I major in bioinformatics. We all know that temperature changes during a year. I find that a disease incidence is really high when temperature is relatively high, while it becomes really low when the temperature is relatively low; that is to say, they are related. I want to find out a way to prove that they are related, not just intuitively feel that they are related. Any suggestions?
$\begingroup$
$\endgroup$
1
-
$\begingroup$ Wrong site - I think you want math.stackexchange, tag statistics... $\endgroup$– rasherCommented Jan 16, 2014 at 1:53
Add a comment
|
3 Answers
$\begingroup$
$\endgroup$
Look into the concept of Granger causality which deals with precisely this. Basically you do a predictive model of one variable to the other. Then you run a hypothesis test to see how much of the variance in the output was explained by the input variable. If enough variance is sufficiently explained, then you can say that temperature 'Granger causes' disease.
Note g-causation is weaker then causation because both variables could be influenced by a third unobserved variable. However, this is in principle the best you can do with only two time series....
answered Mar 19, 2014 at 22:49
$\begingroup$
$\endgroup$
4
Collect your data and then model it. If you want to show a causal relationship, you should design an experiment, as their could be other confounding variables that mask the real cause and effect.
-
1$\begingroup$ The "treatment" in this case is the wheather. How would you manipulate the wheather for your experiment? $\endgroup$ Commented Jan 16, 2014 at 10:13
-
$\begingroup$ Obviously you can't, which means he'll never really be able to prove anything. His causal mechanism could also be sunlight, air quality, a new baseball season, etc. All he'll be able to do is show there's a relationship. (I'm not sure if the original post is about humans or not.) $\endgroup$ Commented Jan 16, 2014 at 10:51
-
$\begingroup$ Sitting back and saying "can't be done" is fine, but is not going to move us forward. There is a large research effort going on for a while now that tries to get (bounds on) causal effects from observational data, which you seem to be ignoring. $\endgroup$ Commented Jan 16, 2014 at 11:00
-
$\begingroup$ You could use something like El Nino as an instrument. Presumably El Nino doesn't affect disease except through its affect on weather. $\endgroup$ Commented Mar 19, 2014 at 23:20
$\begingroup$
$\endgroup$
I think what you are looking for is the methods of causal inference. A standard way to prove causality is through randomization. Some accessible articles on the subject are http://www-stat.wharton.upenn.edu/~rosenbap/ExperAndObsTalk.pdf and http://ftp.cs.ucla.edu/pub/stat_ser/r350.pdf
For biostats specifically, you already have a set up for this by having two groups (a treatment group and a control group) and randomly assigning people to each group.
As a generalization, it is extremely hard to prove that A causes B. I may be wrong, but I don't think that (at least legally) smoking has been proven to cause cancer, just that the probability of getting cancer is increased if you smoke.
