Copyright © 2023 Darren Erik Vengroff

This project generates impact charts from the dataset constructed by the rihdata project. It does so using the impactcharts project, which it turn relies on SHAP.

Before running the analysis is this project, it is necessary to download the data using the rihdata project. Normally, rihdata and rihcharts are installed locally under the same parent directory. If they are not, a small local modification may be needed for things to work.

All of the analysis this project does is orchestrated by a single Makefile. It requires GNU make version 4.3 or higher.

Before you run GNU make, you will have to set up a Python virtual environment with all the project's dependencies. I use poetry to do this. All top-level dependencies are listed in pyproject.toml and the full set of recursive dependencies and version is stored in poetry.lock.

Once you have your virtual environment and the right version of GNU Make installed, the single command

gmake [-j 8]

should build the entire project. The -j argument is optional. If given, it tells make that it can parallelize the operation eight ways across multiple CPU cores. If you have multiple cores, you can choose an appropriate -j number to use them all.

The steps the Makefile will execute are:

1. Fit and hyperparameter tune an XGBoost model for each CBSA that predicts median housing prices based on median income and racial demographics.
2. Fit a series of linear models for the same problem.
3. Generate SHAP values and plots for a large number of individual models and their ensemble for each CBSA and demographic feature.

How long this will take depends on how fast your machine is, how many cores you have and choose to use, your internet connection are. For me, running on 8 eight cores on an M1 MacBook Pro from a clean state it takes less than an hour to complete.

For more details and options, please consult the Makefile.

The results will be a series of plots in a newly created directory called plots-2021/shap/xgb. There is a subdirectory for each CBSA, for example for Atlanta it is Atlanta-Sandy_Springs-Alpharetta,_GA_Metro_Area/12060. The number 12060 is the FIPS code for the CBSA. It will be different for each one.

Within this directory are the impact charts. Each chart shows the impact of the fraction of the population belonging to one racial or ethnic group on median home prices. This is the impact of that variable alone, corrected for the impact of median household income. For example the impact on the fraction of Black residents looks like this

The basic idea of this chart is that neighborhoods that have almost no Black residents at all have higher median home prices than those that have large numbers of Black residents. We can see from the shape of the curve just how much Blackness in a neighborhood creates what level of effect.

Median household income's impact is in a final chart in the subdirectory. For Atlanta, it looks like this

In this case we see that higher income areas tend to have higher median home prices, and that except for very low income areas this affect looks almost linear.

It is important to realize that impact charts are built using Machine Learning models along with an approach called SHAP to specifically isolate the effects of one variable from another.

Taken together, these two charts show us that whether there is any correlation between Blackness and income or not.

This project produces hundreds of charts like this, and they are all a little different, and show different effects in different areas.

Like in rihdata, we can use choose to do the analysis for the top N for N other than the default of 50. For example, for 10, we could run

gmake -j 8 N=10

Before running this, we must be sure that we explicity ran

gmake N=10

in the rihdata project first to ensure we have all the data and metadata we need.