R · Quarto · Shiny · CUNY DATA 607

Housing Affordability
Dynamics in New York State

A multi-source analysis integrating Census ACS, FRED macroeconomic data, and HUD Fair Market Rent benchmarks to investigate what drives housing burden across all 62 New York counties — with a novel application of machine unlearning to assess how 2022–2024 interest rate shocks shaped predictive models.

R Quarto Shiny Machine Learning Machine Unlearning Housing Policy CUNY DATA 607

Language

R · Quarto

Data Sources

Census ACS · FRED · HUD FMR

Models

Linear Regression · Random Forest

Course

CUNY DATA 607 Final Project

Status

Published · Live

Links

RPubs Report ↗ Shiny App ↗ GitHub Repo ↗

01About

This is the final project for CUNY's DATA 607 — but the research question came from genuine curiosity, not a rubric. Housing affordability is one of the defining pressures on New York residents, and I wanted to understand it rigorously: what actually drives the gap between what people earn and what they pay to live here?

The project integrates three heterogeneous government data sources — Census ACS 5-Year estimates for 62 counties (2009–2024), FRED API data for mortgage rates and inflation, and HUD Fair Market Rent schedules as a policy-grounded rental benchmark — into a single unified panel dataset built from scratch via API calls and GitHub-hosted CSVs.

The workflow follows an OSEMN framework (Obtain, Scrub, Explore, Model, Interpret) — progressing from raw API ingestion and cleaning, through exploratory visualization, into predictive modeling with tidymodels and a novel experiment in machine unlearning.

The machine unlearning component — selectively removing 2022–2024 high-interest-rate observations by zeroing out case weights — was the methodological centerpiece. It let me ask: how much of what these models "know" is rate-cycle specific versus structurally stable? That question has real implications for how HUD FMR benchmarks should be adjusted across rate-cycle transitions.

02Research Questions

Two questions structure the modeling work — one about prediction, one about sensitivity. Both are answered through the same unified county-year panel of 62 New York counties from 2009 to 2024.

RQ 1

What structural and macroeconomic factors best predict a county's housing burden?

R² Local structural factors — home prices, population density, and regional rent levels — explain approximately 96% of housing burden variation across counties.
↓ Macroeconomic variables such as mortgage rates and the Fed Funds Rate contributed minimally once local conditions were accounted for.
🌲 Random Forest outperformed Linear Regression on held-out data, confirming nonlinear interactions between geographic and economic variables.

RQ 2

How sensitive are HUD FMR predictions to removal of high-interest-rate years (2022–2024)?

$ Machine unlearning produced a mean statewide prediction shift of −$70/month — modest but real influence of the rate-hike era on what models learned.
↕ Sensitivity was heavily concentrated in high-cost downstate counties. Rural and mid-size upstate markets were largely unaffected, reflecting lower rate-cycle sensitivity.
✦ Both models maintained high R² post-unlearning, confirming FMR is primarily driven by structural county factors, not exclusively by interest rate environment.

03Interactive App

Live Shiny Dashboard

Explore housing burden by county, year, and income group. Interact with the maps, filters, and model outputs directly.

Open in new tab ↗

04What I practiced

01 Multi-API data acquisition — Pulled county-level ACS 5-year data via the Census API, macroeconomic indicators from the FRED API with custom retry logic, and HUD FMR schedules from GitHub-hosted CSVs — all integrated into one panel dataset from scratch.
02 Predictive modeling with tidymodels — Built Linear Regression and Random Forest pipelines using recipes, workflows, and cross-validation. Used the {vip} package to interpret variable importance across both models.
03 Machine unlearning — Implemented selective case weight reduction (1.00 → 0.00) on 2022–2024 observations to simulate removal of rate-shock era data and measure counterfactual prediction shifts across all 62 counties.
04 Geospatial analysis — Mapped county-level housing burden and FMR sensitivity using {leaflet}, {tigris}, and {sf}, including choropleth views normalized by population to avoid density bias.
05 Shiny app development — Built a multi-tab interactive dashboard deploying the model outputs and exploratory views, hosted live on ShinyApps.io for public access.
06 Quarto publishing — Documented the full OSEMN workflow in a reproducible Quarto document with embedded R outputs, published to RPubs as a navigable HTML report.

Language	R · Quarto
Data	tidycensus, fredr, httr, jsonlite
Wrangling	tidyverse, lubridate, zoo, dplyr
Modeling	tidymodels, vip, randomForest
Spatial	leaflet, tigris, sf
Viz	ggplot2, patchwork, scales, gt
App	Shiny · ShinyApps.io
Publish	RPubs · GitHub
Dataset	Census ACS 5-Yr · FRED · HUD FMR (62 NY counties, 2009–2024)

Explore the full project

Published report, interactive app, and all source code are publicly available.

RPubs Report ↗ Shiny App ↗ GitHub ↗ All Projects

Housing Affordability Dynamics in New York State

Explore the full project

Housing Affordability
Dynamics in New York State