I already posted last week about the modeling techniques I plan to
use, so this week, I am going to focus on the value of Tableau to do
sophisticated data analysis in a relatively simple way.
I used Tableau to generate something like 20 candidate data
visualizations for the paper. It sounds like a lot (and it was), but it was
relatively fast to do, and Tableau integrates some basic statistics that
can help uncover what I like to think of as "visual clues" to where and
how to focus the models later on. The scatterplot matrix below is a
good example of how Tableau improves on a simple, but powerful
visualization. Scatterplot matrices can be used to quickly and
efficiently evaluate a range of bivariate relationships all at once.
However, where there are close to 30,000 observations, the graph
can be a little "blobby," as below. Tableau includes a trendline
function, though, and it's easier to apply, remove, or modify than it
would be in R, in my opinion (prettier, too). Moreover, you can pan
over it and get an R-square value that tells you how good the trend
line actually is.
Tableau also makes it easy to investigate different types of
relationships: linear, logarithmic, exponential, polynomial or power.
In the chart below, I started out with a linear trend line, and it was
okay, but not great. R values between 0.8 and 0.9 or so. When I
switched to a polynomial (factor 3) trendline, it matched far more
closely, had fewer points outside the confidence bands, and, most
importantly, delivered a useful insight. You can see from the gold
points and line that unemployed people who lived in Medicaid
expansion jurisdictions were able to pretty quickly get their coverage.
In particular, the data suggest that public coverage rates rise most
sharply in the first four or five years of expansion. This makes sense
because there is probably a natural saturation point, particularly
when the unemployment level is relatively stable, so the proportion
of the population that is unemployed is not very different from year
to year.
