VEP associated to statistics

[bkhelifi]

Dear all,

Before starting the discussion into the Semantics WG, here is the VEP UCD for statistical quantities. The proposals are:

stat.distribution

A VEP was proposed in 2021 about stat.histogram: https://voparis-gitlab.obspm.fr/vespa/ivoa-standards/semantics/vep-ucd/-/blob/master/VEP-UCD-005.txt.
The latest proposal was to use stat.distribution. I propose that we support this proposal of addition, as it is also our.

** stat.error.negative and stat.error.positive

Combination of UCDs can sometimes explain a quantity. With the current list of UCDs (https://ivoa.net/documents/UCD1+/20241218/index.html), the closest combination could be stat.error;stat.min and stat.error;stat.max.
However, I think that it does not bring the meaning of a negative and positive error.
I propose that we do a VEP UCD for these new terms, except if objections.

** stat.lowerlimit and stat.upperlimit
With the current list of UCDs, we have not found a combination to describer upper and lower limit. The main reason is that without them, one can not publish at the same time a 'stat.minand astat.lowerlimit`, because of a lack of unambiguous definition.

Except if someone finds one, we propose:

• the addition of stat.lowerlimit and stat.upperlimit
• the modification of the definition of stat.min and 'stat.max`, from "Minimum or lower limit" (resp. "Maximum or upper limit") to "Minimum" (resp. "Maximum")

stat.confidenceLevel

This UCD is already an official one (see https://ivoa.net/documents/UCD1+/20241218/index.html). No action is required, except to modify our note
With it, one can define the confidence level used for the error computation (e.g. stat.confidenceLevel;stat.error) and the lower/upper limits (e.g. stat.confidenceLevel;stat.lowerlimit).

[loumir]

stat.distribution

A VEP was proposed in 2021 about stat.histogram: https://voparis-gitlab.obspm.fr/vespa/ivoa-standards/semantics/vep-ucd/-/blob/master/VEP-UCD-005.txt.

this proposal for a UCD was to encode a full distribution into one column of a table.
the draw back is that you should also have other metadata about this distribution , like the nb of slots, the nature of the quantity , etc.
This is the main reason why this was not adopted straightforward.

but may be a term like this would be appropriate to designate a pdf data product :
stat_distribution_pdf instead of pdf would stress the fact that it deals with statistical behavior of the observed data.
stat_distribution_draws instead of draws would also work .

[loumir]

This UCD is already an official one (see https://ivoa.net/documents/UCD1+/20241218/index.html). No action is required, except to modify our note
With it, one can define the confidence level used for the error computation (e.g. stat.confidenceLevel;stat.error) and the lower/upper limits (e.g. stat.confidenceLevel;stat.lowerlimit).

These two terms have a P position tag and cannot be attached this way.
What is the quantity whose confidence level is measured exactly ?

is there a dataset example available ?

[iannevans]

Dear all,

Before starting the discussion into the Semantics WG, here is the VEP UCD for statistical quantities. The proposals are:

The link to the proposed VEP is missing.

stat.distribution

A VEP was proposed in 2021 about stat.histogram: https://voparis-gitlab.obspm.fr/vespa/ivoa-standards/semantics/vep-ucd/-/blob/master/VEP-UCD-005.txt. The latest proposal was to use stat.distribution. I propose that we support this proposal of addition, as it is also our.

I'm unclear what you are suggesting here. One comment in response to the stat.histogram proposal was to use stat.distribution as a more generic term with stat.distribution.histogram being more specific. Something like that might allow a named distribution (e.g., Gaussian, Poisson, Exponential) that will apply to many common cases, as well as expansion to provide a histogram. The latter could for example be a pdf data product.

** stat.error.negative and stat.error.positive

Combination of UCDs can sometimes explain a quantity. With the current list of UCDs (https://ivoa.net/documents/UCD1+/20241218/index.html), the closest combination could be stat.error;stat.min and stat.error;stat.max. However, I think that it does not bring the meaning of a negative and positive error. I propose that we do a VEP UCD for these new terms, except if objections.

** stat.lowerlimit and stat.upperlimit With the current list of UCDs, we have not found a combination to describer upper and lower limit. The main reason is that without them, one can not publish at the same time a 'stat.minand astat.lowerlimit`, because of a lack of unambiguous definition.

Except if someone finds one, we propose:

• the addition of stat.lowerlimit and stat.upperlimit
• the modification of the definition of stat.min and 'stat.max`, from "Minimum or lower limit" (resp. "Maximum or upper limit") to "Minimum" (resp. "Maximum")

stat.confidenceLevel

This UCD is already an official one (see https://ivoa.net/documents/UCD1+/20241218/index.html). No action is required, except to modify our note With it, one can define the confidence level used for the error computation (e.g. stat.confidenceLevel;stat.error) and the lower/upper limits (e.g. stat.confidenceLevel;stat.lowerlimit).

The note already states that the existing UCD stat.confidenceLevel should be used to identify the confidence level for confidence limits, so I don't think a change is needed.

If you want to do (e.g.) stat.confidenceLevel;stat.lowerlimit we would need to modify the stat.lowerlimit proposal from tag P to tag Q. Unfortunately, stat.error is already defined with tag P so stat.confidenceLevel;stat.error would not be legal.

[iannevans]

This UCD is already an official one (see https://ivoa.net/documents/UCD1+/20241218/index.html). No action is required, except to modify our note
With it, one can define the confidence level used for the error computation (e.g. stat.confidenceLevel;stat.error) and the lower/upper limits (e.g. stat.confidenceLevel;stat.lowerlimit).

These two terms have a P position tag and cannot be attached this way. What is the quantity whose confidence level is measured exactly ?

is there a dataset example available ?

I think you might want to do (e.g.) stat.error;stat.confidenceLevel or stat.error.positive;stat.confidenceLevel and stat.error.negative;stat.confidenceLevel in order to define what confidence level is associated with reported confidence limits. For example, 2.345 +1.234 (90% confidence) / -2.342 (95% confidence). Typically we don't report confidence intervals that have different positive and negative confidence levels, but it's not impossible. We typically do report different confidence levels for different quantities. For example, we often have table data products where we report <quantity_n>, <lower confidence limit_n>, <upper confidence limit_n> as separate columns, and the confidence limits will vary depending on n.

For lower and upper limits, the situation is more confused. Some folks consider a lower or upper limit to be just another confidence level (typically a more "extreme" confidence level) whereas other might consider an upper limit in terms of detectability , which is more complex because then there must be associated type 1 and type 2 error thresholds associated with the detection limit.

Perhaps the easiest solution would be to revise stat.confidenceLevel to have tag Q?

[bkhelifi]

Hi Ian,
Associated to your last reply, Gammapy proposes e.g. a FluxPoints object containing the err, err_n, err_p, ul and the pdf (see https://docs.gammapy.org/dev/api/gammapy.estimators.FluxPoints.html). And the same for the FluxMaps (see https://docs.gammapy.org/dev/api/gammapy.estimators.FluxMaps.html).

About:

If you want to do (e.g.) stat.confidenceLevel;stat.lowerlimit we would need to modify the stat.lowerlimit proposal from tag P to tag Q. Unfortunately, stat.error is already defined with tag P so stat.confidenceLevel;stat.error would not be legal.

this is a very fair point. One should modify the tags! Mireille willl do a proposal