New Variables

[ld-archer]

For Future:

• [x] Indicator of loneliness: predictor for psyche
  • r[2-8]lnlys - loneliness summary mean score
  • Summary of 4 vars, see p719 of Harmonized ELSA docs
• [ ] Family/relationship struggles: predictor for psyche
• [x] Total family wealth: predictor for many
  • h*atotb - total all assets inc. 2nd home
• [ ] Rename exercise vars and exstat to be based on activity (more accurate than saying exercise)
• [x] Alzheimers: predictor of mortality
  • r*alzhe
• [x] Dementia: predictor of mortality
  • r*demene
• [x] Unemployed: predictor for smoking, drinking, psyche(?), bmi(?)
  • Look at labor force status: r*lbrf_e
• [x] Cohabiting but not married: predictor for smoke_start
  • Look at partnership var: r*part
• [ ] Job Type: smoking, drinking, wealth/earnings
  • r*soc2000
• [ ] Whether owns home: drinking, smoking, financials
  • h*ahown

_Originally posted by @ld-archer in https://github.com/ld-archer/E_FEM/issues/42#issuecomment-730249663_

[ld-archer]

Loneliness Loneliness in ELSA is assessed in 4 ways:

• Whether R lacks companionship
• Whether R often feels left out
• Whether R often feels isolated from others
• Whether R often feels not in tune with other people

The average summary score of the responses to these 4 questions is then reported in r[2-8]lnlys. There is also a 3-item loneliness summary mean (r[2-8]lnlys3), which doesn't include the question of feeling in tune with other people. Not sure which is better, will probably end up including the one with the least missing data.

Update Going with 4 item score, as it has slightly more complete data. Both vars are ordinal, and contain non-integer values from 1 to 3. I think this will require categorising before we can use this for prediction. Will need to round the values to whole numbers and assign dummys for prediction models.

Minimal Anything that is missing in wave 1 causes trouble for the minimal population, as it is derived from people in wave 1. Previously, this hasn't been a big issue, and I have just done some crude imputation (e.g. setting all values to 0), but this isn't really any good. Instead, I'm going to try and copy the values from wave 2 onto wave 1 for these specific variables only.

DONE

[ld-archer]

Unemployed There is a specific variable for this, so won't need to use or abuse the labour force status var: r*unemp

There is a special missing code for those who are not in the labour force (i.e. retired/disabled/not working or trying to find work) which could be useful for sample selections, otherwise there is also a binary var for 'in the labo(u)r force'. When including and recoding the job type vars, it might make sense to replace the current logic around working status and retirement. Cross that bridge when we come to it.

• [x] Implemented and testing Lots of missing data but had to be a bit clever about it due to how the variable is organised in ELSA. I wanted to keep the people who were not in the labour force as missing, but impute those who were missing for other reasons. After fumbling about a bit I found that those who were plain missing (.), who we wanted to impute, were at least 1 of:

• currently in work

• over age of 65

• retired (or at least retemp == 1, may not be exactly the same thing)

• [x] Done

[ld-archer]

Homeowner This is relatively simple again, just not fully clear on how we can transition this. Or what accounting we need to do to make sure we don't have any logical nonsense in the outputs.

• [ ] reshape_long
• covariate definitions
  • [ ] ELSA
  • [ ] minimal
• [ ] sample selections
• [ ] Vars.cpp/.h
• [ ] settings.csv
• [ ] settings.txt
• [ ] cross-validation_ELSA

[ld-archer]

Going to close this issue and open another one for adding vars that is more specific to our immediate plans for the FEM.