I think my dog Juno got stoned at the park — again.
On at least two previous occasions and again just the other day, Juno started to stagger around, and for added effect fall down after chewing a considerable number of leaves off a plant at the park.
Clearly there is something in that plant that caused my dog to act like she was trying out to be the Grateful Dead’s mascot … or is there?
Consider the facts as best I know them:
- She has only acted this way on very hot days.
- She has eaten this plant before with no visible reaction.
- I don’t know how much of the plant she has eaten each time.
I feel pretty certain that ingestion of this plant has some correlation to her aberrant behavior, but I cannot say that it is the cause of it (and I am pretty certain there are no federal grants available for me to perform a controlled study… but then again you never know…).
Here is my problem — if I assume that it is the plant causing her to act this way I could very well be overlooking the real cause. Maybe she eats the plant because she is dehydrated and it is dehydration that is the problem. Or maybe she has some neurological defect. Or maybe my husband just forgot to feed her breakfast and she is famished.
And of course, with only a couple of observation as my “n” I can’t truly know much of anything anyway.
I really have no idea what is causing her to act this way.
Correlation v. causation is a big and serious challenge in healthcare data analysis as well, and as the requirements and pressure to report and act on nascent healthcare data continue to increase, we must understand correlation v. causation fully and consider it seriously.
When we want to know if a certain cause X produces a certain effect Y, we set up a study in which cause X is produced and its effects Y are observed.
But, just because we establish or observe an association or correlation, we have not established a cause-effect relation between the variables.
Correlation is not causation.
Imagine a group of physicians you may be working with. You have developed a “report card” for them. One physician has a much higher mortality rate than any other physician. It is all too easy a jump to “this is a bad doctor who is killing patients.” Perhaps that is true…but it is highly (as in very) unlikely.
Rather, it may be that their patient population is different than that of their peers (this is of course why risk-adjustment is so important) and it is highly likely that the sample size or “n” of one physician’s patients is not big enough to do any sort of real analysis.
Or perhaps, the systems in which the physician works are poorly organized and there are multiple causes for the high mortality rates (a root-cause analysis and further study is required).
Without high quality data, lots of observations and rigorous analysis and study we simply cannot act impulsively on what may appear to be on the surface a clear cause-and-effect relationship in our healthcare data. Because quite frankly, fixing the wrong thing is just about as bad (or worse) than fixing nothing at all.
For most cause-and-effect situations, especially those complicated by the involvement of human beings, a single effect can have many possible and actual causes. The elusive answer to the question “what causes cancer?” is yet another example of how other variables confound our ability to find a clear answer. Is it diet, lifestyle, heredity, environment, age — a combination?
The point is, there may be a number of reasons and causes for the outcomes we observe and simply finding a correlation in our data between two or more variables does not mean that we know with certainty why something is happening — exactly what caused it.
And so, as much as I want to believe that the nefarious plant at the park is causing my beloved dog to act stoned I really just don’t know. But if she keeps it up, I just may have to conduct a first person study to see what she likes about it so very, very much.