Causation relies on the philosophical concept of counterfactuals, and
you can read a fair amount about this on the Internet. The true effect
of a treatment would be measured if we had the magical power to
simultaneously assign a patient to both the treatment and control arm.
The effect for that patient would then be his/her response on the
treatment minus his/her response on the control. Note that this is not
the same as a crossover trial because we need to assign both treatments
simultaneously. Visualize two parallel universes. In the first universe,
a patient is assigned to the treatment and in the second universe, a
patient is assigned at the exact same time to the control, and
everything else in the two universes except the assignment are
identical. If the patient has a blood cholesterol of 180 in the first
universe and 190 in the second universe, you can safely state that the
treatment caused a 10 point drop in that patient. Repeat this experiment
across multiple patients to get the average cholesterol drop caused by
the treatment.
Researchers typically cannot measure things in parallel universes, so
they have to rely on something else.
Suppose we have a randomized trial with the outcomes on the patients
being Yij where i=1 or 2 representing treatment or control and
j=1,...,2*n representing the results of the total of 2*n patients. For
the jth patient, either Y1j or Y2j is observed, but not both. The
missing value is the value observed in the parallel universe. If we had
the data then the estimated effect would be the average of all the
differences Y1j-Y2j.
The missing values in each pair, though, are missing completely at
random (MCAR). MCAR clearly applies here (In statistical analysis,
data-values in a data set are missing completely at random (MCAR) if the
events that lead to any particular data-item being missing are
independent both of observable variables and of unobservable parameters
of interest--Wikipedia).
In the MCAR case, you can safely substitute the average of values where
you did observe a response. This leads to YBAR1-YBAR2 being an unbiased
estimator of the average of Y1j-Y2j. So the effect seen in a randomized
trial is comparable to what you would have seen if you had the power to
assign someone simultaneously to both treatments.
In observational studies, of course, treatment assignment is likely to
be associated with unobservable parameters of interest, which makes
claims of causation much more difficult. There are some methods based on
the concept of Missing At Random (MAR) that might help here.
Steve Simon, [log in to unmask], Standard Disclaimer.
Sign up for the Monthly Mean, the newsletter that
dares to call itself average at www.pmean.com/news
On 1/28/2011 9:06 AM, Djulbegovic, Benjamin wrote:
> Dear all
> I'd like to post this question to the group that I have been thinking
> about for some time... Is there a scientific method that allows us to
> LOGICALLY distinguish the cause-effect from the coincidence? David Hume,
> one of the most influential philosophers of all times, concluded that
> there is no such a method. This was before RCTs were "invented". Many
> people have made cogent arguments that (a well done) RCT is the ONLY
> method that can allow us to draw the inferences about causation. Because
> this is not possible in the observational studies, RCTs are considered
> (all other things being equal) to provide more credible evidence than
> non-RCTs. However, some philosophers have challenged this supposedly
> unique feature of RCT- they claim that RCTs cannot (on theoretical and
> logical ground) establish the relationship between the cause and effect
> any better than non-RCTs. I would appreciate some thoughts from the group:
> 1. Can RCT distinguish between the cause and effect vs. coincidences?
> (under which -theoretical- conditions?)
> If the answer is "no", is there any other method that can help establish
> the cause and effect relationship?
> I believe the answer to this question is of profound relevance to EBM.
>
> Thanks
> Ben Djulbegovic
>>
>>
>>
|