Open ld-archer opened 2 years ago
First step to this (before doing any of the replacement) is to assess what difference we can see in outcomes for each quintile. This will help to assess how viable this line of questioning is, and also to give us a framework for assessing our intervention in the future. We therefore need to make sure the wealth_quint
variable is properly defined in the model as well as being able to look at quintiles as a subgroup in the outputs.
wealth_quint
to Vars.cpp & Vars.hVisualising the baseline outputs by wealth quintile has worked out quite nicely:
Going to add a step here and define different variables for any severe condition (acutely life-threatening i.e. Cancer, Stroke) and mild conditions (i.e. Diabetes, Arthritis). This decision is based on a similar idea from this paper: Michaud & van Soest 2008, Health and wealth of elderly couples: Causality tests using dynamic panel data models. Will then visualise.
Can now start looking into IPF for copying the health and risk behaviour status of wealthier quintiles onto the lowest wealth group.
First attempt at this intervention is up and running. Most of the work has been done in the FEM_R/Wealth_Health_Socioeconomics.Rmd
file, which so far mainly consists of replacing the risk behaviour distribution of the poorest quintile (1) with that of the median (3). This will be expanded soon to produce populations that replace more than just risk behaviours (health and disability status eventually) as well as work for different quintiles. Will collect the logic in a function first and generalise it to make this easier.
Just to note on the method, originally planned to use IPF but when looking at the practicalities of this it didn't seem like the best option for a first pass. The main problems foreseen with IPF is that the seed population would have to have been aggregated, which would need to be individualised again after modifying the totals which I don't think would have been a trivial task. Instead using Multiple Imputation for now and planning to improve or expand later on after testing these first few iterations a bit.
Have created 3 different replacement scenarios:
risk_quint3
: Swapped risk behaviour info from quint 3 (median) onto quint 1 (poorest)risk_quint4
: Swapped risk behaviour info from quint 4 onto quint 1rdh_quint4
: Swapped risk behaviour, disability status, and health status from quint 4 onto quint 1All three show nice effects in both survival and risk behaviour, to differing levels. See the R document for the code to generate the replaced populations and visualisations of differences in prevalence of things like survival, anydisease, disability, severeCondition etc. (couldn't upload)
Best idea yet that avoids a lot of the issues around the FEM causal pathway and the difficulty in predicting wealth or income is to do a 'replacement counterfactual'.
This is where we divide our replenishing population into quintiles based on wealth, then 'replace' or impute the risk behaviour and health status of one of the wealthier quintiles onto the poorest quintile, and simulate the population. We are assuming here that wealth is related to all of these characteristics, and so by changing them to look like the wealthier groups we are effectively intervening on wealth without having to use that explicitly as an input to our transition models (which is complicated given the 'causal pathway' of our model - risk behaviours -> chronic disease & disability -> economic outcomes).
This has to be a cohort simulation as trying to scale this up to the whole population is much much more complex.
First steps: