Dear All,
Paul is correct that (effect) size matters. I would like to add that there
is a further benefit of experimental studies such as RCTs where the
investigators allocate the intervention:
EXPERIMENTAL STUDIES RULE OUT COMMON CAUSES BUT OBSERVATIONAL STUDIES DO
NOT.
An observational study might find a strong (large effect size) correlation
between ashtrays and lung cancer, but ashtrays do not cause lung cancer.
Smoking (the common cause) causes people to buy ashtrays and also causes
lung cancer. If we randomised people to have ashtrays or not, we would
presumably discover that ashtrays did not cause lung cancer ...
These are issues that philosophers have thought about and made some
headway. I hope to have shown that some philosophers can also explain them
in a relatively straightforward manner.
Best wishes,
Jeremy
On 30/01/2011 23:51, "Paul Glasziou" <[log in to unmask]> wrote:
>Good idea Tom,
>Ben - one simple option is the ISEHC Newsletter, from which some articles
>are also picked up for the EBM Journal (and I'm pretty sure this would
>be).
>Even an annotated reference list would be useful. To add my two cents:
>Randomisation comes closer to eliminating confounding than other methods,
>and so can be trusted to correctly identify smaller effects.
>But for very big effects (smoking and lung cancer) observatioinal studies
>can be sufficient (but see Bradford-Hill guidelines).
>The crucial issue is whether the effect size is greater than the
>plausible biases
>(see Glasziou et al . When are randomised trials unnecessary? Picking
>signal from noise. BMJ. 2007 Feb 17;334(7589):349-51. and
>Howick J et a The evolution of evidence hierarchies: what can Bradford
>Hill's 'guidelines for causation' contribute? J R Soc Med. 2009
>May;102(5):186-94)
>Cheers
>Paul Glasziou
>________________________________________
>From: Evidence based health (EBH) [[log in to unmask]]
>On Behalf Of Tom Jefferson [[log in to unmask]]
>Sent: 30 January 2011 17:23
>To: [log in to unmask]
>Subject: Re: Can RCT help establish causation?
>
>Ben and all. I think it is important that your summing up and the many
>excellent contributions to this debate get written up. As you pointed
>out, there is a movement underway to try to subvert the value of the RCT.
>The RCT, if properly designed and conducted, represents the only
>near-perfect tool to test causation at our disposal as it eliminates the
>play of alternative explanations of the results observed. It is as near
>as we are going to get (at least in our lifetime) to a Galilean
>experiment.
>
>The validity of the inferences from its results are still limited by the
>problem of induction (Hume's problem), but Fisher's probabilistic "patch"
>makes their limits at least explicit and quantifiable.
>
>Thank you for a delightful debate.
>
>Tom.
>
>On 30 January 2011 17:29, Djulbegovic, Benjamin
><[log in to unmask]<mailto:[log in to unmask]>> wrote:
>Dear all
>First, I want to thank many of you who responded either directly to me or
>to the entire group with the phenomenal line of thoughtful insights
>related to the question if RCT can establish causation. Once again,
>participation in this group has proved so rewarding there is no
>question, big or small, theoretical or practical that goes unanswered at
>rather deep level.
>
>The immediate reason for my question was motivated by increasing number
>of writings in the philosophical literature that question feasibility of
>developing hierarchy of evidence- one of the most important contribution
>of EBM, I should add. While there are several lines of this critique,
>the most fundamental- and thus most important- is the claim that, in
>principle, the inferences drawn from RCTs are not epistemologically
>superior to observational studies. In particular, the criticism was
>leveled against RCT as the purported mechanism to establish causation (as
>randomization allows equal distribution of all known and UKNOWN
>confounders apart from the intervention of interest). The critics (in
>particular Worrall) has asserted that ²given that there are indefinitely
>many possible confounding factors...the probability that the groups are
>unbalanced ...is high" and hence inferences from RCTs are not
>epistemologically different from those drawn from the observational
>research.
>
>Some of you graciously sent me several papers refuting some of Worralıs
>critique (I am still reading some of the material you sent- so, some what
>I am writing here may still change), but most poignant repudiation of
>Worralıs criticism came from Stephen Senn. His paper (Senn, SJ. Baseline
>Balance and Valid Statistical Analyses: Common Misunderstandings, Applied
>Clinical Trials 2005; 14: 24-27) should be required reading for all
>EBMers (although after reading one will probably say, this is so
>obvious!, it certainly was an eye opener for me, not exactly someone who
>is in this business since yesterday!). As Stephen pointed out (proved),
>³there is no point worrying about the distribution of unmeasured
>covariates² as we can still draw the valid inferences based on
>information we have (what we do NOT know CANNOT be factored in the way we
>draw our conclusions or make our inferences). The same argument cannot
>apply to observational studies as we donıt have formal machinery
>(randomization) to draw the accurate inferences about the probability
>statements.
>Finally, several folks cogently argued that randomization is indeed the
>necessary (although not sufficient reason, as it may be subverted etc) to
>establish causation.
>This all indicate that indeed experimental, RCT design remains
>epistemologically superior to observational studies (no big news for the
>members of this group, but fundamentally really big news).
>Thanks again to all of you for the phenomenal discussion and contribution
>Ben djulbegovic
>
>
>
>--
>Dr Tom Jefferson
>Scientific Editor PLoS ONE
>Reviewer, Cochrane Acute Respiratory Infections Group
>tel 0039 3292025051
|