Proposal for representing Aggregate Statistical Data

This document can be accessed here.
<https://docs.google.com/document/d/139jXakeQk4ChwCkGjqq5wJfCPMDnwIV94oCH-JzJrhM/edit?usp=sharing>


Look forward to feedback.

Guha

Representing aggregate statistics


Examples of aggregate statistical reports include those from Census
Organizations (e.g., American Community Survey), Health Organizations
(e.g., CDC Wonder) and many others. This is a schema, currently in use on
DataCommons.org for representing facts stated in these reports. This
document describes certain general mechanisms for representing statistical
populations and associated observations. This document will be followed
later by a companion proposal suggesting some basic common vocabulary
useful for representing the kind of data released by the US Census, CDC,
etc.

Our interest is not in describing a data set or mapping columns in csv
files, but in representing the actual data itself. Other efforts have
focused on characterizing data cubes in terms of dimensions, etc. While we
draw upon their work, our goals are different.

Examples of the kind of statistics we would like to represent include:

1. In 2016, there were 1213 people in East Podunk, California, who were
male, married, with a median age of 22.
2. In 2017, there were 20 deaths in Falooda County where the cause of death
was XYZ

We will refer to ‘number of people who are male, hispanic’, ‘number of
deaths where cause of death was XYZ’, etc. as variables. Since the number
of possible variables increases combinatorially, clearly, we can’t have a
properties for each variable (or worse, property for each variable x
years). We need a way of compositional way of constructing variable
references. We use the concept of a StatisticalPopulation to do this
construction.

A StatisticalPopulation is a set of instances of a certain given type that
satisfy some set of constraints. The property populationType is used
specify the type. Any property that can be used on instances of that type
can appear on the statistical population. An instance of
StatisticalPopulation whose populationType is C1, which has the properties
p1, p2, … with values v1, v2, … corresponds to the set of objects of type
C1 what have the property p1 with value v1, property p2 with value v2, etc.
The properties numConstraints and constrainingProperties are used to
specify which of the populations properties are used to specify the
population. In the two examples above:


Node: SP1
type: StatisticalPopulation
populationType: Person
location: EastPodunkCalifornia
gender: Male
maritalStatus: Married
numConstraints: 3
constrainingProperties: location, gender, race


Node: SP2
type: StatisticalPopulation
populationType: MortalityEvent
location:   FaloodaCounty
causeOfDeath: XYZ
numConstraints: 2
constrainingProperties: location, causeOfDeath


SP1 is an abstract set in the sense that it does not correspond to a
particular set of people who satisfy that constraint at a certain point in
time, but rather, to an abstract specification, about which we can make
observations that are grounded at a particular point in time. We now turn
our attention to the representation of these observations.

 Instances of the class Observation are used to specify observations about
an entity (which may or may not be an instance of a StatisticalPopulation),
at a particular time. The principal properties of an Observation are
observedNode, measuredProperty, measuredValue (or median, etc.) and
observationDate (measuredProperty can, but need not always, be w3c rdf data
cube "measure properties", as in lifeExpectancy example here:
https://www.w3.org/TR/vocab-data-cube/#dsd-example.) In the two examples
above:


Node: Obs1
type: Observation
observedNode: SP1
measuredProperty: age
median: “23 years”
observationDate: “2016”

Node: Obs2
type: Observation
observedNode: SP1
measuredProperty: count
measuredValue: 1213
observationDate: “2016”

Node: Obs3
type: Observation
observedNode: SP2
measuredProperty: count
measuredValue: 20
observationDate: “2017”


Observations can also have properties related to the measurement technique,
margin of error, etc. To elaborate on Obs2 above, we can have:

Node: Obs2
type: Observation
observedNode: SP1
measuredProperty: count
measuredValue: 1213
observationDate: “2016”
marginOfError: 22
measurementMethod: CensusACS5yrSurvey


Notes:
1. Care needs to be exercised when querying StatisticalPopulations, to make
sure that the query specifies all the constraining properties.
2. We do not yet have a way of using properties which are named in the
opposite direction e.g. we handle "alumniOf" (relating a person to an org),
but if the only existing property was "alumni" (relating an org to a
person).