Obesity investigation

Inspired by the observation that obesity seems to decline in prevalence with increasing altitude in the USA, is the same true in Australia?

Data

I found aust_health_tracker_data_lga.xlsx at https://atlasesaustralia.com.au/ahpc/

This has obesity-related information.

Altitude information is fetched from wikidata.

Economic indicators are fetched from the ABS.

Naive approach

The naive approach -- just correlate altitude with obesity percentages -- suggests not.

How to reproduce it:

• Run altitude.sparql on query.wikidata.org This should produce altitudes.csv

• Run Join helath stats with altitude.ipynb . About half-way down, it creates chart-ready-data.csv

• Run analysis-charts.twb in Tableau.

I was exploring a little, not being overly scientific about this, so there are some weird charts in Tableau as well.

More sophisticated approach

The second half of Join health stats with altitude.ipynb fetches economic indicators.

There's a small bug in it where if it fetches a file-not-found page from the ABS it just carries on anyway without saving valid data. I deleted the invalid files by hand.

The method is:

• Remove numeric counts from the source data; rely on percentages instead.

• Do robust scaling to put all source information on a standard scale.

• Create a lasso model for each of the targets (obesity/overweight/gender/age range); use cross-validation to find the appropriate alpha value to maximise the accuracy of the model.

• Read off the coefficients from the lasso model in decreasing absolute value.

This kind of model should cope with collinearity -- if there are two highly collinear data columns, one will be given a zero coefficient, and the one with greater relevant variability should end up with an appropriate coefficient.