Forums

On 13/07/12 21:59, Nicolas Le Novère wrote:
> Frederic,
>
> Everything you write is absolutely correct. And we can actually encode the
> moments in distrib. But this is a level of sophistication that is light
> years away from what I am talking about. Unfortunately everyone on the list
> has probably googled "moment" now and read the relevant pages on Wikipedia.
> So we cannot run a proper poll. But I am fairly confident that less than 1%
> of SBML user or SBML support developers know what a "moment" is. 100% of
> them know what a mean and a standard deviation are.
I think it is also worth adding that the moment is the property of the
distribution. What people are actually going to want to annotate models
with is probably an *estimate* of the moment (or some other property of
the distribution - such an estimate is commonly known as a statistic)
based on experimental measurements of the parameter, not the moment
itself, since that is a property of the distribution and the true value
is unknown.

Coming up with an unbiased estimator (that is, an estimator sigmahat of
the unknown true value sigma of a parameter which has a true mean error
of zero, i.e. E(sigmahat - sigma) = 0) requires knowledge (or
assumption) of the distribution, or at least of some properties of the
distribution, in general.

See for example
http://en.wikipedia.org/wiki/Unbiased_estimation_of_standard_deviation
for background on how unbiased estimates of the true standard deviation
can be obtained under the assumption that the data is normally distributed.

Correctly obtaining just about any commonly used unbiased estimator
other than the mean requires knowledge of the distribution.

Actually interpreting the data based on that estimator also depends on
the distribution. Even basic assumptions (like, for example, that values
close to the mean are likely, and values a long way from the standard
deviation are unlikely) are distribution dependent.

Therefore, to come up with statistics, the modeller needs to have some
idea of the distribution, and the modelling software needs some idea of
that distribution to make use of the statistics provided by the modeller.

The modelling language is how the modeller and the tools processing the
model communicate, and so it needs to capture the modeller's thoughts
about the assumptions they made about distribution to obtain their
estimates.

Best wishes,
Andrew

>
> On 13/07/12 08:47, Frederic BOIS wrote:
>> To me too. If tools want to understand mean, sd etc. it's not much to
>> understand distrib.
>> Rather than "unknown distribution" I would prefer "moments" for those.
>> You can have known
>> moments with an unknown distribution, the reverse, etc. and those
>> things are quite separate.
>> It just happens that the two first moments are also the parameters of
>> the most common (Normal)
>> distribution, but that's not case for most of the others.
>>
>> Frederic
>>
>>
>>>>> "Chris J. Myers"<myers@ece.utah.edu> 12/07/2012 20:40>>>
>> My preference would likely be (2), in the distrib package. I don't
>> think the argument against it is significant over (3) as if you make a
>>
>> new package than you would need to understand that one. I think a
>> nice way to do this would be to introduce an "unknown" distribution
>> with four parameters in the same way we discussed introducing other
>> distributions. This could then take four parameters, those you list.
>>
>> Chris
>>
>> On Jul 12, 2012, at 9:32 AM, Nicolas Le Novère wrote:
>>
>>> Hello,
>>>
>>> The wish has been expressed for a long time to be able to encode
>>> uncertain
>>> values in SBML models. One particular need is for population models,
>>> and
>>> this will be cater for by the "distrib" package. This package will
>>> allow to
>>> define and sample from "distributions" of values, either using
>>> mathematical
>>> functions, or list of numbers. There is another situation, where we
>>> just
>>> want to store uncertain values. For instance because this is how it
>>> was
>>> reported in the literature and we use SBML as a mean to store
>>> knowledge
>>> about biochemical pathways. Or because we want to use the SBML file
>>> as an
>>> intermediate in a workflow, the final values being determine at a
>>> later
>>> stage.
>>>
>>> We need to be able to encode at lease 4 information:
>>>
>>> * the arithmetic mean
>>> * the standard deviation
>>> * a lower value
>>> * an upper value
>>>
>>> This information of course is mostly a form of annotation. If you do
>>> not
>>> know the distribution of the value, you cannot do much in terms of
>>> mathematical interpretation of the model. And if the distribution is
>>> available, then the distrib package should be used. So the role of
>> the
>>> corresponding structures would be rather like the "reversible"
>>> attribute on
>>> the reaction element.
>>>
>>> After playing with the idea of using regular distributions in the
>>> distrib
>>> package, with things like "first moment" for the mean and "second
>>> moment"
>>> for the stddev, it seems - after discussions at HARMONY2012 - that
>> the
>>> easiest way would rather to just add three optional attributes to
>> any
>>> element carrying a value (the value being identical to the mean if
>> the
>>> attribute stdev is defined).
>>>
>>> The big question is the package to which those attribute should
>>> belong to:
>>> 1) core (remember those are optional attributes)?
>>> 2) distrib?
>>> 3) another package? (at some point it was suggested to develop a
>> stat
>>> package, for instance to extend the MathML).
>>>
>>> All solutions have their appeal, and all solutions have their
>>> drawbacks.
>>>
>>> 1) is harder to implement due to the stiffness of the core, and the
>>> perception that the burden will be high on all shoulders.
>>> 2) will tie the attributes to distrib, and therefore they will only
>> be
>>> accessible to tools understanding distrib.
>>> 3) will force the creation of a new package, with all the management
>>> hassle, but would make easier future independent evolutions (i.e.
>>> adding
>>> median).
>>>
>>> Opinions? Ideas?
>>>
>>> --
>>> Nicolas LE NOVERE, Computational Systems Neurobiology, EMBL-EBI,
>> WTGC,
>>> Hinxton CB101SD UK, Mob:+447833147074, Tel:+441223494521 Fax:468,
>>> lenov@ebi.ac.uk, Skype:n.lenovere, twitter:@lenovere
>>> http://www.ebi.ac.uk/~lenov/, http://www.ebi.ac.uk/compneur/
>>>
>>>
>>>
>>> ____________________________________________________________
>>> To manage your sbml-discuss list subscription, visit
>>> https://utils.its.caltech.edu/mailman/listinfo/sbml-discuss
>>>
>>> For a web interface to the sbml-discuss mailing list, visit
>>> http://sbml.org/Forums/
>>>
>>> For questions or feedback about the sbml-discuss list,
>>> contact sbml-team@caltech.edu
>> ____________________________________________________________
>> To manage your sbml-discuss list subscription, visit
>> https://utils.its.caltech.edu/mailman/listinfo/sbml-discuss
>>
>> For a web interface to the sbml-discuss mailing list, visit
>> http://sbml.org/Forums/
>>
>> For questions or feedback about the sbml-discuss list,
>> contact sbml-team@caltech.edu
>> ____________________________________________________________
>> To manage your sbml-discuss list subscription, visit
>> https://utils.its.caltech.edu/mailman/listinfo/sbml-discuss
>>
>> For a web interface to the sbml-discuss mailing list, visit
>> http://sbml.org/Forums/
>>
>> For questions or feedback about the sbml-discuss list,
>> contact sbml-team@caltech.edu
>>
>

____________________________________________________________
To manage your sbml-discuss list subscription, visit
https://utils.its.caltech.edu/mailman/listinfo/sbml-discuss

For a web interface to the sbml-discuss mailing list, visit
http://sbml.org/Forums/

For questions or feedback about the sbml-discuss list,
contact sbml-team@caltech.edu