11
$\begingroup$

I am new to statistics and causality. To my knowledge, to talk about causality, one must have some sort of intervention. I knew it as "no causation without manipulation".

Now I am curious: I see many information saying that causality can be checked just by using regression, ie this video https://www.youtube.com/watch?v=LDFYx90zF1k

So, I have two questions: Is it possible to use regression to check causality? Does this always hold "no causation without manipulation"? In other words, we can only talk about causation only after doing some sort of manipulation??

$\endgroup$
2
  • 3
    $\begingroup$ Not at all. A very large part of the causal inference literature is based on observational studies with no intervention. Of course, you can't easily prove causation, even with an intervention. Generally a causal hypothesis is "accepted" after a large body of evidence supports it in the absence of a satisfactory competing hypothesis (e.g. smoking and lung cancer) $\endgroup$ – Robert Long Feb 1 '20 at 16:26
  • 1
    $\begingroup$ See this question: stats.stackexchange.com/questions/2245/… $\endgroup$ – Noah Feb 1 '20 at 19:12
11
$\begingroup$

This is a vast topic and your question could be nominated for closure for being too broad, but I will make a few extended comments. It is not intended as full answer. People write whole books on this topic. For example, see the work of Judea Pearl, Sander Greenland and Miguel Hernán for starters.

Is it possible to use regression to check causality?

Certainly a regression model can be used to test a causal hypothesis, and this is done very frequently indeed. However, it is impossible, to prove causality with any statistical model. A researcher may hypothesize a particular causal path, and then use a regression model to test their hypothesis. The problem here is that the causal hypothesis, for example, "Does X cause Y ?", cannot be directly tested in the sense of getting an answer "yes" or "no". In frequentist statistics, for example, the test will usually provide an answer to the question: "if there is no association between X and Y, then the probability of obtaining these data, or data more extreme, is x%" which obviously is not the real question that the researcher has. This is a very common misunderstanding. On the other hand, if this probability is very low, then it is consistent with the researcher's causal hypothesis.

Directed Acyclic Graphs (DAGs), sometimes also known as causal diagrams are a very useful tool in causal inference, particularly in the context of regression analysis, and they can help to reduce a multitude of biases that may occur when trying to estimate a causal effect using regression. See this question and answers, for details:

How do DAGs help to reduce bias in causal inference?

Does this always hold "no causation without manipulation"? In other words, we can only talk about causation only after doing some sort of manipulation??

No. There is a large body of research on causal inference in observational data, where it is not possible to manipulate or perform an intervention.

It is true that a manipulation, or intervention, is strongly preferable to a purely observational study, however this is often not ethical. For example, we might be interested in the causal effect of some potentially harmful exposure on some health outcome, and it would be unethical to devise an experiment where some of the study group were exposed and others were not. Another reason why we can't conduct an experiment might be practicality. For example, we might be interested the effect of certain "trajectories" of childhood bmi on later life bmi. It would not be practical to devise an intervention to manipulate BMI in one arm of a study.

As mentioned above this is a vast area. Check this question and answers for further details about manipulations:

Statistics and causal inference?

$\endgroup$
4
  • $\begingroup$ Don't you think that the presence of a "large body of research on causal inference in observational data" falls short of establishing that you can actually infer causation from observational data? After all, that body of research may be of poor quality, or maybe its excellent quality but just doing the best it can in the face of an impossible task. $\endgroup$ – Hasse1987 Feb 4 '20 at 22:14
  • 3
    $\begingroup$ @Hasse1987 Maybe you misunderstood my point. The next sentence in the comment you quoted says "Of course, you can't easily prove causation, even with an intervention". In observational studies in particular, causality is never actually proved. Instead, a weight of evidence builds up over time from which it is hard to escape a causal link. Perhaps the best example is that smoking causes cancer. A more controversial one would be anthropogenic climate change. $\endgroup$ – Robert Long Feb 5 '20 at 11:47
  • $\begingroup$ ? First, the next sentence begins "It is true that a manipulation..." Second, your answer to the OP's question is a bare "No." with the only justification offered being what I just quoted about the size of the field (repeated again in the final para). The only other paragraph explains why observational studies are the only option, whether for ethical or other grounds, but obviously there being no nice alternative doesn't say anything about whether causal inferences may be drawn from what is available. So what is exactly is the point you think I misunderstood? $\endgroup$ – Hasse1987 Feb 5 '20 at 23:50
  • 2
    $\begingroup$ @Hasse1987 my point is that RCTs provide better evidence than observational studies, but that doesn't mean causation can't be inferred from the latter. That's why I started by answer with No, and that's what I thought you might have misunderstood. Feel free to post your own answer if you disagree, but comments to questions or answers are not the place for extended discussions. $\endgroup$ – Robert Long Feb 6 '20 at 7:38

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.