Re: Teaching physics to biology students

Gregory L. Hansen wrote:

In article <1142950293.575497.4400@xxxxxxxxxxxxxxxxxxxxxxxxxxxx>,
Edward Green <spamspamspam3@xxxxxxxxxxx> wrote:
Gregory L. Hansen wrote:


Another point, mentioned about medical research which is in the same
vein as routine data runs in physics: am I right in thinking that a
significant proportion of medical research is simply statistical
investigation? A carefully controlled statistical experiment may point
to the existence of an undisclosed mechanism. A loosely controlled one
may merely be suggestive, or may do more harm than good -- since the
"suggestion" will be taken as evidence by people prejudiced towards the
hypothesis.

In some sense, *every* experiment is a statistical investigation! Even
simple physical measurements must be reported with an error bar.

True. I was thinking, as Ken suggested, of experiments which might
test if there were any correlation at all.

Now I think we might want to distinguish between experiments ... ahem,
"studies" ... which merely sought causation, and those which sought
effect. What's the difference? I suggest the difference may be that
in a statistical study looking for an "effect", steps have been taking
to reasonably randomize all uncontrolled sources of variation across
the treated and untreated populations. Since we can never know all
sources of variation, this can only be done by randomized assignment to
treated and untreated sub-populations. This cannot be done given a
pre-existing (self-treated) population. Yet I think this latter
situation often occurs in published studies, taking as evidence that
such studies frequently see the light of day through third party
popular publication.

I was thinking along the lines of epidemiological research and drug
testing, and naturally you're testing the null hypothesis or whatever. I
assume that's what you mean by a statistical investigation.

Now that you've forced me to clarify myself, I would say a mere
controlled study to test whether there is an effect, as discussed
above, would qualify as the mininum experiment which _wasn't_ merely
.... hmm .... should find a better term than "statistical".

Now you've got me going, let's try a preliminary classification of
statistical studies:


I. Analysis of existing data.

Finds correlations. Suggests possible avenues for investigation of
effects. Error analysis: confidence interval for measured correlation.
Non-inclusion of zero rejects zero-correlation. Without further
Bayesian argument, zero information on direction of causation.

II. Controlled experiment to reject null hypothesis.

Finds effects. Suggests possible avenues for investigation of strength
of effects. Error analysis: confindence interval for measured
correlation. Non-inclusion of zero rejects zero-correlation. Form of
experiment induces Bayesian prior in favor of causation.

III. Controlled experiment to test strength of effects.

Finds refined estimates of functional form of effects. Error analysis:
error bars for functional parameters. Existence of effect no longer at
issue.


You are right, sir. Types I,II and III could all be considered "merely
statistical". But I had in mind type I, the weakest of the lot.

But not all
research is really like that. For instance, determining a protein
structure by crystallizing it and scattering neutrons through it.
Or determining the metabolites of caffeine.

Good counter-examples.

If you're measuring
something like a drug-drug interaction, I think you'd really want to be
able to produce something like a graph of the biological lifetime of drug
A versus doseage of drug B rather than resorting to "We've rejected the
null hypothesis." Physicians will want to know how to adjust the doses
they give to their patients, simply knowing that it must be adjusted is
not enough.

OK... in our newly minted jargon, I'd say that was an example of at
least type III statistical investigation. I'd also buy into the idea
that all experiments involves statistical elements. Our "type-III"
experiment trails off into or includes the type of experiment we
normally think of, where statistical error analysis is a secondary
tool: we know damn well there is an effect and approximately where its
value lies, we just want to quantify our remaining uncertainty.
However, perhaps many new results in particle physics start off near
the type-II,type-III threshold?

What I was suggesting, however, is that there are a significant
proportion of "type I" investigations out there in the medical field.
I did not define "significant", so I said almost nothing. ;-) However,
I'm going to go out on a limb here. I estimate that at least 50% of
published medical studies which are subsequently picked up by the
popular media for sound-byte status are type-I statistical
investigations.

That is, with our refined classification, what I meant to suggest by my
off-hand remark.

.

Relevant Pages

  • Re: Quick and earnest question about MMX
    ... In physics, one can never measure anything precisely, so we put ... modern experimental physics, specifically error analysis. ... variations of the data, so there is no SIGNIFICANT variation. ... a century which have shown agreement with relativity to many ...
    (sci.physics.relativity)
  • Re: Teaching physics to biology students
    ... mentioned about medical research which is in the same ... vein as routine data runs in physics: am I right in thinking that a ... *every* experiment is a statistical investigation! ... I was thinking along the lines of epidemiological research and drug ...
    (sci.physics)
  • Re: Quick and earnest question about MMX
    ... which result would then serve as the cornerstone of relativity. ... In physics, one can never measure anything precisely, so we put errorbars on measured values. ... So it is true that the MMX did not have what one could call a "null result". ... modern experimental physics, specifically error analysis. ...
    (sci.physics.relativity)
  • Re: The problem of Doppler shift and SR
    ... "Nicolaas Vroom" wrote in message ... and error analysis for such experiments takes into account any minimal other GR effects. ... You are making a comparison between SR and Newton's physics. ...
    (sci.physics.relativity)
  • Re: The problem of Doppler shift and SR
    ... than Newton's physics), and error analysis for such experiments takes into ... account any minimal other GR effects. ... Can you give me more details what type of LAB experiments ...
    (sci.physics.relativity)