Thomas Beale wrote:
> Colin Sutton wrote:
>> The 'coding' is surely 'Accuracy' ('Measurement' has 'Accuracy') where
>> this can be None|~|Unknown|Percentage(value)!SD(Distribution type,value)
>> which would cover any measurement (e.g.height,heart rate), not just
>> pathology lab values
>>
> this seems pretty close to a correct model. Slight corrections I would
> suggest are:
> - I am still uncomfortable with '~', since it seems to mean
> "approximate", but "we don't know how approximate"...
> - does "None" mean a) none recorded (i.e. don't know, i.e. same as '~')
> or b) no accuracy, i.e. an exact value (reasonable for some things, e.g.
> the answer to the question "number of previous pregnancies")?
> - in the case of a statistical distribution, one value may not be enough
> to characterise the limits, since the distribution may be asymmetric (I
> don't remember enough beyond normal/T/Chi2 to remember if there are
> distributions that need even more parameters).
In terms of statistical confidence limits/intervals, the parameters are:
the type of limits/interval (frequentist "confidence interval" or
Bayesian "credible interval", the confidence value (typically 95%, but
often not), and the underlying assumed *error* distribution (normal,
Poisson, Student's T, Weibull etc etc).
However, confidence intervals/limits don't indicate where in a
population distribution a particular value lies - quantiles are more
often used for this - the actual quantile of the value (eg for growth
measurements read against a normogram), or the values of the quartiles
or 5th and 95th percentile, or variations on that.
> The question for us in openEHR is how much to implement of such a model:
> we have to be driven by real use cases.
If you really want to nail this problem, a workshop involving a range of
people (from lab scientists to pathologists to clinicians to
epidemiologists and biostatisticians) is required, I think. It could be
in the form of a virtual workshop via email, but you really need to
gather together a diverse group, state the problem/s to be solved (eg
lab values only, or physical measures only, or to include other things
like measures derived from psychological scales or population or study
measures like odds ratios and relative risks or age-standardised
rates?), and get them to generate use cases and explore the issues.
Happy to be involved, but as an epidemiologist, I'd feel more
comfortable if some mathematical statisticians and some lab scientists
were involved too.
Tim C
