Align predicted educational level with observed outcomes

[andrewbaxter439]

Creating as an issue to document changes...

There's some concern that highest-educational-attainment prediction models seem to under-predict high-attainment at later ages and over-predict medium-attainment:

Insensitivity of some models to predict the higher age groups may be partially down to low numbers of observations in estimate population:

Would suggest having an age-26/27 cutoff, where all leaving (first) education after this are assigned 'High' educational attainment?

[andrewbaxter439]

A possible thing to incorporate would be a model for older ages where 'Low' is no longer a likely outcome (>20 or >25?). A cutoff age could apply a simpler model for these years:

[andrewbaxter439]

• Interact age under/over 22 with only second-order polynomials
• see if keeping folks in education more is possible - low numbers of students compared to observed population
• Why are low values going up in simulation?
• Check whether validation data is valid - 10 percentage-points increase across the years seems strange?

(added @dav-sonn as discussed on 07/10)

[andrewbaxter439]

Further note - this seems now to be aligning better with validation data, though still some way to go:

@dav-sonn and @Mariia-Var - revisiting calculation methods to line things up and the low education is no longer a problem I think! Updating E2a with quad terms is a good start, but E1a and E1b may be next things to look at.

[andrewbaxter439]

Two curious further observations.

1. The proportion of the 17-65 population currently "Students" is fairly stable:

2. the mean age of students (again in the 17-65 population) increases linearly (my fixes don't change this):

The gradient is about ~0.34, meaning every three years the average age of current students increases by 1.

[Mariia-Var]

Two curious further observations.

1. The proportion of the 17-65 population currently "Students" is fairly stable:

2. the mean age of students (again in the 17-65 population) increases linearly (my fixes don't change this): The gradient is about ~0.34, meaning every three years the average age of current students increases by 1.

@andrewbaxter439 ,

I have a suspicion that this can be because some students are not exiting the education process, i.e. stay there forever.
(We used to have this type of problem in the EU models, though education is modelled differently there...)
Could you please tabulate the students data and see if there are some students that do not exit education?

[matteorichiardi]

One way to check whether your hypothesis is true is to create an ad hoc variable with the time in education, and put a break if the value exceeds 1 for individuals above max education age. Remember that if individuals go back to education and they are below max education age they are treated like first-time students, so they can remain in education until max education age. If they go back to education after max education age they should be stay only one year (training course). So:

timeIneducation = timeInEducation ++
if (timeInEducation > 1 & age > maxEduAge) {
// insert break point here
}

Another way is to put the break point when individuals return to education, and then follow one instance of those individuals.

A third way is to add a validation graph with the proportion of students by age. If individuals never exit education if they enter above max education age, this proportion should be increasing.

A fourth way is to inspect the method that forces individuals to exit education if they are above max education age. If there is any error, it should be easy to spot...

All the above are not mutually exclusive, of course.

Matteo